{"name":"Blits.ai AI Use Case Library","license":"https://creativecommons.org/licenses/by/4.0/","citation":"Blits.ai AI Use Case Library, https://www.blits.ai/ai-use-cases (CC BY 4.0)","generatedAt":"2026-09-28","useCases":[{"title":"Agentic AI for autonomous, intent based network operations","shortTitle":"Autonomous network operations","seoTitle":"Autonomous network operations with agentic AI","metaDescription":"AI agents run closed loops over telecom networks within guardrails set by engineers. Deutsche Telekom cut the time to manage major events by more than 95%.","definition":"AI agents that run closed loops over a telecom network: they take an intent from the operator (for example a latency or availability target for a service), observe the network, diagnose deviations and execute corrective actions across radio, transport and core, within guardrails set by engineers and with human approval for major changes.","aliases":["autonomous networks","intent based networking","self healing networks","zero touch network operations","agentic network operations"],"industries":["telecommunications"],"functions":["network-operations","it-and-engineering"],"patterns":["agentic-workflow","anomaly-detection","prediction-and-scoring","classification-and-routing"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"network","problem":"Telecom networks have become too complex to run by hand. A 5G network mixes several radio\nlayers, cloud native core functions, transport, many vendors and new services such as network\nslices with their own performance promises. Vodafone notes that manual tuning and step by step\nautomation scripts are no longer enough as networks grow more complex, and Deutsche Telekom puts it\nplainly: traditional rule based automation falls short in addressing real time challenges.\n\nThe result is that engineers spend their time on repeated manual checks, tuning and firefighting.\nWhen a large event fills an area or a server fails, the right fix is often known, but applying it\ndepends on someone noticing and acting across several vendor tools that do not share data, a\nbarrier Telstra describes in its own network. Vodafone, Google Cloud and TM Forum describe the goal\nas a shift from manual, reactive operations to intent based autonomy, where operators state the\noutcome and the network works out and executes the actions.","problemStats":[],"howItWorks":"1. **Express the intent.** Operators set target outcomes (availability, latency, throughput,\n   energy) per service, area or customer, instead of individual configuration steps.\n2. **Observe.** Agents continuously collect performance, alarm, inventory and external data, such\n   as public event listings, and build a live view of each service across domains.\n3. **Analyse and decide.** Specialised agents detect deviations or predict them, diagnose the likely\n   cause, and select actions, often testing them first in a digital twin.\n4. **Act within guardrails.** Low risk, reversible actions (reallocating resources, adjusting\n   parameters, moving workloads to healthy hardware) run automatically; major changes go to an\n   engineer for approval.\n5. **Document and learn.** Every action and its outcome is logged, explained and used to improve\n   future decisions, and the closed loop checks that the intent is met again.","valueDrivers":["speed","customer-experience","cost-to-serve","employee-productivity"],"kpis":["processing-time-reduction","automation-rate","interactions-handled","cost-reduction"],"indicativeValue":{"referenceOrg":"A mobile operator with 300 staff in network operations and optimisation","inputs":[{"key":"staff","label":"Staff in network operations and optimisation","low":300,"high":300,"unit":"full time employees","note":"The reference operator."},{"key":"costPerFte","label":"Fully loaded cost per employee","low":80000,"high":120000,"unit":"USD per year","note":"Editorial assumption. Replace with your own cost."},{"key":"automatedShare","label":"Share of routine operations work taken over by closed loop automation","low":0.05,"high":0.15,"unit":"fraction of working time","note":"Conservative. Deutsche Telekom reports a more than 95% reduction in the time to manage major events, but that covers one task type, not all operations work."}],"formula":"staff * costPerFte * automatedShare","currency":"USD","period":"per year","resultLabel":"Operations capacity released","caveat":"Staff capacity only. It leaves out faster recovery from incidents, better service level performance, revenue from assured services such as network slices, energy savings, and the substantial investment in data, cloud platforms and integration that autonomy requires."},"macroEstimates":[{"statement":"Vodafone, Google Cloud and TM Forum cite an STL Partners estimate that the potential upside of autonomous networks is circa USD 800 million per operator annually.","sourceTitle":"Vodafone, Google Cloud and TM Forum Unveil Framework for Self-Optimising Autonomous Networks","sourceUrl":"https://www.vodafone.com/news/newsroom/technology/vodafone-google-cloud-tm-forum-unveil-framework-for-self-optimising-autonomous-networks","year":2026}],"feasibility":{"complexity":"high","complexityNote":"Autonomy needs a unified data foundation across vendors and domains, programmable network interfaces, a reliable inventory, a way to test actions safely and a governance model that engineers trust. Most operators get there one closed loop at a time.","dataPrerequisites":["Real time performance, alarm and configuration data across radio, transport and core","Accurate inventory and service topology","History of past incidents, actions and outcomes","External context such as event calendars, weather and planned works"],"integrations":["Network controllers, SON platforms and orchestration systems per domain","Network data lake or data fabric","Digital twin or simulation environment","Trouble ticketing and change management","Service and slice management systems"]},"implementation":{"steps":[{"title":"Pick closed loops with bounded risk","detail":"Start with loops where the action is reversible and the benefit is clear, such as capacity adjustments for planned events or moving workloads away from failed hardware."},{"title":"Build the shared data foundation","detail":"Agents need one view across domains and vendors. Invest in the data layer and inventory before adding more agents."},{"title":"Define intents and guardrails together","detail":"Write down the target outcome, the allowed actions, the limits and the approval rules for each loop, with the engineers who will be accountable."},{"title":"Prove it in shadow and in a twin","detail":"Run agents in recommendation mode and test their actions in a digital twin before letting them act on the live network."},{"title":"Raise autonomy step by step","detail":"Move each loop from recommend to act with approval to act within limits as measured accuracy allows, and keep an easy way to switch back."},{"title":"Coordinate the agents","detail":"As loops multiply, add coordination so agents in different domains do not work against each other, and keep one audit trail."}],"guardrails":["Explicit allow list of actions per agent, with limits and rollback","Human in the loop approval for major network changes and for any action outside the limits","Automatic stop and rollback when service indicators degrade after an action","Every action logged with its reason, data and outcome, and explainable to engineers","Change freezes and maintenance windows enforced for autonomous actions"],"humanInTheLoop":"Engineers define intents, allowed actions and limits, approve major changes and review the audit trail. Autonomy levels are set per loop and raised only on evidence. Operations leaders can pause any agent instantly.","kpisToInstrument":["Time to detect and resolve deviations from intent, per loop","Share of events handled end to end without human action","Actions rolled back and incidents caused by automation","Service level attainment for assured services","Engineer time spent on routine tasks"],"failureModes":[{"title":"Agents that fight each other","detail":"A radio agent and an energy agent undo each other's changes. Coordinate loops and give them shared intents."},{"title":"Autonomy without a trusted data foundation","detail":"Agents act on stale inventory or partial data. Fix data and topology first."},{"title":"Unexplained actions","detail":"Engineers cannot see why an agent acted and stop trusting it. Require explanations and full logs."},{"title":"A small error at network scale","detail":"One wrong automated change is repeated across thousands of elements. Use canaries, rate limits and automatic rollback."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 2 lists AI systems intended as safety components in the management and operation of critical digital infrastructure as high risk; Recital 55 ties this to the digital infrastructure in the Annex to Directive (EU) 2022/2557, which includes providers of public electronic communications networks. Recital 55 defines such safety components as systems that directly protect the physical integrity of the infrastructure or the health and safety of persons and property, and excludes components used solely for cybersecurity. Loops that only optimise performance or capacity are usually not safety components, but a loop that protects physical integrity or life safety services can be, so operators should assess each closed loop and document the outcome."},"regulations":["eu-ai-act","nist-ai-rmf","iso-42001","nis2","eecc"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 2 covers AI systems intended as safety components in the management and operation of critical digital infrastructure."},{"title":"Recital 55, safety components of critical infrastructure","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/recital/55/","note":"Explains which critical digital infrastructure is meant and what counts as a safety component, and excludes components used solely for cybersecurity."},{"title":"NIS2 Directive, securing network and information systems","issuer":"European Commission","region":"europe","url":"https://digital-strategy.ec.europa.eu/en/policies/nis2-directive","note":"Risk management and incident duties for telecom operators that also apply to automated operations."}],"controls":["Inventory of autonomous loops with owners, intents, allowed actions and autonomy levels","Pre deployment testing in a digital twin or staging network","Complete, tamper evident audit trail of agent actions","Kill switch and rollback procedures tested regularly","Periodic review of autonomy levels against measured accuracy"],"incidents":[]},"blitsAi":{"howToBuild":"Closed loops that touch the network run in the operator's orchestration and domain controllers.\nBlits.ai provides the agent layer around them: **agentic workflows** with an agent loop that\ncalls tools, **custom functions** and **MCP servers** that expose controller and ticketing APIs,\na **tool execution policy** that allow lists what each agent may do, and **human in the loop\napproval** for actions above a configurable threshold. **Agentic tasks** implement \"do X when Y\"\nloops with scheduled rechecks and dry run mode.\n\nEngineers talk to the agents through an **AI agent** in **Microsoft Teams** or the **web chat\nwidget** on an internal portal, grounded in a **knowledge base** of runbooks and a **SQL knowledge\nbase** over operational data held in PostgreSQL. Every workflow run has a full audit trail, **test suites** replay scenarios before changes go live, **monitors**\ncheck agent behaviour, and the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"Are autonomous networks real or still a vision?","answer":"Parts are live. Deutsche Telekom's RAN Guardian agent went live in Germany in November 2025 and reduced the time to manage major events from hours to around a minute. Telstra showed a self healing proof of concept that moved network applications away from failed hardware in minutes. Full end to end autonomy across all domains is still ahead."},{"question":"What does intent based mean?","answer":"The operator states the outcome, such as a latency target for a 5G service, and the system works out and executes the actions to achieve and keep it. du and Nokia describe autonomous network slicing that adjusts radio policies to keep premium service levels."},{"question":"Who is accountable when an agent changes the network?","answer":"The operator. Vodafone's framework with Google Cloud and TM Forum stresses policies, explicit guardrails and human in the loop approval for major network changes. Every agent needs an owner, an allow list and an audit trail."}],"related":["network-fault-triage-copilot","predictive-network-maintenance","ran-energy-optimization","network-planning-and-capacity-optimization","aiops-incident-triage"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched for the telecommunications vertical and verified against operator and vendor sources."},{"date":"2026-09-25","note":"Consolidation pass: added NIS2, European Electronic Communications Code to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: problem statement tied to cited Vodafone, Deutsche Telekom and Telstra sources; EU AI Act basis refined with Recital 55 and the critical infrastructure definition; Blits.ai capabilities aligned with the feature inventory; source dates and an archived copy added to evidence; SEO title and meta description added."}],"slug":"autonomous-network-operations","url":"https://www.blits.ai/ai-use-cases/autonomous-network-operations","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":1250050,"min":100,"max":2500000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"telstra-smartfix-proactive-network-fixes","pooled":true},{"id":"deutsche-telekom-ran-guardian-and-mindr-agents","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":95,"min":95,"max":95,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"deutsche-telekom-ran-guardian-and-mindr-agents","pooled":true}]}],"indicativeValueResult":{"low":1200000,"high":5400000},"evidence":["deutsche-telekom-ran-guardian-and-mindr-agents","du-nokia-autonomous-network-slicing","kddi-nokia-performance-degradation-detection","stc-nokia-cognitive-son","telstra-self-healing-network-proof-of-concept","telstra-smartfix-proactive-network-fixes"]},{"title":"AI academic advising assistant for course selection and degree requirements","shortTitle":"Academic advising assistant","seoTitle":"AI academic advising chatbot for students","metaDescription":"Harvard says Student Compass is not meant to replace human advising. Elon calls ElonGPT a supplemental resource for course and requirement questions.","definition":"An AI assistant that answers students' questions about degree requirements, course selection, prerequisites and majors, grounded in the institution's own catalog and advising documents, so students get quick answers to routine questions and are directed to a human advisor for anything that needs judgment, is time sensitive, or falls outside what the assistant can see.","aliases":["academic advising chatbot","AI course selection assistant","degree planning chatbot","virtual academic advisor","major exploration chatbot"],"industries":["education"],"functions":["customer-service"],"patterns":["conversational-agent","rag-knowledge-assistant"],"channels":["web-chat"],"audience":"customer-facing","autonomy":"assist","adoptionStage":"emerging","problem":"Academic advising does not scale well: a large share of what students ask is routine and answerable\nfrom documents that already exist, such as what satisfies a general education requirement, which\ncourses a major requires, or whether a prerequisite has been met. But that information is often\nspread across a catalog, a handbook and several department websites, so students either spend time\nhunting for it themselves or take up an advisor's limited appointment time on a question a document\ncould have answered.\n\nHarvard's Assistant Director Brooks B. Lambert-Sluder wrote that Student Compass \"is not meant to\nreplace human advising, but rather to improve students' access to information that already exists\nonline, and to direct them to appropriate advising resources.\" Elon calls ElonGPT \"a supplemental\nresource to assist you with general advising questions\" and tells students to verify what it says\nagainst the academic catalog or a human advisor. The risk is in where an institution draws that\nline, and how well the tool recognises when a question has crossed it.","problemStats":[],"howItWorks":"1. **Ground the assistant in official documents.** The academic catalog, degree requirement pages,\n   student handbook and relevant department websites are loaded as the assistant's only knowledge\n   source, so it answers from current policy rather than general knowledge about how universities\n   usually work.\n2. **Answer with citations.** The assistant links to the specific source it drew each answer from,\n   the way Harvard's Student Compass does, so a student can verify the answer rather than trust it\n   blindly.\n3. **Route around what it cannot see.** Live, personal data the assistant does not have access to,\n   such as a specific student's transcript, registration status or holds, stays out of scope; the\n   student is directed to the institution's own degree audit or registration tool for that.\n4. **Escalate on cue.** Time sensitive questions, exceptions, and anything nuanced enough that it\n   usually goes to a human advisor in person, triggers a clear handoff to that advisor or advising\n   office rather than a best effort answer.\n5. **Review what it gets wrong.** Advising staff should test the assistant against the questions\n   that actually come up in advising meetings on an ongoing basis, not just at launch, and correct\n   the underlying documents or scope when it gives a wrong or overconfident answer. This matters in\n   practice: The Harvard Crimson's own testing of Student Compass found it handled straightforward,\n   policy grounded questions well but stumbled on more nuanced ones, and Joseph K. Blitzstein, the\n   Statistics department's director of undergraduate studies, separately tested it against real\n   concentration advising questions and found it often gave wrong answers.","valueDrivers":["employee-productivity","customer-experience","inclusion-and-access"],"kpis":["time-saved-per-task","containment-rate","customer-satisfaction","escalation-rate-reduction"],"indicativeValue":{"referenceOrg":"A university with 20,000 undergraduate students","inputs":[{"key":"students","label":"Undergraduate students","low":20000,"high":20000,"unit":"students","note":"The reference university."},{"key":"advisingContactsPerStudent","label":"Advising contacts per student per year","low":2,"high":5,"unit":"contacts per student per year","note":"Editorial assumption, replace with your own advising office contact volume."},{"key":"shareRoutine","label":"Share of contacts that are routine, document answerable questions","low":0.3,"high":0.5,"unit":"fraction of contacts","note":"Editorial assumption, informed by Harvard's Student Compass and Elon's ElonGPT, which both position the assistant for routine, document based questions only, not the judgment based advising that stays with a human. Harvard peer advising fellow Alex I. Draghia told The Harvard Crimson, \"Over 50 percent of the questions I get from the students I'm PAFing are questions that have an absolute answer somewhere on Harvard's website\", an anecdote about questions to peer advising fellows specifically, above the top of this range, not a substitute for it."},{"key":"costPerHumanContact","label":"Cost of a human handled advising contact","low":10,"high":25,"unit":"USD per contact","note":"Editorial assumption for advisor time per routine question. Replace with your own cost."}],"formula":"students * advisingContactsPerStudent * shareRoutine * costPerHumanContact","currency":"USD","period":"per year","resultLabel":"Advising staff time cost avoided on routine questions","caveat":"Gross cost avoided on routine questions only. It leaves out the cost of building and maintaining the assistant, keeping its source documents current, and any drop in quality if students who needed a real advising conversation settle for a chatbot answer instead."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering from a well maintained catalog is straightforward retrieval; the hard part is keeping the source documents current across every department, answering nuanced, concentration specific questions correctly (a Statistics department's director of undergraduate studies tested Student Compass against these and found it often gave wrong answers), and being honest that it cannot see a specific student's transcript or registration status.","dataPrerequisites":["A current, complete academic catalog and degree requirement pages, with an owner per document","A student handbook and relevant department advising pages","A defined escalation policy naming which topics and question types route to a human advisor"],"integrations":["Academic catalog and degree audit system (read only, for grounding content, not personal data)","Advising office scheduling or contact system, for the escalation handoff","Student information system, only if the institution chooses to connect authenticated, per student data rather than keeping the assistant general"]},"implementation":{"steps":[{"title":"Scope it to general advising, not personal data","detail":"Start with catalog and policy questions that are the same for every student, before considering any integration with individual transcript or registration data, which raises the stakes of a wrong answer considerably."},{"title":"Load only current, owned documents","detail":"Every source document needs an owner and a review date; an assistant answering from an outdated requirement page is worse than no assistant, because it looks authoritative."},{"title":"Require citations on every answer","detail":"Have the assistant link to the specific page or document it drew an answer from, so students can verify it and staff can see exactly what it is drawing on when it gets something wrong."},{"title":"Test it against real advising questions","detail":"Before launch, and after every change, run the assistant against the questions that actually come up in advising appointments, not just the questions the catalog makes easy to answer."},{"title":"Design the handoff, not just the refusal","detail":"When the assistant cannot or should not answer, it should point to the specific human or office that can, with enough context that the student does not start over."}],"guardrails":["Answers only from approved, current documents, with a refusal or handoff when nothing in scope covers the question","No access to an individual student's transcript, grades or registration status unless the institution has deliberately built and secured that integration","Every answer cites its source document so a student, or a reviewing advisor, can verify it","Clear routing to a human advisor for time sensitive, exception based or judgment heavy questions"],"humanInTheLoop":"Advising staff own the source documents and their currency, periodically test the assistant against the kind of nuanced questions that come up in real advising meetings, and review a sample of conversations, particularly ones the assistant answered confidently but a department later flagged as wrong, the way a Statistics department's director of undergraduate studies found Student Compass gave incorrect answers when he tested it against real concentration specific advising questions.","kpisToInstrument":["Share of questions answered without escalation to a human advisor, and repeat contacts on the same topic within a set window","Advisor reported accuracy on a sample of the assistant's answers, by topic","Student satisfaction with assistant answers versus human advising, on comparable question types","Volume and topic of questions the assistant declines or escalates"],"failureModes":[{"title":"Confident but wrong on nuanced questions","detail":"A Harvard statistics department's director of undergraduate studies tested Student Compass against questions that come up in real advising meetings and found it often gave wrong answers. Keep a subject expert testing the assistant against real questions on a schedule, not just at launch."},{"title":"Students treat a citation as a guarantee","detail":"Citing a source makes an answer look more authoritative, even when the assistant has misread or combined that source incorrectly. Keep the \"verify with your advisor\" message visible, not buried, on every answer."},{"title":"Stale source documents","detail":"Catalogs and requirement pages change every term. Without an owner and a review cadence per document, the assistant will answer confidently from an outdated requirement."},{"title":"Scope creep into decisions it should not make","detail":"Course selection and degree planning touch real stakes (graduation timing, financial aid eligibility). Keep the assistant to information and routing, and treat any move toward it making or approving a plan as a new, higher risk feature that needs its own review."}]},"risk":{"euAiAct":{"tier":"limited","basis":"An assistant that answers informational questions about courses and requirements, without deciding admission, assigning students to an institution, or evaluating learning outcomes, falls under the transparency duty of Article 50: students must be told they are talking to AI. It would move toward Annex III point 3 (education and vocational training) if it were used to determine access or admission to an institution or programme (point 3(a)), or to evaluate learning outcomes, including when those outcomes are used to steer the learning process (point 3(b))."},"regulations":["eu-ai-act","gdpr","uk-gdpr"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system, unless this is obvious from the context."},{"title":"Guidance for generative AI in education and research","issuer":"UNESCO","region":"global","url":"https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research","note":"UNESCO's global, non binding guidance on generative AI in education, which proposes that governments regulate GenAI tools, including by mandating the protection of users' data privacy and setting an age limit for independent use."}],"controls":["AI disclosure at the start of every conversation","Human advisor review before the assistant's scope expands to a new topic or to personal student data","Escalation rules that are tested against real advising questions, not just designed on paper","A visible, non buried reminder to verify time sensitive or high stakes answers with an advisor"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** built as a **RAG knowledge assistant** over the institution's\nown catalog, degree requirement pages and student handbook, using **hybrid retrieval** so answers\nare grounded in current policy. A **flow** can hold the fixed parts of the journey, such as\ncollecting the student's programme before routing the question, while open ended requirement and\npolicy questions go to the agent.\n\n**Input and output guardrails** screen questions and answers against admin authored policies, to\ncatch off topic or unsafe exchanges rather than relying on the model alone to stay on script.\n**Human handover** routes time sensitive, exception based or judgment heavy questions to a human\nagent, live in the admin console or through Salesforce, Freshdesk or Zoho SalesIQ, and a flow can\nretrieve an AI summarized conversation history before the handover block so the advisor has\ncontext. The assistant is\ndelivered through **web chat**, with a mobile app reachable through the **REST or WebSocket API**\nchannel, and **multi language** support for institutions with non English speaking students. Run\nthe **test suite** against a set of real advising questions, including the nuanced ones a document\nalone cannot answer, as a practice after every change to the source documents."},"faq":[{"question":"Can an AI advising chatbot replace a human academic advisor?","answer":"Harvard's Assistant Director Brooks B. Lambert-Sluder wrote that Student Compass \"is not meant to replace human advising\", and Elon calls ElonGPT \"a supplemental resource\" for general advising questions. Both point students back to a human advisor, Elon by telling students to verify what the chatbot says against the Academic Catalog or an advisor."},{"question":"How reliable are these assistants on real advising questions?","answer":"Mixed, on the one deployment tested independently. The Harvard Crimson's own testing found Student Compass handled straightforward, policy grounded questions well, with its limits showing on more nuanced ones. Joseph K. Blitzstein, the Statistics department's director of undergraduate studies, separately tested it against real concentration advising questions and found it often gave wrong answers; Angela S. Allan, associate director of History and Literature, said it fell short on the program's more detailed, idiosyncratic questions."},{"question":"What should an academic advising assistant never be allowed to do on its own?","answer":"Make or approve a degree plan, decide admission or progression, or answer from a specific student's transcript or registration data unless that integration has been deliberately built and secured. Harvard's Student Compass draws only from select policy documents and Faculty of Arts and Sciences websites, and ends every answer with a reminder to confirm time sensitive or high stakes decisions with a Resident Dean or the Advising Programs Office. Elon's ElonGPT is offered as a supplemental resource for general advising questions, and Elon tells students to verify what it says against the Academic Catalog, My Progress in OnTrack, or an advisor."},{"question":"What is the difference between this and a general student enrollment chatbot?","answer":"A student enrollment and services assistant typically covers admissions, financial aid, registration and deadlines for admitted or prospective students. An academic advising assistant answers a narrower, ongoing question for currently enrolled students: which courses satisfy which requirements, and which major fits their interests and plan."}],"related":["student-enrollment-and-services-assistant","ai-tutor-for-students"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by the editor after the Opus targeted check's remaining text fixes."},{"date":"2026-09-28","note":"First version, researched from Harvard's own statements as reported by The Harvard Crimson, and from Elon's and the University of Utah's own published pages."},{"date":"2026-09-28","note":"Editorial review: corrected misattributed findings (the \"often gave wrong answers\" result is Joseph K. Blitzstein's, not the Crimson's own testing), separated Harvard's April \"not meant to replace human advising\" statement from Lambert-Sluder's later \"starting point\" statement, limited the Elon claim to what its page says, and removed an unsupported Blits.ai citation capability claim."},{"date":"2026-09-28","note":"Round 2 review fixes: corrected the EU AI Act basis from Annex III point 4 to point 3 (education and vocational training), reworded the UNESCO guidance note since the guidance is non binding and proposes rather than mandates, separated the Harvard \"not meant to replace human advising\" quote from the Elon \"supplemental resource\" wording in the problem section, FAQ and metaDescription, corrected the shareRoutine note to say Draghia's anecdote is above the range and about peer advising fellow questions, relabelled advisingContactsPerStudent so the routine filter is not applied twice, dropped the unsupported human handover summary claim in favour of what FEATURE_INVENTORY.md lists, and fixed the Harvard evidence record's claimant reasoning for the Blitzstein finding."}],"slug":"academic-advising-assistant","url":"https://www.blits.ai/ai-use-cases/academic-advising-assistant","benchmarks":[],"indicativeValueResult":{"low":120000,"high":1250000},"evidence":["elon-university-elongpt-advising","harvard-student-compass-academic-advising","university-of-utah-uguide-advising"]},{"title":"AI agent for account and card servicing","shortTitle":"Account and card servicing","seoTitle":"AI agents for bank account and card servicing","metaDescription":"AI agents that block cards, send statements and answer account questions. Microsoft reports Commonwealth Bank resolved about 84.6% of self service chats end to end.","definition":"An AI agent that resolves routine account and card requests end to end, such as balances, statements, card blocks and replacements, PIN resets and limit changes, across app, web, messaging and phone, and hands anything sensitive or unusual to a human with the full context.","aliases":["banking servicing chatbot","card servicing assistant","virtual banking assistant"],"industries":["banking","payments"],"functions":["customer-service","operations"],"patterns":["conversational-agent","voice-agent","agentic-workflow","rag-knowledge-assistant"],"channels":["mobile-app","web-chat","whatsapp","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"mainstream","segment":"front-office","problem":"Routine servicing requests arrive in a constant stream in retail banking: where is my\ntransaction, send me a statement, my card is lost, raise my limit. Each request is simple, but\ntogether they fill queues, push up waiting times at the moments customers are most anxious (a\nlost card, a payment that did not arrive) and take human agents away from the conversations that\nneed judgment, such as hardship, fraud victims and complaints.\n\nMany first generation banking chatbots were rule based: the CFPB describes them as using decision\ntree logic or a database of keywords to trigger preset, limited responses. A preset answer can\nexplain a procedure, but the customer still has to finish the task somewhere else. The step change\nis an agent that is authenticated, can act in the core banking and card systems within strict\nlimits, and knows when to stop and hand over. DBS describes a similar shift: its DBS Joy assistant\nused to give customers instructions on how to find the information they needed, and now answers\nfrom their own transaction and account data.","problemStats":[{"statement":"The US Consumer Financial Protection Bureau cites an estimate that in 2022 over 98 million users, about 37% of the US population, engaged with a bank's chatbot.","sourceTitle":"Chatbots in consumer finance","sourceUrl":"https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/","year":2023}],"howItWorks":"1. **Understand the request.** The agent detects the intent (\"freeze my card\", \"why was I\n   charged twice\") in the customer's own words and language, on any channel.\n2. **Authenticate proportionally.** Information requests need a logged in session; actions that\n   move money or change a card need step up authentication, such as an app confirmation or voice\n   biometrics on the phone.\n3. **Act through approved tools.** The agent calls a small allow list of banking APIs (card\n   block, replacement order, statement request, limit change within set bounds) and confirms the\n   result back to the customer.\n4. **Answer policy questions from approved content.** Fees, product terms and procedures come\n   from retrieval over the bank's own documents, so answers are grounded and current.\n5. **Hand over well.** Vulnerability signals, complaints, disputes and anything outside the\n   allow list go to a human specialist with a summary of the conversation, so the customer never\n   repeats themselves.","valueDrivers":["cost-to-serve","customer-experience","inclusion-and-access","employee-productivity"],"kpis":["containment-rate","interactions-handled","contact-deflection","handling-time-reduction","customer-satisfaction","customer-satisfaction-uplift"],"indicativeValue":{"referenceOrg":"A retail bank with 1 million digitally active customers","inputs":[{"key":"customers","label":"Digitally active customers","low":1000000,"high":1000000,"unit":"customers","note":"The reference bank."},{"key":"contactsPerCustomer","label":"Assisted contacts per customer per year","low":2,"high":4,"unit":"contacts per customer per year","note":"Editorial assumption for a digitally active retail bank. Replace with your own contact volume."},{"key":"servicingShare","label":"Share of contacts that are routine servicing","low":0.4,"high":0.6,"unit":"fraction of contacts","note":"Editorial assumption, replace with the share from your own contact reason report."},{"key":"containment","label":"Share of servicing contacts the agent resolves","low":0.3,"high":0.6,"unit":"fraction of servicing contacts","note":"Conservative against the benchmarks on this page (Microsoft reports about 84.6% of self service messaging interactions resolved end to end at Commonwealth Bank; DBS reports nine in ten digibot queries resolved digitally), because both figures are for messaging and this range also covers the phone channel."},{"key":"costPerContact","label":"Cost of a human handled contact","low":3,"high":6,"unit":"USD per contact","note":"Editorial assumption for a blended chat and phone contact. Replace with your own fully loaded cost."}],"formula":"customers * contactsPerCustomer * servicingShare * containment * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of running the AI, the integration work, the revenue effect of faster service and any reduction in complaint handling."},"macroEstimates":[{"statement":"McKinsey estimates that generative AI could add between USD 200 billion and USD 340 billion in value to global banking each year.","sourceTitle":"Capturing the full value of generative AI in banking","sourceUrl":"https://www.mckinsey.com/industries/financial-services/our-insights/capturing-the-full-value-of-generative-ai-in-banking","year":2023}],"feasibility":{"complexity":"medium","complexityNote":"Answering questions is easy; acting is the hard part. The work is in the integrations with core banking and card platforms, step up authentication and a clean handover into the contact centre.","dataPrerequisites":["Approved, current product terms, fee tables and servicing procedures","A catalog of servicing intents with volumes from the contact centre","Customer and card data reachable through APIs, not screens"],"integrations":["Core banking system (balances, transactions, statements)","Card management platform (block, replace, limits, PIN)","Identity and step up authentication (app push, one time passcode, voice biometrics)","Contact centre platform for handover with conversation context","CRM or case management for follow up tasks"]},"implementation":{"steps":[{"title":"Pick the first intents by volume and risk","detail":"Take the contact reason report and choose five to ten high volume, low risk intents (statement request, card freeze, transaction lookup). Leave money movement for a later wave."},{"title":"Define the action allow list","detail":"For every action write down the API, the authentication level it needs, the limits (for example a maximum limit increase) and what the agent says when a limit is reached."},{"title":"Ground the answers","detail":"Load only approved product and fee content into the knowledge base, with an owner and a review date per document, and make the agent refuse when the answer is not in it."},{"title":"Design the handover","detail":"Decide which signals trigger a human (vulnerability, complaint, dispute, repeated failure) and pass a summary and the authenticated identity so the customer does not repeat anything."},{"title":"Test before customers do","detail":"Build a test set of real conversations per intent, including edge cases and attempts to make the agent act outside its limits, and run it on every change."},{"title":"Launch in one channel, then widen","detail":"Start in the logged in app, where authentication is strongest, measure containment and satisfaction per intent, then add web, messaging and voice."}],"guardrails":["Actions only through an allow list of APIs, each with its own authentication level and limits","Step up authentication before any card action or money movement","Answers only from approved content, with a refusal when the content does not cover the question","Automatic handover on vulnerability signals, complaints and disputes","Masking of card numbers and personal data in logs and model prompts"],"humanInTheLoop":"Humans own the exceptions: disputes, hardship, suspected fraud victims and complaints. They also review a sample of contained conversations every week to catch answers that were fluent but wrong, and approve every new intent and action before it goes live.","kpisToInstrument":["Containment rate per intent, counting repeat contacts within seven days as not contained","Handover rate and handover reasons","Customer satisfaction on contained conversations versus human handled ones","Share of actions completed without error, from the core system logs","Complaints that mention the assistant"],"failureModes":[{"title":"Fluent but wrong policy answers","detail":"Answers drawn from outdated or unapproved content. Prevent with document ownership, review dates and refusal when retrieval finds nothing."},{"title":"Containment that is really abandonment","detail":"Customers give up rather than get helped, which looks like containment in the dashboard. Count repeat contacts and measure satisfaction per intent."},{"title":"Handover without context","detail":"The customer has to start again with a human, which is worse than no assistant. Pass the summary and the authentication state."},{"title":"Scope creep into risky actions","detail":"New actions are added without their own risk review. Treat every new action as a change with sign off."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1): people must be informed that they are interacting with an AI system, unless that is obvious from the context. Servicing existing accounts and cards is not an Annex III use. It would become high risk under Annex III point 5(b) if the agent itself evaluated the creditworthiness of a natural person, for example to decide a credit limit increase."},"regulations":["eu-ai-act","gdpr","dora","pci-dss","uk-consumer-duty","apra-cps-230","eu-psd2","eu-accessibility-act"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Chatbots in consumer finance","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/","note":"Warns that chatbots which cannot resolve a problem and block access to a human (\"doom loops\") can harm customers and lead to violations of consumer financial law."}],"controls":["AI disclosure at the start of every conversation","Inventory entry for the assistant with an accountable owner and a documented action allow list","Immutable audit trail of every action the agent took, with the authentication level used","Change control and regression tests for every new intent or action","Outcome monitoring for vulnerable customers and complaint trends"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a small set of **custom functions** that call the\nbank's servicing APIs as REST calls, each scoped to one action, combined with a **knowledge\nbase** that holds only approved product and fee content, retrieved with hybrid search. Regulated\njourneys such as a card block follow a **flow** with deterministic steps that calls the bank's\nown step up authentication through a custom function; open questions go to the agent.\n\nThe same agent serves **web chat, WhatsApp, SMS and voice**, and the bank's mobile app through\nthe REST or WebSocket API channel, with streaming speech recognition and synthesis on the phone.\n**Guardrails** check input and output, **PII masking** and card number tokenization happen at\nthe gateway before text reaches a model, and **human handover** passes the conversation to a\nlive agent platform such as Salesforce, Freshdesk or Zoho SalesIQ, with the history (optionally\nsummarized by AI) retrieved in the flow. **Test suites** run multi turn conversations per intent on\nevery change, and analytics show interactions, top intents and satisfaction. The platform is\nmodel agnostic, so the bank can choose or switch the underlying model per agent."},"faq":[{"question":"What share of servicing requests can an AI agent resolve?","answer":"It depends on the intent mix, the channel and whether the agent can act as well as answer. Microsoft reports that at Commonwealth Bank about 84.6% of self service messaging interactions were resolved end to end in May 2026, and DBS reports that DBS digibot resolved nine in every ten queries digitally in the first half of 2026. Both figures are for messaging at large banks, so plan more conservatively for voice and for a first launch."},{"question":"Is a banking servicing chatbot high risk under the EU AI Act?","answer":"Usually not. It falls under the transparency duty of Article 50: customers must know they are talking to AI. It would become high risk under Annex III point 5(b) if it evaluated the creditworthiness of a natural person, for example by deciding a credit limit increase itself, so keep credit decisions in the bank's existing credit process."},{"question":"Which requests should stay with humans?","answer":"Disputes, suspected fraud or scams, hardship and financial difficulty, complaints and any conversation where the customer shows signs of vulnerability."}],"related":["first-line-contact-centre-agent","account-servicing-execution","card-dispute-and-chargeback-intake","fraud-alert-confirmation","atm-and-self-service-device-assistance","digital-onboarding-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, migrated from the Banking AI Playbook and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added PSD2, European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: attributed the Commonwealth Bank figure to Microsoft, added DBS digibot and DBS Joy results and a CFPB statistic and guidance, removed unsupported claims about contact mix and answer only chatbots, corrected the DBS agentic timeline, sharpened the EU AI Act basis (Annex III point 5(b)), aligned the Blits.ai section with the feature inventory, added SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: dropped the Commonwealth Bank two million conversations figure (it counts human handled calls too), kept the self service qualifier in the meta description, tied the rule based chatbot claim to the CFPB and DBS, removed the gated authentication block from the Blits.ai section, and moved the Article 50 guidance link to EUR-Lex."}],"slug":"account-and-card-servicing-agent","url":"https://www.blits.ai/ai-use-cases/account-and-card-servicing-agent","benchmarks":[{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":87.3,"min":84.6,"max":90,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dbs-joy-and-digibot-virtual-assistants","pooled":true},{"id":"commonwealth-bank-customer-service-orchestration","pooled":true}]},{"kpi":"contact-deflection","label":"Contact deflection","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":7,"min":7,"max":7,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dbs-joy-and-digibot-virtual-assistants","pooled":true}]},{"kpi":"customer-satisfaction-uplift","label":"Satisfaction uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":17,"min":17,"max":17,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dbs-joy-and-digibot-virtual-assistants","pooled":true}]}],"indicativeValueResult":{"low":720000,"high":8640000},"evidence":["commonwealth-bank-customer-service-orchestration","dbs-joy-and-digibot-virtual-assistants"]},{"title":"AI agent for apartment leasing inquiries and resident service","shortTitle":"Leasing and resident service agent","seoTitle":"AI leasing agent for apartments and residents","metaDescription":"AI agents answer rental prospects, support tours and handle resident requests and reminders. Equity Residential, AvalonBay and Asset Living use AI in leasing.","definition":"An AI agent that answers rental prospects and residents by chat, text, email and phone for a property manager: it answers questions about apartments and policies, books tours, takes maintenance requests, sends renewal and payment reminders, and hands anything that needs judgment to leasing or service staff.","aliases":["AI leasing assistant","virtual leasing agent","AI resident assistant","property management chatbot","tenant service chatbot"],"industries":["real-estate"],"functions":["customer-service","sales","operations"],"patterns":["conversational-agent","voice-agent","agentic-workflow","rag-knowledge-assistant"],"channels":["sms","email","web-chat","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Rental housing runs on a constant stream of routine conversations: is the two bedroom still\navailable, can I tour on Saturday, is parking included, my sink is leaking, when is my renewal\noffer coming. Questions can arrive at any hour, and a prospect who contacts several buildings may\nmove on to whichever answers first.\n\nOn site teams split their day between prospects, residents, rent collection and paperwork, and an\noperator with hundreds of communities has to keep response times and service consistent across\nall of them. AvalonBay supports its resident focused on site associates with centralized customer\ncare: an AI assistant answers prospects' common questions before self guided tours, and renewals\nare handled AI first, with a central associate following up when the AI cannot answer. Equity\nResidential lists AI responses to customer inquiries among its operating technology in its 10-K,\nand its chief operating officer credited centralization, automation and AI together for a 15% cut\nin on site payroll, as reported by Multifamily Dive.","problemStats":[],"howItWorks":"1. **Answer every lead at once.** The agent replies to inquiries from listing sites, the\n   property website, text, email and phone, using current availability, pricing and community\n   information.\n2. **Book the next step.** It schedules guided or self guided tours, sends access details and\n   follows up after the tour, and records everything in the CRM.\n3. **Serve residents.** It takes maintenance requests with the details the technician needs,\n   creates the work order, gives status updates and answers questions about the lease, rent and\n   community rules.\n4. **Remind and renew.** It sends payment reminders and renewal offers prepared by staff, and\n   answers questions about them.\n5. **Hand over.** Emergencies, complaints, fair housing sensitive questions, accommodation\n   requests, disputes and anything outside its knowledge go to a person immediately, with the\n   conversation attached.","valueDrivers":["cost-to-serve","revenue-growth","customer-experience","speed","employee-productivity"],"kpis":["interactions-handled","response-time-reduction","conversion-rate-uplift","containment-rate","hours-saved","cost-reduction"],"indicativeValue":{"referenceOrg":"A property manager with 50,000 apartment homes","inputs":[{"key":"homes","label":"Apartment homes managed","low":50000,"high":50000,"unit":"homes","note":"The reference property manager."},{"key":"contactsPerHome","label":"Prospect and resident contacts per home per year","low":10,"high":20,"unit":"contacts per home per year","note":"Editorial assumption covering leasing inquiries, tours, maintenance requests and account questions. Replace with your own contact volume."},{"key":"automatedShare","label":"Share of contacts the agent handles without staff","low":0.3,"high":0.6,"unit":"fraction of contacts","note":"Editorial assumption, replace with your own data after a pilot."},{"key":"minutesPerContact","label":"Staff minutes per contact","low":4,"high":8,"unit":"minutes per contact","note":"Editorial assumption, replace with a time study of your leasing and service teams."},{"key":"staffCostPerHour","label":"Fully loaded cost of a leasing or service staff hour","low":25,"high":35,"unit":"USD per hour","note":"Editorial assumption, replace with your own cost."}],"formula":"homes * contactsPerHome * automatedShare * minutesPerContact / 60 * staffCostPerHour","currency":"USD","period":"per year","resultLabel":"Leasing and service staff time released","caveat":"Staff time only. It leaves out the revenue effect of faster responses and reminders on occupancy and on time rent (Asset Living credits its occupancy gain to around the clock responsiveness and its on time rent gain to payment reminders), and the cost of the platform, integrations and the centralization that usually comes with it."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering questions is simple; the value needs live availability, tour scheduling, work orders and resident accounts, which means integration with the property management system and CRM, and a clear split of work between the agent, a central team and on site staff.","dataPrerequisites":["Current availability, pricing, fees and community information per property","Lease terms, policies and house rules, reviewed for fair housing compliance","Maintenance categories with emergency definitions and triage questions"],"integrations":["Property management system (units, residents, leases, ledgers, work orders)","CRM and listing sites for leads and tours","Calendar, access control or smart locks for self guided tours","Telephony, SMS and email for the agent and handover to staff"]},"implementation":{"steps":[{"title":"Start with leasing inquiries","detail":"Leads are high volume and time sensitive, and answers come from structured data. Measure response time, tours booked and conversion per property before and after."},{"title":"Write down what stays human","detail":"Emergencies, reasonable accommodation requests, complaints, disputes, eviction related questions and anything touching screening decisions go to staff. Make the agent say so and hand over."},{"title":"Review content for fair housing","detail":"Check that answers, follow ups and targeting treat every prospect the same way, and test the agent with scenarios that probe for steering or discouraging language."},{"title":"Add maintenance intake","detail":"Teach the agent to recognize emergencies (gas, flooding, no heat) and escalate at once, and to collect the details technicians need for routine requests."},{"title":"Add reminders and renewals","detail":"Send payment reminders and renewal offers prepared by staff, with opt outs and quiet hours, and route hardship conversations to a person."}],"guardrails":["Answers only from current property data and approved policies, with a handover when unsure","Immediate escalation of emergencies and safety issues to on call staff","No screening, approval or pricing decisions by the agent","Consistent answers for every prospect, tested for fair housing risks","Consent, opt out and quiet hours for texts and calls"],"humanInTheLoop":"Leasing and service staff own tours, applications, screening, renewals pricing and every exception. A central team reviews handovers and a sample of conversations weekly, and on site staff handle residents in person.","kpisToInstrument":["Median first response time to leads, by hour of day","Tours booked and leases signed per lead, per property","Share of conversations handled without staff, and handover reasons","Maintenance requests with complete information at first contact","Resident satisfaction and complaints mentioning the agent"],"failureModes":[{"title":"Stale availability or pricing","detail":"The agent promises a unit or price that is gone. Read live data from the property management system, not a copied list."},{"title":"Missed emergencies","detail":"A burst pipe is logged as a routine request. Use explicit emergency triage questions and escalate on any doubt."},{"title":"Discriminatory patterns","detail":"Answers or follow ups differ by a prospect's characteristics, which can breach fair housing law. Standardize answers and test regularly."},{"title":"Residents who cannot reach a person","detail":"Automation that blocks access to staff drives complaints and churn. Make handover easy and visible."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"An agent that answers questions, books tours and takes requests falls under the transparency duty of Article 50. It becomes high risk under Annex III point 5(b) if it evaluates the creditworthiness of applicants, for example in tenant screening, and under point 5(a) if a public body uses it to decide eligibility for social housing or other public assistance."},"regulations":["eu-ai-act","gdpr","uk-gdpr","us-tcpa"],"guidance":[{"title":"42 U.S. Code 3604, discrimination in the sale or rental of housing (Fair Housing Act)","issuer":"United States Congress","region":"north-america","url":"https://www.law.cornell.edu/uscode/text/42/3604","note":"The Fair Housing Act makes it unlawful to discriminate in the terms or services of a rental, to publish statements that indicate a preference based on a protected characteristic, to tell someone a dwelling is unavailable when it is available, and to refuse reasonable accommodations for people with disabilities. An agent that answers prospects and residents speaks for the housing provider on all four points."},{"title":"HUD Issues Fair Housing Act Guidance on Applications of Artificial Intelligence (May 2024, archived)","issuer":"US Department of Housing and Urban Development","region":"north-america","url":"https://archives.hud.gov/news/2024/pr24-098.cfm","note":"Historical context, not current guidance. In May 2024 HUD issued two guidance documents on how the Fair Housing Act applies to tenant screening and to targeted housing ads when AI and algorithms are used. The announcement was moved to the HUD archive on February 3, 2025, and the two guidance documents are no longer published at their hud.gov addresses; the PDFs linked from this archived announcement remain on archives.hud.gov. The statute itself still applies."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Annex III, high risk AI systems (point 5, essential private and public services)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Lists creditworthiness evaluation of natural persons and eligibility for public assistance as high risk uses."}],"controls":["AI disclosure in the first message and on calls","Fair housing review of content and regular testing for inconsistent treatment","Consent records and opt outs for text messages and calls","Logging of every conversation, action and handover","Clear separation between the agent and any screening or pricing system (the Fair Credit Reporting Act applies once a deployment obtains or uses consumer reports to screen applicants)"],"incidents":[{"title":"Louis v. SafeRent Solutions, tenant screening algorithm settlement","url":"https://www.cohenmilstein.com/case-study/louis-et-al-v-saferent-solutions-et-al/","note":"Plaintiffs alleged that a tenant screening score gave disproportionately low scores to Black and Hispanic applicants using housing vouchers, in breach of the Fair Housing Act. The court granted final approval of a $2.275 million settlement with injunctive relief in November 2024. It concerns screening, not a leasing agent, and shows why the agent should stay out of screening decisions."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the property management\nsystem and CRM as REST calls (availability, tour slots, work orders, account balances), and a\n**knowledge base** per property with policies and community information, retrieved with hybrid\nsearch. Maintenance intake runs as a **flow** with deterministic emergency triage questions\nbefore the agent takes over routine requests. **Agentic workflows** triggered on a schedule send\nreminders and follow ups, and **agentic tasks** act when a condition is met, such as a work order\nthat is still open after a set time.\n\nThe same agent serves **web chat, SMS, WhatsApp, email and voice**, with streaming speech\nrecognition and synthesis on the phone and call transfer to the leasing office. **Human\nhandover** and live takeover pass conversations to staff with context, **guardrails** and **PII\nmasking** protect resident data, **test suites** run fair housing and emergency scenarios on every\nchange, and **per bot analytics** and **per agent metrics** show volumes, satisfaction and agent\nperformance. The platform is\nmodel agnostic."},"faq":[{"question":"Which property managers use AI leasing agents?","answer":"Equity Residential lists AI responses to customer inquiries among its operating technology in its 10-K, AvalonBay describes on site associates supported by centralized shared services and a technology platform with automation and AI, and Asset Living uses EliseAI for leasing, collections, maintenance and renewals across more than 450,000 units."},{"question":"What results do operators report?","answer":"Asset Living reports more than 130,000 personalized payment reminders sent in one quarter, higher on time rent payments and occupancy, and about 78 hours of extra staff capacity per community per month. On its earnings call, as reported by Multifamily Dive, Equity Residential attributed a 15% cut in on site payroll to centralization, automation and AI together, so the AI share is not separable."},{"question":"Does fair housing law apply to an AI leasing agent?","answer":"Yes. The Fair Housing Act covers what a housing provider says and does, whoever or whatever says it: an agent that steers prospects, tells some of them a unit is unavailable or mishandles an accommodation request creates the same exposure as a staff member. HUD stated in 2024 guidance, since archived, that the Act applies when AI and algorithms are used in tenant screening and housing advertising. Keep screening decisions out of the agent and test that it treats every prospect the same way."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the real estate scope with evidence from Equity Residential, AvalonBay and Asset Living, checked against the sources."},{"date":"2026-09-27","note":"Editor review. Fair Housing Act statute added as guidance and the 2024 HUD AI guidance marked as archived; unsourced renter behavior claim softened; SafeRent settlement added as a related incident; FCRA scope, spelling and platform wording corrected. Archived HUD guidance PDFs located on archives.hud.gov and the AvalonBay Nareit source pinned to an Internet Archive copy."},{"date":"2026-09-27","note":"Second editor review. Problem statement now attributes each AvalonBay and Equity Residential detail to its own source, the Equity Residential payroll forecast credits AI and other automation together, the value caveat separates responsiveness from reminders, FCRA kept as a control note only, and analytics wording aligned with the platform inventory."}],"slug":"apartment-leasing-and-resident-service-agent","url":"https://www.blits.ai/ai-use-cases/apartment-leasing-and-resident-service-agent","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":130000,"min":130000,"max":130000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"asset-living-eliseai-leasing-and-resident-communications","pooled":true}]}],"indicativeValueResult":{"low":250000,"high":2800000},"evidence":["asset-living-eliseai-leasing-and-resident-communications","avalonbay-ai-leasing-and-customer-care","equity-residential-ai-leasing-and-resident-service"]},{"title":"AI agent for ATM and self service device assistance","shortTitle":"ATM and device assistance","seoTitle":"AI assistant for ATM disputes and card problems","metaDescription":"An AI agent for failed ATM withdrawals, retained cards and PIN blocks. NatWest starts ATM disputes in Cora; rules such as US Regulation E set deadlines.","definition":"An AI agent that helps customers with problems at or around ATMs and other self service devices, such as a withdrawal that did not pay out, a retained card, a blocked PIN or finding a working machine with cash, over the app, chat or phone, and that opens and tracks the claim or hands it to a person when it cannot be resolved.","aliases":["ATM dispute assistant","cash machine help bot","self service device support agent"],"industries":["banking"],"functions":["customer-service","operations"],"patterns":["conversational-agent","voice-agent","agentic-workflow"],"channels":["mobile-app","web-chat","voice","kiosk"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"front-office","problem":"ATM problems arrive at the worst moment: the customer needs cash now, the machine kept their\ncard or debited the account without paying out, and the branch is closed. The customer calls,\nwaits, explains the machine location and time, and is told a claim will take days. Behind the\nscenes, the bank checks the claim against the device and transaction records.\n\nThe work splits into three kinds of request. Simple information (where is the nearest machine\nthat has cash and accepts deposits). Card and PIN problems after a retained card or too many PIN\nattempts. And cash disputes, which are regulated: in the United States, for example, Regulation E\nin the general case gives the bank 10 business days to investigate an incorrect amount from an\nelectronic terminal, or up to 45 days if it provisionally credits the account, with longer limits\nfor new accounts and for withdrawals outside the US. Each kind needs a different mix of\nlookups, authentication and human review.","problemStats":[],"howItWorks":"1. **Recognize the device problem.** The agent detects the intent (\"the ATM took my card\", \"no\n   cash came out\") and asks for the machine, time and amount, pulling the candidate transaction\n   from the account instead of asking the customer to type it.\n2. **Authenticate before acting.** Card and PIN actions and any claim require step up\n   authentication in the app or on the phone.\n3. **Resolve the simple cases.** It locates working machines, explains what happens to a\n   retained card, blocks it and orders a replacement through approved card APIs.\n4. **Open the cash dispute correctly.** For a withdrawal that did not pay out it opens the claim\n   with the required data, explains the investigation timeline and any provisional credit the\n   rules require, and tracks the status.\n5. **Hand over when needed.** Claims that fail automatic reconciliation, suspected fraud or\n   distressed customers go to a person with the case already assembled.","valueDrivers":["customer-experience","cost-to-serve","speed","compliance"],"kpis":["containment-rate","processing-time-reduction","cycle-time-days","customer-satisfaction","customer-satisfaction-uplift","interactions-handled"],"indicativeValue":{"referenceOrg":"A retail bank with a network of about 2,000 ATMs","inputs":[{"key":"deviceContacts","label":"Contacts per year about ATM and device problems","low":50000,"high":100000,"unit":"contacts per year","note":"Editorial assumption, replace with your own contact reason and claim volumes."},{"key":"resolvedShare","label":"Share of those contacts the agent resolves or files without a human","low":0.3,"high":0.5,"unit":"fraction of contacts","note":"Editorial assumption; no deployment on this page discloses a rate for ATM journeys."},{"key":"costPerContact","label":"Cost of a human handled contact or manually filed claim","low":4,"high":8,"unit":"USD per contact","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"deviceContacts * resolvedShare * costPerContact","currency":"USD","period":"per year","resultLabel":"Contact and claim handling cost avoided","caveat":"Handling cost only. It leaves out the cost of the AI and integrations, faster reconciliation in back office operations, fewer regulatory breaches on dispute deadlines and the customer value of getting help when branches are closed."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Machine location and card actions are standard integrations. Cash disputes are harder: the agent needs the transaction record, the device journal or reconciliation result and the dispute rules of each market, and the claim must be auditable.","dataPrerequisites":["ATM and device inventory with location, status, cash and deposit capability","Transaction data that links a withdrawal to a device and time","Dispute rules per market, including deadlines and provisional credit","Contact reason data for device related contacts"],"integrations":["ATM monitoring or device management system","Card management platform (block, replace, PIN unblock)","Core banking transactions and the dispute or claims system","Step up authentication in the app and on the phone","Contact centre platform for handover"]},"implementation":{"steps":[{"title":"Split the intents","detail":"Separate information requests, card and PIN problems and cash disputes. Launch the first two, which need no investigation, before the dispute journey."},{"title":"Wire the dispute journey to the rules","detail":"Encode deadlines, required data and provisional credit logic per market as configuration owned by the disputes team, not as model instructions, and test it against past claims."},{"title":"Use the data the bank already has","detail":"Prefill the machine, time and amount from the transaction record and reconciliation data so the customer confirms rather than types, which cuts errors in filed claims."},{"title":"Close the loop","detail":"Send status updates on open claims in the same channel and let customers ask about a claim without calling."}],"guardrails":["Step up authentication before any card, PIN or claim action","Card numbers and PINs never pass through the model; card data is tokenized before it reaches the conversation","Dispute deadlines and provisional credit decided by rules, not generated by the model","Suspected fraud, repeated claims and distressed customers always go to a person","An auditable record of every claim the agent opened, with the data it used"],"humanInTheLoop":"Dispute analysts own every claim that does not reconcile automatically and every refusal. The disputes team signs off the rules the agent follows and reviews a sample of filed claims each week for completeness.","kpisToInstrument":["Share of device contacts resolved or correctly filed without a human","Median days from claim to resolution","Share of filed claims that were complete on first submission","Dispute deadline breaches","Customer satisfaction on device journeys"],"failureModes":[{"title":"Wrong promises on refunds","detail":"The agent tells a customer money will be back by a date the rules do not support. Generate timelines from rules and approved wording only."},{"title":"Card data in the conversation","detail":"Customers type card numbers or PINs into chat. Detect and tokenize them before they reach the model or logs, and tell the customer not to share a PIN."},{"title":"Stale device data","detail":"The agent sends a customer to a machine that is out of cash. Use live device status and say when data may be out of date."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing assistant must tell people they are interacting with an AI system unless that is obvious (Article 50(1)). It does not evaluate creditworthiness (Annex III point 5(b)) or eligibility for public assistance benefits (point 5(a)), so it is not high risk; biometric verification whose sole purpose is to confirm identity is excluded from Annex III point 1(a)."},"regulations":["eu-ai-act","gdpr","uk-gdpr","eu-psd2","pci-dss","dora","uk-consumer-duty","eu-accessibility-act"],"guidance":[{"title":"§ 1005.11 Procedures for resolving errors (Regulation E)","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/regulations/1005/11/","note":"The US rule for investigating electronic fund transfer errors, including receipt of an incorrect amount of money from an electronic terminal such as an ATM. In the general case the bank has 10 business days to investigate, or up to 45 days if it provisionally credits the account. For transfers within 30 days of the first deposit to a new account, the limits are 20 business days (instead of 10) and 90 days (instead of 45). For transfers not initiated in the US, such as a withdrawal abroad, and for point of sale debit card transfers, the 45 day limit becomes 90 days. An example of the dispute rules the agent must follow."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."}],"controls":["Dispute rules per market as reviewed configuration with a named owner","PCI DSS scoping of every component that could see card data","Audit trail of every card action and claim with the authentication level used","Regression tests on past claims for every change","Monitoring of deadline breaches and complaints about device journeys"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the device journeys are **flows** with deterministic steps for authentication,\ncard actions and claim filing, calling the bank's device, card and dispute systems through\n**custom functions**, with an **AI agent** for the open questions around them. Location and\nprocedure questions are answered from a **knowledge base** of approved content.\n\nThe same journeys run in **web chat**, inside the bank's mobile app through the **REST or\nWebSocket API channel**, and on the **phone**, with real time streaming speech recognition and\nDTMF input on voice. Card numbers typed in free text are detected and\n**tokenized at the gateway**, **PII masking** keeps personal data out of prompts, and\n**human handover** passes the assembled claim to the disputes team. **Test suites** replay\npast claims before each release, and analytics show interactions, top intents, satisfaction\nand unexpected answers per channel."},"faq":[{"question":"Do banks already use AI assistants for ATM problems?","answer":"Yes, as the entry point. NatWest tells customers whose ATM withdrawal did not pay out to start the dispute by typing \"ATM dispute\" to Cora, its AI assistant, in online or mobile banking. NatWest reports a 150% improvement in customer satisfaction for Cora+ overall, but public outcome figures for ATM journeys specifically are scarce."},{"question":"Can the agent refund a failed withdrawal automatically?","answer":"Only where the bank's rules allow it, for example when reconciliation confirms the machine did not dispense. Everything else is an investigation with regulated deadlines, owned by a human analyst."},{"question":"How quickly must a bank resolve an ATM cash dispute?","answer":"It depends on the market. In the general case, US Regulation E requires the bank to decide within 10 business days of the notice, or within 45 days if it provisionally credits the account within those 10 business days. For transfers within 30 days of the first deposit to a new account, the limits are 20 business days (instead of 10) and 90 days (instead of 45). For transfers not initiated in the US, such as a withdrawal abroad, and for point of sale debit card transfers, the 45 day limit becomes 90 days. The rule covers consumer accounts only. The agent should quote these timelines from configured rules, never from the model."}],"related":["account-and-card-servicing-agent","card-dispute-and-chargeback-intake","branch-and-appointment-booking-agent","first-line-contact-centre-agent","retail-store-and-kiosk-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the editor after a targeted blocker check."},{"date":"2026-09-25","note":"First version, written from the banking catalog. Kept as draft because only one public deployment record was found; the catalog's ProSight source is vendor commentary without a deployment or containment figure."},{"date":"2026-09-25","note":"Consolidation pass: added European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: tied the dispute deadline statement to Regulation E (10 business days, 45 with provisional credit) and made the guidance note precise; removed the unsourced claim that reconciliation is often manual; named the Annex III points in the EU AI Act basis; added UK GDPR and PSD2; added the satisfaction uplift KPI and the NatWest figure to the FAQ plus a FAQ on dispute deadlines; limited channels in the Blits.ai section to those in the feature inventory; added seoTitle and metaDescription. Stays draft: one public deployment record."},{"date":"2026-09-27","note":"Review fixes: rewrote the meta description so NatWest is no longer tied to US Regulation E and the retained card and PIN scope reads as agent capability; qualified the Regulation E deadlines as the general case and added the longer limits (20 business days, 90 days) for new accounts and withdrawals outside the US in the problem, the FAQ and the guidance note; limited the Blits.ai analytics claim to what the feature inventory lists; softened the unsourced claim about how banks check device claims."}],"slug":"atm-and-self-service-device-assistance","url":"https://www.blits.ai/ai-use-cases/atm-and-self-service-device-assistance","benchmarks":[{"kpi":"customer-satisfaction-uplift","label":"Satisfaction uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":150,"min":150,"max":150,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"natwest-cora-ai-assistant","pooled":true}]}],"indicativeValueResult":{"low":60000,"high":400000},"evidence":["natwest-cora-ai-assistant"]},{"title":"AI agent for branch finding and appointment booking","shortTitle":"Branch and appointment booking","seoTitle":"AI agents for branch and appointment booking","metaDescription":"An AI agent finds the right branch and books the specialist. Bank of America schedules appointments via Erica; Best Buy and MOGUL.sg book appointments with AI.","definition":"A conversational agent that finds the nearest suitable location, checks opening hours and which services it offers, books an in person or video appointment with the right specialist, and records the reason for the visit so staff are prepared. In banking it answers \"where is my nearest branch\" and books the mortgage or business banker; the same job exists in retail, healthcare and property.","aliases":["branch locator chatbot","appointment booking assistant","branch appointment scheduling","specialist booking agent"],"industries":["cross-industry","banking","retail-and-ecommerce","healthcare","real-estate"],"functions":["customer-service","sales"],"patterns":["conversational-agent","agentic-workflow","rag-knowledge-assistant","voice-agent"],"channels":["web-chat","mobile-app","whatsapp","voice"],"audience":"customer-facing","autonomy":"autonomous","adoptionStage":"early-adopters","problem":"When not every branch offers every service, the question \"where do I go\" is hard to answer.\nOpening hours vary, and specialists such as mortgage or business bankers may work by\nappointment in only a few locations. Customers who turn up at the\nwrong place, at the wrong time or without the right documents waste a trip, and staff meet them\nunprepared.\n\nBooking by phone ties up contact centre time for a simple task, and web booking forms often do not\nknow which specialist handles which need. The same pattern appears wherever physical visits need\nto be planned: a retailer's service desk, a clinic, a property viewing.","problemStats":[],"howItWorks":"1. **Understand the need.** The agent asks what the visit is for (a mortgage, a business account,\n   a cash deposit, help with the app) because that decides where and with whom.\n2. **Find the right location.** It searches location data for the nearest branch or ATM that\n   offers that service, with opening hours, accessibility details and any temporary closures.\n3. **Offer alternatives.** Where a video call, the app or a phone call would serve the customer\n   better, it says so, and books that instead if the customer prefers.\n4. **Book the slot.** It checks the specialist calendars, offers real times and books the\n   appointment through the scheduling system, with a confirmation and a reminder.\n5. **Prepare the visit.** It records the reason for the visit and tells the customer which\n   documents to bring, so the specialist is ready.\n6. **Pass special needs to a person.** Accessibility requests and anything unusual go to staff\n   rather than being handled by a rule.","valueDrivers":["customer-experience","cost-to-serve","revenue-growth","inclusion-and-access"],"kpis":["interactions-handled","first-contact-resolution","customer-satisfaction","conversion-rate-uplift","handling-time-reduction"],"indicativeValue":{"referenceOrg":"A bank that books 100,000 branch and specialist appointments a year","inputs":[{"key":"appointments","label":"Appointments booked per year","low":100000,"high":100000,"unit":"appointments per year","note":"The reference bank."},{"key":"agentShare","label":"Share of bookings moved from phone and branch staff to the agent","low":0.3,"high":0.6,"unit":"fraction of appointments","note":"Editorial assumption, replace with your own channel mix."},{"key":"minutesPerBooking","label":"Staff minutes per booking handled by a person","low":4,"high":8,"unit":"minutes per booking","note":"Editorial assumption covering the call, calendar lookup and confirmation."},{"key":"costPerMinute","label":"Fully loaded cost of a staff minute","low":0.6,"high":1,"unit":"USD per minute","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"appointments * agentShare * minutesPerBooking * costPerMinute","currency":"USD","period":"per year","resultLabel":"Booking handling cost avoided","caveat":"Counts booking time only. It leaves out fewer wasted visits and no shows, better prepared appointments that convert more often, the cost of the AI and the calendar integration."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Mostly reads location data and writes to one scheduling system. The effort goes into clean, current branch and service data, and a calendar integration that reflects real specialist availability.","dataPrerequisites":["Branch and ATM locations with opening hours, services, accessibility details and closures","A mapping from visit reasons to services and specialist roles","Specialist calendars and booking rules","Document checklists per visit reason"],"integrations":["Scheduling or appointment system","Location data or maps service","CRM for visit reasons and follow up","Messaging for confirmations and reminders","Contact centre for handover"]},"implementation":{"steps":[{"title":"Clean the location data","detail":"Make one owned source for every location's hours, services, accessibility and closures. A branch assistant can only be as current as the branch data behind it, so fix this first."},{"title":"Map reasons to specialists","detail":"Write down which visit reasons need which service and role, and which can be handled by video or in the app instead."},{"title":"Integrate the real calendar","detail":"Offer only slots the scheduling system confirms, and write the booking back with the reason for the visit."},{"title":"Add reminders and rescheduling","detail":"Send a confirmation and a reminder with the document checklist, and let the customer move or cancel the appointment in the same conversation."},{"title":"Measure kept appointments","detail":"Track bookings, no shows and outcomes by channel so you can see whether guided booking actually improves visits."}],"guardrails":["Services and hours only from the owned location data, never inferred by the model","Bookings only into slots the scheduling system confirms","Accessibility needs and special requests passed to staff, not handled by rule","Personal data in bookings disclosed, minimised and stored in region"],"humanInTheLoop":"Branch staff own the appointment once booked and handle accessibility and special requests. Location data owners keep hours and services current, and a sample of conversations is reviewed monthly for wrong locations or services.","kpisToInstrument":["Bookings completed by the agent and share of all bookings","No show rate by booking channel","Wrong location or service reports from customers and staff","Handover rate and reasons","Customer satisfaction after the visit"],"failureModes":[{"title":"Promising a service the branch does not offer","detail":"The customer arrives and cannot be helped. Keep service lists owned and current, and answer only from them."},{"title":"Phantom slots","detail":"The agent offers times that are no longer free. Read and write the live calendar."},{"title":"Accessibility handled by rule","detail":"A wheelchair user is sent to a branch with steps. Pass accessibility needs to a person and keep accessibility data current."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1): people must be told they are interacting with an AI system unless that is obvious. Finding locations and booking appointments does not fall under any Annex III category. If a healthcare version starts to triage patients by urgency, or a public body uses it to decide eligibility for a public service, reassess it against Annex III point 5."},"regulations":["eu-ai-act","gdpr","eu-accessibility-act"],"guidance":[{"title":"Directive (EU) 2019/882 on the accessibility requirements for products and services","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2019/882/oj","note":"The European Accessibility Act sets accessibility requirements for consumer banking services and ecommerce services, so digital booking channels in those sectors need to meet them."}],"controls":["AI disclosure and a clear route to a person","Owned, dated location and service data with change control","Consent and retention rules for booking data","Monitoring of wrong location reports and no shows"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **custom function** that queries the location data\n(for example a **SQL knowledge base** of branches, hours and services) and another that reads\nand writes the scheduling system over REST. A **flow** handles the booking steps with **show\noptions** for time slots and a **multiple entity check** for the visit details, and the agent\nhandles free questions such as \"which branch near the station opens on Saturday\".\n\nThe same agent runs on **web chat**, **WhatsApp** and **voice**, inside your own app through the\n**REST or WebSocket API channel**, and as a **digital human** in a browser on a branch or lobby screen.\nConfirmations and reminders go out by **email**, sent from a **scheduled workflow**. **Human\nhandover** passes accessibility and special requests to staff, and the **GDPR toolkit** handles consent and data\nremoval. **Monitors** run scheduled checks that the agent returns correct hours for sample\nbranches, and **analytics** show bookings and handovers."},"faq":[{"question":"Is a branch booking agent worth it when most banking is digital?","answer":"It depends on your visit mix. Where branch visits are mostly for specialist needs such as mortgages or business banking, sending the customer to the right person, prepared, saves a wasted trip on both sides. Bank of America, whose Erica assistant has served nearly 50 million users since 2018, uses it to schedule appointments as a handoff to human service."},{"question":"Who already books appointments with AI agents?","answer":"Bank of America uses Erica to schedule appointments as a handoff to high touch service channels. Outside banking, Google Cloud reports that Best Buy uses AI to guide shoppers through appointment scheduling, and MOGUL.sg books property viewings through a WhatsApp agent. In healthcare, Hemominas in Brazil worked with Xertica on a chatbot for donor search and scheduling. Public outcome data specific to booking is scarce."},{"question":"What should you get right first?","answer":"Stale location data is the risk to plan for first. Wrong opening hours or a service a branch no longer offers send the customer on a wasted trip however good the conversation is, so fix data ownership before tuning the agent."}],"related":["patient-appointment-scheduling-and-reminders-agent","outbound-reminder-and-confirmation-agent","home-loan-assistant-and-prequalification","atm-and-self-service-device-assistance","account-and-card-servicing-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written industry neutral from the Banking AI use case catalog; evidence from Bank of America, Best Buy, MOGUL.sg and Hemominas."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsourced claims about shrinking branch networks and the main cause of wrong answers, made the Article 50 basis precise, added the European Accessibility Act, named only listed Blits.ai channels, added Hemominas to the FAQ, dated the evidence sources and added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: corrected the Hemominas year to 2024 with an archived copy of the Google Cloud list, reworded the Bank of America handoff line closer to its release, softened the claim about specialist availability and kept the Blits.ai build within listed features."}],"slug":"branch-and-appointment-booking-agent","url":"https://www.blits.ai/ai-use-cases/branch-and-appointment-booking-agent","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":3000000000,"min":3000000000,"max":3000000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-erica-virtual-assistant","pooled":true}]}],"indicativeValueResult":{"low":72000,"high":480000},"evidence":["bank-of-america-erica-virtual-assistant","best-buy-appointment-scheduling-agent","hemominas-donor-scheduling-chatbot","mogul-whatsapp-viewing-appointments"]},{"title":"AI agent for card dispute intake","shortTitle":"Card dispute intake","seoTitle":"AI agent for card dispute and chargeback intake","metaDescription":"An AI agent finds the charge, separates fraud from merchant disputes and opens complete cases. Visa processed 106 million disputes in 2025, up 35% since 2019.","definition":"A customer facing AI agent that handles the \"I do not recognise this charge\" moment: it finds the transaction, separates suspected fraud from merchant disputes and simple confusion, explains the customer's rights and timelines, collects the details and evidence the rules require, and opens a correctly classified dispute case for the operations team.","aliases":["transaction dispute chatbot","unrecognised charge assistant","dispute intake agent","chargeback claim intake"],"industries":["banking","payments"],"functions":["customer-service","fraud-prevention","operations"],"patterns":["conversational-agent","voice-agent","classification-and-routing","document-processing","agentic-workflow"],"channels":["mobile-app","web-chat","voice","whatsapp"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"Card disputes keep rising, and every one of them starts with a customer who is worried about\nmoney. Many are not fraud at all: Visa's Andrew Torre told CNBC that many disputes start with a\ncardholder who does not recognise a charge on the statement. Others are genuine fraud\nthat needs the card blocked now, and others again are merchant problems (goods not received, a\nrefund that never arrived) that follow the card scheme's dispute rules.\n\nIntake is where much of the later cost and risk starts. A dispute filed under the wrong category,\nor without the details the network requires, can bounce between teams, miss deadlines and end in\na write off. Timelines and refunds are regulated: in the US by Regulation E for debit cards and\nRegulation Z for credit cards, and in the EU by PSD2, which requires a refund of an unauthorised\npayment by the end of the following business day unless the provider has reasonable grounds to\nsuspect fraud. A vague or wrong promise to the customer therefore becomes a compliance issue. The US Consumer Financial Protection Bureau has also\nwarned that chatbots and highly scripted representatives may only recognise a dispute when the\ncustomer uses specific words.","problemStats":[{"statement":"Visa reports that it processed 106 million disputes globally in 2025, a 35% increase since 2019.","sourceTitle":"Visa Unveils New Services to Modernize Dispute Resolution Process","sourceUrl":"https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22261.html","year":2026},{"statement":"Salesforce reported in April 2024 that Americans had disputed USD 83 billion in charges the previous year.","sourceTitle":"Salesforce Launches AI-Powered Capabilities to Help Banks Quickly Resolve Transaction Disputes","sourceUrl":"https://www.salesforce.com/news/stories/financial-services-cloud-ai-capabilities/","year":2024}],"howItWorks":"1. **Recognise the dispute in the customer's own words.** \"I never bought this\", \"they charged me\n   twice\" and \"my refund never came\" are all disputes, whatever the phrasing or language.\n2. **Find the transaction.** In an authenticated session the agent lists recent transactions and\n   enriches the one in question with the merchant's clear name, location and order details.\n   Because many disputes start with a charge the cardholder does not recognise, some of them can\n   be resolved right here, without a case.\n3. **Triage.** The agent separates three paths: unauthorised use (block the card, fraud claim),\n   an authorised purchase that went wrong (merchant dispute), or confusion (explain and close).\n   Signs of a scam, where the customer was tricked into paying, go to the scam team instead.\n4. **Collect what the rules need.** It asks only the questions the chosen path requires (dates,\n   contact with the merchant, cancellation proof) and accepts receipts and screenshots, which\n   document AI reads into structured fields.\n5. **Set expectations from rules, not from the model.** Timelines, provisional credit and next\n   steps come from a deterministic rules engine per product and market, so the promise is right\n   and auditable.\n6. **Open the case.** The agent creates the case in the dispute system with a suggested reason\n   category and the evidence attached, for an analyst to confirm, and gives the customer a\n   reference number.\n7. **Hand over when it matters.** Vulnerable customers, hardship, high values, repeat disputes and\n   anything the agent cannot classify go to a human with the full summary.","valueDrivers":["cost-to-serve","customer-experience","compliance","risk-reduction"],"kpis":["handling-time-reduction","first-contact-resolution","accuracy","automation-rate","customer-satisfaction","interactions-handled"],"indicativeValue":{"referenceOrg":"A card issuer with 1 million active cards","inputs":[{"key":"activeCards","label":"Active cards","low":1000000,"high":1000000,"unit":"cards","note":"The reference issuer."},{"key":"disputesPerCard","label":"Disputes raised per card per year","low":0.01,"high":0.03,"unit":"disputes per card per year","note":"Editorial assumption, replace with your own dispute volume."},{"key":"agentShare","label":"Share of disputes started with the agent","low":0.3,"high":0.6,"unit":"fraction of disputes","note":"Editorial assumption; depends on how prominent the digital route is in the app and on the phone menu."},{"key":"minutesSaved","label":"Intake and rework minutes saved per dispute","low":5,"high":10,"unit":"minutes per dispute","note":"Editorial assumption. Quavo reports that its clients cut handle time per assignment by nearly 30% (vendor claim, grade D), which supports a modest saving."},{"key":"costPerMinute","label":"Fully loaded cost of a dispute agent minute","low":0.8,"high":1.2,"unit":"USD per minute","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"activeCards * disputesPerCard * agentShare * minutesSaved * costPerMinute","currency":"USD","period":"per year","resultLabel":"Dispute intake and rework cost avoided","caveat":"Counts intake and rework time only. It leaves out write offs avoided through correct classification and on time filing, disputes prevented by explaining unfamiliar charges, the cost of running the AI and the integration work."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The conversation is the easy part. The work is in reading transactions and merchant data in real time, encoding timeline and provisional credit rules per product and market, and writing clean cases into the dispute system that analysts trust.","dataPrerequisites":["Transaction history with enriched merchant names, locations and, where available, order details","Dispute rules per product and market (categories, required information, timelines, provisional credit)","Historical disputes with outcomes, to test triage and the suggested categories","Approved customer wording for rights, timelines and next steps"],"integrations":["Card management and core banking (transactions, card block and reissue)","Dispute or case management system (case creation, status, documents)","Merchant data enrichment and network dispute services where the issuer uses them","Identity and step up authentication","Contact centre platform for handover with context"]},"implementation":{"steps":[{"title":"Map the intake paths","detail":"From last year's disputes, list the real paths (fraud, not received, not as described, duplicate, cancelled subscription, unrecognised) with the questions and documents each one needs. This becomes the agent's playbook and your test set."},{"title":"Put the rules outside the model","detail":"Encode timelines, provisional credit and eligibility per product and market in a rules service. The agent calls it and repeats its answer; it never works out a deadline itself."},{"title":"Resolve confusion first","detail":"Show the merchant's clear name, logo, location and order details before offering a dispute. Visa's Order Insight service rests on the same idea: surfacing transaction details clears up confusion over legitimate charges before they become disputes."},{"title":"Make the case analyst ready","detail":"Agree with dispute operations exactly which fields and documents a case needs, and have the agent suggest a category with its reasoning for the analyst to confirm, not apply."},{"title":"Route fraud and scams separately","detail":"Unauthorised use triggers a card block and the fraud path; a customer who was tricked into paying goes to the scam team, because the rules and the customer's rights differ."},{"title":"Pilot on one product and one channel","detail":"Start with debit or credit cards in the logged in app, compare case quality and rework against the human channel for a month, then widen."}],"guardrails":["Timelines, provisional credit and eligibility come only from the rules service, never from the model","Suggested reason categories are confirmed by an analyst before the case is filed with the network","Card numbers are tokenized and personal data masked before any text reaches a model","Automatic handover on vulnerability signals, hardship, suspected scams and high values","The agent never tells a customer a dispute will succeed"],"humanInTheLoop":"Dispute analysts confirm the category, decide on provisional credit where rules leave room, and own every case once it is opened. Specialists take over for scam victims, vulnerable customers and complaints, and a weekly sample of agent opened cases is reviewed for completeness and correct triage.","kpisToInstrument":["Share of disputes resolved without a case (confusion explained), with repeat contact within 30 days counted as not resolved","Case completeness at first submission and analyst rework rate","Accuracy of suggested categories against the analyst's final category","Time from first message to case opened","Customer satisfaction on dispute conversations versus the phone channel"],"failureModes":[{"title":"Wrong promises on timelines or credit","detail":"The model paraphrases a policy and gets a date or an amount wrong. Keep all such statements in the rules service and test them on every change."},{"title":"Scam victims treated as disputes","detail":"A customer who authorised a payment under false pretences is pushed into a fraud chargeback that is unlikely to succeed. Train triage on scam signals and route them to specialists."},{"title":"Dispute not recognised","detail":"The customer describes a problem in words the agent treats as a question, and no dispute is opened. The CFPB names this as a legal risk; test with real, messy phrasing."},{"title":"Friendly fraud made easier","detail":"A frictionless flow invites false claims. Surface repeat patterns to analysts and keep evidence requirements proportionate to value."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing assistant must tell people they are interacting with an AI system (Article 50(1)). It triages and opens cases but does not evaluate creditworthiness (Annex III point 5(b), which in any case excludes systems used to detect financial fraud) or decide access to an essential service, so it is not high risk under Annex III."},"regulations":["eu-ai-act","gdpr","uk-gdpr","pci-dss","uk-consumer-duty","dora","eu-psd2"],"guidance":[{"title":"Chatbots in consumer finance","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/","note":"Warns that chatbots which fail to recognise and resolve disputes can breach federal consumer financial law."},{"title":"Regulation E, section 1005.11, procedures for resolving errors","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/regulations/1005/11/","note":"Sets the investigation timelines and provisional credit rules for electronic fund transfer errors in the US; the agent's promises must match them."},{"title":"Regulation Z, section 1026.13, billing error resolution","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/regulations/1026/13/","note":"Sets the billing error procedure for US credit cards, including completing the investigation within two complete billing cycles."},{"title":"Payment Services Directive (EU) 2015/2366, Article 73, refunds for unauthorised payment transactions","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32015L2366","note":"Requires the payer's provider to refund an unauthorised payment by the end of the following business day, unless it has reasonable grounds to suspect fraud and reports them to the national authority."}],"controls":["AI disclosure at the start of the conversation and a clear route to a person","Versioned rules service for timelines and credit, with change control and tests","Audit trail of every statement about rights, timelines and credit made to the customer","Analyst confirmation of the reason category before network filing","Monitoring of dispute outcomes and complaints by channel"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the intake runs as a **flow** for the regulated steps (identity check, choosing the\ntransaction, required questions, **receive attachment** for receipts) with an **AI agent**\ninside it for free text understanding and explanations. **Custom functions** call the card\nplatform, the rules service and the case system over REST, so the agent reads transactions,\nblocks a card and opens a case only through the functions it has been given. Approved dispute\npolicy content sits in the **knowledge base** with hybrid retrieval.\n\n**PII masking** and card number tokenization at the gateway keep card data away from the model.\n**Guardrails** check input and output, and **human handover** passes the conversation to the\ncontact centre when triage points to a scam, vulnerability or a high value. **Test suites**\nbuilt from real dispute conversations are run before every change, **execution tracing** shows\nhow each conversation was triaged, and **analytics** track volumes and satisfaction. The same\nagent serves **web chat**, **WhatsApp** and **voice**, and reaches the issuer's mobile app\nthrough the **REST or WebSocket API channel**; the platform is model agnostic."},"faq":[{"question":"Can an AI agent decide who gets a chargeback?","answer":"It should not. The agent collects facts and suggests a category; an analyst confirms it and the card scheme rules decide the outcome. The issuer tools Visa announced in April 2026 follow the same split, with AI supporting analysts through predictions and document summaries rather than deciding cases."},{"question":"How is this different from chargeback automation in the back office?","answer":"Intake is the customer conversation that creates the case. Chargeback and representment operations work the case afterwards: evidence packages, network filings and deadlines. Good intake makes the back office cheaper because cases arrive complete and correctly classified."},{"question":"Do AI assistants already handle disputes in production?","answer":"Yes. Klarna says its AI assistant handles disputes alongside refunds and returns, and Commonwealth Bank routes fraud disputes through a guarded, deterministic path inside its AI orchestration. Published results cover customer service as a whole, not disputes alone."}],"related":["chargeback-and-representment","fraud-alert-confirmation","account-and-card-servicing-agent","scam-payment-interception","complaints-handling-agent","agentic-payment-initiation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI use case catalog and verified against Visa, Klarna, Commonwealth Bank, Quavo and CFPB sources."},{"date":"2026-09-25","note":"Consolidation pass: added PSD2 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; replaced unsourced claims about how many charges are resolved without a case with Visa's statements; softened the CFPB paraphrase to match the report; added Regulation Z guidance and UK GDPR; made the EU AI Act basis precise; limited the Blits.ai build notes to inventoried capabilities; corrected the Quavo summary and source date."},{"date":"2026-09-27","note":"Blits.ai build text now names the REST and WebSocket API channel instead of web chat inside the mobile app, and no longer claims test suites run automatically; the unrecognised charge point is attributed to Visa's Andrew Torre in CNBC; unsourced dispute examples removed; PSD2 Article 73 added to the prose and guidance; FAQ scoped to Visa's issuer tools."}],"slug":"card-dispute-and-chargeback-intake","url":"https://www.blits.ai/ai-use-cases/card-dispute-and-chargeback-intake","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2300000,"min":2300000,"max":2300000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"klarna-ai-assistant-customer-service","pooled":true}]}],"indicativeValueResult":{"low":12000,"high":216000},"evidence":["commonwealth-bank-customer-service-orchestration","klarna-ai-assistant-customer-service","visa-dispute-resolution-services"]},{"title":"AI agent for cloud cost optimization and FinOps","shortTitle":"Cloud cost optimization agent","seoTitle":"AI agent for cloud cost optimization","metaDescription":"An AI FinOps agent rightsizes workloads and buys commitments. Akamai saved 40 to 70% with Cast AI; VERMEG cut on demand AWS costs more than 39% with nOps.","definition":"An AI agent that continuously reads an organization's cloud usage and billing data, uses machine learning to separate normal spend from waste, and either rightsizes resources and buys a mix of committed capacity matched to forecast usage on its own within set limits, or proposes higher risk changes for an engineer to approve.","aliases":["FinOps AI agent","AI cloud cost optimization","autonomous FinOps","cloud spend optimization agent","AI commitment management"],"industries":["cross-industry","technology"],"functions":["it-and-engineering"],"patterns":["anomaly-detection","prediction-and-scoring","agentic-workflow"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Cloud bills grow with usage nobody is watching in real time: instances sized for a launch that\nnever get downsized, storage nobody deleted, reserved capacity that no longer matches the\nworkload. A platform team can see the bill at the end of the month, but by then the waste has\nalready been paid for, and working out which of thousands of resources across dozens of accounts\nare the problem is its own project.\n\nCommitted discounts make this harder, not easier. AWS Savings Plans commit a consistent amount\nof usage for a one or three year term, so many teams either skip them and pay full on demand\nprice, or commit too much and pay for capacity they stop using when a workload changes. AWS\nreports that as of May 2026, across a sample of more than 71,000 anonymized, opted in AWS\ncustomers, the median Cost Efficiency score is 83 while the mean is 79, a gap it attributes to a\nlong tail of accounts that are less optimized than the typical one.","problemStats":[{"statement":"Savings Plans offer low prices on Amazon EC2, AWS Lambda and AWS Fargate usage in exchange for a commitment to a consistent amount of usage for a 1 or 3 year term.","sourceTitle":"AWS Savings Plans compute pricing","sourceUrl":"https://aws.amazon.com/savingsplans/compute-pricing/","year":2026},{"statement":"AWS reports that as of May 2026, across more than 71,000 anonymized, opted in AWS customers, the median Cost Efficiency score is 83 while the mean is 79, a gap it attributes to a long tail of less optimized accounts.","sourceTitle":"The AWS State of Cost Efficiency Report","sourceUrl":"https://aws.amazon.com/blogs/aws-cloud-financial-management/the-aws-state-of-cost-efficiency-report/","year":2026}],"howItWorks":"1. **Read usage and billing continuously.** The agent ingests cost and usage data across\n   accounts, regions and workload types, instead of a monthly export someone reviews by hand.\n2. **Learn the normal pattern.** Machine learning models the expected usage per resource and\n   flags spend that deviates from it, including idle, oversized and orphaned resources.\n3. **Model the commitment portfolio.** For reserved capacity, the agent forecasts usage and\n   recommends or automatically manages the mix of on demand, reserved and spot capacity that\n   covers it, adjusting as usage changes instead of locking in a single upfront bet.\n4. **Act within limits.** Low risk changes that are reversible, such as stopping an idle\n   resource or buying a commitment that carries a buyback or exchange option, execute\n   automatically under a policy; changes that touch running production workloads go to an\n   engineer as a proposal with the expected saving and the risk.\n5. **Report the outcome.** Savings, coverage and utilization are tracked back to the team and\n   workload that owns the cost, so the finance and engineering view of spend stays the same one.","valueDrivers":["cost-to-serve","employee-productivity"],"kpis":["cost-reduction","cost-savings","automation-rate"],"indicativeValue":{"referenceOrg":"A technology company spending USD 5 million a year on public cloud","inputs":[{"key":"annualCloudSpend","label":"Annual public cloud spend","low":5000000,"high":5000000,"unit":"USD per year","note":"The reference organization. Replace with your own annual cloud bill."},{"key":"wasteShare","label":"Share of spend that is waste or poorly optimized","low":0.15,"high":0.3,"unit":"fraction of spend","note":"Editorial assumption. AWS reports a median Cost Efficiency score of 83 and a mean of 79 across more than 71,000 opted in customers (AWS State of Cost Efficiency Report). That score measures the share of optimizable spend that is already well optimized, not the share of the total bill that is waste, so it does not map directly to this input; it is used only as a signal that most accounts still leave savings on the table."},{"key":"capturedSavings","label":"Share of the identified waste the agent captures","low":0.4,"high":0.6,"unit":"fraction of identified waste","note":"Editorial assumption, not derived directly from the evidence on this page. VERMEG's reported figure is a cut in on demand AWS spend; part of that reduction moves into committed (Reserved Instance) spend rather than disappearing as a net saving, so it does not map cleanly onto \"share of identified waste captured\". Akamai's 40 to 70% figure covers only the workloads Cast AI manages, not an organization's total cloud bill. Both are used only as a signal that a well run agent captures a meaningful share of identified waste, not as a direct calibration of this range; replace with your own track record."}],"formula":"annualCloudSpend * wasteShare * capturedSavings","currency":"USD","period":"per year","resultLabel":"Annual cloud cost avoided","caveat":"Gross savings only. It leaves out the subscription cost of the optimization platform, the engineering time to configure and monitor it, and the performance risk of aggressive rightsizing if guardrails are too loose."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Reading usage and billing data is a fast, low risk integration. The work is in tagging resources to the right owner and workload, setting the policy for what the agent may change on its own, and building trust before commitment purchases run unattended.","dataPrerequisites":["Billing and usage export from every cloud account in scope","Resource tags that map spend to a team, product or workload owner","Existing reserved capacity and commitment inventory","A policy for what may change automatically versus what needs approval"],"integrations":["Cloud billing and cost management APIs (AWS, Google Cloud, Azure)","Container orchestration platform for rightsizing (Kubernetes)","Infrastructure as code or the compute provisioning API, for executing approved changes","Chat tool for savings reports and approval requests","Finance or FP&A system for chargeback and budget tracking"]},"implementation":{"steps":[{"title":"Start with visibility, not action","detail":"Connect billing and usage data, tag spend to owners, and publish a shared view of waste before the agent changes anything. This is also what earns engineering trust."},{"title":"Automate the reversible changes first","detail":"Idle resource cleanup, storage tier changes and commitment purchases that come with a buyback or exchange guarantee are safe to automate early; changes to a running production workload are not."},{"title":"Write the action policy","detail":"Decide by resource type and environment what the agent may do alone, what needs an engineer's approval, and what stays out of scope entirely (for example anything touching a regulated workload)."},{"title":"Rightsize with a rollback plan","detail":"Every automated rightsizing change should be reversible within minutes; monitor error rates and latency after a change and roll back automatically if they move."},{"title":"Attribute savings to the team that owns the workload","detail":"Report savings and remaining waste by team and product, not only as one company wide number, so the incentive to keep costs down sits with the people who can act on it."}],"guardrails":["Read only access to billing and usage data; write access limited to an explicit allow list of actions","A policy layer that separates automatic actions from actions that need an engineer's approval","Every automated change is reversible, logged and attributed to the agent and the policy that allowed it","Commitment purchases carry a buyback or exchange option rather than a fixed, unchangeable term","Alerting on cost, error rate and latency after every automated change, with automatic rollback on regression"],"humanInTheLoop":"Engineers approve any change to a running production workload and any commitment above a set size. Platform and finance teams review the action policy periodically and adjust what the agent may do alone as trust builds.","kpisToInstrument":["Realized savings versus identified savings, by team and workload","Share of recommendations executed automatically versus approved manually versus rejected","Commitment coverage and utilization, to catch both under and over commitment","Incidents or rollbacks caused by an automated change"],"failureModes":[{"title":"Savings that break performance","detail":"An aggressive rightsizing change starves a workload at its next traffic peak. Keep a margin on autoscaled resources and monitor after every change."},{"title":"Commitment lock in","detail":"A long term commitment is bought against a workload that later moves or shuts down. Prefer commitments with an exchange or buyback option and reforecast regularly."},{"title":"Tag debt hides the owner","detail":"Untagged or mistagged resources cannot be attributed, so savings and accountability stall. Enforce tagging at resource creation and flag untagged spend separately."},{"title":"Automation nobody trusts","detail":"Engineers turn off the automatic actions after one bad surprise. Start narrow, publish a clear action log, and expand scope only after a track record."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"An internal tool that optimizes infrastructure spend and makes no decision about a natural person, so the default case falls outside Annex III. The relevant Annex III entry to check against is point 2, AI safety components in the management and operation of critical digital infrastructure: a cost agent stays outside it as long as its policy keeps it to cost actions (rightsizing, commitment purchases, idle cleanup) rather than acting as a safety component of the infrastructure itself. The Akamai deployment on this page shows the scope can extend to core production infrastructure, so an operator should confirm this against its own policy rather than assume it."},"regulations":["nist-ai-rmf","iso-42001"],"guidance":[],"controls":["Inventory entry for the agent with an owner, its action policy and the resource scope it may touch","Change log of every automated action, with the resource, the saving and who or what approved it","Periodic review of the action policy as workloads and risk tolerance change","Rollback tested and documented for every category of automated action"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered on a schedule, with **custom functions**\n(REST calls) that read usage and billing data from the cloud provider's cost management API,\ncheck cost, error rate and latency after a change to catch a regression, and execute approved\nchanges through the provisioning API. An **AI agent** with **structured output** classifies\neach finding, and the **tool execution policy** limits which of these actions the agent may\ncall on its own, so only actions on the allow list run without a person in the loop.\n\nChanges above a configurable threshold go through **human in the loop confirmation with\napprove and reject controls**, carrying the recommendation, the expected saving and the\nresource it touches. **Run history with analytics, a full audit trail and downloadable run\ndata** keeps every recommendation, action and approval for the platform team and for finance\nto reconcile against the bill. The platform is model agnostic, so the model behind the\nclassification can be changed without rebuilding the workflow."},"faq":[{"question":"Can an AI agent be trusted to change production infrastructure on its own?","answer":"It depends on the policy the operator sets. Akamai uses Cast AI's automated bin packing, instance rightsizing and Spot instance automation to optimize the cost of its core Kubernetes infrastructure, and the interviewee describes the value of being able to \"turn on and forget\" the platform. nOps, in VERMEG's deployment, held only read only permissions and managed VERMEG's AWS commitments (Reserved Instances) with a buyback guarantee, without altering any infrastructure or configuration. Neither case study states whether it kept a separate human approval step for individual changes; a safe general design restricts unattended action to reversible changes and routes anything that touches a running production workload to an engineer for approval."},{"question":"How much can an AI agent actually save on cloud costs?","answer":"It depends heavily on how optimized the starting point already is. Akamai reports savings of 40 to 70% depending on the workload using Cast AI, covering only the workloads Cast AI manages, not Akamai's entire cloud bill. nOps reports that VERMEG cut its on demand AWS costs by more than 39% over ten months; part of that cut moved into committed Reserved Instance spend rather than disappearing as a net saving, so it is not directly comparable to Akamai's figure."},{"question":"Does this replace a FinOps team?","answer":"No. It removes the manual, repetitive part of the work, tracking usage, spotting waste and managing commitments, so a smaller team can cover more infrastructure and spend its time on architecture decisions and the policy that governs what the agent may do."}],"related":["aiops-incident-triage"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched and written from vendor case studies with named organizations, quotes verified against the live or archived source."},{"date":"2026-09-28","note":"Editorial pass: corrected vendor attribution and naming (Cast AI, not Cast.ai) in the meta description, FAQ and indicative value note; fixed the claimant on the VERMEG 39% figure to nOps; rewrote the FAQ on unattended production changes to describe each vendor accurately instead of a shared claim neither source supports; corrected blitsAi.howToBuild to match the platform feature inventory (human in the loop, run history analytics, tool execution policy); added an AWS pricing citation and corrected the AWS customer sample wording; named EU AI Act Annex III point 2 explicitly."},{"date":"2026-09-28","note":"Editorial pass: corrected the Akamai evidence year to 2024 on the live page's own datePublished markup (2024-04-23), the note that justified 2025 was wrong; removed the unsourced claim that Akamai removed a separate approval step once trust was established, here and in the Akamai evidence summary, both now say only what the case study supports (automated bin packing, rightsizing, Spot automation, and the \"turn on and forget\" quote); corrected the VERMEG evidence summary and this page's FAQ to say nOps managed VERMEG's AWS commitments (Reserved Instances) with a buyback guarantee, not \"Reserved Instances and Savings Plans\", which the source does not support; fixed the VERMEG summary wording to match its source (\"software solutions to the worldwide financial services industry\"); qualified howItWorks step 4 so buying or exchanging a commitment is called reversible only when it carries a buyback or exchange option, consistent with the commitment lock in failure mode; rewrote the capturedSavings note to stop calling the 0.4 to 0.6 range \"conservative against the evidence\", since part of VERMEG's reported cut moves into committed spend rather than disappearing as a net saving; corrected blitsAi.howToBuild wording to \"a configurable threshold\" and \"Run history with analytics\" to match the platform feature inventory verbatim; noted in both evidence files that archive.org was offline, so their Wayback captures could not be rechecked directly."},{"date":"2026-09-28","note":"Review fix round: removed the VERMEG evidence record's metric under the cost reduction KPI, since the 39% figure is a cut in on demand spend only, part of which moves into committed Reserved Instance spend, and that KPI measures fully loaded process cost, not a shift between pricing tiers; reworded this page's FAQ on savings to stop presenting that 39% as directly comparable to Akamai's savings figure. Added a second, reachable Cast AI source (its own product page) to the Akamai evidence record, since the case study itself never uses the words AI, machine learning or model, and noted in its verification that the product page's \"advanced predictive model for Kubernetes, trained on a massive dataset\" quote is the basis for treating Cast AI's automation as model driven."}],"slug":"cloud-cost-optimization-agent","url":"https://www.blits.ai/ai-use-cases/cloud-cost-optimization-agent","benchmarks":[{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40,"min":40,"max":40,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"akamai-kubernetes-cost-optimization","pooled":true}]}],"indicativeValueResult":{"low":300000,"high":900000},"evidence":["akamai-kubernetes-cost-optimization","vermeg-cloud-cost-reduction"]},{"title":"AI agent for complaints recognition, investigation and response","shortTitle":"Complaints handling","seoTitle":"AI agents for complaints handling and triage","metaDescription":"AI agents classify and investigate complaints for a human handler. Lloyds Banking Group cut classification to 1 second; NatWest is testing one with the FCA.","definition":"An AI agent that recognizes when a customer interaction is a complaint, logs it against the regulatory definition, classifies its root cause and severity, gathers the evidence, drafts the acknowledgement and the response for a human handler to approve, and tracks every statutory deadline until the case is closed.","aliases":["complaint management AI","complaint triage agent","dispute resolution assistant"],"industries":["cross-industry","banking","payments","insurance","telecommunications"],"functions":["case-management","customer-service","regulatory-compliance"],"patterns":["classification-and-routing","summarization","content-generation","agentic-workflow","rag-knowledge-assistant"],"channels":["agent-desktop","internal-tools","email","web-chat","voice"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"middle-office","problem":"In regulated industries a complaint is not whatever the customer calls a complaint. Financial\nregulators define it broadly (any expression of dissatisfaction, about the firm's service or\nproducts), and a complaint made in a phone call or a chat counts as much as a letter. Firms must\nacknowledge quickly, resolve within set deadlines, explain the outcome in writing and report\nvolumes and root causes. Missing one is a breach of the rules, not just a service failure.\n\nMuch of the handler's work sits around the decision: reading the history across systems, pulling\nstatements and call notes, working out what went wrong, and writing a response that is accurate,\nfair and clear. Meanwhile complaints hidden in ordinary conversations are\nnever logged, so they are never fixed. Customers now also use generative AI to write complaints,\nwhich the [UK Financial Ombudsman Service](https://www.financial-ombudsman.org.uk/businesses/resolving-complaint/our-insight/embracing-ais-transformational-impact-consumer-complaints)\nsays can produce long, unfocused submissions with fabricated laws, misquoted regulations or\ninvented past decisions.","problemStats":[],"howItWorks":"1. **Recognize.** Every channel (calls, chats, emails, letters, social) is screened for\n   expressions of dissatisfaction against the regulatory definition, so a complaint inside an\n   ordinary call is flagged instead of lost.\n2. **Log and classify.** The agent opens the case, sets the product, root cause and severity,\n   flags vulnerability and possible systemic issues, and starts the deadline clock.\n3. **Investigate.** It gathers the evidence from the relevant systems (transactions, call notes,\n   previous contacts, policies in force at the time) and writes a summary of what happened.\n4. **Draft.** It drafts the acknowledgement and the final response from approved wording, with\n   the reasoning and the redress calculation shown separately for the handler.\n5. **Decide and send, by a human.** A complaint handler reviews the evidence, decides the\n   outcome and approves the letter. The agent tracks deadlines and feeds root causes to\n   the teams that can fix them.","valueDrivers":["compliance","employee-productivity","speed","customer-experience"],"kpis":["processing-time-reduction","handling-time-reduction","time-saved-per-task","accuracy","hours-saved","cost-savings"],"indicativeValue":{"referenceOrg":"A retail bank that handles 50,000 complaints a year","inputs":[{"key":"complaints","label":"Complaints handled per year","low":50000,"high":50000,"unit":"complaints per year","note":"The reference bank."},{"key":"minutesSaved","label":"Handler minutes saved per complaint on classification, evidence gathering and drafting","low":5,"high":20,"unit":"minutes per complaint","note":"The low end is the only reported figure on this page, which covers classification alone (Lloyds Banking Group reports complaint classification in 1 second instead of about 5 minutes). The high end adds evidence gathering and drafting, for which no deployment has disclosed a figure; editorial assumption, replace with your own."},{"key":"costPerHour","label":"Fully loaded cost of a complaint handler hour","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption, replace with your own cost."}],"formula":"complaints * minutesSaved / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Complaint handler time released","caveat":"Handler time only. It leaves out lower redress and ombudsman fees from better first responses, fewer missed deadlines, the value of fixing root causes earlier and the cost of the AI and integrations."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Complaint classification is live at Lloyds Banking Group and investigation summaries are in testing at NatWest. The difficulty is reaching the evidence across many systems, keeping drafts factually right, recognizing complaints in unstructured conversations and proving to the regulator that nothing is missed.","dataPrerequisites":["The regulatory complaint definition and internal taxonomy of products, causes and severities","Historical complaints with outcomes and final response letters","Approved response templates, clauses and redress rules","Access to call transcripts and chat logs for recognition"],"integrations":["Complaint or case management system","Contact centre transcripts and chat platforms","Core banking, card and product systems for evidence","Document generation and correspondence","Management information and regulatory reporting"]},"implementation":{"steps":[{"title":"Start with recognition and logging","detail":"Screen transcripts and messages for complaints the contact centre did not log, and measure how many are found. This reduces conduct risk before any drafting is automated."},{"title":"Summarise the evidence for handlers","detail":"Build the investigation summary next, with links to every source, and measure how often handlers agree with it. NatWest's pilot follows this pattern: the agent investigates across data sources and presents a summary to a handler for approval."},{"title":"Draft responses from approved wording","detail":"Generate drafts from templates and clauses, with the decision and any redress set by the handler, and track edit rates per section."},{"title":"Close the loop on root causes","detail":"Aggregate causes across cases weekly and route them to product and process owners; spot systemic issues that affect many customers early."},{"title":"Test with your regulator in mind","detail":"Keep an evaluation set of real cases scored for task accuracy and hallucination, tracked over time. NatWest runs its trial inside the FCA's AI Live Testing with daily tracking of such metrics."}],"guardrails":["Every outcome and every response letter is approved by a human complaint handler","The agent may escalate a case to a complaint but never downgrade a flagged complaint on its own","Drafts cite the evidence they rely on; unsupported statements are flagged for the handler","Deadlines are computed by rules and alerted, never estimated by the model","Vulnerability and systemic issue flags route to specialist teams"],"humanInTheLoop":"Complaint handlers decide every outcome and approve every letter. Quality assurance samples AI classifications and drafts each week, and a senior owner signs off the recognition rules and thresholds, because a missed complaint is a conduct failure.","kpisToInstrument":["Complaints recognized in conversations that were not logged manually","Time from receipt to acknowledgement and to final response","Handler agreement with classifications and edit rate on drafts","Deadline breaches and ombudsman referral and overturn rates","Root causes identified and fixed"],"failureModes":[{"title":"Containment over recognition","detail":"A customer facing assistant tuned for containment answers a complaint as a question and never logs it. Screen every conversation against the complaint definition."},{"title":"Plausible but wrong responses","detail":"A draft misstates facts or policy and the handler, under time pressure, sends it. Show evidence next to every claim and track edit rates."},{"title":"Automated unfairness","detail":"Triage deprioritises complex or vulnerable cases. Test routing outcomes by customer group and keep humans on vulnerability."},{"title":"AI written complaints overwhelm triage","detail":"Long AI drafted submissions with invented legal references slow handling. Summarise the customer's actual points and check references before responding."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Complaint handling is not listed in Annex III, so internal classification and drafting for a handler who decides is minimal risk. Where the agent talks to customers to take the complaint, Article 50(1) requires telling them they are dealing with AI. Only a system that also assessed creditworthiness or priced life and health insurance (Annex III point 5(b) or 5(c)) would be high risk for that part."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","eu-psd2","dora","apra-cps-230","telecom-consumer-rules"],"guidance":[{"title":"DISP 1: Treating complainants fairly","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.handbook.fca.org.uk/handbook/DISP/1/","note":"The UK rules for complaint handling by financial firms, including prompt acknowledgement, the eight week time limit, final response requirements and complaint reporting."},{"title":"RG 271 Internal dispute resolution","issuer":"Australian Securities and Investments Commission","region":"asia-pacific","url":"https://asic.gov.au/regulatory-resources/find-a-document/regulatory-guides/rg-271-internal-dispute-resolution/","note":"ASIC's standards for internal dispute resolution by Australian financial firms, including what counts as a complaint and maximum response timeframes."},{"title":"Embracing AI's transformational impact on consumer complaints","issuer":"Financial Ombudsman Service","region":"europe","url":"https://www.financial-ombudsman.org.uk/businesses/resolving-complaint/our-insight/embracing-ais-transformational-impact-consumer-complaints","note":"The UK ombudsman's view of how AI is changing complaints, including consumers' use of generative AI and firms' automated triage."}],"controls":["Recognition rules mapped to the regulatory complaint definition, owned by a senior manager","Human approval of every outcome and final response, recorded in the case","Audit trail of classifications, evidence used and draft changes","Regular accuracy and hallucination testing on real cases","Root cause and systemic issue reporting to governance forums"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai recognition runs on the conversations the platform already handles: an **AI\nagent** with **structured output** labels dissatisfaction in chat and voice, and **human\nhandover** routes the case to the complaints team with a summary of the conversation. Emails\nand letters can reach the same process through the email channel and the **REST API**.\n\nInvestigation and drafting run as an **agentic workflow** that gathers evidence through\n**custom functions** and **SQL knowledge bases**, retrieves approved wording from a\n**knowledge base**, and stops for **human in the loop** approval before anything is sent, with\na full audit trail per run. **PII masking** masks personal data in prompts, **test\nsuites** with LLM based grading check accuracy on real cases, and **monitors** alert when a\nrecognition check fails. The platform is model agnostic, with EU and UAE data residency."},"faq":[{"question":"Can AI decide complaint outcomes?","answer":"It should not. NatWest's agent investigates and presents a summarised view to a complaint handler for approval, and the bank says all AI generated summaries are subject to strict human oversight. The other reported use on this page, at Lloyds Banking Group, is classifying complaints."},{"question":"Where does AI save the most time in complaints?","answer":"Among the deployments on this page, the only reported gain is in classification: Lloyds Banking Group reports that complaint classification takes 1 second instead of about 5 minutes. NatWest is testing an agent that gathers evidence from several data sources into one summary for the handler, but has not disclosed results."},{"question":"What is the biggest risk?","answer":"A complaint that is never recognized, for example because a customer facing bot treats it as a question to contain. Rules such as the FCA's DISP 1 require complaints to be recorded and resolved within set time limits, a prompt written acknowledgement and a final response within eight weeks, unless the complaint is resolved by the third business day, so recognition deserves as much testing as drafting."}],"related":["complaints-root-cause-analysis","correspondence-triage-and-routing","outbound-notice-drafting","email-and-ticket-reply-drafting","first-line-contact-centre-agent","card-dispute-and-chargeback-intake"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog and verified against the sources. The catalog's ABA Banking Journal source does not discuss complaints and was not used; NatWest's pilot is sourced from the bank itself."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: Lloyds evidence now cites the bank's 2025 results presentation (grade B, with a classification time metric) instead of press coverage that overstated the complaints share of the GBP 50 million; corrected the EU AI Act basis, added UK GDPR, sourced the ombudsman statement, tightened the value range, FAQ and ASIC note, and added the SEO title and description."},{"date":"2026-09-27","note":"Second fact check: howToBuild no longer names the complaint handling block (gated per deployment) and describes recognition with an AI agent and structured output; the FAQ now says the only reported gain is in classification and that NatWest has not disclosed results; the minutes saved range is lowered to 5 to 20, anchored on the Lloyds Banking Group classification figure; added PSD2; dropped the Resultsense press source from the Lloyds evidence."},{"date":"2026-09-27","note":"Review fixes: scoped the classification gain FAQ answer to the deployments on this page rather than the whole market, and corrected the DISP 1 description, complaints must be recorded and resolved within set time limits, not every complaint acknowledged, since a complaint resolved by the third business day gets a summary resolution communication instead."}],"slug":"complaints-handling-agent","url":"https://www.blits.ai/ai-use-cases/complaints-handling-agent","benchmarks":[{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":5,"min":5,"max":5,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"lloyds-banking-group-ai-complaints-processing","pooled":true}]}],"indicativeValueResult":{"low":166666.6666666667,"high":1166666.6666666667},"evidence":["lloyds-banking-group-ai-complaints-processing","natwest-agentic-ai-complaints-handling"]},{"title":"AI agent for corporate credit analysis and credit memo drafting","shortTitle":"Credit underwriting and memos","seoTitle":"AI credit memo drafting for corporate banks","metaDescription":"AI agents spread borrower financials and draft credit memos for bankers to challenge and sign. DBS rolled one out to about 1,500 staff after a 150 user pilot.","definition":"An AI agent that gathers a corporate borrower's documents and data, spreads the financials into the bank's template, calculates ratios and covenant headroom, pulls bureau and news information, and drafts a committee ready credit memo in which every figure links to its source, for the relationship and credit teams to challenge, complete and sign.","aliases":["credit memo generator","AI credit analyst","financial spreading automation","credit proposal drafting"],"industries":["banking"],"functions":["lending-and-credit","underwriting","risk-management"],"patterns":["document-processing","agentic-workflow","rag-knowledge-assistant","content-generation"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"specialized-businesses","problem":"A corporate credit memo is a long document built from many sources: audited accounts, management\naccounts, projections, industry research, bureau data, internal exposure and conduct records, and\nthe bank's own credit policy. Relationship managers sift through annual reports, industry research\nand internal records to build each memo; DBS says preparing credit memos and related credit\nactivities can take up to 40% of a relationship manager's time.\n\nThat is time bankers do not spend with clients. An agent can do the assembly, spreading and first\ndraft, and a shared template with deterministic calculations also keeps memos consistent between\nauthors. But credit is a regulated decision: the value only holds if every number is traceable and\npeople still own the judgement and the approval.","problemStats":[{"statement":"DBS says preparing credit memos and related credit activities can account for up to 40% of a relationship manager's time.","sourceTitle":"DBS scales agentic AI to transform way of working for corporate bankers","sourceUrl":"https://www.dbs.com/newsroom/DBS_scales_agentic_AI_to_transform_way_of_working_for_corporate_bankers_freeing_up_time_for_more_strategic_client_engagements","year":2026}],"howItWorks":"1. **Collect the file.** The agent gathers financial statements, projections and supporting\n   documents from the client portal, email and document store, and lists what is missing.\n2. **Extract and spread.** Document AI extracts line items and maps them into the bank's spreading\n   template, flagging items that need an analyst's judgement.\n3. **Analyse.** It calculates ratios, trends and covenant headroom, and pulls bureau data, news,\n   internal exposure and conduct records through approved connectors.\n4. **Check against policy.** Retrieval over the credit policy and sector guidelines highlights\n   exceptions and required approvals.\n5. **Draft the memo.** It writes each section of the memo in the bank's format, with a link from\n   every figure to its source page and a list of risk flags and open questions.\n6. **Iterate and sign.** The relationship manager and credit risk manager challenge the draft, ask\n   the agent for deeper research, edit, and take it through the normal approval.","valueDrivers":["employee-productivity","speed","risk-reduction","compliance"],"kpis":["time-saved-per-task","processing-time-reduction","productivity-gain","users-served","accuracy"],"indicativeValue":{"referenceOrg":"A corporate bank preparing 2,000 credit memos a year for new facilities and annual reviews","inputs":[{"key":"memos","label":"Credit memos per year","low":2000,"high":2000,"unit":"memos per year","note":"The reference bank."},{"key":"hoursPerMemo","label":"Hours of relationship manager and analyst work per memo","low":20,"high":40,"unit":"hours per memo","note":"Editorial assumption, replace with your own time study."},{"key":"timeSaved","label":"Share of that time saved","low":0.15,"high":0.3,"unit":"fraction of time","note":"Editorial assumption. The upper bound equals the at least 30% goal DBS has set, which is a target and not yet a measured result.","sourceUrl":"https://www.dbs.com/newsroom/DBS_scales_agentic_AI_to_transform_way_of_working_for_corporate_bankers_freeing_up_time_for_more_strategic_client_engagements"},{"key":"hourlyCost","label":"Loaded cost of a banker or analyst hour","low":90,"high":140,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."}],"formula":"memos * hoursPerMemo * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Banker and analyst time released, valued at loaded cost","caveat":"Values released time only. It leaves out the cost of the agent, data licences and model risk validation, and any effect on credit quality, faster time to yes for clients or revenue from the time bankers win back."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Many systems, a regulated decision and model risk validation. Spreading accuracy on messy financial statements, integration with the loan origination system and the bank's approval workflow, and a clear audit trail are the hard parts.","dataPrerequisites":["The bank's spreading template, memo template and credit policy in machine readable form","Historical memos and spreads to test against","Access to bureau, rating and news sources licensed for this use","Internal exposure, limit and conduct data per client group"],"integrations":["Loan origination or credit workflow system","Document management and client portal","Bureau, rating agency and news data providers","Core lending and limits systems","CRM for client context"]},"implementation":{"steps":[{"title":"Start with annual reviews","detail":"Annual reviews of existing clients have prior memos and spreads to compare with, which makes accuracy measurable and the change less risky than new to bank credit."},{"title":"Decompose the memo into tasks","detail":"Break the memo into its tasks (spreading, ratio analysis, industry section, peer comparison, policy exceptions) and automate them one by one; DBS says its agents handle more than 70 tasks."},{"title":"Make traceability non negotiable","detail":"Link every figure to a source page and every statement to a document or data source, and block the draft from moving on while any figure lacks a source."},{"title":"Validate like a model","detail":"Run the agent on past files, compare spreads and ratios with the approved versions, and take the results through model risk validation before live use."},{"title":"Pilot with a small group, then scale","detail":"Pilot with experienced relationship and credit managers, capture their corrections, and widen the rollout only when error rates are stable; DBS went from 150 pilot users to about 1,500."}],"guardrails":["The agent drafts; credit decisions and approvals follow the bank's existing authority matrix","Every figure and statement in the memo links to its source","Numbers are calculated by deterministic code, not generated by the language model","Retrieved documents and news are treated as data, never as instructions","Version history of each draft, with the human edits, is retained with the credit file"],"humanInTheLoop":"Relationship managers and credit risk managers review, challenge and complete every draft, and the approval follows the normal credit authority. Model validation reviews the agent before use and periodically after, and credit risk samples memos to check that the analysis did not become thinner.","kpisToInstrument":["Elapsed time from complete file to memo ready for approval","Analyst and banker hours per memo, from time studies","Spreading accuracy against approved spreads on a sample","Share of memo figures edited by humans, by section","Credit committee questions or returns per memo, before and after"],"failureModes":[{"title":"Fluent but wrong numbers","detail":"A misread statement or unit error flows into ratios and the narrative. Use deterministic calculation, reconciliation checks and source links on every figure."},{"title":"Automation bias","detail":"Reviewers accept the draft instead of analysing the credit. Track edit rates and committee challenges, and keep sections that need judgement explicitly blank for the banker."},{"title":"Stale or unlicensed data","detail":"News or ratings are outdated or not licensed for AI use. Record the as of date and licence of every external source."},{"title":"Scope creep into individual lending","detail":"The same agent is reused for sole traders or personal guarantors, which changes the regulatory tier. Assess each new borrower segment before use."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 5(b) makes AI used to evaluate the creditworthiness of natural persons high risk. Credit analysis of companies is outside that point, but the tier can change when the same system evaluates the creditworthiness of natural persons, such as sole traders, partners who are personally liable or personal guarantors. Design the scope explicitly and document it."},"regulations":["eu-ai-act","eba-loan-origination","gdpr","apra-cps-230"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) covers creditworthiness evaluation and credit scoring of natural persons, which sets the boundary for this use case."},{"title":"EBA Guidelines on loan origination and monitoring","issuer":"European Banking Authority","region":"europe","url":"https://www.eba.europa.eu/activities/single-rulebook/regulatory-activities/credit-risk/guidelines-loan-origination-and-monitoring","note":"Expectations on creditworthiness assessment, the information to collect and the use of automated models in credit decisions."},{"title":"MAS consultation paper on proposed Guidelines on AI Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Consultation issued 13 November 2025, not final guidelines. Proposed expectations on AI inventory, risk materiality and human oversight, relevant for agents that support credit decisions."}],"controls":["Model inventory entry and independent validation before live use","Documented borrower scope, with a check that no natural persons are assessed without the high risk controls","Source logging and draft version history retained with the credit file","Periodic back testing of spreads and memo quality","Clear accountability, with the approver named in the authority matrix owning the decision"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow**: an agent loop with **custom functions** for the\nspreading and ratio calculations (deterministic code in the isolated sandbox), connectors to the\nloan origination, exposure and bureau systems through REST calls or **MCP**, and a **knowledge\nbase** with hybrid retrieval over the credit policy and sector guidelines. Borrower documents in\nPDF, XLSX and DOCX go through document ingestion into vector storage, and **structured output**\nfills the memo template section by section.\n\n**Human in the loop approval** can be required before steps that write to the credit workflow, and\neach run keeps a full **audit trail**. **Guardrails** and **PII\nmasking** protect client data in prompts, **test suites** replay past credit files to catch\nregressions, and the platform is model agnostic, so the bank can choose a model per task and keep\ndata in the EU or UAE region."},"faq":[{"question":"How much faster can AI make a credit memo?","answer":"Public results are still mostly targets. DBS has set a goal of reducing the time spent by at least 30% and rolled its agentic solution out to about 1,500 employees. Measure your own baseline per memo type before claiming a saving."},{"question":"Is AI credit memo drafting high risk under the EU AI Act?","answer":"Not for companies. Annex III point 5(b) covers creditworthiness evaluation of natural persons, so corporate credit analysis is outside it, but assessing sole traders or personal guarantors can bring the system into scope."},{"question":"Who is accountable for the memo?","answer":"The relationship manager and credit risk manager who complete it, and the approver in the credit authority matrix. The agent prepares a draft; it does not recommend approval on its own."}],"related":["sme-cash-flow-underwriting","credit-early-warning-monitoring","client-briefing-and-call-report-copilot","business-onboarding-and-ubo-discovery","underwriting-risk-assessment-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog and verified against the DBS announcement; the 30% figure is recorded as a goal, not a result, and the catalog's unnamed US bank pilot was dropped for lack of a source."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: DBS figures rechecked on the live release; the unsourced claim that analysts spend days on spreading was replaced with the DBS 40% statement; the EU AI Act basis and FAQ were tightened to natural persons under Annex III point 5(b); Banestes evidence now follows the Google Cloud wording and source date; Blits.ai build notes aligned with the feature inventory; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: the problem section now follows the DBS release (sifting annual reports, industry research and internal records; up to 40% of relationship manager time) and no longer states unsourced claims about spreading, copying text or memos differing by author; the MAS guidance is labelled as a consultation; the DBS task count and Blits.ai audit and approval wording were tightened."},{"date":"2026-09-27","note":"Review fixes: dropped the superseded SR 11-7 and the MAS AI risk management guidelines, which are not yet final, from risk.regulations, keeping only regulations already in force; the Banestes evidence now cites the primary Google Workspace customer story with the Banestes CTO's quote instead of stating no primary story was found."}],"slug":"credit-memo-drafting-agent","url":"https://www.blits.ai/ai-use-cases/credit-memo-drafting-agent","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1500,"min":1500,"max":1500,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dbs-agentic-credit-memo","pooled":true}]}],"indicativeValueResult":{"low":540000,"high":3360000},"evidence":["banestes-gemini-credit-analysis","dbs-agentic-credit-memo"]},{"title":"AI agent for data quality monitoring and observability","shortTitle":"Data quality monitoring agent","seoTitle":"AI agent for data quality monitoring","metaDescription":"An AI agent flags data quality problems before they reach a report. Monte Carlo reports SeatGeek halved root cause effort and Contentsquare cut detection time 17%.","definition":"An AI agent that watches data pipelines and tables continuously, uses machine learning to learn the normal pattern of freshness, volume, schema and distribution for each one, flags anomalies before they reach a dashboard or a downstream model, and traces the lineage back to the change that caused them so an engineer can fix the source, not just the symptom.","aliases":["AI data observability","data quality agent","ML enabled anomaly detection for data","data incident detection AI","automated data quality monitoring"],"industries":["cross-industry","technology","retail-and-ecommerce"],"functions":["it-and-engineering","analytics-and-reporting"],"patterns":["anomaly-detection","classification-and-routing","summarization"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","problem":"A broken dbt model, a schema change upstream, or a partner feed that silently stops updating do\nnot throw an error. The pipeline runs, the table populates, and the first sign of trouble is\noften a business user asking why a report looks wrong. By the time someone notices, the bad data\nmay already have reached a dashboard, a finance close or a model that scores customers.\n\nFinding the cause is its own project. Brian London, SeatGeek's Director of Data Engineering,\ndescribed the pattern before the team adopted data observability: \"the way we would find out\nthere was a problem, most of the time, is one of the business users would post a Slack message,\nsaying that they're getting results that don't make sense.\" Monte Carlo's case study on the\ndeployment reports that SeatGeek's data teams were losing full days root causing data anomalies\nthat their business users had already found.","problemStats":[],"howItWorks":"1. **Learn the baseline.** For every monitored table and pipeline, the agent learns the normal\n   pattern of freshness, row volume, null rates, distributions and schema, from historical runs.\n2. **Detect anomalies in real time.** New data is compared against that baseline as it lands, and\n   deviations are flagged before a scheduled report or model run consumes the data.\n3. **Trace the lineage.** Field level lineage shows which upstream tables, jobs and models feed\n   the affected asset, so root causing an anomaly means following a lineage graph instead of\n   manually querying every candidate source.\n4. **Rank and route.** Anomalies are grouped into incidents, ranked by the number of downstream\n   assets and users they affect, and routed to the team that owns the source.\n5. **Confirm and learn.** An engineer confirms the cause and the fix; confirmed incidents refine\n   future ranking and give the team a record of recurring failure points to fix at the source.","valueDrivers":["risk-reduction","employee-productivity","speed"],"kpis":["mttr-reduction","productivity-gain","error-reduction"],"indicativeValue":{"referenceOrg":"A data team that logs 10 data quality incidents a month across its pipelines","inputs":[{"key":"incidentsPerMonth","label":"Data quality incidents per month before monitoring","low":10,"high":10,"unit":"incidents per month","note":"The reference organization, based on the baseline Monte Carlo reports for SeatGeek before it adopted data observability."},{"key":"hoursPerIncident","label":"Engineering hours lost root causing one incident","low":4,"high":8,"unit":"hours per incident","note":"Editorial assumption for a mid sized data team. Replace with your own incident retrospective data."},{"key":"reduction","label":"Reduction in root cause effort from automated anomaly detection and lineage","low":0.15,"high":0.5,"unit":"fraction of hours","note":"The range spans the two vendor reported results on this page, which are not directly comparable: Monte Carlo reports SeatGeek cut root cause resource drain, an effort measure, by 50%, while Contentsquare's 17% and 16% figures measure elapsed detection and resolution time, not effort. Treat this as a rough range to replace with your own incident retrospective data; results depend on how much of the pipeline has lineage mapped."},{"key":"hourlyCost","label":"Fully loaded cost of a data engineer","low":70,"high":110,"unit":"USD per hour","note":"Editorial assumption. Replace with your own rate."}],"formula":"incidentsPerMonth * 12 * hoursPerIncident * reduction * hourlyCost","currency":"USD","period":"per year","resultLabel":"Annual data engineering time released from data incident response","caveat":"Engineering time only. It leaves out the subscription cost of the observability platform, the revenue and trust cost of bad data that does reach a report or a model, and any reduction in the total number of incidents rather than just the time to resolve them."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Connecting to a warehouse or lakehouse and turning on default anomaly detection is fast. Field level lineage and low noise alerting take longer, and depend on consistent naming, documented ownership and a data catalog the agent can use for context.","dataPrerequisites":["Read access to the data warehouse, lakehouse or pipeline orchestrator","Table and pipeline ownership recorded somewhere the agent can read","History of past incidents and their confirmed root cause, to tune ranking","A data catalog or glossary, so anomalies can be described in business terms"],"integrations":["Data warehouse or lakehouse (Snowflake, BigQuery, Databricks and similar)","Orchestration tool for pipeline and job metadata (Airflow, dbt and similar)","Business intelligence tool, to trace which dashboards an anomaly reaches","Chat tool for incident alerts and ChatOps triage","Ticketing system for confirmed incidents that need a fix"]},"implementation":{"steps":[{"title":"Start with the tables people already distrust","detail":"Connect the sources feeding the reports and models with a known history of quiet failures first, so the first alerts are on something a stakeholder will recognize and value."},{"title":"Tune before you trust","detail":"Default anomaly thresholds are noisy on a new data set. Spend the first weeks tuning sensitivity per table and suppressing known, benign patterns before routing alerts widely."},{"title":"Map lineage where it matters most","detail":"Prioritize lineage for the pipelines with the most downstream consumers, since that is where a fast root cause saves the most time and the most trust."},{"title":"Route to the owner, not a shared queue","detail":"An anomaly that lands in a queue nobody owns gets ignored. Route each alert to the team whose pipeline caused it, with the lineage and the affected downstream assets attached."},{"title":"Close the loop on confirmed incidents","detail":"Record the confirmed cause and the fix for every incident, and use that history to reduce noise and to find the pipelines that fail repeatedly and need to be rebuilt, not just fixed."}],"guardrails":["Read only access to source systems; the agent never writes back to production data","Alert thresholds tuned per table, with known benign patterns suppressed rather than silenced entirely","Every alert shows its evidence (the metric, the baseline, the lineage) so it can be checked in seconds","Anomalies are routed to a named owning team, never a shared, unowned queue","Sensitive fields excluded from anomaly previews and alert payloads"],"humanInTheLoop":"An engineer confirms every incident's root cause and decides the fix; the agent detects, ranks and traces lineage, it does not change data or pipelines itself. Data platform leads review alert noise and confirmed incident patterns periodically to retune thresholds and prioritize fixes.","kpisToInstrument":["Data incidents detected before a downstream user reports them, versus after","Time from anomaly detection to root cause confirmation","Alert to confirmed incident ratio, to track noise","Recurrence rate of incidents from the same source pipeline"],"failureModes":[{"title":"Alert fatigue from an untuned baseline","detail":"A new table triggers noisy alerts until enough history exists to learn its normal pattern. Start monitoring in a silent mode and tune before routing alerts."},{"title":"Lineage gaps hide the real source","detail":"An anomaly is traced only as far as the lineage graph reaches, and stops short of the true upstream cause. Prioritize mapping lineage for high impact pipelines first."},{"title":"Confirmed incidents never feed back","detail":"The same failure recurs because nobody tracked it back to a source that needs rebuilding. Keep a record of confirmed causes and review recurring ones on a cadence."},{"title":"Anomaly detection on the wrong signal","detail":"A model flags a real, expected change (a new market launch, a seasonal pattern) as an anomaly. Let owners mark expected changes so the baseline updates instead of alerting every time."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"An internal data engineering tool that flags anomalies in pipelines and tables; it is not a use listed in Annex III and makes no decision about a natural person. If the monitored data feeds a high risk system, such as a credit or employment decision, the AI Act obligations attach to that downstream system, not to this monitoring layer."},"regulations":["gdpr","nist-ai-rmf","iso-42001"],"guidance":[],"controls":["Inventory entry for the monitoring agent with an owner and the data sources in scope","Read only access enforced and reviewed periodically","Confirmed incident log kept for audit and for retuning alert thresholds","Sensitive and personal data excluded from alert previews by default"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered on a schedule or called by the\norchestrator via API. **Custom functions** (SQL queries and REST calls) pull freshness,\nvolume, distribution and schema metrics from a warehouse such as Snowflake, BigQuery or\nDatabricks, and pipeline and lineage metadata from the orchestration tool. Another custom\nfunction, outside the agent's own tool set, writes each run's readings into a table in a\n**SQL knowledge base** (PostgreSQL or SQLite). An **AI agent** with **structured output**\nqueries that table as its baseline, compares each new reading against it, ranks the resulting\nincidents by downstream impact, and summarizes the likely cause. A **tool execution policy**\nlimits which tools the agent may call, restricting it to the read functions that pull metrics\nand lineage.\n\nRanked incidents route to the owning team through the ready made **Microsoft Teams**\nintegration, or to Slack through a custom function's REST call or the integration catalog,\nwith the anomaly, the lineage and the affected downstream assets attached. **Human in the\nloop** approval lets the owning engineer approve or reject the follow up action the agent\nproposes, such as opening a ticket or posting the incident, rather than let it go out\nunattended. **Run history with a full audit trail, analytics and downloadable run data**\nkeeps every detection and every approval decision available for later analysis. A separate\n**monitor** can run a scheduled health check on the monitoring agent itself. The platform is\nmodel agnostic, with EU and UAE data residency for the metadata the agent processes."},"faq":[{"question":"Is this the same as an AIOps incident triage agent?","answer":"No, though the two are close cousins. AIOps incident triage correlates application and infrastructure alerts during a live outage. A data quality monitoring agent watches tables and pipelines for freshness, volume and schema problems, often before anyone would call it an incident at all, and traces the issue through data lineage rather than a service map."},{"question":"How much manual data quality work does this actually remove?","answer":"Monte Carlo reports that SeatGeek, a ticketing marketplace, cut its root cause resource drain by 50% and went from about 10 data incidents a month to zero in the second quarter after enabling its platform, and that Contentsquare cut its time to detect a data incident by 17% and its time to resolution by 16% in one month. Results depend on how much lineage is mapped and how noisy the starting baseline is."},{"question":"What should be monitored first?","answer":"The tables and pipelines that feed reports or models people already distrust. Early wins on data stakeholders already care about build the credibility to expand coverage."}],"related":["aiops-incident-triage"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched and written from vendor case studies with named organizations, quotes verified against the live source."},{"date":"2026-09-28","note":"Editorial pass: scoped the Monitors and baseline claims in blitsAi.howToBuild to what the feature inventory supports, attributed the SeatGeek and Contentsquare figures to Monte Carlo throughout, widened the indicative value reduction range, used the precise \"second quarter\" wording, and corrected the evidence publish date notes."},{"date":"2026-09-28","note":"Fixed adversarial review blockers: Contentsquare evidence year set to 2023 (the case study's publish date, not the unrelated Series F sentence) with the note rewritten; SeatGeek evidence title made neutral instead of stating the vendor's result as fact; blitsAi.howToBuild rewritten so the run history baseline is a custom function writing to a SQL knowledge base table, and human in the loop is approval of a follow up action, not confirmation of a diagnosis, and Teams/Slack and dashboard/PII wording softened to match the feature inventory. Also fixed the problem paragraph's conflated SeatGeek quote, added a note that the Contentsquare and SeatGeek indicativeValue figures measure different things, tightened the FAQ's \"one month\" wording, removed the loose related link, and reworded metaDescription so it no longer implies SeatGeek and Contentsquare used an AI agent."}],"slug":"data-quality-monitoring-agent","url":"https://www.blits.ai/ai-use-cases/data-quality-monitoring-agent","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"seatgeek-data-quality-observability","pooled":true}]},{"kpi":"mttr-reduction","label":"Time to repair reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":17,"min":17,"max":17,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"contentsquare-data-incident-detection","pooled":true}]}],"indicativeValueResult":{"low":5040,"high":52800},"evidence":["contentsquare-data-incident-detection","seatgeek-data-quality-observability"]},{"title":"AI agent for device and connectivity troubleshooting on voice and chat","shortTitle":"Device and connectivity troubleshooting","seoTitle":"AI agent for broadband and router troubleshooting","metaDescription":"AI agents fix broadband and device faults by chat or voice. Singtel reports 73% of mobile and home troubleshooting cases resolved without a Customer Care officer.","definition":"An AI agent that diagnoses and fixes a customer's broadband, mobile, TV or device problem by conversation on the phone or in chat, running line tests and remote resets through the operator's systems, guiding the customer step by step, and booking an engineer or handing over to a technician when the fault needs a person.","aliases":["AI technical support agent","broadband fault voice agent","router troubleshooting bot","virtual technician","tech support chatbot"],"industries":["telecommunications"],"functions":["customer-service","field-service"],"patterns":["conversational-agent","voice-agent","agentic-workflow","rag-knowledge-assistant","computer-vision"],"channels":["voice","web-chat","mobile-app","whatsapp"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"\"My internet is not working\" is one of the reasons customers contact a telecom operator. Many\nfaults are resolved by the same few steps: check for an outage, run a line test, restart the\nrouter, check the cabling, change a setting on the phone. Yet each call ties up a trained agent,\nand a fault that could have been fixed remotely can still end in an engineer visit.\n\nThis use case adds an agent that can call the same diagnostics itself, read what the\ncustomer describes or photographs, and adapt the steps as it goes.","problemStats":[],"howItWorks":"1. **Identify the customer and the service.** The agent authenticates the caller and finds the\n   affected line, device or package.\n2. **Rule out the network first.** It checks for known outages or maintenance in the area and,\n   if there is one, explains it and gives an estimated fix time instead of troubleshooting.\n3. **Run diagnostics.** Through approved tools it runs a line test, reads the router or modem\n   status and recent faults, and can trigger a remote reset or reprovisioning.\n4. **Guide the customer.** It walks the customer through the remaining steps in plain language,\n   on voice or chat, and can interpret a photo of a device's lights or error screen.\n5. **Escalate with the evidence.** If the fault persists it books an engineer or passes the case\n   to a human technician with the diagnostics already run, so nobody repeats the tests.","valueDrivers":["cost-to-serve","customer-experience","speed","employee-productivity"],"kpis":["containment-rate","first-contact-resolution","interactions-handled","time-saved-per-task","handling-time-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"An operator with 2 million broadband and mobile customers","inputs":[{"key":"customers","label":"Broadband and mobile customers","low":2000000,"high":2000000,"unit":"customers","note":"The reference operator."},{"key":"faultContactRate","label":"Technical fault contacts per customer per year","low":0.4,"high":0.8,"unit":"contacts per customer per year","note":"Editorial assumption, replace with the technical share of your contact reason report."},{"key":"resolvedShare","label":"Share of fault contacts the agent resolves without a human","low":0.25,"high":0.5,"unit":"fraction of fault contacts","note":"Editorial assumption: set below the Singtel benchmark on this page (73% of mobile and home troubleshooting cases resolved without a Customer Care officer in its initial results) because voice calls and complex home network faults are expected to resolve less often; replace with your own pilot data."},{"key":"costPerContact","label":"Cost of a human handled technical contact","low":6,"high":10,"unit":"USD per contact","note":"Editorial assumption for a technical support contact, which usually takes longer than a general one. Replace with your own fully loaded cost."}],"formula":"customers * faultContactRate * resolvedShare * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled technical contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the larger saving from engineer visits that are avoided or better prepared, the cost of the AI and the diagnostic integrations, and any repeat contacts from faults the agent closed too early."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Knowledge based troubleshooting is simple. The value comes from connecting diagnostics (line tests, device management, outage status) and engineer booking, which live in network and field systems that were not built for real time use by a conversation.","dataPrerequisites":["Troubleshooting guides per device and technology with an owner and review date","Outage and planned maintenance status by area, available by API","Fault contact reasons with their eventual root cause, for testing","A list of fault types that must always go to a technician"],"integrations":["Network diagnostics and line test tools","Device management for routers and set top boxes (status, reboot, reprovision)","Outage and maintenance status service","Field service scheduling for engineer appointments","Contact centre platform for handover with diagnostic results"]},"implementation":{"steps":[{"title":"Start with the faults that have a known fix","detail":"Take the top fault reasons by volume and pick those with a reliable remote fix, such as router restarts, settings on the phone and account provisioning errors. Leave intermittent and physical faults to people at first."},{"title":"Connect diagnostics before you tune the conversation","detail":"An agent that can run a line test and read router status can resolve faults that questions alone cannot. Build the diagnostic tools and their error handling first."},{"title":"Always check for an outage first","detail":"Make the outage check the first tool call on every fault conversation, so the agent never walks a customer through resets during a network incident."},{"title":"Define the escalation evidence","detail":"Agree with field operations what a technician or engineer needs from the agent (tests run, results, customer availability) and pass it in every escalation."},{"title":"Pilot on selected calls and listen","detail":"Route a small share of routine fault calls to the agent, monitor calls, measure repeat faults within 14 days and expand only when they hold."}],"guardrails":["Outage check before any troubleshooting step","Only allow listed remote actions, each with a confirmation to the customer","Handover to a person on request at any time, on voice and chat","Detection of vulnerable customers and telecare or medical alarm users, who go straight to a person","Personal data masked before text reaches a model or the logs"],"humanInTheLoop":"Technicians and engineers own every fault the agent cannot fix and every physical repair. Supervisors monitor live calls during the pilot, review a weekly sample of resolved cases against repeat faults, and approve every new remote action before the agent can use it.","kpisToInstrument":["Resolution rate per fault type, counting a repeat fault contact within 14 days as unresolved","Engineer visits booked per 1,000 fault contacts, and visits that found no fault","Average handling time of escalated cases, compared with cases without the agent","Customer satisfaction on resolved and escalated conversations","Share of fault conversations that started during a known outage"],"failureModes":[{"title":"Troubleshooting during an outage","detail":"Customers are asked to reset routers while the network is down. Make the outage check mandatory and visible."},{"title":"False resolution","detail":"The customer says it works now, the fault returns the next day. Measure repeat faults, not the end of the conversation."},{"title":"Endless loops","detail":"The agent repeats the same steps with a frustrated caller. Limit attempts and escalate with the evidence."},{"title":"Missing the vulnerable customer","detail":"A customer relying on a telecare alarm is left in automation. Detect these signals and hand over at once."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing troubleshooting agent must tell people they are interacting with AI unless that is obvious (Article 50). It is not high risk as long as it is not used as a safety component in the management and operation of critical digital infrastructure (Annex III, point 2); running line tests and resets for one customer's service does not make it one."},"regulations":["eu-ai-act","gdpr","uk-gdpr","telecom-consumer-rules","eecc"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Ofcom fines Virgin Media £23.8 million for putting vulnerable customers at risk of harm","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/phones-and-broadband/vulnerable-customers/ofcom-fines-virgin-media-23.8-million-for-putting-vulnerable-customers-at-risk-of-harm","note":"Ofcom fined Virgin Media for failing to identify and protect telecare customers during its move to digital landlines, under General Condition C5.2 on the fair treatment of vulnerable consumers. The same duty applies when an automated agent handles their faults."}],"controls":["AI disclosure at the start of every call and chat","Documented list of remote actions with owners, tests and rollback","Call recording and transcript retention in line with local rules","Regression tests on real fault scenarios for every change to prompts, tools or model","Monitoring of repeat faults and complaints that mention the agent"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** for the outage check, line test,\nrouter status and remote reset, and for booking engineer appointments, grounded in a\n**knowledge base** of troubleshooting guides with hybrid retrieval. A **flow** enforces the\norder of steps (outage check first, confirmation before a remote action), and customers can\nupload a photo of the device in chat.\n\nThe agent runs on **voice** with streaming speech recognition and synthesis, DTMF and call\ntransfer, on **web chat** and **WhatsApp**, and in the mobile app through the API channels. **Human handover** passes the\ndiagnostics and summary to a technician, supervisors can monitor and take over live calls,\n**guardrails** and **PII masking** protect the conversation, **test suites** replay real fault\nscenarios on every change, and **monitors** check key journeys on a schedule. The platform is\nmodel agnostic with EU and UAE data residency."},"faq":[{"question":"What share of technical support contacts can an AI agent resolve?","answer":"Few operators publish results, and published voice results are scarce. Singtel, which launched its agentic assistant across chat and voice, reports that 73% of mobile and home troubleshooting cases were resolved without a Customer Care officer, without saying which channel produced that figure. Genesys reports that first contact success on Vodafone Germany's messaging channels rose from 16% at launch to 44%, for a TOBi chatbot built on IBM Watson before generative AI. Virgin Media O2 began with selected routine broadband fault calls in September 2026 and has not published results yet."},{"question":"Should an AI agent troubleshoot on the phone or only in chat?","answer":"Both work. Chat lets customers read instructions and send photos, such as the flashing router light that Vodafone Germany's TOBi recognises, while voice reaches customers whose home internet is down. Singtel launched its agent across chat and voice, and Vodafone runs TOBi on digital channels and telephony. Vodafone Germany's published figure covers messaging only, and Virgin Media O2 started with a narrow set of voice calls that its teams monitor."},{"question":"How does this relate to outage communication?","answer":"They share the outage check. The troubleshooting agent must know about an outage before it starts, and the outage communication agent tells affected customers before they call."}],"related":["network-outage-communication-agent","field-technician-copilot-and-dispatch","bill-explanation-and-billing-dispute-agent","order-to-activation-and-esim-onboarding-assistant","first-line-contact-centre-agent","predictive-network-maintenance"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by the editor after a targeted blocker check."},{"date":"2026-09-25","note":"First version, written for the telecommunications vertical from Singtel, Vodafone and Virgin Media O2 sources checked word for word."},{"date":"2026-09-25","note":"Consolidation pass: added European Electronic Communications Code to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: neutral problem statement, FAQ claims attributed to their claimants, Annex III point 2 named in the AI Act basis, UK GDPR added, Ofcom note corrected; Singtel evidence regraded B after checking the Singtel release; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: FAQ no longer claims the deployments started in chat or that voice is newer, and says which channels each published figure covers; softened unsourced claims in the problem and implementation steps; Article 50 now links to EUR-Lex."}],"slug":"device-and-connectivity-troubleshooting-agent","url":"https://www.blits.ai/ai-use-cases/device-and-connectivity-troubleshooting-agent","benchmarks":[{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":3,"nUpTo":0,"median":70,"min":44,"max":73,"byClaimant":{"organization":1,"vendor":2,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"singtel-shirley-agentic-ai-agent","pooled":true},{"id":"vodafone-tobi-virtual-assistant","pooled":true},{"id":"vodafone-germany-tobi-messaging","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":22535000,"min":70000,"max":45000000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"vodafone-tobi-virtual-assistant","pooled":true},{"id":"singtel-shirley-agentic-ai-agent","pooled":true}]}],"indicativeValueResult":{"low":1200000,"high":8000000},"evidence":["singtel-shirley-agentic-ai-agent","virgin-media-o2-broadband-fault-voice-agent","vodafone-germany-tobi-messaging","vodafone-tobi-virtual-assistant"]},{"title":"AI agent for drafting employee performance reviews","shortTitle":"Performance review drafting","seoTitle":"AI performance review drafting for managers","metaDescription":"AI drafts a first version of each performance review from collected feedback. Windmill reports an 83% drop in review hours at Rho and 93% employee preference.","definition":"An assistant that gathers an employee's work history, goals and peer feedback from the systems a manager already uses, and drafts a first version of the performance review for the manager to edit, rewrite or reject, so the manager starts from a grounded summary instead of a blank form and a stack of six months of context to recall from memory.","aliases":["AI performance review generator","AI assisted review writing","review drafting copilot","AI 360 feedback summarizer","performance management AI agent"],"industries":["cross-industry","technology"],"functions":["human-resources"],"patterns":["content-generation","summarization","agentic-workflow"],"channels":["internal-tools","email"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"A formal review asks a manager to remember and evaluate six or twelve months of a person's work,\nusually while running the same exercise for every other person they manage at once. Windmill\ndescribes this playing out at Rho as the company grew: managers spent significant time gathering\nwork artifacts and feedback across separate tools, reviews stretched longer than intended, and\nleaders lacked a unified view of performance across teams.\n\nWindmill describes this at Case Status, where what Case Status called the \"blank page problem\",\nmanagers and employees staring at an empty review form and trying to reconstruct months of work\nfrom memory, played out against a prior process, a patchwork of Google Docs and forms, that was\nboth time consuming and made it hard to capture the full scope of an employee's contributions.\nWindmill also argues that mid sized companies often see the largest gains from this kind of tool,\nbecause they typically lack the dedicated HR resources of large enterprises to run and chase a\nformal cycle by hand.","problemStats":[],"howItWorks":"1. **Collect the year's context continuously.** The assistant connects to the tools work already\n   happens in, such as project trackers, chat and documents, and to goals set earlier in the\n   cycle, so it has real material to draw from rather than starting from nothing at review time.\n2. **Gather structured feedback.** Peers, the manager and, where used, the employee's own self\n   review answer a short set of questions; the assistant chases outstanding responses so the\n   cycle does not stall on one missing input.\n3. **Draft the review.** The assistant synthesises the work history and feedback into a first\n   draft organised against the organization's review template and competencies, citing the\n   specific examples it drew from.\n4. **The manager edits and owns it.** The manager rewrites, adds judgement the data cannot\n   capture, and is accountable for the final rating and text; the draft is a starting point, not\n   the answer.\n5. **Coordinate the cycle.** The assistant tracks who still owes feedback, reminds them, and gives\n   HR or the manager's own manager visibility into where the cycle stands, without a dedicated\n   coordinator running it by hand.","valueDrivers":["employee-productivity","cost-to-serve","speed"],"kpis":["handling-time-reduction","processing-time-reduction"],"indicativeValue":{"referenceOrg":"An organization with 2,000 employees who receive a formal performance review each cycle","inputs":[{"key":"employees","label":"Employees reviewed per cycle","low":2000,"high":2000,"unit":"employees","note":"The reference organization."},{"key":"cyclesPerYear","label":"Formal review cycles per year","low":1,"high":2,"unit":"cycles per year","note":"Editorial assumption, replace with your own; many organizations run one annual and one mid year cycle."},{"key":"hoursSavedPerReview","label":"Manager and employee hours saved per review","low":1,"high":3,"unit":"hours per review","note":"Editorial assumption, replace with your own. Rho's CFO is quoted comparing the AI assisted draft to a review he estimated would otherwise have taken him 3 hours, which anchors the upper bound; the Case Status figure is treated here as a reduction in elapsed cycle time, not in hours worked, so it is not used to derive this range."},{"key":"hourlyCost","label":"Blended fully loaded hourly cost of a manager or employee","low":40,"high":80,"unit":"USD per hour","note":"Editorial assumption. Replace with your own blended rate."}],"formula":"employees * cyclesPerYear * hoursSavedPerReview * hourlyCost","currency":"USD","period":"per year","resultLabel":"Manager and employee review time cost avoided","caveat":"Gross avoided time cost only. It leaves out the platform cost, the value of more consistent and timely reviews, and any effect on retention, development or promotion decisions, which this page's evidence does not measure."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Drafting from feedback already collected in a form is straightforward. The value comes from connecting to where work actually happens (chat, project tools, documents) so the draft has real, specific material; without that, the assistant just reformats whatever a person typed into a box, which saves little time.","dataPrerequisites":["The organization's review template, competencies and rating scale","Goals set earlier in the cycle for each employee, where the organization uses them","Access to the work context tools the assistant should draw from, scoped per employee","A clear policy on what AI generated text may and may not be used for"],"integrations":["HR information system for employee, manager and cycle data","Chat and project tools (for example Slack, Jira, GitHub) the assistant summarises from","Document storage for prior reviews and goal documents","Identity provider for single sign on and access scoping"]},"implementation":{"steps":[{"title":"Decide what the draft may and may not touch","detail":"Write down, before launch, that the AI draft is a starting point only, the manager owns the final text and rating, and the system is not used directly in pay or promotion decisions."},{"title":"Connect real work context, not just a form","detail":"Wire the assistant to the tools work happens in for a pilot group before rolling out broadly, so the draft cites specific, verifiable examples rather than generic language."},{"title":"Keep feedback collection structured","detail":"Use a short, consistent set of questions for peer and self feedback so the assistant can synthesise reliably, and let it chase outstanding responses automatically."},{"title":"Require a human edit before anything is final","detail":"Do not allow a draft to be submitted unedited; require the manager to open and modify the text, and log that the review was edited before submission."},{"title":"Pilot with one team's cycle","detail":"Run a full cycle with one function first, measure cycle time, hours and employee preference, and fix template and integration gaps before the next team joins."},{"title":"Decide how sensitive topics are handled","detail":"Performance issues that touch conduct, health or personal circumstances should route to a person and a private conversation, not into an AI drafted written record."}],"guardrails":["The manager reviews and edits every draft; nothing is sent to the employee unedited","The system is not used to set pay, bonus or promotion decisions directly from its output","Feedback and drafts are visible only to the people the organization's existing review process would show them to","Draft text is clearly marked as AI assisted until the manager has reviewed and approved it","Conversation and draft logs have a defined retention period and restricted access"],"humanInTheLoop":"The manager owns every rating and every word of the final review. HR owns the template, the cycle policy and what counts as a sensitive topic that must go to a person instead of into a draft, and reviews a sample of cycles for consistency and fairness across teams.","kpisToInstrument":["Total manager and employee hours spent on the cycle, before and after","Cycle length from opening to closing the review period","Employee preference for the AI assisted process versus the prior one","Share of drafts materially rewritten by the manager, as a check that editing is real","On time completion rate across the organization"],"failureModes":[{"title":"Generic, uneditable prose","detail":"A draft that reads fluently but says nothing specific gets rubber stamped rather than improved. Require citations to specific work items in the draft and sample reviews for genuine editing."},{"title":"The draft becomes the decision","detail":"Under time pressure, managers submit the draft with only cosmetic changes, so the AI's synthesis quietly becomes the evaluation. Track the share of drafts materially edited and make manager training explicit about this risk."},{"title":"Feedback collected without consent context","detail":"Peers do not realise their comments will be summarised and shown to the subject, which damages trust in the feedback process. Be explicit up front about what is collected and how it is used."},{"title":"Sensitive matters written into a permanent record","detail":"A conduct or health related issue gets synthesised into formal review text instead of handled as a private conversation. Define these topics in advance and route them to a person, not the drafting flow."}]},"risk":{"euAiAct":{"tier":"high","basis":"Annex III point 4(b) lists AI systems intended to monitor and evaluate the performance and behaviour of workers as high risk. Synthesising an employee's work history and feedback into a performance evaluation is very plausibly profiling of a natural person under GDPR Article 4(4), which expressly covers analysing or predicting a person's \"performance at work\". Article 6(3)'s last subparagraph makes an Annex III system high risk regardless of the derogations whenever it performs such profiling, so a tool built this way is high risk by default however much the manager edits the output. The derogations in Article 6(3), including a narrow procedural task or improving the result of a previously completed human activity, do not fit drafting an evaluation from scratch; the closest is point (d), a preparatory task ahead of a human assessment, which only has a chance of applying to a design that avoids profiling altogether, for example one that only surfaces raw facts without synthesising a judgement. Where that derogation is argued, the documentation duty under Article 6(4) falls on the provider of the system, and only on the deploying organization when it builds the tool itself. Because the tool is high risk by default, Article 26(7) requires informing affected workers and their representatives before it is put into use in the workplace, whatever the tool's output is used for; using the same system's output directly in pay, promotion or termination decisions removes any doubt and triggers the full high risk regime. Annex III's high risk obligations apply from 2 December 2027."},"regulations":["eu-ai-act","gdpr","uk-gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) lists monitoring and evaluating the performance and behaviour of workers as a high risk employment use."},{"title":"Article 6, classification rules for high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/6/","note":"Sets out the narrow procedural task, prior human activity and preparatory task derogations in Article 6(3), the rule in its last subparagraph that profiling of natural persons always makes an Annex III system high risk regardless of those derogations, and the documentation duty in Article 6(4), which falls on the provider of the system."},{"title":"Article 26, obligations of deployers of high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/26/","note":"Article 26(7) requires informing affected workers and their representatives before a high risk AI system is put into use in the workplace."}],"controls":["Written policy that the AI draft is a starting point and the manager owns the final rating and text","Documented Article 6(3) and 6(4) assessment from the system's provider, covering whether the design profiles workers","Inventory entry for the assistant with an accountable HR owner","Sampled review of edited versus unedited drafts, to check editing is real","Defined list of sensitive topics that route to a person instead of the drafting flow"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that read work context from\nconnected tools (project trackers, chat, document storage) through the **integration catalog**,\ncombined with a **knowledge base** holding the review template and competency framework so\nevery draft is structured the same way. An **agentic workflow** collects structured peer and\nself feedback through **web chat**, **email** or **Microsoft Teams**, chases outstanding\nresponses, and produces the draft with **structured output** mapped to the organization's\nreview fields.\n\nEvery draft is written to the HR system only after **human in the loop approval** from the\nmanager; nothing reaches the employee without the manager's approval. **Guardrails** block conduct, health or personal\ncircumstance content from the draft, and a **human handover** rule or agent tool escalates it to\na person, while **PII masking** limits what personal data reaches the model. **Run history with\na full audit trail** records each drafting run and the manager's approve or reject decision, and\n**test suites** evaluate the drafting flow against a reference set of feedback inputs. The\nplatform is model agnostic and offers EU and UAE data residency."},"faq":[{"question":"Does AI decide an employee's rating?","answer":"The sources on this page do not describe the AI setting a rating, and neither Windmill customer story says who decides it. Design the system so the manager decides: require the manager to open, edit and approve every draft before it reaches the employee, and keep the rating a manager judgement that the tool never sets on its own."},{"question":"How much time does AI assisted drafting actually save?","answer":"Windmill reports an 83% reduction in total hours spent on reviews at Rho and an 84% faster full cycle at Case Status, both alongside a reported 93% of employees preferring the new process to the prior one. Both figures cover the whole cycle (reminders, feedback collection, synthesis) rather than drafting on its own, and neither evidence record on this page is a drafting only deployment. Windmill's companies list page separately describes JPMorgan Chase using AI for drafting only, and cites a Boston Consulting Group figure of a 40% reduction in writing time when staff use AI to draft reviews; that is Windmill's own secondhand claim about a different deployment, not evidence recorded on this page."},{"question":"Is this high risk under the EU AI Act?","answer":"By default, yes. Annex III point 4(b) lists monitoring and evaluating worker performance as high risk, and Article 6(3)'s last subparagraph makes an Annex III system high risk regardless of the derogations whenever it profiles natural persons; synthesising someone's work history and feedback into a performance evaluation is plausibly profiling under GDPR's definition. Only a design that avoids profiling and fits a narrow procedural or preparatory task can argue for the derogation, and even then the documentation duty under Article 6(4) falls on the provider of the system, not the deploying organization, unless the organization built it itself."},{"question":"What should stay out of an AI drafted review?","answer":"Conduct issues, health matters and anything tied to a protected characteristic should be handled in a private conversation and only entered into the formal record by a person, not synthesised automatically from feedback text."}],"related":["hr-and-policy-assistant","internal-talent-marketplace-matching","meeting-summarization-and-action-items","employee-onboarding-assistant"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched and written from Windmill's Rho and Case Status customer stories, verified against the sources with usecases:source."},{"date":"2026-09-28","note":"Editorial review: corrected deployment years and source dates to each page's real datePublished, removed the employee adoption KPI (the 93% figure measures process preference, not tool adoption), removed the unsourced JPMorgan Chase claim from the FAQ and step 1, rewrote the EU AI Act section to cover the Article 6(3) profiling override and assign the Article 6(4) documentation duty to the provider, and reworded the problem section and the hours saved assumption to only paraphrase what the cited sources say."},{"date":"2026-09-28","note":"Adversarial review fixes: rewrote blitsAi.howToBuild to match FEATURE_INVENTORY.md exactly (run history records the approve or reject decision, not an edit step; guardrails block sensitive content and a separate handover rule or tool escalates it; test suites run against a reference set, not \"after every change\"); attributed the drafting time saving argument to Windmill in the FAQ instead of stating it as fact, and removed the unattributed narrower deployment prediction; restored the \"often\" qualifier and corrected the \"blank page problem\" attribution to Case Status in the problem section; moved the Article 26(7) informing duty out of the pay or promotion conditional in risk.euAiAct.basis and added the Annex III effective date; reworded the hoursSavedPerReview note to anchor the upper bound on Rho's CFO quote instead of an invalid percent to hours conversion; corrected the Rho evidence verification note, which had missed the 83% figure in the page's stat tiles and JSON-LD and misattributed a Windmill page disclaimer to Rho's own site; and fixed the Case Status evidence year note, which contradicted itself on whether the deployment predates the page's datePublished."},{"date":"2026-09-28","note":"Second adversarial review fixes: reworded blitsAi.howToBuild so the human in the loop sentence claims only that a draft needs the manager's approval before it reaches the employee, not that the platform forces an edit; and reworded the FAQ on drafting time savings so it no longer claims the cited sources have no drafting only comparison, since Windmill's companies list page separately covers a JPMorgan Chase drafting only rollout and a Boston Consulting Group figure, now attributed to Windmill as a secondhand claim rather than recorded as evidence."}],"slug":"performance-review-drafting-agent","url":"https://www.blits.ai/ai-use-cases/performance-review-drafting-agent","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":84,"min":84,"max":84,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"case-status-ai-performance-review-cycle","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":83,"min":83,"max":83,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"rho-ai-performance-review-drafting","pooled":true}]}],"indicativeValueResult":{"low":80000,"high":960000},"evidence":["case-status-ai-performance-review-cycle","rho-ai-performance-review-drafting"]},{"title":"AI agent for early collections and hardship support","shortTitle":"Collections and hardship agent","seoTitle":"AI debt collection agent with hardship routing","metaDescription":"AI collections agents call early arrears, take payments within set rules and pass hardship to people. SameDay Auto Finance's vendor reports 43% higher collections.","definition":"A voice and messaging agent that contacts customers in early arrears and answers their inbound calls, takes payments and sets up payment arrangements within preapproved rules, and recognises signs of hardship or vulnerability so those customers go straight to a trained person.","aliases":["AI debt collection agent","voice AI for collections","arrears and hardship assistant","digital collections agent"],"industries":["cross-industry","banking","payments","telecommunications","energy-and-utilities","automotive","professional-services"],"functions":["collections-and-recovery","customer-service"],"patterns":["voice-agent","conversational-agent","agentic-workflow","classification-and-routing"],"channels":["voice","sms","whatsapp","email","web-chat"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"lending","problem":"Many customers who miss a payment are not refusing to pay: their payday moved, a payment failed or\ntheir circumstances changed. A short, timely conversation in the first days of arrears resolves\nmany of these cases, but collections teams often lack the capacity to reach every account in that\nwindow, and contact outside office hours is limited. Accounts then roll into later buckets where\nrecovery is harder and more expensive.\n\nAt the same time collections is a heavily regulated conversation. Contact\nfrequency, tone, disclosures and the treatment of customers in financial difficulty are all set out\nin rules, and getting it wrong with a vulnerable customer causes real harm. Automation that only\npushes for payment makes this worse; automation that listens and routes well can make it better.","problemStats":[],"howItWorks":"1. **Reach out at the right time.** The agent contacts accounts in early arrears on the channel and\n   at the time each customer is most likely to respond, within contact frequency rules.\n2. **Verify and disclose.** It confirms identity before discussing the debt and gives the required\n   disclosures, including that it is an AI agent.\n3. **Understand the reason.** It asks why the payment was missed and classifies the answer, such as\n   a failed payment, a changed pay date or a change in circumstances.\n4. **Resolve within rules.** It takes a payment, sends a secure payment link, moves a due date or\n   sets up a short arrangement, but only within limits the lender has approved.\n5. **Route hardship and vulnerability to people.** Mentions of job loss, illness, bereavement,\n   domestic abuse or distress, or any request for help, go to a trained specialist with a summary,\n   and collection activity pauses.\n6. **Record everything.** Every contact, disclosure, promise to pay and arrangement is logged with\n   its basis for audit and complaint handling.","valueDrivers":["cost-to-serve","risk-reduction","customer-experience","compliance"],"kpis":["recovery-rate-uplift","cost-reduction","containment-rate","interactions-handled","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A lender with 50,000 accounts entering early arrears each year","inputs":[{"key":"accounts","label":"Accounts entering early arrears per year","low":50000,"high":50000,"unit":"accounts per year","note":"The reference lender."},{"key":"contactsPerAccount","label":"Outbound and inbound contacts per account in early arrears","low":4,"high":8,"unit":"contacts per account","note":"Editorial assumption. Replace with your own contact data."},{"key":"automatedShare","label":"Share of those contacts the agent completes without a person","low":0.4,"high":0.7,"unit":"fraction of contacts","note":"Editorial assumption, deliberately below full automation because hardship cases must reach people."},{"key":"costPerContact","label":"Cost of a human handled collections contact","low":3,"high":6,"unit":"USD per contact","note":"Editorial assumption. For comparison, SameDay Auto Finance's vendor reports 75% lower collection call costs in early delinquency."}],"formula":"accounts * contactsPerAccount * automatedShare * costPerContact","currency":"USD","period":"per year","resultLabel":"Collections contact cost avoided","caveat":"Contact cost only. It leaves out the usually larger effect of fewer accounts rolling into later arrears and charge off, the cost of running the AI and the payment integration, and the effort of compliance review."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Conversation design and integrations (arrears data, payments, arrangement rules, dialler and consent) are manageable. The effort is in conduct rules per market, vulnerability detection and a clean handover to specialists.","dataPrerequisites":["Arrears data per account with contact history, consent and preferred channel","Written arrangement rules the agent may offer, with limits per product","Contact frequency and time of day rules per market","Vulnerability and hardship triggers agreed with the specialist team"],"integrations":["Collections or loan servicing system","Payment gateway for card and bank payments, or payment links","Telephony and messaging channels, including consent management","Case management or CRM for hardship referrals","Complaint handling"]},"implementation":{"steps":[{"title":"Start in the first days past due","detail":"Reminder and resolution conversations in the first bucket are high volume, low risk and the cheapest place to prevent roll forward."},{"title":"Write the arrangement rules down","detail":"List exactly which arrangements the agent may offer (date change, short plan, split payment), with limits, and what it says when a request is outside them."},{"title":"Design hardship routing with the specialists","detail":"Agree the phrases and situations that pause collection and route to a person, test them on real transcripts, and err on the side of routing."},{"title":"Build compliance in","detail":"Encode identity checks, disclosures, AI disclosure, contact limits and quiet hours in the flow, not in the prompt, and log each one."},{"title":"Pilot against a control group","detail":"Run the agent on part of the book, compare roll rates, promises kept, complaints and satisfaction with a human handled control, then widen."}],"guardrails":["Identity verification before any mention of the debt","Clear disclosure that the customer is speaking with an AI agent","Arrangements only within preapproved rules; anything else goes to a person","Immediate handover and a pause in collection on any hardship or vulnerability signal","Contact frequency, time of day and channel consent enforced by the system","No threats, pressure tactics or misleading statements, checked by output guardrails"],"humanInTheLoop":"Trained specialists handle every hardship, vulnerability, dispute and complaint case and every arrangement outside the rules. Quality teams review a sample of AI conversations each week against the conduct standard, and compliance approves every change to scripts or arrangement rules.","kpisToInstrument":["Roll rate from the first to the second arrears bucket versus control","Promise to pay kept rate","Share of conversations routed for hardship, and specialist agreement with the routing","Complaints and conduct breaches per thousand conversations","Cost per account resolved"],"failureModes":[{"title":"Missed vulnerability","detail":"The agent keeps pressing for payment when a customer mentions illness or job loss. Route on broad signals and review missed cases every week."},{"title":"Arrangements that fail","detail":"Easy plans accepted to end the call and then broken. Check affordability within the rules and track kept rates per arrangement type."},{"title":"Contact that becomes harassment","detail":"Automation makes it cheap to call too often. Enforce frequency limits in the system, per customer across channels."},{"title":"Payment data in transcripts","detail":"Card numbers read aloud end up in logs. Use payment links or secure capture and mask card data before it reaches the model."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing collections agent must disclose that it is AI (Article 50). It is not listed in Annex III as long as it applies preapproved arrangement rules and does not itself evaluate creditworthiness; an affordability model that decides who gets which arrangement for individuals should be assessed separately against Annex III point 5(b)."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","pci-dss","telecom-consumer-rules","us-tcpa"],"guidance":[{"title":"FG21/1: guidance for firms on the fair treatment of vulnerable customers","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publication/finalised-guidance/fg21-1.pdf","note":"Expectations for identifying and supporting vulnerable customers, including in automated channels."},{"title":"Debt Collection Practices (Regulation F)","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/final-rules/debt-collection-practices-regulation-f/","note":"US rules under the Fair Debt Collection Practices Act on communications with consumers, harassment, misleading statements and unfair practices. They govern debt collectors as the FDCPA defines that term, which generally covers third party collectors rather than creditors collecting their own debts."},{"title":"RG 96 Debt collection guideline: for collectors and creditors","issuer":"Australian Securities and Investments Commission and Australian Competition and Consumer Commission","region":"asia-pacific","url":"https://asic.gov.au/regulatory-resources/find-a-document/regulatory-guides/rg-96-debt-collection-guideline-for-collectors-and-creditors/","note":"Joint ACCC and ASIC guideline on how Australian consumer protection law applies to debt collection, for creditors collecting their own debts and for external collection agencies."},{"title":"Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Declaratory ruling on AI generated voices under the TCPA (FCC 24-17)","issuer":"Federal Communications Commission","region":"north-america","url":"https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf","note":"AI generated voices count as \"artificial or prerecorded voice\" under the TCPA, so US outbound AI calls need the consent the TCPA requires."}],"controls":["Approved scripts and arrangement rules under change control","Logged identity check, disclosures and consent for every contact","Weekly quality sampling with a specific check for missed vulnerability","Complaint monitoring linked to AI conversations","Card data masked or captured outside the conversation"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** on **voice** (telephony with real time streaming speech\nrecognition and synthesis, and call transfer), **SMS**, **WhatsApp** and **email**, with the regulated steps\n(identity check, disclosures, arrangement offer) built as deterministic **flows** and **DTMF**\ninput where customers prefer the keypad. **Custom functions** read arrears data and write\narrangements to the servicing system, and payments go through **payment links** from the payments\nservice, so card numbers never enter the conversation; on voice, **DTMF** capture with the sensitive\ndata flag keeps digits out of the transcript, and in text channels **credit card number detection\nand tokenization** and **PII masking** at the gateway tokenize or redact card numbers a customer types.\n\n**Sentiment analysis** and **guardrails** help detect distress, and **human handover** sends\nhardship and vulnerability cases to a specialist with a conversation summary, including **live\ntakeover** of a voice call. **Agentic tasks** schedule follow ups such as a check after a promised\npayment date. **Test suites** replay hardship scenarios on every change, and **analytics** show\ninteractions, satisfaction and sentiment. The platform is model agnostic and supports\nseveral languages, including Arabic."},"faq":[{"question":"What results do lenders report from AI collections agents?","answer":"The figures on this page come from vendor case studies about US auto lenders and a collection agency. Skit.ai reports 43% higher collections and 75% lower call costs in early delinquency at SameDay Auto Finance, and 63% lower collection costs at Day Knight & Associates, whose own executive says collections doubled. Test against a control group before relying on such numbers."},{"question":"Should an AI agent handle customers in financial hardship?","answer":"It should recognise them and pass them to a trained person quickly, not negotiate hardship on its own. Hardship and vulnerability need judgment, flexibility and often referral to support that an agent should not decide."},{"question":"Is an AI collections agent allowed to call customers?","answer":"Generally yes, within the same rules as human collectors: identity checks, disclosures, contact frequency limits, quiet hours and channel consent, plus disclosure that it is AI. Some markets add rules on automated calls: in the US, the FCC treats AI generated voices as artificial voices under the Telephone Consumer Protection Act, which sets consent rules for such calls. Check local telemarketing and collection law."}],"related":["loan-restructuring-recommendations","outbound-notice-drafting","financial-wellbeing-coach","credit-early-warning-monitoring","proactive-outbound-engagement-agent","utility-billing-and-move-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: industries now include automotive and professional services, where its evidence comes from; added Telephone Consumer Protection Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added SEO title and description; Day Knight's doubled collections now attributed to the agency's own executive; added the captive lender's 88% containment figure; replaced ASIC responsible lending with the ACCC and ASIC debt collection guideline; corrected the Regulation F note and the FAQ; removed unsupported claims from the problem text; limited the Blits.ai build to capabilities in the feature inventory."},{"date":"2026-09-27","note":"Added the FCC declaratory ruling FCC 24-17 on AI generated voices under the TCPA to the guidance, pointed Article 50 to the EUR-Lex text, aligned the Regulation F note with the CFPB wording and removed unsupported general claims from the problem text."}],"slug":"collections-and-hardship-agent","url":"https://www.blits.ai/ai-use-cases/collections-and-hardship-agent","benchmarks":[{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":69,"min":63,"max":75,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"sameday-auto-finance-voice-collections","pooled":true},{"id":"day-knight-associates-multichannel-collections","pooled":true}]},{"kpi":"recovery-rate-uplift","label":"Recovery uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":71.5,"min":43,"max":100,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"day-knight-associates-multichannel-collections","pooled":true},{"id":"sameday-auto-finance-voice-collections","pooled":true}]}],"indicativeValueResult":{"low":240000,"high":1680000},"evidence":["day-knight-associates-multichannel-collections","sameday-auto-finance-voice-collections"]},{"title":"AI agent for first line contact centre service","shortTitle":"First line contact centre","seoTitle":"AI contact centre agent for first line service","metaDescription":"An AI agent answers routine calls and chats and hands the rest to people. See results from Commonwealth Bank, Klarna and Vodafone, plus risks and a playbook.","definition":"An AI agent that answers the first line of inbound customer contact on phone, chat and messaging, resolves general and routine questions end to end in the customer's own language, and routes everything complex, sensitive or regulated to the right human team with the context attached.","aliases":["AI contact centre agent","virtual agent","conversational IVR","customer service chatbot","AI front door"],"industries":["cross-industry","banking","payments","telecommunications","travel-and-hospitality","retail-and-ecommerce","wealth-and-asset-management"],"functions":["customer-service"],"patterns":["conversational-agent","voice-agent","rag-knowledge-assistant","classification-and-routing"],"channels":["voice","web-chat","mobile-app","whatsapp","social-messaging"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"mainstream","segment":"front-office","problem":"Much inbound contact is general: opening hours and fees, how do I, where is my request, what\ndoes this letter mean. In a bank it is also time critical: a customer abroad without a local SIM,\na card that does not work, a payment that has not arrived. Human agents answer the same\nquestions all day while callers wait in a queue, and the conversations that need care (a\nbereavement, a scam victim, a complaint) wait behind them.\n\nTouch tone menus and first generation chatbots did not fix this. Menus route by the option a\ncaller picks, and those chatbots matched keywords to a fixed list of answers, so anything outside\nthe list ends in a dead end or a transfer. The shift is an agent that understands free speech and text, answers from the\norganization's approved knowledge, can look up the customer's own case, and knows exactly when\na human must take over.","problemStats":[],"howItWorks":"1. **Greet and understand.** The agent identifies the intent and language from the customer's\n   first words, on the phone or in chat, and discloses that it is AI.\n2. **Answer general questions from approved knowledge.** Fees, procedures, product terms and\n   service status come from retrieval over the organization's own content, with a refusal when\n   the content does not cover the question.\n3. **Look up what is personal.** After authentication it checks the status of the customer's\n   request, order or case through read only APIs and explains it.\n4. **Route deliberately.** Regulated journeys (disputes, fraud reports, complaints) follow a\n   defined path; vulnerability, strong emotion or repeated failure trigger a human.\n5. **Hand over with context.** The human receives the transcript, a summary, the authenticated\n   identity and what was already tried, so the customer does not repeat anything.","valueDrivers":["cost-to-serve","customer-experience","inclusion-and-access","employee-productivity"],"kpis":["containment-rate","interactions-handled","response-time-reduction","customer-satisfaction","first-contact-resolution","handling-time-reduction"],"indicativeValue":{"referenceOrg":"A retail bank contact centre that receives 2 million contacts a year across phone and chat","inputs":[{"key":"contacts","label":"Inbound contacts per year","low":2000000,"high":2000000,"unit":"contacts per year","note":"The reference contact centre."},{"key":"generalShare","label":"Share of contacts that are general or routine","low":0.5,"high":0.7,"unit":"fraction of contacts","note":"Editorial assumption, replace with your own contact reason report."},{"key":"containment","label":"Share of those contacts the agent resolves without a human","low":0.3,"high":0.45,"unit":"fraction of general contacts","note":"The low end is conservative; the high end sits above most benchmarks on this page (44 to 47% for Airbnb, Ingka, JetBlue and Vodafone Germany), while a few operators report higher containment (66% at Together Credit Union, 70% for Vodafone TOBi, 84.6% for Commonwealth Bank's self service messaging in May 2026), because early months run lower."},{"key":"costPerContact","label":"Cost of a human handled contact","low":3,"high":6,"unit":"USD per contact","note":"Editorial assumption for a blended chat and phone contact. Replace with your own fully loaded cost."}],"formula":"contacts * generalShare * containment * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of the AI and the integrations, the value of shorter waiting times, the cost of customers who give up instead of being helped, and any capacity the organization chooses to reinvest in human service rather than save."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering from knowledge is straightforward. The effort is in clean, owned knowledge content, authentication on the phone, integration with the contact centre platform for a warm handover, and deciding which intents must never be contained.","dataPrerequisites":["A contact reason report with volumes per intent and channel","Approved, current knowledge articles with an owner and review date each","A written list of intents and signals that always go to a human","Historical transcripts to build test sets"],"integrations":["Telephony or contact centre platform (routing, queues, transfer with context)","Messaging channels (web chat, app, WhatsApp, social)","Authentication (app confirmation, one time passcode, voice biometrics)","CRM or case management, read only at first","Knowledge management system"]},"implementation":{"steps":[{"title":"Choose intents by volume and by risk","detail":"From the contact reason report pick the top general intents and write down, just as explicitly, the intents the agent must route and never contain: complaints, disputes, fraud, hardship and bereavement."},{"title":"Clean the knowledge before you connect it","detail":"Retire duplicate and outdated articles, give every article an owner and a review date, and make the agent refuse when retrieval finds nothing. Many wrong answers are content problems."},{"title":"Design the handover first","detail":"Agree with the contact centre what a human receives (summary, identity, attempted steps) and how the customer keeps their place in the queue. Always offer a clear route to a person."},{"title":"Test on real conversations","detail":"Build test sets from historical transcripts per intent, including angry customers, vulnerable customers and attempts to push the agent off policy, and run them on every change."},{"title":"Launch in chat, then voice","detail":"Chat is easier to monitor and correct. Add voice once containment and satisfaction per intent are stable, and measure repeat contacts, not just containment."}],"guardrails":["Answers only from approved knowledge, with a refusal and a handover when it is not covered","Mandatory routing of complaints, disputes, fraud, hardship and vulnerability signals to people","A visible route to a human at any time, on every channel","AI disclosure at the start of every conversation","Masking of personal and card data before text reaches a model or the logs"],"humanInTheLoop":"Human agents take every conversation the agent routes and can take over live. A quality team reviews a weekly sample of contained conversations for fluent but wrong answers and signs off every new intent before it goes live.","kpisToInstrument":["Containment per intent, counting a repeat contact within seven days as not contained","Handover rate and handover reasons per intent","Customer satisfaction on contained conversations versus human handled ones","Time to first meaningful response and total time to resolution","Complaints that mention the assistant"],"failureModes":[{"title":"Containment that is really abandonment","detail":"Customers give up rather than get helped, which looks like success on the dashboard. Count repeat contacts and measure satisfaction per intent."},{"title":"Replacing people instead of routing to them","detail":"Klarna reported that its assistant handled two thirds of chats, then said in 2025 that it had gone too far and began hiring human agents again so customers can always reach a person. Keep human capacity for the moments that matter."},{"title":"Confident wrong answers","detail":"An agent that invents a policy creates liability, as the Air Canada tribunal case showed. Ground answers in approved content and refuse when unsure."},{"title":"Missed vulnerability","detail":"A customer in distress is kept in automation. Detect vulnerability signals and hand over early."}]},"risk":{"euAiAct":{"tier":"limited","basis":"An AI system that interacts directly with people must be designed so that they know they are dealing with AI, unless that is obvious from the context (Article 50(1)). It is not high risk under Annex III as long as it does not evaluate eligibility for essential public assistance benefits and services (point 5(a)), creditworthiness (point 5(b)), risk and pricing for life and health insurance (point 5(c)) or emergency calls (point 5(d)). This holds only if emotion or vulnerability signals are inferred from what the customer says (text or transcript content), not from voice or other biometric features; an agent that infers emotion from a caller's voice is an emotion recognition system (Article 3(39)), which is high risk under Annex III point 1(c) and triggers the deployer disclosure duty in Article 50(3)."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","dora","apra-cps-230","eu-accessibility-act"],"guidance":[{"title":"Regulation (EU) 2024/1689 (AI Act), Article 50: transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Chatbots in consumer finance","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/","note":"Warns that deficient chatbots that prevent access to live, human support can lead to law violations and customer harm."},{"title":"FG22/5: Final non-Handbook Guidance for firms on the Consumer Duty","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publication/finalised-guidance/fg22-5.pdf","note":"Says firms will likely need a real time human interface, such as a phone service, for security, fraud and other complex or sensitive journeys, and gives an automated phone system without a route to other support as an example of poor practice."}],"controls":["Inventory entry with an accountable owner and a documented list of contained and routed intents","Content governance with owners and review dates for every knowledge article","Transcript logging and retention in line with record keeping rules","Regression tests on every change to prompts, content or model","Monitoring of outcomes for vulnerable customers and of complaint trends"],"incidents":[{"title":"Incident 639: Air Canada Chatbot Reportedly Provides Inaccurate Bereavement Fare Information, Leading to Customer Overpayment","url":"https://incidentdatabase.ai/cite/639/","note":"A Canadian small claims tribunal held the airline responsible for what its website chatbot told a customer about bereavement fares, a reminder that the organization owns every answer its agent gives."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** grounded in a **knowledge base** with hybrid retrieval\nover the organization's approved content, with **flows** for journeys that must follow fixed\nsteps and **custom functions** for read only lookups such as request or case status. Language\ndetection and multi language content let one agent serve customers in their own language.\n\nThe same agent runs on **voice** (streaming speech recognition and synthesis, DTMF, call\ntransfer), **web chat, WhatsApp, SMS and Facebook Messenger**, and inside the organization's\nown app through the API channel. **Human handover** passes the conversation transcript to a\nlive agent platform (Salesforce, Freshdesk or Zoho SalesIQ) or\ntransfers the call on voice, and supervisors can monitor and take over live, including on\ncalls. **Guardrails** check input and output, **PII masking** runs at the gateway, **test\nsuites** hold multi turn conversation sets built from real transcripts that you run before\neach change goes live, **monitors** check key journeys on a schedule,\nand **analytics** show interactions, satisfaction, sentiment and top intents, with\nconversation logs to review every handover. The platform is model agnostic, with EU and UAE\ndata residency."},"faq":[{"question":"What share of contacts can an AI agent resolve on its own?","answer":"It depends on the intent mix and the channel. Microsoft reports that about 84.6% of Commonwealth Bank's self service messaging interactions were resolved end to end in May 2026, and that Vodafone's TOBi fully resolves 70% of inquiries arriving through digital channels; Posh reports that Together Credit Union's voice agent contains 66% of inbound calls. Airbnb itself reports a lower figure: nearly 45% of issues that begin with its AI assistant were resolved without a human agent in Q2 2026. Treat these as upper references and count repeat contacts before you celebrate."},{"question":"Should we remove the option to speak to a person?","answer":"No. Klarna said in 2025 that customers must always have the option to speak with a human and began recruiting customer service agents again. The US Consumer Financial Protection Bureau warns that chatbots that block access to human support can break the law, and the UK Financial Conduct Authority expects a real time human route for complex or sensitive journeys."},{"question":"How is this different from account and card servicing?","answer":"The first line agent answers and routes general inbound contact across the whole contact centre. Account and card servicing is the deeper, authenticated layer that performs transactions such as blocking a card. You can launch the first line first and add servicing actions behind it once authentication and handover work."}],"related":["account-and-card-servicing-agent","live-agent-assist","complaints-handling-agent","correspondence-triage-and-routing","citizen-information-assistant","insurance-policy-servicing-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog and verified against the sources. Made industry neutral because the job works the same way in every sector."},{"date":"2026-09-25","note":"Added travel and hospitality to industries, with airline and holiday rental evidence (Air India, JetBlue, Lufthansa Group, Pegasus Airlines, Airbnb)."},{"date":"2026-09-25","note":"Added retail and ecommerce to industries, with Ingka Group (IKEA) Billie chatbot evidence."},{"date":"2026-09-25","note":"Added telecommunications evidence (Vodafone TOBi and SuperTOBi, Vodafone Germany, BT Group EE Aimee, Mobily, Telkomsel) and a telecom figure to the first FAQ answer."},{"date":"2026-09-25","note":"Consolidation pass: industries now include wealth and asset management, where its evidence comes from; added European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: attributed the Commonwealth Bank, Together Credit Union and Vodafone figures to the vendors that report them and added Airbnb's own figure; spelled out the Annex III points in the EU AI Act basis; added FCA FG22/5 guidance to support the human route answer; corrected the Air Canada incident title; limited the Blits.ai section to capabilities in the feature inventory; removed unsupported generalizations; added an SEO title and meta description. In the evidence: removed a Vodafone call time metric from a pilot of a different tool, removed OpenAI as NatWest Cora vendor, corrected the Klarna launch wording, and added source dates."},{"date":"2026-09-27","note":"Review fixes: the Blits.ai section now names the supported handover integrations (Salesforce, Freshdesk, Zoho SalesIQ, call transfer on voice) and no longer implies a test gate; removed insurance from industries (no insurance evidence); Article 50 now links to EUR-Lex; full Air Canada incident title; softened an unsourced claim about wrong answers. In the evidence: removed a Blits.ai payments provider record that was a demonstration, not a deployment; corrected channels on two Blits.ai records; recorded Vodafone Germany's 44% as containment rate, as its source describes it; described Telkomsel's 5 million transactions as capacity."}],"slug":"first-line-contact-centre-agent","url":"https://www.blits.ai/ai-use-cases/first-line-contact-centre-agent","benchmarks":[{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":7,"nUpTo":0,"median":47,"min":44,"max":84.6,"byClaimant":{"organization":2,"vendor":5,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"commonwealth-bank-customer-service-orchestration","pooled":true},{"id":"vodafone-tobi-virtual-assistant","pooled":true},{"id":"together-credit-union-voice-agent","pooled":true},{"id":"ingka-group-billie-customer-service-chatbot","pooled":true},{"id":"airbnb-ai-customer-support-assistant","pooled":true},{"id":"jetblue-asapp-digital-customer-support","pooled":true},{"id":"vodafone-germany-tobi-messaging","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":6,"nUpTo":1,"median":9600000,"min":10000,"max":3000000000,"byClaimant":{"organization":4,"vendor":2,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-erica-virtual-assistant","pooled":true},{"id":"vodafone-tobi-virtual-assistant","pooled":true},{"id":"lufthansa-group-self-service-ai-agents","pooled":true},{"id":"ingka-group-billie-customer-service-chatbot","pooled":true},{"id":"klarna-ai-assistant-customer-service","pooled":true},{"id":"bt-group-ee-aimee-virtual-assistant","pooled":false},{"id":"air-india-aig-virtual-assistant","pooled":true}]},{"kpi":"response-time-reduction","label":"Response time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":90.75,"min":82,"max":99.5,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"mobily-agentic-ai-self-service","pooled":true},{"id":"klarna-ai-assistant-customer-service","pooled":true}]},{"kpi":"first-contact-resolution","label":"First contact resolution","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":60,"min":60,"max":60,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"vodafone-supertobi-generative-ai-assistant","pooled":true}]}],"indicativeValueResult":{"low":900000,"high":3780000},"evidence":["air-india-aig-virtual-assistant","airbnb-ai-customer-support-assistant","bank-of-america-erica-virtual-assistant","bpi-bea-chat-digital-assistant","bt-group-ee-aimee-virtual-assistant","commonwealth-bank-customer-service-orchestration","ingka-group-billie-customer-service-chatbot","jetblue-asapp-digital-customer-support","klarna-ai-assistant-customer-service","lufthansa-group-self-service-ai-agents","mobily-agentic-ai-self-service","natwest-cora-ai-assistant","pegasus-airlines-flybot-virtual-assistant","telkomsel-veronika-virtual-assistant","together-credit-union-voice-agent","vodafone-germany-tobi-messaging","vodafone-supertobi-generative-ai-assistant","vodafone-tobi-virtual-assistant"]},{"title":"AI agent for first notice of loss claims intake","shortTitle":"First notice of loss agent","seoTitle":"AI agents for first notice of loss (FNOL) intake","metaDescription":"An AI FNOL agent takes the claim report by phone, chat or app and opens the claim. Lemonade says its bot takes 96% of FNOL unaided; Travelers added a voice agent.","definition":"An AI agent that takes the first notice of loss from a policyholder by phone, chat or app, identifies the policy, collects the facts of the incident and the evidence the claim type needs, opens the claim in the claims system and tells the customer what happens next, handing complex, injured or vulnerable claimants to a human handler.","aliases":["FNOL agent","AI claims intake","claims reporting bot","voice agent for claims reporting","digital first notice of loss"],"industries":["insurance"],"functions":["claims","customer-service"],"patterns":["conversational-agent","voice-agent","agentic-workflow","document-processing"],"channels":["voice","web-chat","mobile-app","whatsapp"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"claims","problem":"The first notice of loss is the moment the insurer's promise is tested. A customer who has just\nhad a car accident, a burst pipe or a stolen bag calls or logs in to report it, often upset and\noften at night or at the weekend. Today much of that intake is a scripted conversation with a\nhuman handler who types answers into the claims system: policy number, date and place, what\nhappened, who was involved, photos and documents to send later.\n\nThe work is repetitive but it drives everything downstream. Missing or inconsistent facts at\nintake cause call backs, wrong triage, slower settlement and missed fraud or recovery signals.\nAfter a storm, many customers report losses at the same time, and queues grow exactly when\ncustomers need the insurer most. Intake by web form alone does not solve it: forms can be long,\nand many customers still prefer to call (Travelers launched its voice agent for exactly those\ncustomers).","problemStats":[],"howItWorks":"1. **Identify the customer and the policy.** The agent recognises the caller or logged in user\n   and matches the policy, for example by reading back a spoken policy number or registration.\n2. **Take the story in the customer's words.** It asks what happened and extracts the structured\n   facts the claim type needs (date, place, cause, parties, damage, injuries), asking follow up\n   questions only for what is missing.\n3. **Collect evidence while the customer is there.** It asks for photos, receipts or reports\n   through a link or upload and checks that they are readable and relevant.\n4. **Open the claim and set expectations.** It creates the claim in the claims system through an\n   API, returns the claim number and explains the next steps and timelines for this claim type.\n5. **Hand over when it should.** Injuries, vulnerability signals, disputes about cover, complex\n   commercial losses and anything the customer asks a person for go to a handler with the\n   transcript and extracted facts attached.","valueDrivers":["customer-experience","cost-to-serve","speed","inclusion-and-access"],"kpis":["containment-rate","automation-rate","interactions-handled","handling-time-reduction","customer-satisfaction","accuracy"],"indicativeValue":{"referenceOrg":"A motor and home insurer that receives 300,000 claims a year","inputs":[{"key":"claims","label":"Claims reported per year","low":300000,"high":300000,"unit":"claims per year","note":"The reference insurer."},{"key":"assistedShare","label":"Share of claims reported through a human handler today","low":0.4,"high":0.6,"unit":"fraction of claims","note":"Editorial assumption; replace with your own channel mix."},{"key":"containment","label":"Share of those intakes the agent completes without a handler","low":0.3,"high":0.6,"unit":"fraction of assisted intakes","note":"Conservative against the evidence on this page (Lemonade reports that its claims bot takes the first notice of loss without human intervention 96% of the time), because that figure comes from an insurer whose customers already file claims by chatting with its bot in the app."},{"key":"minutesPerIntake","label":"Handler minutes per first notice of loss, including wrap up","low":15,"high":25,"unit":"minutes per claim","note":"Editorial assumption, replace with your own handling time."},{"key":"costPerHour","label":"Fully loaded cost per handler hour","low":35,"high":55,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"claims * assistedShare * containment * minutesPerIntake / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Handler time avoided at first notice of loss","caveat":"Intake effort only. It leaves out the effect of better captured facts on triage, leakage and fraud detection, surge capacity after catastrophes, the cost of running the agent and the integration work with the claims and policy systems."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The conversation is the easy part. The work is in policy lookup and identity checks, writing a complete claim into the claims system through APIs, handling many claim types with different questions, and a clean handover for injured or vulnerable claimants.","dataPrerequisites":["The question set and required evidence per claim type and line of business","Policy and customer data reachable through APIs","Recorded or transcribed historical intake calls to design and test against","Approved wording for next steps, timelines and cover explanations"],"integrations":["Policy administration system for policy and cover lookup","Claims management system to create and update the claim","Telephony or contact centre platform for voice intake and handover","Document and photo upload with storage linked to the claim","Identity verification and customer authentication"]},"implementation":{"steps":[{"title":"Start with one line and simple claim types","detail":"Pick high volume, low complexity intakes such as glass, minor motor damage or personal possessions, where the facts are predictable and injuries are rare."},{"title":"Write the intake schema per claim type","detail":"List the mandatory facts, the optional ones and the evidence per claim type, and let the agent fill that schema instead of following a fixed script."},{"title":"Connect the systems before going live","detail":"The agent must create a real claim with a real claim number. An agent that only emails a transcript to a queue moves work around without removing it."},{"title":"Design the handover rules","detail":"Decide which signals in what the caller says send them to a person (injury, vulnerability, anger, disputes about cover, commercial losses) and pass the extracted facts so nobody asks twice."},{"title":"Test with real recordings and surge scenarios","detail":"Replay historical calls, noisy lines, accents, partial policy numbers and catastrophe volumes, and track field level accuracy against what a handler recorded."},{"title":"Measure claim quality, not only containment","detail":"Follow each agent intake downstream: call backs for missing facts, triage corrections and customer complaints tell you more than the share of calls without a handler."}],"guardrails":["The agent never tells a customer that a loss is covered or declined; cover decisions stay with the claims process","Automatic handover on injury, vulnerability, distress, disputes about cover and any request for a person","Every extracted field is confirmed back to the customer before the claim is created","Identity and policy match checks before any claim is opened or personal data is disclosed","Personal and health data masked in logs, minimised in model prompts, with retention aligned to claims files"],"humanInTheLoop":"Handlers own every claim the agent opens and review a daily sample of agent intakes against the recording. They take over complex, injured and vulnerable claimants, and claims leadership approves each new claim type before the agent handles it.","kpisToInstrument":["Share of intakes completed without a handler, per claim type","Share of agent opened claims that need a call back for missing or wrong facts","Field level accuracy of extracted facts on a weekly sample","Time from first contact to claim number, and customer satisfaction after intake","Handover rate and handover reasons"],"failureModes":[{"title":"Fast intake, poor claim file","detail":"The agent completes the call but misses facts handlers need, so the work moves downstream. Measure call backs and triage corrections per claim type."},{"title":"Implied promises about cover","detail":"A friendly answer such as \"that will be covered\" creates an expectation the insurer may not meet. Block cover statements in the prompt and in output guardrails."},{"title":"Vulnerable claimants kept in automation","detail":"Bereaved, injured or distressed customers are pushed through a script. Detect the signals in what the customer says and hand over early."},{"title":"Collapse under catastrophe volume","detail":"The agent is sized for normal days and fails during a storm. Load test for catastrophe peaks and keep a simple fallback that still captures the essentials."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing intake agent must be designed so that people know they are interacting with AI (Article 50(1)). Claims intake and claims handling are not listed in Annex III: point 5(c) covers risk assessment and pricing in life and health insurance, not claims. One design choice changes this: detecting distress by inferring emotions from the caller's voice is emotion recognition based on biometric data, which is high risk under Annex III point 1(c) and needs disclosure under Article 50(3). Detecting vulnerability from what the caller says does not. The limited tier assumes that design: every handover signal on this page (injury, distress, anger, vulnerability) is detected from the words of the conversation, and inferring emotions from the voice itself is out of scope."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Published in August 2025 and addressed to national supervisors, it clarifies how the principles and requirements of existing insurance legislation apply to AI systems, following a risk based and proportionate approach."},{"title":"Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."}],"controls":["AI disclosure at the start of every intake conversation","Inventory entry with an accountable claims owner and the list of claim types in scope","Transcript and extracted fields stored with the claim for audit and dispute handling","Regression tests on recorded calls for every change in prompts, models or claim types","Monitoring of outcomes for vulnerable customers and complaints that mention the agent"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** on the **voice** channel (real time streaming speech\nrecognition and synthesis over telephony, with DTMF input) and on **web chat and WhatsApp**, with\nthe insurer's own app connected through the **API channel**. A **flow** holds the intake schema\nper claim type, using **ask question**, **slot filling** and **receive attachment** blocks for the\nfacts, photos and documents. **Custom functions** call the policy system to match the policy and\nthe claims system to open the claim and return the claim number.\n\n**Guardrails** block statements about cover, **PII masking** with configurable patterns keeps\nidentifiers such as policy numbers, contact and payment details out of model prompts and logs,\nand **human handover** escalates to a claims handler through automatic rules,\nwith live takeover also on voice calls. **Test suites** run sets of multi turn intake\nconversations for every claim type on each change, **monitors** run\nscheduled checks against the agent, and **analytics** show interactions and satisfaction per\nbot. The platform is model agnostic and runs\nin EU or UAE regions for data residency."},"faq":[{"question":"How much of first notice of loss can an AI agent take without a human?","answer":"For simple, digital first claim types it can be most of it: Lemonade's annual report states that, as of December 31, 2025, its claims bot took the first notice of loss without human intervention 96% of the time. Incumbent insurers with phone heavy, complex lines should expect lower shares and start with simple claim types such as glass or minor motor damage."},{"question":"Does a voice agent make sense when most insurers push digital claims?","answer":"Yes, because many customers still call. Travelers says it launched a generative AI voice agent for first notice of loss by phone to serve customers who prefer to call, starting with auto damage claims, next to its straight through digital journey."},{"question":"Should the intake agent tell customers whether they are covered?","answer":"No. The agent records the loss and explains the process; cover decisions belong to the claims process with its rules and handlers. Statements that imply cover create expectations and complaints, and should be blocked by guardrails."}],"related":["claims-triage-and-straight-through-processing","photo-based-damage-assessment","claims-fraud-detection","travel-insurance-claims-and-assistance-agent","insurance-policy-servicing-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the insurance vertical with evidence from Lemonade, Travelers, Progressive, Hippo and DOMCURA, verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: rewrote the EU AI Act basis (Annex III points 5(c) and 1(c), Article 50), added UK GDPR, removed unsourced storm volume and form abandonment claims, aligned the Blits.ai build with the feature inventory, corrected the DOMCURA deployment year to 2022 and Hippo wording, added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: DOMCURA deployment year corrected to 2023 (the earliest year the archived source supports), Article 50 guidance now links to EUR-Lex, the EU AI Act basis states that voice based emotion inference is out of scope, Blits.ai test suite and analytics wording aligned with the feature inventory, Hippo and Travelers notes clarified."}],"slug":"claims-first-notice-of-loss-agent","url":"https://www.blits.ai/ai-use-cases/claims-first-notice-of-loss-agent","benchmarks":[{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"domcura-claimens-voice-claims-reporting","pooled":true}]},{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":55,"min":55,"max":55,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"lemonade-ai-jim-claims-automation","pooled":true}]},{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":96,"min":96,"max":96,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"lemonade-ai-jim-claims-automation","pooled":true}]}],"indicativeValueResult":{"low":315000,"high":2475000},"evidence":["domcura-claimens-voice-claims-reporting","hippo-clara-ai-claims-assistant","lemonade-ai-jim-claims-automation","progressive-digital-claims-and-generative-ai-assistant","travelers-generative-ai-fnol-voice-agent"]},{"title":"AI agent for flight disruption and rebooking","shortTitle":"Flight disruption and rebooking","seoTitle":"AI agents for flight disruption and rebooking","metaDescription":"An AI agent that explains flight disruptions and lets passengers rebook or get a refund in one chat. Delta and Lufthansa Group already run rebooking agents.","definition":"An AI agent that tells passengers proactively when their flight is delayed, cancelled or misconnected, explains why, and lets them rebook, request a refund or voucher, or claim care such as meals and hotels in one conversation on app, messaging, web or phone, within the airline's reaccommodation rules and passenger rights, handing complex itineraries and upset customers to a human with the context attached.","aliases":["airline rebooking chatbot","irregular operations assistant","IROPS self service agent","flight cancellation assistant","airline disruption assistant"],"industries":["travel-and-hospitality"],"functions":["customer-service","operations"],"patterns":["conversational-agent","voice-agent","agentic-workflow","content-generation"],"channels":["mobile-app","web-chat","sms","whatsapp","email","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Disruption is when an airline's service is tested hardest. A storm, a strike or an air traffic\ncontrol restriction can cancel many flights at once, and large numbers of passengers then call,\nqueue at transfer desks and post on social media at the same time, most asking the same three questions:\nwhat happened, what are my options, and what am I entitled to. Contact centres are sized for a\nnormal day, so waiting times explode exactly when anxiety is highest, and the passengers who\nmost need a person (families, passengers with reduced mobility, long haul connections) wait\nbehind everyone else.\n\nMost of the work is rule bound. The airline already knows which passengers are affected, which\nalternative flights have seats, what the fare rules allow and, in many markets, what care and\ncompensation the law requires. First generation chatbots could only point to a web page. The\nstep change is an agent that is connected to the reservation and departure control systems, can\npresent and confirm real options, issue a refund or voucher, and explain the reason in plain\nlanguage, while staying inside the airline's reaccommodation policy and passenger rights rules.","problemStats":[],"howItWorks":"1. **Detect and notify.** When operations change a flight, the agent (or staff drafting with\n   AI, as at United) sends a message by app, SMS or email that explains what changed and why,\n   before the passenger has to ask.\n2. **Authenticate and load the trip.** The passenger opens the conversation from the message or\n   the app; the agent identifies the booking, the connections and the loyalty status, so nobody\n   has to search for a reservation.\n3. **Offer real options.** The agent pulls the airline's reaccommodation offer and alternatives\n   with confirmed seats, standby options, a refund or an eCredit, and explains the fare rules\n   and the care the passenger is entitled to (meals, hotel, transport).\n4. **Act through approved tools.** The passenger picks an option; the agent confirms the new\n   flight, issues the refund request, voucher or eCredit and shows where the bags are, through a\n   small allow list of reservation, ticketing and baggage actions.\n5. **Hand over well.** Complex itineraries, group bookings, passengers needing assistance, codeshare\n   and interline cases, and anyone who asks for a person go to a human agent with a summary and\n   the options already shown, so the passenger does not start again.","valueDrivers":["customer-experience","cost-to-serve","speed","employee-productivity","inclusion-and-access"],"kpis":["containment-rate","automation-rate","interactions-handled","hours-saved","customer-satisfaction","customer-satisfaction-uplift","response-time-reduction"],"indicativeValue":{"referenceOrg":"An airline carrying 20 million passengers a year","inputs":[{"key":"passengers","label":"Passengers per year","low":20000000,"high":20000000,"unit":"passengers per year","note":"The reference airline."},{"key":"disruptedShare","label":"Share of passengers whose trip is cancelled, misconnected or significantly delayed","low":0.02,"high":0.05,"unit":"fraction of passengers","note":"Editorial assumption; replace with your own irregular operations data."},{"key":"contactsPerDisrupted","label":"Assisted contacts per disrupted passenger","low":0.5,"high":1,"unit":"contacts per disrupted passenger","note":"Editorial assumption; many passengers accept the automatic reaccommodation, others contact more than once."},{"key":"containment","label":"Share of disruption contacts the agent resolves","low":0.2,"high":0.45,"unit":"fraction of disruption contacts","note":"This range sits at or below the 45% containment rate ASAPP reports for JetBlue's virtual agent across general digital support, because disruption contacts include complex itineraries that need a person. Air India's 97% figure is an automation rate for handled queries, a different metric, and is not used as a containment benchmark here."},{"key":"costPerContact","label":"Cost of a human handled contact","low":4,"high":8,"unit":"USD per contact","note":"Editorial assumption for a blended phone and messaging contact. Replace with your own fully loaded cost."}],"formula":"passengers * disruptedShare * contactsPerDisrupted * containment * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled disruption contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of running the AI and the integrations, the revenue kept by rebooking passengers instead of refunding them, the care and compensation costs (which the agent does not change) and the effect on loyalty and complaints."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Answering \"where is my flight\" is easy. Rebooking is not: the agent needs read and write access to the passenger service system, inventory, ticketing and EMD issuing, fare rules, the reaccommodation engine and the notification platform, plus the legal rules for care and compensation per market. It must also survive peak load on the worst day of the year.","dataPrerequisites":["Reaccommodation and waiver policies per disruption type, in a form the agent can apply","Passenger rights rules per market (care, refund, compensation) with an owner and review date","Real time flight status, delay reasons and the operational reaccommodation offer per passenger","A catalog of disruption contact reasons with volumes from past irregular operations days"],"integrations":["Passenger service system and departure control (bookings, seats, standby lists)","Ticketing, refunds, eCredits and vouchers (EMDs), and the payment service for fare differences","Reaccommodation or disruption management engine","Notification platform (app push, SMS, email, messaging)","Baggage tracking","Contact centre platform for handover with conversation context, and CRM for loyalty status"]},"implementation":{"steps":[{"title":"Start with information, then add actions","detail":"First ship proactive, accurate delay and cancellation messages with the reason and the next step. Measure how many calls they prevent. Add actions (accept the new flight, refund, eCredit, voucher) one by one, each with its own tests and sign off."},{"title":"Encode the rules, not just the knowledge","detail":"Put reaccommodation policy, waivers and passenger rights in deterministic rules or a flow that the agent calls, so the offer a passenger sees is the offer the policy allows. Let the language model explain the options, never invent them."},{"title":"Rehearse the worst day","detail":"Load test the agent, the integrations and the handover queue at the volume of your largest irregular operations day, including the reservation system's rate limits. An agent that fails under peak load is worse than none."},{"title":"Design the handover for disruption","detail":"Route complex itineraries, groups, unaccompanied minors, passengers needing assistance and codeshare or interline tickets to people, with a summary and the options already shown. Give the priority phone line to the cases the agent cannot solve."},{"title":"Close the loop with operations","detail":"Feed contact reasons and failed rebookings back to the operations control centre and the reaccommodation team, so the automatic offer improves and the agent stops getting the same question."},{"title":"Test before passengers do","detail":"Build test conversations per disruption type (weather, crew, technical, strike) and per market rule, including attempts to get a refund or compensation the rules do not allow, and run them on every change."}],"guardrails":["The agent only presents options returned by the reaccommodation engine and fare rules, never options it generates itself","Refunds, compensation and vouchers above set limits need a human approval","Entitlement statements (care, refund, compensation) come only from approved, dated policy content, with a refusal and handover when the content does not cover the case","Automatic handover for passengers needing assistance, groups, complaints and repeated failure","Payment card data tokenized before it reaches the model, and personal data masked in logs"],"humanInTheLoop":"Humans own the exceptions: complex itineraries, passengers with reduced mobility, groups, compensation disputes and complaints. The disruption desk watches live volumes and handover reasons during irregular operations, and a policy owner signs off every change to entitlement content and every new action before it goes live.","kpisToInstrument":["Share of disrupted passengers who rebook or accept an option without a human, per disruption type","Handover rate and handover reasons on irregular operations days","Repeat contacts within seven days on the same booking","Accuracy of entitlement answers on a weekly reviewed sample","Time from disruption notice to a confirmed new itinerary","Complaints and chargebacks that mention the assistant"],"failureModes":[{"title":"Wrong entitlement answers","detail":"The agent promises a refund, compensation or discount the policy does not give, and the airline is held to it, as in the Air Canada tribunal case. Keep entitlement answers in approved content with an owner and review date, and hand over when unsure."},{"title":"Collapse under peak load","detail":"The agent or the reservation integration times out on the busiest day, and passengers are pushed back to a full phone queue. Load test, cache flight status and queue writes."},{"title":"Options that are not real","detail":"The agent shows flights that no longer have seats or that the fare rules do not allow, and rebooking fails at confirmation. Only show options from live inventory and confirm before telling the passenger they are rebooked."},{"title":"Containment that is really abandonment","detail":"Passengers give up and go to the airport desk or book another airline, which looks like containment. Count repeat contacts and measure satisfaction on disruption days separately."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing assistant must tell people they are interacting with AI (Article 50). It is not a high risk use under Annex III: it applies the airline's reaccommodation rules and does not decide on access to an essential public service or on creditworthiness."},"regulations":["eu-ai-act","gdpr","uk-gdpr","pci-dss","eu-accessibility-act"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Passengers must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Air passenger rights (Your Europe)","issuer":"European Union","region":"europe","url":"https://europa.eu/youreurope/citizens/travel/passenger-rights/air/index_en.htm","note":"Summary of the EU rules on reimbursement, rerouting, care and compensation for cancelled, delayed and overbooked flights; the agent's entitlement answers must match them for flights in scope."},{"title":"Flight delays and cancellations","issuer":"UK Civil Aviation Authority","region":"europe","url":"https://www.caa.co.uk/air-passengers/travel-problems-and-rights/flight-delays-and-cancellations/","note":"UK passenger rights guidance for delays and cancellations, the reference for entitlement content on UK flights."}],"controls":["AI disclosure at the start of every conversation and in AI drafted disruption messages where required","Entitlement and policy content versioned, owned and reviewed after every regulatory or policy change","Immutable audit trail of every rebooking, refund, voucher and eCredit the agent issued","Limits per action (refund value, voucher value, number of changes) with human approval above them","Peak load and failover plan for irregular operations days"],"incidents":[{"title":"Incident 639: Air Canada Chatbot Reportedly Provides Inaccurate Bereavement Fare Information, Leading to Customer Overpayment","url":"https://incidentdatabase.ai/cite/639/","note":"A Canadian small claims tribunal held Air Canada responsible in 2024 for its chatbot's wrong statement about a bereavement fare refund (Moffatt v. Air Canada) and ordered it to pay damages. An airline cannot disclaim what its assistant tells passengers about refunds and entitlements."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the airline's\nreservation, reaccommodation, ticketing and baggage APIs, combined with a **knowledge base**\nthat holds only approved policy and passenger rights content, retrieved with hybrid search.\nThe rebooking and refund journeys run as **flows** with deterministic steps and an\nauthentication block, so the offer shown is the offer the rules allow; the agent explains the\noptions and answers open questions. **Agentic workflows** with **human in the loop** approval\nhandle refunds or vouchers above a threshold, and the **payment** service collects fare\ndifferences with card tokenization.\n\nThe same agent serves **web chat, WhatsApp, SMS and voice**, and the airline's own app through\nthe **API channel**, with\nstreaming speech on the phone for passengers who call. **Guardrails** and **PII masking** check\nevery turn, **human handover** passes the summary and the options already offered to the\ncontact centre, **test suites** replay disruption scenarios per market before every change, and\n**monitors** run scheduled health checks against the live agent. **Analytics** show\ninteractions, satisfaction and top intents per channel, custom dashboard widgets can track\nhandovers per disruption type, and the platform is model agnostic, with EU and UAE data\nresidency options."},"faq":[{"question":"What share of disruption contacts can an AI agent resolve?","answer":"No airline on this page publishes a rate for disruption contacts alone, and it depends on whether the agent can act. Across general digital support, ASAPP reports a 45% containment rate for JetBlue's virtual agent in May 2023, and Microsoft reports that 97% of nearly 4 million queries to Air India's AI.g were handled with full automation. Disruption days are harder than average, because more itineraries are complex, so plan for a lower rate and a strong handover."},{"question":"Can an airline be held to what its chatbot says?","answer":"Yes. In Moffatt v. Air Canada (2024) a Canadian tribunal held the airline responsible for its chatbot's wrong answer about a bereavement refund. Keep entitlement answers in approved, dated content and hand over when the content does not cover the case."},{"question":"Should the agent decide what a passenger is entitled to?","answer":"No. Care, refunds and compensation should come from the airline's rules engine and approved policy content, applied the same way to every passenger. The agent explains the options and completes the chosen one; disputes and exceptions go to a person."},{"question":"Is this only for airlines?","answer":"The public deployments on this page are airlines. The same pattern applies to rail and ferry operators and tour operators, wherever a disruption triggers mass rebooking and rule based entitlements."}],"related":["travel-and-hotel-booking-concierge","travel-insurance-claims-and-assistance-agent","outbound-reminder-and-confirmation-agent","complaints-handling-agent","first-line-contact-centre-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with six airline deployments verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: added European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: attributed the JetBlue and Air India figures to the vendors that report them, corrected the Air Canada tribunal description, the UK CAA link and a JetBlue summary quote attribution, added UK GDPR, the SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: meta description no longer credits Delta and Lufthansa Group with proactive alerts, the Blits.ai section names only inventory capabilities, the FAQ says the containment figures cover general service, the ConnectionSaver figure is no longer shown as interactions handled, and evidence dates and languages were corrected."}],"slug":"flight-disruption-and-rebooking-agent","url":"https://www.blits.ai/ai-use-cases/flight-disruption-and-rebooking-agent","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":8005000,"min":10000,"max":16000000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"lufthansa-group-self-service-ai-agents","pooled":true},{"id":"air-india-aig-virtual-assistant","pooled":true}]},{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":97,"min":97,"max":97,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"air-india-aig-virtual-assistant","pooled":true}]},{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":45,"min":45,"max":45,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"jetblue-asapp-digital-customer-support","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":73000,"min":73000,"max":73000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"jetblue-asapp-digital-customer-support","pooled":true}]},{"kpi":"customer-satisfaction-uplift","label":"Satisfaction uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":100,"min":100,"max":100,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"pegasus-airlines-flybot-virtual-assistant","pooled":true}]}],"indicativeValueResult":{"low":160000,"high":3600000},"evidence":["air-india-aig-virtual-assistant","delta-concierge-ai-assistant","jetblue-asapp-digital-customer-support","lufthansa-group-self-service-ai-agents","pegasus-airlines-flybot-virtual-assistant","united-airlines-disruption-messaging-and-rebooking"]},{"title":"AI agent for fraud alert confirmation with cardholders","shortTitle":"Fraud alert confirmation","seoTitle":"AI agents for fraud alert confirmation","metaDescription":"An AI agent asks the cardholder to confirm a flagged card transaction, then lifts the block or freezes the card. Includes in app and verified call examples.","definition":"A customer facing AI agent that contacts the cardholder as soon as the fraud engine flags a card transaction, in the channel they actually respond to, verifies them, asks whether they made the transaction and acts on the answer: releasing the block so a retry succeeds, or freezing the card and starting the fraud claim.","aliases":["suspicious transaction confirmation","transaction verification agent","fraud verification outreach","was this you alert"],"industries":["banking","payments"],"functions":["fraud-prevention","customer-service"],"patterns":["conversational-agent","voice-agent","agentic-workflow"],"channels":["mobile-app","sms","voice","whatsapp"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"front-office","problem":"A fraud model that stops a card transaction cannot be certain the cardholder did not make it. When\nthe customer did make it, the block interrupts a genuine purchase, often while they are still at\nthe checkout, and they need an answer before they give up or pay another way. If confirmation\ndepends on a person calling back, it can take longer than the customer stays at the checkout, and\na call from an unknown number asking about their card looks much like the scams customers are\nwarned about.\n\nThe confirmation step is therefore both a revenue problem (a genuine purchase that is declined is\nspend the issuer may not get back) and a security problem. Scammers send text messages that\nimpersonate legitimate businesses (the reason Commonwealth Bank gives for moving some card\nverification into its app), spoof the bank's\nphone numbers (Westpac has put 94,000 of its numbers on a Do Not Originate list) and can clone a\nvoice well enough to pass a voice identity check, as a journalist showed against Lloyds Bank's\nVoice ID in 2023. The job is to confirm quickly, in a channel the customer trusts, without\ncreating a new route for scammers.","problemStats":[],"howItWorks":"1. **Trigger from the scoring engine.** A flagged authorization or a card placed on hold starts\n   the agent, with the transaction details and the risk reason.\n2. **Pick the trusted channel.** The first choice is a push into the bank's app, where the\n   customer is already authenticated. Commonwealth Bank now asks app users to verify certain online\n   card transactions in the app instead of sending a code, because it can give clearer warnings\n   there than in a text message. Then two way messaging, then an outbound call that is branded and\n   verified (Westpac's SafeCall places calls through its app that show the reason for the call).\n3. **Verify, never collect secrets.** The agent confirms identity through the app or a strong\n   factor, never asks for a passcode, PIN or full card number, and treats the voice on the line as\n   untrusted.\n4. **Ask one clear question.** \"Did you try to pay 84.90 EUR at this merchant at 14:02?\" with the\n   merchant's clear name and location.\n5. **Act on the answer.** Yes: lift the block, allow the retry and tune the rule for this\n   customer. No: freeze the card, order a replacement and open the fraud claim. Unsure, or signs\n   that someone is guiding the customer: route to a scam specialist.\n6. **Handle silence.** No response within the set time keeps the block in place and follows the\n   bank's contact policy.\n7. **Log and learn.** Every alert, answer and action is recorded and fed back to the fraud team as\n   labelled outcomes.","valueDrivers":["risk-reduction","customer-experience","cost-to-serve","revenue-growth"],"kpis":["false-positive-reduction","fraud-loss-reduction","response-time-reduction","automation-rate","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A card issuer with 1 million active cards","inputs":[{"key":"cards","label":"Active cards","low":1000000,"high":1000000,"unit":"cards","note":"The reference issuer."},{"key":"alertsPerCard","label":"Fraud alerts needing customer confirmation per card per year","low":0.2,"high":0.5,"unit":"alerts per card per year","note":"Editorial assumption, replace with your own alert volume."},{"key":"agentResolvedShare","label":"Share of alerts confirmed by the agent without a human call","low":0.3,"high":0.6,"unit":"fraction of alerts","note":"Editorial assumption; depends on app adoption and on how many alerts need a specialist."},{"key":"costPerCall","label":"Cost of a human confirmation call","low":3,"high":6,"unit":"USD per call","note":"Editorial assumption, replace with your own fully loaded cost."},{"key":"genuineShare","label":"Share of alerts that are genuine customer transactions","low":0.7,"high":0.9,"unit":"fraction of alerts","note":"Editorial assumption, replace with your own alert outcomes."},{"key":"recoveredShare","label":"Share of genuine blocked spend recovered by fast confirmation","low":0.2,"high":0.4,"unit":"fraction of genuine alerts","note":"Editorial assumption."},{"key":"avgTransaction","label":"Average value of a flagged genuine transaction","low":50,"high":100,"unit":"USD","note":"Editorial assumption."},{"key":"marginRate","label":"Issuer revenue as a share of spend","low":0.01,"high":0.015,"unit":"fraction of spend","note":"Editorial assumption covering interchange and related income."}],"formula":"cards * alertsPerCard * (agentResolvedShare * costPerCall + genuineShare * recoveredShare * avgTransaction * marginRate)","currency":"USD","period":"per year","resultLabel":"Confirmation call cost avoided plus revenue from recovered genuine spend","caveat":"Leaves out fraud losses prevented by faster freezes, the lifetime value of customers who would otherwise switch cards after a false decline, messaging and telephony costs and the cost of the AI and integration."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The logic is simple; the timing is not. The agent must be triggered by the scoring engine in seconds, reach the customer in a trusted channel and change the card's status through the card platform before the customer gives up at the checkout.","dataPrerequisites":["Real time alert feed with transaction and merchant details and the risk reason","Verified contact channels and app enrolment per customer","Contact policy per alert type (channels, timing, retries, quiet hours)","Labelled outcomes of past alerts to measure false positives"],"integrations":["Fraud scoring engine and alert queue","Card management platform (release, freeze, replace, rule tuning)","App push and in app authentication","Messaging and telephony with branded or verifiable calling","Fraud claim and dispute case system"]},"implementation":{"steps":[{"title":"Move confirmation into the app first","detail":"The app is authenticated and hard to spoof. Make an in app confirmation the default and keep other channels as fallbacks for customers without the app. Commonwealth Bank, for example, now asks app users to verify certain online card transactions in the app instead of sending a one time passcode."},{"title":"Write the no secrets rule into everything","detail":"The agent never asks for a passcode, PIN or card number and tells the customer so in every message. This protects customers from scammers copying your alert."},{"title":"Close the loop with the card platform","detail":"A yes must lift the block in seconds and a no must freeze and reissue. Test both paths end to end, including the retry at the merchant."},{"title":"Connect to scam and dispute journeys","detail":"A customer who is unsure, or who describes being guided by someone, goes to a scam specialist; a confirmed fraud goes straight into the fraud claim with the details captured."},{"title":"Measure false declines, not only fraud","detail":"Track genuine transactions recovered and customers lost after a decline, alongside fraud caught, so the fraud team tunes for both."}],"guardrails":["No collection of passcodes, PINs, full card numbers or remote access in any channel","Outbound calls are verifiable in the app or come from a registered, branded number","Voice alone is never accepted as proof of identity","Freeze and reissue actions only through the card platform's allow listed APIs","Scam signals or uncertainty route to a human specialist"],"humanInTheLoop":"Fraud specialists handle uncertain answers, suspected scams, vulnerable customers and any case where the customer disputes the agent's action. The fraud team reviews alert outcomes weekly to tune rules, and approves any change to the contact policy or the actions the agent may take.","kpisToInstrument":["Median time from alert to customer answer, by channel","Share of alerts resolved without a human call","Genuine transactions recovered after confirmation","Fraud losses on alerted transactions","Complaints and satisfaction after a confirmation contact"],"failureModes":[{"title":"Your alert becomes the scammer's template","detail":"Scammers copy the wording and ask for a code. Never request secrets, say so in every alert and prefer in app confirmation."},{"title":"Slow confirmation","detail":"The answer arrives after the customer has left the checkout. Trigger in real time and prioritise the fastest trusted channel."},{"title":"Voice clone accepted as the customer","detail":"A cloned voice passes a voice check. Confirm through the app or another strong factor, not the voice."},{"title":"Silence treated as consent","detail":"No answer must never release a block. Keep it in place and follow the contact policy."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Confirming flagged transactions with cardholders is not listed in Annex III, and point 5(b) expressly excludes AI used to detect financial fraud from the creditworthiness category, so the system is not high risk. An agent that messages or calls customers must tell them they are dealing with AI under Article 50(1), and synthetic voice output must be marked as AI generated under Article 50(2)."},"regulations":["eu-ai-act","gdpr","pci-dss","uk-consumer-duty","dora","eu-psd2","us-tcpa"],"guidance":[{"title":"Declaratory ruling on AI generated voices under the TCPA (FCC 24-17)","issuer":"Federal Communications Commission","region":"north-america","url":"https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf","note":"Confirms that AI technologies that generate human voices count as an \"artificial or prerecorded voice\" under the TCPA, so US outbound AI voice calls fall under the TCPA's consent rules unless an exemption applies."},{"title":"TCPA Omnibus Declaratory Ruling and Order (FCC 15-72)","issuer":"Federal Communications Commission","region":"north-america","url":"https://docs.fcc.gov/public/attachments/FCC-15-72A1.pdf","note":"Exempts from the TCPA's consent requirements, with conditions, certain calls and texts from financial institutions to mobile numbers about transactions that suggest a risk of fraud, provided they are free to the recipient and limited to three per event over three days."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) excludes AI systems used for detecting financial fraud from the creditworthiness high risk category."}],"controls":["AI disclosure in every automated message and call","Documented contact policy per alert type with quiet hours and retry limits","Audit log of every alert, answer, identity check and card action","Regular tests of the alert channel against impersonation and spoofing","Monitoring of false positive rates and outcomes for vulnerable customers"],"incidents":[{"title":"How I Broke Into a Bank Account With an AI-Generated Voice","url":"https://www.vice.com/en/article/how-i-broke-into-a-bank-account-with-an-ai-generated-voice/","note":"In February 2023 a Vice journalist used an AI generated copy of his own voice to pass Lloyds Bank's Voice ID check and reach his account, which is why voice alone should not verify a customer in a fraud confirmation call."}]},"blitsAi":{"howToBuild":"On Blits.ai the fraud engine triggers an **agentic workflow** through its API token when an alert\nfires, and uses **custom functions** to read the alert and to release or freeze the card through\nthe card platform's APIs. The bank's own notification service sends the push or message that\nstarts the contact. The confirmation conversation then runs in the bank's app (connected through\nthe **REST API or WebSocket API channel**), or on **WhatsApp**, **SMS** or a **voice** line when\nthe customer replies or calls in that channel. A **flow** fixes the confirmation script with an\n**authentication** step and **DTMF** input on the phone, and the **AI agent** handles free text\nanswers and questions.\n\nActions above a set threshold, such as a reissue, can require **human in the loop approval**, and\n**human handover** sends uncertain or scam cases to a specialist with the transcript. **Guardrails**\nblock any request for secrets in generated messages, and **PII masking** and card number\ntokenization keep card data out of prompts. **Monitors** check the alert path on a schedule,\n**test suites** replay confirmation scenarios on every change, and **analytics** show response\ntimes and outcomes per channel."},"faq":[{"question":"Should fraud confirmation use calls, SMS or the app?","answer":"The app first, because the customer is already authenticated there and it is hard to spoof. Commonwealth Bank now asks app users to verify certain online card transactions in the app instead of sending a code, because it can give clearer warnings there. Keep messaging and calls as fallbacks for customers without the app."},{"question":"Can an AI agent unblock a card on its own?","answer":"For a clear \"yes, that was me\" from an authenticated customer, releasing the block within set limits is a reasonable automated action. A freeze after a clear \"no\" can also be automated because it is reversible, while a reissue can go through human approval above a set threshold. Anything uncertain, or with signs of a scam, should go to a person."},{"question":"How do you stop scammers imitating the alert?","answer":"Never ask for passcodes or card details, say so in every alert, and move confirmation into the app or to verifiable calls. Westpac, for example, places branded calls through its app that are verified by Optus and show the reason for the call, and has put 94,000 of its numbers on a Do Not Originate list so scammers cannot display them."}],"related":["real-time-fraud-scoring","fraud-alert-triage","scam-payment-interception","card-dispute-and-chargeback-intake","proactive-outbound-engagement-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI use case catalog and verified against Commonwealth Bank, Westpac, Revolut, Macquarie Bank, Capital One and FCC sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added PSD2, Telephone Consumer Protection Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: rewrote the Commonwealth Bank and Westpac statements to match their releases, sourced the scam threats in the problem, named Lloyds Bank and the date in the voice clone incident, sharpened the EU AI Act basis (Article 50(1) and 50(2)), corrected the Macquarie Bank note and channels, replaced the unlisted webview channel in the Blits.ai section with the REST API or WebSocket API channel, and added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: removed outbound messaging and calling from the Blits.ai section, rewrote unsourced claims in the problem as editorial framing, matched the Commonwealth Bank text message reason to its release, dropped the unsourced false positive claim from the value inputs, aligned the FAQ on freezes and reissues with the playbook, removed the implied bank AI examples from the meta description, set adoption to emerging and added the FCC fraud alert exemption (FCC 15-72)."}],"slug":"fraud-alert-confirmation","url":"https://www.blits.ai/ai-use-cases/fraud-alert-confirmation","benchmarks":[{"kpi":"fraud-loss-reduction","label":"Fraud loss reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":53,"min":30,"max":76,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"commonwealth-bank-scam-and-fraud-interventions","pooled":true},{"id":"revolut-card-scam-detection","pooled":true}]},{"kpi":"false-positive-reduction","label":"False positive reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40,"min":40,"max":40,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"macquarie-bank-fraud-protection-and-self-service","pooled":true}]}],"indicativeValueResult":{"low":193999.99999999997,"high":2069999.9999999998},"evidence":["capital-one-eno-assistant","commonwealth-bank-scam-and-fraud-interventions","macquarie-bank-fraud-protection-and-self-service","revolut-card-scam-detection","westpac-scam-call-assistant"]},{"title":"AI agent for fraud alert triage","shortTitle":"Fraud alert triage","seoTitle":"AI agents for fraud alert triage","metaDescription":"How an AI agent works the fraud alert queue: it enriches each alert, closes clear false positives under written rules and briefs analysts with a drafted rationale.","definition":"An AI agent that works the fraud alert queue behind the scenes as the analyst's first pass, without contacting the customer: it enriches each alert with customer, device and payment context, closes clear false positives under documented rules, merges duplicates, and routes genuine risk to an analyst with a drafted rationale.","aliases":["fraud case triage","fraud alert disposition","fraud analyst copilot","fraud queue automation"],"industries":["banking","payments"],"functions":["fraud-prevention","operations"],"patterns":["agentic-workflow","classification-and-routing","summarization","prediction-and-scoring"],"channels":["agent-desktop","internal-tools"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"middle-office","problem":"Every fraud engine produces a queue. Transactions held for review, customer fraud claims, alerts\nfrom device and behavioural tools and warnings from card schemes all land with analysts. Each\nalert has to be worked, whether it turns out to be fraud or a genuine customer, and each one means\nopening several systems, reading the customer's history and deciding whether to call, release or\nblock.\n\nWhen queues grow faster than teams, genuine customers wait for a held payment and real fraud gets\nworked too late. Much of the work on an alert is gathering context from several systems before an\nanalyst can judge it, and the reasoning behind a closed alert is not always recorded in a way that\ncan be audited later.","problemStats":[{"statement":"In a Feedzai survey of 562 fraud and financial crime professionals at financial institutions (March and April 2025), 43% reported increased efficiency within fraud teams from AI.","sourceTitle":"AI Fraud Trends 2025: Banks Fight Back","sourceUrl":"https://www.feedzai.com/pressrelease/ai-fraud-trends-2025/","year":2025},{"statement":"The same Feedzai survey found that 90% of financial institutions use AI to expedite fraud investigations and detect new tactics in real time.","sourceTitle":"AI Fraud Trends 2025: Banks Fight Back","sourceUrl":"https://www.feedzai.com/pressrelease/ai-fraud-trends-2025/","year":2025}],"howItWorks":"1. **Collect the alert.** Alerts from the fraud engine, device intelligence, customer claims and\n   scheme notifications arrive in one queue with a common structure.\n2. **Enrich it.** The agent pulls customer profile, recent transactions, device and location\n   history, previous alerts and any contact the customer has had, through read only tools.\n3. **Group and rank.** Duplicate alerts on the same customer or event are merged, and a model\n   ranks the rest by risk and value at stake.\n4. **Propose a disposition.** For each alert the agent drafts a disposition (release, contact the\n   customer, block, escalate) with the evidence it used. Alerts that meet documented auto clear\n   criteria are closed with that rationale stored.\n5. **Hand over.** Everything else goes to an analyst with the summary, the evidence and the\n   suggested next step; the analyst decides and the decision is logged with the agent's draft.","valueDrivers":["employee-productivity","speed","risk-reduction","customer-experience"],"kpis":["handling-time-reduction","automation-rate","false-positive-reduction","productivity-gain","time-saved-per-task"],"indicativeValue":{"referenceOrg":"A retail bank whose fraud team works 200,000 alerts a year","inputs":[{"key":"alerts","label":"Fraud alerts worked per year","low":200000,"high":200000,"unit":"alerts per year","note":"The reference bank."},{"key":"minutesPerAlert","label":"Analyst minutes per alert today","low":8,"high":15,"unit":"minutes per alert","note":"Editorial assumption. Replace with your own time study."},{"key":"timeSaved","label":"Share of analyst time saved per alert","low":0.2,"high":0.4,"unit":"fraction of handling time","note":"Editorial assumption. The low end matches the 20% cut in alert handling time Feedzai claims for the investigations skill of its Farol agent; the high end stays well below the 75% cut in daily time per person on manual reviews that Oscilar reports for Coast, which measures staff time rather than time per alert. Replace with your own pilot results.","sourceUrl":"https://www.prnewswire.com/news-releases/as-banks-pivot-to-agentic-ai-feedzai-unveils-farol-to-transform-fraud-analysis-and-cut-investigation-times-302888211.html"},{"key":"costPerHour","label":"Fully loaded analyst cost per hour","low":35,"high":60,"unit":"USD per hour","note":"Editorial assumption. Replace with your own."}],"formula":"alerts * minutesPerAlert / 60 * timeSaved * costPerHour","currency":"USD","period":"per year","resultLabel":"Analyst capacity released","caveat":"Counts analyst time only. It leaves out faster release of genuine customers' payments, losses avoided by working real fraud sooner, and the cost of the platform and integrations."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The agent only reads and drafts, which keeps the risk manageable, but it needs read access to many systems and a clear, approved definition of what may be closed without a human.","dataPrerequisites":["Historical alerts with their final dispositions and the reason recorded","Written procedures for each alert type, including what evidence an analyst checks","Access to customer, transaction, device and contact history through APIs"],"integrations":["Fraud detection engine and alert queue","Case management system","Core banking and card platforms (read only)","Device intelligence and authentication logs","Contact centre and digital banking for customer outreach"]},"implementation":{"steps":[{"title":"Profile the queue","detail":"Break the last year of alerts down by source, type, final disposition and handling time. This shows where the false positives sit and which alert types are safe to automate first."},{"title":"Codify the procedures","detail":"Turn each alert type's procedure into explicit checks and evidence requirements that the agent follows, reviewed by the fraud operations lead."},{"title":"Start as a copilot","detail":"Let the agent enrich and draft only, with analysts deciding every alert. Measure agreement between the draft and the analyst's decision per alert type."},{"title":"Introduce auto clear per alert type","detail":"Where agreement is consistently high and the risk is low, approve documented auto clear criteria for that alert type, with a sample of closures reviewed independently every week."},{"title":"Feed decisions back to detection","detail":"Share the reasons alerts turn out false with the detection team, so rules and models are tuned and fewer bad alerts are produced in the first place."}],"guardrails":["Auto clear only for alert types and thresholds approved in writing, never for high value or vulnerable customer cases","Read only access for the agent; blocks and releases need an analyst or a separate approved action","Every disposition stores the evidence and rationale used, whether drafted by the agent or written by a human","Weekly independent sampling of auto cleared alerts, with automatic rollback if the error rate exceeds a limit","Customer data masked in prompts and logs where the model does not need it"],"humanInTheLoop":"Analysts decide every escalated alert and any action that affects a customer's money. Fraud operations approves which alert types may be auto cleared and reviews a sample of those closures every week; quality assurance compares the agent's drafts with final decisions.","kpisToInstrument":["Average handling time per alert, by alert type","Share of alerts auto cleared and the error rate found in sampling","Agreement rate between the agent's draft and the analyst's final decision","Time from alert to decision for confirmed fraud","Fraud that was later confirmed on alerts the agent had proposed to clear"],"failureModes":[{"title":"Rubber stamping","detail":"Analysts accept drafts without reading them because they are usually right. Measure disagreement rates and include known fraud test cases in the queue."},{"title":"Auto clear drifting with the fraud mix","detail":"Criteria that were safe last quarter clear a new attack pattern. Review auto clear performance monthly and tie it to changes in the detection rules."},{"title":"Missing context","detail":"The agent drafts confidently from partial data when a system is unavailable. Make missing sources explicit in the draft and block auto clear when enrichment is incomplete."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Internal triage of fraud alerts is not listed in Annex III, and point 5(b) explicitly excludes fraud detection from the high risk creditworthiness category. Article 50(1) covers any system that interacts directly with people, analysts included, but it does not apply where the use of AI is obvious to a reasonably well informed user, as it is in an internal analyst tool; the marking duties for generated content in Article 50(2) sit with the provider. Reassess if its output feeds credit decisions. Decisions that affect customers remain subject to GDPR and consumer protection rules."},"regulations":["eu-ai-act","gdpr","dora","uk-consumer-duty","uk-psr-app-reimbursement","us-sr-11-7","nist-ai-rmf","iso-42001","eu-psd2"],"guidance":[{"title":"Annex III: High-Risk AI Systems Referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) carves fraud detection out of the high risk creditworthiness category."},{"title":"APP scams","issuer":"Payment Systems Regulator","region":"europe","url":"https://www.psr.org.uk/our-work/app-scams/","note":"Mandatory reimbursement for authorised push payment scams raises the cost of slow or wrong alert decisions."}],"controls":["Written auto clear criteria per alert type, approved by fraud operations and risk","Stored rationale and evidence for every closed alert","Independent weekly sampling of auto cleared alerts with an error rate limit","Access control so the agent cannot move money or change blocks on its own","Inventory entry for the agent and its prompts, with change control"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the triage runs as an **agentic workflow**, triggered through the API for each new\nalert or batch. The agent calls **custom functions** that read the case, the customer's recent\ntransactions and device history, and it can query **SQL knowledge bases** for past alerts. It\nfollows the fraud team's procedures, loaded into a **knowledge base** with hybrid retrieval, and\nreturns **structured output**: a proposed disposition, the evidence and the rationale.\n\nA tool execution policy limits the agent to read only tools, and any action that changes\nsomething requires **human in the loop approval**, where the analyst approves or rejects it.\nEvery run keeps a full audit trail. **PII masking** keeps account and card data out of prompts\nwhere it is not needed, **test suites** replay historical alerts with known outcomes before a\nchange goes live, and **monitors** run scheduled checks against the agent and alert on failure.\nThe platform is model agnostic and available with EU and UAE data residency."},"faq":[{"question":"Can an AI agent close fraud alerts on its own?","answer":"Only for alert types where it has proven it agrees with analysts and the risk is low, under written criteria, with a sample of closures reviewed every week. High value alerts, vulnerable customers and anything that blocks or releases money should stay with a human."},{"question":"How is this different from the fraud scoring model?","answer":"The scoring model decides in real time whether to hold a payment. Triage starts after that: it works the alerts and held payments the model created, gathers context and prepares or makes the disposition, on a timescale of minutes."},{"question":"What results have organizations reported?","answer":"Oscilar reports that Coast, a fleet card provider, cut the time its staff spend on manual reviews from 2 hours per person per day to under 30 minutes after adopting its case management platform with rule based routing, auto assignment and a feedback loop. Upstream of triage, Visa reports that active users of Decision Manager scoring shrank the manual review queue by 25% or more. Both are claims by the platform provider, not independent measurements."},{"question":"What should we be able to show an auditor or supervisor?","answer":"Why each alert was closed. Store the evidence and rationale for every disposition, keep the auto clear criteria under change control, and be able to show sampling results."}],"related":["real-time-fraud-scoring","fraud-alert-confirmation","aml-alert-triage","mule-network-detection","scam-payment-interception"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Blocker fix: removed the fraud alert triage link from Visa's Decision Manager evidence (its 98.7% automation rate measures real time transaction scoring, not alert triage); cleared the Coast productivity gain metric, which measured time saved on one task rather than a share of working time or output."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added PSD2 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsourced claims from the problem statement, added a second survey statistic, tied the time saved assumption to cited figures, corrected the Blits.ai build description to match the platform, added a results FAQ, SEO title and description; Coast evidence dated to 2024."},{"date":"2026-09-27","note":"Second fact check: the Coast result is now recorded as a productivity gain (staff time per day, not time per case) and no longer leads the description; SEB is recorded as announced rather than pilot; added the UK APP scam reimbursement rules; clarified the Article 50 reasoning and the results FAQ."}],"slug":"fraud-alert-triage","url":"https://www.blits.ai/ai-use-cases/fraud-alert-triage","benchmarks":[],"indicativeValueResult":{"low":186666.6666666667,"high":1200000},"evidence":["coast-oscilar-fraud-case-review","seb-feedzai-farol-fraud-agent"]},{"title":"AI agent for freight dispatch and load matching","shortTitle":"Freight dispatch and load matching","seoTitle":"AI load matching and dispatch for freight","metaDescription":"AI agents read freight quote emails and match loads to carriers. C.H. Robinson reports a 40% productivity gain, Uber Freight 12% more bookings among active carriers.","definition":"An AI agent that does the coordination work behind moving a truckload: reading an emailed quote request and replying with a price, ranking which loads to show which carriers, matching pickup and delivery details to an open dock appointment slot, and chasing the exceptions, so a broker's or carrier's own staff plan lanes and handle disputes instead of typing quotes and making appointment calls.","aliases":["AI freight broker automation","digital freight matching","AI dispatch optimization for trucking","automated freight quoting","AI load recommendation engine"],"industries":["logistics-and-transportation"],"functions":["operations"],"patterns":["agentic-workflow","recommendation-and-personalization","document-processing","prediction-and-scoring"],"channels":["email","api","internal-tools"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"freight-brokerage","problem":"Moving a truckload by road still runs on a lot of manual coordination: a shipper emails asking\nfor a price, a broker or carrier planner searches capacity and quotes back, a driver or carrier\nis matched to the load, and someone calls or emails the loading dock to book an appointment\nslot that works for both sides. C.H. Robinson, which describes its scale in the industry as\nunmatched, says it manages more than 37 million shipments a year, over 100,000 a day, across\ntruckload, less than truckload, ocean and air, for 75,000 customers, so even a small share of\nthat volume arriving as free text email adds up to a large manual workload.\n\nTwo things make the coordination hard to remove by force: freight requests do not arrive in a\nclean form (an email, a phone call, an EDI message with missing fields), and every match has\nreal constraints, capacity, lane preferences, dock hours, detention risk, that a spreadsheet or\na simple rules engine does not capture well. The work is also urgent and perishable: an unbooked\nload has to be rescheduled, which strains the relationship with the shipper, and a quote that\ntakes too long to arrive loses the business to a broker who answered first.","problemStats":[{"statement":"C.H. Robinson says it manages more than 37 million shipments a year, over 100,000 daily, serving 75,000 customers globally across truckload, less than truckload, ocean and air transportation.","sourceTitle":"With AI, C.H. Robinson is the disruptor","sourceUrl":"https://www.chrobinson.com/en-us/about-us/newsroom/news/2026/ai-disruptor-ch-robinson/","year":2026}],"howItWorks":"1. **Read the request.** The agent classifies an incoming email, EDI message or API call: is\n   this a quote request, a booking, a change, a status question, and for which lane, mode and\n   equipment.\n2. **Price or match it.** For a quote, the agent prices the lane from current rates and\n   capacity; for a load, it scores which carriers or which loads best fit, using lane history,\n   equipment, prior bookings and stated preferences, not just who searched first.\n3. **Book the real world detail.** For an appointment, the agent matches the load's ready date\n   and cargo details to an open dock or delivery slot, confirms it in the transportation\n   management system, and tells both sides.\n4. **Watch the load and flag exceptions.** Once booked, the agent tracks the shipment against\n   the plan and raises appointment conflicts, missing paperwork or capacity gaps before they\n   become a missed pickup.\n5. **Escalate what needs judgment.** Unusual freight, rate disputes, a carrier that repeatedly\n   falls through, and any request outside the agent's price or capacity limits go to a human\n   planner or account manager with the context already gathered.","valueDrivers":["employee-productivity","cost-to-serve","speed","revenue-growth"],"kpis":["productivity-gain","response-time-reduction","interactions-handled","conversion-rate-uplift","cost-reduction"],"indicativeValue":{"referenceOrg":"A freight broker handling 500,000 truckload shipments a year","inputs":[{"key":"shipments","label":"Truckload shipments per year","low":500000,"high":500000,"unit":"shipments per year","note":"The reference broker."},{"key":"tasksPerShipment","label":"Repetitive coordination tasks per shipment (quote, appointment, status update)","low":1.5,"high":2.5,"unit":"tasks per shipment","note":"Editorial assumption, replace with your own task counts per shipment."},{"key":"automationRate","label":"Share of those tasks completed without a person","low":0.15,"high":0.4,"unit":"fraction of tasks","note":"Editorial assumption, replace with your own automation rate. The low end is close to C.H. Robinson's own primary source, which says it receives over 11,000 truckload pricing emails a day and automatically replies to 2,000 of them, about 18%; the high end covers a more mature, multi year deployment. The 40% figure on this page is a company wide productivity gain, not a task automation rate, and is not used to set this range."},{"key":"minutesSavedPerTask","label":"Minutes saved per automated task","low":5,"high":12,"unit":"minutes per task","note":"Editorial assumption for a mix of quoting and appointment tasks; replace with your own time studies."},{"key":"costPerHour","label":"Fully loaded cost of a broker or dispatcher hour","low":25,"high":45,"unit":"USD per hour","note":"Editorial assumption for North America; replace with your own fully loaded cost."}],"formula":"shipments * tasksPerShipment * automationRate * minutesSavedPerTask / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Coordinator time cost avoided","caveat":"Gross time saved only. It leaves out the cost of building and running the AI and its integrations, the revenue effect of faster quotes and better load matching, and any change in detention, missed pickups or claims."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Reading a quote or a status question is the easy part; the work is in clean, current rate and capacity data, and in integrating with the transportation management system, the dock scheduling tool and whatever email or EDI gateway the shipper or carrier actually uses.","dataPrerequisites":["Historical lane, rate and booking data to train or tune a matching and pricing model","Current capacity, rates and dock appointment availability, reachable through an API, not a screen","A catalog of recurring email and EDI request types with their required fields"],"integrations":["Transportation or freight management system (loads, rates, tenders)","Dock or carrier appointment scheduling system","Email and EDI ingestion and parsing","Track and trace or telematics feed for shipment status"]},"implementation":{"steps":[{"title":"Automate the highest volume, lowest risk task first","detail":"Start with routine transactional quote replies or status questions on well understood lanes, where a wrong answer is cheap to correct, before automating anything that commits capacity or money."},{"title":"Keep pricing and matching inside a stated allow list","detail":"Define the rate bands, lanes and capacity the agent may quote or book on its own, and what it must not do (spot rates outside a margin band, new lanes, named accounts under contract review)."},{"title":"Add appointment and exception handling once quoting is stable","detail":"Layer in dock appointment matching and shipment exception flags once the team trusts the quoting behaviour, so a wrong appointment does not compound a wrong quote."},{"title":"Confirm every automated action back to a person","detail":"Send the shipper, carrier or internal planner a clear confirmation of exactly what the agent quoted, matched or booked, so no one discovers the automation only when something goes wrong."},{"title":"Test before shippers and carriers do","detail":"Build a test set of real email and EDI requests per lane and exception type, including ambiguous or incomplete ones, and run it on every change to the pricing or matching logic."},{"title":"Instrument before you scale to more lanes or modes","detail":"Measure automation rate, error rate and booking conversion on the first lanes before widening to more freight modes or customer segments."}],"guardrails":["Price and capacity actions only within a stated rate, lane and margin allow list","Human review of any automated quote or match above a size, margin or risk threshold","A clear confirmation sent to the customer or carrier of exactly what the agent did","Immutable audit trail of every automated quote, match and appointment","Regular sampling of automated quotes and matches against a human reviewed sample"],"humanInTheLoop":"Planners and account managers own the exceptions: unusual freight, rate disputes, a carrier that repeatedly falls through, and any new lane, customer or task before it is added to the agent's allow list. They also review a sample of automated quotes and matches every week.","kpisToInstrument":["Share of quotes, matches and appointments completed without a person, per lane and task type","Booking or acceptance rate of automated quotes and recommended loads versus the prior process","Time from request to quote or to a booked appointment","Error rate on automated actions caught by manual audit or by a customer complaint","Handover rate and handover reasons"],"failureModes":[{"title":"Quoting or matching on stale data","detail":"The agent prices a lane or matches capacity from rate or availability data that has already changed. Refresh rate and capacity data on a short cycle and refuse to quote when the data is stale."},{"title":"A confident wrong match","detail":"The agent books an appointment or recommends a load that does not actually fit the equipment, cargo or dock hours. Validate hard constraints (equipment type, hazmat, appointment windows) before committing, not only preference signals."},{"title":"Automation that erodes the relationship","detail":"A shipper or carrier who wanted to negotiate gets an automated reply instead. Give every automated quote and match an easy, visible way to reach a person."},{"title":"Scope creep into rate setting","detail":"New lanes or larger discretion are added to the agent's pricing allow list without a margin review. Treat every change to the allow list as a change with sign off."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Article 50(1) applies whenever the agent interacts directly with a shipper or carrier, for example replying to a quote email or confirming an appointment: the recipient must be able to tell they are dealing with an AI system, unless this is obvious from the context. Article 50(2) is a separate duty on the provider: the generated quote or confirmation text itself must be marked in a machine readable format as artificially generated, and that duty does not apply only where the system performs an assistive function for standard editing or does not substantially alter input data the deployer supplied. Whether dispatch is high risk depends on who is being ranked. Matching freight capacity and pricing a quote for a shipper is not a listed Annex III use. But allocating loads or tasks based on an individual's behaviour in a work related relationship is Annex III point 4(b), so a deployment that ranks or assigns work to a named driver or owner operator based on their own behaviour, for example an asset carrier's employed drivers or a platform ranking owner operators on their clicks, saves and booking history, needs a fresh assessment against that point even though the reference design here scores capacity and price, not a person."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Article 50(1): a system that interacts directly with a shipper or carrier must let them know they are dealing with AI, unless this is obvious from the context. Article 50(2) separately requires the provider to mark generated content, such as an automated quote or confirmation, as artificially generated in a machine readable format, an obligation that does not apply where the system only performs an assistive function for standard editing or does not substantially alter the deployer's input data."}],"controls":["Disclosure that a quote, match or confirmation may be automated, with an easy way to reach a person","Rate, lane and capacity allow list with human approval required above it","Immutable audit trail of every automated quote, match and appointment","Regular sampling of automated actions against a human reviewed sample, with an owner for corrections"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the broker's or\ncarrier's transportation management, rating and appointment scheduling systems as REST calls,\neach scoped to one action (get a rate, check capacity, book a slot). Quote replies and\nappointment confirmations run as an **agentic workflow** with **human in the loop approval**\nabove a stated margin or risk threshold, so a quote is only sent automatically inside the\nallow list; a **knowledge base** with lane notes and exception procedures, retrieved with\nhybrid search, answers open questions.\n\nThe **email channel** ingests and classifies incoming quote requests and status questions, and\nthe same agent can serve **Microsoft Teams** or **Slack** for internal planners, or the\n**REST or WebSocket API** for a carrier or shipper portal. **Guardrails** check input and\noutput before a price or a commitment goes out, and **human handover** passes an exception to\na planner with the load and conversation context attached. **Test suites** run regression\nconversations per lane and exception type on every change; **monitors** run scheduled health\nchecks against the agent with alerts on failure, and **analytics** show interactions,\nrecognition rate and satisfaction broken down by channel. The platform is model agnostic, so\na broker can route pricing decisions through a different model than open ended questions."},"faq":[{"question":"Can an AI agent negotiate freight rates on its own?","answer":"Keep automated quotes inside a stated rate and margin band, and hand anything that needs negotiation, an unusual lane or a large account to a human. C.H. Robinson, for example, describes generative AI reading a transactional truckload quote email and supplying the price from its Dynamic Pricing Engine, not negotiating contract rates."},{"question":"How is this different from a transportation management system (TMS)?","answer":"The TMS stays the system of record for loads, rates and appointments. The AI agent reads unstructured requests, such as an email or an EDI message with missing fields, decides what is being asked, and calls the TMS and scheduling systems to act, instead of a person doing that translation by hand."},{"question":"What results have freight companies reported from this kind of AI?","answer":"C.H. Robinson reports Lean AI increased its productivity by more than 40% since 2022, and in 2024 C.H. Robinson said its AI replied to 2,000 quote requests a day. Uber Freight reported in 2023 that a recommendations system lifted bookings by 12% for active carrier users in an A/B test. J.B. Hunt has put agentic AI from Overroute to work across all of its business units on millions of loads, though it has not disclosed a percentage outcome yet."},{"question":"What should stay with a human planner?","answer":"Unusual or high value freight, rate disputes, a carrier that repeatedly falls through, and any lane, customer or task outside what the agent's allow list already covers."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version. Covers freight dispatch and load matching for trucking, sourced from C.H. Robinson, Uber Freight and J.B. Hunt/Overroute, all fetched and quote checked with usecases:source."},{"date":"2026-09-28","note":"Editorial review: dated the C.H. Robinson figure of 2,000 quotes a day to 2024 and cited it to C.H. Robinson's own press release instead of trade press; dated the Uber Freight evidence to its actual 2023 A/B test instead of 2026; corrected the J.B. Hunt/Overroute summary to what the release actually says (no appointment calls or rescheduling); rewrote the EU AI Act basis to separate Article 50(1) (direct interaction) from Article 50(2) (machine readable marking, with its real exceptions) and added Annex III point 4(b) for task allocation based on individual behaviour, so the tier is now context dependent; corrected the automation rate assumption and its note; reworded FAQ[0] as advice; and named real Blits.ai capabilities (monitors, analytics, Microsoft Teams or Slack) in the how to build section instead of unlisted ones."}],"slug":"freight-dispatch-and-load-matching-agent","url":"https://www.blits.ai/ai-use-cases/freight-dispatch-and-load-matching-agent","benchmarks":[{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":12,"min":12,"max":12,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"uber-freight-ai-load-recommendations","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2000,"min":2000,"max":2000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ch-robinson-lean-ai-freight-automation","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40,"min":40,"max":40,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ch-robinson-lean-ai-freight-automation","pooled":true}]}],"indicativeValueResult":{"low":234375,"high":4500000},"evidence":["ch-robinson-lean-ai-freight-automation","jb-hunt-overroute-agentic-freight-execution","uber-freight-ai-load-recommendations"]},{"title":"AI agent for inbound lead qualification and meeting booking","shortTitle":"Inbound lead qualification","seoTitle":"AI SDR agents for inbound lead qualification","metaDescription":"AI SDR agents answer inbound buyers around the clock, qualify them and book meetings. See deployments at 8x8, SUSE, CarMax and Rocket Mortgage.","definition":"An AI agent that engages inbound prospects the moment they arrive on the website, chat, messaging or the sales phone line, answers their first questions, qualifies them against the organization's criteria, and books a meeting or hands a ready conversation to the right salesperson, with the context written into the CRM.","aliases":["AI SDR","AI sales development representative","lead qualification chatbot","inbound sales agent","conversational lead capture"],"industries":["cross-industry","technology","automotive","banking"],"functions":["sales","marketing"],"patterns":["conversational-agent","voice-agent","classification-and-routing","agentic-workflow"],"channels":["web-chat","voice","email","whatsapp"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Inbound interest comes from people who chose to reach out, and much of it is wasted. Web forms ask\nfor a lot and answer nothing; chat requests go unanswered outside office hours; sales development\nreps spend their day separating buyers from job seekers and support requests, while the real buyer\nwaits for a reply and may book with a competitor. On the phone, callers navigate menus to reach a\nsalesperson who first has to ask the same questions again.\n\nThe deployments on this page show the same pattern in software, car retail and mortgages. At 8x8,\nmore than half of website sessions happened outside business hours, and a meaningful share of chat\nrequests went unanswered, according to its vendor Qualified. A sales team that works office hours\nin one language leaves the rest of the day and the rest of the market to a form. A form or a\nscripted chatbot can capture an email address but cannot hold a discovery conversation. Agents\nthat understand the product, ask the next useful question and act in the CRM and calendar are\nmeant to close that gap.","problemStats":[],"howItWorks":"1. **Engage at the moment of intent.** The agent greets the visitor or caller, in their language,\n   with context from the page they are on or the campaign they came from, and says it is an AI.\n2. **Answer first questions from approved content.** Product, pricing principles, availability and\n   next steps come from the organization's approved knowledge, with a refusal when it does not know.\n3. **Qualify with discovery questions.** It asks what a good salesperson would (need, size,\n   timing, budget, location) and recognises known accounts from the CRM and intent data, instead of\n   presenting a long form.\n4. **Route by rules.** Qualification and routing rules decide the next step: book a meeting with\n   the right rep or specialist, transfer a live call, offer self service, or send a support request\n   or job seeker to the right place.\n5. **Write it down.** The conversation summary, answers and score go into the CRM so the rep starts\n   where the agent stopped.\n6. **Follow up with consent.** Visitors who leave without booking get a follow up by email or\n   messaging only where consent and contact rules allow.","valueDrivers":["revenue-growth","speed","cost-to-serve","customer-experience"],"kpis":["conversion-rate-uplift","revenue-uplift","interactions-handled","response-time-reduction","productivity-gain"],"indicativeValue":{"referenceOrg":"A B2B software company with 20,000 inbound leads a year","inputs":[{"key":"leads","label":"Inbound leads per year","low":20000,"high":20000,"unit":"leads per year","note":"The reference company."},{"key":"baseConversion","label":"Lead to closed deal conversion today","low":0.01,"high":0.02,"unit":"fraction of leads","note":"Editorial assumption for B2B inbound. Replace with your own funnel data."},{"key":"uplift","label":"Relative uplift in lead to deal conversion","low":0.05,"high":0.15,"unit":"fraction","note":"Conservative against the evidence on this page (Qualified reports that 8x8 saw 19% better MQL to SQL conversion and 24% more MQL to closed won deals in its first nine months), because vendor case studies select their best results."},{"key":"dealValue","label":"Average first year deal value","low":15000,"high":30000,"unit":"USD per deal","note":"Editorial assumption. Replace with your own average contract value."}],"formula":"leads * baseConversion * uplift * dealValue","currency":"USD","period":"per year","resultLabel":"Additional first year revenue from better inbound conversion","caveat":"Revenue, not margin, and only the conversion effect. It leaves out the cost of the agent, time saved by sales development reps, after hours coverage beyond the leads counted, and the risk that faster qualification also brings forward deals that would have closed anyway."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"A chat that asks questions is simple. The work is in written qualification rules sales agrees with, CRM and calendar integration with the right routing, product answers that stay within approved claims, and consent handling for follow up.","dataPrerequisites":["Written qualification criteria and routing rules agreed between marketing and sales","Approved product, pricing principle and competitive content","CRM account and contact data, and intent data where available","Consent and contact preference records for follow up"],"integrations":["CRM (leads, contacts, accounts, activities)","Calendar and meeting booking for reps and specialists","Website, chat and telephony for live transfer","Marketing automation for consented follow up","Intent or enrichment data providers"]},"implementation":{"steps":[{"title":"Agree the definition of a qualified lead","detail":"Write down with sales what makes a lead worth a meeting, who gets which lead, and what happens to the rest. The agent can only be as consistent as the rules."},{"title":"Start with after hours and overflow","detail":"Put the agent where nobody answers today (nights, weekends, peak campaigns) so the uplift is easy to see and nobody's pipeline is taken away on day one."},{"title":"Ground every product answer","detail":"Load approved product and pricing content with owners, and make the agent hand over rather than improvise on pricing, discounts, legal terms or roadmap."},{"title":"Integrate booking and CRM before launch","detail":"A qualified conversation that does not land in the CRM or a rep's calendar is a lost lead. Test routing for every territory and segment."},{"title":"Review conversations with sales every week","detail":"Sales and marketing read a sample of qualified and disqualified conversations together and adjust questions and rules."},{"title":"Measure against a baseline","detail":"Compare lead to meeting, meeting to opportunity and closed won rates with the period before, or with a control group, not only the number of conversations."}],"guardrails":["No price, discount or contractual commitment unless it comes from a system of record","Qualification and routing by written rules, with the reason stored in the CRM","AI disclosure at the start of the conversation and on voice calls","Follow up only with valid consent and within contact rules and quiet hours","Protection against prompt injection and attempts to extract confidential information"],"humanInTheLoop":"Sales owns the qualification rules and every commercial commitment. Reps take over qualified conversations, and a sales and marketing pair reviews a weekly sample of agent conversations, including disqualified ones, to catch good buyers the rules turned away.","kpisToInstrument":["Lead to meeting and meeting to opportunity conversion versus baseline","Speed to first response and share of inbound answered outside business hours","Share of conversations disqualified and the reasons, with a sample checked by sales","Closed won revenue from agent qualified leads","Opt outs and complaints from follow up"],"failureModes":[{"title":"The agent makes promises","detail":"Prospects push for prices, discounts or commitments and a fluent agent agrees. Keep commercial terms out of the model and test for manipulation."},{"title":"Qualifying out good buyers","detail":"Rigid rules turn away real buyers who answer one question the wrong way. Review disqualified conversations, not only qualified ones."},{"title":"A pipeline that sales does not trust","detail":"If rep and agent disagree on what qualified means, reps ignore the leads. Agree the rules first and show the reasoning in the CRM."},{"title":"Follow up that breaks consent rules","detail":"Automated email and calls to people who did not consent create regulatory and brand risk. Check consent before every follow up."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing sales agent must make clear that people are talking to an AI system, unless that is obvious (Article 50(1)). Qualifying and routing prospects is not an Annex III use. It becomes high risk where the same system takes on an Annex III task, for example evaluating the creditworthiness of natural persons (Annex III point 5(b)) or assessing risk and pricing for life or health insurance (point 5(c)); those decisions then need the high risk controls."},"regulations":["eu-ai-act","gdpr","uk-gdpr","us-tcpa"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Providers must design systems that interact directly with people so that they are informed they are dealing with an AI system, unless this is obvious from the context."},{"title":"Direct marketing and privacy and electronic communications","issuer":"Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/","note":"UK rules on consent for marketing emails, texts and calls, which govern automated follow up of leads."},{"title":"Declaratory ruling on AI generated voices under the TCPA (FCC 24-17)","issuer":"Federal Communications Commission","region":"north-america","url":"https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf","note":"Calls using AI generated voices count as artificial or prerecorded voice calls under the TCPA, so outbound follow up calls need prior consent in the United States."}],"controls":["AI disclosure in chat and at the start of calls","Consent check before every follow up by email, messaging or phone","Qualification rules versioned, with the reason for each decision stored","Audit log of meetings booked and leads routed by the agent","Regular review of disqualified conversations for bias against segments or regions"],"incidents":[{"title":"Incident 622: Chevrolet dealer chatbot agrees to sell Tahoe for $1","url":"https://incidentdatabase.ai/cite/622/","note":"A car dealer's sales chatbot was manipulated into agreeing to sell a vehicle for one dollar and recommending a competitor, the typical failure of a sales agent without commercial guardrails."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** of approved product and pricing\ncontent, retrieved with hybrid search, and the ready made **HubSpot, Microsoft Dynamics 365 or Zoho** tools\n(or **custom functions** for another CRM such as Salesforce) to look up accounts, create leads and\nlog the conversation. Qualification questions and routing rules can run in a **flow** with deterministic\nconditions, while the agent handles the open conversation; booking a meeting is a function call to\nthe calendar.\n\nThe agent runs on **web chat** (with proactive popup messages), **WhatsApp, email and voice**,\nwhere it can transfer a live call to a rep. **Guardrails** block prompt injection by default, and\nan input and output policy you write can stop commitments on price; **PII masking** protects\ncontact data, and the GDPR toolkit handles consent messages and data removal. **Human handover**\nbrings a rep into the conversation, and a flow can send an AI summarized conversation history to\nthe sales team. **Test suites** replay qualification scenarios on every change, and **analytics**\nshow interactions, satisfaction and sentiment. The platform is model agnostic."},"faq":[{"question":"How much does an AI SDR improve inbound conversion?","answer":"Vendor case studies report gains. Qualified reports that 8x8 saw 19% better MQL to SQL conversion and 24% more MQL to closed won deals in its first nine months with an AI SDR agent. These are vendor selected results with self selection in them; measure your own against a baseline or a control group."},{"question":"Should the agent quote prices?","answer":"Only list prices and pricing principles from approved content or a system of record. Discounts and commitments belong to people; the Chevrolet dealer chatbot that \"agreed\" to a one dollar car shows what happens otherwise."},{"question":"Does it work on the phone as well as in chat?","answer":"Yes. CarMax uses AI voice agents on its inbound sales calls to understand what the caller needs, answer common questions such as vehicle availability and pass the caller to the right associate faster, and Rocket Mortgage runs its digital assistant across chat and voice."}],"related":["conversational-shopping-assistant","sales-call-coaching-and-crm-update","proactive-outbound-engagement-agent","home-loan-assistant-and-prequalification","business-connectivity-quoting-and-service-assistant","personalized-marketing-at-scale"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, cross industry customer facing vertical; evidence from 8x8, SUSE, CarMax and Rocket Mortgage."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: attributed the 8x8 figures to Qualified, removed unsupported sectors and superlatives from the problem, made the EU AI Act basis precise (Article 50(1), Annex III points 5(b) and 5(c)), added UK GDPR and TCPA, aligned the Blits.ai analytics wording with the feature inventory, added the Rocket Mortgage three times close rate, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: recorded the 8x8 24% as MQL to closed won deal conversion (as in the detailed results) instead of revenue, removed the Rocket Mortgage four times metric (no comparison group stated), softened the EU AI Act basis and the problem wording, and tightened the Blits.ai build description."}],"slug":"inbound-lead-qualification-agent","url":"https://www.blits.ai/ai-use-cases/inbound-lead-qualification-agent","benchmarks":[{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":19,"min":19,"max":19,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"8x8-ai-sdr-inbound-qualification","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":400000,"min":400000,"max":400000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"rocket-mortgage-digital-assistant","pooled":true}]}],"indicativeValueResult":{"low":150000,"high":1800000},"evidence":["8x8-ai-sdr-inbound-qualification","carmax-inbound-sales-voice-agent","rocket-mortgage-digital-assistant","suse-ai-sdr-inbound-qualification"]},{"title":"AI agent for insurance policy servicing","shortTitle":"Policy servicing agent","seoTitle":"AI agents for insurance policy servicing","metaDescription":"AI agents answer coverage questions and make routine policy changes. Lemonade's 10-K says its bot handles over half of customer inquiries without human help.","definition":"An AI agent that answers policyholders' coverage questions from their own policy documents and completes routine policy changes and document requests (address and vehicle changes, adding a named driver or item, payment method updates, certificates and proof of cover) across chat, messaging and phone, and hands anything complex or sensitive to a human with the context.","aliases":["policyholder service chatbot","insurance customer service agent","coverage question assistant"],"industries":["insurance"],"functions":["customer-service","operations"],"patterns":["conversational-agent","voice-agent","rag-knowledge-assistant","agentic-workflow"],"channels":["web-chat","mobile-app","whatsapp","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"policy-administration","problem":"Most contact with an insurer between purchase and claim is routine: am I covered for this, send me\nmy certificate, I changed car, add my partner, update my card. Each request is simple but depends on\nthe specific policy wording, endorsements and product version, so front line staff spend time\nlooking things up and customers wait on hold at renewal peaks.\n\nMany first generation chatbots answered generic FAQs and could not see the customer's policy or change\nanything, so the conversation ended in a queue anyway. Coverage answers are also regulated: a wrong\n\"yes, you're covered\" becomes a complaint or a dispute when a claim is declined.","problemStats":[],"howItWorks":"1. **Identify and authenticate.** The agent verifies the policyholder in proportion to the request:\n   a logged in session for questions, step up checks before changes.\n2. **Answer from the customer's own policy.** Coverage questions are answered by retrieving the\n   policy schedule, wording and endorsements that apply to this customer, with the clause cited.\n3. **Complete routine changes.** Through an allow list of policy system actions the agent updates\n   details, adds drivers or items within set limits, issues documents and takes payments, and\n   shows any premium change before the customer confirms.\n4. **Know when to stop.** Claims, complaints, cancellations with refunds above a threshold,\n   vulnerability signals and ambiguous coverage questions go to a human with the conversation\n   summary.\n5. **Learn from the gaps.** Unanswered questions and handovers are reviewed to fix content and add\n   new intents.","valueDrivers":["cost-to-serve","customer-experience","inclusion-and-access","employee-productivity"],"kpis":["containment-rate","first-contact-resolution","interactions-handled","customer-satisfaction","handling-time-reduction"],"indicativeValue":{"referenceOrg":"A personal lines insurer with 1 million policyholders","inputs":[{"key":"policyholders","label":"Policyholders","low":1000000,"high":1000000,"unit":"policyholders","note":"The reference insurer."},{"key":"contactsPerPolicyholder","label":"Servicing contacts per policyholder per year","low":0.5,"high":1,"unit":"contacts per policyholder per year","note":"Editorial assumption, excluding claims contacts. Replace with your own contact volume."},{"key":"containment","label":"Share of servicing contacts the agent resolves","low":0.25,"high":0.5,"unit":"fraction of contacts","note":"In line with the evidence on this page (Lemonade's 10-K says over half of inquiries are handled without human intervention)."},{"key":"costPerContact","label":"Cost of a human handled contact","low":4,"high":8,"unit":"USD per contact","note":"Editorial assumption for a blended phone, chat and email contact. Replace with your own fully loaded cost."}],"formula":"policyholders * contactsPerPolicyholder * containment * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled servicing cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of running the agent, integration with the policy system, and effects on retention and complaints, which can go either way depending on answer quality."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering from generic content is easy; answering from the customer's own wording and changing the policy is the work. Legacy policy administration systems may lack APIs for mid term adjustments, and product wordings exist in many versions.","dataPrerequisites":["Policy wordings, schedules and endorsements by product version, linked to each policy","A catalog of servicing intents with volumes from the contact centre","Rules for which changes can be made without underwriting review, and their limits","Approved answers for regulated topics (cancellation rights, complaints, claims)"],"integrations":["Policy administration system (read and mid term adjustment APIs)","Document generation for certificates and proof of cover","Payment provider for premium changes","Contact centre platform for handover with context","Identity verification and customer portal login"]},"implementation":{"steps":[{"title":"Rank intents by volume and risk","detail":"Start with document requests, payment updates and simple coverage questions; leave cancellations with refunds and anything touching claims for a later wave."},{"title":"Link answers to the right wording","detail":"Index wordings by product and version and retrieve by the customer's policy, not by keyword. If the customer is not identified, answer only in general terms and say so."},{"title":"Define the change allow list","detail":"For every change write the API call, the authentication level, the underwriting limits (for example which vehicles or sums insured may be changed without review) and the confirmation the customer sees."},{"title":"Design handover and vulnerability rules","detail":"Hand over on complaints, claims, bereavement, financial difficulty and repeated failure, with the summary and verified identity passed to the human."},{"title":"Test with real conversations","detail":"Build a test set from transcripts, including tricky coverage questions and attempts to push the agent into confirming cover it cannot confirm, and run it on every change."}],"guardrails":["Coverage answers only from the customer's own policy documents, with the clause cited","Explicit wording that the agent does not decide claims, with handover for any claim question","Premium changes shown and confirmed by the customer before they apply","Step up authentication before changes to payment details or named persons","AI disclosure at the start of the conversation and an easy route to a human"],"humanInTheLoop":"Humans handle claims, complaints, vulnerable customers and any change outside the allow list. A service quality team reviews a weekly sample of contained conversations, with extra focus on coverage answers, and approves every new intent before release.","kpisToInstrument":["Containment per intent, counting repeat contacts within seven days as not contained","Accuracy of coverage answers on a weekly audited sample","Handover rate and reasons","Customer satisfaction for AI handled versus human handled contacts","Complaints that mention the assistant"],"failureModes":[{"title":"Confident wrong coverage answers","detail":"The agent answers from the current product wording when the customer holds an older version. Retrieve by policy and version, and refuse when unsure."},{"title":"Changes that should have gone to underwriting","detail":"A mid term change alters the risk (a new driver, a higher sum insured) without review. Encode underwriting limits in the allow list."},{"title":"Containment that is really abandonment","detail":"Customers give up and call instead. Measure repeat contacts and satisfaction per intent."},{"title":"Claims conversations handled as service","detail":"A customer describes a loss while asking about cover. Detect claim signals and hand over to claims."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing assistant must be designed so that people know they are interacting with AI (Article 50(1), applicable from 2 August 2026). It is not high risk as long as it does not carry out risk assessment and pricing in relation to natural persons in life and health insurance (Annex III point 5(c))."},"regulations":["eu-ai-act","gdpr","uk-gdpr","dora","uk-consumer-duty","pci-dss","eu-idd"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Published 6 August 2025 and addressed to national supervisors. Sets out risk based, proportionate expectations for insurers using AI systems, including fairness, transparency and explainability, and human oversight, and mentions chatbots as an example use."}],"controls":["AI disclosure and a visible route to a human in every channel","Inventory entry with an accountable owner and a documented action allow list","Audit trail of every policy change the agent made","Regression tests on coverage questions for each wording release","Monitoring of complaints and outcomes for vulnerable customers"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** of policy wordings and endorsements,\nsearched with **hybrid retrieval**, plus **custom functions** that call the policy administration\nsystem's REST APIs to fetch the customer's own schedule and wording version and to make mid term\nchanges. Regulated journeys, such as changing payment details, run as a **flow** with an\nauthentication step, and **payment links** handle any premium due.\n\nThe same agent serves **web chat, WhatsApp and voice**, and the insurer's own mobile app through the\n**REST or WebSocket API channel**, with streaming speech on the phone and **multi language** support. **Guardrails** check inputs and outputs, **PII masking** and\ncard number tokenization happen at the gateway, and **human handover** passes the summary to the\ncontact centre, including live takeover. **Test suites** replay real coverage questions on every\nwording release, **analytics** show interactions, satisfaction and top intents, and **conversation\nlogs** show each handover in full."},"faq":[{"question":"What share of customer questions can an AI agent handle?","answer":"Published figures sit between about 30% and 60%, though each covers a different scope. Infobip reports that LAQO's assistant, which covers claims and general information about LAQO rather than policy servicing, handles 30% of customer queries. Nsure.com says its copilot handles around 60% of customer questions. Lemonade's 10-K says over half of its customer inquiries are handled by its bot platform without human intervention. Waterdrop reported in its second quarter 2025 results that its AI Customer Service Agent resolved 60% of inquiries on first contact."},{"question":"Can the agent tell a customer whether they are covered?","answer":"It can explain what the customer's own policy wording says, with the clause cited, and should hand over when the answer depends on the facts of a loss. Claims decisions stay with the claims team."},{"question":"Is a policy servicing chatbot high risk under the EU AI Act?","answer":"Usually not; it carries the Article 50 duty to make clear that customers are talking to AI. It would become high risk under Annex III point 5(c) if it were used for risk assessment and pricing in relation to individuals in life and health insurance."}],"related":["claims-first-notice-of-loss-agent","conversational-insurance-quote-and-buy","insurance-renewal-and-retention","first-line-contact-centre-agent","account-servicing-execution"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from insurer filings and vendor case studies verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added Insurance Distribution Directive to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: Nsure.com metric now attributed to the organization (a named VP), LAQO figure attributed to Infobip, Annex III point 5(c) and the Article 50 date added, EIOPA guidance note corrected, Blits.ai build limited to inventory capabilities; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: Blits.ai analytics claim limited to inventory metrics, Waterdrop period attributed to its results release, Annex III wording quoted exactly, EIOPA note corrected, UK GDPR added, LAQO evidence year set to the case study's 2023 publication."},{"date":"2026-09-27","note":"Review fixes: FAQ reworded to say customer questions rather than policy service requests and to state LAQO's actual scope (claims and general information, not policy servicing); LAQO dropped from the indicative containment note."}],"slug":"insurance-policy-servicing-agent","url":"https://www.blits.ai/ai-use-cases/insurance-policy-servicing-agent","benchmarks":[{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":40,"min":30,"max":50,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"lemonade-ai-maya-and-cx-ai","pooled":true},{"id":"laqo-pavle-digital-assistant","pooled":true}]},{"kpi":"first-contact-resolution","label":"First contact resolution","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":60,"min":60,"max":60,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"waterdrop-guardian-ai-insurance-assistants","pooled":true}]}],"indicativeValueResult":{"low":500000,"high":4000000},"evidence":["laqo-pavle-digital-assistant","lemonade-ai-maya-and-cx-ai","nsure-friendly-john-copilot","sun-life-ai-underwriting-and-client-service","waterdrop-guardian-ai-insurance-assistants","zurich-hong-kong-whatsapp-service-agent"]},{"title":"AI agent for IT service desk resolution","shortTitle":"IT service desk resolution","seoTitle":"AI agents for IT help desk ticket resolution","metaDescription":"An AI agent resets passwords, unlocks accounts and routes IT tickets in Teams or Slack. Bank of America more than halved its IT service desk calls.","definition":"An AI agent in Microsoft Teams, Slack or the intranet that takes the high volume IT support queue, such as password and MFA resets, account unlocks, VPN, device and software requests, and resolves common requests by acting in the identity and IT service management systems, handing the rest to the right resolver group with the context attached.","aliases":["IT helpdesk chatbot","virtual IT support agent","employee IT support assistant","ITSM virtual agent"],"industries":["cross-industry","banking","technology","retail-and-ecommerce","healthcare"],"functions":["it-and-engineering","operations"],"patterns":["conversational-agent","agentic-workflow","rag-knowledge-assistant","classification-and-routing"],"channels":["microsoft-teams","internal-tools","web-chat"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"Most organizations run an IT service desk, and much of its volume is the same few requests:\nforgotten passwords, locked accounts, MFA devices, VPN trouble, a new laptop, access to an\napplication. Each one is quick for an analyst, but they arrive in bursts (Monday mornings,\nafter a password policy change, during an outage) and employees wait in a queue for something a\nsystem could have done in seconds. The first line desk spends its day reading and routing tickets\ninstead of fixing the harder problems.\n\nThe first wave of IT chatbots answered with a knowledge article. The employee still had to follow\nthe steps or raise a ticket anyway. The step change is an agent that is connected to the identity\nprovider and the ITSM tool, can reset, unlock, provision and route within strict limits, and\nlogs every action like a human analyst would. In a bank the same agent also has to respect\nentitlement rules and access reviews, because access changes are a control, not just a service.","problemStats":[],"howItWorks":"1. **Understand the request.** The employee writes in their own words in Teams, Slack or the\n   portal (\"my VPN keeps dropping\", \"I need Visio\"). The agent classifies the intent and\n   extracts the details it needs.\n2. **Verify the person proportionally.** The chat session is already signed in through single\n   sign on. Sensitive actions such as a password or MFA reset need a step up check, because\n   the service desk is a known target for social engineering.\n3. **Act through approved tools.** For a small allow list of requests the agent calls the\n   identity provider or ITSM APIs directly: reset, unlock, add to a group, start a software\n   request that runs the normal approval, and confirms the result.\n4. **Answer from the IT knowledge base.** How to questions are answered by retrieval over\n   current, owned IT articles, with a link to the source.\n5. **Route what it cannot finish.** Everything else becomes a ticket with a summary, category\n   and the right assignment group, so no analyst has to read and reroute it.\n6. **Log everything.** Every action is written to the ITSM record with the requester, the\n   action and the result, so audit and access reviews see the same trail as for a human.","valueDrivers":["cost-to-serve","employee-productivity","speed"],"kpis":["containment-rate","automation-rate","contact-deflection","accuracy","employee-adoption","users-served","processing-time-reduction","productivity-gain"],"indicativeValue":{"referenceOrg":"An organization with 20,000 employees","inputs":[{"key":"employees","label":"Employees served by the service desk","low":20000,"high":20000,"unit":"employees","note":"The reference organization."},{"key":"ticketsPerEmployee","label":"IT support contacts per employee per year","low":6,"high":10,"unit":"contacts per employee per year","note":"Editorial assumption, replace with your own ticket and call volume."},{"key":"resolvedShare","label":"Share of contacts the agent resolves without an analyst","low":0.3,"high":0.6,"unit":"fraction of contacts","note":"Conservative against the evidence on this page (IBM reports over 75% of AskIT queries resolved by an assistant that mainly answers from IT support content; Moveworks reports over 74% of issues at Mercari US handled autonomously), because early waves cover fewer intents."},{"key":"costPerTicket","label":"Fully loaded cost of an analyst handled contact","low":15,"high":25,"unit":"USD per contact","note":"Editorial assumption for a blended first line desk. Replace with your own cost per ticket."}],"formula":"employees * ticketsPerEmployee * resolvedShare * costPerTicket","currency":"USD","period":"per year","resultLabel":"Analyst handled contact cost avoided","caveat":"Gross avoided handling cost only. It leaves out the cost of the platform and integrations, the productivity of employees who get unblocked faster, and any reduction in outsourced desk contracts, which usually only happens at renewal."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering IT questions is easy. The work is in the integrations with the identity provider and the ITSM tool, in safe verification before resets, and in keeping the knowledge base current. Organizations with messy assignment groups need to clean them up before routing works well.","dataPrerequisites":["Ticket history with categories and resolutions, to pick the first intents by volume","Current IT knowledge articles with an owner and review date","A clean list of assignment groups and what each one handles","Entitlement rules for what can be granted without approval"],"integrations":["Identity provider (password reset, unlock, MFA, group membership)","ITSM platform such as ServiceNow, Jira Service Management or Freshservice","Collaboration channel (Microsoft Teams, Slack) and the intranet","Device and software management for provisioning requests","HR system for joiner, mover and leaver context"]},"implementation":{"steps":[{"title":"Pick intents from the ticket data","detail":"Export a year of tickets and calls, group them by resolution, and choose the ten to twenty intents that are both frequent and safe to automate. Password, unlock and access requests are usually the top of the list."},{"title":"Write the action allow list","detail":"For each action, record the API, the verification it needs, the limits (which groups, which software, which approvals) and what the agent says when it cannot proceed."},{"title":"Harden the reset flows","detail":"Treat password and MFA resets as the highest risk actions. Require step up verification tied to something the attacker does not have, and alert on unusual reset patterns."},{"title":"Clean and connect the knowledge base","detail":"Retire stale articles, give every article an owner, and make the agent refuse and route when retrieval finds nothing rather than improvising steps."},{"title":"Route with context","detail":"For requests the agent cannot finish, create the ticket with a summary, category and assignment group. Measure how often analysts reassign it and tune from there."},{"title":"Launch where people already ask for help","detail":"Put the agent in the chat tool employees already use, announce it with a few concrete examples, and track adoption and repeat contacts per intent."}],"guardrails":["Actions only through an allow list of API calls, each with its own verification level and limits","Step up verification before any password, MFA or privileged access change","Access requests follow the same approval and least privilege rules as a human request","Answers only from owned, current IT articles, with refusal and routing when nothing matches","Every action written to the ITSM record and an immutable audit log"],"humanInTheLoop":"Analysts own everything outside the allow list, all privileged access, and any request that fails verification. Approvers stay in the loop for software and access that needs approval, and the desk reviews a weekly sample of resolved conversations and reset logs.","kpisToInstrument":["Share of contacts resolved by the agent per intent, counting a repeat contact within seven days as not resolved","Automated routing accuracy, measured by reassignment rate","Employee adoption (share of employees who used the agent in the last 30 days)","Median time to resolution for automated versus analyst handled tickets","Reset volume and anomalies, reviewed by security"],"failureModes":[{"title":"The help desk becomes the attack path","detail":"Social engineering against resets works on bots as well as people. Tie resets to strong verification, rate limit them and alert security on unusual patterns."},{"title":"Deflection instead of resolution","detail":"The agent sends an article, the employee gives up and phones the desk. Measure repeat contacts, not just conversations closed."},{"title":"Access creep","detail":"Convenient self service grants access without the approvals that access reviews assume. Keep approvals and least privilege identical to the manual path."},{"title":"Stale knowledge","detail":"Articles describe old tools and old screens. Give each article an owner and a review date and retire what is not maintained."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1) requires an assistant that talks with people to make clear they are interacting with AI, unless that is obvious from the context. It is not listed in Annex III. The agent does allocate work, but it routes tickets to resolver and assignment groups based on the content of the request, not to individual workers based on their behaviour or personal traits or characteristics, so Annex III point 4(b) does not apply. It also does not decide on recruitment, promotion, credit or access to essential services. Any use that assigns work to individual analysts, or monitors and evaluates them from their behaviour or performance (including through the agent's logs), would need its own assessment."},"regulations":["eu-ai-act","gdpr","dora","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Guidelines on Risk Management Practices, Technology Risk","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/guidelines/technology-risk-management-guidelines","note":"Example of regional expectations on access management, privileged access and logging that an automated service desk has to meet at a financial institution."},{"title":"OWASP Top 10 for LLM Applications","issuer":"OWASP Gen AI Security Project","region":"global","url":"https://genai.owasp.org/llm-top-10/","note":"Covers prompt injection and excessive agency, the two main risks when an agent can change accounts and access."},{"title":"Scattered Spider, cybersecurity advisory AA23-320A","issuer":"CISA and FBI","region":"north-america","url":"https://www.cisa.gov/news-events/cybersecurity-advisories/aa23-320a","note":"Describes a criminal group that phones IT help desks to get passwords and MFA tokens reset, the attack path that automated resets must be designed against."}],"controls":["Inventory entry with an accountable owner and a documented action allow list","Same entitlement and approval rules for automated and manual access changes","Immutable log of every automated action, reviewed in periodic access reviews","Security monitoring on reset and unlock volumes","Change control and regression tests for every new intent or action"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** in the **Microsoft Teams** or **Slack** channel, with\n**custom functions** that call the identity provider and ITSM APIs (REST calls with their own\ncredentials and limits). The integration catalog includes systems such as ServiceNow, Jira and\nOkta, and Freshdesk is available as a ready made tool. How to questions are answered from a\n**knowledge base** of IT articles with hybrid retrieval, and a **flow** handles the deterministic steps of a reset, including an\nauthentication step, before the agent is allowed to call the reset function.\n\n**Agentic workflows** can run multi step requests such as provisioning, with **human in the\nloop approval** above a set threshold and a **tool execution policy** that limits which tools\nthe agent may use. Every run has an audit trail. **Guardrails** check input and output,\n**PII masking** keeps personal data out of prompts, and **human handover** passes the\nconversation to an analyst. **Test suites** run automated conversation evaluations of the agent\nbefore changes go live, **monitors** run scheduled health checks, and **analytics** show\nvolumes, recognition rate and satisfaction. Model agnostic routing and EU or UAE data residency\nare available, and administrators sign in to the platform with single sign on through Microsoft\nEntra or Google."},"faq":[{"question":"What share of IT requests can an AI agent resolve?","answer":"Published figures are high, for assistants that act and for those that mainly answer. Moveworks reports that Mercari US's assistant, which resets passwords, edits email groups and provisions software, handles over 74% of issues autonomously. IBM reports that over 75% of queries submitted to AskIT in its first four months were resolved by the assistant, which mainly surfaces answers from IT support content. Start lower: early waves cover fewer intents."},{"question":"Does it reduce calls to the service desk?","answer":"Bank of America says Erica for Employees, used by over 90% of its employees, has reduced calls into the IT service desk by more than half. Measure it as calls and tickets per employee before and after, not as chatbot conversations."},{"question":"Is it safe to let an AI reset passwords?","answer":"Only with strong verification. CISA and the FBI have warned that criminal groups phone IT help desks to get passwords and MFA tokens reset, so a reset should need the same or stronger proof of identity than a human analyst would ask for, with rate limits and security alerts on unusual patterns."}],"related":["hr-and-policy-assistant","aiops-incident-triage","employee-onboarding-assistant","enterprise-knowledge-search","support-knowledge-article-generation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog as an industry neutral page, with five evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the deployment years of IBM (2023), Mercari US (2021) and Vituity (2020), added the Equinix routing accuracy metric, sharpened the EU AI Act basis, removed APRA CPS 230, cited the CISA advisory on help desk social engineering, aligned the Blits.ai section with the feature inventory, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: removed the Equinix 82% routing share from the automation rate benchmark, rewrote the EU AI Act basis for Annex III point 4(b) on task allocation, removed the MAS AI risk management proposal from the regulations, described IBM AskIT as an assistant that mainly answers, moved the 7-Eleven Vietnam figure to productivity gain and documented its year, and clarified that single sign on is for platform administrators."}],"slug":"it-service-desk-resolution-agent","url":"https://www.blits.ai/ai-use-cases/it-service-desk-resolution-agent","benchmarks":[{"kpi":"employee-adoption","label":"Employee adoption","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":92,"min":90,"max":94,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"mercari-it-support-assistant","pooled":true},{"id":"bank-of-america-erica-for-employees","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":45,"min":40,"max":50,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"7-eleven-vietnam-it-support-chatbot","pooled":true},{"id":"vituity-it-and-hr-assistant","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":96,"min":96,"max":96,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"equinix-it-ticket-triage-assistant","pooled":true}]},{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":74,"min":74,"max":74,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"mercari-it-support-assistant","pooled":true}]},{"kpi":"contact-deflection","label":"Contact deflection","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-erica-for-employees","pooled":true}]},{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":75,"min":75,"max":75,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ibm-askit-service-desk-assistant","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":133000,"min":133000,"max":133000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ibm-askit-service-desk-assistant","pooled":true}]}],"indicativeValueResult":{"low":540000,"high":3000000},"evidence":["7-eleven-vietnam-it-support-chatbot","bank-of-america-erica-for-employees","equinix-it-ticket-triage-assistant","ibm-askit-service-desk-assistant","mercari-it-support-assistant","vituity-it-and-hr-assistant"]},{"title":"AI agent for network outage detection and customer communication","shortTitle":"Outage communication","seoTitle":"AI agents for telecom outage notifications","metaDescription":"An outage agent links network alarms to affected customers and tells them first. Comcast groups modem alarms, finds causes like power cuts and advises customers.","definition":"An AI agent that turns network alarms into a clear picture of which customers are affected by an outage and why, tells them proactively by message, app or phone with a cause and an estimated fix time, answers their questions during the incident, and updates them until service is restored.","aliases":["outage notification agent","proactive outage messaging","service disruption communication","outage IVR","incident customer updates"],"industries":["telecommunications"],"functions":["customer-service","network-operations","field-service"],"patterns":["anomaly-detection","classification-and-routing","content-generation","conversational-agent","voice-agent"],"channels":["sms","mobile-app","voice","whatsapp","web-chat","email"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"front-office","problem":"When a network fails, every affected customer tries to find out what is going on at the same\nmoment. Inbound calls spike, queues fill with people asking the same question, and customers with\nunrelated problems cannot get through. Agents can know little more than the customer when the\nnetwork operations centre, field teams and the contact centre use different systems and update\neach other slowly.\n\nAn affected customer needs three answers: that the operator knows, why it happened and when it\nwill be fixed. Telling them first can prevent some of these calls, but only if the operator can\nwork out quickly which customers are affected, separate a local power cut from a network fault,\nand keep the estimate honest as the repair progresses.","problemStats":[],"howItWorks":"1. **Group the alarms.** Models correlate modem, cell and network alarms by location and time\n   into one incident instead of thousands of individual alerts.\n2. **Find the cause and the affected customers.** The system estimates the likely cause (a fibre\n   cut, a site failure, a commercial power outage) and the list of affected customers and services.\n3. **Tell customers first.** The agent drafts and sends proactive notices in each customer's\n   channel and language, with the cause, what to do (for example contact the power company) and\n   an estimated fix time, using approved message templates.\n4. **Answer questions during the incident.** Inbound calls and chats from affected customers are\n   recognised at once and answered with the latest status, instead of joining the queue.\n5. **Update and close.** As field teams report progress, customers receive updates, and after\n   restoration a closing message and any compensation that applies.","valueDrivers":["customer-experience","cost-to-serve","speed","compliance"],"kpis":["interactions-handled","response-time-reduction","containment-rate","processing-time-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"An operator with 2 million fixed and mobile customers","inputs":[{"key":"customers","label":"Fixed and mobile customers","low":2000000,"high":2000000,"unit":"customers","note":"The reference operator."},{"key":"affectedShare","label":"Share of customers affected by at least one notable outage per year","low":0.2,"high":0.4,"unit":"fraction of customers","note":"Editorial assumption, replace with your own incident history."},{"key":"callsPerAffected","label":"Outage related calls per affected customer","low":0.15,"high":0.3,"unit":"calls per affected customer","note":"Editorial assumption, replace with call volumes from past incidents."},{"key":"avoidedShare","label":"Share of outage calls avoided or answered without a human","low":0.2,"high":0.4,"unit":"fraction of outage calls","note":"Editorial assumption. No public benchmark for outage calls was found. Kore.ai reports about 40% call containment in the first month for a US telecom provider's voice self service as a whole, not for outage calls, so the range stays at or below that.","sourceUrl":"https://www.kore.ai/customer-stories/large-u-s-telecommunications-provider"},{"key":"costPerCall","label":"Cost of a human handled call","low":4,"high":7,"unit":"USD per call","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"customers * affectedShare * callsPerAffected * avoidedShare * costPerCall","currency":"USD","period":"per year","resultLabel":"Outage call handling cost avoided","caveat":"Counts only avoided call handling. It leaves out faster restoration from better diagnosis and dispatch, lower compensation and complaints, the value of keeping lines free for other customers during incidents, and the cost of the AI and network data integration."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The conversation is the easy part. Linking network alarms to affected customers in minutes, and getting honest fix time estimates from field operations, needs integration across network management, inventory, field service and CRM, and a clear owner for every message sent.","dataPrerequisites":["Network topology and inventory that map customers to cells, nodes and fibre routes","Real time alarm and device status feeds","Field service job status and estimated repair times","Approved message templates per incident type and language","Contact preferences and consent for service messages"],"integrations":["Network monitoring and alarm correlation","Network inventory and customer to network mapping","Field service management","Messaging platforms (SMS, app push, WhatsApp, email)","IVR and contact centre routing for inbound calls"]},"implementation":{"steps":[{"title":"Map customers to the network","detail":"Without a reliable link from customers to the network elements that serve them, every notice is a guess. Fix inventory and mapping first."},{"title":"Agree who owns the message","detail":"Decide with network operations and communications who approves the cause, the estimate and the wording for each incident class, and which notices the agent may send on its own."},{"title":"Start with inbound recognition","detail":"Recognising callers in an affected area and giving them the status is low risk and high value. Add proactive notices once the affected customer lists prove accurate."},{"title":"Keep estimates honest","detail":"Take fix times from field service, give ranges rather than exact times and send an update whenever the estimate changes, even if it gets worse."},{"title":"Rehearse major incidents","detail":"Test the whole chain on simulated large outages, including emergency call disruption, so messaging scales and escalation to crisis teams works."}],"guardrails":["Proactive notices only from approved templates, with human approval for major incidents","Estimates always come from field service data, never generated by the model","Messages about disruption to emergency calling follow the crisis communication plan","Vulnerable and priority customers, such as telecare users, get direct contact","Service messages respect contact preferences and never include marketing"],"humanInTheLoop":"Network operations confirm the cause and scope of each incident, and communications or incident managers approve notices for major and sensitive outages. People handle priority customers and complaints, and review every major incident's customer communication afterwards.","kpisToInstrument":["Time from first alarm to first customer notice","Share of affected customers notified before they contacted the operator","Inbound contacts per affected customer during incidents","Accuracy of affected customer lists and fix time estimates","Satisfaction and complaints after major incidents"],"failureModes":[{"title":"Wrong customers notified","detail":"Notices reach unaffected customers or miss affected ones. Validate customer to network mapping and measure list accuracy."},{"title":"Optimistic estimates","detail":"A fix time that keeps slipping destroys trust. Use ranges from field data and update proactively."},{"title":"Silence during the big one","detail":"Messaging fails under the load of a major outage. Load test and rehearse."},{"title":"Treating a power cut as a network fault","detail":"Engineers are sent to working equipment. Detect power outages and tell customers what they can do."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The customer facing agent is limited risk with an Article 50 duty to disclose AI. AI used as a safety component in the management and operation of critical digital infrastructure is high risk under Annex III point 2, so the classification depends on whether the detection part acts on the network or only informs people."},"regulations":["eu-ai-act","gdpr","telecom-consumer-rules","eecc","nis2"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Investigation into Three following voice service outage on 25 June 2025","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/phones-and-broadband/telecoms-infrastructure/investigation-into-three-following-voice-service-outage-on-25-june-2025","note":"Ofcom opened this investigation on 15 December 2025 after a UK wide voice outage that also disrupted calls to emergency services. It shows that sections 105A and 105C of the Communications Act 2003 require providers to prepare for and mitigate failures of network availability, which outage communication supports but does not replace."},{"title":"European Electronic Communications Code (Directive (EU) 2018/1972)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2018/1972/oj","note":"Sets the EU end user rules for electronic communications services, such as contract information and compensation. Its security articles 40 and 41 were deleted from 18 October 2024 and replaced by NIS2."},{"title":"NIS2 Directive, Article 23 (reporting obligations)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2022/2555/oj","note":"Where appropriate, essential and important entities, including providers of public electronic communications networks and services, must notify the recipients of their services without undue delay of significant incidents likely to affect those services."}],"controls":["Documented message templates and approval rules per incident class","Audit log of every notice, with the data it was based on","Load tests and major incident rehearsals at least yearly","Accuracy review of affected customer lists after each major incident","Priority customer register kept current and used in every incident"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the detection usually stays in the operator's network tools; Blits.ai runs the\ncommunication. **Agentic workflows** are triggered through the API by an incident, fetch the\naffected customers and status through **custom functions**, and draft notices from approved\ntemplates. After **human in the loop approval** for major incidents, custom functions hand the\nnotices to the operator's own SMS, push or email gateway. **Agentic\ntasks** recheck the repair status on a schedule and send updates when it changes.\n\nInbound, the same **AI agent** answers on **voice, SMS, WhatsApp, web chat and email**, and in\nthe operator's own app through the **API channel**, recognising callers from an affected area\nand giving them the latest status. **Human handover** routes priority customers and complaints to people, **monitors** check the incident\njourney on a schedule, **test suites** replay incident scenarios, and the platform is model\nagnostic with EU and UAE data residency."},"faq":[{"question":"Can AI tell customers about an outage before they call?","answer":"Yes, if the operator can link alarms to customers. Comcast describes an AI tool, deployed nationwide, that groups modem alarms into one alert and identifies causes such as commercial power outages, so Comcast can notify affected customers with advice. Kore.ai says a large US telecom provider sends outbound outage SMS to reduce repeat calls. Published figures on calls avoided are still rare."},{"question":"Who should approve outage messages?","answer":"Routine local notices can go out automatically from approved templates. Major incidents, especially those affecting emergency calls, need approval from incident or communications managers, because the message is part of the operator's regulatory response."},{"question":"How does this differ from AIOps incident triage?","answer":"AIOps helps engineers find and fix the fault. Outage communication uses the same incident data to keep customers informed and to take pressure off the contact centre while the fix is under way."}],"related":["network-fault-triage-copilot","device-and-connectivity-troubleshooting-agent","outbound-reminder-and-confirmation-agent","first-line-contact-centre-agent","flight-disruption-and-rebooking-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by Len Debets's editor (targeted blocker check)."},{"date":"2026-09-25","note":"First version, written for the telecommunications vertical from Comcast and an anonymous Kore.ai source. Public evidence is thin and has no call reduction figures yet."},{"date":"2026-09-25","note":"Consolidation pass: added European Electronic Communications Code, NIS2 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: dated the Comcast release and corrected its storm recovery claim, added outbound SMS, Deepgram and the 2025 year to the Kore.ai record, grounded the avoided call assumption in the Kore.ai containment figure, removed an unsupported claim that operators avoid most calls, corrected the EECC note (its security articles moved to NIS2), added NIS2 Article 23 guidance, sharpened the Ofcom note, and added an SEO title and meta description."},{"date":"2026-09-26","note":"Second fact check against sources: Comcast, Kore.ai, Ofcom and NIS2 texts reconfirmed; softened three unsourced generalizations in the problem section (what agents know, what customers want, how many calls proactive notices prevent)."},{"date":"2026-09-27","note":"Corrected the Comcast claim: its AI tool groups alarms and finds the cause so Comcast can notify customers; the release does not say the AI sends the notices. Added the Kore.ai outbound outage SMS to the FAQ, the Wayback copy of the Comcast release, and EUR-Lex links for the AI Act and NIS2 guidance, and clarified how notices are sent on Blits.ai."},{"date":"2026-09-27","note":"Fact checked against sources: Comcast release (Wayback copy), Kore.ai story, Ofcom investigation, EUR-Lex texts of the AI Act, NIS2 and the EECC deletion rule, and the Blits.ai feature inventory all reconfirmed; no changes needed."}],"slug":"network-outage-communication-agent","url":"https://www.blits.ai/ai-use-cases/network-outage-communication-agent","benchmarks":[],"indicativeValueResult":{"low":48000,"high":672000},"evidence":["comcast-ai-outage-detection-and-notification"]},{"title":"AI agent for order status, delivery changes and returns","shortTitle":"Order status and returns","seoTitle":"AI agents for order status and returns","metaDescription":"An AI agent answers where is my order, changes deliveries and arranges returns within policy. BARK, Next and Klarna run one; see the results and the playbook.","definition":"An AI agent that answers \"where is my order\", changes delivery details and arranges returns, exchanges and refunds end to end for online and omnichannel shoppers, by reading and writing to the order, carrier and returns systems within the retailer's policy, and hands exceptions such as damaged goods, disputes and upset customers to a person.","aliases":["WISMO agent","where is my order chatbot","returns and refunds assistant","ecommerce customer service agent","post purchase support agent"],"industries":["cross-industry","retail-and-ecommerce","payments"],"functions":["customer-service","operations"],"patterns":["conversational-agent","agentic-workflow","voice-agent","rag-knowledge-assistant"],"channels":["web-chat","mobile-app","whatsapp","voice","email"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"After the checkout, the questions start: where is my order, can I change the delivery date, how do\nI return this, when do I get my money back. BARK, which ships tens of thousands of boxes a month,\ndescribes these questions as quick to answer one by one but overwhelming together. They peak\ntogether (a sale, the holidays, a carrier delay) and the answer sits in three or four systems: the\norder management system, the carrier's tracking feed, the warehouse and the returns platform.\n\nReturns are also expensive and emotional. A clumsy returns experience costs the next sale, and\npolicy answers that are wrong (a refund promised outside the return window, a label for an item\nthat cannot be returned) create cost and disputes. First generation chatbots pointed customers to\na tracking link or a returns page. The step change is an agent that identifies the order, reads the\nlive status, and completes the change or the return itself, inside the retailer's rules.","problemStats":[{"statement":"The National Retail Federation and Happy Returns projected that total retail returns in the United States would reach USD 890 billion in 2024, with retailers expecting 16.9% of annual sales to be returned.","sourceTitle":"NRF and Happy Returns Report: 2024 Retail Returns to Total $890 Billion","sourceUrl":"https://nrf.com/media-center/press-releases/nrf-and-happy-returns-report-2024-retail-returns-total-890-billion","year":2024},{"statement":"In the same NRF and Happy Returns survey, 67% of consumers said a negative return experience would discourage them from shopping with a retailer again.","sourceTitle":"NRF and Happy Returns Report: 2024 Retail Returns to Total $890 Billion","sourceUrl":"https://nrf.com/media-center/press-releases/nrf-and-happy-returns-report-2024-retail-returns-total-890-billion","year":2024}],"howItWorks":"1. **Identify the customer and the order.** The agent matches the customer to the order from the\n   logged in session, an email or phone number and a verification step, so the customer does not\n   have to find an order number.\n2. **Read the live status.** It combines the order management system, the warehouse status and\n   the carrier's tracking events, and explains them in plain words: packed, handed to the carrier,\n   delayed at the depot, out for delivery.\n3. **Change what can be changed.** Within set rules it updates the delivery address or slot,\n   cancels an order that has not shipped, or reschedules a delivery through the carrier's API.\n4. **Run the return.** It checks eligibility against the return policy (window, item type,\n   condition), offers the options the retailer allows (refund, exchange, store credit, drop off or\n   pickup), creates the return and sends the label or QR code.\n5. **Explain the refund.** It tells the customer when and how the money comes back, from the\n   payment and refund status, not from a guess.\n6. **Hand over exceptions.** Damaged or missing items above a value threshold, suspected fraud,\n   complaints, repeat failures and emotional conversations go to a person with the order and the\n   conversation attached.","valueDrivers":["cost-to-serve","customer-experience","speed","revenue-growth"],"kpis":["containment-rate","customer-satisfaction","interactions-handled","response-time-reduction","handling-time-reduction","cost-savings"],"indicativeValue":{"referenceOrg":"An online retailer that ships 5 million orders a year","inputs":[{"key":"orders","label":"Orders shipped per year","low":5000000,"high":5000000,"unit":"orders per year","note":"The reference retailer."},{"key":"contactsPerOrder","label":"Post purchase contacts per order","low":0.1,"high":0.2,"unit":"contacts per order","note":"Editorial assumption for order status, delivery and returns contacts. Replace with your own contact reason data."},{"key":"resolvedShare","label":"Share of those contacts the agent resolves end to end","low":0.3,"high":0.5,"unit":"fraction of contacts","note":"Editorial assumption within the range of the evidence. BARK's agent handled roughly a quarter of all customer conversations in its first year, Klarna's assistant two thirds of its service chats in its first month, and Ingka Group reports that its Billie chatbot resolved about 47% of all enquiries it received from 2021 to 2023. End to end resolution needs write access to orders and returns.","sourceUrl":"https://www.ingka.com/newsroom/ai-and-remote-selling-bring-ikea-design-expertise-to-the-many/"},{"key":"costPerContact","label":"Cost of a human handled contact","low":3,"high":6,"unit":"USD per contact","note":"Editorial assumption for a blended chat, email and phone contact. Replace with your own fully loaded cost."}],"formula":"orders * contactsPerOrder * resolvedShare * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of running the AI and the integrations, the effect on repeat purchase of a better returns experience, fewer refund errors, and any change in return rates."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Reading the order status is easy; acting safely is the work. The agent needs reliable APIs into order management, carriers and the returns platform, identity matching without an order number, and policy rules that are encoded rather than paraphrased.","dataPrerequisites":["Order, shipment and return data reachable through APIs, with carrier tracking events normalised","The return and refund policy as explicit rules (windows, excluded items, conditions, exceptions)","Contact reason data per intent to choose the first scope","Approved content for delivery, returns and warranty questions"],"integrations":["Order management system and ecommerce platform","Carrier tracking and delivery management APIs","Returns platform or warehouse returns process, including label generation","Payment service provider for refund status","Contact centre or helpdesk for handover with context"]},"implementation":{"steps":[{"title":"Start with the contact reason report","detail":"Order status, delivery changes and return requests are often a large share. Pick the intents with the highest volume and the clearest rules, and leave damaged goods and disputes with people at first."},{"title":"Encode the policy, do not paraphrase it","detail":"Turn the return window, excluded categories, condition rules and refund methods into rules the agent's tools enforce. The model explains the outcome; the rule decides it."},{"title":"Solve order matching","detail":"Most customers do not have the order number at hand. Match on the logged in session, email or phone with a one time code, and confirm the item before acting."},{"title":"Give the agent a narrow set of actions","detail":"Address or slot change before dispatch, cancellation before dispatch, return creation, label sending and refund status. Each action has its own limits, such as a maximum order value for an automatic refund."},{"title":"Plan for peaks and carrier incidents","detail":"When a carrier has a regional delay, publish one explanation the agent uses for every affected order instead of letting it improvise, and scale the channel before the peak."},{"title":"Measure resolution, not deflection","detail":"Count a conversation as resolved only if the customer did not come back on the same order within seven days, and read transcripts of the ones that did."}],"guardrails":["Refund and exchange decisions come from policy rules in the tools, never from the model's own reading","Value thresholds above which refunds, reshipments or goodwill gestures need a person","Identity verification before any change to address, delivery or refund destination","The agent never promises a delivery date the carrier has not given","Automatic handover for complaints, suspected fraud, damaged goods above a threshold and distress"],"humanInTheLoop":"People own exceptions and judgment calls: damaged or missing items above a threshold, goodwill gestures, suspected return fraud and complaints. A team lead reviews a weekly sample of resolved conversations and every policy answer that led to a refund dispute, and signs off each new action before it goes live.","kpisToInstrument":["Resolution rate per intent, counting repeat contacts on the same order within seven days as unresolved","Share of returns created end to end by the agent and their error rate","Satisfaction on agent conversations versus human conversations for the same intents","Average response time and time to refund","Refund disputes and complaints that mention the assistant"],"failureModes":[{"title":"Promising what policy does not allow","detail":"A fluent answer that grants a refund outside the window or for an excluded item. Keep eligibility in deterministic rules and have the agent quote the rule it applied."},{"title":"Tracking data the agent cannot interpret","detail":"Carrier events are cryptic and sometimes wrong. Normalise them into a small set of states and let the agent say \"we do not know yet\" rather than guess."},{"title":"Return fraud through an easy channel","detail":"An agent that issues refunds without checks becomes a target. Use value limits, customer history signals and a person for high value or repeat claims."},{"title":"Deflection dressed up as resolution","detail":"Customers who give up look like contained conversations. Track repeat contacts and satisfaction per intent."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing service agent must disclose that the customer is interacting with AI (Article 50). It is not high risk: it does not decide on access to essential services, credit or employment."},"regulations":["eu-ai-act","gdpr","pci-dss","eu-accessibility-act"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Consumer rights directive","issuer":"European Commission","region":"europe","url":"https://commission.europa.eu/law/law-topic/consumer-protection-law/consumer-contract-law/consumer-rights-directive_en","note":"Sets the EU right of withdrawal for distance purchases and the refund rules an agent's answers on returns must respect."}],"controls":["AI disclosure at the start of each conversation","Return and refund rules versioned with an owner, and regression tests on every policy change","Audit log of every order change, return and refund the agent initiates","Masking of payment data and personal data in logs and model prompts","Monitoring of refund value and return volume initiated through the agent, with alerts on outliers"],"incidents":[{"title":"Incident 639: Air Canada chatbot reportedly provides inaccurate bereavement fare information, leading to customer overpayment","url":"https://incidentdatabase.ai/cite/639/","note":"Air Canada's website chatbot gave a customer inaccurate information about bereavement fare refunds. A Canadian small claims tribunal held the airline responsible for what its chatbot said and ordered it to pay damages. The same risk applies to any agent that answers return and refund questions without the policy encoded as rules."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the retailer's order\nmanagement, carrier tracking and returns APIs (REST calls with their own authentication and\nlimits), and a **knowledge base** with the approved delivery, returns and warranty content,\nretrieved with hybrid search. Return eligibility and refund rules run as deterministic steps in a\n**flow** or inside the function, so the model explains the decision rather than making it; larger\nrefunds can go through an **agentic workflow with human in the loop approval** above a set\nthreshold.\n\nThe same agent serves **web chat, WhatsApp, email and voice**, and a mobile app through the REST or\nWebSocket API channel; the chat widget can show orders and receipts as rich cards. **Guardrails**\ncheck input and output, **PII masking** and card number tokenization happen at the gateway, and\n**human handover** escalates the conversation to the service team, including through Salesforce,\nFreshdesk or Zoho SalesIQ. **Test suites** run multi turn order and return conversations as\nregression tests after each change, and **monitors** check the order lookup against a test order\non a schedule. The platform is model agnostic, so the model can be chosen or switched per agent."},"faq":[{"question":"What share of order and returns contacts can an AI agent resolve?","answer":"It depends on whether the agent can act. Agents that only link to a tracking page resolve little; agents that can read the live status and create returns can do much more. On this page, BARK's agent handled roughly a quarter of all customer conversations in its first year, and Klarna's assistant, which handles refunds, returns and disputes, took two thirds of its service chats in its first month (in 2025 Klarna began recruiting human agents again)."},{"question":"Should the AI decide refunds?","answer":"No. Put eligibility and refund rules in deterministic tools with value limits, and let the agent explain the outcome. A Canadian tribunal held Air Canada responsible for what its website chatbot told a customer about bereavement fare refunds, so a refund answer must come from the policy itself."},{"question":"Does a better returns agent hurt sales?","answer":"No source on this page measures the sales effect of a returns agent specifically. What is documented: in the NRF and Happy Returns survey, 67% of consumers said a negative return experience would discourage them from shopping with a retailer again, a measure of stated intent rather than an AI agent's effect. Separately, Sun & Ski Sports extended its agent from returns and order status into product advice, and its vendor reports that shoppers who engage with it convert at three times the rate of those who do not, a self selected comparison, not a controlled measure of the returns agent's effect on sales."}],"related":["parcel-tracking-and-delivery-exception-agent","conversational-shopping-assistant","chargeback-and-representment","complaints-handling-agent","first-line-contact-centre-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, cross industry customer facing vertical; evidence from Sun & Ski Sports, BARK, Next, Best Buy and Klarna."},{"date":"2026-09-25","note":"Consolidation pass: added European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: problem statement now cites BARK instead of an unsourced claim; Ingka Group figure cited with its source; FAQ uses Klarna instead of evidence not on the page; Air Canada incident described as reported; Blits.ai build limited to inventory capabilities; Sun & Ski Sports and Next records corrected; SEO title and description added."}],"slug":"order-status-and-returns-agent","url":"https://www.blits.ai/ai-use-cases/order-status-and-returns-agent","benchmarks":[{"kpi":"customer-satisfaction","label":"Customer satisfaction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":94,"min":90,"max":98,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"bark-scout-order-tracking-agent","pooled":true},{"id":"sun-and-ski-sports-sunny-ai-agent","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2300000,"min":2300000,"max":2300000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"klarna-ai-assistant-customer-service","pooled":true}]},{"kpi":"response-time-reduction","label":"Response time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":82,"min":82,"max":82,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"klarna-ai-assistant-customer-service","pooled":true}]}],"indicativeValueResult":{"low":450000,"high":3000000},"evidence":["bark-scout-order-tracking-agent","best-buy-appointment-scheduling-agent","klarna-ai-assistant-customer-service","next-returns-ai-agent","sun-and-ski-sports-sunny-ai-agent"]},{"title":"AI agent for outbound reminders and confirmations by voice and messaging","shortTitle":"Outbound reminders and confirmations","seoTitle":"AI appointment reminder and confirmation agents","metaDescription":"AI reminder agents let customers confirm, cancel or move a booking by text or phone. An NHS pilot linked AI texts to more attended appointments.","definition":"An AI agent that contacts customers about something they already booked or ordered (an appointment, a delivery, a reservation or a service visit) to remind them, confirm attendance and let them cancel or move it in the same conversation, by phone, SMS, WhatsApp or email. It is operational service outreach, not marketing: nothing is sold, and success is measured in kept appointments and reused slots, not in conversion.","aliases":["AI appointment reminder calls","automated confirmation calls","two way reminder messaging","service visit reminder agent","reservation confirmation agent"],"industries":["cross-industry","healthcare","government"],"functions":["customer-service","operations"],"patterns":["voice-agent","conversational-agent","agentic-workflow","prediction-and-scoring"],"channels":["voice","sms","whatsapp","email"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Every business that books time with customers loses some of it. Patients miss appointments,\nengineers arrive at empty houses, delivery drivers carry parcels back to the depot and restaurant\ntables stay empty. The cost is not only the lost slot: it is the other customer who could have had\nit, and the repeat visit that now has to be arranged.\n\nMany organizations already send reminders, but classic reminders are one way broadcasts at a fixed\nmoment. They cannot answer \"can I come an hour later\", they do not know which customers are likely\nto miss, and a cancellation made the evening before usually leaves the slot empty. Human\nconfirmation calls allow a real conversation but are costly to make for every booking.","problemStats":[{"statement":"The UK government says sending reminders has been shown to reduce missed appointments by up to 80%, and that NHS trusts report better results when communication with the patient is two way.","sourceTitle":"Power to patients as government sets out plan to cut waiting lists","sourceUrl":"https://www.gov.uk/government/news/power-to-patient-as-government-sets-out-plan-to-cut-waiting-lists","year":2025},{"statement":"NHS England reports that eight million (6.4%) of 124.5 million outpatient appointments in England in the previous year were not attended.","sourceTitle":"NHS AI expansion to help tackle missed appointments and improve waiting times","sourceUrl":"https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/","year":2024}],"howItWorks":"1. **The booking system triggers the contact.** Reminders and confirmations start from an existing\n   booking, order or reservation, at times set per service. The agent does not choose who to\n   contact for commercial reasons.\n2. **Prioritise where it matters.** A risk score can decide who gets an extra reminder, a call\n   instead of a text, or an offer of help, as Sheffield Children's did in its AI Predictor pilot.\n3. **Open with who and why.** The agent names the organization, says it is an automated assistant\n   and states the booking it is about, without revealing sensitive details before the person is\n   verified.\n4. **Let the customer act.** The customer can confirm, cancel, move to another offered slot or ask a\n   practical question (address, parking, preparation, delivery window) in the same conversation.\n5. **Write back and reuse.** Confirmations, cancellations and new times go straight back to the\n   booking system, and freed slots are offered to the next customer on the waiting list.\n6. **Escalate what is not routine.** Complaints, distress, clinical questions and customers who ask\n   for a person are handed to staff with the conversation attached.","valueDrivers":["cost-to-serve","customer-experience","speed","inclusion-and-access"],"kpis":["interactions-handled","containment-rate","customer-satisfaction","cost-reduction"],"indicativeValue":{"referenceOrg":"A service organization with 200,000 booked appointments or visits a year","inputs":[{"key":"bookings","label":"Booked appointments or visits per year","low":200000,"high":200000,"unit":"bookings per year","note":"The reference organization."},{"key":"missedRate","label":"Share of bookings missed without notice today","low":0.05,"high":0.1,"unit":"fraction of bookings","note":"Editorial assumption. NHS England reports 6.4% for outpatient appointments; replace with your own rate."},{"key":"reduction","label":"Share of missed bookings avoided or refilled because of the agent","low":0.1,"high":0.2,"unit":"fraction of missed bookings","note":"Editorial assumption, set at or below the NHS pilots on this page. Sheffield Children's expected 8,581 missed appointments against a benchmark rate, not a control group, and recorded just under 6,500, about a 24% reduction. UHCW's move from 10% to 4% came from reminder timing found through process mining, not from an agent. Measure against a control group."},{"key":"valuePerBooking","label":"Value of a booking that is kept or refilled","low":40,"high":120,"unit":"EUR per booking","note":"Editorial assumption covering the margin or cost of a wasted slot or visit. Replace with your own figure."}],"formula":"bookings * missedRate * reduction * valuePerBooking","currency":"EUR","period":"per year","resultLabel":"Value of bookings kept or refilled","caveat":"Counts kept and refilled bookings only. It leaves out staff time saved on manual confirmation calls, the cost of messages, calls and the AI, and any annoyance from contact that customers did not want."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The dialogue is short. The effort goes into a clean trigger from the booking system, a reliable write back of confirmations and changes, contact rules per channel and market, and verifying the person before details are shared.","dataPrerequisites":["Bookings with time, location, service type and contact details","Available alternative slots and a waiting list, if freed slots are to be reused","Channel preferences, consent and opt outs per customer","Historic attendance data if contacts are prioritised by risk"],"integrations":["Booking, scheduling or order management system (read and write)","Telephony and messaging (voice, SMS, WhatsApp, email)","Waiting list or capacity planning","CRM or case management for follow up","Contact centre for handover"]},"implementation":{"steps":[{"title":"Start with the reminders you already send","detail":"Make the existing reminder two way before adding new ones: let the customer confirm, cancel or move it by replying. This alone turns late cancellations into slots you can reuse."},{"title":"Test the timing","detail":"Timing matters as much as wording. University Hospitals Coventry and Warwickshire found a spike in last minute cancellations after two text reminders, and found that reminders 14 days and four days ahead worked best because patients cancelled early enough for the slot to be rebooked."},{"title":"Add risk based escalation","detail":"Use a missed booking score to decide who gets an extra message, a call or practical help, and keep a control group so you can see the effect."},{"title":"Connect cancellations to the waiting list","detail":"A reminder that produces a cancellation is only valuable if someone else gets the slot. Automate the offer to the next suitable customer."},{"title":"Keep marketing out","detail":"Do not add offers or upsell to service reminders. It changes the legal basis for the contact and the trust customers place in the channel."}],"guardrails":["Contact only about an existing booking, order or reservation, never for sales","Disclose that the caller or sender is an automated assistant and name the organization","Verify the person before sharing sensitive details such as a clinic or diagnosis","Honour opt outs and quiet hours on every channel","Never ask for passwords, card details or payment in a reminder"],"humanInTheLoop":"Service owners approve every reminder script, timing and channel. Staff take over complaints, distress, clinical or safety questions and anyone who asks for a person, and review a sample of conversations and outcomes each month, with attention to groups that miss the most bookings.","kpisToInstrument":["Missed booking rate against a control group","Share of cancellations made early enough to refill, and refill rate","Confirmation, cancellation and rebooking rates per reminder","Opt out and complaint rate per reminder type","Reach and outcomes by language, age and deprivation band"],"failureModes":[{"title":"Reminders that look like scams","detail":"Unexpected calls and texts are what fraudsters imitate. Name the organization, never ask for secrets, and point to known channels."},{"title":"Late cancellations with no refill","detail":"The reminder works but the slot stays empty. Time reminders so cancellations come early and connect them to a waiting list."},{"title":"Privacy leaks in the message","detail":"A reminder names a sensitive clinic or service to whoever reads the phone. Keep content minimal until the person is verified."},{"title":"Service outreach turning into marketing","detail":"Offers creep into reminders and the contact now needs marketing consent. Keep scripts service only and review them."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"People must be told they are interacting with an AI system, and synthetic voice or text must be identifiable as such (Article 50). Reminding people of existing bookings and disclosure alone are limited risk. A missed appointment score used by or for a public authority to grant, reduce, revoke or reclaim access to healthcare or other essential public assistance and services, for example deciding who is offered funded transport, can fall within Annex III point 5(a), and profiling of natural persons within Annex III rules out the Article 6(3) exemption. Using the score only to decide who gets extra reminders or support does not by itself place it outside Annex III when that support is itself the assistance being granted."},"regulations":["eu-ai-act","gdpr","uk-gdpr","hipaa","us-tcpa"],"guidance":[{"title":"Declaratory ruling on AI generated voices under the TCPA (FCC 24-17)","issuer":"Federal Communications Commission","region":"north-america","url":"https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf","note":"AI generated voices count as artificial or prerecorded voice under the TCPA, so US reminder calls made with AI need the consent the TCPA requires for such calls."},{"title":"Guide to PECR: electronic and telephone marketing","issuer":"Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/guide-to-pecr/electronic-and-telephone-marketing/","note":"UK rules on marketing by phone, text and email. Routine customer service messages about a current contract or past purchase, such as delivery arrangements, do not count as direct marketing; unsolicited marketing often needs specific consent."}],"controls":["Documented purpose and legal basis for each reminder type","Opt out, quiet hour and frequency checks logged per contact","Approved scripts per reminder type with an accountable owner","Audit trail of every contact and every change written back to the booking system","Monitoring of complaints, opt outs and outcomes by group"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai each reminder type is an **agentic workflow** started on a schedule or by the booking\nsystem through an API token, with the booking passed in as data. The workflow sends the reminder\nthrough the outbound **email** channel or a **custom function** that calls the messaging provider,\nand an **agentic task** can recheck later and follow up when a booking is still unconfirmed.\nReplies reach an **AI agent** on the **SMS**, **WhatsApp** or **email** channel, where **custom\nfunctions** confirm, cancel or move the booking and offer freed slots to the waiting list.\nCustomers who call back reach the same agent on the **voice** channel.\n\nA **flow** fixes the opening (who is contacting, why, and how to opt out) and any verification step\nbefore details are shared, and the **GDPR toolkit** handles consent and data removal. **Guardrails**\nkeep offers and payment requests out of reminders, **human handover** routes complaints and\nclinical questions to staff, and **analytics** with the workflow run history show confirmations,\ncancellations and refills per reminder type."},"faq":[{"question":"How is this different from proactive outreach and activation?","answer":"Proactive outreach contacts customers about something they have not done yet, such as activating a card or accepting an offer, and often needs marketing consent. Reminder and confirmation outreach is about something the customer already booked or ordered and exists to make it happen as planned."},{"question":"Do AI targeted reminders reduce no shows?","answer":"NHS pilots suggest so. In the first 12 months of a pilot at Sheffield Children's, an AI Predictor sent 53,800 extra text reminders to families at high risk of missing appointments, and NHS England reports almost 200 more attended appointments a month against the benchmark. Results without a control group should be read with care."},{"question":"Should reminders be calls or messages?","answer":"Messages for most people, because they are cheap and let the customer reply when it suits them. Calls suit high risk bookings and people who do not use messaging. WellSpan Health first used its AI voice agent to call patients about colorectal cancer screening, and has announced outreach to patients who missed imaging appointments as a next workflow."},{"question":"Are AI reminder calls legal?","answer":"It depends on the market and the purpose. In the US, AI generated voices fall under the TCPA's rules for artificial voices; in the UK, PECR does not count routine customer service messages as direct marketing. Keep reminders strictly about the booking and record the legal basis."}],"related":["patient-appointment-scheduling-and-reminders-agent","branch-and-appointment-booking-agent","proactive-outbound-engagement-agent","parcel-tracking-and-delivery-exception-agent","network-outage-communication-agent","utility-billing-and-move-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with evidence from Sheffield Children's, University Hospitals Coventry and Warwickshire, WellSpan Health, the US Department of Veterans Affairs and an anonymous healow deployment."},{"date":"2026-09-25","note":"Consolidation pass: added Telephone Consumer Protection Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the WellSpan FAQ (missed imaging outreach is announced, not live) and the UHCW timing step; replaced the ICO guidance link with the page on service messages; added UK GDPR and the Annex III point 5(a) boundary to the risk section; limited the Blits.ai build to listed capabilities; added SEO title and description."}],"slug":"outbound-reminder-and-confirmation-agent","url":"https://www.blits.ai/ai-use-cases/outbound-reminder-and-confirmation-agent","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":160000,"min":160000,"max":160000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"wellspan-hippocratic-ai-patient-voice-agent","pooled":true}]}],"indicativeValueResult":{"low":40000,"high":480000},"evidence":["sheffield-childrens-ai-attendance-predictor","uhcw-process-mining-appointment-reminders","va-evirtual-assistant-appointment-reminders","wellspan-hippocratic-ai-patient-voice-agent"]},{"title":"AI agent for outbound sales prospecting and personalized outreach","shortTitle":"Outbound sales prospecting","seoTitle":"AI agents for outbound sales prospecting","metaDescription":"AI agents research target accounts and draft outreach that reps approve. Clay reports Merge SDRs got 10+ hours back a week and 20% higher response rates.","definition":"An AI agent that researches target accounts and contacts, drafts personalized outbound outreach (emails, LinkedIn messages and call scripts) from the campaign, the prospect's context and the sales goals, and sequences the follow ups, with a sales development rep approving or sending every message.","aliases":["AI sales prospecting assistant","outbound prospecting agent","personalized cold outreach generator","account research agent for sellers","AI outbound sequencing"],"industries":["cross-industry","technology","professional-services"],"functions":["sales","marketing"],"patterns":["content-generation","agentic-workflow","recommendation-and-personalization","prediction-and-scoring"],"channels":["email","internal-tools","api"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Outbound prospecting is mostly research and writing. Before a sales development rep sends a\nfirst message, they have to decide which accounts in a large territory deserve attention this\nweek, find the right people, read the company's news, filings and job posts, and work out why the\nproduct matters to this account now. Then they write the message, and the follow ups, and adapt\nit for email, LinkedIn and a call. Lumen's chief revenue officer says research for customer\noutreach typically takes a seller four hours, so reps either cover few accounts well or many\naccounts with generic templates that buyers ignore.\n\nTemplate automation made the problem worse: sequencing tools make it easy to send thousands of\nnear identical emails, and that volume puts reply rates, sender reputation and complaint levels\nat risk. The\nalternative is an agent that does the research and the first draft per account, grounded in real\nsignals and the organization's own positioning, while the rep keeps judgment over who to contact,\nwhat to say and when to stop. Unlike inbound qualification, the prospect has not asked to be\ncontacted, so consent, privacy and anti spam rules shape the design from the start.","problemStats":[{"statement":"Lumen Technologies' chief revenue officer says it typically takes a seller four hours to do research for customer outreach.","sourceTitle":"Lumen's strategic leap: How Copilot is redefining productivity and employee engagement","sourceUrl":"https://customers.microsoft.com/en-us/story/1771760434465986810-lumen-microsoft-copilot-telecommunications-en-united-states","year":2024}],"howItWorks":"1. **Pick the accounts.** The agent scores the target account list against the ideal customer\n   profile and live signals (funding, hiring, product launches, leadership changes, website\n   visits) and proposes which accounts each rep should work this week, with the reason.\n2. **Research the account and the people.** For each account it gathers public and first party\n   context (news, filings, job posts, technology used, past CRM activity and calls) and writes a\n   short brief with the likely need, the relevant product and the right contacts.\n3. **Draft the outreach.** From the brief, the campaign and the approved positioning it drafts a\n   first email, a LinkedIn message and a call opener, each citing the signal it is based on,\n   within brand, claims and tone rules.\n4. **Check the rules before anything leaves.** Contact source, consent or legitimate interest\n   basis, suppression and opt out lists, country rules and quiet hours are checked for every\n   contact, and anything that fails is dropped.\n5. **Rep approves and sends.** The rep edits, approves or rejects each draft; approved messages go\n   out from the rep's own mailbox or sequencing tool, never as anonymous bulk mail.\n6. **Sequence and learn.** The agent proposes follow ups based on replies and new signals, stops\n   the sequence on any reply or opt out, logs everything in the CRM and feeds reply and meeting\n   rates back into account scoring and message variants.","valueDrivers":["revenue-growth","employee-productivity","speed"],"kpis":["conversion-rate-uplift","productivity-gain","hours-saved","cost-reduction","revenue-uplift"],"indicativeValue":{"referenceOrg":"A B2B software company with 20 sales development reps","inputs":[{"key":"reps","label":"Sales development reps","low":20,"high":20,"unit":"reps","note":"The reference company."},{"key":"researchHours","label":"Hours per rep per week on account research and writing outreach","low":8,"high":15,"unit":"hours per rep per week","note":"Editorial assumption. Lumen's chief revenue officer puts research for customer outreach at about four hours a week per seller; this range adds writing first messages and follow ups. Replace with a time study of your own reps."},{"key":"shareSaved","label":"Share of that time the agent saves","low":0.3,"high":0.6,"unit":"fraction of research and writing time","note":"Conservative against the evidence on this page (Clay reports that Merge's SDRs got 10+ hours back every week; Clay reports an estimate by the person who built Oyster's workflows of about 40 hours per rep per month; Lumen's chief revenue officer says research that took four hours now takes 15 minutes), because vendor case studies select their best results and reps still review every draft."},{"key":"weeks","label":"Working weeks per year","low":44,"high":46,"unit":"weeks","note":"Editorial assumption after holidays and training."},{"key":"hourlyCost","label":"Fully loaded cost per SDR hour","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption. Replace with your own fully loaded SDR cost."}],"formula":"reps * researchHours * shareSaved * weeks * hourlyCost","currency":"USD","period":"per year","resultLabel":"SDR capacity released from research and drafting","caveat":"Capacity, not cash: the value is only real if reps spend the time on more or better outreach. It leaves out the cost of the agent and data providers, any change in reply or meeting rates, and the revenue those meetings produce."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Drafting an email is easy. The work is in reliable account and contact data, signals that are current, positioning content marketing and sales agree on, CRM and sequencing integration, and contact rules per country that are checked before every send.","dataPrerequisites":["A written ideal customer profile and target account list with territories","Approved positioning, value propositions, proof points and claims per segment and product","CRM account, contact and activity history, including previous calls and emails","Contact source records, consent or legitimate interest assessments and suppression lists","Access to signal and enrichment data (news, filings, hiring, technology used) under licence"],"integrations":["CRM (accounts, contacts, activities, opportunities)","Sales engagement or sequencing tool and the reps' mailboxes","Enrichment and intent data providers, and web research","Call recording and conversation intelligence for past interactions","Suppression, opt out and consent management systems"]},"implementation":{"steps":[{"title":"Agree who to target and why","detail":"Write down the ideal customer profile, the signals that make an account worth contacting now and the territory rules with sales leadership. The agent can only prioritize as well as these rules."},{"title":"Build the account brief first","detail":"Start with research and briefs that reps read before writing themselves. Reps will trust drafts only after they trust the research, and the brief shows which sources the agent uses."},{"title":"Ground every claim","detail":"Load approved positioning, case studies and claims into the knowledge base and make the agent cite the signal behind each personalization. No invented customer names, numbers or compliments about the prospect."},{"title":"Put the contact rules in code","detail":"Check the lawful basis, opt outs, country rules and send limits for every contact before a draft is created, not after. Log the basis in the CRM."},{"title":"Keep the rep in the loop","detail":"Every first message and every follow up is approved by the rep and sent from their own account. Autonomous sending comes later, if at all, and only for low risk follow ups."},{"title":"Measure against a control","detail":"Compare reply, meeting and opportunity rates, and opt outs, for agent drafted outreach against a control group of reps or accounts, not only the volume sent."}],"guardrails":["No message leaves without rep approval, and every message is sent from a named person","Personalization must cite a verifiable signal; invented facts, flattery and fake familiarity are blocked","Lawful basis, suppression list and opt out checks before every draft, per contact and country","Send volume limits per rep and domain, and automatic stop on reply, opt out or complaint","No sensitive personal data (health, family, politics) used for personalization","Protection against prompt injection from web pages and documents the agent reads"],"humanInTheLoop":"Reps approve, edit or reject every message and decide when to stop a sequence. Sales leadership owns targeting rules and positioning; marketing and legal own claims and the contact policy. A weekly review of a sample of drafts, replies and opt outs catches tone problems, wrong facts and accounts that should not have been contacted.","kpisToInstrument":["Reply rate and positive reply rate versus a control group","Meetings booked and opportunities created per rep per week","Rep time spent on research and writing, from a time study before and after","Share of drafts approved without major edits, and the reasons for rejection","Opt outs, spam complaints, bounce rate and domain reputation"],"failureModes":[{"title":"Personalization that is wrong or creepy","detail":"The agent cites an outdated role, the wrong company news or personal details the prospect never made public. Require a source per fact, freshness limits and a ban on sensitive data."},{"title":"More volume instead of better outreach","detail":"Teams use the time saved to send many more generic messages, reply rates fall and sending domains get blocked. Cap volume and measure quality, not activity."},{"title":"Contacting people you may not contact","detail":"Consent and opt out rules differ by country and channel; a contact scraped from the web is not a lawful basis. Check before every draft and keep the evidence."},{"title":"Reps stop reading the drafts","detail":"Approval turns into a rubber stamp and errors reach prospects. Track edit rates, sample approved messages and keep the rep accountable for what is sent."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Drafting outreach that a rep reviews and sends as their own message is typically minimal risk. If the agent holds conversations with prospects itself, for example by replying to emails or calling, people must be told they are interacting with AI (Article 50, limited risk). It is not an Annex III use case."},"regulations":["eu-ai-act","gdpr","uk-gdpr","us-tcpa"],"guidance":[{"title":"Directive 2002/58/EC on privacy and electronic communications (ePrivacy Directive)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2002/58/oj","note":"Article 13 sets consent rules for unsolicited electronic marketing to individuals; member states set the rules for legal persons, so B2B email rules differ by country."},{"title":"Electronic mail marketing (guide to PECR)","issuer":"Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/guide-to-pecr/electronic-and-telephone-marketing/electronic-mail-marketing/","note":"UK rules for marketing emails and texts, including the difference between individuals and businesses and the soft opt in for existing customers."},{"title":"CAN-SPAM Act: a compliance guide for business","issuer":"Federal Trade Commission","region":"north-america","url":"https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business","note":"US rules for commercial email, including honest headers and subject lines, identification, a working opt out and honouring opt outs promptly; the law makes no exception for business to business email."},{"title":"Declaratory ruling on AI generated voices under the TCPA (FCC 24-17)","issuer":"Federal Communications Commission","region":"north-america","url":"https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf","note":"Calls that use AI generated voices count as artificial or prerecorded voice calls, so outbound calls with an AI voice to US numbers need the called party's prior express consent unless an exemption applies."}],"controls":["Documented lawful basis per contact source and country, with a legitimate interest assessment where relied on","Suppression and opt out lists synchronized across CRM, sequencing tool and agent","Record of who approved and sent each message, with the draft and the sources used","Approved claims library and review of new message templates by marketing or legal","Data retention limits for prospect research and deletion on request","AI disclosure whenever the agent converses with prospects directly"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that researches each target account with the\nbuilt in **web search and web page browsing** tools, reads CRM history through the ready made\n**HubSpot or Microsoft Dynamics 365** tools (or **custom functions** for another CRM such as\nSalesforce), and writes the account brief and drafts as **structured output**. Approved\npositioning, case studies and claims sit in a **knowledge base** with hybrid retrieval, so every\ndraft is grounded in content marketing and sales signed off. The workflow can run on a\nschedule for a territory or be triggered through an **API token** from the tools reps already\nuse.\n\n**Human in the loop confirmation**, with its threshold set to cover sending, holds every message\nuntil the rep approves or rejects it, and the **tool execution policy** limits which tools the\nagent may use on its own. A custom function checks suppression and opt out lists before\ndrafting, **PII masking** at the gateway protects contact data, and input and output\n**guardrails** screen for prompt injection attempts. Run history gives a full **audit trail**\nper account, **test suites** check drafts against tone, claims and fact rules before every\nprompt change goes live, and the platform is model agnostic, with EU and UAE data residency."},"faq":[{"question":"How is this different from an inbound AI SDR?","answer":"An inbound agent talks to people who came to you and asked for something. An outbound prospecting agent works on people who did not, so its job is research and drafting for a rep, and consent, opt out and anti spam rules decide who may be contacted at all."},{"question":"How much time does it save sales development reps?","answer":"Vendor case studies report large savings: Clay reports that Merge's SDRs got 10+ hours back every week and response rates climbed 20%, and it reports an estimate by Petra Hajal, who built the workflows in Oyster's marketing operations team (she now runs an agency that serves Oyster), that each rep saves about 40 hours a month. These are vendor selected results; run your own time study and a control group."},{"question":"Is AI personalized cold email legal under GDPR?","answer":"It can be, but the AI does not change the rules. You need a lawful basis for processing the contact's data, usually a documented legitimate interest for B2B, you must respect ePrivacy or PECR consent rules for electronic marketing, which are stricter for individuals than for companies, and every message needs a working opt out."},{"question":"Should the agent send messages on its own?","answer":"Start with rep approval for every message. Autonomous sending multiplies any error in facts, tone or targeting across thousands of prospects, and in the United States calls with an AI voice need the called party's prior express consent under the TCPA unless an exemption applies."}],"related":["inbound-lead-qualification-agent","personalized-marketing-at-scale","sales-call-coaching-and-crm-update","client-briefing-and-call-report-copilot","proactive-outbound-engagement-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, cross industry sales page distinct from inbound qualification; evidence from Merge, A-LIGN, Oyster, Unifonic, Dun & Bradstreet and ANS."},{"date":"2026-09-26","note":"Fact checked against sources; Merge and Oyster savings attributed to Clay as claimant, A-LIGN and Merge metric periods and baselines corrected, Blits.ai guardrail and PII masking wording aligned with the feature inventory, SEO title and description added."},{"date":"2026-09-27","note":"Oyster's 40 hour estimate attributed to Petra Hajal as Clay describes her; Merge's 20% response rate kept as prose and removed as a conversion metric; ANS and Dun & Bradstreet country sourced from their own sites; A-LIGN cost figure inconsistency and Unifonic wording disclosed in the evidence."},{"date":"2026-09-27","note":"Fact checked again against all sources; Lumen's four hours stated as per week in the value assumptions, TCPA consent wording aligned with FCC 24-17 (prior express consent unless an exemption applies), ANS deployment year marked as an estimate."}],"slug":"outbound-sales-prospecting-agent","url":"https://www.blits.ai/ai-use-cases/outbound-sales-prospecting-agent","benchmarks":[],"indicativeValueResult":{"low":84480,"high":579600},"evidence":["a-lign-outbound-account-research","ans-copilot-seller-account-agent","dun-and-bradstreet-seller-email-generation","lumen-copilot-sales-account-research","merge-clay-account-research-outreach","oyster-intent-based-outbound-automation","unifonic-copilot-sales-outreach"]},{"title":"AI agent for parcel tracking and delivery exceptions","shortTitle":"Parcel tracking and delivery exceptions","seoTitle":"AI agents for parcel tracking and redelivery","metaDescription":"AI agents answer 'where is my parcel' questions and rebook or redirect deliveries. Chronopost says its agent Léonard resolves 85% of simple requests end to end.","definition":"An AI agent that answers \"where is my parcel\" and resolves delivery exceptions for parcel carriers and postal operators, such as missed deliveries, redelivery or a change of address or pickup point, delays, customs holds and lost or damaged parcel claims, on chat, messaging and phone, and hands disputes and claims above set limits to a human with the tracking history attached.","aliases":["where is my parcel chatbot","redelivery assistant","delivery chatbot","parcel tracking assistant","courier customer service bot"],"industries":["logistics-and-transportation"],"functions":["customer-service","operations"],"patterns":["conversational-agent","voice-agent","agentic-workflow","classification-and-routing"],"channels":["web-chat","voice","whatsapp","mobile-app","sms"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"\"Where is my parcel\" questions are a large share of what a carrier's contact centre handles:\nChronopost says routine enquiries such as parcel tracking and delivery times make up 40 per cent\nof its incoming contacts. Many of those contacts come from someone who is not the carrier's customer: the\nrecipient did not choose the carrier, often does not have an account and only has a tracking\nnumber. Volumes follow ecommerce peaks (Evri calls the weeks from Black Friday to just before\nChristmas its busiest period of the year), so contact centres are busiest exactly when the\nnetwork is under the most pressure.\n\nThe answer is usually already in the tracking data, but tracking events are written for the\nnetwork, not for people (\"exception, consignee not available\", \"held at depot\"). What the\nrecipient wants is an action: deliver it tomorrow, leave it with a neighbour, send it to a\npickup point, pay the customs duty, or open a claim because the parcel is damaged or never\narrived. A first generation chatbot could show the tracking page; an agent connected to the\ndelivery management system can change the delivery, and knows when a claim or a complaint needs\na person.\n\nThis page covers the carrier's side: the recipient contacting the parcel company. The retailer's\nside, where the shopper asks the web shop about an order, a delivery change or a return, is\ncovered by the order status and returns agent.","problemStats":[{"statement":"Chronopost reports that routine enquiries such as parcel tracking and delivery times represent 40 per cent of its incoming contacts.","sourceTitle":"CX Award Innovation: Chronopost Wins Gold for AI Agent","sourceUrl":"https://www.geopost.com/en/news/chronopost-wins-gold-cx-award-for-ai-driven-service/","year":2026},{"statement":"Evri describes the period from Black Friday in late November to just before Christmas Day as its busiest period of the year for parcels.","sourceTitle":"Evri dials up new automated phone line to quickly connect customers to UK advisors, as part of a £46m total investment","sourceUrl":"https://www.evri.com/press/evri-dials-up-new-automated-phone-line","year":2023}],"howItWorks":"1. **Identify the parcel.** The agent takes a tracking number, a phone number or an order\n   reference, and on messaging channels it can start from the notification the recipient already\n   received.\n2. **Translate the tracking.** It reads the scan events and the network status (depot, route,\n   weather or peak delays) and explains in plain language where the parcel is and what happens\n   next, including a realistic delivery window.\n3. **Offer the delivery options.** Within the carrier's rules and the shipper's settings, it\n   offers redelivery on another day, a safe place or neighbour, a pickup point or locker, or a\n   change of address, and confirms the change in the delivery management system.\n4. **Handle exceptions.** For customs holds it explains what is needed and links to payment of\n   duties; for damaged, late or missing parcels it collects the details and photos and opens a\n   claim or an investigation with the right team.\n5. **Hand over well.** Claims above set values, suspected fraud or theft, complaints and\n   repeated failures go to a human agent with the tracking history and what the agent already\n   tried.","valueDrivers":["cost-to-serve","customer-experience","speed","inclusion-and-access"],"kpis":["containment-rate","contact-deflection","response-time-reduction","interactions-handled","customer-satisfaction","first-contact-resolution"],"indicativeValue":{"referenceOrg":"A parcel carrier delivering 200 million parcels a year","inputs":[{"key":"parcels","label":"Parcels delivered per year","low":200000000,"high":200000000,"unit":"parcels per year","note":"The reference carrier."},{"key":"contactRate","label":"Assisted customer contacts per parcel","low":0.01,"high":0.03,"unit":"contacts per parcel","note":"Editorial assumption; replace with your own contact rate."},{"key":"trackingShare","label":"Share of contacts about tracking, delivery times and delivery changes","low":0.4,"high":0.5,"unit":"fraction of contacts","note":"The low end matches Chronopost, which reports that routine enquiries such as parcel tracking and delivery times make up 40 per cent of incoming contacts. The high end, which adds delivery change requests, is an editorial assumption; replace with your own contact reasons."},{"key":"containment","label":"Share of those contacts the agent resolves","low":0.5,"high":0.8,"unit":"fraction of contacts","note":"Conservative against the benchmark on this page (Chronopost reports 85 per cent of simple requests resolved end to end)."},{"key":"costPerContact","label":"Cost of a human handled contact","low":2,"high":5,"unit":"EUR per contact","note":"Editorial assumption for a blended phone and chat contact. Replace with your own fully loaded cost."}],"formula":"parcels * contactRate * trackingShare * containment * costPerContact","currency":"EUR","period":"per year","resultLabel":"Human handled contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of running the AI and the integrations, fewer failed deliveries from redelivery and redirect options, lower claim handling cost and the effect on shippers who choose carriers partly on recipient experience."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Tracking answers are straightforward once scan events are mapped to plain language. The work is in the actions (redelivery, redirect, safe place, claims) through the delivery management system, verification of the recipient, the shipper's restrictions and peak season load.","dataPrerequisites":["Tracking events and exception codes mapped to plain language explanations and next steps","Delivery options and restrictions per shipper and product (signature, age check, high value)","Claim rules (liability limits, evidence needed, deadlines) with an owner","Contact reasons by tracking status from the contact centre"],"integrations":["Track and trace and the delivery management system (redelivery, redirect, pickup point, safe place)","Claims and investigations system","Customs and duty payment for cross border parcels","Notification platform (SMS, email, messaging, app)","Contact centre platform for handover with the tracking history"]},"implementation":{"steps":[{"title":"Start from the contact reasons","detail":"Take a year of contact reasons and split them by tracking status. Most \"where is my parcel\" contacts map to a handful of statuses; write the explanation and the allowed next actions for each."},{"title":"Connect the actions, not just the tracking page","detail":"Integrate redelivery, change of delivery point, safe place and claim intake with the delivery management system, with the shipper's restrictions applied (for example no redirect for age restricted or high value goods)."},{"title":"Separate recipients from shippers","detail":"Recipients need tracking and delivery changes; business shippers need bulk status, pickups, invoices and claims. Give them different authentication and different allow lists."},{"title":"Plan for peak and bad weather","detail":"Load test at peak season volume and prepare network wide messages (weather, strikes, depot backlogs) that the agent uses to answer honestly instead of repeating an estimate that has already slipped."},{"title":"Guard the brand","detail":"Put input and output guardrails on every turn and test the agent against provocation and prompt injection, with a regression set that runs after every model or prompt change."},{"title":"Test before customers do","detail":"Build test conversations per tracking status, per exception and per language, including claims the rules do not allow and attempts to redirect someone else's parcel, and run them on every change."}],"guardrails":["Delivery changes only through an allow list of actions, respecting the shipper's restrictions and proof of delivery rules","Verification (tracking number plus postcode or a one time code) before changing a delivery or showing the full address","Claims above a set value, suspected theft or fraud, and complaints go to a human","Input and output guardrails against abuse, prompt injection and off brand answers, tested after every change","Personal data (names, addresses, phone numbers) masked in logs and model prompts"],"humanInTheLoop":"Humans handle claims above the agent's limit, investigations of missing parcels, suspected fraud and complaints. A team reviews a sample of contained conversations each week, watches handover reasons during peak, and approves every new action or change to claim rules.","kpisToInstrument":["Containment per tracking status and per exception, counting repeat contacts within seven days as not contained","Share of delivery changes completed by the agent and the delivery success rate after a change","Handover rate and handover reasons, especially during peak weeks","Customer satisfaction on contained conversations versus human handled ones","Claims opened by the agent that were later rejected or reopened"],"failureModes":[{"title":"The assistant says something that becomes the story","detail":"A customer provokes the bot into swearing or criticising the company and the screenshots go viral, as happened to DPD in January 2024 after a system update. Keep guardrails on every turn and rerun adversarial tests after every update."},{"title":"Estimates that are not true","detail":"The agent repeats a delivery estimate the network already missed. Use network status and exception events, and say when a date is uncertain."},{"title":"Changing the wrong parcel","detail":"A redirect is made by someone who is not the recipient. Verify before any change and respect the shipper's restrictions."},{"title":"Containment that is really abandonment","detail":"The recipient gives up and complains to the retailer or on social media. Measure repeat contacts across the carrier and the retailer where you can, and satisfaction per status."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1): an assistant that talks directly with recipients must be designed so they know they are interacting with an AI system, unless that is obvious from the context. It is not high risk: explaining tracking, changing a delivery and taking in a parcel claim fall under none of the Annex III areas (it does not decide on public assistance benefits, creditworthiness, insurance pricing or employment), and it involves no practice prohibited by Article 5."},"regulations":["eu-ai-act","gdpr","uk-gdpr"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Recipients must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Parcel companies, complaints and where to get help","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/make-a-complaint/complain-about-postal-services/parcelforce","note":"Ofcom rules require UK parcel delivery companies such as DPD and Evri to have a complaints procedure, and its complaints guidance (in effect from 1 April 2023) says complainants should be told the channels, the process and how long it takes, and that trained staff handle the complaint. The agent's complaint handover should follow that procedure."}],"controls":["AI disclosure at the start of every conversation","Verification before any delivery change, with an audit trail of who changed what","Guardrail and adversarial regression tests rerun after every system, model or prompt update","Claim and compensation limits per action, with human approval above them","Monitoring of handover reasons and complaints during peak weeks"],"incidents":[{"title":"Incident 631: Chatbot for DPD Malfunctioned and Swore at Customers and Criticized Its Own Company","url":"https://incidentdatabase.ai/cite/631/","note":"In January 2024 a customer tracing a parcel got the DPD UK chatbot to swear and criticise the company after a system update; DPD disabled the AI element. Guardrails and regression tests must be rerun after every update, not only at launch."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the carrier's track and\ntrace, delivery management and claims APIs, and a **knowledge base** with delivery options,\nclaim rules and network notices, retrieved with hybrid search. Redelivery, redirect and claim\nintake run as **flows** with deterministic steps, a verification step and photo upload through\nthe receive attachment block, so a change is only made when the rules allow it; the agent\nexplains the tracking and answers open questions.\n\nThe same agent serves **web chat, WhatsApp, SMS and voice**, with streaming speech and DTMF\ninput on the phone, and replies in the recipient's language. **Guardrails** check every input\nand output against abuse and prompt injection, **PII masking** protects names and addresses,\nand **human handover** passes the tracking history to the contact centre. **Test suites** with\nadversarial conversations run after every change and **monitors** run on a schedule, and\n**analytics** with custom dashboard widgets and conversation level satisfaction ratings can break results down by\ntracking status. The platform is model agnostic, with EU\nand UAE data residency."},"faq":[{"question":"How many parcel contacts can an AI agent resolve?","answer":"Chronopost reports that its agent Léonard resolves 85 per cent of simple requests end to end with 40 per cent faster responses, and that routine enquiries such as parcel tracking and delivery times make up 40 per cent of incoming contacts. Hermes UK (now Evri) reported 432,000 chats with its chatbot Holly between its November 2018 launch and October 2019, and a 50 per cent reduction in contacts to customer service operatives."},{"question":"What went wrong with the DPD chatbot?","answer":"In January 2024, after a system update, a customer got the DPD UK chatbot to swear and criticise the company, and DPD disabled the AI element. The lesson is to keep output guardrails on every turn and to rerun adversarial tests after every update."},{"question":"How is this different from an order status agent?","answer":"The carrier's agent serves recipients who often have no account and only a tracking number, and it changes the delivery itself. A retailer's order status agent serves its own shoppers and also handles cancellations, returns and refunds."}],"related":["order-status-and-returns-agent","outbound-reminder-and-confirmation-agent","complaints-handling-agent","first-line-contact-centre-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with four carrier deployments and one paused deployment verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: replaced the unsourced claim that tracking is the single largest contact reason with the Chronopost 40 per cent figure and added two cited problem stats (Chronopost, Evri peak season); corrected the Evri FAQ period (November 2018 to October 2019, not a full year); added contact deflection to the KPIs; replaced PCI DSS (duty payment runs through a link, not the agent) with UK GDPR; made the EU AI Act basis precise (Article 50(1), no Annex III area, no Article 5 practice); limited the analytics claim to what the platform offers; tightened the PostNL evidence summary; added seoTitle and metaDescription."},{"date":"2026-09-26","note":"Second fact check against live sources: all quotes and dates confirmed; replaced the general Ofcom postal complaints page with its parcel companies page, which states the complaints procedure rule and the 2023 complaints guidance; added the DPD Belgium 2020 release as the source for the Phil chatbot."},{"date":"2026-09-27","note":"Review fixes: the Chronopost 40 per cent now keeps the source's scope (routine enquiries such as tracking and delivery times) in the problem, FAQ and value inputs; the Chronopost evidence records the July 2026 rollout as a plan, not a result; the Evri evidence no longer claims the chatbot came after the phone line and its stage is production; inferred channels and languages removed from the DPD Germany and PostNL evidence; the definition drops retailers (covered by the order status and returns agent); monitors described as scheduled."},{"date":"2026-09-27","note":"Fact checked again against all live sources (Geopost, Evri, DPD, BBC, The Guardian, AI Incident Database, PostNL annual report, Ofcom, EU AI Act Article 50): quotes, dates and statements confirmed; the high end of the tracking share input is now marked as an editorial assumption; the DPD Red 2021 source carries its English page heading as title."}],"slug":"parcel-tracking-and-delivery-exception-agent","url":"https://www.blits.ai/ai-use-cases/parcel-tracking-and-delivery-exception-agent","benchmarks":[{"kpi":"contact-deflection","label":"Contact deflection","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"evri-holly-parcel-chatbot-and-voice-assistant","pooled":true}]},{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":85,"min":85,"max":85,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"chronopost-leonard-ai-customer-service-agent","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":432000,"min":432000,"max":432000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"evri-holly-parcel-chatbot-and-voice-assistant","pooled":true}]},{"kpi":"response-time-reduction","label":"Response time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40,"min":40,"max":40,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"chronopost-leonard-ai-customer-service-agent","pooled":true}]}],"indicativeValueResult":{"low":800000,"high":12000000},"evidence":["chronopost-leonard-ai-customer-service-agent","dpd-germany-red-consignee-chatbot","dpd-uk-chatbot-ai-element-disabled","evri-holly-parcel-chatbot-and-voice-assistant","postnl-chatbot-and-conversational-ai-customer-service"]},{"title":"AI agent for patient appointment scheduling, reminders and no show reduction","shortTitle":"Patient scheduling and reminders","seoTitle":"AI patient scheduling and appointment reminders","metaDescription":"AI agents book, move and remind patients by phone and text. NHS England reports a 30% fall in missed appointments in a Mid and South Essex prediction pilot.","definition":"An AI agent that books, moves and cancels patient appointments by phone and messaging while following the provider's scheduling rules (referral, triage level, clinician and visit type, preparation), confirms and reminds patients in two way conversations, predicts who is likely to miss an appointment, and offers freed slots to patients on the waiting list. Unlike a general branch and appointment booking agent, it writes into the electronic health record and must respect clinical constraints, so anything clinical goes to staff.","aliases":["patient scheduling agent","hospital appointment reminder AI","no show prediction and outreach","AI patient access agent","missed appointment reduction"],"industries":["healthcare"],"functions":["customer-service","operations"],"patterns":["voice-agent","conversational-agent","prediction-and-scoring","agentic-workflow"],"channels":["voice","sms","web-chat","mobile-app","whatsapp"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Getting a patient into the right slot is harder than booking a table. Appointments depend on a\nreferral, a triage level, the right clinician and room, preparation instructions and sometimes a\ntest that has to happen first. Patient access teams do this by phone, and at peak times callers wait\non hold or give up, so access to care depends on how long someone can stay on the line.\n\nAt the other end, a large share of booked appointments is simply lost. Patients forget, cannot get\ntransport, cannot take time off or no longer need the visit but never cancel. Every missed\nappointment is clinical time that nobody uses while other patients wait. Standard one way text\nreminders help, but they are sent to everyone at the same moment, cannot answer a question and leave\nthe freed slot empty when a patient cancels late.","problemStats":[{"statement":"NHS England reports that of 124.5 million outpatient appointments in England in the previous year, eight million (6.4%) were not attended, at an estimated annual cost of £1.2 billion.","sourceTitle":"NHS AI expansion to help tackle missed appointments and improve waiting times","sourceUrl":"https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/","year":2024},{"statement":"The UK government reports 8 million missed appointments in elective care in 2023 to 2024.","sourceTitle":"Power to patients as government sets out plan to cut waiting lists","sourceUrl":"https://www.gov.uk/government/news/power-to-patient-as-government-sets-out-plan-to-cut-waiting-lists","year":2025}],"howItWorks":"1. **Take the request on any channel.** The agent answers calls and messages to book, move or cancel\n   an appointment, and identifies the patient against the record before it discloses or changes\n   anything.\n2. **Apply the scheduling rules.** It checks the referral, visit type, clinician, location and any\n   preparation or prior test the rules require, and only offers slots the scheduling system returns\n   for that combination.\n3. **Book and confirm.** It writes the booking into the electronic health record or patient\n   administration system and sends a confirmation with the preparation instructions.\n4. **Remind in a conversation.** Reminders go out at the times that work best for that clinic, and\n   the patient can confirm, cancel or move the appointment in the same thread.\n5. **Target extra help.** A model scores which appointments are likely to be missed; high risk\n   patients get an extra reminder, a better suited time, or an offer of help with transport.\n6. **Backfill freed slots.** When a patient cancels, the agent offers the slot to suitable patients on\n   the waiting list so the clinical time is used.\n7. **Hand clinical questions to people.** Symptoms, urgent concerns and anything outside scheduling\n   go to staff or to urgent care guidance, with the conversation attached.","valueDrivers":["customer-experience","cost-to-serve","inclusion-and-access","employee-productivity"],"kpis":["containment-rate","interactions-handled","handling-time-reduction","response-time-reduction","customer-satisfaction-uplift"],"indicativeValue":{"referenceOrg":"A hospital with 600,000 outpatient appointments a year","inputs":[{"key":"appointments","label":"Outpatient appointments per year","low":600000,"high":600000,"unit":"appointments per year","note":"The reference hospital."},{"key":"missedRate","label":"Share of appointments not attended today","low":0.06,"high":0.08,"unit":"fraction of appointments","note":"NHS England reports 6.4% of outpatient appointments in England were not attended, and higher rates in some specialties (11% in physiotherapy, 8.9% in cardiology). Replace with your own rate.","sourceUrl":"https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/"},{"key":"reduction","label":"Share of missed appointments avoided or backfilled","low":0.1,"high":0.3,"unit":"fraction of missed appointments","note":"Conservative against the evidence on this page (NHS England reports a 30% fall in non attendance in the Mid and South Essex pilot). Editorial assumption, replace with your own."},{"key":"valuePerSlot","label":"Value of an outpatient slot that is used instead of wasted","low":120,"high":170,"unit":"EUR per appointment","note":"NHS England's estimate of £1.2 billion a year for eight million missed appointments implies roughly £150 (about EUR 175) per missed appointment; the euro range is set below that figure to stay conservative. Replace with your own cost per slot.","sourceUrl":"https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/"}],"formula":"appointments * missedRate * reduction * valuePerSlot","currency":"EUR","period":"per year","resultLabel":"Value of clinical time recovered from avoided or backfilled missed appointments","caveat":"Counts recovered clinical capacity only. It leaves out booking calls handled by the agent, shorter waits for patients, the cost of the AI, messaging and integration, and the extra appointments that may be needed when more patients attend."},"macroEstimates":[{"statement":"The UK government says focused action, including AI predictions of which appointments are most likely to be missed, will help save up to 1 million missed appointments.","sourceTitle":"PM sets out plan to end waiting list backlogs through millions more appointments","sourceUrl":"https://www.gov.uk/government/news/pm-sets-out-plan-to-end-waiting-list-backlogs-through-millions-more-appointments","year":2025}],"feasibility":{"complexity":"medium","complexityNote":"The conversation is the easy part. The work is in the scheduling rules (which visit types a patient may book, which need a referral or triage first), a reliable read and write integration with the electronic health record, and identity checks strong enough for health data. Clinical triage by the agent would push this into medical device territory, so keep it out of scope.","dataPrerequisites":["Scheduling rules per specialty, visit type and clinician, owned by the service","Live slot availability and waiting list data from the scheduling system","Contact details, preferred language and channel consent per patient","Historic attendance data to train or calibrate a missed appointment model","Approved preparation instructions per procedure"],"integrations":["Electronic health record or patient administration system (scheduling and waiting list)","Patient portal or app (identity and messaging)","Telephony and messaging (voice, SMS, WhatsApp)","Referral management","Contact centre for handover to patient access staff"]},"implementation":{"steps":[{"title":"Start with rebooking and cancellations","detail":"Cancelling and moving existing appointments is lower risk than first bookings and frees slots immediately. Add new bookings per visit type once the rules for that type are written down."},{"title":"Write the scheduling rules before the prompts","detail":"For each visit type, record who may book it, what must happen first and what the patient must be told. The agent calls these rules as tools; it never infers eligibility from the conversation."},{"title":"Make reminders two way and well timed","detail":"Test reminder timing per clinic. University Hospitals Coventry and Warwickshire found that a reminder 14 days ahead with a second one four days ahead let patients cancel early enough to rebook. Let the patient act in the same thread."},{"title":"Use risk scores to offer help, not to penalise","detail":"Use the missed appointment score to send extra reminders, offer better times or transport support. Do not use it to deny or downgrade bookings, and check it for bias by deprivation, ethnicity and age."},{"title":"Close the loop on freed slots","detail":"Connect cancellations to the waiting list so a freed slot is offered at once to suitable patients, and measure how many freed slots are actually used."}],"guardrails":["No clinical advice or triage decisions by the agent; symptoms go to staff or urgent care guidance","Only slots and visit types the scheduling system confirms for that patient","Identity verification before any appointment detail is disclosed or changed","Missed appointment scores used only to offer support, reviewed for bias","Clear AI disclosure and an easy route to a person on every channel"],"humanInTheLoop":"Patient access staff own exceptions: urgent symptoms, complex bookings across several services, safeguarding concerns and patients who ask for a person. Service managers approve the scheduling rules per visit type, and a clinical safety officer signs off the scope and reviews a sample of conversations each month.","kpisToInstrument":["Missed appointment rate per clinic, against a control group or the prior period","Share of freed slots rebooked from the waiting list","Booking, rebooking and cancellation containment per visit type","Call wait and abandonment rates for patient access lines","Missed appointment rates by deprivation band, ethnicity and age"],"failureModes":[{"title":"Booking the wrong visit type","detail":"The patient arrives without the required test or preparation. Enforce rules per visit type through tools and confirm preparation in writing."},{"title":"Reminder fatigue","detail":"Too many messages lead patients to ignore them or cancel at the last minute. Test timing and frequency per clinic."},{"title":"A risk model that widens inequality","detail":"Scores that correlate with deprivation lead to overbooking or deprioritising the same groups. Use scores for support only and monitor outcomes by group."},{"title":"Symptoms missed in a booking call","detail":"A patient mentions a red flag symptom while rebooking. Detect it and route to clinical staff or urgent care guidance at once."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Booking, rescheduling and reminders carry transparency duties: patients must be told they are dealing with AI (Article 50(1)). It becomes high risk if a public authority, or a provider acting on its behalf, uses it to evaluate eligibility for healthcare services (Annex III point 5(a)), or if it acts as an emergency healthcare patient triage system (Annex III point 5(d)). Clinical triage may also make it a medical device, which is high risk under Article 6(1) when the device needs a notified body assessment. Keep the agent to scheduling and use risk scores only to offer support."},"regulations":["eu-ai-act","gdpr","uk-gdpr","hipaa","us-tcpa"],"guidance":[{"title":"Medical devices: software and artificial intelligence (AI)","issuer":"Medicines and Healthcare products Regulatory Agency","region":"europe","url":"https://www.gov.uk/government/publications/software-and-artificial-intelligence-ai-as-a-medical-device","note":"UK guidance on when software, including AI, is regulated as a medical device; relevant if the agent moves from scheduling into triage."},{"title":"Declaratory ruling on AI generated voices under the TCPA (FCC 24-17)","issuer":"Federal Communications Commission","region":"north-america","url":"https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf","note":"AI generated voices count as artificial or prerecorded voice under the TCPA, which matters for automated reminder calls in the US."}],"controls":["Clinical safety case with the agent's scope written down and signed off","Audit trail of every booking, change and cancellation the agent made","Data protection impact assessment covering health data, recordings and risk scores","Regular bias review of the missed appointment model","Regression tests for scheduling rules on every change"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that read slots and write bookings,\ncancellations and waiting list offers in the scheduling system over REST, so the scheduling rules\nlive in tools the agent must call. A **flow** handles identity checks and the fixed parts of a\nbooking with **show options** for slots, while the agent handles free requests such as \"can I move\nmy Tuesday appointment to an evening\". Reminder runs and slot backfill are **agentic workflows**\ntriggered on a schedule or by an API call from the hospital systems, with **human in the loop**\napproval where a service wants staff to confirm.\n\nThe same agent runs on **voice**, **SMS**, **WhatsApp**, **web chat** and, through the **API\nchannel**, in the provider's own app, in the patient's language. **Guardrails** block clinical\nadvice, **PII masking** at the gateway masks identifiers such as phone numbers and, with custom\npatterns, patient numbers, and **human handover** passes symptoms and complex cases to\npatient access staff. **Test suites** replay booking conversations for every visit type on each\nchange, and **EU and UAE data residency** keeps patient data in region."},"faq":[{"question":"Does AI actually reduce missed hospital appointments?","answer":"Public pilots suggest it can. NHS England reports a 30% fall in non attendance at Mid and South Essex NHS Foundation Trust with software that predicts missed appointments and books backup slots, and a drop from 10% to 4% in one patient group at University Hospitals Coventry and Warwickshire after AI analysis changed reminder timing. These are pilot results, and none of them is reported against a control group."},{"question":"How is this different from a general appointment booking chatbot?","answer":"A general booking agent finds a location and a free time. A patient scheduling agent has to respect referral, triage and visit type rules, write into the health record, handle preparation instructions and stay out of clinical advice, which makes the integration and the safety case the main work."},{"question":"Can a voice agent handle patient calls at scale?","answer":"WellSpan Health reports that its AI agent Ana manages more than 160,000 patient calls a month, including inbound calls and primary care scheduling. PolyAI reports 30% containment and a 72% shorter handle time for routine requests at Howard Brown Health, where the agent guides patients through scheduling in MyChart; booking, rescheduling and cancelling through Epic is the announced next phase."},{"question":"Is a missed appointment prediction model high risk?","answer":"Used to send extra reminders or offer help, it is usually not. It needs more care when scores decide who is overbooked or deprioritised, because that can limit access to care for the groups who already miss most appointments."}],"related":["branch-and-appointment-booking-agent","outbound-reminder-and-confirmation-agent","health-prior-authorization-and-claims-adjudication","first-line-contact-centre-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with evidence from NHS England (three trusts), WellSpan Health, Howard Brown Health, Audibel and two anonymous vendor claims."},{"date":"2026-09-25","note":"Consolidation pass: added Telephone Consumer Protection Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: made the EU AI Act basis precise (Article 50(1), Annex III points 5(a) and 5(d), Article 6(1) for medical devices); added UK GDPR; corrected the Howard Brown Health scheduling claim (Epic booking is a planned phase); removed the ambiguous IntentAI 77% containment figure; set the Mid and South Essex and Coventry records to pilot; tightened value notes and the FAQ."},{"date":"2026-09-27","note":"Review fixes: capped the value per slot at EUR 170 so it stays below the sourced NHS England figure; Sheffield Children's text volume moved to the evidence summary (one way reminders, not interactions handled); metaDescription now names the prediction pilot; narrowed evidence records where languages, country and vendor roles were inferred."},{"date":"2026-09-27","note":"Fact checked against sources again (NHS England archive, GOV.UK, PolyAI, WellSpan, Microsoft, Google Cloud, FCC, MHRA): all quotes and figures confirmed; metaDescription rewritten as full sentences; FAQ softened to say pilots suggest a reduction; the PII masking claim narrowed to what the platform masks (patterns at the gateway, not names)."}],"slug":"patient-appointment-scheduling-and-reminders-agent","url":"https://www.blits.ai/ai-use-cases/patient-appointment-scheduling-and-reminders-agent","benchmarks":[{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":30,"min":30,"max":30,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"howard-brown-health-patient-voice-agent","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":72,"min":72,"max":72,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"howard-brown-health-patient-voice-agent","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":160000,"min":160000,"max":160000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"wellspan-hippocratic-ai-patient-voice-agent","pooled":true}]},{"kpi":"response-time-reduction","label":"Response time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":87,"min":87,"max":87,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"audibel-voice-agent-appointment-calls","pooled":true}]},{"kpi":"customer-satisfaction-uplift","label":"Satisfaction uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":4,"min":4,"max":4,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"howard-brown-health-patient-voice-agent","pooled":true}]}],"indicativeValueResult":{"low":432000,"high":2448000},"evidence":["audibel-voice-agent-appointment-calls","howard-brown-health-patient-voice-agent","mid-and-south-essex-missed-appointment-prediction","sheffield-childrens-ai-attendance-predictor","uhcw-process-mining-appointment-reminders","wellspan-hippocratic-ai-patient-voice-agent"]},{"title":"AI agent for payment initiation within a customer mandate","shortTitle":"Agentic payment initiation","seoTitle":"Agentic payments: AI agents that pay on mandate","metaDescription":"An AI agent pays for the customer within limits they set. See live pilots by Santander, ING and DBS with Mastercard Agent Pay, plus the controls banks need.","definition":"An AI agent that initiates and completes payments or purchases on a customer's behalf, within a mandate the customer set in advance (spending caps, allowed merchants or categories, a tokenized credential and rules for when to ask for confirmation), and then confirms and reconciles every transaction it made.","aliases":["agentic payments","agent initiated payments","delegated payments agent","AI agent checkout"],"industries":["payments","banking","retail-and-ecommerce"],"functions":["customer-service","sales","operations"],"patterns":["agentic-workflow","conversational-agent"],"channels":["mobile-app","web-chat","whatsapp","voice","api"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"front-office","problem":"AI assistants have become good at finding things: the right product, the cheapest ticket, the\nbill that is due. The moment money has to move, they stop. The customer is sent to a checkout\npage, types in card details, passes a one time passcode and may give up before paying. Recurring chores such\nas paying a service charge, topping up a transit card or reordering groceries still need a human\nat every step.\n\nLetting software pay on someone's behalf is not new (standing orders, card on file, merchant\ninitiated payments), but those rails assume a fixed amount or a known merchant. An AI agent picks\nthe merchant, the item and the amount itself. That raises questions the existing rails do not\nanswer: did the customer actually authorise this purchase, can the issuer and merchant tell a\nlegitimate agent from a bot, who carries the loss when the agent buys the wrong thing, and how does\nthe customer see and undo what the agent did. Card networks, wallets and banks are now piloting\nanswers: an explicit mandate, a credential issued per agent, authentication at the right moment\nand a record of what was authorised.\n\nThis page covers the payment step. Helping a shopper discover and compare products is covered by\nthe conversational shopping assistant page; servicing an existing account or card (blocks, limits,\nstatements) is covered by the account and card servicing page.","problemStats":[{"statement":"Visa's Trust Index, a Harris Poll survey of U.S. consumers in May 2026, found that only 23% of U.S. consumers trust generative AI to handle payment transactions on their behalf.","sourceTitle":"New Visa Research Finds Consumer Trust is Accelerating the Path to Agentic Commerce","sourceUrl":"https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22736.html","year":2026},{"statement":"In Visa's Business to AI report with Morning Consult, 60% of Americans said they would not allow AI to spend any amount without approval, and only 27% were comfortable letting AI spend autonomously without limits.","sourceTitle":"Visa Defines the Next Era of Commerce: When AI Becomes the Customer","sourceUrl":"https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22266.html","year":2026}],"howItWorks":"1. **Set the mandate.** The customer tells the agent what it may do: a budget or a cap per\n   transaction and per period, allowed merchants or categories, which card or account to use and\n   when it must ask first (for example above a set amount, for a new merchant, or always). The\n   mandate is stored as structured data, not as a line in a prompt.\n2. **Bind a credential to the agent.** The issuer or wallet provisions a tokenized credential for\n   this agent only, so the real card number never reaches the agent, and the token can be limited,\n   paused or revoked on its own.\n3. **Find and decide.** The agent searches, compares and builds the purchase (a product, a ticket, a\n   ride, a bill payment) and checks it against the mandate before anything is paid.\n4. **Confirm when the rules say so.** Inside the mandate the agent proceeds; outside it, or where\n   the mandate requires it, the agent asks the customer, who approves with strong authentication\n   such as a passkey or app confirmation.\n5. **Pay through the rails.** The agent submits the payment through the network, wallet or payment\n   service provider with identifiers that mark it as agent initiated, so the issuer and merchant\n   can see who acted and the issuer can still authorise or decline.\n6. **Confirm and reconcile.** The agent tells the customer what it bought, for how much and from\n   whom, matches the authorisation, capture and any refund against the order, and logs the\n   mandate, the decision and the payment as one record for disputes and audit.","valueDrivers":["customer-experience","revenue-growth","risk-reduction","speed"],"kpis":["conversion-rate-uplift","automation-rate","users-served","interactions-handled","fraud-loss-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"An online retailer with 2 million orders a year","inputs":[{"key":"orders","label":"Online orders per year","low":2000000,"high":2000000,"unit":"orders per year","note":"The reference retailer."},{"key":"agentShare","label":"Share of shoppers who buy through an AI agent","low":0.005,"high":0.03,"unit":"fraction of orders","note":"Editorial assumption, replace with your own. No deployment on this page has published a share of orders completed by an AI agent: the cited deployments are pilots and single milestone transactions. The Mastercard figure elsewhere on this page is a 2030 forecast about the share of shoppers who will use AI agents routinely, not a measured share of orders today, so it is not a benchmark for this input."},{"key":"incrementalShare","label":"Share of agent orders that would otherwise be lost","low":0.1,"high":0.3,"unit":"fraction of agent orders","note":"Editorial assumption for orders that would have been abandoned or placed with a competitor without agent checkout. Replace with your own test results."},{"key":"averageOrderValue","label":"Average order value","low":60,"high":90,"unit":"USD per order","note":"Editorial assumption. Replace with your own figure."},{"key":"grossMargin","label":"Gross margin","low":0.3,"high":0.4,"unit":"fraction of order value","note":"Editorial assumption for a general online retailer. Replace with your own margin."}],"formula":"orders * agentShare * incrementalShare * averageOrderValue * grossMargin","currency":"USD","period":"per year","resultLabel":"Incremental gross margin from orders completed by AI agents","caveat":"Gross margin on incremental orders only. It leaves out the cost of integrating with agent platforms and payment networks, fees charged by agent platforms, returns and disputes on agent orders, and cannibalisation of orders that would have come through the retailer's own site."},"macroEstimates":[{"statement":"A Mastercard report based on four futurists' predictions expects more than one in ten online shoppers to routinely use AI agents to purchase products on their behalf by 2030.","sourceTitle":"Mastercard report predicts that one in 10 people will routinely use AI agents to shop and pay by 2030","sourceUrl":"https://newsroom.mastercard.com/news/europe/en/newsroom/press-releases/en/2026/mastercard-report-predicts-that-one-in-10-people-will-routinely-use-ai-agents-to-shop-and-pay-by-2030/","year":2026}],"feasibility":{"complexity":"high","complexityNote":"The agent itself is the easy part. The work is in the payment plumbing (tokenized credentials per agent, network agent programmes, wallet or PSP integration), in storing and enforcing the mandate outside the model, in authentication that fits both regulation and the customer's patience, and in a dispute process for purchases the customer did not expect.","dataPrerequisites":["A mandate model per customer (caps, periods, merchants or categories, confirmation rules) stored as structured data","Product, price and availability data from merchants in a form agents can read, or access to agent enabled catalogs","Order, authorisation, capture and refund events from the payment provider for reconciliation","Dispute and chargeback history to calibrate confirmation thresholds"],"integrations":["Card network agent programmes or wallet agent services for agent tokens and agent identification","Payment service provider or acquirer (authorisation, capture, refunds, webhooks)","Issuer or bank authentication (passkeys, app confirmation, one time passcode)","Merchant catalogs and order management systems, directly or through agent commerce protocols","Customer notification channels (app push, messaging, email) for confirmations and receipts","Ledger or order system for reconciliation and dispute evidence"]},"implementation":{"steps":[{"title":"Start with narrow, repeatable purchases","detail":"Pick journeys where the item and merchant are predictable and the harm of a mistake is small: reorders, top ups, tickets from one venue, recurring bills. Leave open ended shopping across unknown merchants for later."},{"title":"Model the mandate before the agent","detail":"Write the mandate as data: cap per transaction, cap per period, allowed merchants or categories, currency, expiry and the confirmation rule. Enforce it in code at payment time, so a clever prompt cannot talk the agent past a limit."},{"title":"Use tokens and agent identifiers, never raw card data","detail":"Issue a credential per agent through the network, wallet or PSP, keep card numbers out of the conversation and the model, and send the identifiers that mark the payment as agent initiated so issuers and merchants can see it."},{"title":"Put confirmation where the risk is","detail":"Ask for explicit approval with strong authentication for the first purchase, a new merchant, anything above the threshold or anything the agent is unsure about. Use authorise then capture so a purchase can still be stopped before money settles."},{"title":"Close the loop on every transaction","detail":"Send a receipt in the customer's channel, reconcile authorisation, capture and order, and keep the mandate, the agent's reasoning and the payment together as one record for disputes."},{"title":"Test the limits, not just the happy path","detail":"Build a test set that tries to exceed caps, switch merchants, repeat payments, inject instructions through product pages and pay in the wrong currency, and run it on every change to the agent or its tools."}],"guardrails":["Mandate limits (amount, period, merchant or category, currency, expiry) enforced in code at payment time, outside the model","Explicit customer confirmation with strong authentication above the threshold, for new merchants and for the first purchase","Tokenized credentials per agent that can be paused or revoked on their own; no card numbers in prompts or logs","Idempotency keys and velocity limits so a retry or loop cannot pay twice","A tool policy that lets the agent call only the payment and catalog tools it needs","Prompt injection checks on content the agent reads from merchant pages and product data"],"humanInTheLoop":"The customer is the human in the loop: they set the mandate, approve anything outside it and can pause or revoke the agent at any time. Inside the organization, payments and fraud teams own the mandate rules and thresholds, review disputes on agent purchases every week and approve every new merchant category or payment type before the agent can use it.","kpisToInstrument":["Completed agent purchases and the share that needed customer confirmation","Mandate breaches blocked by the payment layer (should be caught there, never by the customer)","Dispute and refund rate on agent purchases compared with the same customers' own purchases","Fraud losses on agent tokens compared with other card not present spend","Reconciliation breaks between order, authorisation and capture","Customer satisfaction and revocation rate of mandates"],"failureModes":[{"title":"The agent buys the wrong thing within its limits","detail":"A purchase that is technically allowed but not what the customer wanted, such as the wrong size or a near duplicate. Keep caps tight at first, confirm new merchants and items, and make cancellation and refunds one step away."},{"title":"Mandates enforced only in the prompt","detail":"Limits written as instructions to the model can be ignored or talked around. Enforce them in the payment service, which rejects any request outside the mandate whatever the agent says."},{"title":"Injection through merchant content","detail":"A product page or review tells the agent to pay elsewhere or buy more. Treat everything the agent reads as untrusted, restrict payees to the mandate and screen tool inputs."},{"title":"Confirmation fatigue","detail":"If the agent asks for approval on every purchase, customers stop using it; if it never asks, trust breaks on the first mistake. Tune thresholds per journey using dispute and satisfaction data."},{"title":"No trail when a dispute comes","detail":"Without the mandate, the confirmation and the agent's decision stored with the payment, the issuer and merchant cannot resolve a chargeback. Log them together from day one."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing agent must make clear that people are dealing with AI, unless that is obvious from the context (Article 50). Initiating payments within a customer's mandate is not listed in Annex III. It becomes high risk if the same agent evaluates creditworthiness, for example by deciding on a buy now pay later or credit line at checkout (Annex III point 5(b))."},"regulations":["eu-ai-act","eu-psd2","gdpr","pci-dss","dora","uk-consumer-duty"],"guidance":[{"title":"Regulation (EU) 2024/1689 (AI Act), Article 50: transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Commission Delegated Regulation (EU) 2018/389 on strong customer authentication and common and secure open standards of communication","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg_del/2018/389/oj","note":"Sets the technical requirements for strong customer authentication, including dynamic linking to the amount and payee, and the exemptions from it, which shape whether a customer must authenticate at the mandate, at each agent payment or both."}],"controls":["AI disclosure and a clear statement of what the agent may buy, for how much and with which credential","Mandate records with customer authentication, versioning and expiry","Audit trail that links mandate, confirmation, agent decision and payment for every transaction","Per agent token lifecycle management (issue, limit, pause, revoke)","Dispute and refund handling for agent purchases with defined liability between issuer, merchant and agent provider","Regression tests on mandate enforcement and injection resistance before every release"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the payment step runs as an **agentic workflow** with a **tool execution policy**\nthat limits the agent to the catalog, payment and notification tools it needs. **Custom\nfunctions** call the merchant, wallet or bank APIs and check each request against the customer's\nmandate, which is stored as data rather than in the prompt. Payments go through the platform's\npayment service with **Stripe, Mollie, Adyen or Qi**, using **authorize then capture** as the\napproval step, **card tokenization** at the provider, **per tenant amount ceilings, currency\nallowlists, idempotency keys and velocity limits**, and **signature verified webhooks** that keep\npayment status in sync for reconciliation. Customer specific caps and merchant rules sit in the\ncustom functions on top of those platform limits.\n\n**Human in the loop confirmation** above a configurable threshold holds agentic actions until an\nauthorised operator approves or rejects them, and the payment layer refuses payments above the\ntenant threshold. Confirmation by the customer in the conversation is built in the custom\nfunctions or flows. **Agentic tasks** handle deferred instructions that wait for a condition, with\nscheduled rechecks, expiry and a dry run mode that plans without acting. The same agent works in **web chat,\nWhatsApp and voice**, and in your own mobile app through the **REST or WebSocket API channel**.\nCard numbers typed in chat are detected and tokenized at the\ngateway, **guardrails** screen prompt injection, and every run has an **audit trail** with\ntamper evident logging of payment events. **Test suites** replay attempts to break the mandate on\nevery change, **monitors** check the flow on a schedule, and the platform is model agnostic with\nEU and UAE data residency."},"faq":[{"question":"How is agentic payment initiation different from an AI shopping assistant?","answer":"A shopping assistant helps a customer find and compare products and then hands over to a basket or checkout. Agentic payment initiation is the next step: the agent actually pays, using a credential and a mandate the customer set, and confirms and reconciles the payment. Many deployments combine both, but the controls for moving money are different."},{"question":"Is anyone doing this live yet?","answer":"Mostly as pilots. Banco Santander and Mastercard ran a live agent executed payment, ING, Worldline and Mastercard completed an agentic payment in production in the Netherlands, and Mastercard, DBS and UOB completed a transaction in Singapore in which an AI agent booked and paid for a ride. PayPal launched checkout inside Perplexity for U.S. users in November 2025, with the shopper completing checkout in the chat. None of them has published volumes or outcome figures."},{"question":"Does the customer have to approve every payment?","answer":"Not necessarily. The customer sets a mandate, and the agent pays within it; approval with strong authentication is required above a threshold, for new merchants or whenever the mandate says so. Surveys show why this matters: in Visa's research, 60% of Americans would not let AI spend any amount without approval, and in the ING and DBS pilots the customer explicitly approved the purchase."},{"question":"How does strong customer authentication work when an agent pays?","answer":"Current pilots authenticate the customer when the mandate or purchase is confirmed, for example with passkeys, and use a tokenized credential issued to the agent. Whether later payments within a mandate need their own authentication under PSD2 depends on how the payment is structured, so settle it with the issuer and your legal team before launch."}],"related":["conversational-shopping-assistant","account-and-card-servicing-agent","real-time-fraud-scoring","card-dispute-and-chargeback-intake","ledger-and-payment-reconciliation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with seven public evidence records verified against their sources."},{"date":"2026-09-26","note":"Fact checked against sources. Corrected the DBS and UOB, Majid Al Futtaim, Santander, PayPal and Visa summaries to match their releases, removed channels the sources do not state, softened an unsupported abandonment claim, clarified the Article 50 and SCA guidance wording, limited the Blits.ai channel list to inventory channels, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Second review. Corrected the Blits.ai human in the loop description (operators approve, not the customer), replaced an unsupported claim about confirmation across deployments with the ING and DBS pilots, dated the PayPal launch instead of claiming it is still live, removed the ING channel and trimmed the Blits.ai internal record to its portfolio entry, and cited the AI Act on EUR-Lex."}],"slug":"agentic-payment-initiation","url":"https://www.blits.ai/ai-use-cases/agentic-payment-initiation","benchmarks":[],"indicativeValueResult":{"low":18000,"high":648000},"evidence":["dbs-uob-mastercard-agentic-transaction","ing-worldline-agentic-payment-in-production","majid-al-futtaim-mastercard-agent-pay-pilot","paypal-perplexity-instant-buy","santander-mastercard-agent-pay-live-payment","ulta-beauty-agentic-checkout-google-ai-mode","visa-intelligent-commerce-agent-transactions"]},{"title":"AI agent for personalized offers and rewards","shortTitle":"Offers and rewards","seoTitle":"AI agents for card offers and loyalty rewards","metaDescription":"Banks use AI to pick the next best offer or reward for each customer. See how Commonwealth Bank and DBS target rewards, with the compliance risks to manage.","definition":"A customer facing AI agent for banks and card issuers that picks the offer, reward or loyalty action most relevant to each customer at each moment from their transactions and context, delivers it in the app, in messaging or through a colleague, and helps the customer understand, track and redeem rewards in conversation. Unlike campaign personalization, it works inside the customer's own account and loyalty relationship, one moment at a time.","aliases":["next best action","next best offer","personalized card offers","rewards assistant","loyalty chatbot"],"industries":["banking","payments"],"functions":["marketing","sales","customer-service"],"patterns":["recommendation-and-personalization","prediction-and-scoring","conversational-agent"],"channels":["mobile-app","web-chat","whatsapp","agent-desktop"],"audience":"customer-facing","autonomy":"autonomous","adoptionStage":"mainstream","segment":"front-office","problem":"Banks and card issuers run many offers (card linked merchant deals, rewards points, fee waivers,\nproduct upgrades) and mostly send them as campaigns to broad segments. Customers ignore what is\nnot relevant, points go unredeemed and the rewards budget buys little loyalty. Staff in branches\nand contact centres have no view of which conversation matters most for the customer in front of\nthem.\n\nThe opposite failure is just as real: aggressive targeting that pushes credit at customers who\nare struggling, or offers that systematically skip some groups. The job is to choose the one\nrelevant thing for this customer now, including \"nothing to sell, here is help instead\", and to\nmake it explainable.","problemStats":[],"howItWorks":"1. **Build the candidate list.** All offers, rewards and service messages the customer is\n   eligible for, filtered by business rules, consent and suitability.\n2. **Score and choose.** Propensity and value models rank the candidates; an arbitration layer\n   picks the next best action for the customer across all channels, so the app, the contact\n   centre and the branch show the same priority.\n3. **Deliver in context.** The action appears where the customer is: an in app card, a message\n   after a relevant purchase, or a prompt on a colleague's screen during a call.\n4. **Converse about rewards.** An assistant explains the points balance, what a reward is worth,\n   how to redeem and what is needed for the next tier, and completes the redemption through\n   approved APIs.\n5. **Learn.** Responses (accepted, ignored, dismissed) feed back into the models, and outcomes\n   are monitored for fairness and customer harm.","valueDrivers":["revenue-growth","customer-experience"],"kpis":["conversion-rate-uplift","revenue-uplift","interactions-handled","churn-reduction","nps-change"],"indicativeValue":{"referenceOrg":"A card issuer with 1 million active cardholders","inputs":[{"key":"cardholders","label":"Active cardholders","low":1000000,"high":1000000,"unit":"customers","note":"The reference issuer."},{"key":"reachedShare","label":"Share of cardholders who see personalized offers each year","low":0.3,"high":0.5,"unit":"fraction of cardholders","note":"Editorial assumption, depends on app usage and consent."},{"key":"extraAcceptance","label":"Additional offer acceptances per reached cardholder versus generic campaigns","low":0.005,"high":0.015,"unit":"acceptances per reached cardholder per year","note":"Editorial assumption; no deployment on this page discloses a conversion uplift. Measure it against a control group."},{"key":"marginPerAcceptance","label":"Margin per accepted offer","low":50,"high":150,"unit":"USD per acceptance","note":"Editorial assumption, replace with your own offer economics."}],"formula":"cardholders * reachedShare * extraAcceptance * marginPerAcceptance","currency":"USD","period":"per year","resultLabel":"Additional margin from personalized offers","caveat":"Incremental offer margin only. It leaves out loyalty and retention effects, rewards costs saved by better targeting, merchant funded revenue, and the cost of the decisioning platform and data work."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Decisioning engines are mature. The effort is in clean eligibility and consent data, one arbitration across channels, measurement with control groups and the fairness and suitability rules that keep targeting safe.","dataPrerequisites":["Transaction and product holding data per customer","Offer catalog with eligibility rules, costs and expiry","Marketing consent and contact preferences per channel","Response history with control groups"],"integrations":["Decisioning or next best action engine","Rewards and loyalty platform (balance, redemption)","Card linked offer provider where used","App, messaging and agent desktop channels","Consent management platform"]},"implementation":{"steps":[{"title":"Define the action catalog, including service","detail":"List every offer and every service action (a fee refund, a hardship check in, a reward reminder) and give each eligibility, suitability and consent rules. Service actions must be able to win against sales."},{"title":"Arbitrate in one place","detail":"Use one decisioning layer for all channels so the customer does not get three different offers from the app, email and a call. Commonwealth Bank's engine serves branches, the contact centre and digital channels from one decisioning engine."},{"title":"Measure with control groups","detail":"Hold out a random control group from the start, or no one will be able to say what the engine added."},{"title":"Add the rewards conversation","detail":"Let customers ask about points, value and redemption in the app assistant, with balances and redemption through the loyalty platform's APIs."},{"title":"Review fairness and harm regularly","detail":"Check who is shown and who is excluded from offers, and suppress credit offers for customers with signs of financial difficulty."}],"guardrails":["Eligibility, suitability and consent checked by rules before any model ranks an offer","No credit offers to customers with financial difficulty or vulnerability markers","Offers are clearly labelled as offers and respect opt outs on every channel","Reward values and redemption terms come from the loyalty platform, never generated text","Periodic bias review of targeting outcomes across customer groups"],"humanInTheLoop":"Marketing and product owners approve every offer, rule and model change. A conduct review checks targeting outcomes each quarter. Colleagues in branches and contact centres decide whether to raise a suggested conversation at all.","kpisToInstrument":["Acceptance rate versus a control group, per offer","Incremental revenue or margin per treated customer","Reward redemption rate and points balance age","Opt outs and complaints about offers","Offer exposure by customer segment"],"failureModes":[{"title":"Optimising for the wrong customers","detail":"Models learn that struggling customers accept credit offers. Exclude them by rule and monitor outcomes."},{"title":"Invisible uplift","detail":"Without a control group, gains cannot be separated from seasonality. Hold out from day one."},{"title":"Channel conflict","detail":"Each channel runs its own targeting and customers get contradictory offers. Arbitrate centrally."},{"title":"Unexplainable targeting","detail":"A customer or regulator asks why an offer was shown or withheld and nobody can answer. Log the rules and scores behind each decision."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Ranking offers is generally minimal risk and the conversational part carries the Article 50 transparency duty. Using AI to evaluate creditworthiness for a credit offer is high risk (Annex III point 5(b)), and Article 5 prohibits techniques that exploit vulnerabilities due to a person's social or economic situation to distort their behaviour in a harmful way."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty"],"guidance":[{"title":"Direct marketing and privacy and electronic communications","issuer":"Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/","note":"The ICO's hub for UK direct marketing rules under PECR and data protection law, covering the lawful basis and consent for marketing messages, with its detailed direct marketing guide."},{"title":"Article 5: Prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Bans manipulative techniques and the exploitation of vulnerabilities, including those due to a person's economic situation."}],"controls":["Decisioning models in the AI inventory with owners, validation and periodic bias review","Documented suitability rules for credit and high cost products","Consent and opt out enforcement across all channels","Decision logs that explain why each offer was shown or suppressed","Complaint and outcome monitoring for customers in vulnerable circumstances"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the decisioning itself stays in the bank's next best action engine; the\nplatform delivers and explains its decisions. An **AI agent** in the app or on **WhatsApp**\ncalls the engine and the loyalty platform through **custom functions**, shows offers and\npoints with rich cards (loyalty points, credit cards, product recommendations) and completes\nredemptions through **flows** with confirmation.\n\nOffer terms and rewards rules sit in a **knowledge base** so explanations are grounded.\nReward values come from the loyalty platform, and **guardrails** help prevent the agent from\ninventing values or pushing credit to customers who mention financial difficulty, while\n**human handover** routes those conversations to people. **Analytics** track interactions,\nsatisfaction and sentiment, and **test suites** check the suppression rules before each release. The platform is model agnostic."},"faq":[{"question":"Is next best action the same as a recommendation engine?","answer":"It is broader. A next best action engine chooses among offers, service messages and doing nothing, across channels. Commonwealth Bank said in 2022 that its engine made over 35 million decisions a day and used it to suggest the next best conversation to have with each customer, including same day support for customers affected by natural disasters and matching customers to government benefits through its Benefits finder."},{"question":"Can an assistant help customers use their rewards?","answer":"Yes. Bank of America's Erica highlights cash back deals based on the client's spending and notifies clients of their eligibility for its Preferred Rewards program, and DBS plans to add reward point tracking to its digibot assistant in a later phase of its agentic rollout."},{"question":"What is the main compliance risk?","answer":"Unfair or harmful targeting, such as pushing credit at customers who are struggling or systematically excluding groups. Rules for suitability and consent, control groups and regular bias reviews are the defence."}],"related":["personalized-marketing-at-scale","next-best-action-for-advisors","proactive-outbound-engagement-agent","financial-wellbeing-coach","account-and-card-servicing-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog and verified against the sources. The catalog's USD 100 billion unredeemed points figure appears only in a 2017 press release headline and was not used; its Capital One source is a third party blog."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: reworded the Commonwealth Bank and DBS FAQ answers to match the sources (DBS reward tracking is planned, not live), added UK GDPR, corrected the ICO guidance note, limited the Blits.ai analytics claim to the feature inventory, added seoTitle and metaDescription, dated the Commonwealth Bank source and corrected channels and wording in two anonymized Blits.ai records."},{"date":"2026-09-27","note":"Review fixes: dated the Commonwealth Bank 35 million decisions figure (June 2022) and stopped calling it offer decisions, dropped the Pega 50 million figure (conversations held by staff, not handled by AI), limited Commonwealth Bank channels to the sources, removed an anonymized event demo record that was not an offers deployment, trimmed an internal record to its source, softened two Blits.ai claims and removed the MAS AI risk management entry, which is still a consultation proposal on the cited MAS page."},{"date":"2026-09-27","note":"Second fact check against sources: all quotes, FAQ answers, guidance links and EU AI Act references reconfirmed; the Commonwealth Bank record now describes the engine as suggesting the next best conversation, as the Pega source says, instead of offer or message decisions."}],"slug":"offers-and-rewards-agent","url":"https://www.blits.ai/ai-use-cases/offers-and-rewards-agent","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":3000000000,"min":3000000000,"max":3000000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-erica-virtual-assistant","pooled":true}]}],"indicativeValueResult":{"low":75000,"high":1125000},"evidence":["bank-of-america-erica-virtual-assistant","commonwealth-bank-customer-engagement-engine","dbs-joy-and-digibot-virtual-assistants"]},{"title":"AI agent for proactive customer outreach, activation and retention","shortTitle":"Proactive outreach and activation","seoTitle":"AI agents for proactive bank customer outreach","metaDescription":"AI agents that contact bank customers first about card activation, fee alerts and retention. Bank of America reports over 1.7 billion proactive Erica insights.","definition":"An AI agent that holds the conversation when a bank reaches out first to change something about the customer's account or products, triggered by an event or a campaign: low balance and fee avoidance alerts, payment and renewal reminders, card activation, dormant account reactivation and offers the customer already qualifies for, over messaging or voice, while the bank's own systems decide who is contacted and why. Reminders about appointments and deliveries the customer booked, and the in app coach the customer opens, are separate use cases.","aliases":["outbound banking agent","card activation outreach","dormant account reactivation","event triggered customer outreach"],"industries":["banking","payments"],"functions":["marketing","sales","customer-service"],"patterns":["conversational-agent","voice-agent","agentic-workflow","recommendation-and-personalization"],"channels":["whatsapp","sms","voice","mobile-app","email"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"front-office","problem":"Banks pay to acquire customers and products that then sit unused. A new card that never leaves the\ndrawer earns nothing, a savings account opened for a promotion goes dormant, and a customer who\nkeeps paying overdraft fees has a reason to look at another bank. The bank usually knows the\nmoment to act (a card not used 30 days after delivery, a balance heading below zero, a fixed rate\nending), but the tools it has are weak: push notifications and emails that are easy to ignore,\nor human outbound calls that are costly to scale beyond high value products.\n\nWhat is missing is a way to have a short, useful two way conversation at scale: answer the\ncustomer's question, complete the action (activate, set up a transfer, accept an offer they\nalready qualify for) and leave them alone when they say no. Because some of these conversations\ntouch credit and fees, they are regulated conversations, not marketing copy.","problemStats":[{"statement":"Forrester's survey of banking customers, as reported by Mi3 in 2025, found that the thing customers most want (60 percent) is to be alerted when there is not enough money in their account to cover an upcoming expense.","sourceTitle":"Westpac blows app rivals away as Forrester rates Australia among world's best – but Big Four still missing key customer aspirations","sourceUrl":"https://www.mi-3.com.au/01-10-2025/big-four-deliver-ai-nudges-and-cautious-cleverness-their-banking-apps-forrester","year":2025}],"howItWorks":"1. **The bank decides who and why.** Event triggers and campaign lists come from the bank's own\n   systems (core banking, the decision engine, CRM), including eligibility for any offer. The\n   agent does not choose targets.\n2. **Check consent and contact rules.** Before any message or call the agent checks marketing\n   consent where needed, frequency caps, quiet hours and the customer's preferred channel.\n3. **Open with the reason.** The agent says it is an AI assistant, names the bank and states the\n   specific reason for contact (\"your new card has not been activated\").\n4. **Converse and act.** It answers questions from the customer's own account data and approved\n   product content, and completes the action through allow listed APIs after the right\n   authentication: activate the card, set up a transfer to avoid a fee, book a call, accept a\n   pre approved offer.\n5. **Respect no.** An opt out or a \"not now\" is recorded at once and suppresses further contact\n   on that topic.\n6. **Hand over.** Complaints, hardship, vulnerability and complex product questions go to a\n   human with the conversation attached.\n7. **Write back the outcome.** Results flow to CRM so the decision engine learns and the next\n   campaign does not repeat the contact.","valueDrivers":["revenue-growth","customer-experience","cost-to-serve"],"kpis":["conversion-rate-uplift","churn-reduction","revenue-uplift","interactions-handled","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A card issuer that issues 100,000 new cards a year","inputs":[{"key":"newCards","label":"New cards issued per year","low":100000,"high":100000,"unit":"cards per year","note":"The reference issuer."},{"key":"inactiveShare","label":"Share of new cards not used within 90 days","low":0.1,"high":0.2,"unit":"fraction of new cards","note":"Editorial assumption, replace with your own activation data."},{"key":"activationLift","label":"Share of inactive cards activated because of the outreach","low":0.05,"high":0.15,"unit":"fraction of inactive cards","note":"Editorial assumption; measure against a control group that receives only the usual push and email."},{"key":"revenuePerActiveCard","label":"Annual revenue from an active card","low":50,"high":120,"unit":"USD per card per year","note":"Editorial assumption, replace with your own figure."}],"formula":"newCards * inactiveShare * activationLift * revenuePerActiveCard","currency":"USD","period":"per year","resultLabel":"Annual revenue from cards activated by outreach","caveat":"Covers card activation only. It leaves out fee avoidance, reactivation of dormant accounts, accepted offers and retention, the cost of messages, calls and the AI, and any complaints from unwanted contact."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The conversation is straightforward. The hard parts are the triggers and eligibility data, the consent and contact rules per channel and market, and authenticating a customer on a conversation the bank started.","dataPrerequisites":["Event triggers and campaign lists with the reason for contact","Consent, channel preference and suppression data per customer","Eligibility decisions for any offer, made upstream by the bank's decision engine","Approved product and fee content for questions"],"integrations":["CRM or customer engagement platform (lists, outcomes, suppression)","Messaging and telephony channels with opt out handling","Core banking and card platform (activation, transfers, limits)","Identity and step up authentication","Contact centre for handover and callbacks"]},"implementation":{"steps":[{"title":"Start with service triggers, not sales","detail":"Card activation, low balance and payment reminders help the customer and build trust in the channel. Add offers only after service outreach performs well."},{"title":"Keep the campaign brain outside the agent","detail":"Targeting, eligibility and offer terms come from the bank's systems and are passed to the agent as facts. The agent explains and executes; it does not decide who qualifies."},{"title":"Build consent and frequency into the trigger","detail":"Check consent, caps and quiet hours before the first message, per channel and market, and log the check with the contact."},{"title":"Authenticate on the way in","detail":"When the bank starts the contact, the customer must still prove who they are before any action, ideally in the app. Never ask for secrets in an outbound message."},{"title":"Test against a control group","detail":"Hold out a random share of each trigger and compare activation, fees and churn, so the results are yours and not the vendor's."}],"guardrails":["The agent only offers products the bank's decision engine has already approved for that customer","AI disclosure and the reason for contact in the first message or sentence","Opt outs honoured immediately across channels","No requests for passcodes or card details in outbound contact","Frequency caps and quiet hours enforced before sending"],"humanInTheLoop":"Marketing and product owners approve every trigger, script and offer before launch. Humans take over for complaints, hardship, vulnerable customers and complex product questions, and a sample of conversations is reviewed each week for tone, accuracy and fair treatment.","kpisToInstrument":["Activation, reactivation and acceptance rates against a control group","Opt out and complaint rate per trigger","Fees avoided for customers who acted on a low balance alert","Handover rate and reasons","Churn of contacted customers versus control"],"failureModes":[{"title":"Outreach that feels like spam","detail":"Too many contacts or vague reasons drive opt outs and complaints. Cap frequency, lead with the reason and stop when the customer says no."},{"title":"The agent starts deciding eligibility","detail":"Free form offers creep into the conversation. Pass eligibility in as data and block offers that were not approved upstream."},{"title":"Scam lookalike messages","detail":"Outbound bank messages are what scammers imitate. Never ask for secrets, and point customers to the app to act."},{"title":"No measurable effect","detail":"Without a control group every activation looks like a win. Hold out a share of each trigger from day one."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing agent must disclose that it is AI (Article 50(1)). It stays out of Annex III as long as eligibility for credit offers is decided upstream by the bank's own, separately governed credit processes; if the agent itself assessed creditworthiness it would be high risk under point 5(b)."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","dora","us-tcpa"],"guidance":[{"title":"Guide to Privacy and Electronic Communications Regulations","issuer":"Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/guide-to-pecr/","note":"UK rules on marketing by phone, text and email, including consent and opt out. The ICO says the guide is under review after the Data (Use and Access) Act 2025."},{"title":"Declaratory ruling on AI generated voices under the TCPA (FCC 24-17)","issuer":"Federal Communications Commission","region":"north-america","url":"https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf","note":"AI generated voices count as \"artificial or prerecorded voice\" under the TCPA, so US outbound AI calls need the consent the TCPA requires."}],"controls":["Consent, frequency and quiet hour checks logged per contact","Approved scripts and offer terms per trigger with an accountable owner","Audit trail of every contact, answer and action taken","Suppression lists shared across channels","Monitoring of complaints and outcomes for vulnerable customers"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai each trigger is an **agentic workflow** started by the bank's systems through an API\ntoken or on a schedule, with the customer, the reason and any pre approved offer passed in as\ndata. Blits.ai can send the first message itself on the **email channel**, which works inbound\nand outbound. For outreach in the bank's own app, the app starts the contact and hands the\nconversation to the **AI agent** through the **REST or WebSocket API channel**. For WhatsApp, SMS\nand voice, the bank's own messaging or dialler platform sends the first message or places the\ncall, and the agent answers the replies on the **WhatsApp**, **SMS** and **voice telephony**\nchannels. In every case the agent answers from a **knowledge base** of approved product content\nand acts only through **custom functions** that call the bank's APIs, behind an\n**authentication** block.\n\nA **flow** fixes the opening (AI disclosure, reason for contact, opt out) and the consent checks,\nwhile the **GDPR toolkit** handles consent gating and data removal. **Human handover** routes\ncomplaints and hardship to the contact centre, **guardrails** stop the agent from inventing\noffers or asking for secrets, and **test suites** replay each trigger's conversations before\nlaunch. The workflow **run history with analytics** shows outcomes per trigger for comparison with\nthe control group."},"faq":[{"question":"Is this the same as a marketing campaign tool?","answer":"No. The bank's campaign and decision systems still choose who to contact and what they qualify for. The agent is the caller: it explains, answers questions, completes the action and records the outcome."},{"question":"Do banks already reach out proactively with AI?","answer":"Yes, but mostly as proactive alerts and insights in the app and by message, from an assistant the customer can then talk to. Bank of America says clients have received and interacted with more than 1.7 billion proactive, personalized insights from Erica, and Capital One says its Eno assistant looks out for charges that might surprise the customer and alerts them by text, email and app. We found few published results for AI agents that hold two way outbound conversations or make outbound calls in banking, which is why this page treats the pattern as emerging."},{"question":"What consent do outbound AI calls need?","answer":"It depends on the market and on whether the call is service or marketing. In the US the FCC has ruled that AI generated voices fall under the TCPA's rules for artificial voices; in the UK PECR governs marketing calls, texts and emails. Build the check into every trigger."}],"related":["outbound-reminder-and-confirmation-agent","offers-and-rewards-agent","financial-wellbeing-coach","fraud-alert-confirmation","collections-and-hardship-agent","personalized-marketing-at-scale"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI use case catalog and verified against Bank of America, Capital One, CIMB Niaga, Forrester coverage and regulator sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added Telephone Consumer Protection Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added a Forrester customer survey statistic (via Mi3) as a problem statistic; softened two unsourced claims in the problem; made the Eno FAQ answer match Capital One's wording; added UK GDPR and the ICO review notice; aligned the Blits.ai build notes with the feature inventory; corrected the Westpac source title and removed unsupported CIMB Niaga languages; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: the Blits.ai build notes now claim only outbound email as a Blits.ai started contact, with the bank's app, messaging or dialler platform starting app, WhatsApp, SMS and voice contact; the Forrester statistic follows the Mi3 wording more closely; the FAQ says what the adoption stage rests on; softened the ICO guidance note; removed the staff facing CIMB Niaga evidence from this page."},{"date":"2026-09-27","note":"Fact checked again against all sources (Mi3 and Forrester, Bank of America, Capital One, the ICO PECR guide, FCC 24-17, the Google Cloud list and the feature inventory); every claim is supported. Rephrased one unsourced claim about overdraft fees and switching in the problem."}],"slug":"proactive-outbound-engagement-agent","url":"https://www.blits.ai/ai-use-cases/proactive-outbound-engagement-agent","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":3000000000,"min":3000000000,"max":3000000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-erica-virtual-assistant","pooled":true}]}],"indicativeValueResult":{"low":25000,"high":360000},"evidence":["bank-of-america-erica-virtual-assistant","capital-one-eno-assistant","commonwealth-bank-customer-engagement-engine"]},{"title":"AI agent for public transit passenger information and disruption reporting","shortTitle":"Transit passenger information agent","seoTitle":"AI chatbot for transit passenger information","metaDescription":"An AI agent answers transit riders from live service data and logs their reports. Live at Chicago Transit Authority; NJ Transit pilots Navvie for trip information.","definition":"An AI agent on a public transport operator's website, app or messaging channel that answers riders' real time questions (\"when is my bus coming\", \"why is my train delayed\") from live service data, takes a structured report when something is wrong on board or at a station, and flags urgent or safety related reports for fast human follow up, in the rider's own language.","aliases":["transit chatbot","public transport virtual assistant","rider service disruption assistant","where is my bus/train chatbot","transit agency AI assistant"],"industries":["logistics-and-transportation"],"functions":["customer-service","operations"],"patterns":["conversational-agent","classification-and-routing"],"channels":["web-chat","mobile-app","sms"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"emerging","problem":"A public transport operator runs a large, always moving network and has to tell a diverse rider\nbase what is happening on it right now: is my bus coming, why is my train delayed, is the\nelevator at this station working. The Chicago Transit Authority (CTA), an independent government\nagency, operates one of the largest transit systems in the US, more than a million rides on buses\nand trains on an average weekday, 24 hours a day, across the City of Chicago and 35 surrounding\nsuburbs, serving a diverse community that includes many multilingual commuters.\n\nTraditional channels can struggle to keep up: a call centre works through a queue, station\nsignage and a schedule app show the plan more than the live reality, and a rider who wants to\nreport a problem, a dirty train, a broken air conditioner, a safety concern, often has to call\nor wait to flag someone down. NJ Transit, for example, currently reaches riders through station\nand onboard digital signage, its DepartureVision and MyBus systems, mobile apps, websites, SMS\nand push notifications, email alerts, social media, real time and third party data feeds, and\npublic address systems, and the agency says it wants to unify these into one authoritative\nsource of information.","problemStats":[{"statement":"The Chicago Transit Authority operates one of the nation's largest public transportation systems covering the City of Chicago and 35 surrounding suburbs, operating 24 hours a day with over a million rides on buses and trains on an average weekday.","sourceTitle":"Chicago Transit Authority Connects with City: AI Chatbot Bridges Language Barriers and Empowers Riders","sourceUrl":"https://publicsector.google/ai/chicago-transit-authority-launches-a-multi-lingual-chatbot-for-more-a-more-seamless-commute/","year":2025}],"howItWorks":"1. **Understand the request.** The agent classifies whether the rider is asking a schedule or\n   trip question, asking about a known disruption, or reporting a problem on a vehicle, at a\n   station or with staff.\n2. **Answer from live data.** Trip and delay questions are answered from the operator's real\n   time vehicle location, schedule adherence and service alert feeds, not a static timetable.\n3. **Turn a report into a case.** A free text report (\"the AC is broken on the Red Line\") is\n   structured into a category, a location and an urgency, and logged with a reference number the\n   rider can follow up on.\n4. **Flag what is urgent.** Safety concerns and other urgent situations are flagged for fast\n   human follow up, on a stated time target, rather than sitting in the same queue as a routine\n   cleanliness report.\n5. **Hand over what needs a person.** Safety incidents, complaints, unsupported languages and\n   complex itinerary questions go to a human agent with the conversation already captured.","valueDrivers":["customer-experience","cost-to-serve","inclusion-and-access","speed"],"kpis":["contact-deflection","productivity-gain","containment-rate","response-time-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A transit agency carrying 150 million passenger trips a year","inputs":[{"key":"trips","label":"Passenger trips per year","low":150000000,"high":150000000,"unit":"trips per year","note":"The reference agency."},{"key":"contactsPerTrip","label":"Assisted contacts (calls, chats, social posts, in station reports) per trip","low":0.0005,"high":0.0015,"unit":"contacts per trip","note":"Editorial assumption, replace with your own contact volume."},{"key":"automationShare","label":"Share of contacts the agent resolves or logs without a person reaching them first","low":0.15,"high":0.35,"unit":"fraction of contacts","note":"Editorial assumption, replace with your own. No evidence on this page reports a containment, automation or deflection share: the customer service reach and conversation completion figures Google Public Sector reports for CTA measure different things, and NJ Transit's Navvie is still a pilot with no outcome disclosed yet. For scale only, CTA staff review over 250 incidents a week across a system with over a million weekday rides (Google Public Sector), which does not by itself imply a share of contacts."},{"key":"costPerContact","label":"Cost of a human handled contact","low":3,"high":7,"unit":"USD per contact","note":"Editorial assumption for a blended phone, chat and social contact. Replace with your own fully loaded cost."}],"formula":"trips * contactsPerTrip * automationShare * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of running the AI and its integrations, the value of faster and more accurate disruption information to riders, and any change in the number or quality of maintenance and safety reports the agency actually receives."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering from a static timetable is easy; the work is integrating live vehicle and service alert feeds so answers reflect reality during a disruption, and wiring reports into a system that maintenance and operations teams actually work from, in more than one language.","dataPrerequisites":["Real time vehicle location, schedule adherence and service alert feeds","A taxonomy of report types (cleanliness, mechanical, safety, lost property, staff conduct) mapped to the right team","Current fare, accessibility and policy content","Rider language data, to prioritise which languages to support first"],"integrations":["Real time transit operations feed (vehicle location, schedule adherence, service alerts)","Incident or work order system for maintenance and operations teams","Notification channels (SMS, app push, website and station alerts)","Contact centre or social media platform for handover"]},"implementation":{"steps":[{"title":"Start with schedule and disruption questions","detail":"Answer \"when is my bus or train coming\" and active service alerts first: the highest volume, lowest risk questions, and the ones a static timetable answers worst during a disruption."},{"title":"Add structured issue reporting once answers are trusted","detail":"Turn free text reports into a category, a location and an urgency, and give every report a reference number the rider can check back on."},{"title":"Define what counts as urgent, in writing","detail":"Agree with operations and safety teams which report categories and keywords trigger fast human follow up, and the time target for that follow up."},{"title":"Support the languages your riders actually speak","detail":"Prioritise languages by rider population data, not by what is easiest to add first, the way CTA supports English, Spanish, Polish, Simplified Chinese and Filipino/Tagalog."},{"title":"Test before riders do","detail":"Build a test set of real questions and reports per category and per supported language, including ambiguous and urgent ones, and run it on every change."},{"title":"Widen channels once the first one is proven","detail":"Launch on the website or app first, measure containment and report routing accuracy, then add messaging channels and voice."}],"guardrails":["Trip and disruption answers only from live operational data, with a refusal when the feed is stale or unavailable","Every report gets a case reference and a routed owner, tracked to closure","Urgent or safety related reports flagged for human follow up within a stated time target","Input and output guardrails against abuse and off topic prompts, tested after every change","Personal data in reports (names, contact details, photos) masked in logs and model prompts"],"humanInTheLoop":"Operations and safety staff review every flagged urgent report and decide the follow up. A team monitors handover reasons and unmatched report categories weekly, and approves any new report category or supported language before it goes live.","kpisToInstrument":["Share of contacts resolved or logged without a person, per question and report type","Time from an urgent report to a confirmed human follow up","Reports confirmed as real issues by maintenance or operations teams, versus all reports logged","Customer satisfaction on chatbot interactions versus human handled ones","Language coverage of the assistant against the rider population it serves"],"failureModes":[{"title":"Confident but stale disruption information","detail":"The agent repeats a schedule that a live disruption has already overtaken. Ground trip and delay answers only in live feeds, and say clearly when live data is unavailable."},{"title":"Reports that go nowhere","detail":"A report is logged but never reaches a team that acts on it. Wire every report category to an owning team and track reports to closure, not just to intake."},{"title":"An urgent situation misclassified as routine","detail":"A safety relevant report is filed as a routine cleanliness complaint. Err toward escalation on ambiguous language and review missed urgent cases weekly."},{"title":"Coverage gaps that exclude riders","detail":"The agent supports only the languages that were easiest to add, leaving other riders no better off than before. Prioritise languages by rider population, not convenience."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1): riders must be told they are dealing with an AI system, unless that is obvious from the context. Answering trip questions and logging reports is not a listed Annex III use; it would need a fresh assessment if the same agent decided eligibility for a reduced fare, a concession or paratransit access, which touches access to an essential public service."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Riders must be informed that they are interacting with an AI system unless this is obvious from the context."}],"controls":["AI disclosure at the start of every conversation","Case reference and an owning team for every report, tracked to closure","Escalation rules for urgent and safety related reports, with a monitored time target","Guardrail and regression tests rerun after every model or prompt change"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the operator's real\ntime vehicle and service alert feed, plus a **knowledge base** with fare, accessibility and\npolicy content, retrieved with hybrid search. Issue reporting runs as a **flow** with\ndeterministic steps: category, location and urgency, then a **custom function** creates the\ncase in the maintenance or operations system. The urgent path routes through **agent\nhandover** or a **send email** action, or a **custom function** that calls the on call\nsystem's REST API, so a flagged safety report reaches a person fast.\n\nThe same agent serves **web chat, WhatsApp and SMS**, and, through the **API channel**, the\noperator's own mobile app, replying in the rider's own language. **Guardrails** check input\nand output for abuse and off topic prompts, **PII masking** protects names and contact\ndetails in reports, and **human handover** passes an escalated case to the contact centre or\noperations team with the full conversation. **Test suites** run per question and report type\non every change, **monitors** run on a schedule, and **custom dashboards** break down\nconversation outcomes and report categories. The platform is model agnostic."},"faq":[{"question":"What results have transit agencies reported from this kind of chatbot?","answer":"Google Public Sector reports that the Chicago Transit Authority's Chat with CTA chatbot, built with Google and Quantiphi, grew CTA's customer service reach by over 63% and lifted conversation completion by 16% since launch, and that it helps intercept urgent situations within five minutes of a rider's first message. NJ Transit's Navvie is a narrower assistant for trip information only, launched as a pilot by early September 2026 according to Mass Transit, and the agency says results are still being analysed."},{"question":"Does this replace real time apps like a trip planner or a map app?","answer":"No. It answers from the same kind of live vehicle and service alert data those apps use, but inside a conversation, and it adds a way to report a problem and get a case reference, which a trip planner does not do."},{"question":"How does the agent decide a report is urgent?","answer":"Operations and safety teams agree in advance which report categories and language (for example anything describing a safety threat) trigger fast human follow up, with a stated time target, rather than joining the same queue as a routine cleanliness report."},{"question":"What should stay with a person?","answer":"Safety incidents, complaints, lost property claims above a set value, languages the assistant does not yet support, and any itinerary question complex enough that a scripted answer would mislead the rider."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version. Covers passenger information and disruption reporting agents for public transit, sourced from the Chicago Transit Authority (Google Cloud case study) and NJ Transit's Navvie pilot, both fetched and quote checked with usecases:source."},{"date":"2026-09-28","note":"Editorial pass on review findings: dropped the two CTA metrics that did not map to a taxonomy KPI (kept as attributed claims in evidence and FAQ), removed the unsupported EU Accessibility Act listing, corrected the CTA source year to 2025 (Wayback capture), removed unsourced claims about Chicago's language mix and NJ Transit's channel list, rewrote the meta description and the automationShare note, and fixed the howToBuild capabilities and the Filipino/Tagalog language name."},{"date":"2026-09-28","note":"Second editorial pass on adversarial review: rewrote the meta description so it no longer credits Navvie with issue reporting or languages the source does not support, reframed the NJ Transit launch date as reported by Mass Transit rather than dated by inference (evidence summary and FAQ updated to match), replaced the unlisted outbound webhook capability in howToBuild with a custom function calling a REST API, softened the dashboard wording to conversation outcomes, swapped the weak quality score uplift KPI for containment rate, tightened the CTA evidence summary to the source's exact wording (bus, not bus or train; detailed and timely reports to maintenance crews, not structured tickets), moved its internal KPI taxonomy commentary out of the public summary, cleared the unsourced NJ Transit languages field, and called Navvie a narrower assistant for trip information in the FAQ."}],"slug":"public-transit-passenger-information-agent","url":"https://www.blits.ai/ai-use-cases/public-transit-passenger-information-agent","benchmarks":[],"indicativeValueResult":{"low":33750,"high":551250},"evidence":["chicago-transit-authority-chat-with-cta","nj-transit-navvie-chatbot"]},{"title":"AI agent for source of wealth due diligence in private banking","shortTitle":"Source of wealth diligence","seoTitle":"AI agents for source of wealth due diligence","metaDescription":"AI agents draft source of wealth reports from client documents for human review. Bank of Singapore cut report writing from 10 days to one hour.","definition":"An AI agent that reads a prospective private client's documents, extracts and corroborates how their wealth was built, checks plausibility against benchmarks and external sources, and drafts the source of wealth and enhanced due diligence narrative for the relationship manager and compliance analyst, who decide on the risk rating and the relationship.","aliases":["source of wealth report drafting","SoW assessment assistant","enhanced due diligence narrative drafting","private banking KYC agent"],"industries":["wealth-and-asset-management","banking"],"functions":["onboarding-and-kyc","financial-crime-compliance"],"patterns":["document-processing","agentic-workflow","content-generation","summarization"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"front-office","problem":"Private banks must understand and document how a client acquired their total wealth, not only the\nfunds that arrive in the first account. For entrepreneurs, heirs and executives this means reading\nhundreds of pages of financial statements, tax notices, corporate filings, property valuations and\npayslips, reconciling them into a coherent story and writing it up. Bank of\nSingapore described this as work that took its relationship managers about ten days per report.\n\nThe work is subjective and varies with the experience of the person writing it, so reports are\ninconsistent, and gaps that surface in compliance review send the file back and delay the account\nopening. Regulators have pushed in both directions: after major money laundering cases they\nexpect more rigorous source of wealth work, and in Singapore the regulator has also asked private\nbanks to shorten account opening times.","problemStats":[{"statement":"The Monetary Authority of Singapore asked private banks in May 2026 to cut account opening times to within one month by the end of 2026, from an average of six weeks or more, as reported by Global Business Outlook.","sourceTitle":"Bank of Singapore, DBS take AI route to accelerate wealth client onboarding (Global Business Outlook)","sourceUrl":"https://globalbusinessoutlook.com/banking-and-finance/bank-of-singapore-dbs-take-ai-route-to-accelerate-wealth-client-onboarding/","year":2026}],"howItWorks":"1. **Collect documents.** The relationship manager uploads the client's documents and the fact\n   find; the agent classifies them and lists what is missing for the client's profile.\n2. **Extract and reconcile.** Income, business sales, inheritances, investment gains and assets\n   are extracted with dates and amounts and reconciled into a wealth timeline.\n3. **Corroborate.** The agent checks plausibility against benchmarks (for example typical salary\n   for a role, company revenue) and approved external sources such as company registries and news,\n   alongside the separate PEP, sanctions and adverse media screening results.\n4. **Draft the narrative.** A structured source of wealth report is drafted in the bank's template,\n   citing the document and page behind each statement and flagging gaps and inconsistencies.\n5. **Human decision.** The relationship manager verifies and refines the draft; the compliance\n   analyst challenges it, requests more evidence if needed and decides the risk rating and whether\n   to proceed.","valueDrivers":["speed","compliance","employee-productivity","customer-experience"],"kpis":["cycle-time-days","processing-time-reduction","time-saved-per-task","accuracy","interactions-handled"],"indicativeValue":{"referenceOrg":"A private bank onboarding 2,000 new clients a year","inputs":[{"key":"newClients","label":"New private clients needing a source of wealth report per year","low":2000,"high":2000,"unit":"clients per year","note":"The reference bank."},{"key":"hoursPerReport","label":"Staff hours per source of wealth report today","low":8,"high":16,"unit":"hours per report","note":"Editorial assumption for effort, not elapsed time. Bank of Singapore reports elapsed writing time fell from 10 days to one hour; replace with your own effort data."},{"key":"shareSaved","label":"Share of effort saved with an AI drafted report","low":0.5,"high":0.8,"unit":"fraction of effort","note":"Conservative against the benchmark on this page, because verification and compliance review remain."},{"key":"hourlyCost","label":"Blended hourly cost of relationship managers and analysts","low":80,"high":150,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"newClients * hoursPerReport * shareSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Value of staff time released from source of wealth reports","caveat":"Leaves out the larger commercial effect of faster account opening (assets that arrive sooner, fewer abandoned onboardings) and the cost of tooling and external data."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"It touches a regulated financial crime control. It needs reliable document extraction across many formats and languages, approved external data sources, a clear template agreed with compliance, strict data security and a model risk review.","dataPrerequisites":["Source of wealth policy and report template per client type and jurisdiction","Client documents in digital form, with a secure upload path","Benchmark data for plausibility checks (salaries, company financials)","Access to company registries, adverse media and screening results"],"integrations":["Client onboarding or KYC case management system","Document management and secure upload","Screening providers for PEP, sanctions and adverse media","Company registry and financial data providers"]},"implementation":{"steps":[{"title":"Agree the report standard with compliance","detail":"Write down what a complete source of wealth report contains for each client type (entrepreneur, heir, executive, investor) and which evidence each statement needs."},{"title":"Start with extraction and drafting","detail":"Let the agent extract, reconcile and draft with citations to document pages, while all judgments stay with the relationship manager and analyst."},{"title":"Add corroboration sources carefully","detail":"Connect approved external sources one at a time and record which source supported which statement; never let the model rely on its own general knowledge as evidence."},{"title":"Measure quality, not only speed","detail":"Track compliance send backs, missing evidence and analyst edits alongside turnaround time, and compare with manually written reports."},{"title":"Keep data inside a controlled environment","detail":"Client wealth documents are highly sensitive; process them in a private or dedicated environment with strict access control, as Bank of Singapore does on its private cloud."}],"guardrails":["Every statement in the report cites a document page or an approved external source","The model never decides the risk rating or the onboarding outcome","Gaps and inconsistencies are flagged, not smoothed over","Client documents processed in a controlled environment with access limited to the case team","No use of the model's general knowledge as evidence of wealth"],"humanInTheLoop":"The relationship manager verifies and refines every draft before submission; the compliance analyst challenges it and decides the risk rating; senior management approval applies for PEPs and other high risk clients as the bank's policy requires.","kpisToInstrument":["Elapsed time from complete document set to submitted report","Compliance send back rate and reasons","Share of report statements with a valid citation","Analyst edit rate per report section","Time from first contact to account opening"],"failureModes":[{"title":"Plausible but unsupported narrative","detail":"The draft reads well but a key wealth event has no evidence. Require a citation per statement and flag uncited text."},{"title":"Extraction errors in complex documents","detail":"Amounts or dates are misread from scanned statements or foreign language filings. Show source snippets next to extracted values for verification."},{"title":"Over reliance by reviewers","detail":"Analysts approve well written drafts with less challenge. Sample reports for independent review and track challenge rates."},{"title":"Data leakage","detail":"Sensitive documents reach systems or models outside the controlled environment. Enforce data residency, masking and access control."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Anti money laundering due diligence is not listed in Annex III, so an assistant that drafts source of wealth reports for a human decision is not high risk by default. It becomes high risk if it adds remote biometric identification of the client (Annex III point 1(a); verification that only confirms a claimed identity is excluded) or feeds an assessment of a natural person's creditworthiness, for example for lending to the client (Annex III point 5(b)). GDPR Article 22 on solely automated decisions applies if it ever refused a client on its own."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","mas-ai-risk-management","dora","iso-42001","eu-amlr","us-bsa","mas-notice-626"],"guidance":[{"title":"Notice 626 on Prevention of Money Laundering and Countering the Financing of Terrorism (Banks)","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/notices/notice-626","note":"Singapore's anti money laundering requirements for banks, including enhanced customer due diligence for politically exposed persons and other higher risk customers, such as establishing their source of wealth and source of funds."},{"title":"FG17/6: The treatment of politically exposed persons for anti-money laundering purposes","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publications/finalised-guidance/fg17-6-treatment-politically-exposed-persons-peps-money-laundering","note":"UK guidance on applying enhanced due diligence to politically exposed persons in proportion to the risk they present."},{"title":"Artificial Intelligence Model Risk Management (information paper)","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Good practices for AI and generative AI model risk management observed in a MAS thematic review of banks, covering governance and oversight, risk management systems and processes, and development and deployment."}],"controls":["Inventory entry and model risk review for the drafting and extraction components","Report template and evidence standard approved by financial crime compliance","Citation coverage check before a report can be submitted","Independent sampling of AI drafted reports by a second line reviewer","Data residency, encryption and access controls on client documents"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow**: documents uploaded to the case are ingested (PDF,\nWord, Excel, images and Outlook email files), an **AI agent** extracts and reconciles wealth events with\n**structured output**, and **custom functions** call the bank's registry, benchmark and screening\nproviders through REST. The bank's source of wealth policy and templates sit in a **knowledge\nbase**, and the agent drafts the report with a citation per statement.\n\nThe workflow pauses for **human in the loop** review by the relationship manager and the analyst,\nand its **audit trail** keeps every run. **PII masking**, **tenant isolation** and deployment in\nthe **EU or UAE region** keep documents under the bank's control, and the **model agnostic**\nplatform lets the bank choose a model, or a regional one, per step. **Test suites** check\nextraction and citation quality on reference cases before each change goes live."},"faq":[{"question":"How much faster can source of wealth reports be?","answer":"Bank of Singapore says the time to write a report fell from 10 days to one hour with its Source of Wealth Assistant, with relationship managers verifying and refining each draft. Deutsche Bank has also put an agentic source of wealth solution live in Singapore and Hong Kong."},{"question":"Does the AI decide whether a client is accepted?","answer":"No. It drafts and flags; the relationship manager, the compliance analyst and, for high risk clients, senior management decide. Bank of Singapore has relationship managers verify each draft before internal review, and Deutsche Bank says that while tasks can be automated, accountability remains with its people."},{"question":"Can the AI verify wealth on its own?","answer":"It can check plausibility against benchmarks and approved sources and point out gaps, but every statement must be backed by a document or an approved external source, never by the model's general knowledge."}],"related":["pep-and-adverse-media-screening","dynamic-customer-risk-rating","perpetual-kyc","digital-onboarding-assistant","business-onboarding-and-ubo-discovery"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Source of Wealth Diligence) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added the MAS account opening statistic, tightened the EU AI Act basis and two guidance notes, aligned the document list and FAQ with the sources, dated the evidence sources, named the HELIOS platform and removed an unsupported vendor; added SEO title and description."}],"slug":"source-of-wealth-diligence","url":"https://www.blits.ai/ai-use-cases/source-of-wealth-diligence","benchmarks":[{"kpi":"cycle-time-days","label":"Cycle time","unit":"hours","aggregate":false,"higherIsBetter":false,"n":1,"nUpTo":0,"median":1,"min":1,"max":1,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-singapore-source-of-wealth-assistant","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-singapore-agentic-wealth-onboarding","pooled":true}]}],"indicativeValueResult":{"low":640000,"high":3840000},"evidence":["bank-of-singapore-agentic-wealth-onboarding","bank-of-singapore-source-of-wealth-assistant","deutsche-bank-source-of-wealth-kyc-agent"]},{"title":"AI agent for telecom bill explanation and billing disputes","shortTitle":"Bill explanation and disputes","seoTitle":"AI agent for telecom bill questions and disputes","metaDescription":"AI agents explain telecom bills line by line and route disputes to people. BT Group's generative AI platform behind EE's Aimee explains billing charges.","definition":"An AI agent that explains a customer's telecom bill line by line, in plain language and on any channel, answers why a charge changed or appeared, corrects clear errors within set limits and opens a billing dispute with the evidence attached when a human has to decide.","aliases":["bill explainer","billing inquiry chatbot","bill shock assistant","telco billing assistant","billing dispute intake"],"industries":["telecommunications"],"functions":["customer-service","case-management"],"patterns":["conversational-agent","rag-knowledge-assistant","agentic-workflow","voice-agent"],"channels":["mobile-app","web-chat","whatsapp","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"\"Why is my bill higher this month?\" is a routine, high volume question for telecom operators.\nNiCE Cognigy's Mobily case study, for example, lists billing questions first among the\nrepetitive requests Mobily's contact centres handled before it automated them. Bills combine prorated plan changes, roaming and\npremium charges, device instalments, discounts that expire and annual price rises, and the lines\noften come from different systems. The customer sees one total that moved and no explanation, so\nthey call.\n\nHuman agents then spend minutes opening the billing system, the order history and the tariff\nrules to reconstruct what happened. Answers can differ by agent, credits can be given\ninconsistently, and a customer who feels misled can become a complaint, a regulator escalation or\na churn risk. A chatbot without access to the customer's own bill data can only describe how\nbills work in general; it cannot explain this customer's charges.","problemStats":[],"howItWorks":"1. **Authenticate and fetch the bill.** After login or a one time passcode, the agent reads the\n   current and previous bills, recent orders and plan changes through read only billing APIs.\n2. **Explain the difference.** It compares the bills, identifies what changed (a prorated\n   upgrade, roaming, an ended discount, a price rise) and explains each line in plain language,\n   citing the tariff or contract term from approved content.\n3. **Fix what is clearly wrong, within limits.** Where a rule shows an obvious error, such as a\n   duplicate charge, the agent applies a correction up to a set amount and confirms it.\n4. **Open a dispute when judgment is needed.** Anything above the limit, contested or unclear\n   becomes a dispute case with the bill lines, the explanation given and the customer's reason,\n   routed to a billing specialist. Complaint signals follow the complaint process.\n5. **Hand over with context.** A human who takes over sees the bill analysis and the\n   conversation, so the customer does not explain again.","valueDrivers":["cost-to-serve","customer-experience","compliance","employee-productivity"],"kpis":["first-contact-resolution","automation-rate","containment-rate","response-time-reduction","interactions-handled","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A mobile and broadband operator with 5 million consumer customers","inputs":[{"key":"customers","label":"Consumer customers","low":5000000,"high":5000000,"unit":"customers","note":"The reference operator."},{"key":"billContactRate","label":"Billing contacts per customer per year","low":0.3,"high":0.6,"unit":"contacts per customer per year","note":"Editorial assumption, replace with the billing share of your contact reason report."},{"key":"resolvedShare","label":"Share of billing contacts the agent resolves without a human","low":0.3,"high":0.4,"unit":"fraction of billing contacts","note":"Editorial assumption. BT Group reports automation success approaching 50% on several unnamed types of Aimee journey, not specifically billing; replace with your own billing containment, keeping in mind that disputes and complaints must still reach people."},{"key":"costPerContact","label":"Cost of a human handled billing contact","low":4,"high":8,"unit":"USD per contact","note":"Editorial assumption for a blended chat and phone contact, replace with your own fully loaded cost."}],"formula":"customers * billContactRate * resolvedShare * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled billing contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of the AI and the billing integrations, credits the agent gives within its limits, fewer complaints and regulator escalations, and the retention effect of customers who understand their bill."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Explaining a bill well needs clean, read only access to billing, order and tariff data, which in many operators sit in several legacy systems. Corrections and disputes add write access and financial limits, which need their own controls and audit.","dataPrerequisites":["Structured bill data per line item for at least the last two bills","Current and historical tariff, discount and price rise rules in approved content","A billing contact reason report with volumes per cause","A written policy on which corrections the agent may make and up to what amount"],"integrations":["Billing and charging system (read, and write for corrections)","Order management and CRM for plan changes and history","Identity and authentication (app login, one time passcode)","Case or dispute management for billing disputes","Contact centre platform for handover with context"]},"implementation":{"steps":[{"title":"Map the top reasons bills change","detail":"From the contact reason report and a sample of calls, list the ten most common reasons a bill differs from the last one (prorating, roaming, expired discounts, price rises) and write the explanation and evidence for each."},{"title":"Build a bill comparison tool, not a prompt","detail":"Give the agent a function that returns the structured difference between two bills. The model explains; the numbers come from the billing system, never from the model's arithmetic."},{"title":"Set correction limits and dispute routing","detail":"Agree with finance which errors the agent may correct and up to what amount, and route everything else to a dispute case with the bill lines and the customer's reason attached."},{"title":"Ground every explanation in approved terms","detail":"Load current tariffs, contract terms and price rise notices with owners and review dates, and make the agent refuse rather than guess when a charge is not covered."},{"title":"Test on real bills","detail":"Replay anonymised bills with known causes, including edge cases such as mid cycle plan changes and roaming, and check each explanation against the correct answer before launch and on every change."},{"title":"Launch in the app, then widen","detail":"Start where customers are already logged in, measure first contact resolution and repeat contacts per cause, then add messaging and voice."}],"guardrails":["Figures and dates come only from billing system tools, never from model generated arithmetic","Corrections only within documented amount limits, with every credit logged","Complaint and vulnerability signals route to the complaint process and a human","Answers about terms only from approved tariff and contract content, with refusal when not covered","Masking of payment card data and personal data before text reaches a model or the logs"],"humanInTheLoop":"Billing specialists decide every dispute above the agent's limit and every contested charge. A quality team reviews a weekly sample of explanations against the bill data and signs off new correction rules and price rise explanations before they go live.","kpisToInstrument":["First contact resolution for billing contacts, counting a repeat billing contact within 30 days as unresolved","Share of billing conversations resolved without a human, per cause","Accuracy of explanations on a weekly checked sample","Credits issued by the agent, in count and value, against limits","Billing complaints and regulator escalations per 10,000 customers"],"failureModes":[{"title":"Confident wrong numbers","detail":"The model does its own arithmetic and explains a charge that does not exist. Keep all numbers in tools and test explanations against real bills."},{"title":"Explaining away a genuine error","detail":"The agent justifies an overcharge because a rule seems to allow it. Give it a clear path to open a dispute and measure disputes upheld later."},{"title":"Credits as a containment tactic","detail":"Goodwill credits used to end conversations cost more than the contacts saved. Log and cap credits, and review them weekly."},{"title":"Price rise conversations without care","detail":"Customers angry about a price rise need their rights explained, including any right to exit. Route those signals to trained people."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing assistant must be designed so that people know they are interacting with AI (Article 50(1)). Explaining bills, correcting clear errors and opening disputes are not listed in Annex III. The tier changes only if the system is also used to evaluate customers' creditworthiness, for example to set credit limits, which Annex III point 5(b) lists as high risk."},"regulations":["eu-ai-act","gdpr","telecom-consumer-rules","pci-dss","eecc"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"European Electronic Communications Code (Directive (EU) 2018/1972)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32018L1972","note":"Sets EU end user rights for electronic communications contracts, including contract information and consumption monitoring (Article 102(5)), the right to exit without cost when the provider changes contract conditions (Article 105(4)) and itemised billing (Annex VI Part A), which billing explanations must be consistent with."},{"title":"Quicker complaints resolution for telecoms customers, under new Ofcom rules","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/phones-and-broadband/service-quality/quicker-complaints-resolution-for-telecoms-customers-under-new-ofcom-rules","note":"From 8 April 2026, UK providers must tell customers about their right to independent dispute resolution when a complaint is still unresolved after six weeks (previously eight), so an agent must not hold a billing dispute in automation."}],"controls":["AI disclosure at the start of every conversation","Documented correction limits with finance sign off and an audit trail per credit","Regression tests on real bill scenarios for every change to prompts, tools or model","Complaint recognition that starts statutory complaint clocks and dispute resolution letters","Monthly review of disputes that were upheld after the agent's explanation"],"incidents":[{"title":"Incident 639: Air Canada chatbot reportedly provides inaccurate bereavement fare information, leading to customer overpayment","url":"https://incidentdatabase.ai/cite/639/","note":"A Canadian small claims tribunal ordered the airline to pay damages after its chatbot gave wrong fare information, a reminder that an operator owns every billing explanation its agent gives."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the billing, order and\ncase systems through REST, including a bill comparison function that returns structured\ndifferences so the model explains but never calculates. A **knowledge base** with hybrid\nretrieval holds only approved tariff and contract content. A **flow** runs corrections and\ndispute creation as fixed steps with amount limits, backed by custom functions that call the\noperator's identity and case systems.\n\nThe same agent runs in **web chat, WhatsApp and voice**, and inside the operator's mobile app\nthrough the **REST or WebSocket API channel**. **Guardrails** check input and output, **PII\nmasking** and card number tokenization run at the gateway, a flow gets the conversation history\n(optionally AI summarized) and the **agent handover** block escalates the conversation to a\nbilling specialist, who can also take over live from the admin console, and **test suites**\nreplay real bill scenarios on every change. Analytics dashboards with custom widgets track\noutcomes per billing cause, and the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"How many billing questions can an AI agent resolve?","answer":"It depends on how much of the bill it can see and explain. BT Group reports automation success approaching 50% on several types of journey in its EE assistant Aimee, without naming them, and says generative AI on the same platform gives detailed explanations of billing charges. Vodafone reports that initial SuperTOBi tests at one call centre showed a 50% improvement in first time resolution of critical journeys such as complex billing inquiries."},{"question":"Should the agent be allowed to give credits?","answer":"Only for clear errors, within an amount limit agreed with finance, with every credit logged. Contested or larger amounts should become a dispute case for a human, so the agent never uses credits to end a conversation."},{"question":"Is a billing assistant high risk under the EU AI Act?","answer":"No. Explaining bills and handling disputes is not listed in Annex III, so it is a limited risk system with an Article 50 duty to tell people they are talking to AI. It would need a new assessment if it were also used to evaluate customers' creditworthiness, for example to set credit limits, which Annex III lists as high risk."}],"related":["device-and-connectivity-troubleshooting-agent","complaints-handling-agent","churn-prediction-and-retention-offers","plan-upgrade-and-sales-assistant","collections-and-hardship-agent","utility-billing-and-move-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the telecommunications vertical from Vodafone, BT Group, Mobily and Verizon sources checked word for word."},{"date":"2026-09-25","note":"Consolidation pass: added European Electronic Communications Code to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact check: problem prose now traced to Mobily and Vodafone sources; EU AI Act basis corrected (telecom disconnection is not in Annex III, creditworthiness under point 5(b) is); FAQ no longer implies the BT Group automation rate is a billing result and states the Vodafone test scope; Ofcom note updated to the six week rule from 8 April 2026; EECC note names Articles 102, 105(4) and Annex VI; incident title aligned with the AI Incident Database; howToBuild limited to capabilities in the feature inventory; evidence channels, years and dates corrected; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Second fact check: howToBuild no longer names the authentication and complaint handling blocks (gated per deployment); resolved share capped at 0.4 because the BT Group rate is not a billing figure; Mobily sentence attributed to the NiCE Cognigy case study; Vodafone rollout beyond Italy and Portugal recorded as announced, not done; EECC Articles 102(5), 105(4) and Annex VI Part A rechecked against the EUR-Lex text."}],"slug":"bill-explanation-and-billing-dispute-agent","url":"https://www.blits.ai/ai-use-cases/bill-explanation-and-billing-dispute-agent","benchmarks":[{"kpi":"first-contact-resolution","label":"First contact resolution","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":60,"min":60,"max":60,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"vodafone-supertobi-generative-ai-assistant","pooled":true}]},{"kpi":"response-time-reduction","label":"Response time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":99.5,"min":99.5,"max":99.5,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"mobily-agentic-ai-self-service","pooled":true}]},{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bt-group-ee-aimee-virtual-assistant","pooled":false}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bt-group-ee-aimee-virtual-assistant","pooled":false}]}],"indicativeValueResult":{"low":1800000,"high":9600000},"evidence":["bt-group-ee-aimee-virtual-assistant","mobily-agentic-ai-self-service","verizon-ai-customer-experience-transformation","vodafone-supertobi-generative-ai-assistant"]},{"title":"AI agent for travel and expense report audit","shortTitle":"Travel and expense audit","seoTitle":"AI audit for travel and expense reports","metaDescription":"AI reviews every expense line against policy, not a sample. AppZen reports Takeda saved 4,000 auditor hours a quarter; Databricks found $483K in wasted spend.","definition":"An AI agent that checks every travel and expense report line against policy, receipts and prior submissions instead of a small manual sample, flags duplicates, altered receipts and policy violations with the evidence attached, and auto approves the clean majority so auditors spend their time on the reports that are genuinely risky.","aliases":["expense report auditing","T&E audit automation","AI spend audit","expense fraud detection"],"industries":["cross-industry","pharma-and-life-sciences","technology"],"functions":["finance-and-accounting"],"patterns":["document-processing","anomaly-detection","classification-and-routing","agentic-workflow"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"back-office","problem":"Travel and expense spend is high volume, low value per transaction, and hard to police at scale.\nTraditional audit teams cannot review every report line by line, so they sample: only expenses\nabove a value threshold, or a random percentage, get a real look. At Takeda,\nAppZen reports that at times 70 to 100% of expenses triggered audit rules, yet sampling and value\nthresholds still left spend uncovered and exposed to duplicate submissions and employee spend\nleakage that a person reviewing one report at a time cannot see across the whole population.\n\nMultinational organizations add language and currency variation across dozens of countries, each\nwith its own per diem and travel policy, on top of the volume problem. The result is a familiar\ntrade off: either spend more headcount on audit, which does not scale with travel volume, or\naccept that most spend goes unchecked and hope the exceptions surface some other way, for example\nwhen a manager happens to notice.","problemStats":[],"howItWorks":"1. **Capture every report.** Expense reports, receipts and the underlying corporate card\n   transactions feed into one pipeline from the T&E platform, so every line has a receipt image or\n   card record to check against, not just the ones a person opens.\n2. **Score every line against policy.** Each line is checked against the organization's written\n   policy by category, region and grade, rather than a single value threshold, and against the\n   employee's own submission history.\n3. **Cross check for duplicates and altered receipts.** The agent compares receipts across reports\n   and employees for duplicate submissions (including the same meal claimed by two attendees) and\n   checks receipt images for signs of alteration.\n4. **Auto approve the clean majority.** Lines that pass every check within policy limits approve\n   automatically; anything flagged goes to an auditor with the receipt, the rule it triggered and\n   similar past decisions shown together.\n5. **Learn from auditor decisions.** Confirmed and overturned flags feed back as candidate rule\n   adjustments, which a T&E or finance manager reviews and approves before they change what\n   auto approves.\n6. **Feed the policy owners.** Patterns by category, region or employee go to the corporate card\n   and travel policy teams, so recurring issues get fixed at the policy level, not flagged again\n   every month.","valueDrivers":["cost-to-serve","risk-reduction","compliance","employee-productivity"],"kpis":["automation-rate","hours-saved","cost-savings","interactions-handled"],"indicativeValue":{"referenceOrg":"A company with 7,000+ employees submitting about 130,000 expense reports a year","inputs":[{"key":"reportsPerYear","label":"Expense reports submitted per year","low":130000,"high":130000,"unit":"reports per year","note":"AppZen reports that before AppZen, Databricks' \"two auditors manually reviewed nearly 130K expense reports annually\" across \"7,000+ employees globally.\" Replace with your own report volume.","sourceUrl":"https://www.appzen.com/databricks-improved-travel-expense-management/"},{"key":"autoApprovalShare","label":"Share of reports the AI approves without a person","low":0.63,"high":0.72,"unit":"fraction of reports","note":"Range spans the auto approval rates AppZen reports for its two customers: 63% at [Takeda](https://www.appzen.com/resources/case-studies/how-takeda-is-transforming-global-expense-auditing-with-ai) across 63 countries, 72% at Databricks, where AppZen says the team kept adjusting its configuration month after month. Expect a lower rate before your own policy is tuned.","sourceUrl":"https://www.appzen.com/databricks-improved-travel-expense-management/"},{"key":"minutesPerReport","label":"Auditor minutes saved per auto approved report","low":1.9,"high":1.9,"unit":"minutes per auto approved report","note":"Derived from the hours AppZen reports Databricks saved, spread only across the reports that were auto approved: about 1.9 minutes per auto approved report (nearly 3,000 auditor hours saved a year across 130,000 reports at a 72% auto approval rate), which anchors this figure to the reference org above. This is time saved on the reports the AI clears, not a manual review time per report; no source gives that figure. AppZen also reports [Takeda](https://www.appzen.com/resources/case-studies/how-takeda-is-transforming-global-expense-auditing-with-ai) saving 4,000 auditor hours a quarter across \"400K expense audits annually\" at a 63% auto approval rate, but that figure is not used here: AppZen never states that an expense audit is the same unit as an expense report, so it cannot be safely converted into minutes per report and applied to the reference org's report count. Replace with your own time study.","sourceUrl":"https://www.appzen.com/databricks-improved-travel-expense-management/"},{"key":"costPerHour","label":"Fully loaded cost of a T&E auditor","low":25,"high":45,"unit":"USD per hour","note":"Editorial assumption for a blended onshore and offshore audit team."}],"formula":"reportsPerYear * autoApprovalShare * minutesPerReport / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Manual expense audit effort avoided","caveat":"Labour only. It leaves out the wasteful and non compliant spend actually caught (AppZen reports Databricks identified $483K in wasteful spend over twelve months), faster employee reimbursement, and the cost of the platform and policy configuration work."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Checking a receipt against a rule is straightforward; the work is encoding a policy that differs by category, region and grade into machine readable rules, and tuning thresholds so genuine risk is not buried under false positives on routine claims like meals and mileage.","dataPrerequisites":["Twelve months of history of audited expense reports with the outcome and reason","The written expense policy by category, region and grade, in a machine readable form","Card and travel booking feeds to cross check receipts against actual charges"],"integrations":["T&E and expense platform (SAP Concur, Expensify, Emburse, Ramp)","Corporate card and travel booking feeds","HR system for employee grade, location and manager","ERP or payroll for reimbursement posting"]},"implementation":{"steps":[{"title":"Baseline current coverage and findings","detail":"Record what share of reports get a real manual look today, what the sampling rule is, and what the audit team actually finds, so the AI is credited only for the increment."},{"title":"Encode the policy as machine readable rules","detail":"Turn the written expense policy into structured rules by category, region and grade; most early gaps come from policy that only exists as prose no one reads consistently."},{"title":"Start with the highest risk categories","detail":"Pick two or three categories with the most manual audit volume or the most prior findings, such as meals, mileage or entertainment, rather than trying to cover every category at once."},{"title":"Run in parallel with current sampling","detail":"Let the AI score every report next to the existing sample based process for a full close cycle, and compare what each approach would have flagged before changing anything live."},{"title":"Automate approval within limits","detail":"Auto approve only lines that pass every check within an agreed value limit; anything flagged, anything above the limit, and anything the model has not seen before goes to a person."},{"title":"Feed the policy and card teams","detail":"Route recurring flag patterns by category or employee to the people who own travel policy and the corporate card program, so the root cause gets fixed, not just the individual claim."}],"guardrails":["Every report checked against policy, not a value threshold sample; 100% coverage is the point","Only clean, policy compliant lines under an agreed value limit auto approve; anything flagged goes to a human auditor","Duplicate and altered receipt checks run on every submission before reimbursement","Any flag that could support disciplinary or employment action goes to a manager or HR review; the AI never decides or triggers that action itself"],"humanInTheLoop":"Auditors review every flagged line and decide whether to reject it, ask the employee for more information, or approve it. A T&E or finance manager approves changes to policy rules and thresholds, and reviews a sample of auto approved lines every month for drift.","kpisToInstrument":["Auto approval rate by category and region","Auditor hours per period","Wasteful or non compliant spend identified","Reimbursement cycle time","Repeat flag rate by employee and category"],"failureModes":[{"title":"False positives bury real risk","detail":"Too many low value flags on routine claims train auditors to rubber stamp the queue. Tune thresholds by category and track the override rate, not only the flag count."},{"title":"New patterns the model has not seen","detail":"A scheme built around the model's blind spot, such as a generated receipt image, goes through. Sample auto approved lines and periodically red team the checks with new patterns."},{"title":"Flags treated as verdicts","detail":"A flag is acted on as if it were a finding rather than a lead, which is unfair to the employee if the flag is wrong. Require a human decision and a documented reason before any action follows from a flag."}]},"risk":{"euAiAct":{"tier":"high","basis":"Annex III point 4(b) covers AI systems intended to monitor and evaluate the performance and behaviour of persons in a work related relationship. Scoring every line against the employee's own submission history, and instrumenting a repeat flag rate by employee, is that kind of behavioural evaluation, so this design falls under Annex III. The only carve out, Article 6(3), lets a narrow procedural or preparatory task escape high risk with a documented assessment, but the last subparagraph of Article 6(3) removes that carve out whenever the system performs profiling of natural persons. Scoring lines against an individual employee's history is profiling, so the carve out is not available here: keeping a human auditor as the actual decision maker on any personnel action is a required control, not an exit from Annex III."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"Annex III, point 4(b): employment, workers' management and access to self employment","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Lists AI systems intended to make decisions affecting the terms of a work related relationship, its promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate the performance and behaviour of persons in that relationship, as high risk."},{"title":"Article 6(3): the narrow task exception, and why it does not apply here","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"An Annex III system escapes high risk only if it performs a narrow procedural task, improves the result of a previously completed human activity, detects deviations from prior human decision making patterns without replacing or influencing the completed human assessment, or is preparatory. The last subparagraph closes this exception whenever the system profiles natural persons, which per employee, history based scoring does."}],"controls":["Every flag reviewed by a human auditor before it affects reimbursement or any personnel action","Documented, versioned policy rules with an owner and a review date","Consistency sampling across regions, expense categories and employee grades to catch uneven flagging","Full audit trail of every flag, the rule it triggered, and the human decision that followed"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that runs on a schedule, or that the T&E platform\ncalls through the workflow's API token when a new batch of reports is ready. A **custom\nfunction** fetches the new expense reports and the underlying card transactions over REST from\nthe T&E platform. An **AI agent** with **structured output** scores each line against the policy\nrules held in the **knowledge base**, organized into documents by category, region and grade and\nfound through **hybrid retrieval**, and further **custom functions** cross check the line\nagainst the corporate card feed and prior submissions in the ERP or expense platform (for\nexample through the SAP, Workday or Oracle connections in the integration catalog) for\nduplicates.\n\n**Human in the loop confirmation** holds any flagged or above threshold line for an auditor to\napprove or reject; their decision is logged for the next policy review. When an auditor needs\nmore information before deciding, that follow up with the employee happens outside the\nplatform. **PII masking** keeps employee personal data out of prompts sent to the model,\n**guardrails** run content checks on what the agent produces, **test suites** replay a labelled\nset of past reports before any policy change goes live, and **monitors** run scheduled checks\nagainst the agent. **Workflow run history with analytics and downloadable run data** shows auto\napproved versus held lines and the reasons behind each hold; feed in your own time tracking data\nto show auditor hours by period too. The platform is **model agnostic**, so the policy scoring\nmodel can be changed without rebuilding the workflow."},"faq":[{"question":"What share of expense reports can AI approve without a person?","answer":"AppZen reports a 63% auto approval rate across 63 countries at Takeda and a 72% auto approval rate at Databricks, where AppZen says the team kept adjusting its configuration month after month. Expect a lower rate at first, since the policy rules and thresholds need tuning against your own spend patterns before the model earns a wider auto approval limit."},{"question":"Does full coverage replace sampling?","answer":"It changes what sampling is for. Instead of choosing which reports get a real look, every report is checked against policy and prior patterns, and the sample becomes a quality check on the AI itself: auditors periodically review a slice of auto approved lines to catch drift, not a slice of all submissions to find the risky ones."},{"question":"Is expense audit AI high risk under the EU AI Act?","answer":"As designed here, yes. Annex III point 4(b) covers AI systems that monitor and evaluate employee behaviour, and scoring lines against an employee's own history is that kind of evaluation. The Article 6(3) narrow task exception cannot rescue it, because that exception never applies once a system profiles individual people, which per employee, history based scoring does. Keeping a human auditor as the actual decision maker on any personnel action is a required control under Annex III, not a way around it."},{"question":"What should stay with a person?","answer":"Any decision with disciplinary or termination consequences, ambiguous policy interpretation that a rule cannot capture, and cross border cases with tax or immigration implications."}],"related":["supplier-invoice-processing","procurement-spend-classification","internal-audit-copilot","continuous-controls-testing"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version. Evidence from AppZen customer stories for Takeda and Databricks, quotes checked against the live pages with usecases:source."},{"date":"2026-09-28","note":"Editorial pass: shortened metaDescription; recalibrated indicativeValue against the Databricks 130K report and 3,000 hour figures and fixed the referenceOrg employee count; attributed and restored the \"at times\" qualifier on the Takeda 70 to 100% figure; limited the month over month tuning claim to Databricks; corrected howToBuild to match FEATURE_INVENTORY.md (scheduled or API token triggered workflow, REST custom functions, no webhook trigger); reclassified euAiAct to high risk under Annex III point 4(b), since per employee, history based scoring closes the Article 6(3) exception, and updated the guidance and FAQ to match; dropped the professional services industry (no evidence covers it)."},{"date":"2026-09-28","note":"Adversarial review pass: limited human in the loop confirmation to approve or reject, matching FEATURE_INVENTORY.md section 6 (moved the \"request more information\" step outside the platform); credited PII masking alone with keeping personal data out of prompts, and gave guardrails their own role (content checks); rephrased the knowledge base line so it no longer implies metadata filtering by category, region and grade is a listed retrieval capability; attributed the \"improved month after month\" claim to AppZen's own narrative about the Databricks team, not a Databricks quote, in the autoApprovalShare note and the FAQ; relabelled minutesPerReport as minutes saved per auto approved report and rederived it from each company's own auto approval rate so the formula no longer double counts autoApprovalShare (1.9 to 3.8 minutes, replacing 1.4 to 2.4), and added the Takeda source link to both indicativeValue notes; removed the unsupported generalization that sampled checks usually stop at receipt existence; tightened the Article 6(3) guidance note to track the actual carve outs (narrow procedural task, improving a completed human activity, detecting deviations from prior human decision making patterns without replacing or influencing the human assessment, or preparatory)."},{"date":"2026-09-28","note":"Second adversarial review fixes: replaced the custom dashboard widgets claim in howToBuild, which FEATURE_INVENTORY.md does not support for a ranked list such as top flag reasons, with the workflow run history and analytics capability that is actually listed; and removed the Takeda derived 3.8 minute high end from the minutesPerReport input, which combined Takeda's per unit rate with Databricks' report count and roughly doubled the saving the reference org itself reported, anchoring the input to Databricks' own 1.9 minute figure instead. The Databricks evidence record's cost savings metric was also removed: the source says the $483K in wasteful spend was \"identified\", not saved, which does not meet the taxonomy's cost savings definition of cost saved per year net of the cost of running the AI, and no other KPI fits an identified but not confirmed saved figure."}],"slug":"travel-and-expense-audit-agent","url":"https://www.blits.ai/ai-use-cases/travel-and-expense-audit-agent","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":67.5,"min":63,"max":72,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"databricks-expense-audit-automation","pooled":true},{"id":"takeda-expense-audit-automation","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":3500,"min":3000,"max":4000,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"takeda-expense-audit-automation","pooled":true},{"id":"databricks-expense-audit-automation","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":400000,"min":400000,"max":400000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"takeda-expense-audit-automation","pooled":true}]}],"indicativeValueResult":{"low":64837.5,"high":133380},"evidence":["databricks-expense-audit-automation","takeda-expense-audit-automation"]},{"title":"AI agent for travel insurance claims and assistance","shortTitle":"Travel insurance claims and assistance","seoTitle":"AI agents for travel insurance claims","metaDescription":"AI agents take travel insurance claims, read receipts and route emergencies to people. Shift Technology reports a US travel insurer went from 0% to 57% automation.","definition":"An AI agent that helps insured travellers around the clock and in their own language: it answers cover questions, takes claims for delays, cancellations, lost baggage and medical costs, reads the receipts and certificates they upload, settles simple claims within set limits, and connects medical emergencies and complex cases to the assistance team at once.","aliases":["travel claims automation","travel assistance chatbot","trip cancellation and delay claims AI","AI for travel insurance claims"],"industries":["insurance","travel-and-hospitality"],"functions":["claims","customer-service"],"patterns":["conversational-agent","voice-agent","document-processing","agentic-workflow","translation"],"channels":["web-chat","whatsapp","mobile-app","voice","email"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"claims","problem":"Travel insurance claims are numerous and document heavy. A single trip can produce an\nairline delay certificate, hotel and taxi receipts, a baggage report, a foreign hospital invoice and\na doctor's note, in several languages and formats. Many insurers still review each claim by hand:\nthe US travel insurer in the evidence below handled around 400,000 claims a year manually, each\ntaking ten days to three weeks.\n\nAssistance is the other half. Travellers call from abroad, from other time zones, with a missed\nconnection, a lost passport or a medical emergency. Some calls are simple questions about cover and\nwhat to do next, but they can sit in the same queue as the emergencies that need a person\nimmediately, and a single storm, strike or airline disruption can affect many travellers at once.","problemStats":[{"statement":"Shift Technology states that only about 7% of insurance claims are processed straight through, because most claims data is unstructured and does not fit rules based systems.","sourceTitle":"AI in Action: From zero to 50%+ automation in travel insurance","sourceUrl":"https://www.shift-technology.com/resources/case-studies/genai-travel-insurance-automation","year":2025}],"howItWorks":"1. **Recognise the traveller and the policy.** The agent identifies the policy from the booking or\n   policy reference and confirms the trip details and the cover.\n2. **Triage the situation first.** Medical emergencies, safety threats and vulnerable travellers go\n   to the assistance team immediately; everything else continues with the agent.\n3. **Answer cover questions from the wording.** The agent explains what the policy covers for this\n   event, citing the policy wording, and tells the traveller what documents to keep.\n4. **Take the claim and read the documents.** It collects the facts, reads uploaded receipts and\n   certificates, checks them against the event (for example the flight delay) and the limits, and\n   lists anything missing.\n5. **Settle or hand over.** Simple claims within set limits and a clean fraud check are settled\n   automatically; others go to a handler with the documents already extracted and summarised.","valueDrivers":["customer-experience","cost-to-serve","speed","inclusion-and-access"],"kpis":["automation-rate","containment-rate","processing-time-reduction","accuracy","customer-satisfaction","interactions-handled"],"indicativeValue":{"referenceOrg":"A travel insurer handling 100,000 claims a year","inputs":[{"key":"claims","label":"Claims per year","low":100000,"high":100000,"unit":"claims per year","note":"The reference insurer."},{"key":"automationShare","label":"Share of claims decided without a handler","low":0.2,"high":0.4,"unit":"fraction of claims","note":"Conservative against the evidence on this page (Shift Technology reports that an unnamed US travel insurer raised its automation rate from 0% to 57%), because the mix of simple delay and baggage claims differs by portfolio."},{"key":"costPerClaim","label":"Internal cost of handling a simple claim by hand","low":30,"high":80,"unit":"USD per claim","note":"Editorial assumption covering handler time and document review; replace with your own claims expense data."}],"formula":"claims * automationShare * costPerClaim","currency":"USD","period":"per year","resultLabel":"Claims handling cost avoided","caveat":"Claims handling cost only. It leaves out assistance calls answered by the agent, faster payment and its effect on renewals and partner satisfaction, fraud impact, and the cost of the platform and integration."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The claim types are standardised and well suited to document extraction, but the agent must work in many languages and time zones, integrate with policy and claims systems that are often run by partners or assistance companies, and route medical emergencies without delay.","dataPrerequisites":["Policy wordings, limits and exclusions per product and distribution partner","Historical claims with documents and decisions, per claim type","Assistance protocols for medical, security and travel emergencies","Flight status or disruption data where delay claims are covered"],"integrations":["Policy administration and distribution partner systems for policy lookup","Claims management and payment systems","Assistance centre case management and telephony","Flight status or travel disruption data services","Fraud detection, including industry data sharing where available"]},"implementation":{"steps":[{"title":"Separate emergencies from everything else","detail":"Before automating anything, define the signals that send a traveller to a person at once (medical, safety, minors, vulnerability) and test them in every supported language."},{"title":"Automate the simplest claim types first","detail":"Travel delay, baggage delay and small cancellation claims have predictable documents and clear limits. Leave medical expenses abroad for a later wave."},{"title":"Ground cover answers in the policy wording","detail":"Load each product's wording and partner variations into the knowledge base, cite the clause in every answer, and refuse when the wording does not settle the question."},{"title":"Check documents against the event","detail":"Match receipts, dates and amounts to the insured event and the limits, and use external data, such as flight status, where it confirms the claim."},{"title":"Plan for disruption peaks","detail":"Airline strikes and storms can produce a sudden surge of claims and calls. Load test the agent and the document pipeline for those peaks."}],"guardrails":["Immediate handover to the assistance team for medical emergencies, safety threats and vulnerable travellers","Automatic settlement only inside limits per claim type with a clean fraud check","Cover answers cite the policy wording and never promise payment before the claim is assessed","Automatic declines avoided; unclear claims go to a handler with an explanation to the traveller","Medical and passport data masked in logs and prompts, with retention aligned to claims files"],"humanInTheLoop":"Assistance coordinators and medical teams handle every emergency. Claims handlers decide every claim outside the automatic limits and review a weekly sample of automatic settlements, and product owners approve each new product wording before the agent answers on it.","kpisToInstrument":["Time from emergency signal to a person on the line","Share of claims settled automatically, per claim type","Time from claim submission to payment","Accuracy of automatic decisions on an audited sample","Satisfaction after claims and assistance conversations, per language"],"failureModes":[{"title":"An emergency handled like a question","detail":"A traveller describing chest pain is asked for a policy number. Detect medical and safety signals first and test them in every language."},{"title":"Wrong wording for the product","detail":"Partner specific variations of cover are mixed up. Keep wordings per product and partner, and check which applies before answering."},{"title":"Automation that invites fraud","detail":"Fast payment of small claims attracts invented receipts and repeat claims. Keep fraud checks and industry data sharing in the flow."},{"title":"Collapse during mass disruption","detail":"The agent and the document pipeline are not sized for a strike or a storm. Load test for peaks and plan a degraded mode."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing agent must disclose that it is AI (Article 50), unless this is obvious from the context. Travel insurance claims handling is not listed in Annex III; point 5(c) covers risk assessment and pricing in life and health insurance, not the handling of claims. Handing a traveller who reports a medical emergency to the assistance team is not the classification of emergency calls or the patient triage in point 5(d), as long as the agent only hands over and does not set medical priorities. Claim decisions based solely on automated processing are subject to GDPR Article 22 (and its UK equivalent), and medical data is special category data under Article 9."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","iso-42001"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Travellers must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Addressed to national supervisors (August 2025); sets expectations for governance, fairness and human oversight of AI systems used across the insurance value chain, including claims."}],"controls":["AI disclosure at the start of every conversation, in the traveller's language","Tested emergency routing rules with response time monitoring","Documented automatic settlement limits per claim type under change control","Audit log of every automated decision with documents and rules applied","Regular review of answers per product wording and language"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** on **web chat, WhatsApp, email and voice**, and in the insurer's\nown mobile app through the **REST or WebSocket API channel**, with\n**automatic language detection** and multi language answers, and a **knowledge base** that holds the\npolicy wordings per product and partner, retrieved with hybrid search. **Flows** handle the\nemergency triage and the claim intake with **receive attachment** blocks for receipts and\ncertificates, and **custom functions** look up the policy, check flight status and create the claim.\n\nA **custom function** passes the uploaded attachments to the insurer's own document extraction\nservice, and an **agentic workflow** checks the extracted data against the event and the limits and\nuses **human in the loop approval** for any payment above the threshold. **Human handover**, including\nlive takeover on voice, connects emergencies to the assistance team immediately. **Guardrails**\nstop promises of payment, gateway **PII masking** can be configured with custom patterns for medical and passport data, and **test suites**\nreplay emergency and claim scenarios in every language on each change."},"faq":[{"question":"How far can travel insurance claims be automated?","answer":"Travel claims suit automation because claim types and documents are standardised. Shift Technology reports that an unnamed US travel insurer handling about 400,000 claims a year raised its automation rate from 0% to 57% (46% pay and 11% deny decisions), with 98% accuracy on pay decisions; the same vendor puts straight through processing across insurance claims at about 7%. Start with delay, baggage and small cancellation claims and leave medical expenses abroad for a later wave."},{"question":"Is fraud a bigger risk when travel claims are paid faster?","answer":"It can be, because fast payment rewards invented receipts and claims repeated across insurers. In Singapore, 25 insurers set up a shared data analytics initiative with the General Insurance Association of Singapore in 2017; according to a 2022 Shift Technology case study, its member insurers analyse travel and motor claims with Shift Technology's AI to find connections between people, providers and claims that can look genuine in isolation."},{"question":"Should an AI agent handle medical emergencies abroad?","answer":"Only to recognise them and connect the traveller to a person at once. The agent can collect the policy and location details in parallel, but decisions on treatment, evacuation and guarantees of payment belong to the assistance and medical teams."}],"related":["claims-first-notice-of-loss-agent","claims-triage-and-straight-through-processing","flight-disruption-and-rebooking-agent","claims-fraud-detection","insurance-policy-servicing-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the insurance vertical with evidence from an anonymized US travel insurer and the General Insurance Association of Singapore, verified against the sources. Public evidence from named travel insurers is still thin."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsourced claims about claim value and peak volumes, added a cited straight through processing statistic, corrected the evidence year to 2025, added UK GDPR and the Annex III point 5(d) reasoning, named the Singapore association in the FAQ, dropped the mobile app as a native channel, and added seoTitle and metaDescription."},{"date":"2026-09-26","note":"Second fact check against sources: both Shift Technology quotes, the 7% straight through statistic, the Singapore association page, the EIOPA opinion and the Annex III reasoning confirmed; reworded the metaDescription to follow the source's automation rate wording."},{"date":"2026-09-27","note":"Review fixes: the Blits.ai build now reads uploaded documents through a custom function that calls the insurer's own document extraction service, the PII masking claim follows the platform's configurable patterns, the Singapore FAQ dates the Shift Technology claim to its 2022 case study, the Article 50 basis includes the obvious from context exception, and unsourced wording about claim size and night calls was softened."},{"date":"2026-09-27","note":"Fact checked against sources; no changes needed. Rechecked both Shift Technology case studies, the 7% straight through statistic, the EIOPA opinion (including its reference to claims management), Article 50 and the Annex III reasoning, and the Blits.ai capabilities against the feature inventory."},{"date":"2026-09-27","note":"Added a named public evidence record: Allianz Partners' director of partnerships told Signature Travel Network's Horizon Club in April 2026 that AI assistance handles 65 to 70% of its claims and cut claims turnaround time from about 14 days to three to four days. This is the page's second named, public evidence record, and the first at grade B, alongside the General Insurance Association of Singapore record already on the page. Verified against both sources with zero mismatches. No other section contradicts the new evidence, so the rest of the page is unchanged."}],"slug":"travel-insurance-claims-and-assistance-agent","url":"https://www.blits.ai/ai-use-cases/travel-insurance-claims-and-assistance-agent","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"allianz-partners-ai-claims-turnaround","pooled":false}]}],"indicativeValueResult":{"low":600000,"high":3200000},"evidence":["allianz-partners-ai-claims-turnaround","gia-singapore-industry-fraud-analytics"]},{"title":"AI agent for utility billing, payments, meter readings and move in or move out","shortTitle":"Utility billing and home moves","seoTitle":"AI agents for utility billing and home moves","metaDescription":"AI agents explain energy bills, take meter readings and payments, and handle home moves. PolyAI reports 67% containment at PG&E; Octopus Energy's Arlo got 76% CSAT.","definition":"An AI agent for energy and water customers that explains bills and tariffs, takes meter readings, sets up or changes payments, and handles move in and move out (final reads, closing one account and opening the next), across phone, messaging, email and the app, while anyone in payment difficulty, in a vulnerable situation or with a complaint is handed to a person.","aliases":["energy billing chatbot","utility customer service agent","move in move out agent","meter reading assistant","energy supplier AI assistant"],"industries":["energy-and-utilities"],"functions":["customer-service","operations"],"patterns":["conversational-agent","voice-agent","agentic-workflow","rag-knowledge-assistant"],"channels":["voice","web-chat","mobile-app","whatsapp","email","sms"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Alongside outage calls, utility contact centres handle a steady stream of routine account tasks:\nwhy is my bill so high, when will I be charged, here is my meter reading, I am moving house. Each\none needs the customer's own account, meter and tariff data, which is why static FAQs and old IVR\nmenus rarely finish the job. Volumes are also spiky. PolyAI reports that PG&E saw daily calls in the\ntens of thousands during weather emergencies, and Aydem Energy says its call volumes rise sharply in\nseasonal peaks, faster than it can add staff.\n\nHome moves are among the most involved routine tasks. The supplier needs the right date and final\nreadings, has to close one account and open the next without a gap in supply, and often has to\ndeal with an unknown new occupant. Errors here can turn into estimated bills, back billing and\ncomplaints months later. Billing is already the largest complaint category at the UK Energy\nOmbudsman.","problemStats":[{"statement":"The UK Energy Ombudsman reports that billing related disputes remain the most common complaint category, accounting for 58% of the 46,532 cases it accepted in the first half of 2026.","sourceTitle":"Energy Ombudsman H1 Data 2026","sourceUrl":"https://www.energyombudsman.org/news/energy-ombudsman-h1-data-2026","year":2026}],"howItWorks":"1. **Identify the customer and the account.** The agent verifies the customer and finds the supply\n   points, meters and tariffs involved, matching an inbound phone number to the account where the\n   rules allow.\n2. **Explain from the customer's own data.** Bill questions are answered from the actual bill lines,\n   readings, tariff and payment history, not from generic content, and the agent shows how the\n   amount was calculated.\n3. **Take readings and payments.** It validates a meter reading against the expected range, asks for\n   a photo when the value looks wrong, and sets up or changes a payment date or amount within the\n   limits the supplier allows.\n4. **Run the move as a workflow.** For a move out or move in it collects the date, final or opening\n   readings and the forwarding address, closes or opens the account and confirms each step, with a\n   human check before anything irreversible.\n5. **Answer policy questions from approved content.** Tariff terms, price cap rules and support\n   schemes come from retrieval over the supplier's approved documents.\n6. **Hand over at the right moments.** Signs of payment difficulty or vulnerability, disputes,\n   complaints and safety issues (such as a gas smell) go straight to a person or the emergency line,\n   with the conversation attached.","valueDrivers":["cost-to-serve","customer-experience","speed","inclusion-and-access"],"kpis":["containment-rate","contact-deflection","interactions-handled","customer-satisfaction","customer-satisfaction-uplift","hours-saved"],"indicativeValue":{"referenceOrg":"An energy retailer with 1 million residential accounts","inputs":[{"key":"accounts","label":"Residential accounts","low":1000000,"high":1000000,"unit":"accounts","note":"The reference retailer."},{"key":"contactsPerAccount","label":"Assisted contacts per account per year","low":1,"high":2,"unit":"contacts per account per year","note":"Editorial assumption, replace with your own contact volume."},{"key":"inScopeShare","label":"Share of contacts about bills, payments, readings and moves","low":0.4,"high":0.6,"unit":"fraction of contacts","note":"Editorial assumption; billing is the largest complaint category at the UK Energy Ombudsman, but check your own contact reasons."},{"key":"containment","label":"Share of in scope contacts the agent resolves","low":0.25,"high":0.5,"unit":"fraction of in scope contacts","note":"Conservative against the evidence on this page (PolyAI reports 67% containment at PG&E and Aydem Energy reports 75% of WhatsApp inquiries resolved), because moves and disputes are harder than FAQs."},{"key":"costPerContact","label":"Cost of a human handled contact","low":4,"high":8,"unit":"EUR per contact","note":"Editorial assumption for a blended phone, email and chat contact. Replace with your own fully loaded cost."}],"formula":"accounts * contactsPerAccount * inScopeShare * containment * costPerContact","currency":"EUR","period":"per year","resultLabel":"Human handled contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out the cost of the AI and the billing system integration, fewer estimated bills and complaints from better readings and cleaner moves, and the peak capacity the agent adds during outages and price changes."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering is simple; acting on the account is where the work lies. Moves touch the billing system, meter data, industry registration processes and sometimes a credit check, and every write needs clear limits and a way back.","dataPrerequisites":["Account, bill line, tariff and payment data reachable through APIs","Meter data with expected reading ranges per meter","Approved tariff terms, regulatory rules and support scheme content","Vulnerability and priority services flags with the rules for using them","Contact reason data to choose the first intents"],"integrations":["Billing and customer information system","Meter data management","Payments and direct debit platform","Industry registration or switching processes for moves","Contact centre platform for handover and callbacks"]},"implementation":{"steps":[{"title":"Start with readings, bill explanations and payment dates","detail":"These are high volume, low risk and easy to check. Octopus Energy started its email assistant on tariff renewals, payment dates and account details and kept complex complaints, sensitive cases and vulnerable customers with people."},{"title":"Ground every bill answer in the account","detail":"Give the agent tools that return the bill lines, readings and tariff for this customer, and make it show the calculation. Generic answers about \"how bills work\" do not reduce repeat contact."},{"title":"Treat moves as a workflow with checkpoints","detail":"Model move in and move out as steps with validation (date, readings, address) and a confirmation at each step, and keep a human approval for account closure until error rates are known."},{"title":"Build vulnerability detection into the flow","detail":"Define the phrases and signals (missed payments, health conditions, distress) that route to a specialist, and check the Priority Services Register or its local equivalent before any change."},{"title":"Plan for the peak","detail":"Outages and price changes bring the surges. Make sure the agent can carry outage updates at volume, as Aydem Energy and PG&E do, so billing questions still get through."}],"guardrails":["No disconnection, debt or payment plan decisions by the agent; those go to trained staff","Payment changes only within limits the supplier sets, with confirmation to the customer","Readings outside the expected range rejected or sent for review, never billed as given","Safety issues such as a gas smell routed to the emergency line immediately","AI disclosure on every channel and an easy route to a person"],"humanInTheLoop":"Specialists handle payment difficulty, vulnerable customers, disputes, complaints and failed moves. Operations leads approve each new intent and action, and a sample of contained conversations and AI written emails is reviewed every week. Octopus Energy says its human experts keep checking the messages its assistant sends.","kpisToInstrument":["Containment per intent, counting repeat contact within seven days as not contained","Estimated bills and back billing cases after agent handled moves","Customer satisfaction for AI handled versus human handled contacts","Handover rate and reasons, including vulnerability referrals","Complaints that mention the assistant"],"failureModes":[{"title":"Wrong reading, wrong bill","detail":"A mistyped or misread meter value flows into billing. Validate against expected ranges and ask for a photo when in doubt."},{"title":"Broken moves","detail":"The old account is closed before the new one is confirmed, or readings are missing. Use a workflow with checkpoints and human approval for closures."},{"title":"Missing vulnerability","detail":"A customer in difficulty is handled like any other and put on a payment plan they cannot keep. Detect signals early and hand over."},{"title":"Generic bill explanations","detail":"The agent explains how bills work in general instead of this bill, and the customer calls anyway. Give it the bill data and measure repeat contact."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"A customer service agent for bills, readings and moves falls under the transparency duty for systems that interact with people (Article 50(1)): customers must be told they are talking to AI. If the agent assesses creditworthiness, for example to set a deposit when a new customer moves in, that part falls under Annex III point 5(b) and is high risk; keep credit decisions in separately governed systems. The agent is not a safety component in the operation of the gas, water or electricity supply (Annex III point 2), so safety reports such as a gas smell go straight to the emergency line rather than being handled by the agent."},"regulations":["eu-ai-act","gdpr","uk-gdpr"],"guidance":[{"title":"Ethical AI use in the energy sector","issuer":"Ofgem","region":"europe","url":"https://www.ofgem.gov.uk/guidance/ethical-ai-use-energy-sector","note":"Good practice guidance for energy companies, updated in May 2026, including a section on AI in consumer interactions with transparency and proportionate explainability."}],"controls":["AI disclosure and a documented route to a human on every channel","Action allow list with limits for payments, readings and account changes","Audit trail of every account change the agent made, with the verification used","Vulnerability and priority services checks before any change","Change control and regression tests for each new intent"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the billing, meter data\nand payment systems over REST, so every bill answer is grounded in the customer's own data. Tariff\nterms and support schemes sit in a **knowledge base** with hybrid retrieval. Moves run as an\n**agentic workflow** with validation steps and **human in the loop** approval before an account is\nclosed, while a **flow** handles verification and the fixed steps of a meter reading.\n\nThe same agent serves **voice**, **WhatsApp**, **SMS**, **email**, **web chat** and the supplier's\nown app through the **REST API channel**, which lets outage and price change peaks move to messaging. **Guardrails** stop the agent from making\npayment plan or disconnection decisions, **PII masking** protects account data, and **human\nhandover** routes payment difficulty and vulnerability to specialists. **Test suites** replay moves\nand bill questions on every change, and **analytics** show containment and handover reasons per\nintent."},"faq":[{"question":"Do customers accept AI answers from their energy supplier?","answer":"Early evidence suggests so, when the AI is transparent and a human is one reply away. Octopus Energy reports that its email assistant Arlo scored 76% customer satisfaction in a trial, against 72% for comparable human replies, with every AI written email labelled."},{"question":"What share of utility calls can an AI agent resolve?","answer":"PolyAI reports 67% containment for PG&E's voice agent, which handles outage and billing calls, and Aydem Energy reports that 75% of the inquiries reaching its WhatsApp assistant are resolved without an agent. Moves and disputes are likely to resolve less often than outage updates and FAQs."},{"question":"Should the agent handle move in and move out end to end?","answer":"It can collect everything and run the steps, but keep a human check before closing an account until you know the error rate. Public outcome data for AI handled moves is still scarce; PolyAI lists start and stop service at PG&E as a use case being built, not as a result."},{"question":"Can AI help with bill confusion without a chatbot?","answer":"Yes. EDF's AI Bill Explainer breaks each bill down step by step in the customer account, and EDF reports 5.6% fewer billing related contacts among customers who used it in the pilot."}],"related":["bill-explanation-and-billing-dispute-agent","collections-and-hardship-agent","outbound-reminder-and-confirmation-agent","first-line-contact-centre-agent","complaints-handling-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with evidence from Octopus Energy, EDF, PG&E, Aydem Energy, DEWA and an anonymous Kraken deployment."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: problem prose and Energy Ombudsman statistic tied to their sources (first half of 2026), PG&E CSAT uplift limited to outage calls, EDF stage set to production, EU AI Act basis made precise (Article 50(1), Annex III point 2), UK GDPR added, SEO title and description added."}],"slug":"utility-billing-and-move-agent","url":"https://www.blits.ai/ai-use-cases/utility-billing-and-move-agent","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":3,"nUpTo":0,"median":8000,"min":1000,"max":9200000,"byClaimant":{"organization":1,"vendor":2,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dewa-rammas-customer-chatbot","pooled":true},{"id":"octopus-energy-arlo-email-assistant","pooled":true},{"id":"aydem-energy-whatsapp-assistant","pooled":true}]},{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":71,"min":67,"max":75,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"aydem-energy-whatsapp-assistant","pooled":true},{"id":"pge-polyai-voice-agent-billing-and-outages","pooled":true}]},{"kpi":"contact-deflection","label":"Contact deflection","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":5.6,"min":5.6,"max":5.6,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"edf-ai-bill-explainer","pooled":true}]},{"kpi":"customer-satisfaction","label":"Customer satisfaction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":76,"min":76,"max":76,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"octopus-energy-arlo-email-assistant","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":35000,"min":35000,"max":35000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"pge-polyai-voice-agent-billing-and-outages","pooled":true}]},{"kpi":"customer-satisfaction-uplift","label":"Satisfaction uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":22,"min":22,"max":22,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"pge-polyai-voice-agent-billing-and-outages","pooled":true}]}],"indicativeValueResult":{"low":400000,"high":4800000},"evidence":["aydem-energy-whatsapp-assistant","dewa-rammas-customer-chatbot","edf-ai-bill-explainer","octopus-energy-arlo-email-assistant","pge-polyai-voice-agent-billing-and-outages"]},{"title":"AI ambient scribe for clinical documentation","shortTitle":"Ambient clinical documentation","seoTitle":"Ambient AI scribes for clinical documentation","metaDescription":"Ambient AI scribes draft the clinical note from the visit for the clinician to review. Permanente physicians used them in more than 2.5 million patient encounters.","definition":"An AI scribe that listens, with the patient's consent, to the conversation between a clinician and a patient and drafts the clinical note, and often the letter or after visit summary, for the clinician to review, edit and sign in the health record. It documents; it does not diagnose or decide on treatment.","aliases":["ambient AI scribe","AI medical scribe","ambient voice technology","ambient listening for clinical notes","AI clinical note taking"],"industries":["healthcare"],"functions":["operations","knowledge-management"],"patterns":["speech-analytics","summarization","content-generation"],"channels":["internal-tools","mobile-app"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"mainstream","problem":"Clinicians spend a large share of every working day on documentation. During the visit they type\nwhile the patient talks, and after clinic they finish notes and letters in the evening, the time\nclinicians call \"pajama time\". Patients notice the screen between them and their doctor, and health\nsystems that deploy scribes name the documentation load as a driver of clinician burnout.\n\nHuman scribes and dictation help, but they are expensive or still take clinician time. Generative\nAI changed the economics: speech recognition that copes with a real consultation, followed by a\nlanguage model that turns the conversation into a structured note in the clinician's preferred\nformat. The risk moved with it. A fluent note that contains something nobody said, or leaves out a\nsymptom, ends up in the medical record unless the clinician catches it.","problemStats":[],"howItWorks":"1. **Ask for consent.** The clinician tells the patient that an AI scribe will listen and records\n   the consent; the patient can decline or stop it at any time, even mid visit.\n2. **Capture the conversation.** A phone, tablet or workstation app records the visit (in person,\n   phone or video) and streams it to speech recognition with speaker separation.\n3. **Draft the note.** A language model turns the transcript into a note in the clinician's template\n   and specialty format (history, examination, assessment and plan), and optionally a letter,\n   patient instructions or suggested codes.\n4. **Review and sign.** The draft appears in the health record or next to it; the clinician checks\n   it against what happened, edits it and signs it. Nothing enters the record unreviewed.\n5. **Learn from edits.** Edit rates, clinician feedback and quality samples show where the drafts\n   are weak, per specialty and per template.","valueDrivers":["employee-productivity","customer-experience","speed","cost-to-serve"],"kpis":["time-saved-per-task","hours-saved","handling-time-reduction","productivity-gain","users-served","interactions-handled","employee-adoption"],"indicativeValue":{"referenceOrg":"A health system with 1,000 clinicians using an ambient scribe","inputs":[{"key":"clinicians","label":"Clinicians using the scribe","low":1000,"high":1000,"unit":"clinicians","note":"The reference health system."},{"key":"encountersPerClinician","label":"Documented encounters per clinician per year","low":2000,"high":3000,"unit":"encounters per clinician per year","note":"Editorial assumption for outpatient and primary care clinicians. Replace with your own visit volumes."},{"key":"minutesSaved","label":"Documentation minutes saved per encounter","low":0.2,"high":0.5,"unit":"minutes per encounter","note":"Derived from The Permanente Medical Group analysis on this page, with two assumptions of our own: a working day of 8 hours, and that the saving of 1,794 working days \"in one year\" can be set against the more than 2.5 million encounters counted over the 63 week evaluation. That gives about 0.34 minutes per encounter on average (about 0.42 if the encounters are scaled to 52 weeks). The top third of users accounted for 89% of activations, and high users saved two and a half times more per note than infrequent users, so the average is close to what frequent users saved (roughly 0.35 to 0.45 minutes); occasional users saved less. The range stays around that evidence. Replace with your own time data."},{"key":"costPerHour","label":"Fully loaded cost per clinician hour","low":100,"high":150,"unit":"USD per hour","note":"Editorial assumption. Replace with your own fully loaded clinician cost."}],"formula":"clinicians * encountersPerClinician * minutesSaved / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Clinician documentation time released","caveat":"Time released is not cash saved unless it becomes extra appointments or less overtime. The figure leaves out licence and integration costs, clinician review time for drafts, the effect on burnout and retention, and any change in coding completeness."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Mature products exist, so the work is in integration with the health record, consent and information governance, clinical safety assessment, device and network setup in clinics, and training clinicians to review drafts properly. Specialty templates and languages take tuning.","dataPrerequisites":["Note templates and documentation standards per specialty","A consent process and patient information text","A clinical safety case and a data protection impact assessment","A sample of real consultations (with consent) to test draft quality per specialty"],"integrations":["Electronic health record for patient context and filing the signed note","Clinician devices (mobile app, desktop, dictation hardware)","Identity and single sign on for clinicians","Audit logging and records retention for audio and transcripts"]},"implementation":{"steps":[{"title":"Start with willing clinicians in a few specialties","detail":"At The Permanente Medical Group in Northern California, mental health, emergency medicine and primary care doctors were the most likely to use the scribe. Pick specialties with long conversations and heavy notes, and clinicians who want the tool."},{"title":"Settle consent, retention and safety before go live","detail":"Decide how consent is asked and recorded, whether audio is kept or deleted after the note is signed, and complete the clinical safety assessment and data protection impact assessment."},{"title":"Tune templates per specialty","detail":"Build note formats with clinicians in each specialty and let individual clinicians adjust style, so the draft needs editing rather than rewriting."},{"title":"Train clinicians to review, not to trust","detail":"Teach what the tool gets wrong (medication names, negations, who said what) and make clear that the signature means the clinician has checked the note."},{"title":"Measure time and quality, then widen","detail":"Track documentation time, after hours time, edit rates and patient feedback against a baseline, and expand to more specialties once quality holds."}],"guardrails":["No note enters the record without clinician review and signature","Patient consent recorded for every encounter, with an easy way to decline or stop","The scribe documents only; it does not suggest diagnoses or place orders without clinician action","Audio and transcripts retained only as long as policy allows, with access logging","Health data processed in approved regions under HIPAA or GDPR special category rules"],"humanInTheLoop":"The clinician reviews, edits and signs every note and remains accountable for the record. A clinical safety officer owns the risk log, and a quality team samples signed notes against transcripts to find omissions and invented content.","kpisToInstrument":["Documentation time per encounter and after hours time in the record, before and after","Share of encounters where the scribe is used, per clinician and specialty","Edit distance between draft and signed note","Omissions and invented content found in quality samples","Patient consent and decline rates, and patient feedback"],"failureModes":[{"title":"Invented or misattributed content","detail":"The draft contains a symptom, medication or statement nobody said, or attributes the patient's words to the clinician. Sample notes against transcripts and teach clinicians what to check."},{"title":"Automation complacency","detail":"Clinicians sign drafts after a glance because they are usually right. Measure review time and edit rates and make quality feedback visible."},{"title":"Low use after launch","detail":"Most of the time savings go to frequent users; occasional users gain little. Support adoption with training and specialty templates rather than counting licences."},{"title":"Consent that is not real","detail":"Patients are not told clearly or feel they cannot refuse. Script the consent, make declining easy and track decline rates."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"A scribe that only transcribes and summarises for a clinician to review is not listed in Annex III and is usually minimal risk, although the provider of a system that generates text can still owe the Article 50(2) duty to mark output as AI generated, unless an exception such as an assistive function for standard editing applies. If the product qualifies as medical device software under the EU Medical Device Regulation and needs a notified body assessment, for example because it suggests diagnoses or treatment, it becomes high risk under Article 6(1) and Annex I. Health data in audio and notes falls under GDPR Article 9 in every case."},"regulations":["eu-ai-act","gdpr","uk-gdpr","hipaa","nist-ai-rmf","iso-42001"],"guidance":[{"title":"MHRA clarifies regulatory status of ambient voice technologies used in the NHS","issuer":"Medicines and Healthcare products Regulatory Agency","region":"europe","url":"https://www.gov.uk/government/news/mhra-clarifies-regulatory-status-of-ambient-voice-technologies-used-in-the-nhs","note":"Confirms that products used solely for transcription, summarising consultations, drafting letters or suggesting codes for clinician review are not regulated as medical devices in Great Britain, while products that support diagnosis or treatment, or act without clinician review, are."},{"title":"Article 6, classification rules for high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/6/","note":"An AI system that is, or is a safety component of, a product covered by EU harmonisation legislation such as the Medical Device Regulation and needs third party conformity assessment is high risk."},{"title":"Regulation (EU) 2017/745 on medical devices","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2017/745/oj","note":"Decides whether a scribe with clinical functions is medical device software, and its risk class."}],"controls":["Clinical safety case and hazard log for the scribe, owned by a named clinical safety officer","Data protection impact assessment covering audio, transcripts and vendor processing","Consent procedure and patient information in plain language","Audit trail of drafts, edits and signatures per note","Periodic quality sampling of signed notes against source audio or transcripts"],"incidents":[{"title":"OpenAI's transcription tool hallucinates more than any other, experts say, but hospitals keep using it","url":"https://fortune.com/2024/10/26/openai-transcription-tool-whisper-hallucination-rate-ai-tools-hospitals-patients-doctors/","note":"An Associated Press investigation reported that the Whisper speech model can invent text, and that a Whisper based medical transcription tool from Nabla, used by over 30,000 clinicians and 40 health systems, deletes the original audio, so transcripts cannot be checked against the recording. Nabla said clinicians must edit and approve notes."}]},"blitsAi":{"howToBuild":"On Blits.ai the documentation step is an **agentic workflow** that the organization's clinic app\ntriggers through the **REST API** after each consented visit. The audio is transcribed with\n**self hosted transcription and speaker diarization** (WhisperX), which keeps \"who spoke when\" and\ncan run on Blits.ai infrastructure for data sovereignty, or with one of the supported speech to text\nproviders. WhisperX is based on Whisper, the model family in the incident on this page, so keep the\naudio or transcript available for checking drafts.\nAn **AI agent** with **structured output** turns the transcript into the note sections of the\nchosen template, and **custom functions** return the draft to the record system for review.\n\nThe clinician reviews and signs the note in the health record, not in Blits.ai; the workflow only\nreturns a draft. **PII masking** limits the health\ndata that reaches a model, **audit trails** record every run, and **test suites** grade drafts\nagainst reviewed notes per specialty before each change goes live. The platform is model agnostic,\nand EU and UAE data residency keeps audio and notes in region."},"faq":[{"question":"How much time do ambient AI scribes save?","answer":"The Permanente Medical Group in Northern California reports that AI scribes saved its physicians the equivalent of 1,794 working days in one year; over a 63 week evaluation they were used in more than 2.5 million encounters, and high users saved two and a half times more time per note than infrequent users. In an NHS England sponsored study led by Great Ormond Street Hospital, appointments were 8.2% shorter and A&E clinicians saw 13.4% more patients per shift."},{"question":"Is an AI scribe a medical device?","answer":"It depends on what it does. The UK MHRA confirmed in July 2026 that tools used solely to transcribe, summarise, draft letters or suggest codes for clinician review are not medical devices, while tools that support diagnosis or treatment, or act without review, are. In the EU, a scribe that qualifies as medical device software under the Medical Device Regulation and needs a notified body is also high risk under the AI Act."},{"question":"What are the main safety risks?","answer":"Invented content, omissions and misattributed statements in a note that looks complete. Keep clinician review and signature mandatory, sample signed notes against transcripts, and be careful with tools that delete the audio before anyone can check the transcript."}],"related":["medical-coding-automation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with evidence from Kaiser Permanente, Great Ormond Street Hospital and the Veterans Health Administration, verified against the sources."}],"slug":"ambient-clinical-documentation","url":"https://www.blits.ai/ai-use-cases/ambient-clinical-documentation","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":1258500,"min":17000,"max":2500000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"kaiser-permanente-ambient-ai-scribes","pooled":true},{"id":"great-ormond-street-hospital-ai-scribe-trial","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":8.2,"min":8.2,"max":8.2,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"great-ormond-street-hospital-ai-scribe-trial","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":13.4,"min":13.4,"max":13.4,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"great-ormond-street-hospital-ai-scribe-trial","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":7260,"min":7260,"max":7260,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"kaiser-permanente-ambient-ai-scribes","pooled":true}]}],"indicativeValueResult":{"low":666666.6666666667,"high":3750000},"evidence":["great-ormond-street-hospital-ai-scribe-trial","kaiser-permanente-ambient-ai-scribes","veterans-health-administration-ambient-scribe"]},{"title":"AI analytics for smart meter and AMI data","shortTitle":"Smart meter analytics","seoTitle":"AI analytics for smart meter data","metaDescription":"Con Edison used C3 AI to monitor its 5.3 million smart meter deployment; Southern California Gas Company used Bidgely for digital home energy reports.","definition":"AI that turns the flood of readings from smart electricity, gas and water meters into usable information: it monitors meter and network health at scale, estimates which appliances drive a household's usage from the meter signal alone, flags unusual consumption, and targets efficiency and electrification programmes at the customers who will benefit most, instead of a utility treating every meter and every customer the same way.","aliases":["AMI analytics","energy disaggregation","meter data analytics","advanced metering infrastructure analytics"],"industries":["energy-and-utilities"],"functions":["operations","analytics-and-reporting"],"patterns":["anomaly-detection","prediction-and-scoring"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"assist","adoptionStage":"early-adopters","segment":"metering-and-billing","problem":"Con Edison's rollout of 5.3 million smart meters was expected to generate between 100 terabytes\nand 1 petabyte of data annually, far more than any team can review reading by reading. Buried in\nthat volume are the installation and configuration issues that leave a meter or its network module\nunhealthy, which is what Con Edison's deployment set out to find. More generally, the same kind of\ndata holds the individual appliances, such as electric vehicles, heat pumps and air conditioning,\nthat a utility needs to understand as adoption grows, and the customers who would benefit most\nfrom an efficiency programme but are hard to identify from billing data alone.","problemStats":[],"howItWorks":"1. **Ingest at scale.** Readings stream from millions of meters into a data platform, alongside\n   the utility's asset, billing and programme data.\n2. **Watch meter and network health.** Machine learning flags meters and communication modules\n   showing signs of a deployment or configuration problem, so field crews fix the ones that\n   actually need attention.\n3. **Disaggregate usage.** A separate model estimates which appliances, such as HVAC, water\n   heating, electric vehicle charging or pool pumps, are driving each home's consumption from the\n   meter signature alone, without a sensor on the appliance itself.\n4. **Segment and target.** The disaggregated data groups customers by what is actually happening\n   in their home, such as households with an electric vehicle or an ageing HVAC system, so an\n   efficiency, demand response or electrification programme can be targeted instead of broadcast\n   to everyone.\n5. **Feed operations and customer teams.** A prioritised list, a dashboard or a personalised\n   message reaches the team or the customer, closing the loop from raw meter data to action.","valueDrivers":["cost-to-serve","customer-experience","employee-productivity"],"kpis":["energy-savings","error-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A utility with 2 million smart meters and an energy efficiency programme budget","inputs":[{"key":"meters","label":"Smart meters in the analytics programme","low":1500000,"high":2500000,"unit":"smart meters","note":"Range set around the 2 million meter reference organization, well within Con Edison's 5.3 million meter deployment."},{"key":"issueRate","label":"Share of meters with a deployment or health issue found by analytics each year","low":0.005,"high":0.02,"unit":"fraction of meters","note":"Editorial assumption, replace with your own meter health data."},{"key":"costPerIssue","label":"Cost saved per flagged issue by folding it into a planned visit instead of a repeat or emergency visit","low":80,"high":200,"unit":"USD per issue","note":"Editorial assumption for the difference between a planned and an emergency or repeat US utility field visit; replace with your own figure. Neither deployment on this page reports a per issue avoided cost."}],"formula":"meters * issueRate * costPerIssue","currency":"USD","period":"per year","resultLabel":"Avoided field visit and rework cost from early meter health detection","caveat":"Meter health savings only, and it assumes a flagged issue can be folded into a planned visit rather than always triggering a separate site visit, which is an editorial assumption neither deployment on this page confirms. It also leaves out the analytics platform and integration cost, and any separate value from the efficiency and electrification programmes the same data supports, which neither deployment on this page reports as a single company wide figure."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Con Edison's deployment aggregated two years of data from 13 source systems covering 5 million customer accounts into an integrated data image; that integration work is typically harder than the machine learning itself.","dataPrerequisites":["Interval meter reads at the frequency the advanced metering infrastructure supports","Meter and network asset data, such as install date, model and communication module","Customer and premise data linking a meter to a household or business"],"integrations":["Meter data management system","Geospatial information system for network topology","Customer information system or CRM, for targeting and outreach","Demand side management or efficiency programme platforms"]},"implementation":{"steps":[{"title":"Start with meter and network health, not customer analytics","detail":"Con Edison's first phase covered deployment and installation issues plus meter and network health, using two machine learning algorithms and 50 analytics; the company planned further customer insight and distribution and transmission automation applications for later phases."},{"title":"Unify the source systems before modelling","detail":"Con Edison's deployment integrated two years of data across 13 source systems before the machine learning models were configured; budget and staff for that integration effort, not just the model."},{"title":"Disaggregate before you personalise","detail":"Appliance level disaggregation from the meter signal is what lets a programme target the households with an electric vehicle or an ageing HVAC system, rather than mailing everyone the same offer."},{"title":"Close the loop into an actual programme","detail":"Southern California Gas Company's deployment fed a digital only home energy report programme for its medium consumption gas customers, which exceeded its savings goal, not a dashboard; without a programme on the receiving end, better analytics changes nothing for the customer."},{"title":"Scale meter by meter, programme by programme","detail":"Widen coverage and add new applications, such as electrification planning or peak forecast, once the foundational data platform is proven."}],"guardrails":["Customer usage data used for targeting is handled under the utility's own data privacy and retention rules","A person reviews the priority list before a field crew is dispatched based on a flagged meter","Model changes are tested against historical data before being applied to live operations"],"humanInTheLoop":"Operations analysts triage the prioritised list of flagged meters before a field crew is dispatched, and programme managers decide which segments an efficiency or electrification campaign actually targets based on the disaggregated data.","kpisToInstrument":["Meters flagged with a health or deployment issue, and the share confirmed correct on inspection","Programme enrollment and savings among AI targeted customers versus a general mailing","Time from a flagged issue to resolution"],"failureModes":[{"title":"A flood of technically correct but unprioritised flags","detail":"Millions of meters can produce more anomalies than any team can act on; rank by impact and route only the highest priority batches to a person. Con Edison's application produces a prioritised list of meters that require attention."},{"title":"Disaggregation that does not hold up across meter and appliance types","detail":"Appliance signatures vary by region, climate and equipment age; validate disaggregation accuracy on a local sample before using it to target a programme at scale."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 2 covers AI systems intended to be used as a safety component in the management and operation of critical digital infrastructure and the supply of water, gas, heating or electricity. Meter health prioritisation and usage disaggregation for programme targeting are not intended as safety components, so they stay outside that scope regardless of whether a person reviews the output. The tier would instead be high risk if the same kind of analytics were intended as a safety component in network operation or supply, for example directly controlling grid or metering protection systems; a human in the loop is then an Article 14 obligation for that high risk system, not a way to fall outside the category."},"regulations":["eu-ai-act","gdpr","nis2"],"guidance":[],"controls":["A documented separation between advisory analytics and any system that can act on grid protection or metering infrastructure directly","Data minimisation and retention limits on the household level usage data used for disaggregation and targeting"],"incidents":[]},"blitsAi":{"howToBuild":"The meter data platform, the anomaly detection and the appliance disaggregation models are\nspecialist utility analytics products; Blits.ai is not where you build an AMI operations or\ndisaggregation model. What Blits.ai adds is the layer operations and programme teams use to act\non the output: an SQL knowledge base over the meter health and disaggregation results lets an\nagent answer a programme manager's question, such as which customers in a service area show a\nnew electric vehicle signature, in plain language instead of a query the analytics team has to\nrun by hand.\n\nAgentic tasks can watch for a condition, such as a batch of meters newly flagged as unhealthy in\na service area, and draft the field work order with human in the loop confirmation before a crew\nis dispatched. Agentic workflow run history and custom analytics dashboard widgets track flagged\nvolumes and resolution over time, and the platform's model agnostic routing and EU and UAE data\nresidency fit a utility that must keep customer usage data inside a required region."},"faq":[{"question":"What does AI actually do with smart meter data?","answer":"The two deployments on this page show different jobs on the same underlying data. Con Edison used C3 AI to monitor the health of its 5.3 million meter rollout end to end, from an individual meter up to the whole system, while Southern California Gas Company used Bidgely's platform to disaggregate usage from smart meter data for digital energy efficiency reports for its medium consumption gas customers, a programme that exceeded its savings goal."},{"question":"How much data is involved?","answer":"Con Edison's smart meter deployment was expected to generate between 100 terabytes and 1 petabyte of data a year. Separately, and only once, building the analytics platform itself meant aggregating two years of data from 13 source systems covering 5 million customer accounts."},{"question":"Does this replace smart meters themselves?","answer":"No. It is the analytics layer on top of an existing advanced metering infrastructure rollout; both deployments on this page assume the meters are already reporting interval data."},{"question":"Can it detect theft or non technical losses?","answer":"Detecting unusual consumption patterns is a well documented research area for smart meter analytics, but neither deployment on this page reports a theft or loss detection result with a checked figure, so treat that specific claim as unproven until you have your own evidence."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched for the real estate and energy scope, with Con Edison's C3 AI deployment and Southern California Gas Company's Bidgely deployment checked against the primary sources."},{"date":"2026-09-28","note":"Editorial fix pass after adversarial review: replaced the SoCalGas evidence with the archived Wayback text (live page no longer carries the detail), corrected the \"targeting\" framing of the SoCalGas deployment in the meta description, FAQ and implementation steps, split the FAQ on data volume into the yearly forecast and the one time platform build, rewrote the Con Edison implementation step and problem statement to match the source and dropped the uncited reading interval and mass mailing claims, corrected the EU AI Act basis to the intended purpose test, narrowed the indicative value's meter range to the reference organization and relabelled its cost input as an editorial assumption, and changed the year on the Con Edison evidence to 2019 with a cited Wayback bound instead of an uncited 2018."},{"date":"2026-09-28","note":"Second editorial fix pass after adversarial review: rewrote the Con Edison complexityNote to the verified \"13 source systems covering 5 million customer accounts\" quote and dropped the uncited \"geospatial\" system and \"clean, current\" description; reworded the meta description and FAQ 1 to drop the unsupported \"personalised\" claim about SoCalGas's reports and to state the Con Edison figure as the source's own \"5.3 million smart meter deployment\"; corrected the SoCalGas evidence summary to describe a new programme for a previously unreached segment instead of one \"replacing\" paper reports; and set outcomeDisclosed to false on the Con Edison evidence, since its case study reports project scale, not a measured result."},{"date":"2026-09-28","note":"Third editorial fix pass after adversarial review: reworded the problem paragraph so the Con Edison sentence states only the installation, configuration and network health issues the source describes, and moved the appliance disaggregation and programme targeting problem to a general statement instead of attaching it to Con Edison's deployment; dropped the \"at a scale of terabytes to a petabyte a year\" clause from the complexityNote, which wrongly turned the yearly data forecast into a description of the one time historical data integration; and separated the prioritised routing advice in the first failure mode from its Con Edison attribution, since the source only reports a prioritised list, not a ranked routing practice."}],"slug":"smart-meter-analytics","url":"https://www.blits.ai/ai-use-cases/smart-meter-analytics","benchmarks":[],"indicativeValueResult":{"low":600000,"high":10000000},"evidence":["con-edison-c3-ai-smart-meter-analytics","socalgas-bidgely-energy-efficiency"]},{"title":"AI answer engine for readers built on a publisher's own journalism","shortTitle":"Publisher archive answer engine","seoTitle":"AI answer engine for news publishers","metaDescription":"Publishers answer reader questions with AI and citations. The Washington Post and the Financial Times answer only from their own reporting; TIME runs an agent too.","definition":"A generative AI search and answer tool on a publisher's own site or app that answers readers' questions only from that publisher's published journalism and archive, cites the articles it used, and declines to answer when its own reporting does not cover the question.","aliases":["news archive chatbot","AI search for news publishers","ask the newsroom AI","publisher AI answer engine"],"industries":["media-and-entertainment"],"functions":["customer-service"],"patterns":["rag-knowledge-assistant","conversational-agent","summarization"],"channels":["web-chat","mobile-app"],"audience":"customer-facing","autonomy":"autonomous","adoptionStage":"early-adopters","problem":"Readers increasingly ask AI chatbots and search engines for news instead of visiting a news site.\nThe answers are built from publishers' reporting, but the publisher loses the visit, the\nsubscriber relationship and the chance to show its depth. Archives are often hard to use even for\nthe publisher: TIME told Digital Content Next that its content sat in five different databases\ngoing back to the 1920s, some of it only as PDFs of magazines. Paying readers can struggle too:\nthe Financial Times found that FT Professional subscribers were looking for specific sources or\ndata points and struggled to find them quickly.\n\nGeneral AI assistants also get the news wrong. The BBC found significant issues in more than half\nof the answers four assistants gave to questions about the news, and Apple suspended its AI\nsummaries of news notifications after repeated mistakes. Publishers that answer with AI take\non the same risk under their own brand, so the design has to protect accuracy and attribution\nfirst.","problemStats":[{"statement":"The Reuters Institute Digital News Report 2026 finds that weekly use of AI chatbots such as ChatGPT, Perplexity and Google Gemini for news rose year on year from 7% to 10%.","sourceTitle":"Overview and key findings of the 2026 Digital News Report","sourceUrl":"https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/dnr-executive-summary","year":2026},{"statement":"BBC research found that 51% of all AI answers to questions about the news were judged to have significant issues of some form.","sourceTitle":"Groundbreaking BBC research shows issues with over half the answers from Artificial Intelligence (AI) assistants","sourceUrl":"https://www.bbc.co.uk/mediacentre/2025/bbc-research-shows-issues-with-answers-from-artificial-intelligence-assistants","year":2025}],"howItWorks":"1. **Index the journalism.** Published articles, and where rights allow the digitized archive,\n   are indexed with their dates, bylines, sections and corrections. New articles are indexed as\n   they publish.\n2. **Retrieve and rank.** A reader's question is matched against the index with semantic and\n   keyword search, favoring recent and relevant reporting.\n3. **Answer only above a threshold.** If no article scores above a relevance threshold, the tool\n   says it cannot answer rather than filling the gap from general knowledge.\n4. **Answer with sources.** The answer summarizes the retrieved reporting and links every point to\n   the articles it came from, with dates, so readers can check and read on. Some tools also\n   translate answers or read them aloud, as TIME's agent does in 13 languages, which opens the\n   reporting to readers who prefer another language or audio.\n5. **Label and learn.** The tool is labeled as AI generated and experimental where needed,\n   readers can flag bad answers, and editors review flagged answers and the most asked questions.","valueDrivers":["customer-experience","revenue-growth","inclusion-and-access"],"kpis":["users-served","interactions-handled","churn-reduction","customer-satisfaction","accuracy"],"indicativeValue":{"referenceOrg":"A subscription news publisher with 100,000 paying subscribers","inputs":[{"key":"subscribers","label":"Paying subscribers","low":100000,"high":100000,"unit":"subscribers","note":"The reference publisher."},{"key":"usageShare","label":"Share of subscribers who use the answer engine regularly","low":0.05,"high":0.15,"unit":"fraction of subscribers","note":"Editorial assumption, replace with your own usage data."},{"key":"churnPointsAvoided","label":"Reduction in annual churn among regular users","low":0.01,"high":0.03,"unit":"fraction of users (percentage points as a fraction)","note":"Editorial assumption. The Financial Times says Ask FT supports retention of key accounts but publishes no figure, and the return rate Digital Content Next reports for users of TIME's agent is not a controlled comparison."},{"key":"subscriptionValue","label":"Annual revenue per subscriber","low":100,"high":300,"unit":"USD per subscriber per year","note":"Editorial assumption, replace with your own average revenue per subscriber."}],"formula":"subscribers * usageShare * churnPointsAvoided * subscriptionValue","currency":"USD","period":"per year","resultLabel":"Subscription revenue retained","caveat":"A retention effect only, and hard to separate from the fact that engaged readers are the ones who use the tool. It leaves out model and search costs, which rise with usage, editorial time for review, any conversion or advertising value, and the reputational cost of a wrong answer."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Retrieval over clean article data is well understood. The hard parts are archive digitization and rights, a relevance threshold that refuses often enough, source linking, cost control as usage grows, and editorial ownership of a product that speaks in the newsroom's name.","dataPrerequisites":["Structured article data with dates, authors, sections, updates and corrections","Clear rights to use archive, wire and syndicated content in AI answers","An editorial policy for AI generated answers and how corrections apply to them"],"integrations":["Content management system and publishing pipeline for near real time indexing","Subscription and paywall system, to meter or gate answers","Analytics to measure return visits, reading and retention of users"]},"implementation":{"steps":[{"title":"Start where the reporting is deepest","detail":"Launch on a bounded topic or franchise with a rich archive, as TIME did with its Person of the Year coverage before opening its wider archive. The Washington Post also ran smaller AI experiments, including Climate Answers, before Ask The Post AI."},{"title":"Set the refusal threshold first","detail":"Decide the relevance score below which the tool declines to answer, and test it on questions the publication has not covered. Refusing is better than answering from outside the reporting."},{"title":"Link every answer to articles","detail":"Show the source articles with dates next to each answer so readers can verify and read on, which also drives article consumption."},{"title":"Put editors in the loop","detail":"Give editors a daily view of flagged answers and top questions, a way to correct or block answers, and a rule for how corrections to articles flow into answers."},{"title":"Roll out in phases","detail":"Test internally, then with a small group of subscribers, then widely, as the Financial Times did, measuring answer quality, repeat use and cost per answer at each step."}],"guardrails":["Answers only from the publisher's own indexed journalism, never from general model knowledge","A relevance threshold below which the tool declines to answer","Source links with dates for every answer","Clear labeling as AI generated, with a way for readers to report errors","Handling of sensitive topics (elections, health, breaking news) with stricter thresholds or editor written answers"],"humanInTheLoop":"Editors own the product's scope and policy, review flagged and high traffic answers, correct or block answers that fall short of editorial standards, and decide which topics are excluded, especially fast moving breaking news.","kpisToInstrument":["Share of questions answered versus declined","Accuracy of answers on a weekly editor reviewed sample","Click through from answers to articles","Return visits and retention of users versus comparable non users","Cost per answer"],"failureModes":[{"title":"Confident answers from thin coverage","detail":"The tool stretches a few articles into an answer the reporting does not support. Tune the threshold and review declined and borderline questions."},{"title":"Stale or corrected facts","detail":"Old articles contradict newer reporting or corrections. Rank by date, index corrections and show publication dates."},{"title":"Costs that scale with success","detail":"Model costs rise with every answer. Cache frequent answers and balance speed, accuracy and cost, as the Financial Times notes."},{"title":"Intrusive design","detail":"An AI box that crowds out the journalism annoys readers. Keep it optional and secondary to the article."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1) requires that readers know they are interacting with AI. Article 50(4) requires deployers to disclose AI generated text published to inform the public on matters of public interest, unless the content has undergone human review or editorial control and a person holds editorial responsibility. Whether answers generated on demand for a single reader count as text published to inform the public is open to interpretation, but they are rarely reviewed before readers see them, so the conservative choice is to label them."},"regulations":["eu-ai-act","gdpr","uk-gdpr"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Requires disclosure of AI generated text published to inform the public on matters of public interest, with an exception for content under human editorial control."}],"controls":["AI labeling on the answer box and on every answer","Editorial policy for AI answers, owned by a named editor","Rights register for the content the tool may use","Weekly accuracy review on a sample, with results reported to the editor","Logging of questions, retrieved sources and answers for corrections and complaints"],"incidents":[{"title":"Apple Intelligence: iPhone AI news alerts halted after errors","url":"https://www.bbc.com/news/articles/cq5ggew08eyo","note":"Apple suspended AI generated summaries of news notifications, which appeared to come from within news organizations' apps, after repeated mistakes summarizing headlines. Not a publisher's own tool, but the failure mode an archive answer engine must prevent."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** grounded in a **knowledge base** of the publisher's articles,\ningested from documents or **crawled website pages** with recrawls for new stories, and retrieved\nwith **hybrid search** (dense vectors plus BM25), so names and exact phrases match as well as\nmeaning. The agent is instructed to answer only from retrieved articles and to decline when\nretrieval finds nothing relevant; **output guardrails** and **quality guardrails** check answers\nbefore they are shown.\n\nIt runs in the **web chat** widget with the publisher's theming, or inside the publisher's own\nsite and app through the **REST or WebSocket API**, with streaming answers, **multi language**\nsupport and optional spoken answers through **text to speech**. **Thumbs up and down feedback**\nlets readers flag answers, **conversation logs** and **analytics** show what readers ask, **response\nfeedback summaries** show which answers they flagged. **Test suites** evaluate answer quality and\nrefusals against editor approved question sets, and scheduled **monitors** check the live agent.\nThe platform is model agnostic,\nwhich helps manage cost per answer."},"faq":[{"question":"Which publishers run their own AI answer engines?","answer":"The Washington Post launched Ask The Post AI in November 2024 over its reporting since 2016. The Financial Times made Ask FT available to all FT Professional customers in April 2025, and TIME has shown its AI agent on almost every piece of content since November 2025."},{"question":"How do publishers stop the AI from making things up?","answer":"By answering only from their own reporting and refusing below a relevance threshold. The Washington Post does not serve an answer when the tool does not readily find a relevant article, and Ask FT is built to call out when it lacks sufficient information."},{"question":"Do AI answer engines help publishers commercially?","answer":"Early signs point to engagement and retention rather than direct revenue. The Financial Times says Ask FT drives deeper reading that supports renewal of key accounts, and Digital Content Next reports that users of TIME's agent are more likely to return and spend more time on the site, although that compares users with non users rather than a controlled test."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the media scope with evidence from The Washington Post, the Financial Times and TIME, checked against the sources."},{"date":"2026-09-27","note":"Editor review. TIME claims limited to what TIME and Digital Content Next state, engagement figures attributed to Digital Content Next, archive search claim replaced with the sourced FT finding, and Blits.ai analytics wording aligned with the feature inventory."},{"date":"2026-09-27","note":"Second editor review. Removed Blits.ai claims not in the feature inventory (citations, tests on every change), tied the archive problem to the sourced TIME detail, softened the Climate Answers example, added translation and audio to support the access driver, and noted the uncertain reach of Article 50(4)."}],"slug":"publisher-archive-answer-engine","url":"https://www.blits.ai/ai-use-cases/publisher-archive-answer-engine","benchmarks":[],"indicativeValueResult":{"low":5000,"high":135000},"evidence":["financial-times-ask-ft","time-ai-agent-archive-answers","washington-post-ask-the-post-ai"]},{"title":"AI assistant for B2B telecom quoting, sales and service","shortTitle":"B2B quoting and service","seoTitle":"AI assistant for B2B telecom quoting and service","metaDescription":"An AI assistant answers telecom business customers' product and price questions and drafts quotes for sellers to approve. SoftBank reports 70% self resolution.","definition":"An AI assistant that serves business customers of a telecom operator and the sellers who look after them: it answers product, pricing and contract questions, prepares configurations and quotes for connectivity, mobile fleets and devices, drafts responses to tenders, and handles routine service requests and fault tickets, with a sales or service specialist approving anything binding.","aliases":["B2B telco sales assistant","enterprise quoting assistant","CPQ assistant for telecom","business customer service agent","RFP response assistant"],"industries":["telecommunications"],"functions":["sales","customer-service","product-and-pricing"],"patterns":["conversational-agent","rag-knowledge-assistant","agentic-workflow","recommendation-and-personalization","content-generation"],"channels":["web-chat","agent-desktop","email","internal-tools"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"Business customers buy complex bundles: fibre and dedicated lines at several sites, mobile fleets,\ndevices, security and cloud services, each with its own availability, pricing rules and service\nlevels. Small businesses want quick answers and quotes without waiting for a sales call; large\naccounts issue tenders with long questionnaires. Sellers spend hours researching accounts,\nchecking availability and assembling quotes, and service teams handle routine requests that the\ncustomer could have done alone.\n\nThe knowledge needed lives in product sheets, price books, contract templates and past proposals\nspread across systems. Configure, price and quote tools help specialists, but they do not answer a\nsmall business owner's question at night or draft the first version of a tender response.","problemStats":[],"howItWorks":"1. **Answer and qualify.** On the business website or portal, the assistant answers questions\n   about products, prices and contracts, checks basic eligibility such as fibre availability at an\n   address, and qualifies the need.\n2. **Configure and quote.** Using the product catalogue and configure, price and quote tools, it\n   prepares a configuration and an indicative quote within standard price rules.\n3. **Hand over to a seller.** Anything non standard (discounts, multi site designs, tenders) goes\n   to a seller with the configuration, the conversation and the account context.\n4. **Assist sellers.** For sellers the assistant researches the account, suggests the next best\n   offer, and drafts tender answers and proposals from approved content.\n5. **Serve after the sale.** Business customers raise and track orders and fault tickets, and the\n   assistant resolves routine requests, keeping them informed during service interruptions.","valueDrivers":["revenue-growth","employee-productivity","cost-to-serve","speed"],"kpis":["containment-rate","conversion-rate-uplift","time-saved-per-task","handling-time-reduction","interactions-handled"],"indicativeValue":{"referenceOrg":"A telecom operator's business unit with 50,000 small and medium business customers","inputs":[{"key":"businessCustomers","label":"Business customers","low":50000,"high":50000,"unit":"customers","note":"The reference business unit."},{"key":"inquiriesPerCustomer","label":"Routine sales and service inquiries per customer per year","low":2,"high":4,"unit":"inquiries per customer per year","note":"Editorial assumption, replace with your own business contact volumes."},{"key":"resolvedShare","label":"Share of routine inquiries the assistant resolves","low":0.3,"high":0.6,"unit":"fraction of inquiries","note":"Conservative against the benchmark on this page (SoftBank reports a self resolution rate of 70% for its business website agent)."},{"key":"costPerInquiry","label":"Cost of an inquiry handled by a seller or business service agent","low":15,"high":30,"unit":"USD per inquiry","note":"Editorial assumption; business inquiries take longer and are handled by more expensive staff. Replace with your own cost."}],"formula":"businessCustomers * inquiriesPerCustomer * resolvedShare * costPerInquiry","currency":"USD","period":"per year","resultLabel":"Sales and service handling cost avoided","caveat":"Counts only routine inquiries handled without a person. It leaves out additional sales from faster quotes and better prepared sellers, faster tender responses, the cost of the AI and the catalogue and quoting integrations, and discounts that still need approval."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering product questions is easy once content is clean. Quoting needs the product catalogue, availability checks and price rules behind APIs, and binding offers must stay in the approved configure, price and quote process.","dataPrerequisites":["Current business product catalogue, price books and discount rules","Address level availability for fixed products","Approved contract templates, service levels and past tender answers","Account data in the CRM, including installed base and contract dates"],"integrations":["CRM and account management","Configure, price and quote system","Availability and serviceability checks","Service management and ticketing for business customers","Document and proposal repositories"]},"implementation":{"steps":[{"title":"Start with questions and qualification","detail":"Put an assistant on the business website that answers product, price and contract questions from approved content and routes qualified leads, as SoftBank did first because it had the most customer touchpoints."},{"title":"Connect availability and indicative quotes","detail":"Add address availability and standard price quotes through tools, and keep every binding quote and discount in the existing approval process."},{"title":"Give sellers an account and tender assistant","detail":"Let sellers ask for account research, next best offers and draft tender answers from an approved library, and measure time saved per proposal."},{"title":"Bring service into the same assistant","detail":"Let business customers raise and track orders and tickets and get status during service interruptions, so the relationship does not end at the sale."},{"title":"Improve from unresolved questions","detail":"Review the questions the assistant could not answer each week and add the missing content. SoftBank credits this kind of iterative improvement for raising its self resolution rate from about 50% at launch to 70%."}],"guardrails":["Prices, discounts and availability only from tools, never from the model","Binding quotes and non standard discounts approved by a seller","Tender answers drawn only from an approved answer library, with a seller reviewing every submission","Customer contract data visible only to authorised users of that account","Handover to a named seller or service agent on request"],"humanInTheLoop":"Sellers approve every binding quote, discount and tender response, and service specialists handle complex faults and escalations. Product and pricing teams own the content the assistant uses and review unresolved questions weekly.","kpisToInstrument":["Self resolution rate of business inquiries, checked against repeat contacts","Time from inquiry to quote","Win rate of assisted quotes versus unassisted ones","Seller hours spent on research and tender responses","Quote errors found at approval"],"failureModes":[{"title":"Quotes the operator cannot honour","detail":"The assistant quotes a price or service that is not available at the address. Check availability and price in tools and label quotes as indicative."},{"title":"Tender answers that overpromise","detail":"A generated answer commits to a service level the operator does not offer. Use an approved library and seller review."},{"title":"Leaking account data","detail":"A user sees another company's contract details. Enforce account level access in the tools."},{"title":"Leads lost in handover","detail":"Qualified leads sit in an inbox. Route them to named sellers with service levels for follow up."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1): people must be informed that they are interacting with an AI system, unless that is obvious from the context. Quoting, sales support and service for business customers are not listed in Annex III. The use would become high risk under Annex III point 5(b) only if the assistant itself evaluated the creditworthiness of a natural person, such as a sole trader, to decide whether to offer a contract."},"regulations":["eu-ai-act","gdpr","telecom-consumer-rules","eecc"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"European Electronic Communications Code (Directive (EU) 2018/1972)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2018/1972/oj","note":"Several end user protections, such as contract information, also apply to microenterprises and small businesses unless they waive them, so a B2B assistant cannot treat every business customer as a large enterprise."}],"controls":["AI disclosure on the business website and portal","Price and discount approval workflow outside the model","Approved answer library with owners for tender content","Account level access control and audit logging","Weekly review of unresolved questions and quote errors"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** grounded in a **knowledge base** of product sheets, price\nrules and approved tender answers with hybrid retrieval, and connected through **custom\nfunctions** or ready made connectors to the CRM, quoting and availability systems (for example\nSalesforce, Microsoft Dynamics 365, HubSpot or ServiceNow from the integration catalogue).\n**SQL knowledge bases** let sellers ask questions about their installed base in plain language.\n\nCustomers use it in **web chat** on the business site and by **email**; sellers use it in\n**Microsoft Teams**. **Agentic workflows** with **human in the loop approval** draft quotes and\ntender answers for a seller to approve, **human handover** routes qualified leads, **guardrails**\nkeep prices and commitments inside approved content, and **test suites** replay product\nquestions whenever the catalogue changes. The platform is model agnostic with EU and UAE data\nresidency."},"faq":[{"question":"Can business customers really self serve with an AI agent?","answer":"For routine questions, yes. SoftBank reports that the self resolution rate of its website agent for small and medium business customers rose from about 50% at launch to 70%. Sierra, the platform vendor, measures that rate with its own AI monitoring and puts the volume at about 100 inquiries a day. Complex designs and discounts still go to sellers."},{"question":"What does AI do for B2B sellers?","answer":"Mostly research and preparation. Microsoft reports that Lumen cut account research that took a seller up to four hours to 15 minutes with Copilot, and Pega reports a 15% improvement in win rate from Verizon's AI guided selling engine."},{"question":"Should the assistant produce binding quotes?","answer":"Indicative quotes, yes; binding ones, no. Keep discounts and binding offers in the approved quoting process, with a seller signing off."}],"related":["plan-upgrade-and-sales-assistant","corporate-client-servicing-assistant","network-outage-communication-agent","order-to-activation-and-esim-onboarding-assistant","inbound-lead-qualification-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the telecommunications vertical from SoftBank, Vodafone Business, Telefónica, Lumen and Verizon sources checked word for word."},{"date":"2026-09-25","note":"Consolidation pass: added European Electronic Communications Code to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Lumen claimant (Microsoft reports it) and the \"up to four hours\" baseline, separated SoftBank's 70% rate from Sierra's volume figure, corrected the Verizon evidence year to 2022, added source dates, removed an unsourced tender size, refined the EU AI Act basis, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Fully separated the SoftBank self resolution rate from Sierra's AI monitoring and volume claim in the FAQ, since the answer still mixed the two attributions."}],"slug":"business-connectivity-quoting-and-service-assistant","url":"https://www.blits.ai/ai-use-cases/business-connectivity-quoting-and-service-assistant","benchmarks":[{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":70,"min":70,"max":70,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"softbank-smb-sales-ai-agent","pooled":true}]},{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":15,"min":15,"max":15,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"verizon-business-next-best-x-engine","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":100,"min":100,"max":100,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"softbank-smb-sales-ai-agent","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"verizon-business-next-best-x-engine","pooled":false}]},{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"lumen-copilot-sales-account-research","pooled":false}]}],"indicativeValueResult":{"low":450000,"high":3600000},"evidence":["lumen-copilot-sales-account-research","softbank-smb-sales-ai-agent","telefonica-genia-b2b-device-marketplace","verizon-business-next-best-x-engine","vodafone-business-servicenow-service-automation"]},{"title":"AI assistant for benefits eligibility questions and applications","shortTitle":"Benefits eligibility and application assistant","seoTitle":"AI assistant for benefits eligibility and claims","metaDescription":"AI assistants explain benefit rules, guide applications and answer status questions; caseworkers decide. DWP's voice platform handles about 1 million calls a month.","definition":"An AI assistant that helps people understand which public benefits and grants may apply to them, explains the rules and documents in plain language, guides them through the application and checks it for completeness, while the eligibility decision stays with the agency's rules and caseworkers.","aliases":["benefits chatbot","social security virtual assistant","welfare application assistant","benefits navigator"],"industries":["government"],"functions":["citizen-services","case-management","customer-service"],"patterns":["conversational-agent","rag-knowledge-assistant","voice-agent","document-processing"],"channels":["web-chat","voice","mobile-app","whatsapp"],"audience":"customer-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Benefit rules are complex and change often, and the people who need them are often under\npressure: a lost job, a disaster, a new disability, a family change. Many never\nclaim support they are entitled to: in Great Britain alone, DWP estimates that up to 910,000\nfamilies entitled to Pension Credit did not claim it. Others apply for the wrong scheme, and\nincomplete applications bounce back and forth between applicant and caseworker. Phone lines fill\nwith questions about claim progress, payment dates and missing documents.\n\nIt is also an area where automation has already caused serious public harm. The Dutch childcare\nbenefits scandal, where a risk profiling algorithm used nationality as a risk factor, shows what\nhappens when an algorithm's output drives how applicants are treated. An assistant here must widen\naccess and cut rework without quietly deciding who gets support.","problemStats":[{"statement":"DWP estimates that up to 910,000 families in Great Britain who were entitled to Pension Credit did not claim it in the financial year ending 2024, leaving up to £2.5 billion unclaimed.","sourceTitle":"Income-related benefits: estimates of take-up: financial year ending 2024","sourceUrl":"https://www.gov.uk/government/statistics/income-related-benefits-estimates-of-take-up-financial-year-ending-2024/income-related-benefits-estimates-of-take-up-financial-year-ending-2024","year":2025}],"howItWorks":"1. **Explain the schemes.** The assistant answers questions about benefits, grants and support in\n   plain language from the agency's approved rules and guidance, with sources.\n2. **Screen, do not decide.** Where the agency allows it, it runs the official eligibility\n   questions (often a rules engine owned by the agency) and says which schemes look worth applying\n   for, stating clearly that the outcome is not a decision.\n3. **Guide the application.** It explains what each question means, which documents are needed and\n   why, in the applicant's language, and checks the form for gaps before submission.\n4. **Answer status questions.** For authenticated users it reads case status, next payment date\n   and missing evidence from the case system.\n5. **Hand over.** Anyone in crisis, disputing a decision or showing signs of vulnerability goes to\n   a caseworker with the conversation attached.","valueDrivers":["inclusion-and-access","customer-experience","cost-to-serve","speed"],"kpis":["interactions-handled","users-served","accuracy","response-time-reduction","containment-rate","processing-time-reduction"],"indicativeValue":{"referenceOrg":"A regional benefits agency that receives 500,000 applications a year","inputs":[{"key":"applications","label":"Applications received per year","low":500000,"high":500000,"unit":"applications per year","note":"The reference agency."},{"key":"incompleteShare","label":"Share of applications returned for missing information","low":0.15,"high":0.3,"unit":"fraction of applications","note":"Editorial assumption. Replace with your own rework rate."},{"key":"reductionShare","label":"Share of those incomplete applications the assistant prevents","low":0.2,"high":0.4,"unit":"fraction of incomplete applications","note":"Editorial assumption; no public benchmark yet measures this directly."},{"key":"reworkHours","label":"Caseworker hours per incomplete application","low":0.5,"high":1,"unit":"hours per application","note":"Editorial assumption covering contact, chasing documents and re entry."},{"key":"hourlyCost","label":"Fully loaded caseworker cost per hour","low":35,"high":55,"unit":"USD per hour","note":"Editorial assumption. Replace with your own cost."}],"formula":"applications * incompleteShare * reductionShare * reworkHours * hourlyCost","currency":"USD","period":"per year","resultLabel":"Caseworker rework cost avoided","caveat":"Covers rework on incomplete applications only. It leaves out contact centre savings on status questions, faster payment to applicants, higher take up (which raises benefit spending) and the cost of building and governing the assistant."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Answering general questions is straightforward; screening and application support touch the most regulated decisions in government. The work is in keeping the assistant separate from the eligibility decision, integrating with case systems and identity, and proving it treats groups fairly.","dataPrerequisites":["Approved, current benefit rules and guidance with owners and effective dates","The agency's official eligibility logic, preferably as an executable rules service","Application forms and document requirements per scheme","Case status data reachable through APIs for authenticated users"],"integrations":["Identity and authentication (national login or agency account)","Case management and payment systems for status and next payment","Rules engine for eligibility screening, owned by the policy team","Document upload and verification services","Contact centre and caseworker queues for handover"]},"implementation":{"steps":[{"title":"Separate information from decision","detail":"Write down which outputs are information, which are screening and which are decisions, and keep the last with the agency's rules engine and caseworkers. Put this in the service design and the register entry."},{"title":"Start with explanation and status","detail":"Launch with plain language explanations and authenticated status questions, which carry low decision risk and high contact volume (DWP's voice platform answers next payment questions in the IVR)."},{"title":"Use the official rules for screening","detail":"If you add eligibility screening, call the agency's own rules service rather than letting a language model interpret the law, and label the result as indicative."},{"title":"Test for fairness and vulnerability","detail":"Build test sets across languages, disabilities and circumstances, including people in crisis, and check that handover triggers fire."},{"title":"Pilot with caseworkers watching","detail":"Pilot with a limited group and have caseworkers review transcripts weekly. Leeds tested answers against a question set with reference answers verified by domain experts before launch, and plans to review transcripts and feedback during its pilot."},{"title":"Measure take up and rework, not only contacts","detail":"Track incomplete applications, time to decision and take up among eligible groups, not just conversations."}],"guardrails":["The assistant never tells a person they are or are not entitled; screening results are labelled indicative and come from the official rules","Answers only from approved rules and guidance, with sources and effective dates","Automatic handover on crisis, vulnerability, disputes and appeals","Personal data masking in logs and prompts, and data minimisation in what the assistant asks","Equal treatment tests across language and demographic groups before and after every change"],"humanInTheLoop":"Caseworkers decide every application and every change to an award. They review samples of conversations each week, with priority for handovers and complaints, and policy owners approve each new scheme or rule before the assistant explains it.","kpisToInstrument":["Share of applications submitted complete, with and without the assistant","Time from first contact to decision","Accuracy of answers on a weekly expert reviewed sample","Handover rate and reasons, including vulnerability triggers","Take up and outcome differences across language and demographic groups"],"failureModes":[{"title":"Screening becomes the decision","detail":"Applicants told they probably do not qualify stop applying. Nissewaard's register entry names this exact risk. Label screening as indicative and always allow an application."},{"title":"Automated suspicion","detail":"In the Dutch childcare benefits scandal, Amnesty International found that a risk profiling algorithm using nationality led to discrimination and racial profiling of applicants. Keep fraud models out of the assistant and govern them separately."},{"title":"Outdated rules","detail":"Benefit rates and thresholds change every year. Tie content to effective dates and retest on every change."},{"title":"No route to a person","detail":"A chatbot that cannot hand over leaves people in crisis with a phone number at best. Leeds' pilot, which does not escalate to a human, at least points distressed users to emergency helplines and gives the council's number. Add a direct human route before scaling."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 5(a) makes AI high risk when it is used by or on behalf of public authorities to evaluate the eligibility of natural persons for essential public assistance benefits and services, or to grant, reduce, revoke or reclaim them. An assistant that only explains rules and guides applications carries the Article 50 transparency duties (limited risk); one that screens or scores eligibility falls under point 5(a), and a public body deploying it must carry out a fundamental rights impact assessment first (Article 27)."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(a) covers AI used to evaluate eligibility for essential public assistance benefits and services."},{"title":"Article 27, fundamental rights impact assessment for high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/27/","note":"Public bodies deploying high risk AI must assess the impact on fundamental rights before use."},{"title":"Directive on Automated Decision-Making","issuer":"Treasury Board of Canada Secretariat","region":"north-america","url":"https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32592","note":"Requires an algorithmic impact assessment, notice before decisions, explanation after decisions and human involvement for automated decision systems of Canadian federal institutions."},{"title":"AI Playbook for the UK Government","issuer":"UK Government","region":"europe","url":"https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government","note":"Guidance for UK public bodies on building and governing AI services."}],"controls":["Documented boundary between information, screening and decision in the service design and register entry","Fundamental rights or algorithmic impact assessment before launch where screening is involved","Equality monitoring of outcomes and handovers across groups","Right to a human caseworker and to apply regardless of any screening result","Audit trail of every answer, source and screening result shown to an applicant"],"incidents":[{"title":"Xenophobic machines: discrimination through unregulated use of algorithms in the Dutch childcare benefits scandal","url":"https://www.amnesty.org/en/documents/eur35/4686/2021/en/","note":"Amnesty International's analysis of how the Dutch tax authorities' risk profiling of childcare benefit applicants used nationality as a risk factor, resulting in discrimination and racial profiling."},{"title":"Automated Neglect: how the World Bank's push to allocate cash assistance using algorithms threatens rights","url":"https://www.hrw.org/report/2023/06/13/automated-neglect/how-world-banks-push-allocate-cash-assistance-using-algorithms","note":"Human Rights Watch on how Takaful, a World Bank funded cash transfer programme in Jordan, ranks families by algorithm and deprives many people of their right to social security."}]},"blitsAi":{"howToBuild":"On Blits.ai the conversation runs as an **AI agent** grounded in a **knowledge base** of approved\nbenefit rules and guidance, retrieved with **hybrid search**. Eligibility screening is not left\nto the model: a **custom function** calls the agency's own rules service, and the screening\njourney runs as a **flow** with deterministic steps, an authentication block and sensitive data\nflags on the questions it asks. Status and payment questions use custom functions against the\ncase system; document uploads use the receive attachment block.\n\n**Guardrails** stop the agent from stating entitlement, **PII masking** removes personal data\nbefore text reaches a model, and **human handover** routes crisis, vulnerability and disputes to\ncaseworkers with the transcript. The same agent works on **web chat, WhatsApp and voice**, with\nlanguage detection and translation for multilingual service. **Test suites** check fairness and\nhandover scenarios on every change, and the platform runs in **EU and UAE data residency** regions."},"faq":[{"question":"Can an AI assistant decide benefit eligibility?","answer":"It should not. Under the EU AI Act, AI that evaluates eligibility for public assistance is high risk, and public examples keep the decision elsewhere: DWP's voice platform makes no decisions and routes callers to advisers, and Nissewaard's eligibility check is a rules based decision tree whose outcome caseworkers can overrule."},{"question":"What do benefits assistants handle today?","answer":"Mostly explanation, status and routing. Federal Student Aid's Aidan reached over 2.6 million unique customers in just over two years, and about 1 million calls a month pass through DWP's voice platform, which answers some questions in the IVR, signposts callers to GOV.UK or routes them to an adviser."},{"question":"How accurate are they?","answer":"Public figures are sparse and set their own bars. Leeds reports that 83% of answers scored at least 3 out of 5 in quality in testing before launch, with input from domain experts, and DWP reports 97% speech recognition success. Measure accuracy on your own schemes with experts before launch."}],"related":["citizen-information-assistant","benefit-fraud-and-error-detection","public-service-translation","permit-and-licence-application-processing","immigration-and-visa-application-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Unpublished by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from the US federal AI inventory, UK transparency records, the Dutch algorithm register and vendor case studies, with every quote verified against the source."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Leeds handover failure mode and test context, removed an unsupported FEMA claim from the FAQ, tightened incident notes and the EU AI Act basis, added a DWP take up statistic, corrected the Nissewaard register category and vendor, and added an SEO title and meta description."}],"slug":"benefits-eligibility-and-application-assistant","url":"https://www.blits.ai/ai-use-cases/benefits-eligibility-and-application-assistant","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":3,"nUpTo":0,"median":1000000,"min":30000,"max":11000000,"byClaimant":{"organization":2,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"federal-student-aid-aidan-virtual-assistant","pooled":true},{"id":"dwp-conversational-platform","pooled":true},{"id":"region-sud-training-grant-document-verification","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":90,"min":83,"max":97,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dwp-conversational-platform","pooled":true},{"id":"leeds-city-council-money-information-centre-chatbot","pooled":true}]},{"kpi":"response-time-reduction","label":"Response time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":99,"min":99,"max":99,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"youngwilliams-priya-benefits-agent","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2600000,"min":2600000,"max":2600000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"federal-student-aid-aidan-virtual-assistant","pooled":true}]}],"indicativeValueResult":{"low":262500,"high":3300000},"evidence":["dwp-conversational-platform","federal-student-aid-aidan-virtual-assistant","fema-individual-assistance-document-translation","gemeente-nissewaard-benefit-application-eligibility-check","leeds-city-council-money-information-centre-chatbot","region-sud-training-grant-document-verification","youngwilliams-priya-benefits-agent"]},{"title":"AI assistant for citizen information and government services","shortTitle":"Citizen information assistant","seoTitle":"Government AI chatbot for citizen information","metaDescription":"AI assistants that answer residents from official guidance and hand personal cases to staff. DVLA's bot automates around 20% of its web chat enquiries.","definition":"An AI assistant that answers residents' and businesses' questions about government services in plain language, grounded only in official guidance with links to the source, points them to the right online service or office, and hands anything personal, urgent or outside its content to a human with the context attached.","aliases":["government chatbot","citizen services chatbot","public sector virtual assistant","GOV.UK Chat style assistant"],"industries":["government"],"functions":["citizen-services","customer-service","knowledge-management"],"patterns":["rag-knowledge-assistant","conversational-agent","voice-agent","classification-and-routing"],"channels":["web-chat","mobile-app","whatsapp","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"Government information is spread over thousands of pages written by different departments, and\npeople rarely know which agency owns their problem. They phone or visit because they cannot find\nor trust the answer online, so contact centres answer many questions the website already\ncovers: which form, which deadline, which office, what does this letter mean.\nQueues grow at exactly the moments demand spikes, such as tax deadlines, benefit changes or an\nemergency.\n\nKeyword search and scripted FAQ bots only help people who already know the official term.\nGenerative assistants can read a question in everyday words and synthesise an answer across\npages, but a wrong answer from a government channel carries more weight than a wrong answer from\na shop. The work is in grounding, refusal, privacy and a clean route to a human.","problemStats":[],"howItWorks":"1. **Understand the question.** The assistant reads the question in the resident's own words and\n   language, and classifies it (information request, personal case question, urgent need,\n   complaint, out of scope).\n2. **Retrieve from official content only.** It searches an index of approved guidance pages and\n   answers from the retrieved passages, with links to every source, and refuses when the content\n   does not cover the question.\n3. **Protect personal data.** Personal data in the question is detected and masked or the question\n   is rejected; nothing is used to profile the resident.\n4. **Check the answer.** A second check screens the draft for unsupported claims, advice the\n   government cannot give, tone and personal data before it is shown.\n5. **Route what it cannot answer.** Case specific questions go to the authenticated service or a\n   human adviser with the conversation attached; urgent needs are pointed to the right phone line.","valueDrivers":["inclusion-and-access","customer-experience","cost-to-serve","speed"],"kpis":["containment-rate","interactions-handled","users-served","accuracy","customer-satisfaction","time-saved-per-task"],"indicativeValue":{"referenceOrg":"A national agency that receives 2 million phone and chat contacts a year","inputs":[{"key":"contacts","label":"Assisted phone and chat contacts per year","low":2000000,"high":2000000,"unit":"contacts per year","note":"The reference agency."},{"key":"infoShare","label":"Share of contacts that are general information questions","low":0.3,"high":0.5,"unit":"fraction of contacts","note":"Editorial assumption. Replace with your own contact reason analysis."},{"key":"containment","label":"Share of those questions the assistant resolves without a human","low":0.2,"high":0.4,"unit":"fraction of information contacts","note":"The low end follows the benchmark on this page (DVLA reports around 20% of web chat enquiries automated by its scripted, non generative bot); the high end is an editorial assumption for a grounded generative assistant. DVLA's figure is a share of all web chat enquiries, not of information questions only, so the two bases differ. Replace with your own pilot data."},{"key":"costPerContact","label":"Cost of a human handled contact","low":4,"high":8,"unit":"USD per contact","note":"Editorial assumption for a blended phone and chat contact in the public sector. Replace with your own fully loaded cost."}],"formula":"contacts * infoShare * containment * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled contact cost avoided","caveat":"Gross avoided cost only. It leaves out the cost of building and running the assistant, content maintenance, the value of 24/7 access and the extra demand an easier channel can create."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"A grounded information assistant touches no transactional systems, so the build is mostly content and evaluation work: cleaning the guidance corpus, deciding what it must refuse and testing accuracy by topic. Complexity rises once it reads personal case data or books services.","dataPrerequisites":["A current, owned corpus of official guidance with publication and review dates","Contact reason data from the contact centre to choose the first topics","A set of real questions with approved ideal answers for evaluation"],"integrations":["Content management system or publishing feed, so the index updates when guidance changes","Contact centre or webchat platform for handover with the transcript","Web, app, messaging and telephony channels","Analytics for unanswered questions and feedback"]},"implementation":{"steps":[{"title":"Choose topics by volume and harm","detail":"Start with high volume, low harm information topics (opening hours, which form, how to apply) and keep personal case questions, legal advice and urgent needs out of scope at first."},{"title":"Clean and filter the content","detail":"Index only official pages, exclude documents likely to contain personal data, and chunk by heading so answers can cite the exact section. Connect the index to publishing so it updates daily."},{"title":"Write the refusal and routing rules","detail":"Decide what the assistant must never do (give personal advice, guess eligibility, answer outside the content) and where each type of refused question goes."},{"title":"Build an evaluation set before launch","detail":"Collect real questions with ideal answers per topic, add adversarial and jailbreak prompts, and score groundedness, accuracy and completeness on every change."},{"title":"Pilot with a capped group","detail":"Run a limited test (the GOV.UK Chat transparency record describes a test with up to 2,000 users over four weeks), review transcripts, then widen channel by channel."},{"title":"Measure unanswered questions, not only usage","detail":"Track refusals, handovers and repeat contacts per topic and feed gaps back to the content owners."}],"guardrails":["Answers only from retrieved official content, with a link to every source and a refusal when retrieval finds nothing relevant","Personal data detection on input, with masking or rejection before text reaches a model","A second check on every answer for unsupported claims, advice and tone before display","Clear AI disclosure and a visible route to a human or the official phone line","No decisions about eligibility, entitlement or enforcement"],"humanInTheLoop":"Content owners approve the corpus and the topics in scope; a service team reviews samples of conversations every week, with extra review for refusals and complaints. Advisers take every handover with the transcript, and changes to scope go through the same sign off as a change to published guidance.","kpisToInstrument":["Share of questions answered from content, refused and handed over, per topic","Accuracy and groundedness on a weekly human reviewed sample","Repeat contact on the same topic within seven days","Satisfaction and feedback on answers, compared with search and phone","Contact centre volume on the topics in scope"],"failureModes":[{"title":"Confident answers that break the law","detail":"New York City's MyCity chatbot was reported in 2024 to tell businesses they could do things that are illegal. Ground strictly, refuse outside the content and test legal edge cases before launch."},{"title":"Stale guidance in the index","detail":"The assistant repeats last year's rules after a policy change. Tie the index to the publishing pipeline and show the source date."},{"title":"Personal questions answered generically","detail":"Residents ask about their own case and get a generic answer that sounds personal. Detect case questions and route them to the authenticated service."},{"title":"Usage mistaken for success","detail":"High conversation numbers can hide abandonment. Measure unanswered questions and repeat contacts. Microsoft's case study on Montgomery County's Monty 2.0 reports a drop in the share of unanswered queries from between 35% and 45% to between 10% and 15%, next to its conversation count."}]},"risk":{"euAiAct":{"tier":"limited","basis":"An information assistant must tell people they are interacting with AI (Article 50). It is not high risk as long as it does not evaluate eligibility for public assistance benefits or services (Annex III point 5(a)); an assistant that starts to pre assess eligibility should be reassessed."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"AI Playbook for the UK Government","issuer":"UK Government","region":"europe","url":"https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government","note":"Guidance for civil servants and people working in government organisations on using AI, including generative AI, safely, effectively and securely."},{"title":"Algorithmic Transparency Recording Standard Hub","issuer":"Government Digital Service","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"Mandatory for all UK government departments and for arm's length bodies that deliver public or frontline services; the GOV.UK Chat, DVLA and FCDO records on this page were published under it."},{"title":"M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust","issuer":"US Office of Management and Budget","region":"north-america","url":"https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf","note":"Requires US federal agencies to inventory their AI use cases at least annually and to apply minimum risk management practices to high impact AI."}],"controls":["AI disclosure and a statement that answers may be wrong, with links to the official source","Entry in the public AI or algorithm register (for example ATRS in the UK or the Dutch algorithm register)","Data protection impact assessment covering question logs and retention","Red teaming for jailbreaks and legal edge cases before launch and after model changes","Retention limits on conversation logs and access control for reviewers"],"incidents":[{"title":"NYC's AI chatbot tells businesses to break the law","url":"https://themarkup.org/news/2024/03/29/nycs-ai-chatbot-tells-businesses-to-break-the-law","note":"In March 2024 The Markup found New York City's MyCity business chatbot giving answers that contradicted local law on housing, worker rights and consumer protection, for example that bosses could take workers' tips."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** grounded in a **knowledge base** that holds only official\nguidance, ingested from documents and **crawled websites** that can be recrawled, and retrieved\nwith **hybrid search**, so every answer can cite its source. **Guardrails** check input and\noutput, the platform default blocks prompt injection, and **PII masking** at the gateway removes\npersonal data before text reaches a model. Fixed journeys such as \"which office do I need\"\nrun as a **flow**, with language detection and translation blocks for multilingual service.\n\nThe same agent serves **web chat, WhatsApp and voice**, reaches a mobile app through the **REST\nor WebSocket API channel**, and **human handover** passes the conversation to the contact\ncentre. **Test suites** replay real questions with ideal answers on every content or model\nchange, **monitors** check live answers on a schedule, and **analytics** break down untrained\nquestions and unexpected answers. The platform is **model agnostic** and runs\nin **EU and UAE data residency** regions, which matters for public sector hosting rules."},"faq":[{"question":"Can a government chatbot use generative AI safely?","answer":"Yes, if it answers only from official content and shows its sources. GOV.UK Chat retrieves from GOV.UK pages, rejects questions that contain common personal data patterns and runs every answer through a second model check; FCDO's consular triage goes further and only picks from approved templates, so users never see generated text."},{"question":"What share of questions can it handle?","answer":"Reported figures are modest but real. DVLA automates around 20% of web chat enquiries, and FCDO's triage picked the correct response for 76% to 81% of historic enquiries in testing. Treat any number as topic specific and measure it on your own questions."},{"question":"Is a citizen information chatbot high risk under the EU AI Act?","answer":"Not by default. It carries the Article 50 transparency duty. It becomes high risk when it evaluates eligibility for public assistance benefits or services, which is why eligibility questions should route to the official service."},{"question":"Should every agency build its own assistant?","answer":"Not necessarily. Estonia's Bürokratt is one assistant, developed under the Information System Authority, that about 20 public bodies and websites use, from the state portal to the Tax and Customs Board. A shared platform means each agency does not have to build, secure and evaluate its own assistant."}],"related":["benefits-eligibility-and-application-assistant","non-emergency-service-request-routing","tax-questions-and-filing-assistant","public-service-translation","immigration-and-visa-application-assistant","permit-and-licence-application-processing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from UK transparency records, the Dutch algorithm register, government pages and vendor case studies, with every quote verified against the source."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: FCDO triage corrected to pilot with Azure OpenAI as model provider; Boti channels limited to WhatsApp with the city's own page added; Tilburg supplier, start date and Bürokratt organization list corrected; source dates added; Bürokratt FAQ, guidance notes and Blits.ai capabilities aligned with sources; SEO title and meta description added."},{"date":"2026-09-27","note":"Review fixes: DVLA's 50% IVR figure removed as a handling time metric (it measures menu navigation time) and replaced by the adviser chat time saved per handed over chat; handling time reduction replaced by time saved per task in the KPIs; the Montgomery County unanswered query figure attributed to Microsoft's case study; Montgomery topic count and languages, Boti history and DVLA summary aligned with the sources."}],"slug":"citizen-information-assistant","url":"https://www.blits.ai/ai-use-cases/citizen-information-assistant","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":1010000,"min":20000,"max":2000000,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"government-of-the-city-of-buenos-aires-boti","pooled":true},{"id":"montgomery-county-monty-chatbot","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":76,"min":76,"max":76,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"fcdo-consular-enquiry-triage","pooled":true}]},{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":20,"min":20,"max":20,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dvla-contact-centre-conversational-ai","pooled":true}]},{"kpi":"customer-satisfaction","label":"Customer satisfaction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"montgomery-county-monty-chatbot","pooled":true}]},{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2,"min":2,"max":2,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dvla-contact-centre-conversational-ai","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":300000,"min":300000,"max":300000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dvla-contact-centre-conversational-ai","pooled":true}]}],"indicativeValueResult":{"low":480000,"high":3200000},"evidence":["abu-dhabi-tamm-ai-assistant","dvla-contact-centre-conversational-ai","estonian-information-system-authority-burokratt","fcdo-consular-enquiry-triage","gemeente-tilburg-vragen-ai","government-digital-service-gov-uk-chat","government-of-the-city-of-buenos-aires-boti","madrid-destino-visitmadridgpt","montgomery-county-monty-chatbot"]},{"title":"AI assistant for corporate and commercial client servicing","shortTitle":"Corporate client servicing","seoTitle":"AI assistant for corporate banking client service","metaDescription":"An AI assistant in the corporate banking portal answers payment and cut off questions and hands the rest to specialists. DBS Joy served about 4,000 clients a month.","definition":"A conversational assistant inside the corporate banking portal, app and messaging channels that answers finance and treasury teams' servicing questions, such as payment status, balances, cut off times, fees and how to submit an instruction, resolves routine requests end to end and hands the rest to a service specialist who has an AI copilot.","aliases":["corporate banking virtual assistant","transaction banking chatbot","cash management service assistant","business banking virtual agent"],"industries":["banking","payments"],"functions":["customer-service","operations"],"patterns":["conversational-agent","rag-knowledge-assistant","agentic-workflow","summarization"],"channels":["web-chat","mobile-app","agent-desktop","email","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"specialized-businesses","problem":"Corporate clients do not call about one card. Their finance and treasury teams ask where a\npayment is, why a file was rejected, what the cut off time is for a currency, how to add a user or\nreset a token, and what a fee on the analysis statement means. Many of these questions arrive at\nthe same moments (month end, payroll, a failed payment run), and every hour of delay can hold up a\nsupplier payment or a payroll.\n\nService teams for transaction banking are small and specialised, and much of their time goes to\nquestions whose answer already exists in a product guide or a status screen. Consumer style\nchatbots do not help much here: they lack the entitlement model of a corporate portal, cannot see\npayment status, and invent fees or cut off times when their content is thin.","problemStats":[],"howItWorks":"1. **Recognise the user and entitlements.** The assistant runs inside the authenticated portal\n   and only sees the accounts and functions the user is entitled to.\n2. **Answer from approved content.** Product guides, cut off tables, fee schedules and how to\n   articles come from the bank's own knowledge base, retrieved with citations.\n3. **Look things up.** Through read only APIs it checks payment status, balances, file\n   processing results and user administration status.\n4. **Complete simple requests.** Within an allow list (for example a token reset request or a\n   statement copy) it performs the action or opens a service request with the details filled in.\n5. **Hand over with context.** Anything outside the allow list, or where the client asks for a\n   person, goes to a service specialist with a summary, and the specialist's copilot drafts the\n   reply from the same knowledge base.","valueDrivers":["cost-to-serve","customer-experience","speed","employee-productivity"],"kpis":["containment-rate","contact-deflection","interactions-handled","users-served","response-time-reduction","customer-satisfaction-uplift"],"indicativeValue":{"referenceOrg":"A transaction bank serving 5,000 corporate and SME clients through its online platform","inputs":[{"key":"clients","label":"Active corporate and SME clients","low":5000,"high":5000,"unit":"clients","note":"The reference bank."},{"key":"requestsPerClient","label":"Servicing requests per client per year","low":10,"high":20,"unit":"requests per client per year","note":"Editorial assumption, replace with your own service desk volumes."},{"key":"containment","label":"Share of requests the assistant resolves without a specialist","low":0.2,"high":0.4,"unit":"fraction of requests","note":"Editorial assumption. No bank on this page publishes a resolution rate. The high value is capped at the more than 40% of CashPro Chat client interactions Bank of America says Erica handles, which is a handled share used here as an upper bound, not a resolution rate."},{"key":"costPerRequest","label":"Cost of a specialist handled request","low":15,"high":30,"unit":"USD per request","note":"Editorial assumption for a specialised transaction banking service desk. Replace with your own loaded cost."}],"formula":"clients * requestsPerClient * containment * costPerRequest","currency":"USD","period":"per year","resultLabel":"Specialist service cost avoided","caveat":"Gross avoided service cost only. It leaves out the cost of the assistant and integrations, the time saved by the specialist copilot on the requests that are handed over, and the value to the client of faster answers at month end."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering from product content is straightforward. The effort is in the portal's entitlement model, read access to payment and file status, and a clean handover into the service desk tool.","dataPrerequisites":["Current product guides, cut off tables and fee schedules with named owners","Service request categories with volumes from the service desk","Payment, file and user administration status reachable through APIs"],"integrations":["Corporate banking portal and app (authentication and entitlements)","Payment hub and file processing status","Service desk or CRM case management for handover","Specialist desktop for the copilot"]},"implementation":{"steps":[{"title":"Rank requests by volume and risk","detail":"Take a quarter of service desk tickets, group them into intents, and start with the high volume informational ones (payment status, cut off times, how to) before any action."},{"title":"Build on the portal's entitlements","detail":"Let the assistant call only the APIs the logged in user could use in the portal, and test that a user from one entity never sees another entity's data."},{"title":"Own the content","detail":"Give every guide and fee table an owner and a review date, and make the assistant refuse when retrieval finds nothing rather than guess a fee or cut off time."},{"title":"Equip the specialists","detail":"Put a copilot on the specialist desktop that drafts replies from the same content and shows the assistant's conversation summary, so handovers are fast."},{"title":"Review conversations, not only dashboards","detail":"Have experienced service staff review a sample of conversations every week and feed corrections back into the content. DBS uses experienced customer service agents as DBS Joy evaluators, who assess the quality of responses after the chat and suggest improvements."}],"guardrails":["Access limited to the user's portal entitlements, enforced on every API call","Answers on fees, cut off times and terms only from approved content, with citations","Read only by default; any action goes through an allow list with its own authentication level","A visible way to reach a human at any point","Full transcripts retained for disputes and complaints"],"humanInTheLoop":"Service specialists handle every request outside the allow list, every complaint and any case where the client asks for a person. Experienced staff review a sample of assistant conversations each week and approve content changes before they go live.","kpisToInstrument":["Containment per intent, counting a repeat request within seven days as not contained","Time to first answer and time to resolution against the specialist channel","Client satisfaction on assistant conversations and on handed over cases","Share of answers with a citation, and answers corrected by evaluators","Monthly active client users of the assistant"],"failureModes":[{"title":"Invented fees or cut off times","detail":"The assistant answers from general knowledge when content is missing. Force refusal when retrieval is empty and test with questions outside the content."},{"title":"Entitlement leaks","detail":"A user sees another entity's payments through the assistant. Enforce entitlements in the API layer, not in the prompt, and include cross entity tests."},{"title":"Handover that loses context","detail":"The specialist starts again and the client repeats everything. Pass the summary, the identified entity and the steps already tried."},{"title":"Month end overload","detail":"Volume spikes expose slow integrations and time outs. Load test at month end volumes before launch."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A chatbot that interacts with people at client companies must disclose that it is AI (Article 50). It does not evaluate creditworthiness or decide on access to an essential service (Annex III point 5), so it is not high risk."},"regulations":["eu-ai-act","gdpr","dora","mas-ai-risk-management","hkma-genai","apra-cps-230"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Users must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"MAS Guidelines for Artificial Intelligence (AI) Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Consultation paper issued on 13 November 2025 proposing supervisory expectations for all financial institutions on AI oversight, AI inventories and life cycle controls, covering generative AI and AI agents."}],"controls":["AI disclosure in the assistant and a documented route to a human","Inventory entry with an owner, the content sources and the action allow list","Entitlement tests and regression tests on every release","Transcript retention in line with the bank's record keeping rules","Weekly quality review by experienced service staff"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** embedded in the corporate portal through the **web widget**\nor the REST and WebSocket **API channel**, with a **knowledge base** of product guides, fee\nschedules and cut off tables retrieved with hybrid search. **Custom functions** call the bank's\npayment status and service request APIs, which apply the portal's own entitlement checks, so\naccess control stays in the bank's API layer. Regulated steps, such as a token reset request, can run as a **flow** with fixed\nsteps that calls the bank's own authentication service.\n\n**Human handover** passes the conversation, optionally summarised by AI, to the service desk, including Salesforce,\nand **live takeover** lets a specialist join a running conversation. **Guardrails** and **PII\nmasking** apply to every message, **test suites** replay real questions on each content or\nprompt change, and **monitors** check the key answers on a schedule. Analytics show volumes, top\nintents and satisfaction, and the model can be chosen or switched per agent."},"faq":[{"question":"How is a corporate servicing assistant different from a retail chatbot?","answer":"It works inside the corporate portal's entitlement model, answers treasury and payments questions rather than card questions, and hands over to specialised service teams. DBS runs DBS Joy inside its IDEAL platform, and said in November 2025 that about 4,000 corporate clients used it every month."},{"question":"How much of the chat volume can it take on?","answer":"None of the banks cited on this page publishes a resolution rate for its corporate assistant. Bank of America says Erica handles more than 40% of client interactions in CashPro Chat, and that chats with a live agent fell 16% after the Erica integration while chat volume rose 41%. Results depend on the intent mix and on whether the assistant can see payment and file status."},{"question":"How do banks keep the answers accurate?","answer":"By grounding answers in the bank's own knowledge base, filtering responses through rule based checks, and having experienced service staff assess responses and suggest improvements, as DBS describes for DBS Joy."}],"related":["payment-investigations-and-exceptions","client-briefing-and-call-report-copilot","developer-api-integration-assistant","treasury-cash-flow-forecasting","corporate-account-onboarding-orchestration","account-and-card-servicing-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog and verified against DBS and Bank of America press releases."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription, corrected the MAS guidance note (consultation paper), narrowed evidence channels to what the sources state, added the 65% CashPro Chat adoption figure to the evidence summary, and removed Blits.ai capabilities not in the feature inventory (authentication block, handover reason analytics)."},{"date":"2026-09-27","note":"Review fixes: the CashPro Chat 40% is now presented as a handled share, not a resolution rate; the DBS evaluator practice is quoted as DBS describes it, without an attributed weekly cadence; the custom functions sentence no longer claims end user token passthrough; KPIs aligned with the evidence; DBS user figure dated to November 2025."},{"date":"2026-09-27","note":"Review fixes: the resolution rate FAQ answer now scopes the claim to the banks cited on this page, not the whole market; the DBS Joy and CashPro Chat evidence records no longer state an unsourced English only language."}],"slug":"corporate-client-servicing-assistant","url":"https://www.blits.ai/ai-use-cases/corporate-client-servicing-assistant","benchmarks":[{"kpi":"contact-deflection","label":"Contact deflection","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":16,"min":16,"max":16,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-cashpro-chat","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":120000,"min":120000,"max":120000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dbs-joy-corporate-virtual-assistant","pooled":true}]},{"kpi":"customer-satisfaction-uplift","label":"Satisfaction uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":23,"min":23,"max":23,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dbs-joy-corporate-virtual-assistant","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":4000,"min":4000,"max":4000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dbs-joy-corporate-virtual-assistant","pooled":true}]}],"indicativeValueResult":{"low":150000,"high":1200000},"evidence":["bank-of-america-cashpro-chat","dbs-joy-corporate-virtual-assistant"]},{"title":"AI assistant for deal sourcing and M&A due diligence","shortTitle":"Deal sourcing and due diligence","seoTitle":"AI for deal sourcing and M&A due diligence","metaDescription":"AI that screens targets, reads data rooms and drafts diligence memos for deal teams to verify, with evidence from EQT, Freshfields, Datasite and Rogo.","definition":"An AI assistant that screens the market for acquisition or investment targets, builds company profiles, and speeds up due diligence by reading data room documents, extracting key terms and risks and drafting the investment or diligence memo, for the deal team to verify and decide.","aliases":["deal sourcing AI","M&A due diligence AI","private equity target screening","data room document review assistant","investment memo drafting assistant"],"industries":["capital-markets","wealth-and-asset-management","professional-services"],"functions":["analytics-and-reporting","legal","risk-management"],"patterns":["document-processing","summarization","rag-knowledge-assistant","agentic-workflow","prediction-and-scoring"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"front-office","problem":"Deal teams spend much of their time on work that comes before judgment. On the sourcing side,\nprivate equity firms, corporate development teams and bankers track large numbers of companies\nto find the few that fit a thesis, often with analysts who assemble lists and one pagers by\nhand. Testing whether one business fits the thesis can take an analyst 20 to 25 hours, according\nto a startup founder quoted in EQT's ThinQ publication.\nGood targets are missed because nobody noticed the signal in time, and the same research is\nredone for every new mandate because lessons from past deals sit in individual inboxes.\n\nOnce a deal is live, the data room opens and the clock starts. Associates, lawyers and advisers\nread contracts, financial statements, customer agreements and corporate records to find change\nof control clauses, unusual liabilities, customer concentration and missing documents, then\nwrite it up. Sellers must redact personal data before bidders see it. All of this is repetitive,\nhigh stakes and time boxed, and fatigue raises the risk that material issues slip through.","problemStats":[{"statement":"Datasite's chief product officer says sellers might have up to 100,000 documents associated with a deal.","sourceTitle":"Datasite automates M&A and speeds redaction by 80%, saving customers valuable time with Azure Cognitive Services","sourceUrl":"https://www.microsoft.com/en/customers/story/1379631359425784399-datasite-banking-capital-markets-azure-cognitive-services","year":2021},{"statement":"Testing whether a business fits a private equity firm's thesis and has value creation potential can take an analyst 20 to 25 hours, according to Clarum cofounder Anton Otaner.","sourceTitle":"AI Promises to Make Private Equity Faster as Competition Heats Up","sourceUrl":"https://eqtgroup.com/thinq/technology/first-ai-native-private-equity-firm","year":2025}],"howItWorks":"1. **Screen the market against the thesis.** The assistant maps companies from licensed data,\n   filings, news and the firm's own CRM, finds similar companies and ranks them against the\n   investment thesis and signals such as growth, hiring or ownership changes, with the reasons.\n2. **Build the company profile.** For a shortlisted target it drafts a profile: business model,\n   financials, competitors, ownership, management and news, with a source for every figure, and\n   adds what the firm learned from comparable past deals.\n3. **Read the data room.** When diligence starts it classifies the documents, extracts key terms\n   (change of control, exclusivity, termination, liabilities, key customers and suppliers) into a\n   structured table and flags what is missing against the diligence request list.\n4. **Flag risks and redact.** It highlights clauses and figures that deviate from the norm, and on\n   the sell side finds personal and sensitive data for batch redaction before bidders get access.\n5. **Answer questions with citations.** Deal team members ask questions across the whole data room\n   and get answers that link to the exact page, so every statement can be checked.\n6. **Draft the memo.** It drafts the diligence findings and the investment committee memo from the\n   firm's template; the deal team verifies, completes and owns every conclusion.","valueDrivers":["speed","employee-productivity","risk-reduction"],"kpis":["productivity-gain","time-saved-per-task","hours-saved","users-served","accuracy"],"indicativeValue":{"referenceOrg":"A mid market private equity firm that takes 15 companies a year into full due diligence","inputs":[{"key":"deals","label":"Companies taken into full due diligence per year","low":15,"high":15,"unit":"deals per year","note":"The reference firm."},{"key":"hoursPerDeal","label":"Internal hours per deal on document review, extraction and memo drafting","low":300,"high":600,"unit":"hours per deal","note":"Editorial assumption for the deal team's own time, excluding external advisers. Replace with your own time records."},{"key":"shareSaved","label":"Share of those hours the assistant saves","low":0.2,"high":0.4,"unit":"fraction of hours","note":"Conservative against the evidence on this page (Microsoft reports that Datasite's Redaction AI cuts redaction times by up to 80 percent), because redaction is the most mechanical step and review and judgment stay with the team."},{"key":"hourlyCost","label":"Blended cost per deal team hour","low":150,"high":300,"unit":"USD per hour","note":"Editorial assumption for a blended associate and principal cost. Replace with your own."}],"formula":"deals * hoursPerDeal * shareSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Deal team time released in due diligence","caveat":"Internal time only. It leaves out savings on external advisers, the value of deals found earlier or not missed in sourcing, the effect of issues caught or missed on deal value, and the cost of the assistant and data licences."},"macroEstimates":[{"statement":"Allvue research cited by EQT found that 82 percent of private equity and venture capital firms reported using AI in some capacity by the end of 2024, up from 47 percent a year earlier.","sourceTitle":"Why Private Capital Needs to Embrace Artificial Intelligence","sourceUrl":"https://eqtgroup.com/thinq/technology/why-private-capital-needs-to-embrace-artificial-intelligence","year":2025}],"feasibility":{"complexity":"high","complexityNote":"Summarizing one document is easy. The hard parts are licensed market data and entity matching for sourcing, secure access to data rooms under strict confidentiality, reliable extraction from scanned and inconsistent documents, citations for every statement, and a review process that deal teams, lawyers and investment committees accept.","dataPrerequisites":["A written investment thesis and screening criteria per strategy or mandate","Licensed company, financial and transaction data, plus the firm's CRM and past deal records","A diligence request list and a key terms checklist per deal type","Memo and findings templates approved by the investment committee","Confidentiality and information barrier rules per deal"],"integrations":["Market data and company databases (for example PitchBook, S&P Global, FactSet, Preqin)","CRM or deal pipeline system","Virtual data room and document management","Document storage such as SharePoint for memos and past deal files","Collaboration tools where the deal team works"]},"implementation":{"steps":[{"title":"Start with one deal type and one step","detail":"Pick the most repetitive step for your team, such as first pass contract review for one deal type or company profiles for one strategy, and measure it before widening the scope."},{"title":"Write the checklist before the prompt","detail":"Turn the diligence request list and key terms checklist into explicit fields with definitions, so the extraction is complete and comparable across deals."},{"title":"Make citations mandatory","detail":"Every extracted term, figure and memo statement links to the page it came from, and the assistant says when something is not in the data room rather than guessing."},{"title":"Test on closed deals","detail":"Run the assistant on data rooms and outcomes from past deals and compare its findings with what the team and advisers found, including the issues that mattered most."},{"title":"Protect the deal","detail":"Keep each deal's documents in a separate, access controlled space, enforce information barriers, and make sure no deal data trains or reaches an unapproved model."},{"title":"Define who signs what","detail":"Agree with the deal team, counsel and the investment committee which outputs are drafts, who reviews them and how reviewed findings are marked in the memo."}],"guardrails":["Every statement in a profile, table or memo cites its source page, with a refusal when the source is missing","Deal data isolated per deal and user, with information barriers and no training on client data","Extraction checked against a fixed checklist, with confidence flags on uncertain fields","Redaction reviewed by a person before any document is released to bidders","Screening criteria documented, so targets are not excluded for reasons nobody can explain"],"humanInTheLoop":"The deal team owns sourcing decisions, the diligence findings and the memo. Associates check extracted terms against the source, counsel reviews legal findings and redactions, and the investment committee decides on an explicitly human reviewed memo. A sample of AI findings is compared with adviser reports after each deal to calibrate trust.","kpisToInstrument":["Hours per deal on document review and memo drafting, before and after","Recall of material issues on a test set of closed deals, compared with the team's own review","Share of extracted terms corrected by reviewers","Time from data room opening to first findings","Share of sourced targets that reach a first meeting or a term sheet"],"failureModes":[{"title":"A fluent memo with a wrong number","detail":"A figure is misread from a scanned document or taken from the wrong period, and it survives into the investment committee memo. Require citations and check every material figure against the source."},{"title":"Silence taken for comfort","detail":"The assistant finds nothing on an issue because the document is missing or unreadable, and the team reads that as no issue. Report gaps against the request list explicitly."},{"title":"Confidential deal data leaking","detail":"Documents from one deal reach another team, another deal or an external model. Isolate data per deal and control which models and tools may see it."},{"title":"Sourcing that only finds the obvious","detail":"Rankings built on the same data every competitor licenses surface the same companies. Combine proprietary signals and past deal knowledge, and review targets the model ranked low."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Decision support for professional investors and advisers about companies is not a use listed in Annex III and is not a practice prohibited by Article 5. The users are deal professionals who know they are working with an AI tool, and no consumer interacts with it, so the Article 50(1) duty to disclose an AI interaction has little practical effect. Article 50(2) is different: a firm that builds the assistant itself, including on a platform such as Blits.ai and putting it into service under its own name, is the provider of that system and must mark generated text in a machine readable format, unless the system only performs an assistive function for standard editing or does not substantially alter the input data or its semantics, which may cover extraction and redaction. A firm that instead licenses a vendor product, such as Datasite or Rogo, should confirm that the vendor meets this duty. Obligations are otherwise general: AI literacy for the deal team under Article 4 and, where personal data in the data room is processed, the GDPR."},"regulations":["eu-ai-act","gdpr","uk-gdpr","eu-mar"],"guidance":[{"title":"Risk Outlook report: The use of artificial intelligence in the legal market","issuer":"Solicitors Regulation Authority","region":"europe","url":"https://www.sra.org.uk/sra/research-publications/artificial-intelligence-legal-market/","note":"The regulator of solicitors in England and Wales on the risks of AI in legal work, including accuracy, confidentiality and supervision, relevant when AI supports legal due diligence."},{"title":"Guidance on AI and data protection","issuer":"Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/artificial-intelligence/guidance-on-ai-and-data-protection/","note":"How data protection law applies to AI systems, relevant to the employee and customer personal data found in data rooms."}],"controls":["Inventory entry for the assistant with an owner, approved data sources and approved models","Access control and information barriers per deal, with logging of every query and document read","Insider list and inside information handling for deals involving listed companies","Human review and sign off recorded for every finding that enters the investment committee memo","Periodic accuracy testing on closed deals and after every model change"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** with **human in the loop** approval. For sourcing,\nthe agent uses the built in **web search and web page browsing** tools, **custom functions**\nthat call the firm's licensed data providers and **SQL knowledge bases** over the deal pipeline\nand past deal records, and writes company profiles as **structured output** for the deal team.\nFor diligence, data room exports and past memos go into a **knowledge base** that ingests PDF,\nWord, Excel and Outlook email files, and the agent answers from passages found with **hybrid\nretrieval**. The prompt and flow design require a citation to the source document for every\nstatement.\n\nA **flow** drives the key terms checklist deal by deal, and the **agent** drafts findings and\nthe memo in the firm's template, which the team reviews and completes. **Tenant isolation**\nkeeps the firm's data apart from other customers. With a separate bot and knowledge base per\ndeal, **role based access** with per bot roles limits who sees which deal. **PII masking** hides\npersonal data at the gateway, a per tenant **audit log** records user actions, and input and\noutput **guardrails** screen for prompt injection attempts. **Test suites**\nreplay questions on closed deals after every change, and the platform is model agnostic, with\nEU and UAE data residency, so the firm can choose where its deal data is processed."},"faq":[{"question":"Can AI do due diligence on its own?","answer":"No. It can read, extract, compare and draft, and Datasite's chief product officer says its AI features can potentially compress weeks of work into days, but the findings and the decision stay with people. An article in EQT's ThinQ publication reports a consensus that screening and early diligence are ripe for automation, while confirmatory checks and negotiation are not."},{"question":"How much time does AI save in M&A due diligence?","answer":"It depends on the step. Microsoft's case study on Datasite reports that Redaction AI lets sell side bankers and lawyers reduce redaction times by up to 80 percent, a figure Datasite bases on what customers tell it. Review and memo drafting are likely to save less because every finding needs checking, so measure hours per deal before and after."},{"question":"How do private equity firms use AI for deal sourcing?","answer":"EQT, for example, says it uses its Motherbrain platform to source deals, including a model that measures how similar companies are for tasks such as competitor mapping, and an EQT partner describes tools that rank potential targets by attractiveness across a range of criteria. EQT stresses that AI supports, rather than replaces, human decision making."},{"question":"Is it safe to put data room documents into an AI tool?","answer":"Only with controls: data isolated per deal, information barriers, no training on client data, logging of every access and approved model providers in approved regions. Freshfields, for example, runs its Dynamic Due Diligence tool on Google's Gemini models as part of a strategic collaboration with Google Cloud, and says its people use Gemini, NotebookLM Enterprise and Google Workspace daily with strong governance."}],"related":["vendor-due-diligence","investment-research-summarization","credit-memo-drafting-agent","procurement-contract-review","intelligent-document-processing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, deal sourcing and M&A due diligence for investors, banks and law firms; evidence from EQT, Freshfields, Datasite and Rogo."},{"date":"2026-09-26","note":"Fact checked against sources. Added seoTitle and metaDescription, source dates, a fuller EU AI Act basis (Articles 4, 5 and 50), attribution of the 80 percent redaction figure to Microsoft's case study, and FAQ answers and a Blits.ai section rewritten to match what the sources and the feature inventory state."},{"date":"2026-09-27","note":"Review fixes. Removed unsourced quantities from the problem and added the ThinQ figure of 20 to 25 analyst hours for thesis testing, limited the Blits.ai section to capabilities in the feature inventory (document citations by design, per tenant audit log, per deal bots), explained Article 50(2) in the EU AI Act basis, moved the Datasite redaction metric to productivity gain, and aligned FAQ wording with the sources."},{"date":"2026-09-27","note":"Review fix: corrected the Article 50(2) basis, a firm that builds the assistant itself is the provider and carries the marking duty, not its model and platform providers; stated both exemption limbs (standard editing assistance, or no substantial alteration of the input or its semantics); and changed the EU AI Act tier to context dependent to match."}],"slug":"deal-sourcing-and-due-diligence-assistant","url":"https://www.blits.ai/ai-use-cases/deal-sourcing-and-due-diligence-assistant","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":6000,"min":6000,"max":6000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"rogo-investment-banking-research-agents","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"datasite-redaction-ai","pooled":false}]}],"indicativeValueResult":{"low":135000,"high":1080000},"evidence":["datasite-redaction-ai","eqt-motherbrain-deal-sourcing","freshfields-dynamic-due-diligence","rogo-investment-banking-research-agents"]},{"title":"AI assistant for developers integrating a company's APIs","shortTitle":"API integration assistant","seoTitle":"AI assistant for developer portals and API docs","metaDescription":"AI assistants on developer portals answer API questions, suggest endpoints and generate sample calls. Mapbox reports 30% fewer monthly support tickets with one.","definition":"An AI assistant on a developer portal and in its documentation that answers integration questions, recommends the right endpoints, helps debug connections and generates sample calls, grounded in the API catalogue, reference docs and test material, so clients and partners integrate faster with fewer support tickets.","aliases":["developer portal assistant","API documentation chatbot","ask AI for docs","integration support copilot"],"industries":["cross-industry","banking","payments","technology"],"functions":["it-and-engineering","customer-service","onboarding-and-kyc"],"patterns":["rag-knowledge-assistant","conversational-agent","code-generation"],"channels":["web-chat","api","agent-desktop"],"audience":"customer-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"specialized-businesses","problem":"Every company that sells through APIs, from a bank embedding payments and treasury services in a\nclient's ERP to a software platform, depends on outside developers getting their integration to\nwork. Those developers search long reference docs, try calls in a sandbox, hit an error and file a\nticket, then wait. Many of those questions already have an answer somewhere in the documentation\nor in a past ticket. The Mapbox case study on this page notes that by the time a ticket was filed,\nthe answer usually already existed in the documentation.\n\nFor a bank, slow integration delays the start of transaction revenue and ties up implementation\nmanagers and support engineers on repetitive questions. The assistant pattern already runs in\nproduction at software companies such as Mapbox, CircleCI and monday.com; the banking specific part is keeping it strictly away from production data,\nlive credentials and client entitlements.","problemStats":[],"howItWorks":"1. **Index the developer knowledge.** API reference, guides, SDKs, changelogs, sample code, test\n   scripts and resolved support tickets are indexed and refreshed as they change.\n2. **Answer in context.** In the docs, the portal or the sandbox, the assistant answers questions\n   with citations to the exact page or endpoint.\n3. **Recommend and generate.** It suggests which endpoints fit the use case and generates sample\n   requests and code in the developer's language, using sandbox values only.\n4. **Help debug.** Given an error message or a failing request (with secrets removed), it explains\n   the likely cause and the fix.\n5. **Escalate.** Anything it cannot answer, and anything touching production access or\n   entitlements, goes to a support engineer or implementation manager with the conversation.","valueDrivers":["customer-experience","cost-to-serve","speed","revenue-growth"],"kpis":["contact-deflection","response-time-reduction","hours-saved","interactions-handled","containment-rate","processing-time-reduction"],"indicativeValue":{"referenceOrg":"A bank running 400 client and partner API integrations a year","inputs":[{"key":"integrations","label":"Client and partner integrations per year","low":400,"high":400,"unit":"integrations per year","note":"The reference bank."},{"key":"supportHours","label":"Support and implementation hours per integration","low":20,"high":40,"unit":"hours per integration","note":"Editorial assumption, replace with your own ticket and implementation data."},{"key":"deflection","label":"Share of those hours the assistant saves","low":0.15,"high":0.25,"unit":"fraction of hours","note":"Kept below the 30% reduction in monthly support tickets from paid users that Mapbox reports on this page, because that figure covers tickets only and implementation hours usually fall less than ticket volume. Start from the low end unless your own data says otherwise."},{"key":"hourlyCost","label":"Loaded cost of a support or implementation engineer hour","low":90,"high":130,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."}],"formula":"integrations * supportHours * deflection * hourlyCost","currency":"USD","period":"per year","resultLabel":"Support and implementation time released, valued at loaded cost","caveat":"Values engineering time only. It leaves out the revenue from integrations that go live sooner, the cost of the assistant, and the documentation improvements that its unanswered questions reveal."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Mostly a retrieval assistant over public or partner facing documentation. The effort is in keeping the index current, adding the sandbox and ticket knowledge, and enforcing the boundary with production systems.","dataPrerequisites":["Current API reference (for example OpenAPI specifications), guides and changelogs","Sample code and test scripts for the sandbox","Resolved support tickets, cleaned of client data"],"integrations":["Developer portal and documentation site","Sandbox environment (read only for the assistant)","Support ticketing system for escalation"]},"implementation":{"steps":[{"title":"Clean and connect the sources","detail":"Start from the API specifications and guides, add resolved tickets after removing client data, and set a refresh schedule tied to documentation releases."},{"title":"Launch in the docs first","detail":"Put an ask AI entry point on the documentation pages, where developers already are, then add it to the portal and sandbox."},{"title":"Keep it in the sandbox","detail":"Generate samples with sandbox hosts and placeholder credentials only, and refuse questions that ask for production data or client specific configuration."},{"title":"Learn from unanswered questions","detail":"Review questions the assistant could not answer every week and fix the documentation, not only the prompts."}],"guardrails":["Access to documentation and sandbox only, never to production systems, live credentials or client data","Secrets and tokens pasted by developers are masked before they reach the model","Answers cite the documentation page or endpoint they rely on","Anything about production access or entitlements is routed to a human implementation manager"],"humanInTheLoop":"Support engineers and implementation managers handle escalations and anything involving production access. The developer relations or documentation team reviews unanswered and poorly rated questions weekly and owns the content.","kpisToInstrument":["Support tickets per integration, before and after","Time from sandbox access to first successful production call","Questions answered and share rated helpful","Unanswered questions per documentation area"],"failureModes":[{"title":"Invented endpoints or parameters","detail":"The model generates plausible but wrong calls. Ground answers in the specification, cite it, and test generated samples against the sandbox."},{"title":"Secrets in the chat","detail":"Developers paste API keys or tokens. Mask secrets at input and warn the user."},{"title":"Outdated answers after a release","detail":"The index lags a new API version. Tie reindexing to documentation releases and show the version an answer refers to."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A chatbot that interacts with developers must disclose that it is AI (Article 50). Code generation for integration is not listed in Annex III."},"regulations":["eu-ai-act","gdpr","dora","iso-42001"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Developers must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"OWASP Top 10 for Large Language Model Applications","issuer":"OWASP","region":"global","url":"https://genai.owasp.org/llm-top-10/","note":"The 2025 list covers prompt injection, sensitive information disclosure and improper output handling, the main risks of an assistant that generates code."}],"controls":["Documented boundary between the assistant and production systems","Secret masking at input and logging without credentials","Conversation logs retained and reviewed for quality","Change control on the indexed sources"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** built from the API reference,\nguides and resolved tickets, ingested from files and crawled documentation pages that are\nrecrawled after each release, and retrieved with hybrid search so exact endpoint names match. It runs in the **web\nwidget** on the documentation site and developer portal, where code blocks with copy to clipboard\nrender in the chat.\n\n**Guardrails** keep it to integration topics, **PII masking** with custom patterns removes keys\nand tokens before text reaches a model, and **human handover** sends escalations to the support\ndesk with the conversation. **Response feedback** on poorly rated answers shows where the docs need work, and\n**test suites** check answers on known questions after each documentation release. The model can\nbe switched per agent without rebuilding."},"faq":[{"question":"Do developer assistants reduce support tickets?","answer":"Published vendor case studies say so. Mapbox reports a 30% monthly reduction in support tickets and CircleCI 28% faster support response times. At monday.com the assistant answers more than 125,000 technical queries a year. These are vendor case studies, not independent measurements."},{"question":"Are banks doing this?","answer":"Yes. U.S. Bank added a generative AI Developer Assistant to its Developer Portal that recommends APIs, helps troubleshoot and generates sample code, aiming to cut integration time by weeks."},{"question":"What must a bank keep out of the assistant?","answer":"Production data, live client credentials and entitlements. Keep it on documentation and the sandbox, mask any secrets developers paste, and route production questions to a person."}],"related":["developer-coding-assistant","corporate-client-servicing-assistant","corporate-account-onboarding-orchestration","support-knowledge-article-generation","it-service-desk-resolution-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, industry neutral with banking notes, verified against U.S. Bank and three named software company case studies."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: capped the indicative deflection input at the 30% Mapbox benchmark, added contact deflection to the KPIs, updated the OWASP note to the 2025 list, aligned the Blits.ai build notes with the feature inventory, added SEO title and description, and added volume and hours metrics to the CircleCI and monday.com records."},{"date":"2026-09-27","note":"Review fixes: removed unsourced \"cumulative\" periods from the CircleCI and monday.com metrics, attributed the documentation claim to the Mapbox case study, lowered the indicative deflection range to 15 to 25%, pointed the Article 50 guidance to EUR-Lex and dropped the untrained questions analytics claim."}],"slug":"developer-api-integration-assistant","url":"https://www.blits.ai/ai-use-cases/developer-api-integration-assistant","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":3,"nUpTo":0,"median":125000,"min":32000,"max":175000,"byClaimant":{"organization":0,"vendor":3,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"mapbox-docs-ai-assistant","pooled":true},{"id":"monday-com-developer-docs-assistant","pooled":true},{"id":"circleci-docs-ai-assistant","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":20250,"min":500,"max":40000,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"mapbox-docs-ai-assistant","pooled":true},{"id":"circleci-docs-ai-assistant","pooled":true}]},{"kpi":"contact-deflection","label":"Contact deflection","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":30,"min":30,"max":30,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"mapbox-docs-ai-assistant","pooled":true}]},{"kpi":"response-time-reduction","label":"Response time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":28,"min":28,"max":28,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"circleci-docs-ai-assistant","pooled":true}]}],"indicativeValueResult":{"low":108000,"high":520000},"evidence":["circleci-docs-ai-assistant","mapbox-docs-ai-assistant","monday-com-developer-docs-assistant","us-bank-developer-assistant"]},{"title":"AI assistant for digital account onboarding and KYC","shortTitle":"Digital onboarding","seoTitle":"AI assistant for digital onboarding and KYC","metaDescription":"An AI assistant guides applicants through identity and KYC checks and sends unclear cases to reviewers. Deutsche Bank and M-DAQ Global use AI for KYC and KYB checks.","definition":"A customer facing AI assistant that guides a new applicant, a person or a small merchant, through a digital account, card or relationship application: it collects and checks identity and supporting documents, orchestrates the know your customer and anti money laundering checks, prefills what it can and sends only the unclear cases to a human reviewer with a summary. The ownership research for complex corporate clients is a separate back office job.","aliases":["digital account opening assistant","KYC onboarding agent","e-KYC assistant","customer onboarding chatbot"],"industries":["banking","payments","wealth-and-asset-management"],"functions":["onboarding-and-kyc","sales","customer-service"],"patterns":["conversational-agent","document-processing","computer-vision","agentic-workflow"],"channels":["mobile-app","web-chat","whatsapp","internal-tools"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"Onboarding is where a bank wins or loses a new customer, and it is also one of the most\nheavily regulated steps. The applicant has to choose a product, fill in long forms, photograph an identity\ndocument, pass a liveness check and sometimes explain where their money comes from. Every extra\nstep or unclear question costs completions, and every applicant who stalls either leaves or\ncalls the contact centre.\n\nBehind the form, operations and compliance teams do the slow part by hand: checking documents\nthat are blurred or expired, chasing missing information, screening names and, for wealth and\nbusiness clients, researching source of wealth or the structure of a company. Rules differ by\ncountry, so a bank that operates in several markets maintains several processes. Fully manual\nreview does not scale; fully automated rejection loses good customers and still misses\nsophisticated fraud such as synthetic identities and deepfakes.","problemStats":[],"howItWorks":"1. **Help the applicant choose.** The assistant asks a few needs questions and explains the\n   products the applicant is eligible for, in plain language and the applicant's own language,\n   without giving personal advice.\n2. **Collect data once.** It prefills from what the bank already knows (a lead form, an existing\n   relationship or a government digital identity where one is available) and asks only for what\n   is missing.\n3. **Capture and verify identity.** Document capture, authenticity checks, liveness and face\n   matching run through the bank's identity verification provider; the assistant explains\n   failures (\"the photo is blurred, try again in better light\") instead of ending the journey.\n4. **Run the checks.** An agentic workflow calls sanctions and politically exposed person\n   screening, fraud signals and, where needed, source of wealth or company registry research,\n   and records every result with its evidence.\n5. **Decide the route.** Clean cases continue straight to account opening under the bank's\n   rules; unclear or high risk cases go to a human reviewer with a summary of what is missing\n   or inconsistent. The approval decision on edge cases stays with a person.","valueDrivers":["revenue-growth","speed","compliance","cost-to-serve","inclusion-and-access"],"kpis":["processing-time-reduction","conversion-rate-uplift","automation-rate","productivity-gain","cycle-time-days"],"indicativeValue":{"referenceOrg":"A retail bank that receives 100,000 digital account applications a year","inputs":[{"key":"applications","label":"Digital applications started per year","low":100000,"high":100000,"unit":"applications per year","note":"The reference bank."},{"key":"abandonRate","label":"Share of started applications that are abandoned today","low":0.3,"high":0.5,"unit":"fraction of applications","note":"Editorial assumption, replace with your own funnel data."},{"key":"recoveredShare","label":"Share of abandoned applications the assistant recovers","low":0.1,"high":0.2,"unit":"fraction of abandoned applications","note":"Editorial assumption, deliberately conservative because no deployment on this page discloses a completion uplift."},{"key":"valuePerAccount","label":"First year revenue of a new account","low":50,"high":150,"unit":"USD per account","note":"Editorial assumption, replace with your own product economics."}],"formula":"applications * abandonRate * recoveredShare * valuePerAccount","currency":"USD","period":"per year","resultLabel":"First year revenue from recovered applications","caveat":"Revenue from recovered applications only. It leaves out the review hours saved in operations, lower fraud losses, the cost of the identity verification provider and the platform, and the lifetime value of an account beyond the first year."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The conversation is the easy part. The work is in orchestrating identity verification, screening and core account opening systems, keeping the rules configurable per jurisdiction, and producing an audit trail that satisfies compliance and the regulator.","dataPrerequisites":["Product eligibility rules and approved product descriptions per market","The bank's know your customer and anti money laundering policy, including risk ratings and document lists per jurisdiction","Funnel data showing where applicants abandon today","Labelled historical cases of manual review outcomes, to test routing"],"integrations":["Identity verification provider (document authenticity, liveness, face match)","Sanctions, politically exposed person and adverse media screening","Government digital identity schemes where they exist","Customer onboarding or account origination platform and core banking","Case management for manual review, with the assistant's summary attached"]},"implementation":{"steps":[{"title":"Map the funnel and the rules","detail":"Take the current application funnel and mark where applicants drop out and why. Write the verification rules per product and jurisdiction as configuration, not as prompts, so compliance can own and change them."},{"title":"Start with guidance and document help","detail":"The first release explains products, answers questions and helps applicants get their documents right. It changes no decisions, so it can ship early and still move completion."},{"title":"Automate the checks behind the form","detail":"Add the agentic workflow that runs screening and collects evidence, with every tool call logged. Measure how often the reviewer agrees with the prepared summary before you let any case skip review."},{"title":"Set the routing thresholds with compliance","detail":"Agree which cases may proceed without a human and which must be reviewed (high risk countries, politically exposed persons, mismatched data, low verification confidence). Keep human approval for every edge case."},{"title":"Test against fraud as well as friction","detail":"Build a test set with genuine applicants who struggle (poor lighting, unusual names, foreign documents) and with attack patterns (edited documents, replayed selfies, synthetic identities), and run it on every change."}],"guardrails":["Verification rules are configuration owned by compliance, never free text instructions to a model","The assistant cannot approve an edge case; ambiguous or high risk cases always go to a person","Personal data and identity images are masked in logs and never used for model training without consent","Clear disclosure of what data is collected, why, and that the applicant is talking to AI","No personal financial advice during product selection, only eligibility and product facts"],"humanInTheLoop":"Onboarding analysts review every case the rules mark as unclear or high risk, with the assistant's summary and evidence in front of them, and they own the approval. Compliance samples straight through cases every week and signs off every change to thresholds.","kpisToInstrument":["Application completion rate per step and per channel, before and after","Median time from start to account open","Share of cases sent to manual review, and reviewer agreement with the prepared summary","Fraud found after onboarding in accounts that went straight through","Applicant satisfaction and contact centre calls about applications"],"failureModes":[{"title":"Fast onboarding for fraudsters","detail":"Straight through rules tuned for conversion let synthetic identities in. Track fraud on new accounts by route and tighten thresholds when it rises."},{"title":"Silent unfair rejection","detail":"Document or liveness checks fail more often for some groups of applicants, who then give up. Monitor failure rates by document type, age band and language and offer an assisted route."},{"title":"One rulebook for many countries","detail":"A single process copied across markets breaks local rules. Keep verification logic configurable per jurisdiction with a named owner."},{"title":"Unexplainable decisions","detail":"A reviewer or regulator cannot see why a case was routed. Log every check, its result and the rule that applied."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The conversational assistant falls under the Article 50 transparency duty. Biometric verification whose sole purpose is to confirm that a person is who they claim to be is excluded from the Annex III biometric category. The system becomes high risk when the same journey assesses creditworthiness or a credit score of a natural person, for example for a credit card or overdraft (Annex III point 5(b))."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","dora","iso-42001","eu-amlr","us-bsa","mas-notice-626","eu-accessibility-act"],"guidance":[{"title":"Annex III: High-risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 1(a) excludes biometric verification that only confirms identity; point 5(b) makes creditworthiness assessment high risk."},{"title":"NIST SP 800-63 Digital Identity Guidelines","issuer":"NIST","region":"north-america","url":"https://pages.nist.gov/800-63-4/","note":"Reference for identity proofing and assurance levels, useful for setting verification strength per product."}],"controls":["Documented verification rules per jurisdiction with a named compliance owner","Audit trail of every check, tool call and routing decision per application","Bias monitoring of verification failure rates across applicant groups","Human approval for every case outside the straight through rules","AI disclosure and a data collection notice at the start of the journey"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the applicant facing side is an **AI agent** combined with a **flow** for the\nregulated steps: the flow collects fields with slot filling, receives document uploads with\nallowed file types and calls the bank's identity verification, screening and account opening\nservices through **custom functions** (REST calls). Product and\npolicy questions are answered from a **knowledge base** with hybrid retrieval over approved\ncontent only.\n\nThe checks behind the form run as an **agentic workflow** with a tool execution policy,\n**human in the loop** confirmation (approve or reject) before the workflow takes the actions you\nmark as sensitive, and a full audit trail per run. Which cases count as high risk and go to a\nreviewer is decided by rules in the flow or in custom functions that the bank's compliance\nteam owns. The same journey serves **web chat and WhatsApp**, and the bank's mobile\napp through the **REST or WebSocket API channel**, in several languages. **PII masking** at the gateway keeps personal data such as ID numbers out\nof model prompts and logs, **test suites** replay difficult applications on every change, and **EU or UAE data\nresidency** keeps applicant data in region. The platform is model agnostic."},"faq":[{"question":"Can AI approve new customers without a human?","answer":"A bank can let its rules approve clean, low risk cases straight through, with the AI preparing the evidence. Unclear or high risk cases should always reach a person. Deutsche Bank describes its Source of Wealth agent this way: it prepares assessments for staff to review and accountability stays with people."},{"question":"Is identity verification with face matching high risk under the EU AI Act?","answer":"Not by itself. Annex III excludes biometric verification whose only purpose is to confirm that someone is who they claim to be. The journey becomes high risk if it also scores the applicant's creditworthiness."},{"question":"Where does AI help most in onboarding?","answer":"In two places: helping applicants finish (explaining products and fixing document problems in the moment) and preparing the checks for reviewers. The deployments on this page show both, from a neobank chatbot for first time bank users to automated Know Your Business and Source of Wealth research."}],"related":["business-onboarding-and-ubo-discovery","application-and-identity-fraud-detection","perpetual-kyc","pep-and-adverse-media-screening","account-and-card-servicing-agent","conversational-loan-application-intake"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog and verified against the sources. The catalog's Veriff citation did not contain the drop off claim and was not used."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626, European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed an unsourced line from the Albo record and unsourced languages from two records, added the M-DAQ headquarters source and source dates, reworded an unsourced FAQ generalisation, aligned the Blits.ai build notes with the feature inventory, and added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: corrected the Blits.ai build notes so human in the loop confirmation applies to sensitive workflow actions and risk routing comes from compliance owned rules, narrowed the PII masking claim, toned down an unsourced claim about regulation, and aligned the meta description and the M-DAQ title with the sources."},{"date":"2026-09-27","note":"Second review pass: removed the MAS AI risk management guidelines from the regulations, since the source is a consultation paper (closed 31 January 2026) and no final guidelines could be found; MAS Notice 626 already covers Singapore."}],"slug":"digital-onboarding-assistant","url":"https://www.blits.ai/ai-use-cases/digital-onboarding-assistant","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":30,"min":30,"max":30,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"m-daq-global-kyb-onboarding","pooled":true}]}],"indicativeValueResult":{"low":150000,"high":1500000},"evidence":["albo-albot-ai-chatbot","deutsche-bank-source-of-wealth-kyc-agent","m-daq-global-kyb-onboarding"]},{"title":"AI assistant for employee onboarding","shortTitle":"Employee onboarding assistant","seoTitle":"AI employee onboarding assistant for new hires","metaDescription":"An AI onboarding assistant answers new hire questions, tracks the checklist and chases paperwork, access and training. Value model, evidence and EU AI Act risk.","definition":"An assistant that guides each new employee from signed contract through the first months: it answers first week questions in plain language, tracks the personal onboarding checklist, triggers the paperwork, equipment, access and training steps in the systems that own them, and keeps the manager and HR informed of what is still open.","aliases":["new hire assistant","onboarding chatbot","AI onboarding buddy","preboarding assistant","new joiner assistant"],"industries":["cross-industry","government","professional-services","healthcare"],"functions":["human-resources","knowledge-management"],"patterns":["conversational-agent","rag-knowledge-assistant","agentic-workflow"],"channels":["microsoft-teams","internal-tools","mobile-app","email"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Onboarding is where an employer makes its first impression on someone it has already paid to\nrecruit, and it is usually a patchwork. Tasks sit with HR, IT, facilities, payroll, security and\nthe manager, each with its own system and checklist. New hires do not know whom to ask, so they ask\neveryone, or nobody. Laptops arrive late, access requests wait for approval, mandatory training is\nmissed, and managers spend the first weeks answering the same questions every new starter has.\n\nThe cost is real: slower time to productivity, early attrition among people who leave in their\nfirst months, and compliance gaps when a mandatory step is skipped. An assistant helps in two ways.\nIt answers questions from the organization's own onboarding content at any hour, in the new\nhire's language, and it orchestrates the checklist, starting and chasing steps in the owning\nsystems so that nothing depends on memory. It does not replace the manager's welcome or the buddy;\nit frees them for it.","problemStats":[],"howItWorks":"1. **Start at signature.** When the hire is confirmed in the HR system, the assistant creates a\n   personal onboarding plan based on role, location, contract type and start date.\n2. **Preboard.** Before day one it collects documents and details through the HR system's own\n   forms, explains what to expect and answers questions about the first day.\n3. **Trigger the provisioning.** It opens the requests for equipment, accounts, access and badges\n   in IT and facilities systems, following the normal approvals, and tracks them.\n4. **Answer from approved content.** Questions about policies, tools, benefits and \"how do I\" are\n   answered from onboarding and HR content, with links, and routed to a person when the answer is\n   not there or the topic is sensitive.\n5. **Keep the plan moving.** It reminds the new hire of mandatory training and tasks, nudges owners\n   of overdue steps and gives the manager a view of what is complete.\n6. **Check in and hand over.** At set points it asks how things are going, passes concerns to HR or\n   the manager, and hands over to the general HR and IT assistants once onboarding ends.","valueDrivers":["employee-productivity","speed","cost-to-serve","compliance"],"kpis":["cycle-time-days","contact-deflection","time-saved-per-task","employee-adoption","containment-rate"],"indicativeValue":{"referenceOrg":"An organization that onboards 2,000 new employees a year","inputs":[{"key":"newHires","label":"New hires per year","low":2000,"high":2000,"unit":"hires per year","note":"The reference organization."},{"key":"supportHoursPerHire","label":"HR, IT and manager hours spent per hire on onboarding questions and chasing","low":4,"high":8,"unit":"hours per hire","note":"Editorial assumption. Replace with a time study of your own onboarding."},{"key":"shareSaved","label":"Share of those hours the assistant takes over","low":0.2,"high":0.4,"unit":"fraction of hours","note":"Editorial assumption for the whole range, replace with your own measurement. No cited source measures the share of onboarding support hours an assistant takes over; the low end assumes it handles routine questions, the high end adds time spent chasing provisioning steps."},{"key":"hourlyCost","label":"Blended hourly cost of HR, IT and managers","low":45,"high":70,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"newHires * supportHoursPerHire * shareSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Onboarding support time released","caveat":"Counts only support time. It leaves out faster time to productivity for the new hire, lower early attrition and fewer missed compliance steps, which you should estimate separately, and the cost of the platform and integrations."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering onboarding questions is quick to build. Orchestrating the checklist across HR, IT and facilities systems, with the right approvals and a plan per role and country, is where the effort goes.","dataPrerequisites":["Current onboarding content per country and role, with owners","The onboarding checklist per role, location and contract type, with the owner of each step","Start date, role and manager data from the HR system","The list of topics that must go to a person"],"integrations":["HR information system (for example Workday or SAP SuccessFactors)","IT service management for equipment, accounts and access requests","Identity provider for account creation and single sign on","Learning management system for mandatory training","Collaboration tools where employees work, such as Microsoft Teams"]},"implementation":{"steps":[{"title":"Map the journey and its owners","detail":"List every onboarding step from signature to the end of probation, who owns it, which system records it and what usually goes wrong. Fix obviously broken steps before automating them."},{"title":"Launch the question answering first","detail":"Load current onboarding content, filtered by country and role, and let new hires ask questions from before day one. Measure what they ask to find gaps in the content."},{"title":"Connect the checklist","detail":"Create the plan from the HR system and open requests in IT and facilities systems through their normal workflows and approvals, one step type at a time."},{"title":"Give managers a view","detail":"Show the manager what is complete and what is overdue, and send nudges to step owners rather than to the new hire."},{"title":"Design the human moments","detail":"Decide where a person must be present (welcome, first one to one, sensitive questions) and make the assistant route to them instead of trying to answer."},{"title":"Hand over and measure","detail":"At the end of onboarding, pass the employee to the general HR and IT assistants and measure time to productivity, early attrition and missed mandatory steps."}],"guardrails":["Answers only from approved onboarding content, with links, and a handover when the content has no answer","Provisioning through the owning systems and their approval workflows, never by direct changes","No evaluation of the new hire's performance or suitability; the assistant supports, managers assess","Sensitive topics (health, adjustments, grievances, pay disputes) routed to a person","Personal data read only through APIs scoped to the new hire and their manager"],"humanInTheLoop":"HR owns the content and the plan templates, IT and facilities approve provisioning in their own systems, and managers own the welcome, the check ins and every judgment about the new hire. HR reviews a sample of conversations each month for accuracy and for topics that should have been handed over.","kpisToInstrument":["Time from start date to equipment and access ready","Share of onboarding steps completed on time, per owner","Questions answered without a person, and handover reasons","New hire satisfaction with onboarding","Early attrition in the first 90 days, compared with before"],"failureModes":[{"title":"Automating a broken process","detail":"The assistant faithfully chases steps that nobody owns. Map owners and fix broken steps before connecting them."},{"title":"One size fits all","detail":"A contractor in one country gets the plan for an employee in another. Build plans from role, location and contract type."},{"title":"Drift into evaluation","detail":"Check in answers or training completion are used to judge new hires. That changes the risk class and needs its own assessment and consultation."},{"title":"The assistant replaces the welcome","detail":"Managers leave onboarding to the bot. Keep the human moments in the plan and make them visible to the manager."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Answering onboarding questions and orchestrating provisioning is limited risk: under Article 50(1) the assistant must be designed so that employees are told they are interacting with AI, unless that is obvious. It becomes high risk under Annex III point 4(b) if it is used to make decisions on the terms or termination of the work relationship, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate new hires' performance or behaviour, for example to judge probation."},"regulations":["eu-ai-act","gdpr","iso-42001"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) covers AI used for decisions on the terms, promotion or termination of work relationships, to allocate tasks based on behaviour or personal traits, or to monitor and evaluate workers."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Article 50(1) requires AI systems that interact directly with people to be designed so that they are told they are interacting with AI, unless this is obvious from the context."},{"title":"Employment practices and data protection: monitoring workers","issuer":"UK Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/","note":"UK guidance, under review after the Data (Use and Access) Act, on monitoring workers under the UK GDPR. Relevant to what the assistant logs about new hires and who may see it."}],"controls":["Data protection impact assessment covering what is logged about new hires","Content ownership and review dates for onboarding material per country","Consultation with employee representatives where required","Access control on onboarding progress data, limited to HR and the line manager"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** in **Microsoft Teams** (or the **web chat** bubble on the\nintranet, a mobile app through the **API channel**, or **email**), grounded in a **knowledge base**\nof onboarding content with hybrid retrieval and document version control, with separate content\nwhere policies differ by country. An **agentic workflow** started\nfrom the HR system through an **API token** builds the plan and uses **custom functions** to open\nrequests in IT and facilities systems (the integration catalog includes Workday, ServiceNow and\nOkta), with **agentic tasks** that recheck overdue steps on a schedule and nudge their owners.\n\n**Human in the loop** approval holds actions above a threshold you set, a **flow** routes sensitive\ntopics to HR through **human handover**, and **guardrails** with **PII masking** keep personal data\nout of prompts. An **authentication** step in the flow confirms who the employee is,\n**multi language** support serves new hires in their own language, and **test suites** check\nanswers per country before content changes go live. The platform is model agnostic and can run in\nthe EU or UAE region."},"faq":[{"question":"What does an onboarding assistant change in practice?","answer":"It takes routine questions off managers and HR and keeps paperwork, access and training steps moving, but published outcome data for dedicated onboarding assistants is thin. Microsoft says KPMG designed its onboarding agent to reduce follow up calls by 20%, a stated aim without a reported result. American Addiction Centers says Gemini for Google Workspace, a general productivity suite rather than an onboarding assistant, helped cut employee onboarding from three days to 12 hours."},{"question":"Is this the same as an HR chatbot?","answer":"It overlaps, but onboarding is a time bound journey with a checklist across HR, IT and facilities, not only questions. It can run as a mode of the same HR assistant, which takes over once onboarding is complete."},{"question":"Is an onboarding assistant high risk under the EU AI Act?","answer":"Not when it answers questions and orchestrates tasks. It becomes high risk under Annex III point 4(b) if it is used to evaluate new hires or allocate work based on their behaviour or traits, for example to judge probation."}],"related":["hr-and-policy-assistant","it-service-desk-resolution-agent","recruitment-screening-and-interview-scheduling","enterprise-knowledge-search","conversation-roleplay-training"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical with three evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: dated the American Addiction Centers record to 2024 using archived copies, clarified that the KPMG figure is a Microsoft summary, attributed the 20% in the value inputs to Microsoft and narrowed the share of hours saved to 0.2 to 0.4, added contact deflection to the KPIs, completed the Annex III point 4(b) and Article 50(1) wording, limited the Blits.ai build to listed capabilities, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Second fact check: the KPMG 20% is now recorded as a stated aim rather than a result, with the KPMG record moved to announced and the WorkLab article cited; the American Addiction Centers figure now cites the CIO quote of April 2024 and is described as a Gemini for Google Workspace result, not an onboarding assistant; the share of hours saved is marked as an editorial assumption; new meta description."},{"date":"2026-09-27","note":"Third review pass: removed the unsourced claim that the benefits left out of the value model are usually worth more than support time."}],"slug":"employee-onboarding-assistant","url":"https://www.blits.ai/ai-use-cases/employee-onboarding-assistant","benchmarks":[{"kpi":"cycle-time-days","label":"Cycle time","unit":"hours","aggregate":false,"higherIsBetter":false,"n":1,"nUpTo":0,"median":12,"min":12,"max":12,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"american-addiction-centers-employee-onboarding","pooled":true}]}],"indicativeValueResult":{"low":72000,"high":448000},"evidence":["american-addiction-centers-employee-onboarding","kpmg-new-hire-onboarding-agent","usda-forest-service-new-hire-experience-assistant"]},{"title":"AI assistant for goal based financial planning","shortTitle":"Goal based planning","seoTitle":"AI assistants for goal based financial planning","metaDescription":"AI assistants turn client goals into scenarios from a planning engine the firm can audit, explain trade offs plainly and prepare plans for advisor sign off.","definition":"An AI assistant that turns a client's goals into projections and what if scenarios using a rules based planning engine, explains the trade offs in plain language and prepares the plan for an advisor to validate, with every assumption disclosed and reproducible.","aliases":["goal planning copilot","financial plan scenario assistant","retirement planning assistant for advisors","what if planning assistant"],"industries":["wealth-and-asset-management","banking"],"functions":["sales","customer-service","product-and-pricing"],"patterns":["conversational-agent","content-generation","agentic-workflow"],"channels":["internal-tools","web-chat","mobile-app"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"front-office","problem":"Clients think in goals: retire at 60, pay for a child's university, buy a second home. Turning\nthose goals into a plan means collecting a lot of information, running projections across several\nscenarios and then explaining uncertainty in a way a client understands. If a firm measures how much\nof that effort is data gathering and document assembly, a full written plan can look costly to\nprovide for smaller client relationships.\n\nLanguage models are good at the conversation and the explanation, but they are not a reliable\nsource of exact arithmetic and they cannot see everything in a complex personal situation. The design question is how to use them for what they are good at\nwhile the numbers come from an engine the firm can audit.","problemStats":[],"howItWorks":"1. **Gather goals and facts.** A conversation, with the client or the advisor, captures goals,\n   timelines, income, assets, liabilities and constraints, and flags what is missing.\n2. **Run the engine.** A planning engine the firm can audit (rule based cash flow projections, or\n   Monte Carlo scenario models with recorded assumptions and a fixed seed) calculates the plan\n   with documented assumptions for returns, inflation and taxes.\n3. **Explore what ifs.** The client or advisor asks \"what if I retire two years later\" and the\n   assistant reruns the engine and compares the scenarios.\n4. **Explain in plain language.** The model narrates the results, the trade offs and the\n   uncertainty, quoting numbers only from the engine output.\n5. **Advisor validates.** The advisor reviews the plan, adjusts it for what the model cannot see,\n   and signs it off before it is presented as advice.","valueDrivers":["employee-productivity","inclusion-and-access","customer-experience","revenue-growth"],"kpis":["time-saved-per-task","productivity-gain","users-served","customer-satisfaction","conversion-rate-uplift"],"indicativeValue":{"referenceOrg":"A wealth manager with 200 financial planners","inputs":[{"key":"planners","label":"Financial planners","low":200,"high":200,"unit":"planners","note":"The reference firm."},{"key":"plansPerYear","label":"Plans or plan reviews per planner per year","low":50,"high":100,"unit":"plans per planner per year","note":"Editorial assumption, replace with your own planning volumes."},{"key":"hoursSaved","label":"Hours saved per plan on data gathering, scenarios and write up","low":1,"high":2,"unit":"hours per plan","note":"Editorial assumption, replace with your own. No public source on this page states a time saving for AI assisted plans."},{"key":"hourlyCost","label":"Fully loaded planner cost per hour","low":80,"high":150,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"planners * plansPerYear * hoursSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Value of planner time released per year","caveat":"Leaves out the larger but less certain effect of serving clients who did not get a written plan before, the cost of the planning engine and the review time that remains."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"If the firm already runs a planning engine, the work is connecting the conversation to it, keeping every number from the engine, and designing the advisor review so plans remain advice the firm stands behind.","dataPrerequisites":["A validated planning engine with documented capital market and tax assumptions","Client fact find data from CRM and account aggregation","Approved explanations of key concepts (risk, sequence of returns, inflation)","Disclosures and plan templates per market"],"integrations":["Financial planning engine through an API","CRM and account aggregation","Document generation for the plan report","Advisor desktop or client portal"]},"implementation":{"steps":[{"title":"Keep the maths in the engine","detail":"Connect the assistant to the existing planning engine through an API and forbid the model from calculating projections itself. Every number in the narrative must come from an engine call."},{"title":"Start with advisors, not clients","detail":"Use the assistant to prepare fact finds, scenarios and plan drafts for advisors first. Client self service comes later, once explanations and refusals are proven."},{"title":"Write the explanation library","detail":"Agree plain language explanations of assumptions and uncertainty with compliance, so the model narrates within approved wording."},{"title":"Define what the assistant must hand over","detail":"Complex situations (business owners, cross border tax, estate structures, vulnerable clients) go to the advisor with a summary rather than an automated plan."},{"title":"Test scenario consistency","detail":"Build test cases where the right answer is known and check that the assistant calls the engine correctly, reports the numbers exactly and discloses the assumptions."}],"guardrails":["Numbers only from the planning engine, never generated by the language model","Every assumption disclosed and reproducible in the plan output","Projections labelled as illustrations, not promises","Handover of complex or vulnerable client situations to an advisor","Advisor sign off before a plan is presented as advice"],"humanInTheLoop":"The advisor reviews and signs off every plan, can override assumptions with a recorded reason, and owns the advice. The planning engine's assumptions are approved and reviewed periodically by an investment committee or equivalent.","kpisToInstrument":["Time from first conversation to signed off plan","Share of plans where the advisor changed numbers or assumptions, and why","Number of clients with a current written plan","Engine call errors and narrative number mismatches found in tests","Client satisfaction with plan explanations"],"failureModes":[{"title":"Model does the maths","detail":"The narrative contains a number that no engine call produced. Enforce engine only numbers and check the output automatically."},{"title":"False precision","detail":"A single projected value is presented as a promise. Show ranges and scenarios with the assumptions."},{"title":"Complex cases forced through","detail":"A business owner or cross border case gets a generic plan. Detect complexity early and hand over."},{"title":"Guidance crossing into advice","detail":"A client facing version starts recommending products. Keep product recommendations in the advised process with suitability checks."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Planning support for advisors is not listed in Annex III. A client facing version must disclose that the client is talking to AI (Article 50). It becomes high risk if it is used to assess the creditworthiness of individuals (Annex III point 5(b)) or for risk assessment and pricing of life or health insurance for individuals (Annex III point 5(c))."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","mas-ai-risk-management","iso-42001","mifid-ii"],"guidance":[{"title":"Harnessing AI in the Financial Planning Profession","issuer":"CFP Board","region":"north-america","url":"https://www.cfp.net/-/media/files/cfp-board/knowledge/reports-and-research/harnessing-ai-in-the-financial-planning-profession-cfp-board-report.pdf","note":"Scenario based report (October 2025) on how AI may change financial planning by 2030 and what planners should do now."},{"title":"ESMA public statement on the use of AI in the provision of retail investment services","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf","note":"Applies MiFID II conduct and organisational duties when AI supports investment advice, including transparency to clients about its use."},{"title":"PS22/9: A new Consumer Duty","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publication/policy/ps22-9.pdf","note":"The policy statement sets rules for four outcomes under the Duty, including a consumer understanding outcome, which applies directly to how projections and uncertainty are explained."}],"controls":["Planning engine and its assumptions inventoried, validated and version controlled","Automated check that narrative numbers match engine output","Standard disclosures on every projection","Advisor sign off recorded for each plan","Periodic review of explanation wording by compliance"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the conversation runs in an **AI agent** that gathers goals and facts, with a\n**flow** for the fixed parts of the fact find and **custom functions** that call the firm's\nplanning engine through its API, so every projection comes from the engine. A **knowledge base**\nholds the approved explanations of assumptions and concepts, and **structured output** keeps\nnumbers in separate fields that a custom function can compare with the engine response.\n\nThe advisor version runs in **Microsoft Teams** or inside internal tools through the **REST API\nchannel**; a client version can use the **web chat** channel, or the API channel inside the firm's\nown app, with **rich cards** and charts to compare scenarios. **Guardrails** stop\nproduct recommendations and hand complex cases to a human through **human handover**, and **test\nsuites** replay known scenarios on every change. **Multi language** support covers markets where\nclients plan in their own language."},"faq":[{"question":"Can a language model do financial projections?","answer":"It should not. Use it to gather information and explain results, and let a planning engine the firm can audit do the calculations, so every number is reproducible and auditable. Vanguard's Digital Advisor, for example, uses an algorithm to build and rebalance portfolios for a client's goals, and labels its projections and goal forecasts as hypothetical and not guarantees."},{"question":"Does AI replace the financial planner?","answer":"Not in the design on this page: the assistant prepares and explains, and the planner validates and signs off. Fully automated services exist, such as Vanguard's Digital Advisor; Vanguard's Personal Advisor offers ongoing financial planning and access to an advisor for those investing $50,000 or more. CIMB Niaga's agents are described as helping bank staff give tailored advice and proactive guidance."},{"question":"Is a client facing planning assistant regulated advice?","answer":"It depends on what it does. Explaining concepts and running projections is usually guidance. Recommending specific products to a client is investment advice under MiFID II and needs the full suitability process. ESMA's 2024 statement says MiFID II conduct duties still apply when firms use AI, and the firm stays responsible for the outcome."}],"related":["suitability-assessment-assistant","next-best-action-for-advisors","financial-wellbeing-coach","portfolio-drift-monitoring-and-rebalancing","wealth-advisor-knowledge-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Goal Based Planning) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: added MiFID II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the FAQ on planners (Vanguard Digital Advisor is fully automated), removed unsourced claims from the problem, specified Annex III points 5(b) and 5(c), limited channels to those in the feature inventory, added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: rewrote the meta description without the examples, pointed the Consumer Duty guidance to PS22/9 (which states the four outcomes), aligned both FAQs with the literal Vanguard and CIMB Niaga sources, and set the adoption stage to emerging because no evidence on the page shows this exact design in production."},{"date":"2026-09-27","note":"Review fixes: replaced the deterministic/Monte Carlo contradiction with a planning engine the firm can audit (definition, howItWorks, metaDescription, FAQ 1), hedged the unsourced problem and complexityNote claims, and corrected the Vanguard FAQ to match the literal Digital Advisor and Personal Advisor sources."}],"slug":"goal-based-financial-planning-assistant","url":"https://www.blits.ai/ai-use-cases/goal-based-financial-planning-assistant","benchmarks":[],"indicativeValueResult":{"low":800000,"high":6000000},"evidence":["cimb-niaga-proactive-guidance-agents","vanguard-digital-advisor"]},{"title":"AI assistant for HR and policy questions","shortTitle":"HR and policy assistant","seoTitle":"HR chatbot for employee policy questions","metaDescription":"An HR assistant answers leave, pay and policy questions from the organization's own documents. IBM reports a 94% containment rate of common questions for AskHR.","definition":"An employee self service assistant that answers questions on leave, pay and tax forms, benefits, expenses, travel and conduct policies from the organization's own HR documents, personalized to the employee's country and role, and starts simple HR transactions such as leave requests or employment letters in the HR system.","aliases":["HR chatbot","HR virtual agent","employee self service assistant","policy assistant"],"industries":["cross-industry","banking","technology","healthcare"],"functions":["human-resources","knowledge-management"],"patterns":["rag-knowledge-assistant","conversational-agent","agentic-workflow"],"channels":["microsoft-teams","internal-tools","web-chat"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"HR shared services answer the same questions every day: how many days of leave do I have left,\nwhere is my payslip, does the policy cover this expense, what happens to my benefits if I move\ncountry. The answers exist, but they are spread across policy PDFs, intranet pages and HR\nsystems, differ by country, entity and grade, and change every year. Employees raise tickets or\nemail a business partner, and HR spends skilled time on lookups.\n\nA generic chatbot makes this worse if it gives one answer to everyone. The value comes from\nanswers grounded in the current policy that applies to this person, plus the ability to complete\nthe simple transaction behind the question. The risk is the opposite failure: HR data is among\nthe most sensitive an organization holds, and some conversations (grievances, misconduct,\nwellbeing) must reach a person, not a bot.","problemStats":[],"howItWorks":"1. **Know who is asking.** The assistant runs inside the employee's signed in session and reads\n   only that person's attributes that matter for policy: country, entity, grade, contract type.\n2. **Retrieve the applicable policy.** Retrieval is filtered to the documents that apply to\n   that population, and the answer cites the policy and section it came from.\n3. **Read personal data only through the HR system.** Leave balances or payslip locations come\n   from HR system APIs scoped to the requester, never from documents about other people.\n4. **Start simple transactions.** Leave requests, employment verification letters or address\n   changes go through the HR system's normal workflow and approvals.\n5. **Route sensitive topics to people.** Grievances, disciplinary matters, harassment, health and\n   wellbeing are detected and handed to HR or employee assistance, with the employee's consent.","valueDrivers":["cost-to-serve","employee-productivity","speed"],"kpis":["containment-rate","interactions-handled","employee-adoption","cost-reduction","processing-time-reduction"],"indicativeValue":{"referenceOrg":"An organization with 20,000 employees","inputs":[{"key":"employees","label":"Employees","low":20000,"high":20000,"unit":"employees","note":"The reference organization."},{"key":"queriesPerEmployee","label":"HR questions and requests per employee per year that reach HR today","low":2,"high":4,"unit":"queries per employee per year","note":"Editorial assumption, replace with your HR case volume."},{"key":"containment","label":"Share resolved by the assistant without HR staff","low":0.4,"high":0.7,"unit":"fraction of queries","note":"Conservative against the benchmark on this page (IBM reports a 94% containment rate of common questions for AskHR after years of refinement)."},{"key":"costPerQuery","label":"Cost of an HR handled query","low":10,"high":20,"unit":"USD per query","note":"Editorial assumption for a shared services centre. Replace with your own cost."}],"formula":"employees * queriesPerEmployee * containment * costPerQuery","currency":"USD","period":"per year","resultLabel":"HR handling cost avoided","caveat":"Gross handling cost only. It leaves out the platform and integration cost, employee time saved by faster answers, and the effect of fewer errors from outdated policy copies. Savings usually show as capacity redeployed within HR, not as headcount."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"A policy question answerer is quick to build. Personalization by country and grade, access scoping to the individual, and transactions in the HR system are where the effort goes, along with keeping policy documents current.","dataPrerequisites":["Current HR policies per country and entity, each with an owner and effective date","Employee attributes that decide which policy applies (country, entity, grade, contract)","The list of topics that must always go to a person"],"integrations":["HR information system (for example Workday or SAP SuccessFactors) for balances and transactions","Payroll and benefits portals for links and documents","Identity provider for single sign on","HR case management for handover"]},"implementation":{"steps":[{"title":"Map the question volume","detail":"Pull a year of HR tickets and emails and group them. Most volume sits in a few topics: leave, pay, benefits, letters and expenses. Start there."},{"title":"Fix the source documents first","detail":"Remove duplicates and old versions, tag each policy with the population it applies to and an effective date. The assistant can only be as current as this library."},{"title":"Scope retrieval and data to the person","detail":"Filter retrieval by the employee's attributes and read personal data only through HR system APIs scoped to the requester. Test that no answer can reveal another employee's data."},{"title":"Define the human topics","detail":"Agree with HR, legal and employee representatives which topics always go to a person and how the handover works, including confidential routes."},{"title":"Add transactions one at a time","detail":"Start with letters and leave requests that already have approval workflows, and let the HR system enforce the rules rather than the assistant."},{"title":"Consult before launch","detail":"Where works councils or unions have a say in employee monitoring tools, involve them early and document what is and is not logged."}],"guardrails":["Answers cite the policy and section, and the assistant refuses when no applicable policy is found","Personal data only through APIs scoped to the signed in employee","Automatic handover for grievances, misconduct, harassment, health and wellbeing","No decisions on pay, performance, promotion or discipline; the assistant informs, HR decides","Conversation logs with restricted access and a defined retention period"],"humanInTheLoop":"HR owns the policy content, every sensitive topic and every decision about an individual. HR business partners review a weekly sample of answers per country, and policy owners approve changes before they reach the knowledge base.","kpisToInstrument":["Containment per topic, counting a follow up ticket within seven days as not contained","Answer accuracy on a monthly sample checked by HR per country","Employee adoption and satisfaction","Share of sensitive conversations correctly handed over","Time to complete letters and leave requests"],"failureModes":[{"title":"The wrong country's policy","detail":"A correct answer for Germany given to an employee in Singapore. Filter retrieval by the employee's attributes and test each country separately."},{"title":"Leaking another employee's data","detail":"Documents or logs that contain personal data are retrieved for the wrong person. Keep personal data out of the document index and scope every lookup."},{"title":"Automating what needs empathy","detail":"A distressed employee gets a policy extract. Detect sensitive topics and route to a person."},{"title":"Drift into decisions","detail":"The assistant starts screening internal applications or ranking requests. That changes its risk class and needs its own assessment."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Answering policy questions and starting routine requests is limited risk, with the Article 50 duty to disclose AI. It becomes high risk under Annex III point 4 if it is used to make or support decisions on recruitment, promotion, termination, allocating tasks based on individual behaviour or personal traits, or the monitoring and evaluation of workers; an employer deploying it then must also inform workers' representatives and the affected workers before use (Article 26(7)). Sensitive topic detection should work on what the employee writes: inferring emotions of people in the workplace from biometric data such as voice or facial expressions is prohibited under Article 5(1)(f), except for medical or safety reasons."},"regulations":["eu-ai-act","gdpr","uk-gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4 lists employment and worker management uses that would make an HR assistant high risk."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Employees must know they are interacting with AI unless this is obvious from the context."},{"title":"Employment practices and data protection: monitoring workers","issuer":"UK Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/","note":"Relevant to how conversation logs about employees are kept, used and disclosed under UK GDPR. The ICO says the guidance is under review after the Data (Use and Access) Act."}],"controls":["Data protection impact assessment covering logs, retention and access to conversations","Inventory entry with an owner in HR and a documented list of topics routed to people","Access to conversation logs restricted and audited","Consultation with employee representatives where required","Periodic accuracy review per country and policy area"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** in **Microsoft Teams** or the intranet, grounded in a\n**knowledge base** of HR policies with hybrid retrieval and document version control, so every\nanswer can cite the current version. **Custom functions** call the HR system's APIs (the\nintegration catalog includes Workday and SAP) for balances and transactions, passing only the\nrequesting employee's identifier so every lookup is scoped to that person.\n\nA **flow** detects sensitive topics and triggers **human handover** to HR, while **guardrails**\nand **PII masking** keep personal data out of prompts and logs. **Role based access control**\nlimits who can read conversation logs, and the GDPR toolkit covers retention and removal\nrequests. **Multi language** support serves employees in their own language, and **test\nsuites** check answers per country before every policy change goes live. The platform is model\nagnostic and can run in the EU or UAE region."},"faq":[{"question":"How much of HR's question volume can an assistant take?","answer":"A mature deployment takes most routine questions. IBM reports a 94% containment rate of common questions for AskHR, which recorded more than 11.5 million employee interactions in 2024, and says it helped contribute to a 40% reduction in HR operating costs over four years. That took several years of refinement; plan for lower rates at launch."},{"question":"Is an HR chatbot high risk under the EU AI Act?","answer":"Not when it answers policy questions and starts routine requests. It becomes high risk under Annex III point 4 if it is used for recruitment, promotion, termination, allocating tasks based on behaviour or personal traits, or evaluating employees, which needs a separate assessment."},{"question":"Should IT and HR share one assistant?","answer":"Often yes: employees do not care which department owns the answer. Bank of America's Erica for Employees started with IT support and added HR topics such as benefits and payroll forms, and Vituity's assistant covers both IT and HR requests. Keep separate content owners and data scopes behind the single front door."}],"related":["employee-onboarding-assistant","it-service-desk-resolution-agent","enterprise-knowledge-search","policy-drafting-and-gap-analysis","recruitment-screening-and-interview-scheduling"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog as an industry neutral page, with four evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources. Added seoTitle and metaDescription, contact deflection as a KPI and UK GDPR; added Article 26(7) and Article 5(1)(f) to the EU AI Act basis; noted that the ICO guidance is under review; removed an unsupported single sign on claim from the Blits.ai section; corrected the Turing evidence summary."},{"date":"2026-09-27","note":"Sourced the Turing company description to Turing's own about page and terms of service; narrowed the EU AI Act basis to task allocation based on behaviour or personal traits and to emotion recognition from biometric data, in line with Annex III point 4(b) and Article 5(1)(f)."}],"slug":"hr-and-policy-assistant","url":"https://www.blits.ai/ai-use-cases/hr-and-policy-assistant","benchmarks":[{"kpi":"employee-adoption","label":"Employee adoption","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":94.5,"min":90,"max":99,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ibm-askhr","pooled":true},{"id":"bank-of-america-erica-for-employees","pooled":true}]},{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":94,"min":94,"max":94,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ibm-askhr","pooled":true}]},{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40,"min":40,"max":40,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ibm-askhr","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":33,"min":33,"max":33,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"turing-hr-ticket-reply-drafting","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":11500000,"min":11500000,"max":11500000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ibm-askhr","pooled":true}]}],"indicativeValueResult":{"low":160000,"high":1120000},"evidence":["bank-of-america-erica-for-employees","ibm-askhr","turing-hr-ticket-reply-drafting","vituity-it-and-hr-assistant"]},{"title":"AI assistant for insurance brokers and agents","shortTitle":"Broker and agent assistant","seoTitle":"AI assistant for insurance agents and brokers","metaDescription":"AI assistants help insurance agents find product answers and draft follow ups. Manulife uses GenAI sales tools in nine markets; Prudential screens leads with AI.","definition":"An AI assistant for tied agents, independent brokers, advisors and the insurer's own distribution staff that answers product, underwriting and process questions from approved sources, prepares personalized customer engagement and follow ups, validates and prioritizes leads, and drafts meeting notes and emails, so producers spend more time with customers.","aliases":["agent copilot for insurance","producer assistant","insurance sales enablement AI"],"industries":["insurance"],"functions":["sales","knowledge-management"],"patterns":["rag-knowledge-assistant","recommendation-and-personalization","content-generation","summarization"],"channels":["mobile-app","internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"distribution","problem":"Insurance agents and brokers sell many products, each with product guides, underwriting rules and\nchanging campaigns. Finding the right answer (is this condition acceptable, which rider fits, what\ndocuments are needed) means calling a helpdesk or searching intranets, which slows agents down, especially new ones.\n\nAt the same time, agents spend time on admin: preparing for meetings, writing follow ups, logging\nactivity in the CRM and chasing leads that turn out to be poor. Insurers with large agency forces in\nAsia report sales or profit per active agent (Manulife and Prudential both disclose it), and at\nZurich the client and broker facing commercial insurance teams mind more than 100,000 active\nopportunities in their CRM, so small time savings per person add up.","problemStats":[],"howItWorks":"1. **Answer from approved content.** Agents ask questions in their own words and get answers\n   grounded in product guides, underwriting guidelines and procedures, with links to the source.\n2. **Prepare the customer conversation.** From CRM and policy data, the assistant suggests which\n   customers to contact and why (a policy anniversary, a gap in cover, a life event) and drafts a\n   personalized message for the agent to edit.\n3. **Qualify leads.** An AI agent contacts or screens incoming leads, validates contact details and\n   interest, and passes qualified leads to the agent with notes.\n4. **Handle the admin.** After a meeting the assistant drafts notes and follow ups and updates the\n   CRM for the agent to approve.\n5. **Coach.** Performance dashboards and call insights suggest where each agent can improve.","valueDrivers":["employee-productivity","revenue-growth","customer-experience","compliance"],"kpis":["productivity-gain","interactions-handled","employee-adoption","time-saved-per-task","hours-saved","users-served"],"indicativeValue":{"referenceOrg":"An insurer with 2,000 active tied agents","inputs":[{"key":"agents","label":"Active agents using the assistant","low":2000,"high":2000,"unit":"agents","note":"The reference insurer."},{"key":"hoursSavedPerWeek","label":"Hours saved per agent per week","low":0.5,"high":2,"unit":"hours per week","note":"Editorial assumption for product lookups, meeting preparation and follow up drafting. Replace with your own time study."},{"key":"weeksPerYear","label":"Working weeks per year","low":46,"high":46,"unit":"weeks","note":"Editorial assumption."},{"key":"valuePerHour","label":"Value of an agent hour redeployed to selling","low":25,"high":40,"unit":"USD per hour","note":"Editorial assumption, replace with your own figure for agent earnings per hour."}],"formula":"agents * hoursSavedPerWeek * weeksPerYear * valuePerHour","currency":"USD","period":"per year","resultLabel":"Value of agent time redeployed to selling","caveat":"Time value only; it assumes saved time goes into customer contact. It leaves out any sales uplift, licence and platform costs, content maintenance and the compliance review of AI drafted customer messages."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Knowledge answering on approved content is well understood. Personalized engagement and lead qualification need clean CRM and policy data, consent for contact, and compliance review of what the assistant suggests agents say.","dataPrerequisites":["Current product guides, underwriting guidelines and sales procedures with owners","CRM and policy data per agent's book, with contact consent flags","Approved message templates and compliance rules for customer communications","Lead sources and qualification criteria"],"integrations":["Agent portal or mobile app","CRM (for example Dynamics 365 or Salesforce)","Policy administration for in force data","Email, calendar and Microsoft Teams","Telephony or messaging for lead qualification"]},"implementation":{"steps":[{"title":"Start with the question agents ask most","detail":"Mine helpdesk tickets and calls from agents to find the top product and underwriting questions and build the first version around them."},{"title":"Own the content","detail":"Give every product guide and guideline an owner and a review date; stale content is the main reason agents stop trusting the assistant."},{"title":"Keep agents in control of customer messages","detail":"Drafts are suggestions the agent edits and sends; nothing goes to a customer automatically, and compliance approves the templates behind them."},{"title":"Pilot in one market and measure productivity","detail":"Compare active agents using the assistant with a similar group on activity, conversion and time to first sale for new agents."},{"title":"Add lead qualification carefully","detail":"When an AI contacts leads directly it becomes customer facing: add disclosure, consent checks and a clean handover to the agent."}],"guardrails":["Answers only from approved, current content, with sources shown","No automatic sending of customer messages; agents approve every draft","Customer data access limited to the agent's own book","Contact suggestions respect marketing consent and do not target vulnerable customers inappropriately","Clear disclosure when an AI agent contacts a lead directly"],"humanInTheLoop":"Agents decide what to recommend and send. Distribution compliance approves templates and reviews a sample of AI assisted communications, and product owners keep the content current.","kpisToInstrument":["Weekly active agents as a share of the agency force","Questions answered without escalation to the helpdesk","Share of qualified leads followed up by agents","New business per active agent, assistant users versus comparison group","Compliance findings on AI assisted communications"],"failureModes":[{"title":"Confident answers from old guides","detail":"An agent quotes a withdrawn product feature to a customer. Retire content on schedule and show the document date with every answer."},{"title":"Personalization that crosses a line","detail":"Suggestions use data the customer did not expect to be used for sales. Check purpose and consent for every data source."},{"title":"Adoption stalls after launch","detail":"Agents try it once and go back to the helpdesk. Measure weekly active use and fix the top unanswered questions every week."},{"title":"Productivity claims without a comparison","detail":"Agent productivity rises for many reasons. Use a comparison group before attributing gains to the assistant."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"An employee facing assistant for knowledge answers and drafting is not listed in Annex III and is minimal risk. A lead qualification agent that talks to customers must tell them they are dealing with AI (Article 50). Using performance insights to monitor and evaluate individual agents, or to allocate leads based on their behaviour or traits, is high risk under Annex III point 4(b), and any component that does risk assessment or pricing of life or health insurance for individuals is high risk under Annex III point 5(c)."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","iso-42001","eu-idd"],"guidance":[{"title":"Insurance Distribution Directive (IDD)","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/browse/regulation-and-policy/insurance-distribution-directive-idd_en","note":"Conduct, product information and demands and needs rules apply to what agents tell customers, including when an AI drafted it."},{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Addressed to national supervisors; explains how existing insurance rules on governance, risk management and fair treatment of customers apply to AI systems that are not high risk under the AI Act."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) covers monitoring and evaluating the performance of people in work relationships; point 5(c) covers risk assessment and pricing in life and health insurance."}],"controls":["Content ownership and review dates for every source document","Compliance approval of message templates and a monthly sample review","Access control by agent book and role","Consent checks before contact suggestions","Usage and outcome monitoring by market","Performance insights coach agents and are not used on their own for decisions on contracts, commission or lead allocation"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** in the insurer's agent portal or app (through the **web chat**\nor **API channel**) or in **Microsoft Teams**, grounded in a **knowledge base** of product guides\nand guidelines with **hybrid retrieval** and **document version control**, so answers come from\nthe current version. The ready made **Microsoft Dynamics 365** tool, or **custom functions** for\nanother CRM such as Salesforce, read the agent's book and write approved notes back to the CRM.\n\nLead qualification runs as a separate customer facing agent that screens incoming leads on\n**voice or WhatsApp**, with AI disclosure and **human handover** to the assigned agent. Drafted customer messages pass **output\nguardrails** that apply the insurer's communication rules, and **agentic workflows** triggered on a\nschedule can prepare follow up lists for agents to approve. **Test suites** check answers against\nthe latest product content, and **analytics** show usage and unanswered questions."},"faq":[{"question":"What do insurers use AI assistants for in their agency force?","answer":"Manulife has deployed GenAI sales enablement across nine markets, with engagement insights, email drafting and coaching; Prudential uses an AI talkbot to validate leads and a GenAI performance platform for agents; Sun Life gives advisors a GenAI chatbot and a notes assistant."},{"question":"Is there evidence it makes agents more productive?","answer":"Some. Waterdrop reports its Life Planner Copilot handled 300,000 product consultations for its online consultants, and Prudential reports an initial result of 98 per cent of leads validated by its talkbot in the Philippines being adopted by agents for follow up. None of the insurers cited here publish productivity figures against a comparison group."},{"question":"Who is responsible for what the agent tells the customer?","answer":"The agent and the insurer or broker, as today. AI drafts do not change distribution rules under the IDD or conduct rules such as the FCA Consumer Duty, so templates need compliance approval and agents should edit before sending."}],"related":["conversational-insurance-quote-and-buy","underwriting-risk-assessment-copilot","sales-call-coaching-and-crm-update","wealth-advisor-knowledge-assistant","client-meeting-notes-and-crm-update","insurance-renewal-and-retention"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from insurer annual reports, results filings and vendor case studies verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added Insurance Distribution Directive to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsourced figures from the problem statement and cited Zurich, Manulife and Prudential instead; changed the EU AI Act tier to context dependent (Annex III points 4(b) and 5(c)) and added the Annex III guidance; aligned the Blits.ai build with the feature inventory; tightened the Prudential wording; added an SEO title and meta description."},{"date":"2026-09-27","note":"Blits.ai build limited to feature inventory capabilities (no source citation claim; Dynamics 365 as ready made tool, Salesforce through custom functions; lead qualification screens incoming leads); softened two unsourced generalisations."}],"slug":"insurance-broker-and-agent-assistant","url":"https://www.blits.ai/ai-use-cases/insurance-broker-and-agent-assistant","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":300,"min":300,"max":300,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"zurich-copilot-for-sales-crm-updates","pooled":true}]}],"indicativeValueResult":{"low":1150000,"high":7360000},"evidence":["manulife-genai-sales-enablement","prudential-plc-ai-agency-tools","sun-life-ai-underwriting-and-client-service","waterdrop-guardian-ai-insurance-assistants","zurich-copilot-for-sales-crm-updates"]},{"title":"AI assistant for investment suitability assessment and reports","shortTitle":"Suitability assessment","seoTitle":"AI suitability assessment for investment advice","metaDescription":"An AI assistant checks client fit and drafts suitability reports while hard fails stay with rules. Covers MiFID II, ESMA's AI statement and Morgan Stanley.","definition":"An AI assistant that checks whether a proposed product or portfolio fits a client's risk tolerance, objectives, knowledge, experience and financial situation against the firm's rules, flags mismatches, and drafts the suitability rationale and report for the advisor to confirm, while hard rule failures are decided by deterministic checks, not by the model.","aliases":["suitability report drafting","suitability check copilot","appropriateness assessment assistant","statement of suitability generator"],"industries":["wealth-and-asset-management","banking"],"functions":["regulatory-compliance","sales","risk-management"],"patterns":["agentic-workflow","content-generation","classification-and-routing"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"front-office","problem":"Every investment recommendation to a retail client needs a documented suitability assessment:\ndoes the product or portfolio match the client's objectives, horizon, risk tolerance, capacity for\nloss, knowledge and experience, and sustainability preferences, and why. Under MiFID II the client\nreceives a written statement of suitability, and comparable suitability or best interest duties\napply in other major markets.\n\nIn practice the rationale is written by hand from a fact find, a risk profile and product data held\nin different systems. Quality varies by advisor, reports are long and generic, and supervisors find\ngaps only when sampling after the fact. Circumstances also change: a report that was right at the\ntime of advice says nothing about whether the portfolio still fits a year later.","problemStats":[],"howItWorks":"1. **Assemble the facts.** The assistant pulls the client's profile (objectives, horizon, risk\n   tolerance, capacity for loss, knowledge and experience, preferences) and the proposed products\n   or portfolio with their risk and cost data.\n2. **Run the rules.** A deterministic rules engine applies the firm's suitability and product\n   governance rules (risk class limits, target market, concentration, complexity) and returns pass,\n   fail or refer, with reasons.\n3. **Reason about the gaps.** For referrals, the model explains the mismatch in plain language and\n   suggests what information is missing or what the advisor should consider.\n4. **Draft the report.** A suitability rationale is drafted in the firm's template, quoting the\n   client's own stated objectives and the rule results, never inventing facts.\n5. **Advisor and supervisor confirm.** The advisor edits and confirms the report; referrals and\n   hard fails go to a supervisor. Everything is logged for audit.","valueDrivers":["compliance","employee-productivity","risk-reduction","customer-experience"],"kpis":["time-saved-per-task","error-reduction","accuracy","processing-time-reduction"],"indicativeValue":{"referenceOrg":"A wealth manager with 500 advisors giving regulated advice","inputs":[{"key":"advisors","label":"Advisors giving regulated advice","low":500,"high":500,"unit":"advisors","note":"The reference firm."},{"key":"recommendations","label":"Recommendations with a suitability report per advisor per year","low":100,"high":200,"unit":"reports per advisor per year","note":"Editorial assumption, replace with your own advice volumes."},{"key":"minutesSaved","label":"Minutes saved per suitability report","low":15,"high":30,"unit":"minutes per report","note":"Editorial assumption, replace with your own. No public source on this page states a time saving for AI drafted suitability reports."},{"key":"hourlyCost","label":"Fully loaded advisor cost per hour","low":80,"high":150,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"advisors * recommendations * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Value of advisor time released from suitability write up","caveat":"Productivity only. The main value is consistent, complete suitability records and fewer unsuitable recommendations, which this figure does not price, and it leaves out rules engine and review costs."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"This is a regulated control. It needs codified suitability and product governance rules, clean client profile data, a validated risk profiling method, compliance sign off on the report template and a full audit trail.","dataPrerequisites":["Client fact find and risk profile with dates of last update","Product data such as risk class, complexity, costs and target market","Codified suitability and product governance rules per market","Approved suitability report templates and wording"],"integrations":["Advice and portfolio management platform","Product governance and target market database","CRM for client profile and records","Document generation and client delivery"]},"implementation":{"steps":[{"title":"Codify the rules before adding a model","detail":"Write the firm's suitability and target market rules as deterministic checks with clear outcomes. The model explains and drafts; it does not decide hard fails."},{"title":"Draft from facts, not from memory","detail":"The report may only cite the client's recorded profile, the rule results and product data. Anything the model cannot trace to a source is left out or flagged."},{"title":"Pilot on one advice type","detail":"Start with a common, simple advice type (for example a model portfolio recommendation) and compare AI drafted reports with manual ones in supervisory review."},{"title":"Add ongoing suitability monitoring","detail":"Once point in time assessments work, rerun the checks when markets or client circumstances change and flag portfolios that no longer fit for advisor review."},{"title":"Keep compliance in the loop","detail":"Compliance approves templates, rule changes and model changes, and samples reports every month."}],"guardrails":["Hard suitability failures decided by deterministic rules, never overridden by the model","Reports cite only recorded client facts, rule results and product data","Advisor confirmation of every report, supervisor review of referrals and overrides","Full audit log of inputs, rule outcomes, drafts and edits","No use of protected characteristics in suitability reasoning"],"humanInTheLoop":"The advisor confirms each assessment and report and remains responsible for the recommendation; supervisors decide referrals and overrides; compliance owns the rules and templates and samples output.","kpisToInstrument":["Time to a confirmed suitability report","Supervisory findings per hundred reports, before and after","Share of assessments referred or failed, with reasons","Advisor edit rate on drafted rationales","Portfolios flagged by ongoing monitoring and resolved"],"failureModes":[{"title":"Model overrules the rule","detail":"A fluent rationale justifies a product that failed a hard check. Keep the verdict in the rules engine and block contradictory drafts."},{"title":"Boilerplate reports","detail":"Every report reads the same and does not reflect the client's own words. Require quotes from the fact find and supervise for generic text."},{"title":"Stale profiles","detail":"The assessment uses a risk profile that is years old. Check profile dates and require an update before advice."},{"title":"Hidden bias","detail":"Reasoning uses proxies such as age or nationality inappropriately. Test outputs across segments and restrict inputs."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Investment suitability assessment is not listed in Annex III, so the tier depends on design. It becomes high risk where the same system assesses creditworthiness, for example for lending against a portfolio (Annex III point 5(b)). MiFID II suitability duties apply regardless of the AI Act tier."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","mas-ai-risk-management","us-sr-11-7","iso-42001","mifid-ii"],"guidance":[{"title":"Guidelines on certain aspects of the MiFID II suitability requirements","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/document/guidelines-certain-aspects-mifid-ii-suitability-requirements-1","note":"The current guidelines (ESMA35-43-3172, 2022), which replaced the 2018 version (ESMA35-43-1163). They set out how firms collect client information, including sustainability preferences, assess suitability and document it, with specific points on automated advice."},{"title":"ESMA public statement on the use of AI in the provision of retail investment services","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf","note":"Says firms that use AI in investment advice and portfolio management must apply heightened vigilance and diligence, particularly in ensuring the suitability of services and financial instruments for each client."},{"title":"Artificial Intelligence Model Risk Management (information paper)","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Good practices observed in a 2024 MAS thematic review of banks' AI and generative AI model risk management, covering governance and oversight, risk management systems and processes, and development and deployment."}],"controls":["Rules engine and model inventoried with owners in compliance and the business","Version control and approval for rules, templates and prompts","Audit trail per assessment kept for the regulatory retention period","Monthly supervisory sampling with documented findings","Fairness testing of drafted rationales across client segments"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the deterministic checks run as **custom functions** (custom code in an isolated\nsandbox, or REST calls to the firm's existing rules engine), orchestrated by an **agentic\nworkflow** whose **tool execution policy** limits it to reading data and running checks. An **AI\nagent** with **structured output** drafts the rationale in the approved template from the client\nprofile and rule results, with the firm's suitability policy held in a **knowledge base**.\n\nEvery draft pauses for **human in the loop** approval, and referrals route to a supervisor. The\nworkflow's **audit trail** keeps inputs, rule outcomes, drafts and decisions per run. **Guardrails**\nblock drafts that contradict a failed rule, **PII masking** protects client data, and **test\nsuites** with deterministic and LLM based grading replay reference cases after every rule or prompt change.\n**Agentic tasks** can recheck suitability when a condition such as a large drawdown is met."},"faq":[{"question":"Can AI decide whether an investment is suitable?","answer":"It should not make the verdict on its own. Keep hard rules in a deterministic engine, let the model explain and draft, and have the advisor confirm. ESMA expects heightened diligence on suitability when AI is used in advice."},{"question":"Is anyone doing this at scale yet?","answer":"Automated digital advice services such as Vanguard Digital Advisor already use an algorithm to build portfolios from a client's goals, time horizon and risk tolerance, and since October 2025 Morgan Stanley's advisors get AI drafted talking points that flag misalignment with a client's objectives. Public, named deployments of generative AI that write full suitability reports are still rare, so treat vendor claims with care."},{"question":"What about checking suitability after the sale?","answer":"That is where agentic monitoring helps: rerunning the checks when markets or client circumstances change and flagging portfolios for review. It links closely to drift monitoring."}],"related":["goal-based-financial-planning-assistant","next-best-action-for-advisors","client-meeting-notes-and-crm-update","portfolio-drift-monitoring-and-rebalancing","wealth-advisor-knowledge-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Suitability Assessment) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: added MiFID II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: ESMA suitability guidance now points to the current 2022 guidelines; MAS and ESMA notes aligned with the documents; unsourced market list and the Vanguard claim rewritten; added seoTitle and metaDescription."},{"date":"2026-09-26","note":"Second fact check against sources: ESMA 2022 guidelines, ESMA AI statement, MAS paper, BlackRock and InvestmentNews articles and the Vanguard page rechecked; FAQ now describes Vanguard Digital Advisor as algorithm based, as Vanguard does, and dates Morgan Stanley's access to October 2025."}],"slug":"suitability-assessment-assistant","url":"https://www.blits.ai/ai-use-cases/suitability-assessment-assistant","benchmarks":[],"indicativeValueResult":{"low":1000000,"high":7500000},"evidence":["morgan-stanley-aladdin-auto-commentary","vanguard-digital-advisor"]},{"title":"AI assistant for procurement and supplier contract review","shortTitle":"Procurement and contract review","seoTitle":"AI for procurement and supplier contract review","metaDescription":"AI extracts key terms from supplier contracts and flags deviations for a buyer or lawyer to approve. The IRS uses it to draft and check contract file documents.","definition":"An assistant for procurement and vendor management that reads supplier contracts and proposals, extracts the key terms, flags deviations from the organization's standard positions, drafts requests for proposal and evaluation matrices, and prepares negotiation positions, with a procurement or legal owner approving every conclusion.","aliases":["contract review AI","supplier contract analysis","AI contract abstraction","procurement copilot","AI supplier negotiation"],"industries":["cross-industry","banking","government","retail-and-ecommerce","manufacturing"],"functions":["procurement","legal","finance-and-accounting"],"patterns":["document-processing","rag-knowledge-assistant","content-generation","agentic-workflow"],"channels":["internal-tools","microsoft-teams","email"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"A large organization holds thousands of supplier contracts, each with its own prices, renewal\ndates, service levels, liability caps, audit rights, data clauses and exit terms. Buyers and\nlawyers read them one by one: to check a new contract against the playbook, to find every\ncontract affected by a new regulation, or to prepare a renewal. Review queues grow, auto renewals\nslip through, and the long tail of smaller suppliers is barely negotiated at all. An HBR article\nco written by Walmart International's sourcing leaders described this: around 20% of Walmart's\nsuppliers had signed agreements with standard terms that were often not negotiated.\n\nIn financial services the stakes are regulatory. Third party rules such as DORA in the EU and APRA\nCPS 230 in Australia require specific clauses in contracts with ICT and material service providers.\nUnder DORA every ICT services contract must cover a description of the services, data locations\nand termination rights, and contracts for services that support critical or important functions\nmust also include audit and access rights and exit strategies. Both rules also require a register\nof those arrangements: under DORA the supervisor can request it, and under CPS 230 it is submitted\nto APRA.","problemStats":[],"howItWorks":"1. **Ingest and classify.** Contracts, amendments and supplier proposals are loaded from the\n   contract repository or email, split into clauses and classified by clause type.\n2. **Extract key terms.** The AI extracts parties, prices, dates, renewal and notice periods,\n   service levels, liability, audit rights, data protection and exit terms into a structured\n   record, citing the clause for each value.\n3. **Compare with the playbook.** Each clause is compared with the organization's standard and\n   fallback positions and with regulatory must haves; deviations are flagged with a risk rating\n   and a suggested redline.\n4. **Draft sourcing documents.** From the requirement and past tenders the assistant drafts the\n   request for proposal, the evaluation criteria and a first comparison of supplier responses.\n5. **Prepare the negotiation.** It summarises the contract history, benchmarks and open\n   deviations into a negotiation brief; for low value tail spend, some organizations let a bot\n   negotiate within limits the buyer sets, as Walmart has done with tail end suppliers.\n6. **Hand to the owner.** A buyer or lawyer reviews the extraction and the flags, decides, and the\n   decision and evidence are stored with the contract record.","valueDrivers":["cost-to-serve","employee-productivity","compliance","risk-reduction","speed"],"kpis":["processing-time-reduction","time-saved-per-task","cost-reduction","automation-rate","accuracy"],"indicativeValue":{"referenceOrg":"A bank with 2,000 supplier contracts and 400 new contracts or renewals reviewed a year","inputs":[{"key":"reviews","label":"Contract reviews per year (new contracts, renewals and amendments)","low":400,"high":400,"unit":"reviews per year","note":"The reference organization. Replace with your own volume."},{"key":"hoursPerReview","label":"Procurement and legal hours per review today","low":4,"high":10,"unit":"hours per review","note":"Editorial assumption covering reading, playbook comparison and write up."},{"key":"timeSaved","label":"Share of review time saved by AI extraction and deviation flags","low":0.2,"high":0.4,"unit":"fraction of review time","note":"Editorial assumption; reviewers still read flagged clauses and confirm extracted terms."},{"key":"hourlyCost","label":"Blended procurement and legal hour","low":80,"high":150,"unit":"USD per hour","note":"Editorial assumption. Replace with your own rate."},{"key":"tailSpend","label":"Annual tail spend with suppliers that are rarely negotiated","low":5000000,"high":20000000,"unit":"USD per year","note":"Editorial assumption for a mid sized bank."},{"key":"tailSaving","label":"Saving on negotiated tail spend","low":0.01,"high":0.02,"unit":"fraction of spend","note":"Conservative against the evidence on this page (Pactum reports a 3% average gain across Walmart's negotiations), because not every supplier agrees."}],"formula":"reviews * hoursPerReview * timeSaved * hourlyCost + tailSpend * tailSaving","currency":"USD","period":"per year","resultLabel":"Review time released plus savings on tail spend","caveat":"It leaves out the value of avoided auto renewals, fewer missing regulatory clauses in material outsourcing contracts and faster sourcing cycles, and it leaves out the cost of loading and clause tagging the existing contract estate."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Extraction from clean digital contracts works well; scanned legacy contracts, amendments that override earlier terms and a missing clause playbook are where projects stall. Autonomous negotiation is a separate step that needs clear commercial limits.","dataPrerequisites":["A contract repository with contracts, amendments and supplier metadata in one place","A clause playbook with standard, fallback and unacceptable positions","The list of regulatory must have clauses for outsourcing and material service providers","Past tenders, evaluation matrices and awarded contracts as examples"],"integrations":["Contract lifecycle management or document management system","Procurement suite (sourcing, purchase to pay, supplier master)","Third party risk management register","Email and Microsoft Teams for buyer and legal workflows","Electronic signature platform"]},"implementation":{"steps":[{"title":"Write the playbook before the prompts","detail":"Agree standard and fallback positions for the clauses that matter, and the regulatory must haves for outsourcing contracts. The AI can only flag deviations from a position someone has written down."},{"title":"Start with extraction on new contracts","detail":"Extract key terms from incoming contracts with a citation per value and have buyers confirm them. Measure field level accuracy on a sample before trusting it on the back book."},{"title":"Run a back book sweep","detail":"Use the extraction on the existing estate to find renewal dates, missing audit or exit clauses and data location terms, and route gaps to contract owners as tasks."},{"title":"Add deviation flags and redlines","detail":"Compare clauses with the playbook, rate the deviation and suggest a redline. Lawyers approve or change every redline before it goes to a supplier."},{"title":"Draft sourcing documents","detail":"Generate first drafts of requests for proposal and evaluation matrices from templates and past tenders, and a first comparison of responses for the buying team to score."},{"title":"Consider negotiation bots only for the tail","detail":"If you automate negotiation, restrict it to low value suppliers and terms such as payment days and discounts, with limits set by a buyer, as Walmart has done with its tail end suppliers."}],"guardrails":["Every extracted term and deviation flag cites the clause it came from","A named buyer or lawyer approves every conclusion, redline and award recommendation","The AI does not sign, accept terms or commit spend; automated negotiation stays within limits a buyer set in advance","Supplier documents are treated as untrusted input, so instructions hidden in a contract or proposal cannot change the assistant's behaviour","Supplier evaluation scores are decided by the evaluation panel, not by the AI","Contract data stays in region and is not used to train external models"],"humanInTheLoop":"Procurement owns commercial decisions, legal owns clause positions and redlines, and the third party risk owner signs off material outsourcing contracts. The AI prepares, extracts and flags; people decide and their decisions are recorded with the contract.","kpisToInstrument":["Review cycle time from receipt to decision, before and after","Field level extraction accuracy on a monthly sample","Deviations flagged per contract and share accepted by legal","Missing regulatory clauses found in the back book and time to remediate","Savings on renegotiated contracts, net of supplier attrition"],"failureModes":[{"title":"Amendments ignored","detail":"The AI reads the master agreement and misses the amendment that changed the price or the term. Link amendments to their master and extract the effective terms."},{"title":"Confident extraction from bad scans","detail":"Poor OCR on old contracts yields wrong dates or amounts that look authoritative. Flag low confidence fields and route them to a person."},{"title":"Playbook drift","detail":"Positions change but the playbook does not, so the AI flags the wrong deviations. Give the playbook an owner and version it."},{"title":"Automated procurement decisions without accountability","detail":"Tools that score suppliers or proposals can quietly become the decision when evaluators copy their output instead of forming their own view. Keep scoring advisory and record the panel's own reasoning."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Contract review and sourcing are not among the Annex III high risk uses, so an internal assistant that makes no decisions about natural persons is minimal risk (with the Article 4 AI literacy duty). If a negotiation bot chats directly with supplier staff, Article 50(1) applies and it must tell them they are dealing with an AI system, unless that is obvious from the context. Public authorities using AI in procurement should still check national public procurement rules on transparency and equal treatment of bidders."},"regulations":["dora","apra-cps-230","gdpr","eu-ai-act","iso-42001"],"guidance":[{"title":"Digital Operational Resilience Act (Regulation (EU) 2022/2554), Article 30","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2022/2554/oj","note":"Sets the key contractual provisions for ICT third party services at financial entities, which a contract review assistant should check as must have clauses."},{"title":"Operational risk management (CPS 230)","issuer":"Australian Prudential Regulation Authority","region":"asia-pacific","url":"https://www.apra.gov.au/operational-risk-management","note":"Requires formal agreements with material service providers covering specified terms, and a register of those providers submitted to APRA. Amendments that commenced on 1 July 2026 exempt some categories of service provider from specific contractual requirements."}],"controls":["Clause playbook and regulatory clause list owned by legal, versioned and reviewed","Human approval recorded for every redline, deviation acceptance and award recommendation","Sampled accuracy checks on extracted terms, with results kept as evidence","Access control on contract data by business unit and sensitivity","Inventory entry for the assistant with an accountable owner"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** holding the clause playbook,\nregulatory clause lists, templates and past tenders, retrieved with **hybrid search**. Contracts\nand proposals arrive as uploads or through **incoming email as a knowledge source**, and document\ningestion reads PDF, DOCX, image and Outlook email files. An **agentic workflow** extracts the key terms with\n**structured output**, compares them with the playbook, and writes the record back to the contract\nsystem through **custom functions** (REST calls), connectors from the **integration catalog**\n(for example SAP, Oracle, NetSuite, DocuSign) or the ready made SharePoint tool.\n\nBuyers and lawyers work with the agent in **Microsoft Teams** or email, and write backs and\noutgoing redlines are set to require **human in the loop approval**. Input and output\n**guardrails** block prompt injection attempts, **PII masking** protects personal data in contracts, **run history with a\nfull audit trail** keeps each extraction and approval, and **test suites** check extraction\naccuracy on a reference set of contracts after every change. The platform is model agnostic and\noffers EU and UAE data residency."},"faq":[{"question":"Can AI review supplier contracts reliably?","answer":"It is reliable as a first pass that extracts terms and flags deviations with a citation to the clause, and unreliable as the final word. Measure field level accuracy on your own contracts, route low confidence fields to a person, and keep a lawyer's approval on every redline."},{"question":"Can AI negotiate with suppliers?","answer":"For simple terms with low value suppliers, yes. An HBR article co written by Walmart International's sourcing leaders reported in 2022 that its negotiation chatbot had so far closed agreements with 68% of suppliers approached, and the vendor, Pactum, reports a 3% average gain across Walmart's negotiations. Buyers set the limits and strategic suppliers stay with people."},{"question":"How does this help with third party risk rules for banks?","answer":"Rules such as DORA and APRA CPS 230 require specific contractual provisions with ICT and material service providers. A back book sweep finds contracts missing audit, data location or exit clauses so they can be remediated, and the evidence is kept with each contract."},{"question":"How do government buyers use AI in public procurement?","answer":"Mostly as a drafting and review aid inside the existing procedure. The US Administration for Children and Families uses it to find passages in proposals and draft technical evaluation language, the IRS to draft and check contract file documents, and GSA to screen solicitations for missing compliance language. In the ACF and IRS entries the AI drafts and flags while officials make the final determinations. GSA also retired CALI, a machine learning tool for checking proposal compliance that was still in training, in November 2024 without publishing a reason."}],"related":["vendor-due-diligence","supplier-invoice-processing","policy-drafting-and-gap-analysis","internal-audit-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog, made industry neutral and verified against the sources."},{"date":"2026-09-25","note":"Added public procurement evidence from the US federal AI use case inventory (HHS ACF, IRS, GSA) and a public sector FAQ."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription, limited the regulatory paragraph to DORA and CPS 230 as cited (dropped the uncited MAS reference), updated the CPS 230 note for the July 2026 amendments, set the EU AI Act tier to context dependent because a supplier facing negotiation bot triggers Article 50, aligned document ingestion with the platform inventory, and set the Walmart record to production with a summary limited to its public text."},{"date":"2026-09-27","note":"Review fixes: removed the Walmart 68% and 3% figures as KPI metrics (an agreement rate and a gain on negotiated terms, not an automation rate or a process cost reduction) and kept them in prose, corrected the GSA CALI record to a compliance checking tool still in training when retired, removed CALI from the accountability failure mode, limited DORA audit and exit clauses to critical or important functions, and tightened the metaDescription, FAQ attributions and platform wording."},{"date":"2026-09-27","note":"Review fixes: removed the uncited '20 to 30 clauses' figure from the implementation steps; no source gives a count of clauses that matter."},{"date":"2026-09-27","note":"Fact checked against sources: limited the DORA paragraph to what Article 30(2) requires of every contract (subcontracting conditions apply to critical or important functions), stated how each register reaches the supervisor, added the Article 50(1) obviousness exception, named SharePoint as a ready made tool rather than an integration catalog connector, named the IRS in the metaDescription, and aligned the IRS and ACF summaries with their inventory wording."}],"slug":"procurement-contract-review","url":"https://www.blits.ai/ai-use-cases/procurement-contract-review","benchmarks":[],"indicativeValueResult":{"low":75600,"high":640000},"evidence":["gsa-cali-proposal-evaluation","gsa-solicitation-review-tool","hhs-acf-proposal-review-drafting","irs-ai-contract-document-toolbox","walmart-autonomous-supplier-negotiation"]},{"title":"AI assistant for Shariah compliance screening and review","shortTitle":"Shariah compliance screening","seoTitle":"AI for Shariah compliance screening and review","metaDescription":"AI flags riba, gharar and prohibited exposure in contracts and investments, citing the standard and fatwa, while the Shariah board keeps sole authority over rulings.","definition":"An AI assistant that screens Islamic financing contracts, deal structures and investments for Shariah compliance risks such as riba, gharar and exposure to prohibited activities, retrieves the relevant standards and fatwas, drafts the Shariah review documentation and flags issues for the Shariah board, which keeps sole authority over any ruling.","aliases":["Shariah screening AI","Islamic finance compliance assistant","riba and gharar detection","Shariah review copilot"],"industries":["banking","wealth-and-asset-management","insurance"],"functions":["regulatory-compliance","legal","product-and-pricing"],"patterns":["rag-knowledge-assistant","document-processing","classification-and-routing","content-generation"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"specialized-businesses","problem":"Every Islamic financing product, contract and investment has to comply with Shariah principles:\nno interest (riba), no excessive uncertainty (gharar), no exposure to prohibited sectors, and the\nstructure must follow the approved contract type. Shariah compliance teams review contracts clause\nby clause, check structures against the institution's approved standards and the fatwas of its\nShariah board, screen investments against financial ratios, and document every review for the\nboard and for Shariah audit.\n\nMuch of this work is manual and depends on specialists who understand both finance and fiqh. The\nIslamic Financial Services Board reports that the industry keeps growing and that new products and\nstructures increasingly mimic conventional banking characteristics. AI can find clauses, compare\nthem with standards and draft documentation, but an error in the tool can lead to a Shariah non\ncompliance finding, so rulings must stay with qualified scholars.","problemStats":[{"statement":"The Islamic Financial Services Board reports that the global Islamic financial services industry reached approximately USD 4.4 trillion in total assets in 2025.","sourceTitle":"IFSB Releases Islamic Financial Stability Report 2026 Highlighting Emerging Hybrid Risks in Islamic Banking","sourceUrl":"https://www.ifsb.org/press-releases/ifsb-releases-islamic-financial-stability-report-2026-highlighting-emerging-hybrid-risks-in-islamic-banking/","year":2026}],"howItWorks":"1. **Load the reference base.** The institution's approved standards (for example AAOIFI based\n   policies), its Shariah board's fatwas and resolutions, and product templates are indexed.\n2. **Read the contract or structure.** Document AI splits the contract into clauses and\n   identifies the contract type, pricing, penalties, ownership transfer and asset terms.\n3. **Screen.** Each clause is compared with the approved template and standards, and the model\n   flags possible riba, gharar, prohibited activities or deviations, citing the standard.\n4. **Screen investments.** For equities and funds, business activity and financial ratio screens\n   run on current data, with changes in status monitored.\n5. **Draft the review.** It drafts the Shariah review memo with findings and references; the\n   Shariah compliance officer completes it and the Shariah board decides.","valueDrivers":["compliance","employee-productivity","speed"],"kpis":["time-saved-per-task","processing-time-reduction","accuracy","productivity-gain","users-served"],"indicativeValue":{"referenceOrg":"An Islamic bank reviewing 1,500 financing contracts and structures a year","inputs":[{"key":"reviews","label":"Shariah reviews per year","low":1500,"high":1500,"unit":"reviews per year","note":"The reference bank."},{"key":"hoursPerReview","label":"Compliance hours per review","low":3,"high":6,"unit":"hours per review","note":"Editorial assumption, replace with your own time study."},{"key":"timeSaved","label":"Share of review time saved","low":0.2,"high":0.35,"unit":"fraction of time","note":"Editorial assumption; no verified public benchmark was found."},{"key":"hourlyCost","label":"Loaded cost of a Shariah compliance hour","low":70,"high":120,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."}],"formula":"reviews * hoursPerReview * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Shariah compliance time released, valued at loaded cost","caveat":"Values compliance time only. It leaves out faster product approval, fewer Shariah non compliance events and income purification, and the cost of building and validating the reference base."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The retrieval and drafting are standard; the hard parts are a curated, versioned reference base of the institution's own standards and fatwas, Arabic and local language sources, and scholar trust in the outputs.","dataPrerequisites":["The institution's approved Shariah standards, policies and product templates","Shariah board fatwas and resolutions, versioned","Past reviews with findings, for testing","Financial data for investment screening"],"integrations":["Document management and contract repository","Product approval workflow","Market and financial data for investment screening"]},"implementation":{"steps":[{"title":"Curate the reference base with the Shariah board","detail":"Agree which standards, fatwas and resolutions the assistant may use, who maintains them and how superseded rulings are marked, before any screening."},{"title":"Start with templates and deviations","detail":"Begin by comparing contracts with approved templates and highlighting deviations, which is verifiable, before asking the model to judge substance."},{"title":"Test on past reviews","detail":"Run the assistant on past contracts with known findings, including hard cases such as hybrid and cross border structures, and share the results with the board."},{"title":"Draft documentation, not rulings","detail":"Use the assistant to draft review memos and audit working papers with references, leaving conclusions to the compliance officer and rulings to the board."}],"guardrails":["The assistant never issues or implies a Shariah ruling; it flags and cites","Only board approved standards and fatwas are in the reference base, with version control","Every finding cites the clause and the standard or fatwa it relies on","Uncertain and complex structures are routed to scholars without a suggested conclusion"],"humanInTheLoop":"Shariah compliance officers review every finding and complete every memo. The Shariah board retains sole authority over rulings and approves the reference base. Shariah audit samples the assistant's work, and the board is told how the tool works and where it is weak.","kpisToInstrument":["Review time per contract type, before and after","Findings confirmed or rejected by compliance officers","Issues found later in Shariah audit that the assistant missed","Share of findings with a correct citation"],"failureModes":[{"title":"Confident errors on complex structures","detail":"A model may catch explicit interest clauses yet misjudge hybrid or novel structures. Route these to scholars and measure accuracy by structure type."},{"title":"Outdated or foreign rulings","detail":"The assistant cites a superseded fatwa or another institution's standard. Keep only approved, versioned sources in the reference base."},{"title":"Loss of scholar trust","detail":"Opaque outputs make the board reject the tool. Show sources for every finding and involve scholars in testing."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"An internal assistant that screens contracts for compliance with Shariah standards is not listed in Annex III: it assesses contracts, structures and securities, not the creditworthiness of natural persons (Annex III point 5(b)). If a customer facing version answers product questions, it must disclose that people are interacting with an AI system under Article 50(1). National Islamic finance regulators set their own Shariah governance expectations."},"regulations":["eu-ai-act","gdpr","sdaia-ai-ethics","cbuae-ai-guidance","iso-42001","bnm-shariah-governance","aaoifi-shariah-standards"],"guidance":[{"title":"NIST AI Risk Management Framework","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/itl/ai-risk-management-framework","note":"A general structure to document how the assistant works, test it and manage its limits, which supports the transparency a Shariah board needs."}],"controls":["Board approved reference base with version control and an owner","Model documentation shared with the Shariah board, including known limitations","Audit trail of findings, sources and human decisions per review","Periodic Shariah audit sampling of assistant supported reviews"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** for the Shariah compliance team with a **knowledge base** of\nthe institution's approved standards, fatwas and templates, stored in the document library with\n**version control** and retrieved with hybrid search. Contracts are uploaded in PDF or Word and\ncompared with templates; **structured output** returns the findings with the clause and the\nsource for each. **Arabic support**, including Arabic normalisation and regional Arabic\nmodels such as Humain and Fanar, helps with Arabic sources.\n\n**Guardrails** stop the agent from phrasing findings as rulings, **test suites** replay past\nreviews with known findings on every change, and **execution tracing** shows what was retrieved\nfor each answer. The platform is model agnostic and runs in the EU or UAE region for data\nresidency."},"faq":[{"question":"Can AI decide whether a product is Shariah compliant?","answer":"No. It can find clauses, compare them with approved standards and draft documentation, but rulings belong to the Shariah board. The assistant should flag and cite, never conclude."},{"question":"Is automated Shariah screening already in use?","answer":"For listed investments, yes, although not as generative AI: Zoya says it publishes Shariah compliance reports for over 60,000 stocks, applies the AAOIFI screening methodology under the guidance of its Shariah advisors and is trusted by more than 400,000 investors. For bank contracts we did not find a verified public deployment with results."},{"question":"Where does AI struggle in Shariah review?","answer":"With hybrid instruments, cross border structures and new products, where judgement depends on context. These cases should go to scholars without a suggested conclusion."}],"related":["credit-memo-drafting-agent","policy-drafting-and-gap-analysis","regulatory-horizon-scanning","trade-document-examination"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by the editor after a targeted blocker check."},{"date":"2026-09-25","note":"First version, kept as draft. The catalog's Bank Islam Malaysia figure could not be verified (the cited Islamic Finance News page returned 403 and other reports give a different number), so it is not used."},{"date":"2026-09-25","note":"Consolidation pass: added BNM Shariah governance policy, AAOIFI Shariah standards to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added the IFSB 2026 industry size statistic to support the problem statement, removed the MAS AI guidelines (little Islamic banking in Singapore), cited Article 50(1) and Annex III point 5(b), made the Zoya FAQ answer match the source wording, softened an unsourced failure mode claim and added the SEO title and meta description."},{"date":"2026-09-26","note":"Second fact check against sources: all quotes, the IFSB statistic, regulation and guidance links confirmed; removed an unsourced claim that review time holds up product launches, aligned the IFSB wording with the press release and tightened the meta description."}],"slug":"shariah-compliance-screening","url":"https://www.blits.ai/ai-use-cases/shariah-compliance-screening","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":400000,"min":400000,"max":400000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"zoya-shariah-stock-screening","pooled":true}]}],"indicativeValueResult":{"low":63000,"high":378000},"evidence":["zoya-shariah-stock-screening"]},{"title":"AI assistant for student enrollment and student services","shortTitle":"Student enrollment assistant","seoTitle":"AI chatbot for student enrollment and services","metaDescription":"AI assistants answer students by text and chat and nudge them to enroll. Georgia State cut summer melt by 22% with a new student portal and its Pounce chatbot.","definition":"An AI assistant that answers admitted and current students' questions about admissions, financial aid, registration, housing and deadlines by text message and web chat, sends timely reminders for the tasks each student still has to complete, and hands personal or complex cases to staff.","aliases":["university admissions chatbot","summer melt chatbot","student services chatbot","enrollment nudging assistant"],"industries":["education"],"functions":["customer-service","operations"],"patterns":["conversational-agent","rag-knowledge-assistant","classification-and-routing"],"channels":["sms","web-chat","whatsapp","mobile-app"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Between acceptance and the first day of class, students face a string of administrative hurdles:\nfinancial aid forms, verification documents, immunization records, placement tests, housing and\nregistration. Students without someone to guide them can stall at any one of these steps, and\nsome simply never show up, a pattern known as summer melt. Georgia State University describes it\nas a problem for at risk students, especially those from urban school districts.\n\nAdmissions and student service offices cannot answer thousands of repetitive questions at the\nmoment students ask them, often in the evening and at weekends, and mass emails go unread. The\nsame pattern continues after enrollment: students do not know which office to contact, deadlines\npass, and staff spend their time on questions a published policy already answers instead of on\nthe students who need a person.","problemStats":[{"statement":"Georgia State University, listing Lindsay Page and Ben Castleman's book Summer Melt (2014) among its sources, reports that as many as 20 percent of students from urban school districts who are admitted to college and confirm their intent to enroll never attend any post secondary institution.","sourceTitle":"Reduction of Summer Melt (Internet Archive snapshot)","sourceUrl":"https://web.archive.org/web/20260514132005/https://success.gsu.edu/reduction-of-summer-melt/","year":2014}],"howItWorks":"1. **Know each student's open tasks.** The assistant reads from the student information system\n   which steps each admitted or current student has still to complete: aid documents, deposits,\n   immunizations, orientation, registration.\n2. **Nudge at the right time.** It sends short, personal reminders by text message before each\n   deadline, and short surveys (for example intent to enroll) whose answers flow back to staff.\n3. **Answer questions around the clock.** Students reply or ask in their own words; answers come\n   from the institution's approved policies, deadlines and office information.\n4. **Route what needs a person.** Questions about a student's own aid package, a crisis,\n   wellbeing or anything the knowledge base does not cover go to the right office with the\n   conversation attached.\n5. **Learn from the questions.** Staff review the most common questions and failed answers to fix\n   confusing web pages and processes, not only the bot.","valueDrivers":["inclusion-and-access","customer-experience","employee-productivity","cost-to-serve"],"kpis":["interactions-handled","hours-saved","users-served","contact-deflection","customer-satisfaction","response-time-reduction"],"indicativeValue":{"referenceOrg":"A public university with 30,000 students and 6,000 new students a year","inputs":[{"key":"questions","label":"Student questions and replies per year across admissions and student services","low":100000,"high":200000,"unit":"messages per year","note":"Editorial assumption. For scale, Georgia State reports more than 200,000 answers to incoming students in one summer, and Mainstay reports 81,167 messages handled in a year by Adelphi University's assistant."},{"key":"handledShare","label":"Share of messages the assistant handles without staff","low":0.5,"high":0.8,"unit":"fraction of messages","note":"Editorial assumption, replace with your own data after a first term."},{"key":"minutesPerMessage","label":"Staff minutes per message","low":1,"high":2,"unit":"minutes per message","note":"Mainstay's case study for Adelphi University assumes approximately one minute per message; the high value allows for emails and calls that take longer. Editorial assumption, replace with your own."},{"key":"staffCostPerHour","label":"Fully loaded cost of a staff hour","low":30,"high":45,"unit":"USD per hour","note":"Editorial assumption, replace with your own cost."}],"formula":"questions * handledShare * minutesPerMessage / 60 * staffCostPerHour","currency":"USD","period":"per year","resultLabel":"Staff time released for student support","caveat":"Counts staff time only. It leaves out the larger effect that institutions such as Georgia State report, more admitted students actually enrolling, and the cost of the platform, integration and content upkeep."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering general questions is straightforward. The value comes from personal nudges, which need a clean feed of each student's open tasks from the student information system, consent for text messaging and a working handover to several offices.","dataPrerequisites":["Current, owned content per office (admissions, aid, registrar, housing, bursar)","Each student's open enrollment tasks and deadlines from the student information system","Mobile numbers with consent to receive text messages"],"integrations":["Student information system and admissions CRM","Messaging (SMS, WhatsApp) and web chat on the institution's site","Ticketing or case routing to each student service office","Single sign on for questions about a student's own record"]},"implementation":{"steps":[{"title":"Map the drop off points","detail":"List the steps between acceptance and the first day (or between terms) where students stall, and the questions they ask at each. Use last year's data on who did not show up."},{"title":"Clean up the content first","detail":"Give each office ownership of its answers with a review date. The assistant is only as good as the policies and deadlines behind it."},{"title":"Design nudges with the offices","detail":"Agree the reminder calendar and wording with admissions, aid and the registrar, keep messages short and personal, and respect quiet hours and opt outs."},{"title":"Define the handover","detail":"Decide which topics go to which office and within what time, and pass the conversation along so students do not repeat themselves."},{"title":"Measure against a comparison group","detail":"Where possible, compare enrollment and task completion with students who did not receive the assistant, as Georgia State did in a randomized trial, rather than counting messages alone."}],"guardrails":["Answers only from approved institutional content, with a refusal and a route to staff otherwise","No decisions on admission, aid or placement; the assistant informs and reminds","Opt in and opt out for text messaging, with quiet hours","Crisis and wellbeing keywords routed to trained staff immediately","Personal data minimized in messages and masked in logs"],"humanInTheLoop":"Staff in each office own their content and handle every case that concerns a student's own record, money or wellbeing. A cross office group decides campaign strategy, timing and wording, as the task force Adelphi University set up does. Someone should also review unanswered questions and handovers every week.","kpisToInstrument":["Enrollment or task completion rate versus a comparison group","Messages handled without staff and handover rate by topic","Response rate to reminders and surveys","Student satisfaction with answers","Opt out rate from text messaging"],"failureModes":[{"title":"Nudges that become noise","detail":"Too many or generic messages lead to opt outs, and billing reminders crowd out help. Coordinate campaigns centrally and keep them relevant to each student's open tasks."},{"title":"Stale answers","detail":"Deadlines and policies change each term. Give every answer an owner and a review date."},{"title":"Bot as a wall","detail":"Students with urgent or personal problems cannot reach a person. Make handover easy and fast, and monitor repeat questions."},{"title":"Counting messages, not outcomes","detail":"Message volume says little about whether more students enrolled or completed tasks. Measure outcomes against a comparison group."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"An assistant that answers questions and sends reminders falls under the transparency duty of Article 50. It becomes high risk under Annex III point 3(a) if it is used to determine access or admission or to assign students to institutions, and under point 3(c) if it assesses the level of education a student will receive. Keep admission and placement decisions with staff."},"regulations":["eu-ai-act","gdpr","uk-gdpr","us-tcpa","nist-ai-rmf"],"guidance":[{"title":"Annex III, high risk AI systems (point 3, education and vocational training)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Lists AI that determines access or admission to educational institutions as high risk."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Protecting Student Privacy","issuer":"US Department of Education, Student Privacy Policy Office","region":"north-america","url":"https://studentprivacy.ed.gov/","note":"Guidance and resources on student privacy law (FERPA), relevant when the assistant uses education records to personalize messages."}],"controls":["AI disclosure in the first message and on the chat widget","Consent records for text messaging and an easy opt out","Content ownership and review dates per office","Data protection impact assessment covering education records used for personalization","Weekly review of unanswered questions and handovers"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** grounded in a **knowledge base** per office (admissions, aid,\nregistrar, housing), built from uploaded policies and **crawled website pages** with hybrid\nretrieval, so answers match the institution's own wording. **Custom functions** read each\nstudent's open tasks from the student information system through REST calls. A scheduled\n**agentic workflow** or **agentic task** sends each reminder before its deadline through a\n**custom function** that calls the institution's messaging provider, or through the outbound\n**email** channel.\n\nStudents' replies and questions reach the same agent on the **SMS**, **WhatsApp** or **web chat**\nchannel, with multi language support for international students and families. **Human\nhandover** rules route personal cases to the right office with the conversation attached.\n**Guardrails**, including the deterministic content scanner's self harm lexicon, flag risky\nmessages, and handover rules send them to trained staff. **PII masking** keeps student data out\nof model prompts, **analytics** show the top questions and unanswered topics each week, and the\n**GDPR toolkit** handles consent and data removal requests."},"faq":[{"question":"Can a chatbot really reduce summer melt?","answer":"Georgia State University reports that in the first summer its Pounce chatbot delivered more than 200,000 answers to incoming students, and that a new student portal and Pounce together reduced summer melt by 22 percent, an additional 324 students in class on the first day. Separately, in a randomized control trial the university saw a four percent overall decrease in the share of confirmed freshmen who did not enroll, and it says those gains came from the students who had access to Pounce."},{"question":"Do students actually respond to text message nudges and surveys?","answer":"Mainstay's case study for Austin Peay State University reports that the university's first intent to enroll campaign by text, in 2019, received a 40% response rate, and that staff could see which students planned to attend within one day. The vendor says the same task would have taken over a month by paper survey. It gives no enrollment figures, so this shows engagement, not outcomes."},{"question":"How much staff time does it save?","answer":"Mainstay's case study for Adelphi University reports 81,167 messages handled by its assistant Adele in the past year, which it estimates at 1,353 staff hours assuming approximately one minute per message. That is the vendor's estimate, not a time study. Georgia State's assistant vice president of undergraduate admissions said the university would have needed 10 more full time staff to handle the volume without Pounce."},{"question":"Is an admissions chatbot high risk under the EU AI Act?","answer":"Not if it only informs and reminds; then the transparency duty applies. It is high risk under Annex III point 3(a) if it determines access or admission, so admission decisions should stay with admissions staff."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the education scope with evidence from Georgia State University, Adelphi University and Austin Peay State University, checked against the sources. Editor review fixed the attribution of the Georgia State summer melt result, dropped the modelled Adelphi hours figure as a metric and corrected claim attributions."},{"date":"2026-09-27","note":"Second editor review separated the Georgia State randomized trial result (four percent) from the 22 percent summer melt result, limited the Pounce description to what the source states, aligned the Blits.ai build notes with the feature inventory (reminders through a custom function or outbound email, scanner as a guardrail), softened the unsourced claim about first generation students and added the Austin Peay survey response to the FAQ."}],"slug":"student-enrollment-and-services-assistant","url":"https://www.blits.ai/ai-use-cases/student-enrollment-and-services-assistant","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":140583.5,"min":81167,"max":200000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"georgia-state-university-pounce-enrollment-chatbot","pooled":true},{"id":"adelphi-university-adele-student-assistant","pooled":true}]}],"indicativeValueResult":{"low":25000,"high":240000},"evidence":["adelphi-university-adele-student-assistant","austin-peay-state-university-enrollment-chatbot","georgia-state-university-pounce-enrollment-chatbot"]},{"title":"AI assistant for tax questions and filing support","shortTitle":"Tax questions and filing assistant","seoTitle":"AI assistant for tax questions and filing","metaDescription":"Tax authorities use AI assistants for refund, payment and filing questions. HMRC logged 5.48 million interactions in 2024/25; IRS voice bots took 3 million calls.","definition":"An AI assistant that answers taxpayers' questions about taxes, deadlines, refunds and payments, lets authenticated taxpayers check their status or set up a payment plan within set rules, and guides them through filing, while assessments, penalties and disputes stay with the tax authority's staff and systems.","aliases":["tax authority chatbot","taxpayer virtual assistant","tax voice bot","tax filing chatbot"],"industries":["government"],"functions":["citizen-services","customer-service","finance-and-accounting"],"patterns":["conversational-agent","voice-agent","rag-knowledge-assistant","classification-and-routing"],"channels":["web-chat","voice","mobile-app","whatsapp"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"Tax authorities face sharply seasonal demand. Around filing deadlines and after\nevery policy change, phone lines and webchat fill with the same questions: where is my refund,\nhow do I get a reference number, can I pay in instalments, what does this notice mean. Long waits\npush people to give up, file late or file wrongly, which creates more work later in compliance\nand correspondence.\n\nThe questions are repetitive but the stakes are not trivial: a wrong answer about a deadline or a\nrelief can cost the taxpayer money. That is why the tax authority assistants on this page (HMRC's\ndigital assistant and the IRS voice bots and chatbots) classify intent and return approved\ncontent, and add authenticated actions such as payment plans only behind identity checks.","problemStats":[],"howItWorks":"1. **Recognise the intent.** The assistant classifies the question (refund status, payment plan,\n   notice, registration, how to file) on chat or on the phone.\n2. **Answer from approved content.** General questions get answers written or approved by the tax\n   authority, with links to the guidance; unclear questions get a choice of likely meanings.\n3. **Authenticate for account questions.** For refund status, balance or a payment plan, the\n   taxpayer verifies identity (shared secrets, PIN or national login) before the assistant reads\n   the account.\n4. **Act within rules.** Within fixed limits the assistant can set up or change a payment plan or\n   grant a payment extension, and confirms the result.\n5. **Escalate.** Complex, disputed or personal situations go to a human adviser, who sees the whole\n   conversation and completes identity checks.","valueDrivers":["cost-to-serve","customer-experience","compliance","inclusion-and-access"],"kpis":["interactions-handled","containment-rate","contact-deflection","accuracy","users-served","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A national tax authority with 3 million assisted contacts a year","inputs":[{"key":"contacts","label":"Assisted phone and chat contacts per year","low":3000000,"high":3000000,"unit":"contacts per year","note":"The reference authority."},{"key":"routineShare","label":"Share of contacts on routine topics (refund status, payments, forms)","low":0.4,"high":0.6,"unit":"fraction of contacts","note":"Editorial assumption. Replace with your own contact reason data."},{"key":"containment","label":"Share of routine contacts the assistant resolves","low":0.2,"high":0.4,"unit":"fraction of routine contacts","note":"Editorial assumption. For context, HMRC reports that webchat escalations to advisers fell 20% while assistant interactions grew 18.8%."},{"key":"costPerContact","label":"Cost of a human handled contact","low":5,"high":10,"unit":"USD per contact","note":"Editorial assumption. Replace with your own fully loaded cost."}],"formula":"contacts * routineShare * containment * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled contact cost avoided","caveat":"Gross contact cost only. It leaves out the effect on filing accuracy and late payment, the value of shorter queues at peak and the cost of building and running the assistant."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Information answers are low complexity; authenticated actions such as payment plans need integration with taxpayer account systems, strong identity checks and strict rules on what the assistant may change.","dataPrerequisites":["Approved answers and guidance per tax and topic, with owners and effective dates","Contact reason data by season, to pick intents and plan for peaks","Rules for payment plans and extensions that can be applied without judgment"],"integrations":["Taxpayer account, refund and payment systems through APIs","Identity verification and step up authentication","Telephony IVR and webchat platforms with handover to advisers","Notice and correspondence systems, to explain letters by reference"]},"implementation":{"steps":[{"title":"Start where the queues are","detail":"Use contact reason data to pick the few intents that dominate peak season (refund status, payments, notices), as the IRS did when it put voice bots on its Economic Impact Payment, notice and payment lines."},{"title":"Keep answers approved","detail":"Let the model classify and retrieve, and serve content the business owner approved. The IRS inventory stresses that its bots do not generate answers."},{"title":"Add authenticated self service","detail":"Add account questions and payment plans behind identity checks, with limits (amount, term) written as rules, and confirm every change back to the taxpayer."},{"title":"Design escalation with context","detail":"Pass the conversation to the adviser, as HMRC's webchat advisers see the assistant history, and complete identity checks before personal discussion."},{"title":"Prepare for peaks and changes","detail":"Retest before each filing season and after each budget, and plan capacity for the spike."}],"guardrails":["No assessment, penalty or dispute outcome is decided by the assistant","Account data only after identity verification, at the level the action needs","Actions such as payment plans only within written limits, with confirmation to the taxpayer","Answers from approved content with effective dates; refusal when the topic is out of scope","Personal data and tax identifiers masked in logs and model prompts"],"humanInTheLoop":"Advisers handle escalations, disputes, hardship and anything outside the rules. Content owners approve every answer and every change after a budget; a team samples conversations weekly and reviews failed intents and complaints.","kpisToInstrument":["Containment per intent, counting repeat contacts within seven days as not contained","Escalations to advisers, before and after launch","Intent recognition accuracy on a labelled test set","Payment plans set up through the assistant and their default rate","Satisfaction on assistant and adviser conversations"],"failureModes":[{"title":"Wrong deadline or relief answers","detail":"A fluent but wrong answer costs the taxpayer money and trust. Serve approved content, show effective dates and refuse outside scope."},{"title":"Peak season collapse","detail":"The assistant is tested in quiet months and fails at the deadline. Load test and retest before each season."},{"title":"Authentication friction","detail":"Taxpayers fail identity checks and fall back to the phone. Offer several proportionate methods and measure drop off."},{"title":"Private helpers without oversight","detail":"Third party filing assistants (such as ClearTax's WhatsApp agent) help people file but are not the authority; make official guidance easy for them to use and keep the authority's own channel authoritative."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A taxpayer assistant must tell people they are interacting with an AI system (Article 50). It is not listed in Annex III as long as it only informs and applies fixed rules. It becomes high risk under Annex III point 5(a) if it evaluates eligibility for, or grants, reduces, revokes or reclaims, public assistance benefits (which can include benefits paid through the tax system). Recital 59 says systems used for administrative proceedings by tax and customs authorities are not high risk law enforcement systems; audit selection and risk scoring are covered on a separate page."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Providers must design assistants so that taxpayers are informed they are interacting with an AI system, unless that is obvious from the context."},{"title":"Recital 59, AI systems used by tax and customs authorities","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/recital/59/","note":"Systems for administrative proceedings by tax and customs authorities should not be classified as high risk law enforcement systems."},{"title":"Algorithmic Transparency Recording Standard Hub","issuer":"Government Digital Service","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"HMRC publishes its assistant under this standard, including accuracy on known intents."},{"title":"M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust","issuer":"US Office of Management and Budget","region":"north-america","url":"https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf","note":"Sets the rules for US federal AI use and requires agencies to inventory their AI use cases at least annually and publish the inventory, where the IRS lists its voice bots and chatbots."}],"controls":["AI disclosure and a route to a human adviser on every channel","Entry in the public AI inventory or transparency register","Written limits for every action the assistant can take on an account","Change control tied to the budget and filing season calendar","Retention limits and access control on conversation logs containing tax data"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** of approved tax guidance, retrieved\nwith **hybrid search**, plus **flows** for the fixed journeys: an authentication block before any\naccount question, and deterministic steps for payment plans that call the authority's systems\nthrough **custom functions** with their own limits. Where a classic intent model is preferred,\nthe platform's NLU bots with nightly retraining can sit in front of the agent.\n\nThe same logic serves **web chat, WhatsApp, voice and a mobile app** (through the API channel), with streaming speech\nrecognition, DTMF capture for reference numbers and call transfer on the phone. **PII masking**\nremoves tax identifiers before text reaches a model, **guardrails** keep answers on approved\ncontent, and **human handover** passes the transcript to advisers. **Test suites** replay each\nseason's top questions before the deadline, and **analytics** show recognition and handover per intent."},"faq":[{"question":"How much volume do tax assistants handle?","answer":"HMRC's digital assistant had 5.48 million interactions in the 2024/25 tax year up to 6 March 2025, and IRS voice bots had answered over 3 million calls by June 2022, about a year after the first one went live in May 2021."},{"question":"Do tax authorities use generative AI for these answers?","answer":"Not in the deployments on this page. The IRS inventory states its chatbots and voice bots return content predetermined by content owners, and HMRC matches intents to approved answers. The IRS tested a generative AI chatbot for volunteer tax preparers as a proof of concept (listed as retired in 2025), while the private filing service ClearTax runs a WhatsApp agent on Azure OpenAI models."},{"question":"How accurate is intent recognition?","answer":"HMRC reports 83.03% accuracy on known intents in its March 2025 test set. Plan for the rest with disambiguation questions and an easy route to a human."}],"related":["citizen-information-assistant","tax-compliance-risk-scoring","benefits-eligibility-and-application-assistant","public-service-translation","first-line-contact-centre-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from UK transparency records, the US federal AI inventory, the IRS newsroom and vendor case studies, with every quote verified against the source."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsupported claims (most seasonal demand, most tax authorities, voice bots within months of launch), corrected the EU AI Act basis (Annex III point 5(a), Recital 59), added Recital 59 guidance, the contact deflection KPI and the IRS voice bot and chatbot platforms (Nuance, eGain), and added the SEO title and meta description."}],"slug":"tax-questions-and-filing-assistant","url":"https://www.blits.ai/ai-use-cases/tax-questions-and-filing-assistant","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":4240000,"min":3000000,"max":5480000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"hmrc-ask-hmrc-digital-assistant","pooled":true},{"id":"irs-taxpayer-voicebots-and-chatbots","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":83.03,"min":83.03,"max":83.03,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"hmrc-ask-hmrc-digital-assistant","pooled":true}]},{"kpi":"contact-deflection","label":"Contact deflection","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":20,"min":20,"max":20,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"hmrc-ask-hmrc-digital-assistant","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":200000,"min":200000,"max":200000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"cleartax-whatsapp-tax-filing-agent","pooled":true}]}],"indicativeValueResult":{"low":1200000,"high":7200000},"evidence":["cleartax-whatsapp-tax-filing-agent","hmrc-ask-hmrc-digital-assistant","irs-machine-translation","irs-taxpayer-voicebots-and-chatbots"]},{"title":"AI assistant for telecom order to activation and eSIM onboarding","shortTitle":"Order to activation and eSIM onboarding","seoTitle":"AI assistant for telecom activation and eSIM","metaDescription":"AI assistants take telecom customers from order to working service, including porting and eSIM setup. Singtel's Shirley completed 76% of roaming sign ups unassisted.","definition":"An AI assistant that takes a new or existing customer from order to a working service: it collects and checks the order details, guides number porting, eSIM download or SIM activation and installation appointments, tracks the order and fixes or escalates the step that is stuck, on messaging, app, web or phone.","aliases":["activation assistant","eSIM activation bot","number porting assistant","order tracking chatbot","new customer onboarding agent"],"industries":["telecommunications"],"functions":["sales","onboarding-and-kyc","customer-service","operations"],"patterns":["conversational-agent","agentic-workflow","classification-and-routing","document-processing"],"channels":["whatsapp","mobile-app","web-chat","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"front-office","problem":"The weeks between \"I want to join\" and \"it works\" generate many of a new customer's first\ncontacts. Customers ask where their SIM or router is, why the number transfer has not happened,\nhow to install an eSIM, why the service is not active yet, or when the engineer will come. Each\nstep (identity check, credit check, porting, provisioning, delivery, installation) sits in a\ndifferent system, and nobody in the contact centre sees the whole order at a glance.\n\nA stuck order also gives a new customer a reason to cancel within the cooling off period. eSIM\nremoves the wait for a SIM card in the post, but customers still get stuck on QR codes, device compatibility and\ntransfers between phones, and a failed download can need a person to reissue the profile.","problemStats":[],"howItWorks":"1. **Take or find the order.** The assistant either captures a new order (plan, device, number\n   transfer, address) or finds the existing one after identity verification.\n2. **Check each step.** Through order, porting, logistics and provisioning tools it sees which\n   step is complete, which is pending and which has failed.\n3. **Guide the customer.** It walks the customer through their part: confirming the porting\n   code, scanning the eSIM QR code on a compatible device, inserting the SIM, or choosing an\n   installation slot.\n4. **Fix or escalate the stuck step.** Within its permissions it reissues an eSIM profile,\n   resubmits a porting request or rebooks an appointment; anything needing a person (a failed\n   identity check, a credit referral, fraud signals) goes to a specialist with the order history.\n5. **Confirm the service works.** It checks activation and sends the customer a clear summary of\n   what they bought and what happens next.","valueDrivers":["customer-experience","speed","cost-to-serve","revenue-growth"],"kpis":["automation-rate","users-served","processing-time-reduction","conversion-rate-uplift"],"indicativeValue":{"referenceOrg":"An operator that activates 400,000 new mobile and broadband connections a year","inputs":[{"key":"activations","label":"New connections and activations per year","low":400000,"high":400000,"unit":"activations per year","note":"The reference operator."},{"key":"contactRate","label":"Assisted contacts per activation between order and working service","low":0.3,"high":0.5,"unit":"contacts per activation","note":"Editorial assumption, replace with the order and activation share of your contact reason report."},{"key":"resolvedShare","label":"Share of those contacts the assistant resolves without a human","low":0.3,"high":0.6,"unit":"fraction of contacts","note":"Conservative against the benchmark on this page (Singtel reports 76% of roaming sign up requests completed without a Customer Care officer), because porting and failed identity checks need people."},{"key":"costPerContact","label":"Cost of a human handled order contact","low":5,"high":8,"unit":"USD per contact","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"activations * contactRate * resolvedShare * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled order contact cost avoided","caveat":"Gross avoided contact cost only. It leaves out fewer cancellations in the cooling off period, faster time to revenue, the cost of the AI and the order and provisioning integrations, and the fraud controls that activation and porting need."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The order journey crosses identity, credit, porting, logistics, provisioning and field service systems, often with batch updates. Actions such as reissuing an eSIM or resubmitting a port are also the ones fraudsters target, so identity and fraud controls must be strong.","dataPrerequisites":["Order status per step, available by API and close to real time","eSIM and device compatibility information with an owner and review date","Porting rules and error codes with their fixes","Contact reasons for new customers in their first 60 days"],"integrations":["Order management and CRM","Number porting system","eSIM profile management and SIM provisioning","Logistics tracking and field service scheduling","Identity verification and fraud screening"]},"implementation":{"steps":[{"title":"Map the order journey and where it breaks","detail":"List every step from order to working service, the system that owns it and the most common failures, using contacts from customers in their first 60 days."},{"title":"Start with status and guidance","detail":"Order status, delivery tracking, installation slots and eSIM installation guidance are read only and high volume. Launch these before any action that changes the order."},{"title":"Add fixes with fraud controls","detail":"Reissuing an eSIM, resubmitting a port or changing a delivery address are attractive to fraudsters. Put step up verification and fraud checks in front of each of them."},{"title":"Close the loop on activation","detail":"Have the assistant confirm the service is active and summarise the purchase, so the first bill does not become a second contact."},{"title":"Measure per step","detail":"Track completion and drop out per journey step, not only per conversation, and fix the systems behind the steps that fail most."}],"guardrails":["Step up identity verification before any eSIM reissue, SIM swap, port or address change","Fraud screening on every action that moves a number or a SIM","Credit and identity decisions stay in the existing governed processes","Clear confirmation of plan, price and contract terms before the order is placed","Handover to a person on request and after repeated failure of any step"],"humanInTheLoop":"Specialists handle failed identity checks, credit referrals, suspected fraud and porting disputes. The fraud team approves every action the assistant can perform on a SIM or number, and operations review stuck orders the assistant escalated each day.","kpisToInstrument":["Share of orders activated without an assisted contact","Time from order to working service, per product and channel","eSIM and porting failures resolved by the assistant versus escalated","Cancellations within the cooling off period","Fraud attempts caught at SIM and porting steps"],"failureModes":[{"title":"SIM swap and porting fraud","detail":"A weakly verified request moves a victim's number to a fraudster. Verify strongly and screen every SIM and port action."},{"title":"Status the systems do not know","detail":"The assistant reports a step as done because a batch update has not run. Show the time of the last update and escalate stale orders."},{"title":"Device compatibility guesswork","detail":"The assistant tells a customer their phone supports eSIM when it does not. Use a maintained compatibility list, not the model's memory."},{"title":"Dead ends for new customers","detail":"A new customer stuck in automation cancels. Offer a person after repeated failure."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing assistant must disclose that it is AI (Article 50). Biometric verification whose sole purpose is to confirm that a person is who they claim to be is excluded from remote biometric identification in Annex III point 1(a); creditworthiness assessment of individuals (Annex III point 5(b)) would be high risk and belongs in a separate governed process."},"regulations":["eu-ai-act","gdpr","telecom-consumer-rules","eecc"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Simpler and quicker broadband switching is here","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/phones-and-broadband/switching-provider/simpler-broadband-switching-is-here","note":"Under One Touch Switch, UK landline and broadband customers only contact their new provider, and providers must compensate customers if the switch goes wrong or leaves them without service for more than one working day, which sets the bar for an activation assistant."}],"controls":["AI disclosure at the start of every conversation","Documented list of order actions with verification level and fraud checks per action","Audit trail of every SIM, eSIM and porting action","Regression tests on order scenarios, including fraud attempts, for every change","Daily review of stuck and escalated orders"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that read and update the order,\nporting, eSIM and logistics systems, and a **flow** with an authentication block for any action\non a SIM or number. A **knowledge base** with hybrid retrieval holds device compatibility and\ninstallation guides, customers can **upload files and photos** in the chat, and\n**agentic tasks** can recheck a pending port or delivery on a schedule and act when the\nstatus changes.\n\nThe agent runs on **WhatsApp, web chat and voice**, and inside the operator's own app through\nthe **REST and WebSocket API channels**. **Human handover** passes\nfailed checks and suspected fraud to specialists with the order history, **guardrails** and\n**PII masking** protect identity data, **test suites** replay order scenarios on every change,\nand the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"Can an AI assistant really complete sign ups and activations?","answer":"For simpler orders, yes. In its initial results after launch, Singtel reports 76% of roaming sign up requests completed without a Customer Care officer. Jio's WhatsApp assistant, built with Haptik, covers the 5G journey from lead to porting and plan purchase, and Haptik reports it acquires 8,000 new Jio Fiber and 5G customers a day. Complex orders with credit checks still need people."},{"question":"Is there public evidence for AI in eSIM onboarding specifically?","answer":"None of the deployments on this page publishes results for eSIM onboarding specifically. The evidence covers sign up, porting and plan purchase in general; eSIM guidance and profile reissue are a natural extension, with strong identity checks, because reissuing an eSIM moves the number to a new device just as a SIM swap does."},{"question":"How do we stop the assistant from enabling SIM swap fraud?","answer":"Treat every SIM, eSIM and porting action as high risk: step up verification, fraud screening and an audit trail for each, and route anything unusual to a specialist."}],"related":["plan-upgrade-and-sales-assistant","device-and-connectivity-troubleshooting-agent","retail-store-and-kiosk-assistant","telecom-fraud-detection","business-connectivity-quoting-and-service-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Unpublished by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the telecommunications vertical from Singtel, Jio and Verizon sources checked word for word."},{"date":"2026-09-25","note":"Consolidation pass: added European Electronic Communications Code to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; softened unsourced claims in the problem and FAQ; cited Annex III points 1(a) and 5(b); corrected the Ofcom note to the page text; limited the build channels to those Blits.ai offers; confirmed the Jio year (2023) from an archived copy."},{"date":"2026-09-26","note":"Second fact check against live sources (Haptik, Sierra, Singtel, Verizon, Ofcom, AI Act): all quotes confirmed; replaced the unsourced claim that eSIM makes activation instant and put the Singtel result in the past tense in the meta description."},{"date":"2026-09-26","note":"Third fact check against live sources (Haptik, Sierra, Singtel, Verizon, Ofcom, AI Act, feature inventory): all quotes and prose confirmed; the FAQ now ties the Singtel 76% to its initial results after launch, as the release does, instead of to the first six weeks."}],"slug":"order-to-activation-and-esim-onboarding-assistant","url":"https://www.blits.ai/ai-use-cases/order-to-activation-and-esim-onboarding-assistant","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":76,"min":76,"max":76,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"singtel-shirley-agentic-ai-agent","pooled":true}]}],"indicativeValueResult":{"low":180000,"high":960000},"evidence":["jio-whatsapp-customer-lifecycle","singtel-shirley-agentic-ai-agent","verizon-ai-customer-experience-transformation"]},{"title":"AI assistant for telecom plan upgrades, add ons and sales","shortTitle":"Plan upgrade and sales assistant","seoTitle":"AI sales assistant for telecom plan upgrades","metaDescription":"AI assistants help telecom customers choose and buy plans, add ons and devices. See results from Singtel, Telenet and Orange France, the value math and the rules.","definition":"An AI assistant that helps existing and prospective customers choose, compare and buy the right mobile, broadband or TV plan, device or extra, in the app, in messaging, on the phone or through a human advisor, using the customer's usage and eligibility and the operator's current offers, and that completes the order or passes a ready quote to a person.","aliases":["telco sales assistant","upgrade assistant","guided selling for telecom","add on recommendation agent","advisor sales copilot"],"industries":["telecommunications"],"functions":["sales","customer-service","marketing"],"patterns":["recommendation-and-personalization","conversational-agent","rag-knowledge-assistant","voice-agent","agentic-workflow"],"channels":["mobile-app","web-chat","whatsapp","voice","agent-desktop"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"Telecom offers are hard to compare. Plans differ by data, speed, contract length, device\ninstalments, bundles and promotions that change often, and eligibility depends on the\ncustomer's current contract, credit and address. Customers who cannot compare offers may pick a\nplan that does not fit, or call to ask, and the answer can depend on how current the advisor's\nknowledge of the promotions is.\n\nOperators already score who is likely to upgrade, but the moment of decision happens in a\nconversation: a question in the app, a chat about roaming before a trip, a call about a slow\nconnection. Rule based chatbots could list plans, not reason about which one fits this customer,\nand advisors lose time during sales calls looking up the current offer and writing up the call.","problemStats":[],"howItWorks":"1. **Understand the need.** The assistant asks about usage, household, devices and budget, or\n   reads the customer's actual usage and contract after authentication.\n2. **Check eligibility and current offers.** It retrieves the offers, promotions and upgrade\n   eligibility that apply to this customer from the product catalogue and decisioning engine,\n   never from memory.\n3. **Recommend and compare.** It proposes a small number of options with the reasons and the\n   total cost, including what changes on the next bill, and compares devices on request.\n4. **Complete or hand over.** For simple purchases (a roaming pass, an extra, a plan change) it\n   completes the order with confirmation; for contracts, devices on credit or complex bundles it\n   prepares the quote and passes it to an advisor or the checkout.\n5. **Assist advisors.** In stores and contact centres the same engine suggests the next best\n   offer and answers product questions for the advisor during the conversation.","valueDrivers":["revenue-growth","customer-experience","employee-productivity"],"kpis":["conversion-rate-uplift","revenue-uplift","interactions-handled","users-served","handling-time-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"An operator with 3 million postpaid mobile and broadband customers","inputs":[{"key":"customers","label":"Postpaid customers","low":3000000,"high":3000000,"unit":"customers","note":"The reference operator."},{"key":"engagedShare","label":"Share of customers who have a sales conversation with the assistant or an assisted advisor each year","low":0.2,"high":0.4,"unit":"fraction of customers","note":"Editorial assumption, replace with your own digital and assisted sales reach."},{"key":"incrementalConversion","label":"Additional upgrades or extras bought per engaged customer","low":0.01,"high":0.02,"unit":"fraction of engaged customers","note":"Editorial assumption and deliberately low against the benchmark on this page (Pega reports a 75% increase in offer acceptance at Telenet), because that figure is relative, has no stated baseline or control group and comes from one vendor story."},{"key":"marginPerSale","label":"Incremental annual margin per upgrade or extra","low":30,"high":60,"unit":"USD per sale per year","note":"Editorial assumption, replace with your own margin per upgrade."}],"formula":"customers * engagedShare * incrementalConversion * marginPerSale","currency":"USD","period":"per year","resultLabel":"Incremental annual margin from assisted upgrades","caveat":"Counts only the first year's margin on additional sales. It leaves out lower churn from better fitting plans, the cost of the AI and catalogue integration, cannibalisation of sales that would have happened anyway, and discounts given to close."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Recommending is easy; recommending correctly is not. The work is in a clean, current product and promotion catalogue, real eligibility rules and an order API. Selling on credit or under long contracts adds pre contract information and consent duties.","dataPrerequisites":["A current product, price and promotion catalogue with start and end dates","Upgrade eligibility and credit rules available by API","Customer usage and contract data for authenticated recommendations","Marketing consent and contact preferences per customer"],"integrations":["Product catalogue and pricing engine","Decisioning or next best action engine","Order management and checkout","CRM with contract, usage and consent data","Contact centre and store systems for advisor assistance"]},"implementation":{"steps":[{"title":"Fix the catalogue before the conversation","detail":"Put every plan, extra and promotion, with its eligibility and end date, into one source the assistant reads through a tool. Stale promotions are the fastest way to lose trust."},{"title":"Start with simple, reversible purchases","detail":"Roaming passes, data extras and plan changes without a new contract are low risk. Add devices on credit and new contracts later, with the full pre contract information flow."},{"title":"Show the total cost","detail":"Make the assistant state the monthly and total cost and the change on the next bill for every option, so customers do not feel sold to and complaints stay low."},{"title":"Give advisors the same brain","detail":"Use the same recommendation and product answers in stores and contact centres, so a customer hears the same offer on every channel."},{"title":"Measure against a control group","detail":"Hold out a random share of customers or conversations to measure real incremental conversion, not sales that would have happened anyway."}],"guardrails":["Prices, promotions and eligibility come only from catalogue tools, never from the model","Total cost and contract length stated before any order is confirmed","Marketing consent checked before any proactive or outbound offer","No offers to customers flagged as vulnerable or in financial difficulty without human review","Explicit confirmation before every order, with a record of what the customer saw"],"humanInTheLoop":"Advisors complete contract and device on credit sales from the prepared quote. The commercial team approves every new offer the assistant may present, and a quality team reviews a weekly sample of sales conversations for mis selling and unclear pricing.","kpisToInstrument":["Incremental conversion against a holdout group","Order cancellations and returns within the cooling off period","Complaints about sales or pricing that mention the assistant","Advisor handling time on sales calls with and without assistance","Revenue and margin per assisted sale"],"failureModes":[{"title":"Stale or wrong offers","detail":"The assistant quotes a promotion that ended. Keep all offers in a tool with end dates and test daily."},{"title":"Mis selling","detail":"Customers are pushed to bigger plans they do not need. Recommend on actual usage, show the total cost and review samples."},{"title":"Ignoring consent","detail":"Proactive offers reach customers who opted out of marketing. Check consent in the tool, not in the prompt."},{"title":"Credit decisions by the back door","detail":"The assistant decides who may buy a device on credit. Keep credit checks in the existing, governed process."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"A sales assistant is limited risk with an Article 50 duty to disclose AI. If it assesses the creditworthiness of individuals for devices on credit, that part is high risk under Annex III point 5(b), so keep credit decisions in the existing governed process. Selling that uses manipulative or deceptive techniques, or exploits a customer's age, disability or economic situation, to materially distort a purchase decision in a way likely to cause significant harm is prohibited under Article 5(1)(a) and (b)."},"regulations":["eu-ai-act","gdpr","telecom-consumer-rules","eecc","us-tcpa"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Directive on privacy and electronic communications (Directive 2002/58/EC)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2002/58/oj","note":"Sets consent rules for unsolicited electronic marketing, which apply to proactive offers sent by messaging, email or automated calls."},{"title":"Customers to get clearer broadband information","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/phones-and-broadband/bills-and-charges/customers-to-get-clearer-broadband-information","note":"Ofcom guidance says UK broadband providers must describe the underlying network technology in clear terms before a customer agrees to buy, whether online, by phone or in person, so a sales assistant has to give the same information."}],"controls":["AI disclosure at the start of every sales conversation","Offer approval workflow with owners, start and end dates","Consent and vulnerability checks enforced in tools","Record of the offer, price and terms shown before each order","Monthly review of cancellations, returns and sales complaints"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that read the product catalogue,\neligibility and the decisioning engine, and that place orders through the order API. A\n**knowledge base** with hybrid retrieval answers product questions from approved content, rich\n**cards and carousels** in the chat widget compare plans and devices, and a **flow** handles the\norder confirmation with the total cost and consent check. Payments for extras can be taken in\nthe conversation with **payment links**.\n\nThe same agent serves **web chat, WhatsApp and voice**, reaches the operator's own app through\nthe **REST and WebSocket API**, and can run for advisors in **Microsoft Teams** or the agent\ndesktop through the same API. **Guardrails** block unapproved\nclaims, **PII masking** protects customer data, **human handover** passes a prepared quote to an\nadvisor, and **test suites** replay sales conversations whenever offers change. The platform is\nmodel agnostic with EU and UAE data residency."},"faq":[{"question":"Does AI actually increase telecom sales?","answer":"There is early evidence, mostly vendor reported. Pega reports that Telenet saw a 75% increase in offer acceptance and a 33% increase in cross sell with AI decisioning, and Singtel reports that customers bought more than 200 roaming add ons independently through its assistant Shirley, among the initial results after launch. Neither states a control group, so measure against a holdout group before you trust the uplift."},{"question":"Should the assistant sell to customers or help advisors sell?","answer":"Many operators do both. Simple extras and plan changes suit self service, while devices on credit and new contracts often go through advisors, who benefit from the same offer and product information: T-Mobile gives store and call centre staff its PromoGenius app, and Orange France gives 3,000 sales advisors an AI assistant during customer calls."},{"question":"What rules apply to AI generated offers?","answer":"The usual telecom and consumer rules: clear pre contract information, total cost, cooling off rights and marketing consent for proactive offers. An AI assistant has to follow them in every conversation, which is easier to prove when prices and terms come from tools."}],"related":["churn-prediction-and-retention-offers","order-to-activation-and-esim-onboarding-assistant","retail-store-and-kiosk-assistant","business-connectivity-quoting-and-service-assistant","bill-explanation-and-billing-dispute-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the telecommunications vertical from Telenet, Singtel, T-Mobile, Orange France, Mobily, Jio and Verizon sources checked word for word."},{"date":"2026-09-25","note":"Consolidation pass: added European Electronic Communications Code to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: attributed the Telenet figure to Pega, corrected the Singtel roaming add on wording, softened unsourced wording in the problem, clarified the Ofcom note and the channels in the Blits.ai section, dated the Orange source and added seoTitle and metaDescription."},{"date":"2026-09-26","note":"Second fact check against sources: aligned the Singtel add on wording with the release's initial results, added the Article 5 prohibition on manipulative selling to the EU AI Act basis and added the Telephone Consumer Protection Act for automated outbound offers."}],"slug":"plan-upgrade-and-sales-assistant","url":"https://www.blits.ai/ai-use-cases/plan-upgrade-and-sales-assistant","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":3,"nUpTo":0,"median":1000000,"min":70000,"max":45000000,"byClaimant":{"organization":2,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"vodafone-tobi-virtual-assistant","pooled":true},{"id":"orange-france-maia-advisor-assistant","pooled":true},{"id":"singtel-shirley-agentic-ai-agent","pooled":true}]},{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":75,"min":75,"max":75,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"telenet-next-best-action-decisioning","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":83000,"min":83000,"max":83000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"t-mobile-promogenius-retail-agent","pooled":true}]}],"indicativeValueResult":{"low":180000,"high":1440000},"evidence":["jio-whatsapp-customer-lifecycle","mobily-agentic-ai-self-service","orange-france-maia-advisor-assistant","singtel-shirley-agentic-ai-agent","t-mobile-promogenius-retail-agent","telenet-next-best-action-decisioning","verizon-ai-customer-experience-transformation","virgin-media-o2-lumi-ai-advisor-assistant","vodafone-tobi-virtual-assistant"]},{"title":"AI assistant for telecom retail stores, from associate copilot to digital human kiosk","shortTitle":"Retail store and kiosk assistant","seoTitle":"AI assistant for telecom stores and kiosks","metaDescription":"Telecom store assistants answer staff on offers and devices. Salesforce reports 95% accuracy at launch for Bouygues Telecom; T-Mobile's app has over 83,000 users.","definition":"An AI assistant for telecom shops that gives store associates quick, sourced answers on plans, promotions, devices and the customer's account during the conversation, and that can also greet and serve customers directly on an in store screen or kiosk, sometimes as a digital human, handing them to an associate when they are ready to buy or need help.","aliases":["store associate copilot","retail assistant for telecom","digital human kiosk","in store AI assistant","shop floor product assistant"],"industries":["telecommunications"],"functions":["sales","customer-service","knowledge-management"],"patterns":["rag-knowledge-assistant","digital-human","conversational-agent","recommendation-and-personalization"],"channels":["kiosk","internal-tools","agent-desktop","web-chat"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"front-office","problem":"Telecom stores sell technical products whose offers keep changing: new devices, discounts and\ntrade in values, and plans whose eligibility depends on the customer's contract. At T-Mobile, a\ndaily promotions report sent to retail representatives had become complex and hard to search,\ntrade in values sat in other systems, and device details meant a visit to manufacturers'\nwebsites, so representatives often left the customer conversation to find an answer. At Bouygues\nTelecom, contact centre reps sifted through more than 500 articles, sometimes skimming up to 12\npages for one inquiry, and turned to supervisors or colleagues when unsure, which gave customers\ninconsistent answers.\n\nMany customers, meanwhile, doubt their own technical knowledge and may feel embarrassed to ask a\nperson for help, as UneeQ notes in its Deutsche Telekom case study; unsure buyers tend to pick the\ncheapest option and can be disappointed. A static screen in a store can show a brochure, but it\ncannot answer a question about the customer's own situation.","problemStats":[],"howItWorks":"1. **Answer the associate in seconds.** On a tablet or the store system, the associate asks in\n   plain language about a promotion, a device comparison or a policy, and gets a summarised answer\n   from approved content and live data, with the source.\n2. **Bring the customer's context.** After the customer is identified, the assistant shows their\n   plan, contract end date, open cases and eligible offers, so the conversation starts informed.\n3. **Build comparisons to show.** It assembles device and plan comparisons that the associate can\n   show or send to the customer.\n4. **Serve customers at the screen.** On a kiosk or a digital human screen, customers ask\n   questions, compare options and check eligibility themselves, and the assistant calls an\n   associate or books a slot when they want to buy.\n5. **Keep every channel consistent.** The same knowledge serves the contact centre, online sales\n   and stores, so a customer hears the same answer everywhere.","valueDrivers":["employee-productivity","revenue-growth","customer-experience"],"kpis":["accuracy","users-served","time-saved-per-task","conversion-rate-uplift","customer-satisfaction","employee-adoption"],"indicativeValue":{"referenceOrg":"An operator with 300 stores","inputs":[{"key":"stores","label":"Stores","low":300,"high":300,"unit":"stores","note":"The reference operator."},{"key":"lookupsPerDay","label":"Product and account lookups per store per day","low":20,"high":40,"unit":"lookups per store per day","note":"Editorial assumption, replace with your own store activity."},{"key":"openDays","label":"Trading days per year","low":300,"high":360,"unit":"days per year","note":"Editorial assumption."},{"key":"hoursSaved","label":"Associate time saved per lookup","low":0.03,"high":0.07,"unit":"hours per lookup","note":"Editorial assumption of about two to four minutes per lookup. Salesforce reports that Bouygues Telecom's Iris answers in seconds a task that once took minutes of manual searching.","sourceUrl":"https://www.salesforce.com/customer-stories/bouygues-telecom/agentic-service-faqs/"},{"key":"hourlyCost","label":"Fully loaded associate cost","low":20,"high":30,"unit":"USD per hour","note":"Editorial assumption, replace with your own cost."}],"formula":"stores * lookupsPerDay * openDays * hoursSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Associate time released","caveat":"Values only the associate time saved on lookups. It leaves out higher conversion and basket size from better informed conversations, shorter queues, faster onboarding of new staff, and the cost of the AI, devices, kiosks and integrations. Released time is only a saving if staffing or sales change as a result."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"An associate copilot over a clean knowledge base is quick to deliver. Account context and live promotions need CRM and catalogue integration, and a customer facing digital human adds hardware, speech in a noisy store, accessibility and privacy design.","dataPrerequisites":["One current source for promotions, prices and trade in values with start and end dates","Device specifications from manufacturers or a product data feed","Knowledge articles rewritten for retrieval, with owners and review dates","Store and customer data access rules per role"],"integrations":["Product catalogue and promotions system","CRM and customer account data","Point of sale and order systems","Store tablets, kiosks or digital human screens","Queue management or appointment booking in store"]},"implementation":{"steps":[{"title":"Clean the knowledge first","detail":"Rewrite and standardise the articles associates use, as Bouygues Telecom did with Salesforce for its more than 500 articles, and give each an owner and review date."},{"title":"Set an accuracy bar before launch","detail":"Agree a minimum accuracy on a test set of real associate questions and launch only when the assistant meets it. Bouygues Telecom required its agent to exceed 90% accuracy or outperform a supervisor."},{"title":"Start with associates, then customers","detail":"Associates can judge and correct answers; customers cannot. Launch the copilot first and add a customer facing kiosk or digital human once answers are reliable."},{"title":"Design the kiosk for a real shop floor","detail":"Test speech in store noise, offer touch and text alternatives, avoid showing personal data on a public screen, and make calling an associate one tap."},{"title":"Share content across channels","detail":"Use one knowledge source for stores, contact centre and online sales, so an update reaches every channel at once."}],"guardrails":["Answers only from approved content and live tools, with the source shown to associates","Prices and promotions only from the catalogue, with end dates enforced","No personal account data on a public screen without identification and privacy screening","AI disclosure on every customer facing screen and digital human","No emotion recognition or camera analysis of customers without a lawful basis and clear notice"],"humanInTheLoop":"Associates decide what to tell and sell to the customer and remain responsible for the sale. Content owners approve every article and promotion the assistant uses, and store managers report wrong answers, which are reviewed weekly.","kpisToInstrument":["Answer accuracy on a weekly checked sample of associate questions","Weekly active associates as a share of store staff","Time to answer common questions, before and after","Conversion and basket size in stores with and without the assistant","Kiosk conversations handed to associates and resulting sales"],"failureModes":[{"title":"Outdated promotions","detail":"Associates quote an ended offer from the assistant. Keep promotions in a tool with end dates, not in documents."},{"title":"A gimmick instead of a tool","detail":"A digital human draws attention but cannot answer real questions. Measure conversations that lead to help or a sale, not visits."},{"title":"Privacy on the shop floor","detail":"Account details appear on a screen others can see. Design the screen and identification flow for a public space."},{"title":"Different answers per channel","detail":"The store says one thing and the app another. Use one knowledge source for all channels."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A digital human or kiosk that talks to customers must be designed so that they are told they are interacting with an AI system (Article 50(1)). An associate copilot over product content is not listed in Annex III and is minimal risk. Inferring the emotions of employees at work is prohibited (Article 5(1)(f)); emotion recognition of customers by camera is high risk under Annex III point 1(c), biometric categorisation by sensitive or protected attributes is high risk under Annex III point 1(b), and categorisation that infers race, political opinions, religion or sexual orientation is prohibited under Article 5(1)(g). Both need separate legal review."},"regulations":["eu-ai-act","gdpr","telecom-consumer-rules","eecc","eu-accessibility-act"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context, which includes digital humans on store screens."},{"title":"Customers to get clearer broadband information","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/phones-and-broadband/bills-and-charges/customers-to-get-clearer-broadband-information","note":"UK providers must give clear information about the broadband technology before a customer buys, including in person, so in store assistants must use the same approved descriptions."}],"controls":["AI disclosure on every customer facing screen","Content ownership and review dates for all articles and promotions","Role based access to customer data in the store","Accuracy tests before launch and after every content or model change","Privacy impact assessment for kiosks and any camera or microphone in store"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the associate copilot is an **AI agent** over a **knowledge base** of product,\npromotion and policy content with hybrid retrieval and document version control, with **custom\nfunctions** for live promotions, trade in values and the customer's account. Associates use it\nthrough the web widget on store tablets or in **Microsoft Teams**, and rich in chat cards show\ndevice and plan options that the associate can show the customer.\n\nFor customers, the same agent runs as a **digital human**: a photorealistic avatar with lip\nsynced speech, streamed to an in store screen, with **voice** input and text as an alternative.\n**Human handover** escalates to a staff member, **guardrails** check answers against the\noperator's policies, **test suites** of real associate questions are rerun after every content\nchange, and the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"How accurate does a store assistant need to be?","answer":"Set the bar before launch. Bouygues Telecom required its Iris agent to exceed 90% accuracy or outperform a supervisor before it entered the contact centre; Salesforce reports it reached 95% on day one, and Iris has since been extended to 500 retail stores."},{"question":"Should we start with a digital human or an associate copilot?","answer":"Starting with associates carries less risk, because they can catch mistakes; the documented deployments at T-Mobile (PromoGenius, over 83,000 unique users among retail and call centre staff) and Bouygues Telecom (Iris) are both associate tools. Customer facing digital humans, such as Deutsche Telekom's Selena, who explains home broadband options, and Max, who answers questions on the Telekom website and app, guide customers to products; in a store, plan a clear handover to an associate when the customer is ready to buy."},{"question":"What privacy issues come with kiosks?","answer":"Screens in a public space should not show account details without identification and privacy design, microphones and cameras need a clear legal basis and notice, and inferring the emotions of employees at work is prohibited under Article 5 of the EU AI Act."}],"related":["plan-upgrade-and-sales-assistant","live-agent-assist","order-to-activation-and-esim-onboarding-assistant","atm-and-self-service-device-assistance","enterprise-knowledge-search"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the telecommunications vertical from T-Mobile, Bouygues Telecom and Deutsche Telekom sources checked word for word."},{"date":"2026-09-25","note":"Consolidation pass: added European Electronic Communications Code, European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Bouygues Telecom accuracy bar (more than 90% or better than a supervisor), removed unsupported claims about Deutsche Telekom digital humans in stores and operator practice, made the EU AI Act basis precise (Article 50(1), Article 5(1)(f), Annex III point 1), aligned the Blits.ai build with the feature inventory, and added the 6,000 Iris users metric."},{"date":"2026-09-27","note":"Fact checked against sources: split the Annex III point 1 biometric basis into emotion recognition (1(c), high risk), sensitive attribute categorisation (1(b), high risk) and the Article 5(1)(g) prohibition; gave Selena and Max their own sourced roles instead of a shared, unsourced app claim; removed the unsupported automatic test suite trigger."}],"slug":"retail-store-and-kiosk-assistant","url":"https://www.blits.ai/ai-use-cases/retail-store-and-kiosk-assistant","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":44500,"min":6000,"max":83000,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"t-mobile-promogenius-retail-agent","pooled":true},{"id":"bouygues-telecom-iris-service-agent","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":95,"min":95,"max":95,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"bouygues-telecom-iris-service-agent","pooled":true}]}],"indicativeValueResult":{"low":1080000,"high":9072000},"evidence":["bouygues-telecom-iris-service-agent","deutsche-telekom-uneeq-digital-humans","t-mobile-promogenius-retail-agent"]},{"title":"AI cash flow forecasting for corporate treasury","shortTitle":"Treasury cash forecasting","seoTitle":"AI cash flow forecasting for corporate treasury","metaDescription":"AI tools sort a company's transactions and forecast its cash positions. J.P. Morgan reports Prysmian halved manual work and Domino's cut data cleanup by up to 90%.","definition":"Machine learning and conversational analytics, offered by some banks inside their cash management platforms, that categorise a company's cash flows, forecast positions across accounts and currencies, and answer treasurers' questions in plain language, so the treasury team decides on funding and idle balances with better information and less spreadsheet work.","aliases":["AI cash forecasting","cash flow intelligence","treasury analytics assistant","liquidity forecasting"],"industries":["banking","cross-industry","logistics-and-transportation","retail-and-ecommerce","manufacturing"],"functions":["treasury","finance-and-accounting","analytics-and-reporting"],"patterns":["prediction-and-scoring","classification-and-routing","conversational-agent","agentic-workflow"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"specialized-businesses","problem":"Where cash forecasting runs on spreadsheets, someone pulls bank reports and extracts, maps\ntransactions to categories by hand and rolls the result forward, cycle after cycle. The published\ncases describe exactly this. J.P. Morgan's case study on Prysmian says its two person North\nAmerican treasury team relied on spreadsheets for cash forecasting and daily reconciliation, and\nits case study on Domino's describes manual data entry and categorisation that consumed valuable\ntime. The work is slow, depends on a few people and can produce forecasts too coarse to act on,\nso companies hold buffers of idle cash, as Amtrak did before improving its projections.\n\nMuch of the data to do better already sits in the bank's systems, because the company's payments\nand receipts pass through them. The opportunity is to categorise those flows automatically, learn their\npatterns, and give the treasurer a forecast and a way to question it, while the decision to fund,\nsweep or invest stays with the treasurer.","problemStats":[],"howItWorks":"1. **Connect the data.** Account and transaction data from the bank's platform, and optionally\n   other banks and the ERP, flow in daily.\n2. **Categorise flows.** Models sort every transaction into the company's own categories (payroll,\n   suppliers, card receipts, taxes, intercompany), and the treasury team corrects the ones that\n   are wrong.\n3. **Forecast.** Per category, account and currency, the tool projects positions over a chosen\n   horizon and shows the forecast against actuals as they arrive.\n4. **Ask in plain language.** A conversational layer turns questions such as \"show balances by\n   account for the last three months\" into queries and charts over the same data. This is the\n   least mature step: J.P. Morgan describes its treasury analytics assistant as a prototype, while\n   Bank of America offers CashPro Chat, a virtual service advisor, in its CashPro platform.\n5. **Recommend within limits.** J.P. Morgan describes, as a future direction, GenAI that gives\n   treasurers recommendations and might one day act on their behalf within parameters they set.\n   A design option that follows from this, not a feature any cited bank has shipped, is to run\n   scenarios (a delayed receipt, a currency move) and propose sweeps or investment of idle\n   balances for the treasurer to approve.","valueDrivers":["employee-productivity","speed","risk-reduction","revenue-growth"],"kpis":["productivity-gain","cost-savings","hours-saved","users-served","forecast-accuracy"],"indicativeValue":{"referenceOrg":"A mid sized company with a treasury team of five","inputs":[{"key":"teamSize","label":"Treasury staff involved in forecasting","low":5,"high":5,"unit":"people","note":"The reference company."},{"key":"hoursPerWeek","label":"Hours per person per week on data preparation and forecasting","low":6,"high":12,"unit":"hours per person per week","note":"Editorial assumption, replace with your own time study."},{"key":"weeks","label":"Working weeks per year","low":46,"high":46,"unit":"weeks","note":"Editorial assumption."},{"key":"gain","label":"Share of that time saved","low":0.25,"high":0.4,"unit":"fraction of time","note":"Editorial assumption, applied only to the data preparation and forecasting hours above. Set below the single cases on this page, which J.P. Morgan reports as half of one Prysmian team member's manual forecasting and reconciliation time and Domino's weekly manual data cleanup down by up to 90%."},{"key":"hourlyCost","label":"Loaded cost of a treasury hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."},{"key":"releasedCash","label":"Idle balance released for investment thanks to better forecasts","low":1000000,"high":5000000,"unit":"USD","note":"Editorial assumption. Amtrak describes investing balances it had set aside once forecasts improved; size this from your own buffers."},{"key":"netYield","label":"Net yield on the released balance","low":0.02,"high":0.04,"unit":"fraction per year","note":"Editorial assumption, replace with your own short term investment yield."}],"formula":"teamSize * hoursPerWeek * weeks * gain * hourlyCost + releasedCash * netYield","currency":"USD","period":"per year","resultLabel":"Treasury time released plus yield on released idle cash","caveat":"Leaves out the cost of the tool, the value of avoided overdrafts or short term borrowing, and any fees the bank charges. The released cash figure depends heavily on how conservative the current buffers are."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The models are well understood; forecast quality depends on data coverage. Treasuries with many banks and entities need data from all of them, and categories must match how the company thinks about its cash.","dataPrerequisites":["Transaction history of at least a year across the main accounts","An agreed category scheme for inflows and outflows","Known large or irregular flows (tax dates, dividends, funding installments) entered as events"],"integrations":["Bank cash management platform or multibank data aggregation","ERP and treasury management system for payables, receivables and plans","Market data for foreign exchange and investment rates"]},"implementation":{"steps":[{"title":"Baseline the current forecast","detail":"Record how long the weekly forecast takes and how far it has been from actuals per category over the last quarters, so improvements can be measured."},{"title":"Agree categories with the treasury team","detail":"Define categories that match decisions (payroll, suppliers, receipts, taxes, intercompany), and separate large irregular flows so they do not distort the daily pattern, as Amtrak did."},{"title":"Run in parallel","detail":"Run the AI forecast next to the spreadsheet for several cycles and compare both with actuals before the team relies on it."},{"title":"Add the conversational layer on governed data","detail":"Let treasurers query the same governed data in plain language, with every answer showing the query and data it used."},{"title":"Keep actions behind approval","detail":"If the tool suggests sweeps or investments, route them as proposals with limits the treasurer sets, never as automatic instructions."}],"guardrails":["Forecasts are decision support; no payment, sweep or investment is executed without explicit approval","Forecast accuracy per category is shown next to the forecast, not hidden in a report","Conversational answers show the underlying query and data source","Drift monitoring on categorisation and forecast error, with an owner who acts on alerts"],"humanInTheLoop":"Treasurers review forecasts, correct categories and decide on funding, sweeps and investments. Any automated action runs only within limits the client has explicitly authorised, with a reversible audit trail. The bank's model owner monitors accuracy and drift across clients.","kpisToInstrument":["Hours per week spent on forecast preparation, before and after","Forecast error per category and horizon against actuals","Share of transactions categorised automatically without correction","Idle balances invested or buffers reduced as a result of better forecasts"],"failureModes":[{"title":"Garbage categories","detail":"Early miscategorisation trains the model on wrong labels. Review categories closely in the first cycles and lock the scheme once stable."},{"title":"Blind spots from missing banks","detail":"Flows at other banks or in cash pools are missing and the forecast looks precise but is wrong. Show coverage and warn when material accounts are absent."},{"title":"Over trust in a single number","detail":"The team treats a point forecast as certain. Show ranges and scenario results, and keep liquidity buffers under human policy."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Forecasting a company's cash flows is not listed in Annex III and makes no decision about a natural person, so the forecasting model itself carries no obligations beyond AI literacy (Article 4). The conversational layer interacts directly with treasury staff, so under Article 50(1) they must be informed that they are dealing with an AI system unless that is obvious from the context. Without a conversational layer the use case is minimal risk."},"regulations":["eu-ai-act","dora","us-sr-11-7","mas-ai-risk-management","nist-ai-rmf"],"guidance":[{"title":"NIST AI Risk Management Framework","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/itl/ai-risk-management-framework","note":"A practical structure to map, measure and manage the accuracy and drift risks of a forecasting model offered to clients."},{"title":"MAS Guidelines for Artificial Intelligence (AI) Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Consultation paper of November 2025 proposing supervisory expectations for AI inventories, risk materiality assessment, evaluation and testing, and monitoring at financial institutions."}],"controls":["Model inventory entry with an owner, validation results and drift monitoring","Documented limits for any automated sweep or investment, set by the client","Audit trail of forecasts, overrides and approved actions","Clear client terms that forecasts are informational and not advice"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the forecasting model itself stays in the bank's analytics stack; Blits.ai adds the\nconversational and agentic layer around it. An **AI agent** answers treasurers' questions by\nquerying a **SQL knowledge base** over the governed transaction and forecast tables, and\n**custom functions** call the forecasting and scenario APIs. Answers can come back as charts and\ninsight panels in the web widget, or through the **API channel** inside the bank's portal.\n\nAn **agentic workflow** can check positions on a schedule and prepare a sweep or investment\nproposal, with **human in the loop approval** above a threshold the treasurer sets. **Guardrails**\nkeep the agent to treasury topics, **execution tracing** records every query behind an answer, and\n**test suites** check answers on known questions after each change. Models are selectable per\nagent, and the platform runs in EU or UAE regions for data residency."},"faq":[{"question":"How much manual work does AI cash forecasting remove?","answer":"The published client cases are case studies by the bank that sells the tool and should be read as such. J.P. Morgan reports that Prysmian halved the manual forecasting and reconciliation work of one treasury team member (about 10 hours a week) and saved an estimated USD 100,000 a year, and that Domino's cut weekly manual data cleanup by up to 90%. According to a trade press report of a Bloomberg interview, about 2,500 corporate clients used the tool a year after launch."},{"question":"Does the AI move money on its own?","answer":"It should not by default. Forecasts are decision support, and any sweep or investment is a proposal the treasurer approves, or runs within limits the client has explicitly authorised."},{"question":"Is this only for large corporates?","answer":"The published cases are all large companies (Prysmian, Domino's, Amtrak), so there is no public evidence yet for smaller firms. The teams can be small, though: J.P. Morgan's case study describes Prysmian's North American treasury as a team of two. For a smaller company, the deciding factors are whether its bank offers such a tool and how much of its cash flows through that bank."}],"related":["corporate-client-servicing-assistant","ledger-and-payment-reconciliation","governed-text-to-sql-analytics","sme-cash-flow-underwriting","client-briefing-and-call-report-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version. The catalog's \"around 90 percent\" figure was not on the cited J.P. Morgan page; it is sourced here to a trade press report of a Bloomberg interview, alongside named client cases."},{"date":"2026-09-25","note":"Consolidation pass: industries now include logistics and transportation and retail and ecommerce and manufacturing, where its evidence comes from; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: EU AI Act tier corrected to limited because the conversational layer triggers Article 50(1); problem and how it works rewritten to cite the Prysmian, Domino's and Amtrak cases and to mark scenario suggestions as a future direction; FAQ now attributes figures to J.P. Morgan and trade press; MAS guidance noted as a November 2025 consultation; Prysmian saving marked as an estimate, Bank of America summary aligned with its press release, source dates added; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: step 5 now matches what J.P. Morgan actually says (recommendations, acting within parameters the treasurer sets) and presents scenarios and sweeps as a design option; the Domino's challenge is attributed to J.P. Morgan's case study; the FAQ on company size no longer claims evidence for smaller firms; the conversational layer is marked as a prototype at J.P. Morgan next to Bank of America's CashPro Chat; time saving range lowered to 25 to 40%; KPI accuracy replaced by forecast accuracy; Prysmian country added."},{"date":"2026-09-27","note":"Fact checked against sources again: all quotes, sources, the EU AI Act basis and the MAS and NIST guidance confirmed; the Prysmian 50% is now scoped to one treasury team member's manual forecasting and reconciliation time (about 10 hours a week) in its evidence record and in the FAQ."}],"slug":"treasury-cash-flow-forecasting","url":"https://www.blits.ai/ai-use-cases/treasury-cash-flow-forecasting","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":1,"median":70,"min":50,"max":90,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dominos-cash-flow-intelligence","pooled":false},{"id":"jpmorgan-cash-flow-intelligence","pooled":true},{"id":"prysmian-cash-flow-intelligence","pooled":true}]},{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":100000,"min":100000,"max":100000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"prysmian-cash-flow-intelligence","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2500,"min":2500,"max":2500,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"jpmorgan-cash-flow-intelligence","pooled":true}]}],"indicativeValueResult":{"low":40700,"high":310400},"evidence":["amtrak-cash-flow-intelligence","bank-of-america-cashpro-forecasting","dominos-cash-flow-intelligence","jpmorgan-cash-flow-intelligence","prysmian-cash-flow-intelligence"]},{"title":"AI cash flow underwriting for small business loans","shortTitle":"SME cash flow underwriting","seoTitle":"AI cash flow underwriting for SME lending","metaDescription":"Cash flow underwriting decides small business loans from bank and accounting data. MYbank approves in under a second; NAB opened 45% of SME loan accounts this way.","definition":"An underwriting engine that assesses a small business's repayment capacity from live bank transactions, point of sale and payment flows, receivables and accounting data instead of audited accounts, and returns a decision recommendation with the evidence and reasons behind it.","aliases":["small business cash flow lending","MSME credit decisioning","SME lending automation"],"industries":["banking"],"functions":["lending-and-credit","underwriting","risk-management"],"patterns":["prediction-and-scoring","document-processing","agentic-workflow","conversational-agent"],"channels":["api","mobile-app","web-chat","internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"lending","problem":"Many small businesses struggle to get bank credit. Traditional underwriting asks for audited\nfinancial statements, tax returns and collateral, which many micro and small firms do not have, and\nrelies on manual spreading and credit memos for loans that are small relative to the effort. The\nresult is a high cost to serve, slow decisions and owners who turn to more expensive finance or go\nwithout. At MYbank, which lends on its own data and models, over 72 percent of the 3 million\nborrowers it added in 2023 had never had a business loan from a bank before.\n\nMost of these businesses do have a detailed financial record: their bank account, card acquiring,\nmarketplace or accounting software. Cash flow underwriting reads that record directly. It can decide\nsimple, small facilities in minutes and give credit officers a much better picture for larger ones,\nbut only if the data connections, the model and the credit policy are designed together.","problemStats":[],"howItWorks":"1. **Connect the data.** With the owner's consent the engine pulls bank transactions, acquiring or\n   marketplace sales, and accounting data through APIs, or reads uploaded statements with document\n   AI.\n2. **Build the cash flow picture.** Transactions are categorised into revenue, payroll, suppliers,\n   taxes and existing debt service; seasonality, volatility and concentration are measured.\n3. **Score and size.** A model estimates default risk, and policy rules translate free cash flow\n   into an affordable limit and tenor.\n4. **Decide or refer.** Small, clean applications within policy are approved automatically; the\n   rest go to a credit officer with a prepared summary, the key ratios and the reasons.\n5. **Keep the owner informed.** Status updates and requests for missing documents go out on the\n   owner's channel of choice.\n6. **Keep watching.** The same data feeds monitor the borrower after the loan is drawn.","valueDrivers":["speed","inclusion-and-access","cost-to-serve","revenue-growth","risk-reduction"],"kpis":["automation-rate","processing-time-reduction","cycle-time-days","users-served"],"indicativeValue":{"referenceOrg":"A bank receiving 20,000 small business loan applications a year","inputs":[{"key":"applications","label":"Small business loan applications per year","low":20000,"high":20000,"unit":"applications per year","note":"The reference bank."},{"key":"automatedShare","label":"Share of applications decided through the automated cash flow path","low":0.2,"high":0.33,"unit":"fraction of applications","note":"iTnews reported that NAB's QuickBiz platform had been held up for deciding one in every three small business loans, and a NAB executive later said 45 percent of small business lending accounts were opened through it. Not every application on such a platform is decided without an underwriter, so the high bound stays at one in three. The low bound of 0.2 is an editorial floor, not a reported figure: replace it with your own measured automation rate.","sourceUrl":"https://www.itnews.com.au/news/nab-watches-cloud-based-quickbiz-lending-process-gain-traction-530744"},{"key":"hoursSaved","label":"Underwriter hours saved per application on that path","low":3,"high":6,"unit":"hours per application","note":"Editorial assumption for spreading, analysis and memo writing on a small facility. Replace with your own time study."},{"key":"hourlyCost","label":"Fully loaded cost of an underwriter hour","low":50,"high":80,"unit":"USD per hour","note":"Editorial assumption. Replace with your own cost."}],"formula":"applications * automatedShare * hoursSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Underwriting effort released","caveat":"Counts underwriting effort only. It leaves out additional lending volume from faster decisions, changes in credit losses, data access fees, and the cost of building and validating the model."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Data connections, categorisation quality and credit policy design carry most of the effort. Decisions are regulated in many markets, the model needs validation, and the automated path must be tightly bounded by product, size and risk grade.","dataPrerequisites":["Historical small business applications with repayment outcomes","Consent based access to bank transaction, acquiring or accounting data","A written credit policy for the automated path (eligible products, maximum amounts, exclusions)","Reason codes that credit and compliance have approved"],"integrations":["Open banking or account aggregation provider","Accounting software connectors (for example Xero, MYOB, QuickBooks)","Acquiring, marketplace or point of sale data where the bank has it","Loan origination system and decision engine","Credit bureau and business registry"]},"implementation":{"steps":[{"title":"Define the automated lane","detail":"Write down which products, amounts, industries and risk grades may be decided without a person. Start narrow, such as unsecured facilities up to a set limit for existing customers."},{"title":"Get the categorisation right","detail":"Test transaction categorisation on a few hundred real businesses per sector. Revenue, owner drawings and transfers between own accounts are where errors hide."},{"title":"Backtest and validate","detail":"Score past applicants and compare with outcomes, then take the model through independent validation and add it to the model inventory before any live decision."},{"title":"Prepare the referral pack","detail":"For cases outside the lane, generate a summary with cash flow charts, key ratios and the reasons for referral, so credit officers start from analysis instead of raw statements."},{"title":"Launch with existing customers","detail":"The bank already holds their transaction history, which removes the consent step and gives a cleaner first measurement of approval, speed and loss rates."},{"title":"Extend to connected new customers","detail":"Add accounting and acquiring connections for new to bank businesses once the lane performs as expected."}],"guardrails":["Automatic approvals only inside the documented lane; everything else goes to a credit officer","Specific, recorded reasons for every decline, reduced limit or referral","Consent recorded for every external data source used in a decision","Deterministic affordability and exposure limits outside the model","Monitoring that compares automated approvals with manually underwritten ones"],"humanInTheLoop":"Credit officers own every decision outside the automated lane, every appeal and every exception to policy. Credit risk reviews the performance of automated approvals each month and can close the lane for a segment at any time.","kpisToInstrument":["Share of applications decided in the automated lane","Median time from application to decision, by lane","Default and arrears rates of automated versus manual approvals","Consent and data connection completion rate","Decline reasons distribution and appeal overturn rate"],"failureModes":[{"title":"Misread cash flow","detail":"Transfers between the owner's own accounts or a one off asset sale look like revenue. Categorisation must be tested per sector and suspicious patterns flagged."},{"title":"A lane that creeps wider","detail":"Pressure to grow volume pushes larger or riskier loans into automatic decisions. Change the lane only through credit committee with fresh backtests."},{"title":"Declines without a real reason","detail":"A small business told only that it \"did not meet criteria\" cannot act and may complain. Give specific reasons tied to the data."},{"title":"Blind spots after drawdown","detail":"Underwriting uses live data but monitoring still waits for annual accounts. Connect the same feeds to early warning."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 5(b) makes AI systems that evaluate the creditworthiness of natural persons or establish their credit score high risk. Scoring a company is outside that point, but a sole trader is a natural person, and a model that also assesses the personal credit of owners, partners or guarantors evaluates natural persons. The tier therefore depends on who the borrower is and whose creditworthiness the model assesses."},"regulations":["eu-ai-act","gdpr","eba-loan-origination","us-sr-11-7","mas-ai-risk-management","dora","us-ecoa-reg-b"],"guidance":[{"title":"Guidelines on loan origination and monitoring","issuer":"European Banking Authority","region":"europe","url":"https://www.eba.europa.eu/regulation-and-policy/credit-risk/guidelines-on-loan-origination-and-monitoring","note":"Covers creditworthiness assessment for micro and small enterprises and the governance of automated models used in credit decisions."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) on creditworthiness of natural persons decides whether a given small business model is high risk."},{"title":"MAS Guidelines for Artificial Intelligence Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Consultation paper of November 2025 proposing supervisory expectations for financial institutions in Singapore on AI inventories, risk materiality, fairness, explainability and human oversight, which apply to credit models."}],"controls":["Model inventory entry, independent validation and an approved scope for automatic decisions","Written automated lane policy approved by credit committee","Consent and data lineage records for every decision","Reason code library reviewed by compliance","Monthly performance review of automated approvals against manual ones"],"incidents":[]},"blitsAi":{"howToBuild":"The scoring model and the decision engine stay in the bank's credit stack. Blits.ai carries the\napplication and the conversation around it: an **AI agent** on web chat, WhatsApp or, through the API\nchannel, the bank's own app guides the owner through eligibility, asks for consent to connect accounts, collects statements\nthrough **receive attachment** blocks and answers product questions from a **knowledge base** of\napproved terms. **Custom functions** call the bank's aggregation, origination and decision APIs,\nand the regulated steps (consent, declarations, acceptance) run as deterministic **flows**.\n\nFor referred cases an **agentic workflow** assembles the credit officer's pack from the decision\nengine output and the documents, and waits for **human in the loop approval** before anything is\nsent to the customer. **SQL knowledge bases** let credit staff ask questions about the pipeline in\nplain language, **PII masking** protects owner data at the gateway, and **test suites** replay\napplications on every change. EU and UAE hosting supports data residency."},"faq":[{"question":"How fast can a small business loan be decided with cash flow data?","answer":"For small, simple facilities, very fast. MYbank describes a loan that takes under three minutes to apply for and under one second to approve with no human involved, and a NAB executive said QuickBiz credit decisions often came the same day as the conversation with the banker. Larger facilities still need a credit officer, but with a prepared analysis instead of raw statements."},{"question":"Does cash flow underwriting replace financial statements?","answer":"For micro and small loans it often can, because transaction data shows revenue, costs and debt service more currently than annual accounts. For larger facilities it complements statements and supports continuous monitoring after drawdown, as OakNorth Bank does by comparing each borrower with peers in the same sector and location."},{"question":"Is SME credit scoring high risk under the EU AI Act?","answer":"Annex III point 5(b) covers creditworthiness of natural persons, so scoring a company is not listed. Sole traders are natural persons, and models that assess owners or guarantors personally also fall inside it, so classify each product by whose creditworthiness is assessed."}],"related":["alternative-data-credit-scoring","credit-memo-drafting-agent","credit-early-warning-monitoring","adverse-action-explanations","treasury-cash-flow-forecasting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added ECOA and Regulation B to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed an unsourced claim from the problem, lowered the automated share range to match NAB, corrected the NAB same day wording, made the EU AI Act basis precise, marked the MAS guidelines as a consultation, and added SEO title and description."}],"slug":"sme-cash-flow-underwriting","url":"https://www.blits.ai/ai-use-cases/sme-cash-flow-underwriting","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":53000000,"min":53000000,"max":53000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"mybank-310-sme-lending","pooled":true}]}],"indicativeValueResult":{"low":600000,"high":3168000},"evidence":["mybank-310-sme-lending","nab-quickbiz-automated-sme-lending","oaknorth-bank-continuous-credit-monitoring","sumitomo-mitsui-banking-corporation-oaknorth-credit-intelligence"]},{"title":"AI clinical trial patient matching and prescreening","shortTitle":"Clinical trial patient matching","seoTitle":"AI clinical trial patient matching","metaDescription":"AI checks records against trial criteria; staff confirm. Yale Cancer Center cut screening time per reviewed chart by 41%; Mount Sinai runs AI matching systemwide.","definition":"AI that reads structured data and clinical notes in the health record, compares each patient with the inclusion and exclusion criteria of open clinical trials, and gives research staff and treating clinicians a ranked list of likely eligible patients with the evidence for each criterion, so that people confirm eligibility and invite the patient.","aliases":["AI trial prescreening","clinical trial recruitment AI","patient to trial matching","eligibility screening with language models"],"industries":["healthcare","pharma-and-life-sciences"],"functions":["operations","analytics-and-reporting"],"patterns":["document-processing","classification-and-routing"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","problem":"Many clinical trials struggle to enroll enough patients, and slow enrollment extends trial timelines.\nEligibility criteria are long and specific (biomarkers, prior treatments, lab values, stage), and\nthe facts needed to check them are scattered across notes, pathology reports and lab results.\nResearch coordinators screen charts by hand, one trial and one patient at a time, so they see only a\nfraction of the patients who might qualify, and patients treated outside the flagship hospital are\nconsidered less often.\n\nThe result is lost opportunity on both sides: patients are not offered trials that could help them,\nsponsors wait longer for results, and trial participation stays concentrated at flagship academic sites.\nLanguage models can read clinical notes at scale, but eligibility errors in either direction matter,\nso the design has to keep people in charge of the final decision.","problemStats":[{"statement":"Researchers at Yale Cancer Center write that cancer clinical trial enrollment remains critically low at 5% to 7% of adult patients, despite exponential growth in the number of available trials.","sourceTitle":"Clinical Trial Patient Matching: A Real-Time, Common Data Model and Artificial Intelligence-Driven System for Semiautomated Patient Prescreening in Cancer Clinical Trials","sourceUrl":"https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:41512229%20AND%20SRC:MED&resultType=core&format=json","year":2026}],"howItWorks":"1. **Encode the criteria.** Each trial's inclusion and exclusion criteria are turned into checks,\n   some structured (age, diagnosis codes, lab values) and some that need the notes (prior lines of\n   therapy, performance status, biomarkers).\n2. **Find the population.** Rules on structured data narrow the whole patient population to\n   candidates, for example patients with a relevant diagnosis and an upcoming visit.\n3. **Read the record.** A language model or NLP pipeline reads notes and reports for each candidate\n   and marks each criterion as met, not met or unknown, with the passage that supports it.\n4. **Present a ranked list.** Research staff and treating clinicians see likely eligible patients and\n   the evidence, ideally before the patient's next visit.\n5. **Confirm and invite.** Staff verify eligibility in the record, discuss the trial with the treating\n   clinician and approach the patient; outcomes feed back to improve the criteria and the model.","valueDrivers":["inclusion-and-access","speed","employee-productivity"],"kpis":["handling-time-reduction","interactions-handled","accuracy","detection-rate-improvement"],"indicativeValue":{"referenceOrg":"A cancer center whose research staff prescreen 15,000 charts a year","inputs":[{"key":"charts","label":"Charts prescreened by hand per year","low":15000,"high":15000,"unit":"charts per year","note":"Editorial assumption for the reference cancer center, replace with your own."},{"key":"minutesPerChart","label":"Minutes of manual review per chart","low":3,"high":15,"unit":"minutes per chart","note":"Yale Cancer Center measured 3.1 minutes per chart for its prescreening workflow. Cleveland Clinic writes that a manual chart review can take more than 30 minutes per record, depending on the complexity of the criteria and the volume of history; the high value is an editorial assumption well below that. Replace with your own time data."},{"key":"reviewAvoided","label":"Share of manual chart review avoided","low":0.4,"high":0.8,"unit":"fraction of review minutes","note":"Conservative against Yale Cancer Center's report of a tenfold reduction in chart review workload and 41% less screening time per chart that was still reviewed."},{"key":"costPerHour","label":"Fully loaded cost per research coordinator hour","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption. Replace with your own."}],"formula":"charts * minutesPerChart / 60 * reviewAvoided * costPerHour","currency":"USD","period":"per year","resultLabel":"Research staff screening time released","caveat":"Screening effort only, and likely small next to the main value: more patients offered trials and faster enrollment, neither of which is included. It also leaves out the cost of encoding criteria, integrating the record and validating the tool."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Reading the record is feasible with current models; the effort goes into access to structured and unstructured data (often through a common data model such as OMOP), turning free text criteria into checks for each new trial, and fitting the output into how research teams already work.","dataPrerequisites":["Access to structured record data and clinical notes, pathology and lab reports","Trial protocols with inclusion and exclusion criteria, and a list of open trials","Visit schedules to time outreach before appointments","Past screening decisions to measure accuracy"],"integrations":["Electronic health record or clinical data warehouse (for example on the OMOP model)","Clinical trial management system for trials and enrollment status","Research staff worklists and secure messaging to treating clinicians"]},"implementation":{"steps":[{"title":"Start with trials that struggle to enroll","detail":"Pick a few open trials with clear criteria and slow accrual, and measure how many eligible patients the current process finds."},{"title":"Split criteria into structured and text checks","detail":"Use structured data to narrow the population cheaply and reserve language model reading for criteria that only the notes contain."},{"title":"Show evidence per criterion","detail":"For every criterion show met, not met or unknown and the passage behind it, so staff can verify in seconds rather than rereading the chart."},{"title":"Measure against manual screening","detail":"Compare accuracy, missed patients and time per chart with the manual process on the same trials, as Yale Cancer Center and Cleveland Clinic did, before scaling."},{"title":"Scale across trials and sites","detail":"Add trials and community sites, and check that patients at every site and in every group are identified at similar rates."}],"guardrails":["Eligibility is always confirmed by research staff or the investigator before a patient is approached","The treating clinician is involved before any patient contact","Evidence shown for every criterion, with unknowns marked rather than guessed","Access to records for prescreening limited to what research rules and local approvals allow","Identification rates monitored by site, sex, age and ethnicity to catch unequal access"],"humanInTheLoop":"Research coordinators and investigators decide who is eligible and who is approached, together with the treating clinician. The tool prioritizes and explains; it never enrolls or contacts a patient itself.","kpisToInstrument":["Eligible patients identified per trial per month, compared with manual screening","Screening minutes per chart and charts reviewed per enrollment","Accuracy of eligibility suggestions on a verified sample, including missed eligible patients","Enrollment and time to first patient per trial","Identification and enrollment rates by site and demographic group"],"failureModes":[{"title":"Missed eligible patients","detail":"Criteria encoded too strictly, or facts hidden in scanned documents, exclude patients who qualify. Measure sensitivity against manual screening, not only precision."},{"title":"Confident but wrong eligibility","detail":"The model marks a criterion as met from an outdated or negated note. Show the source passage and date and require verification."},{"title":"More candidates, no more enrollments","detail":"Lists grow but staff and clinicians have no time to act. Fit the output to visit schedules and worklists and track enrollments, not matches."},{"title":"Conflicts of interest in evaluation","detail":"Health systems that invest in the vendor they evaluate may overstate results. Look for independent or prospective evaluations."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Prescreening for research that staff verify is not listed in Annex III and is usually minimal risk. The Article 2(6) exclusion covers only systems developed and put into service for the sole purpose of scientific research and development, so an operational recruitment tool used across a health system usually falls inside the Act. If the software recommends trials to a clinician as a treatment option for an individual patient, it may qualify as medical device software under the Medical Device Regulation; where that needs a notified body assessment, it is high risk under Article 6(1). Processing health records for research falls under GDPR Article 9 and national research rules."},"regulations":["eu-ai-act","gdpr","hipaa","nist-ai-rmf"],"guidance":[{"title":"Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle","issuer":"European Medicines Agency","region":"europe","url":"https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle_en.pdf","note":"Says AI used in clinical trials should meet applicable ICH E6 good clinical practice requirements, and that where a use could have high regulatory impact or high patient risk and the method has not been previously qualified by the EMA for that context of use, the model documentation may be treated as clinical trial data and requested at marketing authorization, clinical trial application or GCP inspection. A reflection paper, not binding, and general to AI in clinical trials rather than specific to recruitment."},{"title":"Regulation (EU) No 536/2014 on clinical trials on medicinal products for human use","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2014/536/oj","note":"Sets the EU rules on trial conduct, informed consent and subject protection that recruitment processes supported by AI must respect."}],"controls":["Documented approval of prescreening under the institution's research governance and privacy rules","Versioned criteria per trial with an owner and review against protocol amendments","Audit trail of suggestions, staff decisions and patient contacts","Periodic accuracy and fairness review per trial and site"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai prescreening runs as a scheduled **agentic workflow**, for example each week for\npatients with upcoming visits. A clinical data warehouse on PostgreSQL is\nregistered as a **SQL knowledge base** so the agent can narrow the population with structured\nqueries (other databases are reached through custom functions), trial protocols sit in a\n**knowledge base** with hybrid retrieval, and **custom functions** fetch the notes and reports for\neach candidate. An **AI agent** with **structured output** marks each criterion as met, not met or\nunknown, with the supporting passage.\n\nResearch staff review the ranked list, and any outreach action the workflow takes requires\n**human in the loop approval**, so no patient is approached without a person deciding. **PII masking**\nat the gateway, with custom patterns per bot, covers traffic through the platform; the custom\nfunctions should fetch only the record fields each criterion needs. The **audit trail**\nrecords every run, **test suites** compare suggestions with verified screening decisions, and\n**monitors** run scheduled health checks on the screening agent and alert on failure. EU and UAE data residency keeps patient data in region."},"faq":[{"question":"Does AI trial matching increase enrollment?","answer":"The deployments on this page do not report enrollment with and without the tool. Yale Cancer Center's tool screened 98,348 patients across 29 trials since September 2022 and facilitated 117 enrollments, and Cleveland Clinic found 22 eligible patients for a rare disease trial in one week, where the usual process had prescreened nine in a year. Mount Sinai deployed matching systemwide in 2026 and has promised published results; until comparisons are published, treat an enrollment increase as something to measure, not to assume."},{"question":"How accurate is it?","answer":"Good enough to prioritize, not to decide. On one trial, Cleveland Clinic's research staff confirmed all 22 patients the system identified as eligible (100% positive predictive value), but missed eligible patients are not reported on the Cleveland Clinic page. On its validation trial Yale Cancer Center reports 94% retrospective and 88% prospective accuracy with 100% sensitivity. Both keep staff verification in the loop, so measure missed eligible patients as well as false matches."},{"question":"What should we watch out for?","answer":"Unequal identification across sites and groups, criteria that fall out of date after protocol amendments, and evaluations run by organizations with a financial interest in the vendor, as Cleveland Clinic discloses for its investment in Dyania Health."}],"related":["clinical-and-regulatory-document-drafting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with evidence from Yale Cancer Center, Mount Sinai and Cleveland Clinic, verified against the sources; editor passes tightened the FAQs (no claimed enrollment increase, enrollment answer limited to the evidence on this page, Cleveland Clinic sensitivity marked as not reported), problem claims, EU AI Act basis, EMA note (qualification condition and clinical trial application added), PII masking wording and Blits.ai build notes."}],"slug":"clinical-trial-patient-matching","url":"https://www.blits.ai/ai-use-cases/clinical-trial-patient-matching","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":49626,"min":904,"max":98348,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"yale-cancer-center-clinical-trial-patient-matching","pooled":true},{"id":"cleveland-clinic-ai-trial-prescreening","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":100,"min":100,"max":100,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"cleveland-clinic-ai-trial-prescreening","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":41,"min":41,"max":41,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"yale-cancer-center-clinical-trial-patient-matching","pooled":true}]}],"indicativeValueResult":{"low":12000,"high":210000},"evidence":["cleveland-clinic-ai-trial-prescreening","mount-sinai-oncology-trial-matching","yale-cancer-center-clinical-trial-patient-matching"]},{"title":"AI coding assistant for software developers","shortTitle":"Developer coding assistant","seoTitle":"AI coding assistant for enterprise developers","metaDescription":"AI coding assistants draft, test and review code under developer control. Bank of America reports efficiency gains over 20%, and ANZ and Accenture ran trials.","definition":"An AI assistant in the developer's IDE and code review flow that completes and generates code, explains unfamiliar modules, drafts unit tests and reviews pull requests for common defects, while generated code goes through the same review, testing and change controls as any other code.","aliases":["developer copilot","AI pair programmer","code assistant","AI code review"],"industries":["cross-industry","banking","capital-markets","technology","professional-services"],"functions":["it-and-engineering"],"patterns":["code-generation"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"mainstream","problem":"Large organizations run thousands of engineers, and a big share of their time goes to work that is\nnecessary but not differentiating: boilerplate, glue code, tests, reading code someone else wrote,\nand first pass review. Banks and insurers add a long tail of internal frameworks and legacy\nservices that new joiners have to learn before they are productive.\n\nCoding assistants take part of that load. The question for an engineering leader is no longer\nwhether developers will use them, but how to capture the gain safely: keeping proprietary code\nout of external training, stopping insecure or unlicensed code from reaching production, and\nmeasuring real delivery rather than lines of code accepted.","problemStats":[],"howItWorks":"1. **Inline completion and chat in the IDE.** The assistant suggests code as the developer types\n   and answers questions about the code base, using the open files and repository as context.\n2. **Tests and explanations.** Developers ask it to draft unit tests, explain a module or\n   propose a fix for a failing build.\n3. **Review assistance.** On a pull request it summarizes the change and flags likely defects,\n   which the human reviewer accepts or rejects.\n4. **Normal controls apply.** Generated code goes through peer review, static analysis, secret\n   and licence scanning, tests and change approval, exactly like human code.\n5. **Agentic tasks, carefully.** Newer tools can plan and apply multi file changes or run\n   commands. These should run in sandboxes with limited permissions, never directly against\n   production.","valueDrivers":["employee-productivity","speed"],"kpis":["productivity-gain","hours-saved","employee-adoption","users-served"],"indicativeValue":{"referenceOrg":"An engineering organization with 500 developers","inputs":[{"key":"developers","label":"Developers with the assistant","low":500,"high":500,"unit":"developers","note":"The reference organization."},{"key":"loadedCost","label":"Fully loaded cost per developer","low":100000,"high":150000,"unit":"USD per developer per year","note":"Editorial assumption. Replace with your own blended cost, including contractors."},{"key":"affectedShare","label":"Share of developer time spent on tasks the assistant helps with","low":0.3,"high":0.5,"unit":"fraction of working time","note":"Editorial assumption. Coding, tests and reading code, excluding meetings, design and incidents."},{"key":"timeSaved","label":"Time saved on those tasks","low":0.1,"high":0.2,"unit":"fraction of task time","note":"Editorial assumption, set below the Bank of America and ANZ figures on this page (Bank of America reports efficiency gains of over 20%; ANZ's controlled experiment measured about 42% less time on algorithmic Python challenges). Most CME Group developers using Gemini Code Assist report at least 10.5 hours a month, which falls inside this range. The only field randomized trial, Accenture's, measured a different thing (8.69% more pull requests), so replace this with your own control group result."}],"formula":"developers * loadedCost * affectedShare * timeSaved","currency":"USD","period":"per year","resultLabel":"Developer capacity released","caveat":"Released capacity, not cash: it only becomes value if the time goes into more delivery. It leaves out licence and review costs, the extra review effort generated code can create, and quality effects in either direction."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Rolling out a commercial assistant is technically simple. The effort is in the contract and data terms, security review, secure configuration of repositories and secrets, training, and a measurement plan that looks at delivery rather than acceptance rates.","dataPrerequisites":["Repository access rules and a list of repositories excluded from assistant context","Secure coding standards and approved libraries the assistant should follow","Baseline delivery metrics (cycle time, pull request throughput, change failure rate)"],"integrations":["IDEs and the source control platform","CI pipeline with static analysis, secret scanning and licence scanning","Identity provider for licence assignment and single sign on","Model gateway or vendor tenant with zero retention terms"]},"implementation":{"steps":[{"title":"Settle data terms first","detail":"Choose a tenant where your code is not retained or used for training, confirm where prompts are processed, and record the tool as a third party service in your risk register."},{"title":"Pilot with a control group","detail":"Give the tool to a representative set of teams and compare them with similar teams without it over several weeks, on delivery metrics rather than surveys alone."},{"title":"Strengthen the pipeline","detail":"Make secret scanning, dependency and licence checks and static analysis mandatory gates before merge, since more code will arrive faster."},{"title":"Train reviewers as well as authors","detail":"Teach engineers to treat suggestions as untrusted input, to check generated tests actually test something, and to reject code they do not understand."},{"title":"Scale and keep measuring","detail":"Roll out by team, track adoption and delivery metrics per cohort, and revisit settings when new agentic features arrive."}],"guardrails":["Zero retention and no training on the organization's code, confirmed contractually","Generated code passes the same review, testing and change approval as human code","Mandatory secret, dependency, licence and static analysis scanning before merge","Agentic features run in sandboxes without production credentials","Sensitive repositories excluded from assistant context where required"],"humanInTheLoop":"A developer accepts or rejects every suggestion and remains the author of record. A second engineer reviews every change before merge, and release managers approve production changes as before. Agent generated multi file changes are reviewed like a new colleague's first pull request.","kpisToInstrument":["Pull request throughput and lead time per team, compared with a control group","Change failure rate and escaped defects","Security findings per thousand lines in generated versus human code","Weekly active users among licensed developers","Developer satisfaction, surveyed quarterly"],"failureModes":[{"title":"Measuring the wrong thing","detail":"Acceptance rates and lines generated rise while delivery does not. Measure throughput, lead time and quality against a control group."},{"title":"Faster insecure code","detail":"Suggestions reproduce insecure patterns or hard coded secrets. Scanning gates and reviewer training catch them; the tool alone does not."},{"title":"Source code leakage","detail":"Engineers paste proprietary code into public chat tools when the approved tool is weak. Provide a good approved tool and block the alternatives."},{"title":"Agents with too much reach","detail":"An agent with shell or database access acts outside its task. Limit permissions and never give it production credentials."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"A coding assistant used by developers is not a prohibited practice under Article 5 and is not listed in Annex III. Developers know they are working with an AI tool, so the Article 50 disclosure duty has no practical effect for the deploying organization, and the marking of generated content under Article 50(2) falls on the tool's provider. What remains is AI literacy (Article 4). Using an AI system to monitor or evaluate individual developers' performance would fall under Annex III point 4(b), and the software the assistant helps build may itself fall under the Act."},"regulations":["eu-ai-act","dora","iso-42001","nist-ai-rmf","apra-cps-230"],"guidance":[{"title":"SP 800-218, Secure Software Development Framework (SSDF) Version 1.1","issuer":"NIST","region":"north-america","url":"https://csrc.nist.gov/pubs/sp/800/218/final","note":"Baseline secure software development practices, such as code review, testing and vulnerability response, that apply to generated code as much as to code written by hand."},{"title":"Guidelines for secure AI system development","issuer":"UK National Cyber Security Centre","region":"europe","url":"https://www.ncsc.gov.uk/collection/guidelines-secure-ai-system-development","note":"Guidelines for providers of AI systems in four areas (secure design, development, deployment, and operation and maintenance), relevant when you build your own tooling or agents around the assistant."},{"title":"Article 4, AI literacy","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/4/","note":"Providers and deployers must take measures on the AI literacy of staff who use AI systems (the amended wording shown on this page asks them to support it rather than ensure a sufficient level). Here that means training developers on the tool's limits."}],"controls":["Third party risk assessment and register entry for the assistant vendor","Contractual zero retention and data location terms","Mandatory scanning gates in the CI pipeline","Developer training on reviewing generated code","Periodic review of agent permissions and repository exclusions"],"incidents":[{"title":"Incident 768: ChatGPT reportedly implicated in Samsung data leak of source code and meeting notes","url":"https://incidentdatabase.ai/cite/768/","note":"Samsung engineers were reported to have leaked source code and internal meeting notes in March 2023 by entering them into ChatGPT for help with their work. It shows why an approved tool with proper data terms matters."},{"title":"Incident 1152: LLM-Driven Replit Agent Reportedly Executed Unauthorized Destructive Commands During Code Freeze, Leading to Loss of Production Data","url":"https://incidentdatabase.ai/cite/1152/","note":"An agentic coding tool with access to a live production database reportedly deleted it during a code freeze, despite instructions not to change code without permission. Agentic features need sandboxes and least privilege."}]},"blitsAi":{"howToBuild":"Blits.ai is not an IDE coding assistant, and a dedicated tool is the right choice for inline\ncompletion and code review. Where Blits.ai fits is the engineering knowledge around the code: an\n**AI agent** grounded in a **knowledge base** of internal standards, architecture decisions,\nrunbooks and API documentation, with hybrid retrieval, available to engineers in **Microsoft\nTeams** or **Slack**. The integration catalog includes GitHub, Jira and Confluence.\n\nFor repeatable engineering tasks, such as drafting release notes or checking a change against\ninternal standards, **agentic workflows** run with **human in the loop approval** and a\n**tool execution policy**, and every run keeps an audit trail. **PII masking** applies at the\ngateway before text reaches a model, input and output **guardrails** check content, **model\nagnostic** routing lets the organization choose\nwhich model sees its code, and EU or UAE data residency is available."},"faq":[{"question":"How much faster do developers get with a coding assistant?","answer":"Published results vary widely with how they are measured. Bank of America reports efficiency gains of over 20% for its developers, Accenture's randomized trial with GitHub saw 8.69% more pull requests, and ANZ measured about 42% less time on algorithmic Python challenges. Set exercises tend to overstate everyday gains, so measure your own teams against a control group."},{"question":"Does our code train the vendor's model?","answer":"It depends on the vendor, the plan and the contract, and terms differ between them. Confirm in writing that your code is not used for training and how long prompts are retained, check where prompts are processed, and treat the tool as a material third party service where your regulator expects that."},{"question":"Is generated code a regulatory problem for a bank?","answer":"Not in itself. Rules such as DORA require ICT change management, testing and security controls, and these apply whoever or whatever wrote the code. The risk is the pipeline receiving more code than it can review, so strengthen scanning and review before scaling."}],"related":["legacy-code-modernization","developer-api-integration-assistant","aiops-incident-triage","governed-text-to-sql-analytics","enterprise-knowledge-search"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog as an industry neutral page, with five evidence records verified against their sources. The catalog's Citizens Bank figure was not found in the cited article and was dropped."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources. Added seoTitle and metaDescription. Citigroup evidence moved to Citi's own earnings call transcript (grade B); ANZ stage corrected to production and its metric period clarified; Bank of America no longer listed as in house; EU AI Act basis expanded (Article 5, Article 50, Annex III point 4(b)); NIST guidance replaced with SP 800-218 because SP 800-218A covers AI model development; guidance, incident and FAQ wording aligned with what the sources say."},{"date":"2026-09-27","note":"Second fact check. Removed the MAS AI risk management guidelines from the regulations, since MAS has only consulted on them. Citigroup's 30,000 developers are now recorded as access to the tools rather than a users served metric. Accenture stage corrected to production. Incident 1152 given its full title and framed as reported, ANZ figure rounded to about 42% on the page, the time saved assumption note and the guardrail wording in the Blits.ai section made precise."},{"date":"2026-09-27","note":"Third review pass: CME Group evidence record redated to 2025, since Wayback shows the entry was added in the April 2025 update, not present in January 2025; corrected the time saved assumption note so it only claims to sit below the Bank of America and ANZ figures, with the CME Group figure noted as falling inside the range."}],"slug":"developer-coding-assistant","url":"https://www.blits.ai/ai-use-cases/developer-coding-assistant","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":3,"nUpTo":0,"median":20,"min":8.69,"max":42.36,"byClaimant":{"organization":2,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"anz-github-copilot-study","pooled":true},{"id":"bank-of-america-coding-assistant","pooled":true},{"id":"accenture-github-copilot-trial","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10.5,"min":10.5,"max":10.5,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"cme-group-gemini-code-assist","pooled":true}]}],"indicativeValueResult":{"low":1500000,"high":7500000},"evidence":["accenture-github-copilot-trial","anz-github-copilot-study","bank-of-america-coding-assistant","citigroup-developer-coding-tools","cme-group-gemini-code-assist","meta-testgen-llm-unit-tests"]},{"title":"AI command center for hospital bed and staff capacity planning","shortTitle":"Hospital capacity command center","seoTitle":"Hospital bed capacity command center AI","metaDescription":"An AI command center predicts patient flow and bed assignments. Hopkins raised bed use from 85% to about 94%; Humber cut the average ED wait for a bed by 34%.","definition":"An AI powered operations center that predicts patient admissions, discharges and transfers across a hospital or health system, and helps a team of coordinators sitting in one room sequence real time bed assignments, staffing levels and patient moves, so patients get into the right bed faster and existing capacity is used fully without adding beds.","aliases":["hospital command center","capacity command center","patient flow AI","hospital operations center"],"industries":["healthcare"],"functions":["operations","analytics-and-reporting"],"patterns":["prediction-and-scoring","classification-and-routing","anomaly-detection"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"hospital operations","problem":"Johns Hopkins Hospital, like many hospitals, faced backlogs at 85% bed utilization before it\nbuilt its command center, and small delays cascade into big ones. A patient\nwaits on a gurney in the emergency department because no ward bed is ready; an operating room holds\na finished case because there is nowhere to send the patient; a referring physician sends a\ntransfer request elsewhere because nobody can say quickly whether a bed exists. Before a command\ncenter, the staff who track admissions, discharges, transport and cleaning are scattered across the\nbuilding, working from pen and paper, whiteboards and markers.\n\nThe result is that decisions can lag reality by hours: a bed that becomes available at 9am, for\nexample, might not appear on anyone's list until the afternoon huddle. Humber River Hospital's own\nstaff put the case for building a command center plainly: \"we soon realized we were going to\nexceed our new capacity by 2020\" and set out to improve capacity \"without seeking government\nfunding to expand a hospital we had just barely opened.\" Johns Hopkins Medicine reports having\n\"essentially opened 16 beds on a daily basis\" without building a new wing or adding new staff,\naccording to Jim Scheulen, its chief administrative officer for emergency medicine and capacity\nmanagement.","problemStats":[],"howItWorks":"1. **Bring the data into one model.** Live feeds from the electronic health record's admission,\n   discharge and transfer stream, the bed management system, the operating room schedule and\n   ambulance dispatch are combined into a single, continuously updated picture of the hospital.\n2. **Predict.** Machine learning models forecast occupancy by unit for the next shift, day and week,\n   and estimate which patients are likely to be discharged soon.\n3. **Prioritize and recommend.** The system flags situations that need attention, such as an\n   emergency department patient waiting past target or a unit nearing capacity, and suggests the\n   next action: which bed to assign, which porter to send, which transfer request to accept.\n4. **Coordinate in one room.** Staff from admitting, patient transport, environmental services and\n   the referral line sit together, watch the same shared displays, and act on the recommendations as\n   they appear, instead of each working from a different, stale version of the truth.\n5. **Learn and rebalance.** Actual outcomes feed back into the forecasting models, and alert\n   thresholds are retuned as the hospital's patient mix, seasonal demand and physical capacity\n   change.","valueDrivers":["cost-to-serve","speed","employee-productivity","customer-experience"],"kpis":["processing-time-reduction"],"indicativeValue":{"referenceOrg":"A 600 bed academic hospital","inputs":[{"key":"beds","label":"Licensed beds","low":600,"high":600,"unit":"beds","note":"The reference hospital."},{"key":"freedBedShare","label":"Share of licensed beds freed through better patient flow","low":0.01,"high":0.014,"unit":"fraction of beds","note":"The high end matches Johns Hopkins' own reported benchmark after five years of tuning: the equivalent of 16 beds a day, about 1.4% of its 1,162 licensed beds. Humber River Health's own site reports a larger first year result, the equivalent of 35 beds against its roughly 688 bed footprint, about 5%, but that single first year figure is not used as the cap here since it is not yet a multi year benchmark. The low end is conservative against both."},{"key":"costPerBedDay","label":"Fully loaded cost avoided per bed day of freed capacity","low":1500,"high":2500,"unit":"USD per bed day","note":"Editorial assumption for a US academic hospital's marginal cost of a staffed bed day; replace with your own."},{"key":"daysPerYear","label":"Days per year","low":365,"high":365,"unit":"days","note":"Calendar year."}],"formula":"beds * freedBedShare * costPerBedDay * daysPerYear","currency":"USD","period":"per year","resultLabel":"Annual value of capacity freed without adding beds","caveat":"Gross capacity value only. It leaves out the cost of building and staffing the command center itself, the software licence, and any change in case mix or payer rate that comes with treating more patients in the freed capacity."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The technology is rarely the hard part. Integrating live feeds from the EHR, bed management, OR scheduling and ambulance dispatch into one data model, and getting departments that have never shared a dashboard to work from the same numbers in one room, is a multi year operating model change, not a software rollout.","dataPrerequisites":["Real time admission, discharge and transfer feed from the electronic health record","Bed and unit status from the bed management and environmental services systems","Operating room schedule and case status","Ambulance dispatch and inter hospital transfer request data"],"integrations":["Electronic health record (admission, discharge, transfer feed)","Bed management and environmental services systems","Operating room scheduling system","Ambulance dispatch and inter hospital transfer systems"]},"implementation":{"steps":[{"title":"Time the current patient journey before buying anything","detail":"Measure how long it actually takes today from an admit decision to a bed assignment, and from a finished OR case to a transfer, by unit and shift. This baseline is what proves the value later and tells you which bottleneck to attack first."},{"title":"Build one shared data model","detail":"Connect the EHR's ADT feed, the bed board, the OR schedule and transport systems into a single live view before adding any predictive model on top of it."},{"title":"Predict, then prioritize, one alert at a time","detail":"Start with a next shift occupancy forecast and a single at risk alert (for example, an emergency department patient waiting past target), prove it changes behaviour, then add more."},{"title":"Put every department in one room, physically or virtually","detail":"Admitting, transport, environmental services and the referral line need to see the same numbers at the same time and be empowered to act on them without escalating every decision."},{"title":"Set targets and instrument every one before scaling to more units","detail":"Agree a target for each metric (time to bed assignment, transfer acceptance rate, discharge before noon) with the unit that owns it, and only widen to more units once the first one holds."}],"guardrails":["Every bed assignment and transfer decision stays with a named clinical or administrative owner; the AI recommends, it does not assign","Escalation rules for clinically urgent transfers (stroke, trauma) bypass queue based recommendations and go straight to the relevant team","Forecast accuracy is checked against actual admissions and discharges on a regular schedule, and alert thresholds are retuned when it drifts"],"humanInTheLoop":"Coordinators in the command center act on every recommendation; nothing moves a patient, assigns a bed or accepts a transfer without a person confirming it. Unit and department leaders review forecast accuracy and override patterns regularly, and any new alert type or automated recommendation is approved before it goes live.","kpisToInstrument":["Time from admit decision to bed assignment, by unit and shift","Transfer delay from the operating room after a procedure","Ambulance and inter hospital transfer acceptance rate and decline reasons","Forecast accuracy against actual admissions and discharges"],"failureModes":[{"title":"Optimizing the room, not the ward","detail":"Staff in the command center chase a dashboard metric that looks good centrally but does not reflect what a specific unit is experiencing. Review metrics with the units that own the work, not only centrally."},{"title":"Alert fatigue","detail":"Too many predictive alerts, or alerts that are frequently wrong, and staff start ignoring all of them, including the ones that matter. Track false alarm rate per alert type and retire or retune alerts that staff routinely dismiss."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The tier depends on what the system is scoped to do. A design limited to occupancy and discharge forecasting and to sequencing bed assignments for patients already admitted is operational decision support for hospital logistics, outside Annex III. Annex III point 5(d) covers AI used \"to dispatch, or to establish priority in the dispatching of, emergency first response services\", including medical aid and emergency healthcare patient triage systems. On a plain reading, that point can apply when a system dispatches, or sets the priority of dispatching, ambulance or critical care transport itself (work similar to what the Johns Hopkins center's Lifeline transport staff do for helicopter and ambulance transfers), or when it assesses the clinical urgency of an emergency patient, that is, triage. Sequencing which already admitted ED patient gets the next ward bed, and deciding whether to accept an inter hospital transfer request on capacity grounds, are not listed activities under 5(d) as written; whether either counts as dispatching or triage in a given deployment is a case by case legal question, not a settled fact, and should be assessed with counsel before relying on this tier. For public hospitals, Annex III point 5(a) (access to essential public services, including healthcare) can also be relevant. Scoping the system to bed sequencing and transfer acceptance only, and keeping every ambulance dispatch and ED triage decision with clinical staff outside the AI's recommendation, is what keeps a deployment in the lower tier."},"regulations":["eu-ai-act","gdpr","hipaa","nist-ai-rmf","iso-42001"],"guidance":[{"title":"Article 6: Classification Rules for High-Risk AI Systems","issuer":"Future of Life Institute","region":"europe","url":"https://artificialintelligenceact.eu/article/6/","note":"Paragraph 1a addresses one route to high risk status, an AI system that is itself a safety component of a regulated product (the Annex I route): it says a system used solely for non safety related user assistance, performance optimization, service efficiency or convenience does not qualify as such a safety component. It does not decide whether a system falls under an Annex III listed use case, which is the separate route assessed above for point 5(d). A logistics and staffing tool for hospital operations is unlikely to be a product safety component either way, but the Annex III analysis above is the one that matters here."},{"title":"AI Risk Management Framework","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/itl/ai-risk-management-framework","note":"A framework hospitals can use to map, measure and manage the risk of predictive capacity tools, including the risk that a forecast gets treated as a decision rather than a recommendation."}],"controls":["Named accountable owner for every bed assignment and transfer decision; the AI recommends only","Escalation path for clinically urgent cases that bypasses queue based recommendations","Regular comparison of forecast accuracy against actual admissions, discharges and transfers","Access logging and data governance for the shared ADT and EHR feed powering the command center"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai, the shared operating picture is built with an **agentic workflow** whose **custom\nfunctions** call the hospital's EHR, bed management and OR scheduling systems as REST or SQL\nintegrations, combined with a **SQL knowledge base** (PostgreSQL or SQLite) that lets the\ncoordination team query occupancy, transfer and staffing data in plain language instead of a\nfixed dashboard; a database that is not already PostgreSQL or SQLite needs a replica in one of\nthose engines first. An **AI\nagent** turns each system's status into a plain language brief and a recommended next action for\nthe shift, while a **flow** with deterministic steps encodes the escalation rules for clinically\nurgent transfers, so time critical cases are never left to a recommendation alone.\n\nThe workflow's **human in the loop approval** step, when configured to require it, keeps every\nbed and transfer decision with a named owner: nothing writes back to the hospital's systems\nwithout a person confirming it. The\noccupancy and discharge forecasts themselves come from the hospital's own or a third party's\nforecasting models, called through the custom functions, not from Blits.ai; **monitors** then run\nscheduled recurring health checks on the agent and workflow themselves, with an alert if a check\nfails or recovers, and **analytics** give the team a per bot dashboard of interactions, recognition\nrate and satisfaction for the coordination team's own use of the agent. The platform is **model\nagnostic**, so the model behind the agent's briefings can change without rebuilding the\nintegrations, and **EU data residency** keeps patient flow data in region for European health\nsystems."},"faq":[{"question":"What is a hospital capacity command center?","answer":"A control room, physical or virtual, where staff from admitting, transport, environmental services and referral intake work from one live, AI predicted view of every bed, patient and transfer in the hospital, instead of tracking patient flow by phone and whiteboard from separate departments."},{"question":"How much capacity can a command center free without adding beds?","answer":"Johns Hopkins Medicine reports opening the equivalent of 16 beds a day and improving bed utilization from 85% to about 94%, a figure reported after five years of tuning. Humber River Health reports a larger result in its first year alone, the equivalent of 35 beds. A first deployment can move faster than expected, as Humber's did, but plan the first year target from the mature, multi year benchmark rather than assume a first year result as large as Humber's."},{"question":"Does the AI decide which patient gets a bed?","answer":"No. The system forecasts occupancy and recommends an assignment or action; a person in the command center makes and confirms every bed assignment and transfer decision."},{"question":"Is this the same as clinical triage software?","answer":"No. A capacity command center manages beds, staff and patient flow across the hospital. Software that reads a medical image to flag an urgent clinical finding, such as an AI radiology worklist triage tool, is a different, separately regulated category of AI."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by the editor after the Opus targeted check's remaining text fixes."},{"date":"2026-09-28","note":"First version, researched from Johns Hopkins Medicine's and Humber River Health's own primary sources."},{"date":"2026-09-28","note":"Editorial review: sourced the GE Healthcare vendor and the 2017 launch year for Humber from a Humber River Hospital press release; corrected the Humber bed count and freed capacity share to sourced figures; capped indicativeValue.freedBedShare at the Hopkins mature benchmark and removed the unsupported claim that a first deployment captures less; rewrote the FAQ on Humber's first year result versus Hopkins' five year one; changed euAiAct so the tier depends on the design, for Annex III point 5(d), emergency dispatch and triage; corrected the Blits.ai section against FEATURE_INVENTORY.md; replaced the mismapped Humber metric about hospital harm with a metric about bed cleaning time on the handling time reduction KPI; fixed the Johns Hopkins evidence summary against its source; removed an unsourced claim from the problem section."},{"date":"2026-09-28","note":"Second editorial review, against an adversarial checklist: rewrote the problem section's Hopkins claim to the exact quote (attributed to Jim Scheulen) instead of a stronger paraphrase; rewrote metaDescription so \"ED wait\" states the measured metric (waits for an inpatient bed); removed the Johns Hopkins 83% operating room metric, which described fewer delayed transfers rather than a shorter transfer time, and moved the finding to the evidence summary as a qualitative note; narrowed and hedged the euAiAct basis so only ambulance dispatch and clinical triage are said to plausibly fall under Annex III 5(d), reworded the Lifeline reference so it no longer implies Hopkins' own deployment is high risk, and separated the Article 6 paragraph 1a note (the Annex I safety component route) from the Annex III analysis; corrected the Johns Hopkins evidence summary and verification note (removed the uncited Becker's Hospital Review comparison and the unverifiable \"undated section\" claim, and fixed which figures are actually used where); corrected the Humber evidence note's description of humbercommandcentre.ca and softened \"Generation 3\" from completed to under way, matching the 2022 source; softened the unsourced 9am example in the problem section; noted the SQL knowledge base's PostgreSQL/SQLite limits and made human in the loop approval conditional on configuration in the Blits.ai section."},{"date":"2026-09-28","note":"Third editorial fix pass after adversarial review: restored the \"about 94%\" qualifier and changed \"inpatient bed\" to \"a bed\" in the metaDescription to match the hrh.ca quote, which does not say whether the bed is an ED or inpatient bed; remapped the Humber 45% bed cleaning metric from the handling time reduction KPI to the cycle time reduction KPI, since the quote credits bed planning for a shorter elapsed turnaround rather than hands on cleaning effort, and removed the handling time reduction KPI from the use case's kpis list since no remaining metric supports it; and rewrote the problem section's opening to tie the \"hospitals run close to full\" statement and the pen and paper description to Johns Hopkins' own reported baseline, and replaced the unsourced line about building capacity being slow and expensive with Humber River Hospital's own quoted reason for building its command center."}],"slug":"hospital-bed-and-staff-capacity-command-center","url":"https://www.blits.ai/ai-use-cases/hospital-bed-and-staff-capacity-command-center","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":36,"min":34,"max":38,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"johns-hopkins-capacity-command-center","pooled":true},{"id":"humber-river-health-command-centre","pooled":true}]}],"indicativeValueResult":{"low":3285000,"high":7665000},"evidence":["humber-river-health-command-centre","johns-hopkins-capacity-command-center"]},{"title":"AI copilot for corporate client briefings and call reports","shortTitle":"Client briefing and call reports","seoTitle":"Relationship manager AI copilot for call reports","metaDescription":"AI builds client briefing packs and drafts call reports that bankers approve. Evidence: Bank of America's meeting prep tool, Standard Chartered's AI client insights.","definition":"An AI copilot for relationship managers, mainly in corporate and commercial banking, whose main job is preparation: before a client meeting it assembles a briefing pack from filings, news, internal notes, product holdings and upcoming maturities, and afterwards it turns the banker's notes into a structured call report and CRM update. Unlike a meeting notetaker, which centres on capturing the conversation, it centres on the credit and cross sell context around the meeting; wealth advisor tools that also prepare meetings overlap with it. The banker reviews every output.","aliases":["relationship manager copilot","meeting preparation assistant","call report automation","banker briefing pack generator"],"industries":["banking","wealth-and-asset-management","capital-markets"],"functions":["sales","knowledge-management"],"patterns":["rag-knowledge-assistant","summarization","content-generation","agentic-workflow"],"channels":["internal-tools","microsoft-teams","email"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"specialized-businesses","problem":"Preparing for a corporate client meeting means assembling information from many places. The\nrelationship manager pulls the latest financials and filings, scans news about the client and its\nsector, checks product holdings, limits and upcoming maturities in several systems, and looks for\nshare of wallet gaps. After the meeting the notes have to become a call report and a CRM update,\nwhich can happen late, briefly or not at all.\n\nWhen this work is manual, preparation depends on how much time each banker has, CRM data stays\nthin, and cross sell opportunities are missed because the information sits in different systems.\nA copilot takes over the gathering and first drafting across those sources, cites where every fact\ncame from, and leaves the judgement and the client conversation with the banker.","problemStats":[],"howItWorks":"1. **Trigger from the calendar or CRM.** A scheduled client meeting, or a banker's request, starts\n   the preparation a day or two ahead.\n2. **Gather from approved sources.** The copilot retrieves internal notes, product holdings,\n   exposures and maturities from the CRM and core systems, and public filings and news from\n   licensed or approved external sources.\n3. **Draft the briefing pack.** It writes a short pack: client snapshot, recent events, open\n   items, maturities and renewals, possible needs, and a suggested agenda, with a source link on\n   every fact.\n4. **Capture the meeting.** With the client's consent, or from the banker's own notes or\n   dictation, it produces a structured summary of decisions and next steps.\n5. **Draft the call report and CRM update.** It fills the call report template and proposes CRM\n   updates and follow up tasks; the banker edits and approves before anything is saved.","valueDrivers":["employee-productivity","revenue-growth","customer-experience"],"kpis":["time-saved-per-task","employee-adoption","users-served","hours-saved","productivity-gain"],"indicativeValue":{"referenceOrg":"A commercial bank with 300 relationship managers","inputs":[{"key":"bankers","label":"Relationship managers using the copilot","low":300,"high":300,"unit":"bankers","note":"The reference bank."},{"key":"meetingsPerBanker","label":"Client meetings per banker per year","low":150,"high":250,"unit":"meetings per banker per year","note":"Editorial assumption, replace with your own CRM activity data."},{"key":"minutesSaved","label":"Preparation and call report time saved per meeting","low":30,"high":60,"unit":"minutes per meeting","note":"Editorial assumption. Deliberately far below the up to four hours per meeting that Bank of America says its meeting tool can save, because that figure is a stated potential, not a measured result."},{"key":"hourlyCost","label":"Loaded cost of a relationship manager hour","low":80,"high":120,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."}],"formula":"bankers * meetingsPerBanker * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Relationship manager time released, valued at loaded cost","caveat":"Values released time, not revenue. It leaves out the cost of the copilot and its data feeds, the revenue effect of better prepared meetings, and the gain from more complete CRM data, which is often the larger benefit but hard to measure."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The drafting is the easy part. The work is in reaching CRM, exposure and product data through APIs, licensing news and filings content for AI use, and respecting information barriers between client teams.","dataPrerequisites":["CRM with client hierarchy, contacts, notes and pipeline reachable through an API","Product holdings, limits, exposures and maturities per client group","Licensed or approved news and filings sources","A call report template and the bank's rules on what must be recorded"],"integrations":["CRM (read and write, with banker approval)","Core banking, lending and treasury systems for holdings and maturities","News, filings and market data providers","Calendar and email or collaboration suite","Meeting transcription, where client consent is obtained"]},"implementation":{"steps":[{"title":"Start with one segment and one meeting type","detail":"Pick annual review meetings in one commercial segment, where the pack content is predictable, and agree with a group of bankers what a good briefing looks like."},{"title":"Map sources and entitlements","detail":"List every source the pack draws on, who may see it, and which information barriers apply. The copilot inherits the banker's entitlements; it never widens them."},{"title":"Make every fact traceable","detail":"Require a source link on every figure and statement in the pack, and let the copilot say \"not found\" rather than fill gaps from general model knowledge."},{"title":"Draft call reports, do not file them","detail":"Generate the call report and CRM changes as drafts in the banker's queue; saving requires an explicit approval, so accountability for the record stays with the banker."},{"title":"Measure time and quality together","detail":"Track preparation time, call report completeness and timeliness, and banker ratings of the packs, and review a sample of packs each month for errors."}],"guardrails":["The copilot uses only the requesting banker's entitlements and respects information barriers","Every fact in a briefing carries a link to its source; unsupported statements are removed","External news and documents are treated as data, never as instructions, to resist prompt injection","No call report or CRM update is saved without the banker's approval","Meeting recording and transcription only with client consent, and with retention rules applied"],"humanInTheLoop":"The banker reviews every briefing and approves every call report and CRM update. Advice and any product recommendation to the client remain the banker's responsibility. A team lead samples packs and call reports each month for accuracy and completeness.","kpisToInstrument":["Preparation time per meeting, from a time study before and after","Share of meetings with a call report filed within 48 hours","Banker adoption, weekly active users against licensed users","Error rate found in monthly sampling of briefings","Follow up tasks created and completed after meetings"],"failureModes":[{"title":"Confident but stale briefings","detail":"The pack repeats an old exposure or a superseded news item. Show the as of date of every source and refresh data on the day of the meeting."},{"title":"Leakage across information barriers","detail":"Content from a deal team or another client group appears in a pack. Enforce entitlements at retrieval time and test barrier cases in the regression set."},{"title":"Rubber stamped call reports","detail":"Bankers approve drafts without reading them and errors enter the CRM. Sample reports, and require edits on key fields such as next steps."},{"title":"Prompt injection from external content","detail":"A news article or document contains text that steers the model. Strip instructions from retrieved content and keep tool permissions narrow."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Bankers interact with the copilot directly, but Article 50(1) does not bite here: it requires telling people they are dealing with an AI system unless that is obvious to a reasonably well informed person, and an internal tool that is openly presented and labelled as an AI assistant meets that bar by design. The copilot never interacts with the client. Article 50(2) marking of generated text falls on the provider of the system, including a bank that builds it in house, but the copilot turns a banker's own notes into a call report, an assistive function for standard editing of the banker's input that does not substantially alter it, so the Article 50(2) exception applies and no machine readable marking is required. It is not an Annex III use: credit context about corporate clients is not the creditworthiness assessment of natural persons in Annex III point 5(b), so it falls outside the high risk tier. If a deployment starts to score individuals for credit, the tier changes. AI literacy duties under Article 4 still apply. If meeting capture is used, recording and transcription rules under data protection law apply separately."},"regulations":["eu-ai-act","gdpr","dora","mas-ai-risk-management","nist-ai-rmf","iso-42001"],"guidance":[{"title":"MAS Guidelines for Artificial Intelligence (AI) Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Consultation paper of 13 November 2025 proposing supervisory expectations for AI oversight, AI inventories and risk materiality assessments at all financial institutions, explicitly covering generative AI and AI agents. Comments were due by 31 January 2026; this page cites the proposals, not final guidelines."},{"title":"OWASP Top 10 for Large Language Model Applications","issuer":"OWASP","region":"global","url":"https://genai.owasp.org/llm-top-10/","note":"Prompt injection through retrieved news and documents is the main technical risk for a copilot that reads external content."}],"controls":["Inventory entry with an accountable owner and a list of approved data sources","Entitlement checks at retrieval time, including information barriers","Source logging for every briefing, retained with the call report","Consent capture before any meeting recording","Monthly quality sampling with results reported to the business owner"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** for bankers, reachable in **Microsoft Teams** or a web\nwidget, that uses **custom functions** (REST calls) to read CRM, exposure and maturity data and a\n**knowledge base** with hybrid retrieval over approved internal documents. Public news can come\nthrough the built in web search tools or a connected **MCP** server for a licensed data provider.\n**Structured output** turns the pack and the call report into fixed templates.\n\nAn **agentic workflow** can prepare packs on a schedule before meetings and put CRM updates\nbehind **human in the loop approval**, so nothing is written back without the banker's\nconfirmation. **Guardrails** and **PII masking** protect client data in prompts, execution\ntracing records each agent turn for review, and **test suites** replay sample meetings on every\nprompt change. The platform is model agnostic and can run in the EU or UAE region."},"faq":[{"question":"How much time does an AI copilot save a relationship manager?","answer":"None of the deployments on this page reports a measured, quantified time saving. Bank of America says its meeting tool can save advisors up to four hours per meeting, but presents that as potential, not as a measured result, and Scotiabank's statement that a client report which took weeks now takes seconds comes from a proof of concept. Measure your own baseline preparation and call report time before rollout and compare on the same meeting types."},{"question":"Does the copilot give advice to clients?","answer":"No. It prepares information and drafts records for the banker. Advice and recommendations stay with the banker, who reviews every pack and approves every call report."},{"question":"What makes adoption stick?","answer":"Trust in the content matters as much as the tool, and a briefing is only as good as the CRM data behind it. Standard Chartered cleaned out most duplicated and outdated contacts as part of moving its corporate and investment bankers onto one CRM platform, the base for its AI client insights. Start with a meeting type bankers find tedious, show the source of every fact, and track weekly active use against licensed users."}],"related":["client-meeting-notes-and-crm-update","credit-memo-drafting-agent","corporate-client-servicing-assistant","sales-call-coaching-and-crm-update","treasury-cash-flow-forecasting","credit-early-warning-monitoring"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog and verified against primary sources; the catalog's DBS briefing pack claim was not found on the cited page and was dropped."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: rewrote unsourced claims in the problem and FAQ, sharpened the EU AI Act basis and the MAS guidance note, corrected the Scotiabank stage to announced (proof of concept), tightened the FNB and Standard Chartered summaries, and added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: moved the First National Bank email drafting record to reply drafting, dropped the Standard Chartered CRM headcount as a users served metric, rewrote the meta description so it no longer attributes the full copilot to named banks, corrected the Article 50 reasoning, and rewrote the adoption FAQ."}],"slug":"client-briefing-and-call-report-copilot","url":"https://www.blits.ai/ai-use-cases/client-briefing-and-call-report-copilot","benchmarks":[],"indicativeValueResult":{"low":1800000,"high":9000000},"evidence":["bank-of-america-merrill-ai-meeting-journey","scotiabank-client-insight-report-agents","standard-chartered-cib-client-insights-crm"]},{"title":"AI copilot for field technicians and dispatch optimization","shortTitle":"Field technician copilot and dispatch","seoTitle":"AI copilot for field technicians and dispatch","metaDescription":"AI can help decide if a fault needs a visit and predict the work. Openreach says AI updates to customers help prevent over 3,000 cancelled orders a month.","definition":"AI that decides whether a fault needs a site visit at all, predicts what work and parts a job will need, helps plan and update appointments, and gives technicians on site guided diagnosis and instant answers from manuals and past jobs, so more jobs are fixed on the first visit.","aliases":["field service copilot","technician assistant","truck roll avoidance","AI dispatch optimization","field workforce copilot"],"industries":["telecommunications"],"functions":["field-service","operations","customer-service"],"patterns":["prediction-and-scoring","rag-knowledge-assistant","conversational-agent","classification-and-routing"],"channels":["mobile-app","sms","internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"network","problem":"Technician visits are expensive: a van, a skilled person and often a customer who took time off to\nbe at home. Some visits should not happen at all, because the fault could have been fixed remotely\nor was not a fault. Others happen but fail: the\ntechnician finds a blocked duct, a missing part or a job that needs a different skill, and the\ncustomer has to wait for a second appointment.\n\nWhen installations run into trouble, customers can be left without clear information. Openreach\ndescribes how issues such as blocked underground ducts, complex build requirements or changed survey\nfindings could in the past mean missed appointments, repeat visits and uncertainty. nbn made the same point\nfrom the other side in 2017: it wanted its field workforce to connect more homes rather than attend\nto problems that do not exist. Meanwhile technicians on site search manuals, call colleagues or phone a help desk\nfor answers that already exist somewhere in the company.","problemStats":[],"howItWorks":"1. **Triage before dispatch.** When a fault is reported, a model uses line tests, device telemetry\n   and history to decide whether it can be fixed remotely, needs a visit, or needs a specific skill\n   or part.\n2. **Predict the job.** For visits, the system predicts the likely work type, duration and parts,\n   so the right technician arrives with the right kit, and flags jobs likely to need follow on work.\n3. **Plan and communicate.** Scheduling uses the predictions to build realistic routes and slots;\n   when a job hits a snag, the customer and the retail provider get a clear update with the\n   expected next step and timing, and can ask questions in their own words.\n4. **Assist on site.** The technician asks a copilot on a phone or tablet for procedures, wiring\n   diagrams, known issues and similar past jobs, and gets step by step diagnosis grounded in\n   approved documentation.\n5. **Close the loop.** Job notes, photos and outcomes are summarised into the ticket automatically\n   and feed back into triage and prediction models.","valueDrivers":["cost-to-serve","employee-productivity","customer-experience","speed"],"kpis":["cost-savings","processing-time-reduction","productivity-gain","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A fixed line operator that dispatches 1 million technician visits a year","inputs":[{"key":"visits","label":"Technician visits per year","low":1000000,"high":1000000,"unit":"visits per year","note":"The reference operator."},{"key":"avoidedShare","label":"Share of visits avoided through remote fixes, better triage and fewer repeat visits","low":0.02,"high":0.06,"unit":"fraction of visits","note":"Editorial assumption. The evidence on this page reports fewer cancellations and better triage but no verified share of visits avoided; replace with your own repeat visit and no fault found rates."},{"key":"costPerVisit","label":"Fully loaded cost of a technician visit","low":100,"high":200,"unit":"USD per visit","note":"Editorial assumption. Replace with your own cost per visit."}],"formula":"visits * avoidedShare * costPerVisit","currency":"USD","period":"per year","resultLabel":"Technician visit cost avoided","caveat":"Visit cost only. It leaves out time saved on site, fewer customer contacts about delayed jobs, fewer cancelled orders, and the cost of the platform and device rollout for technicians."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Triage and job prediction need joined up data from fault, test, workforce and order systems. The copilot for technicians is simpler but only works if manuals and procedures are current and usable offline or on weak coverage.","dataPrerequisites":["Fault and order history with visit outcomes, including no fault found and repeat visits","Line test and device telemetry results before dispatch","Job notes and completion codes from technicians","Current procedures, manuals and wiring diagrams in a searchable form"],"integrations":["Workforce management and scheduling","Fault management and line test systems","Order management and customer notification channels","Technician mobile apps","Knowledge management for field procedures"]},"implementation":{"steps":[{"title":"Measure wasted visits","detail":"Quantify no fault found, repeat and failed visits by fault type. That tells you where triage and prediction will pay back first."},{"title":"Start with triage before dispatch","detail":"Use tests and history to recommend remote fix or visit, and let dispatchers compare the recommendation with their decision before automating anything."},{"title":"Predict jobs that will go wrong","detail":"Predict when an installation will need extra work and use the prediction to update customers and plan the follow up early."},{"title":"Give technicians a grounded copilot","detail":"Put procedures and past job knowledge behind a copilot on the technician's device, with answers that cite the source and work on weak coverage."},{"title":"Keep humans in charge of people decisions","detail":"Use the models to plan work, not to rank or discipline individual technicians, and consult works councils or unions early."}],"guardrails":["Dispatch recommendations are advisory until measured against human decisions","No automated performance scoring or disciplinary use of technician data","Copilot answers cite approved procedures and refuse on safety critical steps without a source","Customer updates generated from predictions are reviewed for tone and accuracy on a sample basis"],"humanInTheLoop":"Dispatchers and team leaders own dispatch decisions and schedules, technicians own the work on site and can overrule the copilot, and safety procedures always follow the official manual. A sample of remote fix decisions is checked each week against later repeat faults.","kpisToInstrument":["Visits avoided through remote resolution, with repeat faults within 14 days counted against them","First time fix rate and repeat visit rate","Missed and cancelled appointments","Time on site per job type","Technician satisfaction with the copilot"],"failureModes":[{"title":"Remote fixes that do not stick","detail":"The model keeps customers off the visit list but the fault returns. Count repeat faults against avoided visits."},{"title":"Predictions nobody tells the customer about","detail":"The system knows a job will slip but customers still wait at home. Connect predictions to proactive updates."},{"title":"Copilot that fails in the field","detail":"Answers depend on coverage the technician does not have. Design for weak signal and offline use."},{"title":"Surveillance by stealth","detail":"Job data starts being used to rate individuals, and trust collapses. Set and publish purpose limits."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Triage and a knowledge copilot for technicians are normally minimal risk. Annex III point 4(b) lists AI systems that allocate tasks based on individual behaviour or personal traits, or that monitor and evaluate the performance and behaviour of workers, as high risk, so dispatch systems that do this need the full high risk controls, and under Article 26(7) employers must inform workers' representatives and the affected workers before using them. A conversational assistant that answers customers directly carries the Article 50 duty to tell people they are dealing with AI."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4 covers employment and worker management, including task allocation based on individual behaviour and performance monitoring."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"AI systems that interact directly with people must be designed so those people know they are dealing with an AI system, unless that is obvious."},{"title":"Article 26, obligations of deployers of high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/26/","note":"Paragraph 7 requires employers to inform workers' representatives and the affected workers before putting a high risk AI system into use at the workplace."}],"controls":["Documented purpose limits for technician data, agreed with employee representatives where required","Data protection impact assessment for location and job performance data","Measurement of dispatch recommendations against outcomes before automation","Review and ownership of field procedures used by the copilot"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the technician copilot is an **AI agent** with a **knowledge base** of procedures,\nmanuals and resolved jobs using **hybrid retrieval**, reachable from the technician's phone through\nthe **REST API channel** inside the operator's own app, **Microsoft Teams** or **WhatsApp**, or\nthrough the **web chat widget** with **voice input** for moments when hands are busy. **Custom functions** read job details, line tests\nand device status from the operator's workforce and fault systems.\n\nFor customers, an agent on **SMS** or **WhatsApp**, built with **flows**, answers questions about a\ndelayed job in the customer's own words using the latest job status, with **human handover** to the\nservice team. **Agentic workflows** triggered via API when a job changes can prepare the update.\n**Guardrails** check inputs and outputs against the operator's policies, **PII masking**\nprotects customer data in prompts, and **test suites** replay real questions before every\nchange."},"faq":[{"question":"Can AI decide whether a technician needs to visit?","answer":"It can recommend it. In 2017 nbn described a machine learning Tech Lab meant to help determine whether a fault can be dealt with remotely or needs a field technician, so its workforce spends time connecting homes rather than attending problems that do not exist. The 2017 post reports no results, so start with dispatchers comparing the recommendation with their own decision."},{"question":"What happens when an installation runs into problems?","answer":"Openreach uses a prediction tool, Crystal Ball, to predict what work a delayed installation will need and whether it will take more or less than ten days. That drives a text message update to the customer, typically within 24 hours of the engineer reporting the issue. Openreach says its AI tools help prevent more than 3,000 cancellations a month."},{"question":"Is AI dispatch high risk under the EU AI Act?","answer":"It depends on the design. Planning jobs by fault type and location is normally not high risk; allocating work based on individual technicians' behaviour or monitoring their performance falls under Annex III point 4 and is high risk. A chatbot that answers customers must also be designed so they know they are talking to AI, unless that is obvious (Article 50)."}],"related":["predictive-network-maintenance","device-and-connectivity-troubleshooting-agent","network-fault-triage-copilot","network-outage-communication-agent","enterprise-knowledge-search"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched for the telecommunications vertical and verified against operator sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: nbn record marked as announced and dated 2017, problem and FAQ wording aligned with the sources, Article 26(7) and Article 50 added to the risk basis, Blits.ai build limited to listed capabilities, SEO title and description added."},{"date":"2026-09-27","note":"Review fixes: Openreach vendor named as the CXone platform only, nbn claim scoped to the 2017 post, voice input tied to the web chat widget, Article 26 guidance added, Article 50 wording and meta description made more precise."},{"date":"2026-09-27","note":"Second review pass: reworded the Blits.ai guardrails claim to match the feature inventory (policy checks, not grounding)."}],"slug":"field-technician-copilot-and-dispatch","url":"https://www.blits.ai/ai-use-cases/field-technician-copilot-and-dispatch","benchmarks":[],"indicativeValueResult":{"low":2000000,"high":12000000},"evidence":["nbn-tech-lab-fault-dispatch-prediction","openreach-crystal-ball-installation-prediction"]},{"title":"AI copilot for hotel revenue management","shortTitle":"Hotel revenue management copilot","seoTitle":"Hotel revenue management AI copilot","metaDescription":"An AI copilot forecasts hotel demand and prices. RIMC Hotels & Resorts Group reports a 28.44% RevPAR gain at its Polish property; Duetto reports 33% at Hôtel Swexan.","definition":"An employee facing AI system that forecasts demand for a hotel or portfolio by date, room type and segment, recommends or automatically adjusts room prices and availability controls within limits a revenue manager sets, and scores group and event enquiries for true profitability, so a small revenue team can run pricing that used to need daily manual adjustment in the property management system.","aliases":["hotel revenue management AI","AI dynamic pricing for hotels","hotel demand forecasting copilot","revenue management system AI"],"industries":["travel-and-hospitality"],"functions":["product-and-pricing","analytics-and-reporting"],"patterns":["prediction-and-scoring","agentic-workflow"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"revenue-management","problem":"A hotel's price should change with demand: a room that rents cheaply on a quiet Tuesday can be\nworth much more on a night a conference fills the market. Before a revenue management system,\nthat means one person watching pickup, competitor rates and events across every room type and segment, and\nupdating prices by hand in the property management system. RIMC Hotels & Resorts Group describes\nthis before it adopted Duetto: pricing managed manually, with adjustments made through\nthe property management system, a process it calls time consuming and inefficient, especially for\npricing and capacity control.\n\nThe problem compounds with scale and complexity. Hôtel Swexan runs a 134 room luxury property\nwith more than 20 room types, eight premium suites and five food and beverage outlets, with one\nperson, Director of Revenue Jessica Schiele, responsible for the whole pricing strategy. Before an\nAI assisted revenue management system, every hour she spent pushing rates by hand was an hour not\nspent on analysis or strategy, and macro level tools such as closing the whole house or a blanket\nminimum stay were the only practical levers, because granular, segment level rules took too long\nto maintain manually.","problemStats":[],"howItWorks":"1. **Pull the data continuously.** The system reads live occupancy, rate and booking pace from the\n   property management or central reservation system, plus competitor rates and, where used,\n   group and food and beverage data.\n2. **Forecast demand.** It predicts occupancy and demand by date, room type and segment, updating\n   the forecast as new bookings and cancellations arrive.\n3. **Recommend or execute a price.** Within a floor, a ceiling and a maximum daily rate change the\n   revenue manager configures, it either recommends a rate for approval or pushes it directly to\n   the property management system.\n4. **Score group and event enquiries.** For a group or function space request, it weighs room\n   revenue, food and beverage minimums, room rental and the displacement of other business to\n   recommend a profitable quote, not just an available rate.\n5. **Flag what needs a person.** Forecast misses, a recommendation outside the configured range, a\n   new room type or segment and any group quote below the profitability threshold go to the\n   revenue manager to review or approve.","valueDrivers":["revenue-growth","employee-productivity","speed"],"kpis":["revenue-uplift","forecast-accuracy","time-saved-per-task","productivity-gain"],"indicativeValue":{"referenceOrg":"A hotel group with 20 properties and average annual room revenue of USD 3 million per property","inputs":[{"key":"properties","label":"Properties in the portfolio","low":20,"high":20,"unit":"properties","note":"The reference group."},{"key":"roomRevenuePerProperty","label":"Annual room revenue per property before the change","low":2000000,"high":4000000,"unit":"USD per property per year","note":"Editorial assumption, replace with your own room revenue."},{"key":"revparUplift","label":"RevPAR uplift attributable to AI assisted revenue management","low":0.03,"high":0.1,"unit":"fraction of room revenue","note":"Conservative against the evidence on this page. RIMC Hotels & Resorts Group reports a 28.44% RevPAR increase at its Polish property, with no period or baseline stated for that figure, and, separately, an increase in RevPAR at every hotel in its portfolio compared to the previous year since adopting Duetto in 2022. Duetto's case study reports 33% total hotel RevPAR growth at Hôtel Swexan, 2025 versus 2024, with no date given for when Hôtel Swexan itself went live on Duetto. RIMC's own account describes its pricing before Duetto as fully manual, adjusted through the property management system, with no revenue management system of any kind in place before 2022: a group already running a mature revenue management system and adding only an AI forecasting layer on top of it should expect a smaller further gain than a group's first system delivered. Neither hotel's own case study uses the words AI or machine learning about its pricing; both run Duetto's revenue management platform, and Duetto's own product page for that platform, GameChanger, describes pairing data with \"AI-driven rate recommendations\" (checked with usecases:source, 2026-09-28)."}],"formula":"properties * roomRevenuePerProperty * revparUplift","currency":"USD","period":"per year","resultLabel":"Additional room revenue from AI assisted pricing","caveat":"Gross room revenue uplift only. It leaves out the cost of the revenue management system and its integration, any change in distribution cost or food and beverage revenue, and the fact that part of a RevPAR gain in any single case can come from market conditions rather than the system itself."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The forecasting and pricing logic itself is largely the vendor's; the work is integrating the property management or central reservation system, a channel manager and a competitor rate feed cleanly enough that the forecast is trustworthy, and agreeing the floor, ceiling and approval rules before turning on any automatic pricing.","dataPrerequisites":["At least one to two years of historical occupancy, rate and booking pace by room type and segment","A competitor set and a rate shopping feed","Group, event and food and beverage data if group profitability scoring is in scope","Agreed rate floors, ceilings and maximum daily change per room type and segment"],"integrations":["Property management system or central reservation system","Channel manager and distribution system","Competitor rate shopping tool","Point of sale or catering system for group and food and beverage profitability","Business intelligence or reporting tool for the revenue manager's dashboard"]},"implementation":{"steps":[{"title":"Connect the data before you connect the automation","detail":"Integrate the property management system, channel manager and rate shopping feed first, and run the forecast in shadow mode against actual outcomes before any price is pushed automatically."},{"title":"Set the guardrails first","detail":"Agree the rate floor, ceiling and maximum single day change per room type and segment with the revenue manager before switching on automatic pricing, not after."},{"title":"Start with one segment or room type","detail":"Prove forecast accuracy and RevPAR impact on transient, individual bookings before extending automation to groups, suites or a new property."},{"title":"Add group and event profitability scoring once transient pricing is stable","detail":"Bring in food and beverage minimums, room rental and displacement so any team member, not just the revenue manager, can quote a group at a profitable rate."},{"title":"Review overrides every week","detail":"Look at every case where the revenue manager overrode the recommendation, and why; a pattern of overrides in the same direction means the forecast or the guardrails need adjusting."}],"guardrails":["Automatic pricing bounded by a floor, a ceiling and a maximum single day rate change, set by the revenue manager","Any recommendation outside the configured range, or for a new room type or segment, requires human approval before it goes live","Group and event quotes below the agreed profitability threshold route to a person, not straight to the quote"],"humanInTheLoop":"The revenue manager sets and periodically revisits the pricing thresholds and rules, approves anything the system flags as outside them, and owns the final call in unusual situations such as a citywide event, a competitor's distress pricing or a local disruption the forecast has not seen before.","kpisToInstrument":["Forecast accuracy against actual occupancy and pickup, by room type and segment","RevPAR, ADR and occupancy against the competitor set, before and after","Override rate: how often the revenue manager changes or rejects a recommendation, and why","Time the revenue team spends on manual pricing and monitoring tasks"],"failureModes":[{"title":"A forecast blind to what has not happened before","detail":"A model trained on historical patterns misses a new event, a cancellation wave or a local disruption. Keep an anomaly alert and a fast human override path for exactly these moments."},{"title":"Automated pricing that moves against a competitor's algorithm","detail":"Automated systems in the same market can drift in either direction: repeatedly undercutting each other compresses rates for everyone, and, when several hotels lean on the same vendor's pricing recommendations, the arrangement can itself become an antitrust question. Two federal appeals courts have reached different outcomes on different shared pricing software complaints: the Ninth Circuit upheld the dismissal of a Sherman Act suit over shared Las Vegas Strip hotel pricing software in 2025, and the Third Circuit revived a similar suit over Atlantic City casino hotel pricing software in 2026, based on that complaint's own allegations (see the incidents on this page). Enforce both a rate floor and a ceiling, and review competitor reactions, not just your own pickup."},{"title":"Chasing short term pickup at the expense of the base","detail":"The system discounts aggressively on a soft looking date and displaces higher value corporate or group business that would have booked later. Weigh displacement, not only occupancy, in the recommendation."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"A demand forecasting and pricing tool used by hotel staff is not listed in Annex III and is not a system that decides on a natural person's access to an essential service; it prices a hotel room, not a person. It is not customer facing, so the Article 50 transparency duty for conversational AI does not apply. The AI literacy obligation on staff who use AI systems (Article 4) still applies."},"regulations":["eu-ai-act"],"guidance":[{"title":"Article 4, AI literacy","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Providers and deployers must take measures to ensure staff and other people who operate and use an AI system on their behalf have a sufficient level of AI literacy."}],"controls":["Rate floor, ceiling and maximum daily change configured per room type and segment, reviewed periodically","Every automatic price change logged with the forecast and inputs that produced it, for audit","Regular comparison of AI recommended rates against revenue manager overrides to catch systematic bias"],"incidents":[{"title":"Gibson v. Cendyn Group: Ninth Circuit affirms dismissal of a Las Vegas Strip hotel pricing algorithm suit","url":"https://cdn.ca9.uscourts.gov/datastore/opinions/2025/08/15/24-3576.pdf","note":"A Sherman Act Section 1 class action alleged that competing hotels on the Las Vegas Strip fixed room prices by all licensing Cendyn Group's algorithmic pricing software. The Ninth Circuit's summary states it is \"affirming the district court's dismissal,\" holding that several competitors independently choosing the same pricing software, followed by higher prices, does not by itself show an anticompetitive agreement. The US Department of Justice's Antitrust Division appeared as amicus curiae; the opinion does not state which side the government argued for. Decided August 15, 2025 (No. 24-3576, 9th Cir.)."},{"title":"Cornish-Adebiyi v. Caesars Entertainment: Third Circuit revives an Atlantic City hotel pricing algorithm suit","url":"https://www2.ca3.uscourts.gov/opinarch/243006p.pdf","note":"A parallel Sherman Act Section 1 suit against Atlantic City casino hotels that used the same Cendyn Group pricing software was dismissed by the district court in 2024. On July 29, 2026 the Third Circuit reversed that dismissal and remanded, writing that it will \"reverse the District Court's dismissal of the Complaint,\" and holding that the complaint's well pleaded allegations of a hub and spoke agreement run through Cendyn's shared pricing software are sufficient to proceed (No. 24-3006, 3rd Cir.). The opinion cites Gibson v. Cendyn Group, the Ninth Circuit's ruling in the Las Vegas Strip case, without disagreeing with it: the two courts reached different outcomes on different complaints with different allegations, not opposite readings of the same rule."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow**, not a conversation. It runs on a schedule, and\n**custom functions** (REST calls, SQL queries and custom code in an isolated JavaScript sandbox)\nconnect it to the property management or central reservation system, the channel manager, a\nrate shopping feed and an external forecasting or revenue management engine that produces the\ndemand forecast and a recommended rate. A **SQL knowledge base** holds the structured\noccupancy, rate and booking pace history those calls read. Within a floor and ceiling the\nrevenue manager configures, a custom function pushes the recommended rate back to the property\nmanagement system.\n\n**Human in the loop confirmation** holds any recommendation above the configured threshold, or\nfor a new room type or segment, for a person to approve or reject before it reaches the\nproperty management system. Every run is recorded in the workflow's run history with a full\naudit trail, and **test suites** run against the workflow before a change to its configuration\ngoes live. The platform is **model agnostic**, so the LLM behind the workflow's own reasoning\ncan be swapped between providers without rebuilding it, and **EU and UAE data residency** keeps\nrate and booking data in region for groups that require it."},"faq":[{"question":"How much RevPAR gain can an AI revenue management copilot deliver?","answer":"It depends heavily on the starting point. RIMC Hotels & Resorts Group reports a 28.44% RevPAR increase at its Polish property, with no period or baseline stated for that figure, and, separately, a RevPAR increase across its whole portfolio compared to the previous year since adopting Duetto in 2022; Duetto's case study reports 33% total hotel RevPAR growth at Hôtel Swexan, 2025 versus 2024. Treat both as single case study results rather than a guaranteed uplift: RIMC moved from pricing it describes as fully manual, adjusted through the property management system, with no revenue management system in place before 2022, and the Swexan case study does not state when its own Duetto deployment began relative to that comparison. Neither hotel's own case study calls its pricing AI or machine learning; both run Duetto's revenue management platform, which Duetto's own product page describes as pairing data with \"AI-driven rate recommendations\"."},{"question":"Does the AI set prices on its own?","answer":"It depends on the deployment. Hôtel Swexan's case study describes automated pricing strategies that execute around the clock \"without manual intervention,\" based on occupancy thresholds and segment logic the Director of Revenue, Jessica Schiele, defines in advance, with no need to change rates by hand. The oversight in that setup is Schiele setting and revising those rules and running a daily check the case study describes as a few minutes rather than a few hours, not a person approving each price. Setting an explicit rate floor, a ceiling and a maximum daily change before switching on automatic pricing, as the implementation guidance on this page recommends, is good practice, but neither cited deployment states those specific limits publicly."},{"question":"What data does a hotel need before it can use one?","answer":"At least a year or two of historical occupancy, rate and booking pace by room type and segment, a defined competitor set with rate shopping data, and, for group profitability scoring, food and beverage and function space data."},{"question":"Is this the same as a hotel guest chatbot?","answer":"No. A revenue management copilot is an internal tool for the revenue team that sets prices; it does not talk to guests. A guest facing booking or service assistant is a separate use case."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by the editor after the Opus targeted check's remaining text fixes."},{"date":"2026-09-28","note":"New use case: AI copilot for hotel revenue management, with two verified deployments (RIMC Hotels & Resorts Group and Hôtel Swexan) on Duetto's revenue management platform."},{"date":"2026-09-28","note":"Editorial review: corrected the RIMC RevPAR figure's scope in the meta description, attributed the Hôtel Swexan RevPAR figures to Duetto's case study rather than the hotel, removed the unsupported claim that both deployments enforce a rate floor, ceiling and escalation rule, dropped Swexan from the \"moved from manual pricing\" framing, and rewrote the Blits.ai build section to match the platform feature inventory."},{"date":"2026-09-28","note":"Adversarial review fixes: removed the unsourced \"early adopter\" label for RIMC Hotels & Resorts Group and Hôtel Swexan from the indicative value note and FAQ 1, and stopped attaching \"since adopting Duetto in 2022\" to the unrelated 28.44% RevPAR figure. Added two risk.incidents entries with the actual case outcomes (Ninth Circuit affirmed dismissal in Gibson v. Cendyn Group, 2025; Third Circuit reversed dismissal in Cornish-Adebiyi v. Caesars Entertainment, 2026) and tied the antitrust failure mode to them instead of leaving it unsourced. Dropped \"exactly\" from the RIMC problem framing and reworded FAQ 2 so the \"without oversight\" contrast is not attributed to the Hôtel Swexan case study, which actually stresses \"without manual intervention.\" Noted for the editor: REGULATIONS has no competition or antitrust id, so risk.regulations still lists only the EU AI Act."},{"date":"2026-09-28","note":"Round 2 review fixes: confirmed with usecases:source that neither the RIMC nor the Hôtel Swexan case study uses the words AI, artificial intelligence or machine learning about its pricing; both evidence records and the indicativeValue and FAQ 1 text now cite Duetto's own GameChanger product page, which describes pairing data with \"AI-driven rate recommendations,\" as the source for the AI claim, and both evidence records now carry that page as a second source. Removed the RIMC \"time saved per task\" metric (30 minutes is a per day, per property rate, not a rate per task; kept in the evidence summary as prose and flagged as a taxonomy gap). Gibson v. Cendyn Group incident note no longer claims the DOJ argued for the plaintiffs; the opinion only shows the DOJ appeared as amicus curiae. Cornish-Adebiyi v. Caesars Entertainment incident note drops \"with prejudice\" and the DOJ statement of interest claim, neither of which is in the cited opinion, and restates the Third Circuit's holding as specific to that complaint's hub and spoke allegations rather than a general rule. Removed the \"two circuits disagree\" framing from the incident note and failureModes[1], since the Third Circuit opinion cites Gibson without disagreeing with it; both now say the courts reached different outcomes on different complaints."}],"slug":"hotel-revenue-management-copilot","url":"https://www.blits.ai/ai-use-cases/hotel-revenue-management-copilot","benchmarks":[{"kpi":"revenue-uplift","label":"Revenue uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":30.72,"min":28.44,"max":33,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"hotel-swexan-duetto-revenue-management","pooled":true},{"id":"rimc-hotels-resorts-duetto-revenue-management","pooled":true}]}],"indicativeValueResult":{"low":1200000,"high":8000000},"evidence":["hotel-swexan-duetto-revenue-management","rimc-hotels-resorts-duetto-revenue-management"]},{"title":"AI copilot for insurance pricing and actuarial analysis","shortTitle":"Pricing and actuarial copilot","seoTitle":"AI for insurance pricing and actuarial teams","metaDescription":"AI pricing copilots automate data preparation and model search while actuaries sign off the rates. Generali France reports modelling five times faster with Akur8.","definition":"AI that speeds up the work of pricing and actuarial teams, from automated, transparent risk and demand model building to natural language analysis of rate filings, experience data and reserving diagnostics, while actuaries select the models, sign off the rates and own the professional judgment.","aliases":["AI for insurance pricing","actuarial modelling automation","ratemaking copilot"],"industries":["insurance"],"functions":["product-and-pricing","risk-management","analytics-and-reporting"],"patterns":["prediction-and-scoring","code-generation","agentic-workflow","summarization"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"pricing","problem":"Pricing and actuarial teams work to fixed review cycles with a high volume of work. Building or\nupdating a risk model in traditional tools can take weeks of data preparation, variable selection\nand code, often split across separate tools such as SAS, Python or R, where every change to the\ndata means manual code updates that are hard to version and audit. Competitor filings, experience\nstudies and reserving reviews compete for the same people.\n\nThe constraint is not only speed. Pricing models must be explainable to supervisors, tested for\nunfair discrimination and consistent with fair value rules, so machine learning that improves\naccuracy but cannot be explained is hard to defend. The opportunity is AI that removes the manual\nwork while keeping models transparent and actuaries in control.","problemStats":[],"howItWorks":"1. **Prepare the data.** The platform imports policy, claims and quote data, handles missing values\n   and builds candidate features, including approved external data.\n2. **Build transparent models faster.** Automated search explores thousands of variable\n   combinations and interactions and proposes interpretable models (for example generalized linear\n   or additive models), while actuaries choose and adjust the final model.\n3. **Set rates.** Actuaries compare pricing scenarios and their effect on volume, loss ratio and\n   fairness tests, then approve the rates.\n4. **Ask questions in plain language.** A generative assistant answers questions over rate\n   filings, experience data and reserving outputs, and drafts code or documentation for review.\n5. **Deploy and monitor.** Approved rates go to the rating engine through an audited deployment,\n   and model performance is monitored against actual experience.","valueDrivers":["speed","employee-productivity","risk-reduction","compliance"],"kpis":["productivity-gain","cycle-time-days","processing-time-reduction","accuracy","error-reduction"],"indicativeValue":{"referenceOrg":"An insurer with a pricing and actuarial team of 20 people","inputs":[{"key":"actuaries","label":"Pricing and actuarial staff","low":20,"high":20,"unit":"people","note":"The reference insurer."},{"key":"modellingShare","label":"Share of their time spent on model building and data preparation","low":0.3,"high":0.5,"unit":"fraction of working time","note":"Editorial assumption. Replace with your own time allocation."},{"key":"timeSaved","label":"Share of that time the AI removes","low":0.5,"high":0.75,"unit":"fraction of modelling time","note":"Conservative against the evidence on this page (Generali France's actuarial studies manager reports modelling five times faster; Akur8 reports that Europ Assistance cut work that took weeks to one or two days)."},{"key":"costPerPerson","label":"Fully loaded annual cost per person","low":120000,"high":180000,"unit":"USD per year","note":"Editorial assumption. Replace with your own fully loaded cost."}],"formula":"actuaries * modellingShare * timeSaved * costPerPerson","currency":"USD","period":"per year","resultLabel":"Actuarial capacity released","caveat":"Capacity released, not cash saved. It leaves out the usually larger value of better risk selection and faster rate changes on the loss ratio, software licences, and the validation and governance effort that pricing models need."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Automated modelling platforms already run in production at insurers such as MAIF and Generali France; the effort is in data quality, integration with rating engines, model governance and fairness testing. Generative assistants for actuarial analysis are newer and need controls on data access and on code they generate.","dataPrerequisites":["Clean, joined policy, exposure, claims and quote data at the level pricing needs","Approved external data sources with documented use conditions","Model governance standards and fairness testing methods","Rate filings and experience studies in a searchable form"],"integrations":["Data warehouse or lakehouse holding policy and claims data","Rating engine for deployment","Model risk management inventory","Version control and documentation repository"]},"implementation":{"steps":[{"title":"Pick one product with a rate review due","detail":"Run the new approach in parallel with the existing process on a real rate review so the results can be compared on accuracy, time and explainability."},{"title":"Keep models explainable by design","detail":"Prefer methods that produce interpretable models (GLM or GAM style) or add robust explanations, because regulators and fair value reviews will ask why each factor is there."},{"title":"Build fairness testing into the workflow","detail":"Test rating factors and outcomes for proxies of protected characteristics before approval, as rules such as Colorado's SB21-169 require for covered lines."},{"title":"Govern generated code and analysis","detail":"Treat code or documentation drafted by a generative assistant like a junior's work: reviewed, tested and versioned before it touches production models."},{"title":"Monitor after deployment","detail":"Compare predicted and actual experience monthly and set triggers for review."}],"guardrails":["Actuaries select and sign off every model and rate; nothing deploys without approval","Fairness and proxy testing before any new factor is approved","Full lineage from data to deployed rate, with version control","Generative assistants have read only access to data and cannot deploy","Price optimization constrained by fair value and renewal pricing rules where they apply"],"humanInTheLoop":"Actuaries own model selection, rate approval and professional sign off. Model risk management validates pricing models independently, and compliance reviews fairness test results and rate filings before they go to regulators.","kpisToInstrument":["Elapsed days from data extract to approved model, per review","Models built or refreshed per actuary per quarter","Predictive lift of new models versus the current rates on holdout data","Fairness test results per model version","Actual versus expected loss ratio after deployment"],"failureModes":[{"title":"More accurate, less explainable","detail":"A complex model wins on lift but cannot be explained in a filing. Set explainability requirements before modelling starts."},{"title":"Proxy discrimination","detail":"External data or interactions act as proxies for protected characteristics. Test outcomes by group and document the reasons for every factor."},{"title":"Speed without governance","detail":"Faster models mean more changes than validation can keep up with. Scale validation capacity with modelling capacity."},{"title":"Unreviewed generated code","detail":"Code drafted by an assistant contains a subtle error that flows into rates. Require review and tests for every change."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Pricing and risk assessment of natural persons for life and health insurance is high risk under Annex III point 5(c). Pricing for property and casualty products, and actuarial analysis that does not price individuals, are not listed, although supervisors still expect sound model governance."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","nist-ai-rmf","iso-42001","solvency-ii","eu-idd"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(c) makes AI for risk assessment and pricing of natural persons in life and health insurance high risk."},{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Covers fairness, data governance, explainability and human oversight for AI in insurance, including pricing models outside the high risk list."},{"title":"SB21-169, Protecting Consumers from Unfair Discrimination in Insurance Practices","issuer":"Colorado Division of Insurance","region":"north-america","url":"https://doi.colorado.gov/for-consumers/sb21-169-protecting-consumers-from-unfair-discrimination-in-insurance-practices","note":"Holds insurers accountable for testing external consumer data, algorithms and predictive models so they do not unfairly discriminate on the basis of a protected class; amended Regulation 10-1-1 sets governance requirements for life, private passenger auto and health insurers."},{"title":"PS21/5: General insurance pricing practices market study, feedback to CP20/19 and final rules","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publication/policy/ps21-5.pdf","note":"Limits UK home and motor renewal prices to the equivalent new business price, which constrains price optimization models."}],"controls":["Model inventory entries with owners, validation status and approval history","Fairness and proxy testing records per model version","Rate change approval workflow with actuarial sign off","Access controls separating analysis tools from production deployment","Post deployment monitoring with documented triggers"],"incidents":[{"title":"Suckers List: How Allstate's Secret Auto Insurance Algorithm Squeezes Big Spenders","url":"https://themarkup.org/allstates-algorithm/2020/02/25/car-insurance-suckers-list","note":"Reporting by The Markup and Consumer Reports on a price adjustment algorithm Allstate filed in Maryland, which regulators rejected as discriminatory; Allstate said its rating plans comply with state laws and regulations. A reminder that pricing models are judged on outcomes, not only on predictive accuracy."}]},"blitsAi":{"howToBuild":"Blits.ai does not replace a pricing platform or rating engine. It fits the analysis around them: an\n**AI agent** with **SQL knowledge bases** over experience and filing data answers actuaries'\nquestions in plain language, and a **knowledge base** of rate filings, pricing guidelines and model\ndocumentation, searched with **hybrid retrieval**, answers questions over those documents.\n\n**Agentic workflows** can be triggered on a schedule to prepare recurring analyses (experience\nmonitoring, actual versus expected reports, competitor filing summaries), with **human in the loop\napproval** before results are used, and the file generation tool produces the output documents.\nConnect the databases with read only accounts and **custom functions**, limit the tools a workflow\nmay call with the tool execution policy, use **execution tracing** to inspect every turn and the\nqueries it ran, and run **test suites** that check answers against known figures. The platform is\nmodel agnostic, and data can stay in EU or UAE regions."},"faq":[{"question":"How much faster does AI make insurance pricing?","answer":"Figures published by the vendor Akur8 report large gains: Generali France's actuarial studies manager says modelling is five times faster, and Akur8's case study says Europ Assistance now completes pricing work that took weeks in one or two days. These are not independently audited and do not measure effects on loss ratios."},{"question":"Is AI pricing high risk under the EU AI Act?","answer":"For life and health insurance of individuals, yes, under Annex III point 5(c). For motor, home and commercial lines the Act does not list pricing as high risk, but fairness, explainability and national pricing rules such as the FCA's renewal pricing rules still apply."},{"question":"Do actuaries still decide?","answer":"Yes. The tools on this page automate data preparation, variable search and analysis, while actuaries select models and sign off rates. Kinsale reports that AI tool use is most prevalent in its IT, actuarial and analytical teams."}],"related":["underwriting-risk-assessment-copilot","model-risk-validation-copilot","insurance-renewal-and-retention","governed-text-to-sql-analytics","conversational-insurance-quote-and-buy"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Unpublished by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from insurer filings and vendor case studies verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added Solvency II, Insurance Distribution Directive to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact check against sources: rewrote unsourced problem statements, attributed the Europ Assistance figure to Akur8, corrected the Colorado SB21-169 and Allstate incident notes, aligned the Blits.ai section with the feature inventory, added seoTitle and metaDescription."}],"slug":"insurance-pricing-and-actuarial-copilot","url":"https://www.blits.ai/ai-use-cases/insurance-pricing-and-actuarial-copilot","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":5,"min":5,"max":5,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"generali-france-akur8-pricing-models","pooled":true}]},{"kpi":"cycle-time-days","label":"Cycle time","unit":"days","aggregate":false,"higherIsBetter":false,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"europ-assistance-akur8-pricing","pooled":false}]}],"indicativeValueResult":{"low":360000,"high":1350000},"evidence":["accelerant-ai-underwriting-and-actuarial-tools","europ-assistance-akur8-pricing","generali-france-akur8-pricing-models","kinsale-ai-submission-routing","maif-akur8-pricing-models"]},{"title":"AI copilot for marketing content with compliance pre review","shortTitle":"Marketing content and compliance","seoTitle":"AI for marketing content and compliance review","metaDescription":"A copilot drafts marketing copy from approved facts and checks it against ad rules before review. In a pilot, Ally's marketers reported 34% average time saved.","definition":"A copilot that drafts campaign copy, product explainers and social posts on brand and in the customer's language from approved product facts, then runs a first pass compliance check against advertising rules and required disclosures, flagging unsupported claims and missing warnings before a human in marketing compliance approves publication.","aliases":["AI marketing copywriting","financial promotions review AI","marketing compliance review","generative AI content production","MLR pre review"],"industries":["cross-industry","banking","insurance","payments","wealth-and-asset-management","pharma-and-life-sciences"],"functions":["marketing","regulatory-compliance","legal"],"patterns":["content-generation","rag-knowledge-assistant","classification-and-routing","translation"],"channels":["internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Marketing teams in regulated industries produce a steady stream of assets: emails, landing pages,\nsocial posts, product explainers, often in several languages and formats. Each one must be on\nbrand and must pass compliance review: in financial services a promotion must be clear, fair and\nnot misleading, show rates and fees correctly and carry the required risk warnings; in pharma,\nmedical, legal and regulatory review plays the same role. Review queues and rounds of redrafting\nbetween marketing and compliance add to the time it takes to launch a campaign.\n\nGenerative AI can draft quickly, but in these industries a fast draft that invents a rate, implies\na guarantee or drops a risk warning creates regulatory and conduct risk. The job is to speed up\ndrafting and give compliance a consistent first pass, while keeping product facts locked and a\nhuman approver accountable for every published asset.","problemStats":[],"howItWorks":"1. **Brief.** The marketer gives the product, audience, channel, language and message. The copilot\n   retrieves the approved product facts (rates, fees, eligibility), the brand voice guide and the\n   required disclosures for that product and channel.\n2. **Draft within the facts.** It drafts variants that use only the approved facts, inserting\n   numbers and disclosures from the source rather than generating them, and adapts length and tone\n   to the channel.\n3. **Compliance first pass.** A separate check compares the draft with the advertising rules and\n   house policy: unsupported or superlative claims, missing or misplaced risk warnings, balance of\n   benefits and risks, and terms that need a qualifier. Each flag cites the rule it relies on.\n4. **Localize.** Approved copy is translated and culturally adapted, and the compliance check runs\n   again in the target language.\n5. **Human approval.** Marketing edits, and a compliance reviewer approves or rejects with\n   comments; nothing is published without that approval.\n6. **Record.** The final asset, its claims, sources, flags and approvals are stored as a versioned\n   record for audit and for later complaints.","valueDrivers":["speed","employee-productivity","compliance","cost-to-serve"],"kpis":["productivity-gain","cycle-time-days","cost-savings","processing-time-reduction","time-saved-per-task"],"indicativeValue":{"referenceOrg":"A retail bank producing 1,500 marketing assets a year across two languages","inputs":[{"key":"assets","label":"Marketing assets produced per year","low":1500,"high":1500,"unit":"assets per year","note":"The reference organization. Replace with your own volume."},{"key":"hoursPerAsset","label":"Marketing and compliance hours per asset today, including review rounds","low":4,"high":8,"unit":"hours per asset","note":"Editorial assumption."},{"key":"timeSaved","label":"Share of those hours saved","low":0.15,"high":0.3,"unit":"fraction of time","note":"Conservative against the evidence on this page. Ally reports an average time saving of 34% in its marketing experiment and says the largest reductions came in early creative tasks such as research, first drafts and naming; none of the sources measures review time separately."},{"key":"hourlyCost","label":"Blended marketing and compliance hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption. Replace with your own rate."},{"key":"agencySpend","label":"External copywriting and translation spend per year","low":200000,"high":600000,"unit":"USD per year","note":"Editorial assumption."},{"key":"agencyReduction","label":"Reduction in that external spend","low":0.1,"high":0.25,"unit":"fraction of spend","note":"Editorial assumption; Klarna reports a 25% cut in external marketing supplier spend, used here as the high end."}],"formula":"assets * hoursPerAsset * timeSaved * hourlyCost + agencySpend * agencyReduction","currency":"USD","period":"per year","resultLabel":"Staff time released plus reduced agency spend","caveat":"It leaves out any revenue effect of more or better targeted campaigns, the value of fewer compliance breaches and complaints, and the cost of maintaining the approved fact base and rule library."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Drafting is easy to start. The work that makes it safe is a maintained source of approved product facts, a written rule library for the compliance check, and a workflow that records approvals.","dataPrerequisites":["Approved product facts (rates, fees, eligibility, terms) with an owner and effective dates","Brand voice and style guide with examples of approved assets","Required disclosures and risk warnings per product, channel and market","The advertising rules and house policy the compliance team applies, written as checkable rules"],"integrations":["Content management system or digital asset management","Marketing automation and campaign tools","Approval workflow used by marketing compliance","Product information source for current rates and fees"]},"implementation":{"steps":[{"title":"Lock the facts first","detail":"Build a single source of approved product facts and disclosures that the copilot reads from. Numbers and warnings are inserted from that source, never generated."},{"title":"Turn the compliance manual into checks","detail":"With compliance, write the rules the first pass applies (claims that need substantiation, banned phrases, warning placement, balance of risks and benefits) and test them on past approved and rejected assets."},{"title":"Start with one channel and product family","detail":"Pick high volume, lower risk assets such as service emails or social posts for savings products, and measure drafting time, review rounds and rejection reasons."},{"title":"Keep approval where it is","detail":"Plug the copilot into the existing approval workflow so every asset still gets a named approver. In Ally's experiment, an AI drafted blog article still went through its established regulatory review."},{"title":"Measure review rounds, not just drafting","detail":"Drafting gains are the easiest to see. Measure time in review as well, track first time approval rate and rejection reasons, and feed recurring flags back into the rules."},{"title":"Extend to images and languages carefully","detail":"Add image generation and translation once text is stable, with disclosure of AI generated imagery where required and a compliance check in each language."}],"guardrails":["Rates, fees, returns and other product numbers come only from the approved fact source, never from the model","The model may not promise guarantees, returns or outcomes the product terms do not support","Every asset gets a named human approver in marketing compliance before publication","Compliance flags cite the rule they rely on, and a pass by the first check never replaces human review","AI generated or manipulated images and video are marked as required by law and house policy","A versioned record of each asset's claims, sources and approvals is retained"],"humanInTheLoop":"Marketers own the message and edit every draft. A compliance reviewer approves every published asset and every material change, and product owners approve the fact base. The AI drafts and flags; it does not approve.","kpisToInstrument":["Time from brief to approved asset, before and after","Review rounds per asset and first time approval rate","Compliance flags per asset and share confirmed by reviewers","Post publication issues (withdrawn assets, complaints, regulator queries)","External agency and translation spend"],"failureModes":[{"title":"Invented facts","detail":"The model writes a rate, fee or benefit that is not in the approved facts. Insert numbers from the fact source and block publication if a number in the draft has no source."},{"title":"Rubber stamp review","detail":"Reviewers trust the automated first pass and skim. Keep reviewers accountable, sample assets the check passed, and show them what the check did and did not cover."},{"title":"Outdated disclosures","detail":"A rate or warning changes but the fact source does not. Give every fact an owner and effective date and expire stale ones."},{"title":"Volume without control","detail":"Faster drafting multiplies assets and variants beyond what compliance can review. Plan review capacity together with production."}]},"risk":{"euAiAct":{"tier":"limited","basis":"An internal drafting and review aid that makes no decisions about people. Article 50 transparency duties apply to generated content: providers must mark synthetic content, and deployers must disclose deep fake images, audio or video. Personalized targeting of individuals is governed mainly by data protection and consumer law rather than the AI Act."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","iso-42001","mifid-ii","eu-idd"],"guidance":[{"title":"Regulating financial promotions and adverts","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/firms/financial-promotions-adverts","note":"All financial promotions must be clear, fair and not misleading regardless of media type, which is the standard AI drafted promotions are reviewed against in the UK."},{"title":"RG 234 Advertising financial products and services (including credit)","issuer":"Australian Securities and Investments Commission","region":"asia-pacific","url":"https://asic.gov.au/regulatory-resources/find-a-document/regulatory-guides/rg-234-advertising-financial-products-and-services-including-credit-good-practice-guidance/","note":"Guidance, reissued in June 2026, that helps promoters and publishers of financial product and credit advertising avoid false or misleading statements, a useful basis for the compliance rule library."},{"title":"Article 50: Transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Covers marking of AI generated content and disclosure of deep fakes, relevant when campaigns use generated imagery, audio or video."}],"controls":["Approved product fact base with owners, effective dates and change control","Written rule library for the compliance first pass, reviewed by compliance","Named approval recorded for every published asset","Retention of each asset version with its claims, sources, flags and approvals","Inventory entry for the copilot with an accountable owner"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the copilot is two **AI agents**: a drafting agent and a separate compliance check\nagent, each with its own **prompt versioning**. Both draw on a **knowledge base** with the\napproved product facts, brand guide, disclosures and the compliance rule library, retrieved with\n**hybrid search**, and **custom functions** pull current rates and fees from the product system so\nnumbers are inserted, not generated. **Structured output** returns the draft, its claims with\nsources and the compliance flags with the rule each relies on.\n\nAn **agentic workflow** routes drafts to the compliance reviewer with **human in the loop\napproval**, and **run history with a full audit trail** keeps every version and decision.\nMarketers work in **Microsoft Teams** or the web; **machine translation** and multi language\nsupport cover localization, and **guardrails** block banned claims before a draft reaches a\nreviewer. **Test suites** replay past approved and rejected assets after every change, and the\nplatform is model agnostic, so drafting and checking can run on different models."},"faq":[{"question":"How much time does generative AI save in regulated marketing?","answer":"Ally reports an average time saving of 34% in a month long experiment with its marketers; it says the largest reductions, of up to two to three weeks, came in early stages such as research, first drafts and naming. Klarna reports cutting its image development cycle from six weeks to seven days, including brand and legal compliance checks. None of these sources measures review time separately, so track review rounds as well as drafting time."},{"question":"Can AI approve financial promotions?","answer":"No. It can check drafts against written rules and flag issues with the rule cited, but a named compliance reviewer must approve every published asset. Promotions must still be clear, fair and not misleading, and the firm remains accountable."},{"question":"How do you stop the model inventing rates or guarantees?","answer":"Keep product numbers in an approved fact source and insert them into the copy rather than letting the model write them, block any draft that contains a number without a source, and add banned claims such as guarantees to the compliance rules."}],"related":["personalized-marketing-at-scale","regulatory-horizon-scanning","policy-drafting-and-gap-analysis","investment-research-summarization","offers-and-rewards-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog, made industry neutral and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added MiFID II, Insurance Distribution Directive to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact check against sources: corrected the Ally claims so regulatory review is tied to the example the source describes, removed unsupported statements about where review time sits, updated the ASIC RG 234 note to the June 2026 reissue, fixed Klarna, JPMorgan Chase and Ally evidence details, and added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Second fact check: tied Ally's four hours to one figure to creating and editing only, attributed the early stage finding to Ally's reduction of up to two to three weeks rather than the 34% average, removed the unsourced claim about review time, and set the JPMorgan Chase record to production."}],"slug":"marketing-content-compliance-copilot","url":"https://www.blits.ai/ai-use-cases/marketing-content-compliance-copilot","benchmarks":[{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10000000,"min":10000000,"max":10000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"klarna-generative-ai-marketing-production","pooled":true}]},{"kpi":"cycle-time-days","label":"Cycle time","unit":"days","aggregate":false,"higherIsBetter":false,"n":1,"nUpTo":0,"median":7,"min":7,"max":7,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"klarna-generative-ai-marketing-production","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":34,"min":34,"max":34,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ally-financial-generative-ai-marketing-content","pooled":true}]}],"indicativeValueResult":{"low":74000,"high":510000},"evidence":["ally-financial-generative-ai-marketing-content","jpmorgan-chase-ai-marketing-copy","klarna-generative-ai-marketing-production"]},{"title":"AI copilot for model risk validation and monitoring","shortTitle":"Model risk validation","seoTitle":"AI for model risk validation and monitoring","metaDescription":"AI checks model documentation, tests generative AI with an LLM judge and drafts reports. See what the ECB runs and what UOB and Standard Chartered piloted.","definition":"A copilot for independent model validation and review, whether run by a bank's validation function, an external tester or a supervisor, that checks model documentation against the model risk standard, generates and scores challenger tests (for generative AI, often with an LLM as a judge calibrated against human experts), watches production models for drift and drafts and consistency checks the validation report. An accountable validator owns every conclusion.","aliases":["AI model validation assistant","model validation automation","model monitoring and drift detection","LLM evaluation for model risk","LLM as a judge for model validation","AI assisted model review"],"industries":["banking","insurance","capital-markets","wealth-and-asset-management"],"functions":["risk-management","regulatory-compliance"],"patterns":["agentic-workflow","document-processing","content-generation","anomaly-detection"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"second-line","problem":"Model inventories keep growing: credit, fraud, pricing, stress testing, anti money laundering and\nnow generative AI applications, each needing independent validation before use and periodic\nrevalidation after. Canada's OSFI describes a rapid rise in model applications, amplified by AI\nand machine learning. When validation capacity does not keep pace, backlogs build up and lower\nrisk models wait, or get reviewed with the same depth as critical ones.\n\nMuch validation effort is mechanical: checking that documentation covers every required section,\nrerunning the developer's tests, writing standard sections of the report and chasing monitoring\nresults. Generative AI adds a harder problem: testing open ended outputs at scale for accuracy,\nbias, leakage and robustness. When the US banking agencies replaced SR 11-7 in April 2026, they\nleft generative and agentic AI models outside the scope of the revised guidance because these\nmodels are novel and rapidly evolving, so banks must set those validation standards themselves.","problemStats":[{"statement":"Risk.net's 2026 Model Risk Benchmarking study of 44 banks found that banks are automating the testing of generative AI, but scope varies widely: LLM as judge testing offers model testing at scale, but few lenders use it to allow autonomous sign off.","sourceTitle":"Model Risk Benchmarking 2026 (Risk.net topic page, archived; the article itself is paywalled)","sourceUrl":"https://web.archive.org/web/20260921001318/https://www.risk.net/topics/model-risk-benchmarking-2026","year":2026},{"statement":"In the Bank of England and FCA 2024 survey, 46% of responding firms said they have only a partial understanding of the AI technologies they use, largely because of third party models.","sourceTitle":"Artificial intelligence in UK financial services 2024","sourceUrl":"https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024","year":2024}],"howItWorks":"1. **Intake.** The model owner submits the model, its documentation, data and code into the\n   validation workflow; the copilot records it against the inventory entry and risk tier.\n2. **Documentation check.** The copilot reads the documentation and checks it against every\n   requirement of the model risk standard, listing gaps with the missing section and the rule.\n3. **Challenger testing.** It generates test plans, edge cases and scenarios (for generative AI:\n   synthetic inputs, adversarial prompts, grounding and bias test sets) and runs them through\n   approved tooling. For generative AI outputs, an LLM as a judge scores each output against a\n   checklist derived from the requirements (hallucinations, contradictions, completeness, policy\n   compliance), and human experts score a sample on the same scale to calibrate the judge.\n4. **Ongoing monitoring.** For production models it tracks drift, stability and performance\n   against thresholds and flags models due for revalidation.\n5. **Draft the report.** It drafts the standard sections of the validation report with every\n   result linked to its evidence, and checks findings for consistency with earlier reports and\n   similar models.\n6. **Validator decides.** The validator reviews, challenges, adds findings and signs the\n   conclusion. The copilot never approves a model.","valueDrivers":["compliance","risk-reduction","employee-productivity","speed"],"kpis":["processing-time-reduction","productivity-gain","time-saved-per-task","accuracy"],"indicativeValue":{"referenceOrg":"A bank that completes 150 model validations and revalidations a year","inputs":[{"key":"validations","label":"Validations and revalidations per year","low":150,"high":150,"unit":"validations per year","note":"The reference bank. Replace with your own validation plan."},{"key":"hoursPerValidation","label":"Validator hours per validation","low":80,"high":200,"unit":"hours per validation","note":"Editorial assumption; depends heavily on model tier. Replace with your own records."},{"key":"timeSaved","label":"Share of validator time saved on documentation checks, testing and drafting","low":0.1,"high":0.25,"unit":"fraction of time","note":"Editorial assumption. No public measured benchmark of AI assisted validation was found; keep this conservative."},{"key":"hourlyCost","label":"Fully loaded cost of a model validator","low":90,"high":160,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"validations * hoursPerValidation * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Validator capacity released","caveat":"Capacity only. It leaves out the value of clearing the validation backlog sooner, catching drift earlier and the cost of validating the copilot itself, which is a model in the inventory."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The copilot touches the core of model governance, must preserve validation independence and is itself subject to model risk management. Integration with model development platforms, data and monitoring is substantial.","dataPrerequisites":["A model inventory with risk tiers, owners and validation history","The model risk standard and documentation templates in machine readable form","Access to model artefacts, test data and production monitoring metrics","Past validation reports and findings to test the copilot against"],"integrations":["Model inventory and model risk management platform","Model development and MLOps platforms (code, data, experiments)","Production monitoring and data quality tooling","Document management for validation reports and evidence"]},"implementation":{"steps":[{"title":"Start with documentation completeness","detail":"The lowest risk, highest volume task: check documentation against the standard and list gaps. Validators can confirm results quickly, which builds trust."},{"title":"Add monitoring triage","detail":"Summarise monitoring results across the inventory and flag models breaching thresholds or due for revalidation, so validators spend time where risk moved."},{"title":"Generate tests, run them in approved tooling","detail":"Let the copilot propose test plans and generative AI test sets, but execute them in the bank's validated tooling and keep the test set under version control."},{"title":"Draft report sections last","detail":"Only once checks and tests are trusted, draft standard report sections with links to evidence; conclusions and findings stay with the validator."},{"title":"Validate the validator","detail":"Register the copilot in the model inventory, validate it with a team not using it, and monitor its accuracy against validator decisions."}],"guardrails":["The accountable validator signs every conclusion; the copilot cannot approve, reject or tier a model","Every statement in a draft links to the test, data or document it rests on","The copilot is an inventoried model with its own validation and monitoring","Model developers cannot configure or prompt the copilot used by the validation team","Test sets and prompts are version controlled and reviewed"],"humanInTheLoop":"Independent validators own every test choice, finding and conclusion, and model risk committees approve models for use. The copilot prepares evidence and drafts; its outputs are reviewed like work from a junior validator.","kpisToInstrument":["Validation cycle time by model tier","Validation backlog and overdue revalidations","Documentation gaps found per model, and validator agreement with the copilot's gap list","Time from a monitoring breach to validator review","Share of drafted report text kept after validator review"],"failureModes":[{"title":"Loss of independence","detail":"The same AI setup helps build and validate a model, so its blind spots repeat. Separate configurations and owners for development and validation."},{"title":"False comfort from generated tests","detail":"Generated tests cover what is easy to test, not what matters. Validators must review coverage against the model's use and risks."},{"title":"Boilerplate reports","detail":"Drafted sections read well but miss model specific issues. Keep findings and conclusions validator written."},{"title":"Unvalidated copilot","detail":"The copilot is treated as a tool, not a model, and drifts unnoticed. Inventory, validate and monitor it."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"A validation copilot supports internal governance and is not itself an Annex III use, and its drafts are internal, so Article 50 transparency duties do not normally apply. It often helps validate models that are high risk under Annex III (point 5(b), creditworthiness and credit scoring of natural persons; point 5(c), life and health insurance pricing), and the testing and documentation it supports feed the provider obligations of Articles 9, 11 and 15."},"regulations":["eu-ai-act","mas-ai-risk-management","nist-ai-rmf","iso-42001","pra-ss1-23","solvency-ii"],"guidance":[{"title":"SR 26-2, Revised Guidance on Model Risk Management","issuer":"Federal Reserve, OCC and FDIC","region":"north-america","url":"https://www.federalreserve.gov/supervisionreg/srletters/SR2602a1.pdf","note":"Issued on 17 April 2026, it supersedes and replaces SR 11-7 and SR 21-8, with a risk based approach to model risk management tailored to each bank's model risk profile. Generative and agentic AI models are explicitly outside its scope; traditional and non generative AI models are covered."},{"title":"SS1/23, Model risk management principles for banks","issuer":"Prudential Regulation Authority","region":"europe","url":"https://www.bankofengland.co.uk/prudential-regulation/publication/2023/may/model-risk-management-principles-for-banks-ss","note":"The PRA's expectations for model risk management as a risk discipline in its own right, for UK banks with internal model approval; the PRA has also held a roundtable on model risk management for AI and machine learning."},{"title":"Guideline E-23, Model Risk Management","issuer":"Office of the Superintendent of Financial Institutions","region":"north-america","url":"https://www.osfi-bsif.gc.ca/en/guidance/guidance-library/guideline-e-23-model-risk-management-2027","note":"Canadian model risk management expectations for all federally regulated financial institutions, banks and insurers alike, explicitly covering AI and machine learning models. It takes effect on 1 May 2027."},{"title":"Artificial Intelligence Model Risk Management: observations from a thematic review","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Good practices observed in a mid 2024 thematic review of banks, including independent validation of higher risk AI before deployment, monitoring for data and model drift, and controls for generative AI."}],"controls":["Copilot registered in the model inventory with its own validation and monitoring","Segregation between development and validation configurations","Evidence links and reviewer sign off for every drafted report section","Periodic comparison of copilot outputs with validator decisions","Version control for test sets, prompts and model versions used"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** started per validation, with **custom functions**\nthat read model artefacts, monitoring metrics and test results from the bank's model platforms\nand a **knowledge base** holding the model risk standard, templates and past reports. The agent\nreturns **structured output**: a documentation gap list, a proposed test plan and draft report\nsections with evidence links.\n\nFor generative AI assistants and agents that run on Blits.ai, including the copilot itself,\n**test suites** with deterministic and LLM based grading, multi turn conversation sets and\n**monitors** with scheduled checks can serve as part of the challenger testing and ongoing\nmonitoring. **Human in the loop approval**, **prompt versioning**\nand per run audit trails keep the validator in control and the record complete, and because the\nplatform is model agnostic the validation team can use a different model from the developers."},"faq":[{"question":"Does using AI in validation undermine independence?","answer":"It can if the same tools and configurations are used to build and to validate a model. Keep the validation copilot under the validation team's control, validate it like any other model and leave every conclusion with an accountable validator."},{"question":"Who uses AI in model validation today?","answer":"Adoption is early and partial. Risk.net's 2026 Model Risk Benchmarking study of 44 banks found banks automating the testing of generative AI with widely varying scope, and few lenders using LLM as judge testing to allow autonomous sign off. ECB Banking Supervision runs Medusa, which it describes as an AI application for consistency checks of internal model assessment reports. In Singapore's Global AI Assurance Pilot, PwC tested generative AI tools at Standard Chartered and UOB with an LLM as a judge; none of these published a measured result of the AI assisted checks themselves."},{"question":"Can an LLM as a judge replace human validators?","answer":"No. In the Standard Chartered pilot the LLM judge let the tester score many outputs against the requirements, but human subject matter experts scored a subset on the same framework to check its calibration, and turning the requirements into judge test prompts took substantial technical effort. In the UOB pilot the judge checked each clause of an answer against retrieved passages of the source document, and scoring rested on ground truths that PwC built and UOB reviewed. Treat the judge as a model in its own right: validate it, version its prompts and keep conclusions with an accountable validator."},{"question":"Where should a validation team start?","answer":"With documentation completeness checks and monitoring triage, which are high volume and easy to verify, before moving to generated tests and drafted report sections."}],"related":["ai-model-inventory","alternative-data-credit-scoring","insurance-pricing-and-actuarial-copilot","continuous-controls-testing","market-abuse-surveillance-triage"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by the editor after the strict review's four blockers were fixed."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog. Draft because no named public deployment of AI assisted model validation could be verified."},{"date":"2026-09-25","note":"Consolidation pass: added PRA SS1/23, Solvency II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: replaced SR 11-7 guidance with SR 26-2 (April 2026), which supersedes it and excludes generative and agentic AI; removed the unverifiable 44 banks figure from the Risk.net stat; sourced the problem statement; added Annex III points and Articles 9, 11, 15 to the AI Act basis; tightened guidance notes and the Blits.ai build; added seoTitle and metaDescription. Stays draft: no public evidence yet."},{"date":"2026-09-27","note":"Removed SR 11-7 from the regulations, since SR 26-2 supersedes it; added the 1 May 2027 effective date of OSFI Guideline E-23; aligned the SR 26-2 note with the letter's wording. Stays draft: no public evidence yet."},{"date":"2026-09-27","note":"Added one evidence record: General Bank of Canada's use of the ValidMind platform to validate an internally built model (grade C, a vendor case study naming the customer), with a 70 percent validation time reduction and a 90 percent cost saving against project estimates. Updated the FAQ on named deployments to reflect it. Stays draft: one grade C record is below the bar for review status (two named organizations, or one of grade A or B)."},{"date":"2026-09-28","note":"Added three evidence records: ECB Banking Supervision's Medusa tool for drafting and consistency checks of internal model assessment reports (grade B, the ECB describing its own tool), and PwC's LLM as a judge testing of generative AI tools at Standard Chartered and UOB in the AI Verify Foundation's Global AI Assurance Pilot (grade C each, pilots). Widened the definition to cover external testers and supervisors, added LLM as a judge scoring and consistency checks to how it works, added two aliases, rewrote the FAQ on adoption and added one on LLM judges, and updated the meta description. Moved to review: four named organizations, one of them grade B."},{"date":"2026-09-28","note":"Review fixes: replaced the unverifiable \"almost one third of banks\" figure in the problem stat and FAQ with the headline and standfirst of Risk.net's 44 bank study, cited through an archived copy of its topic page because the article is paywalled; limited the calibration and prompt effort points in the LLM judge FAQ to the Standard Chartered pilot and described the UOB method separately; dropped the 70 and 90 percent metrics from the General Bank of Canada record, because the case study credits a workflow platform and vendor support rather than AI, and explained in the time saved note why the 70 percent is not used."}],"slug":"model-risk-validation-copilot","url":"https://www.blits.ai/ai-use-cases/model-risk-validation-copilot","benchmarks":[],"indicativeValueResult":{"low":108000,"high":1200000},"evidence":["european-central-bank-medusa-internal-model-reports","standard-chartered-pwc-llm-judge-genai-validation","uob-pwc-llm-judge-genai-chatbot-testing"]},{"title":"AI copilot for network operations centre fault triage","shortTitle":"NOC fault triage copilot","seoTitle":"AI for NOC alarm correlation and fault triage","metaDescription":"An AI NOC copilot groups network alarms into probable faults and proposes a fix. Bell Canada ranks issues by customer impact; Orange adopted alarm correlation.","definition":"AI in the network operations centre (NOC) that correlates alarms and performance data from radio, transport, core and fixed networks into a small number of probable faults, ranks them by customer impact, proposes the likely root cause and fix from runbooks, vendor documentation and past tickets, and routes the ticket to the right team, while an engineer decides what to change.","aliases":["NOC copilot","network alarm correlation","telecom network AIOps","network fault management AI","root cause analysis for telecom networks"],"industries":["telecommunications"],"functions":["network-operations","operations"],"patterns":["anomaly-detection","classification-and-routing","rag-knowledge-assistant","summarization","agentic-workflow"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"network","problem":"Operator networks are large and multi vendor: Vodafone's first anomaly detection deployment alone\ncovered more than 60,000 4G cells in Italy, and Orange Global Networks has thousands of IP routers in\n800 points of presence. Each element raises its own alarms, so one broken fibre or a failed power\nsupply produces a storm of alarms across radio, transport and core at the same time. NOC engineers\nspend the first part of every incident working out which alarms belong together, which domain owns\nthe problem and whether customers are affected at all, while the ticket queue keeps growing.\n\nThe knowledge that shortens a fault (vendor manuals, runbooks, the fix that worked last month) is\nspread over many tools that do not talk to each other. Telstra describes the multi vendor problem\ndirectly: disparate vendor platforms cannot talk to each other, which slows fault diagnosis. Some\ndegradations raise no alarm at all: Nokia, describing its work with KDDI, calls these \"silent\ncells\", which do not hurt service quality at once but eventually degrade the end user experience.\nThe result is long mean time to repair, engineers who chase noise, and customers who report faults\nbefore the operator sees them.","problemStats":[],"howItWorks":"1. **Collect and normalise.** Alarms, performance counters, logs, topology and inventory from every\n   vendor platform stream into one data layer, with the network topology as a graph.\n2. **Correlate.** Models group alarms that share a topological or temporal cause into one probable\n   incident, so the NOC sees one problem instead of hundreds of alarms. Anomaly detection adds\n   degradations that raised no alarm.\n3. **Rank by customer impact.** The incident is scored with traffic, affected services and\n   customers, so a small fault on a busy site outranks a large one on an idle site.\n4. **Explain and propose.** A language model assistant summarises the evidence, retrieves matching\n   runbook steps, vendor documentation and similar past tickets, and proposes the likely root\n   cause and the next diagnostic or corrective step, with its sources.\n5. **Route and track.** The ticket goes to the owning team (radio, transport, core, field) with the\n   correlated evidence attached; the engineer approves any change, and the outcome is fed back to\n   improve correlation and ranking.","valueDrivers":["employee-productivity","speed","customer-experience","cost-to-serve"],"kpis":["alert-volume-reduction","processing-time-reduction","productivity-gain","interactions-handled"],"indicativeValue":{"referenceOrg":"A mobile and fixed operator with a 24 hour NOC handling 100,000 network incidents a year","inputs":[{"key":"incidents","label":"Network incidents and trouble tickets triaged by the NOC per year","low":60000,"high":150000,"unit":"incidents per year","note":"Editorial assumption for a national operator. Replace with your own ticket volume."},{"key":"triageMinutes","label":"Engineer minutes spent on correlation and diagnosis per incident","low":30,"high":60,"unit":"minutes per incident","note":"Editorial assumption. Replace with a time study of your own NOC."},{"key":"timeSaved","label":"Share of triage time saved by correlation and assisted diagnosis","low":0.2,"high":0.4,"unit":"fraction of triage time","note":"Editorial assumption. Orange expects topology based correlation to cut the daily alarms its NOC has to address by 70%, but fewer alarms do not translate one to one into less engineer time, so the range stays well below that."},{"key":"hourlyCost","label":"Fully loaded cost of a NOC engineer hour","low":50,"high":90,"unit":"USD per hour","note":"Editorial assumption. Replace with your own fully loaded cost."}],"formula":"incidents * triageMinutes / 60 * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"NOC engineering time released","caveat":"Engineering time only. It leaves out the larger effect of shorter outages on customer experience, churn and service level penalties, the contacts avoided when faults are fixed before customers call, and the cost of the data platform and integration work."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The model is the easy part. The work is in getting clean, timely alarm, performance, topology and inventory data out of many vendor systems, and in earning the trust of engineers who have seen many correlation tools promise more than they delivered.","dataPrerequisites":["Alarm and event streams from every vendor element manager and OSS","Performance counters per cell, link and node at a useful granularity","An accurate network topology and inventory, ideally as a graph","Historical trouble tickets with root cause and resolution codes","Runbooks and vendor documentation in a searchable form"],"integrations":["Fault management and OSS platforms per domain (radio, transport, core, fixed access)","Performance management and network data lake","Network inventory and topology systems","Trouble ticketing and workforce management","Collaboration tools used by the NOC for incident communication"]},"implementation":{"steps":[{"title":"Start with one domain and one pain","detail":"Pick the domain with the worst alarm noise (often radio access) and measure today's alarms per real incident, time to diagnose and repeat tickets, so you have a baseline."},{"title":"Fix the data before the model","detail":"Get topology and inventory right and stream alarms in near real time. Correlation built on a wrong topology produces confident nonsense."},{"title":"Run in shadow mode","detail":"Let the system group alarms and propose root causes next to the existing process for several weeks, and have senior engineers grade each proposal before anyone relies on it."},{"title":"Ground the assistant in your own knowledge","detail":"Load runbooks, vendor manuals and resolved tickets into retrieval with owners and review dates, and make the assistant cite its sources and say when it does not know."},{"title":"Put suppression under change control","detail":"Every rule or model that hides alarms from engineers is a risk. Version it, test it against past major incidents and review it like any network change."},{"title":"Widen to cross domain incidents","detail":"Once radio works, add transport and core so the system can trace a customer impact to its real cause across domains."}],"guardrails":["The copilot proposes; engineers approve every configuration change and every ticket closure","Suppressed alarms remain visible on request and are sampled for review every week","Answers cite the runbook, document or past ticket they come from, with a refusal when nothing matches","Read only access to network elements for the assistant, with any write actions routed through existing change management","Major incidents always trigger the normal human escalation path, whatever the model says"],"humanInTheLoop":"NOC engineers own diagnosis and every change to the network. The copilot correlates, ranks and drafts; engineers accept, correct or reject its proposals, and their corrections are the training signal. Senior engineers review suppression rules and a sample of closed incidents every week.","kpisToInstrument":["Alarms per actionable incident, before and after correlation","Mean time to detect and mean time to repair per domain","Share of root cause proposals accepted by engineers","Incidents first reported by customers rather than detected by the NOC","Missed incidents, where a suppressed or low ranked alarm turned out to matter"],"failureModes":[{"title":"Suppression that hides a real outage","detail":"A correlation rule groups a new failure under an old pattern and it never reaches an engineer. Sample suppressed alarms and replay past major incidents on every model change."},{"title":"Stale topology","detail":"Correlation depends on knowing what is connected to what. When inventory lags behind the network, root cause proposals point at the wrong element."},{"title":"Fluent but wrong diagnosis","detail":"A language model explains a fault convincingly from an outdated runbook. Require citations and track acceptance rates per runbook."},{"title":"Tool nobody opens","detail":"The copilot lives in yet another screen. Put its output inside the ticket and the NOC wall, not beside them."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The main test is Annex III point 2, which lists AI systems intended as safety components in the management and operation of critical digital infrastructure as high risk. A copilot that prepares diagnoses for engineers who decide every change is normally not such a safety component, and is then minimal risk. The tier rises when the system is designed to protect the safe operation of the network, for example by acting on it automatically to prevent or contain outages. Article 6(3) can exempt an Annex III system that only performs a preparatory task to an assessment and poses no significant risk of harm, provided the provider documents that assessment and registers the system."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","nis2"],"guidance":[{"title":"Regulation (EU) 2024/1689 (AI Act), Annex III high risk AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng","note":"Point 2 covers AI systems intended as safety components in the management and operation of critical digital infrastructure."},{"title":"Directive (EU) 2022/2555 (NIS2 Directive)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2022/2555/oj/eng","note":"Under Article 2(2), providers of public electronic communications networks and services fall under NIS2 regardless of their size, with risk management and significant incident reporting duties that the NOC tooling supports and must not weaken."}],"controls":["Inventory entry for the copilot with an accountable owner and documented data sources","Change control and regression tests on correlation and suppression rules","Audit trail of every proposal, the engineer's decision and the outcome","Weekly sampling of suppressed alarms and closed incidents by senior engineers"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the NOC copilot is an **AI agent** available in the engineers' tools (for example\n**Microsoft Teams** or an internal portal through the **REST API**). A **knowledge base** holds\nrunbooks, vendor manuals and resolved tickets with hybrid retrieval, and a **SQL knowledge base**\nlets the agent query incident and performance tables in plain language. **Custom functions** read\nthe correlated incident, topology and ticket history from the operator's OSS and ticketing APIs.\n\nActions that change anything, such as reassigning or closing a ticket, run as **agentic\nworkflows** with **human in the loop approval**, and a **tool execution policy** keeps the agent\nlimited to read only tools on network systems. Output **guardrails** check answers against a\npolicy that requires a cited source, **test suites** rerun questions built from past incidents on every change,\nscheduled **monitors** check the agent's answers, and the platform is model agnostic with EU and\nUAE data residency options."},"faq":[{"question":"How much alarm noise can AI correlation remove?","answer":"It depends on the network and the quality of topology data. After a two year production trial, Orange said topology based auto correlation from Augtera will cut the daily alarms its NOC has to address by 70%, stated as the expected effect of the rollout rather than a measured result. Measure your own alarms per real incident before and after, rather than relying on that figure."},{"question":"Does a generative AI copilot replace the correlation engine?","answer":"No. Correlation and anomaly detection remain statistical and topology based. The language model sits on top: it explains the incident, finds the relevant runbook and past tickets, and drafts the next step. Bell's setup follows this pattern: custom AI and machine learning models correlate network data with customer experience to prioritise issues, and Gemini models support incident analysis, historical context retrieval and access to vendor documentation."},{"question":"Is a NOC copilot high risk under the EU AI Act?","answer":"Usually not while engineers make every change. It can become high risk under Annex III point 2 when the system itself acts as a safety component in operating critical digital infrastructure, so the design choice about autonomy decides the tier."}],"related":["aiops-incident-triage","autonomous-network-operations","predictive-network-maintenance","network-outage-communication-agent","field-technician-copilot-and-dispatch"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched for the telecommunications vertical and verified against operator and vendor sources."},{"date":"2026-09-25","note":"Consolidation pass: added NIS2, European Electronic Communications Code to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: Orange's 70% alarm reduction reclassified as an expected effect, not a measured result (FAQ, value note and KPI list updated to alert volume reduction); problem statement tied to cited network sizes and the KDDI silent cell wording corrected; EU AI Act basis refined with Article 6(3); EECC removed because NIS2 deleted its security Articles 40 and 41; Vodafone rollout described as planned; SEO title and description added."},{"date":"2026-09-27","note":"Regulation corrected against EUR-Lex: NIS2 applies to public electronic communications providers regardless of size; the Article 6(3) derogation restated with its conditions, with Annex III point 2 as the main test; AI Act and NIS2 guidance now link to EUR-Lex; Bell FAQ wording aligned with the release."},{"date":"2026-09-27","note":"Fact checked against sources again (Bell Canada, Orange, Nokia for Vodafone and KDDI, Telstra, Deutsche Telekom, EUR-Lex): meta description now names Bell Canada; test suite wording in the Blits.ai section matched to the feature inventory; Gemini recorded as model provider on the Bell Canada record."}],"slug":"network-fault-triage-copilot","url":"https://www.blits.ai/ai-use-cases/network-fault-triage-copilot","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":95,"min":95,"max":95,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"deutsche-telekom-ran-guardian-and-mindr-agents","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":100,"min":100,"max":100,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"deutsche-telekom-ran-guardian-and-mindr-agents","pooled":true}]}],"indicativeValueResult":{"low":300000,"high":5400000},"evidence":["bell-canada-network-ai-ops","deutsche-telekom-ran-guardian-and-mindr-agents","kddi-nokia-performance-degradation-detection","orange-augtera-noc-alarm-correlation","telstra-self-healing-network-proof-of-concept","vodafone-nokia-network-anomaly-detection"]},{"title":"AI copilot for plant operators and maintenance technicians","shortTitle":"Plant operator and maintenance copilot","seoTitle":"AI maintenance assistant for plant technicians","metaDescription":"An AI assistant answers questions on machine faults from manuals, fault logs and sensor data. BMW, Georgia-Pacific and Textron Aviation run one for technicians.","definition":"A generative AI assistant for the people who run and repair machines in plants, workshops and service centres: it answers fault and procedure questions from equipment manuals, fault reports, shift logs and live machine data, in the technician's language, with links to the sources, so faults are diagnosed faster and expert knowledge is not lost when experienced staff retire.","aliases":["industrial copilot","shop floor AI assistant","maintenance troubleshooting assistant","operator assistant for manufacturing","aircraft maintenance AI assistant"],"industries":["manufacturing","automotive"],"functions":["operations","knowledge-management"],"patterns":["rag-knowledge-assistant","conversational-agent","summarization","translation"],"channels":["internal-tools","mobile-app"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"production","problem":"When a machine stops, every minute counts, and the answer usually exists somewhere: in\ndocumentation that can run to tens of thousands of pages (Textron Aviation counts more than\n60,000 pages for over 50 aircraft models), in last month's shift log, in a fault report from\nanother plant, or in the head of a technician who is on a different shift. Junior operators and\ntechnicians spend their time searching binders and portals or phoning the one expert who knows,\nwhile the line waits.\n\nThe knowledge is also leaving. Georgia-Pacific describes equipment that is 50 years old and, for\nmany machines, no proper documentation of operating procedures: the know how sits with\nexperienced employees, and they are retiring. New plants add a language problem: BMW notes that manuals are often\nnot available in Hungarian at its Debrecen plant. Keyword search over document portals does not\nsolve this, because the question is phrased as a symptom, not as a document title.","problemStats":[],"howItWorks":"1. **Gather the knowledge.** Equipment manuals, work instructions, fault reports, maintenance\n   records and shift logs are loaded and refreshed daily; interviews with retiring experts can be\n   recorded and turned into procedure documents.\n2. **Ask in plain language.** The operator or technician describes the symptom or error code on a\n   tablet, phone or line terminal, in their own language.\n3. **Retrieve and combine.** The assistant finds the relevant passages and, where connected, reads\n   the machine's current state and recent trends from sensor data.\n4. **Answer with sources.** It summarises the likely causes and the steps to check, with links to\n   the exact manual page or video frame, and answers follow up questions.\n5. **Hand over when needed.** Safety critical work, lockout procedures and anything the sources do\n   not cover go to the responsible engineer, and confirmed fixes are written back as new knowledge.","valueDrivers":["employee-productivity","speed","risk-reduction","cost-to-serve"],"kpis":["time-saved-per-task","search-time-reduction","mttr-reduction","first-time-fix-rate","employee-adoption","time-to-proficiency-reduction","accuracy"],"indicativeValue":{"referenceOrg":"A manufacturing site with 300 operators and maintenance technicians","inputs":[{"key":"staff","label":"Operators and technicians who use the assistant","low":300,"high":300,"unit":"people","note":"The reference site."},{"key":"lookupsPerWeek","label":"Troubleshooting or procedure lookups per person per week","low":3,"high":5,"unit":"lookups per person per week","note":"Editorial assumption, replace with your own estimate from help desk calls or a short time study."},{"key":"weeks","label":"Working weeks per year","low":46,"high":46,"unit":"weeks","note":"Editorial assumption."},{"key":"minutesSaved","label":"Minutes saved per lookup","low":5,"high":15,"unit":"minutes","note":"Conservative against the benchmark on this page (Microsoft reports that at Textron Aviation troubleshooting that took up to 20 minutes takes one to two minutes), because many lookups are shorter than a full troubleshooting search."},{"key":"hourlyCost","label":"Fully loaded cost per technician hour","low":40,"high":60,"unit":"USD per hour","note":"Editorial assumption, replace with your own labour cost."}],"formula":"staff * lookupsPerWeek * weeks * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Technician time released","caveat":"Values technician time only. It leaves out the usually larger value of shorter machine downtime and less off quality production, the cost of preparing and maintaining the content, and licence and model costs."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"A first version over manuals and fault reports is quick to build. The effort is in document quality and ownership per site, in connecting live machine data safely, and in keeping answers about safety procedures correct.","dataPrerequisites":["Equipment manuals and work instructions per machine and site, with owners","Fault reports, maintenance records and shift logs in machine readable form","An asset register that links machines, documents and sensor tags","Optionally, recorded interviews with experienced staff"],"integrations":["Document management and maintenance systems (enterprise asset management)","Plant historian or IoT platform for live machine data","Identity and access management for plant staff","Tablets, phones or line terminals on the shop floor"]},"implementation":{"steps":[{"title":"Start where downtime is expensive and knowledge is thin","detail":"Pick one line or equipment family with frequent faults, many new staff and scattered documentation, and collect the questions technicians actually ask."},{"title":"Clean and own the sources","detail":"Load manuals, work instructions and recent fault reports with an owner and review date per document, remove outdated versions and mark safety critical procedures."},{"title":"Test with your best technicians","detail":"Let senior technicians put their hardest questions to it and score the answers, as Textron Aviation did before scaling, and fix the gaps in the content rather than the prompt."},{"title":"Add live machine data carefully","detail":"Connect sensor data for the assets in scope so answers can reflect the machine's current state and recent trends, as Georgia-Pacific does. Keep the connection read only, so the assistant has no write access to controls."},{"title":"Roll out plant by plant with local content","detail":"Reuse the platform, add each plant's documents and languages, and measure adoption and time to repair per site."}],"guardrails":["Answers only from approved sources, with a link to the source and a refusal when none is found","Safety critical procedures (lockout, high voltage, confined spaces) quoted from the approved document, never paraphrased","Read only access to machine data; no control actions from the assistant","Access by role and plant, so confidential process data stays with the right people","In regulated maintenance such as aviation, the assistant points to the approved maintenance data and never replaces it as the basis for the work"],"humanInTheLoop":"The technician decides and does the work; the assistant only informs. Maintenance engineers own the content per equipment family, review answers that technicians flag as wrong, and approve new documents before they are loaded. Safety procedures stay under the plant's safety management, and in aviation the work is still signed off against the approved maintenance data.","kpisToInstrument":["Mean time to repair on equipment in scope, before and after","Time to find an answer, from telemetry or a time study","Share of answers rated helpful, and flagged wrong answers per week","Weekly active users among operators and technicians","Weeks for new technicians to work independently"],"failureModes":[{"title":"Outdated or conflicting manuals","detail":"The assistant faithfully quotes an old revision. Keep one current version per document and retire the rest."},{"title":"A paraphrased safety step","detail":"A summarised lockout procedure drops a step. Quote safety procedures verbatim and link the source."},{"title":"Pilot that never reaches the second plant","detail":"Each plant builds its own tool. BMW consolidated parallel plant pilots into one company wide application; plan for a shared platform early."},{"title":"Adoption stalls on the shop floor","detail":"The assistant is only available on office PCs. Put it on the devices technicians carry and in their language."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1): staff must know they are interacting with an AI system, unless that is obvious from the context. Answering maintenance questions is not an Annex III use. It would become high risk under Annex III point 4(b) if the usage data were used to monitor and evaluate the performance of individual workers, so keep usage analytics aggregated."},"regulations":["eu-ai-act","gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) covers AI used to monitor and evaluate the performance and behaviour of workers."},{"title":"AI Risk Management Framework","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/itl/ai-risk-management-framework","note":"Voluntary framework for mapping and managing AI risk, including the reliability of generated answers."}],"controls":["Document ownership, review dates and version control for every source","Logging of questions, retrieved sources and answers for review","Role based access by plant and function","Regression tests with real technician questions before each content or model change"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** over a **knowledge base** that ingests manuals, work\ninstructions and fault reports (PDF, Word, spreadsheets and images) with version control, and\nretrieves with hybrid search so exact error codes and part numbers are found as well as\nsymptoms. A **SQL knowledge base** or read only **custom functions** add live machine data and\nwork order history, and **multi language** support with automatic language detection lets each\ntechnician ask in their own language.\n\nTechnicians reach the agent in **Microsoft Teams**, the web widget on a tablet or through the\nREST API channel inside existing shop floor apps, with voice input for hands busy work.\n**Guardrails** check answers against the plant's policies, **human handover** passes open\nquestions to a person on the responsible team, **test suites** replay real technician questions on every change and\nanalytics show which questions go unanswered. The platform is model agnostic, with EU and UAE\ndata residency."},"faq":[{"question":"Who uses generative AI assistants for maintenance?","answer":"BMW made its Factory Genius assistant available across its plants in 2025, Georgia-Pacific runs ChatGP for machine operators on Amazon Bedrock with live sensor data, and Textron Aviation built TAMI for aircraft technicians on Azure OpenAI Service."},{"question":"How much time does it save?","answer":"Microsoft reports that at Textron Aviation troubleshooting that took up to 20 minutes now takes one to two minutes, the benchmark recorded on this page. Few other figures are public, so measure time to repair and time to find an answer on your own equipment before and after."},{"question":"How is this different from a field service copilot?","answer":"A field service copilot supports technicians who visit customer sites and includes dispatch. This assistant serves operators and maintenance staff on machines inside a plant, workshop or service centre, and often combines documents with live machine data."}],"related":["field-technician-copilot-and-dispatch","industrial-asset-predictive-maintenance","enterprise-knowledge-search"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the discovery workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the manufacturing and automotive scope with BMW Group, Georgia-Pacific and Textron Aviation evidence checked against the sources."},{"date":"2026-09-27","note":"Editor pass. Recorded the Textron Aviation time saved as a metric, attributed the Georgia-Pacific outcome to AWS, removed an unsourced page count, the energy and utilities industry tag and an unsourced read only claim, and added a note on approved maintenance data in aviation."}],"slug":"plant-operator-and-maintenance-copilot","url":"https://www.blits.ai/ai-use-cases/plant-operator-and-maintenance-copilot","benchmarks":[{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"textron-aviation-tami-maintenance-assistant","pooled":false}]}],"indicativeValueResult":{"low":138000,"high":1035000},"evidence":["bmw-group-factory-genius-maintenance-assistant","georgia-pacific-chatgp-operator-assistant","textron-aviation-tami-maintenance-assistant"]},{"title":"AI copilot for SAR and STR narrative drafting","shortTitle":"SAR and STR drafting","seoTitle":"AI copilot for SAR and STR narrative drafting","metaDescription":"Generative AI drafts SAR and STR narratives from the case file for investigators to verify. At Nexo, Unit21 reports 57% of alert reviews automated by AI agents.","definition":"Generative AI that drafts the narrative of a single suspicious activity or suspicious transaction report from the investigation file (who, what, when, where, why and how), with every fact linked to its source record, so the investigator verifies, edits and files instead of starting from a blank page. It works case by case, unlike the periodic data returns of regulatory reporting.","aliases":["SAR narrative generation","STR drafting assistant","suspicious transaction report copilot","AML investigation summary"],"industries":["banking","payments"],"functions":["financial-crime-compliance","case-management"],"patterns":["content-generation","summarization","rag-knowledge-assistant","agentic-workflow"],"channels":["internal-tools","agent-desktop"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"middle-office","problem":"When an investigation concludes that activity is suspicious, the investigator must write a\nreport to the financial intelligence unit. The narrative is the only free text part of the\nreport, and it has to explain who was involved, which accounts and transactions, over what\nperiod, what typology it resembles and why it is suspicious, clearly and completely. FinCEN's\nnarrative guidance warns that incomplete, incorrect or disorganized narratives make further\nanalysis by law enforcement difficult, if not impossible.\n\nWriting it is slow. Investigators copy transaction tables, reconstruct timelines from statements\nand case notes, and write prose under deadline pressure, since reports must be filed within a\nfixed period after detection (in the United States, as FinCEN's guidance restates, no later than\n30 calendar days after initial detection). Quality varies between investigators, and quality assurance\nteams send drafts back for missing facts or unclear reasoning. Time spent writing is time not\nspent investigating.","problemStats":[],"howItWorks":"1. **Assemble the facts.** The agent gathers the case file: alerts, customer due diligence,\n   transactions, counterparties, previous reports, investigator notes and any external requests.\n2. **Build the timeline.** It orders the relevant transactions and events and computes totals,\n   date ranges and counterparties, using data queries rather than the language model for numbers.\n3. **Draft the narrative.** It writes the narrative in the structure the financial intelligence\n   unit expects, with each statement referencing the record it comes from.\n4. **Check completeness.** It checks the draft against the filing guidance (subjects,\n   instruments, dates, amounts, locations, reason for suspicion) and flags anything missing.\n5. **Investigator attests.** The investigator verifies each fact, edits the reasoning, and files.\n   The draft, the edits and the final version are retained.","valueDrivers":["employee-productivity","compliance","speed"],"kpis":["time-saved-per-task","processing-time-reduction","productivity-gain","error-reduction","hours-saved"],"indicativeValue":{"referenceOrg":"A bank filing 5,000 suspicious activity reports a year","inputs":[{"key":"reports","label":"Reports filed per year","low":5000,"high":5000,"unit":"reports per year","note":"The reference bank."},{"key":"hoursPerNarrative","label":"Investigator hours spent drafting each narrative today","low":1.5,"high":4,"unit":"hours per report","note":"Editorial assumption. Replace with your own time study."},{"key":"timeSaved","label":"Share of drafting time saved","low":0.3,"high":0.5,"unit":"fraction of drafting time","note":"Editorial assumption. Verification and editing time remain with the investigator, and public results so far are vendor reported and measure alert review rather than drafting alone (for example Uphold's 44% faster median alert review in a pilot); replace with results from your own pilot."},{"key":"costPerHour","label":"Fully loaded investigator cost per hour","low":45,"high":80,"unit":"USD per hour","note":"Editorial assumption. Replace with your own."}],"formula":"reports * hoursPerNarrative * timeSaved * costPerHour","currency":"USD","period":"per year","resultLabel":"Investigator capacity released","caveat":"Counts drafting time only. It leaves out quality assurance rework avoided, better reports for law enforcement, and the cost of validating and running the drafting system."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Drafting from a well structured case file is within reach of current models. The work is in grounding every fact, computing numbers outside the model, protecting highly sensitive data and keeping the investigator accountable for the content.","dataPrerequisites":["Structured case files with transactions, subjects and investigator notes","Examples of high quality narratives approved by quality assurance","The financial intelligence unit's filing guidance and field definitions","Access controls that match SAR confidentiality requirements"],"integrations":["AML case management system","Transaction and customer data stores","Financial intelligence unit filing system or e filing format","Quality assurance workflow"]},"implementation":{"steps":[{"title":"Start from the filing guidance","detail":"Turn the financial intelligence unit's guidance and your quality assurance checklist into a structure and a completeness check the draft must satisfy."},{"title":"Compute, then write","detail":"Calculate totals, date ranges and counts with queries, and give them to the model as facts. Never let the model do arithmetic in the narrative."},{"title":"Require citations","detail":"Make every statement in the draft reference a record in the case file, and show the references to the investigator in the editing view."},{"title":"Measure against quality assurance","detail":"Compare drafts, edited versions and quality assurance outcomes on a sample of cases before rolling out, and track edit distance and rework rates afterwards."},{"title":"Lock down confidentiality","detail":"Run the model in an environment approved for SAR data, with no retention by third parties and access limited to the investigation team."}],"guardrails":["The investigator verifies every fact and is the named author of the filed report","Numbers come from data queries, not from the language model","Every statement references a source record; unsupported statements are flagged","SAR confidentiality, with access limited to the investigation team and no data retained by model providers","Draft, edits and final version retained for audit"],"humanInTheLoop":"The system drafts; the investigator decides whether to file, verifies the facts, rewrites the reasoning where needed and attests. Quality assurance reviews a sample as today, and the money laundering reporting officer owns the use of the tool.","kpisToInstrument":["Drafting time per report, before and after","Quality assurance rework rate and reasons","Share of draft statements changed or removed by investigators","Completeness check failures per draft","Time from case conclusion to filing"],"failureModes":[{"title":"Confident errors in facts","detail":"A wrong amount or date in a filed report undermines the bank and the investigation. Compute numbers outside the model and require citations."},{"title":"Boilerplate reasoning","detail":"Drafts converge on generic typology language that tells law enforcement little. Measure how much investigators rewrite the reasoning section and coach the prompts on good examples."},{"title":"Automation of the decision","detail":"The draft makes filing look like the default. Keep the decision to file separate from the drafting step and record it explicitly."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Drafting internal reports for a human investigator is not listed in Annex III (the law enforcement uses in point 6 cover systems used by or for law enforcement authorities, not a bank's own reporting), and the text is not published to inform the public, so the deployer disclosure duty for generated text in Article 50(4) does not apply. Confidentiality rules for suspicious activity reports and GDPR apply in full."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","dora","mas-ai-risk-management","nist-ai-rmf","iso-42001","eu-amlr","us-bsa","mas-notice-626"],"guidance":[{"title":"Guidance on Preparing a Complete and Sufficient Suspicious Activity Report Narrative","issuer":"Financial Crimes Enforcement Network (FinCEN)","region":"north-america","url":"https://www.fincen.gov/system/files/shared/sarnarrcompletguidfinal_112003.pdf","note":"Guidance of November 2003 that explains the five essential elements a narrative must cover (who, what, when, where and why, plus how) and the 30 day filing deadline, with examples of sufficient and insufficient narratives; a useful completeness checklist for any drafting tool."},{"title":"Supporting Artificial Intelligence Adoption in AML/CFT","issuer":"Hong Kong Monetary Authority","region":"asia-pacific","url":"https://brdr.hkma.gov.hk/eng/doc-ldg/docId/getPdf/20251118-3-EN/20251118-3-EN.pdf","note":"Circular of 19 November 2025. It reports that more than 30% of authorized institutions already use AI in transaction monitoring and announces supervisory workshops that include the use of generative AI to compile suspicious transaction reports."},{"title":"Joint Statement Encouraging Innovative Industry Approaches to AML Compliance","issuer":"FinCEN and the US federal banking agencies","region":"north-america","url":"https://www.fincen.gov/news/news-releases/treasurys-fincen-and-federal-banking-agencies-issue-joint-statement-encouraging","note":"Statement of 3 December 2018. Innovative pilot programs should not in themselves subject banks to supervisory criticism, even if they ultimately prove unsuccessful."}],"controls":["Named investigator attests every filed report","Environment and model provider terms approved for SAR confidentiality","Citation and completeness checks on every draft","Retention of draft, edits and final version","Inventory entry with owner, prompts under change control and periodic quality review"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the drafting runs as an **agentic workflow**, triggered from the case management\nsystem through the API. The agent reads the case through **custom functions** and **SQL\nknowledge bases**, so totals, dates and counts come from queries, and it follows the filing\nguidance and approved example narratives stored in a **knowledge base** with hybrid retrieval.\nIt returns **structured output**: the narrative with record references, and a completeness\nchecklist.\n\nThe investigator reviews the draft and approves or rejects it through **human in the loop\napproval** before anything goes back to the case system, where the investigator edits and\nfiles; every run keeps a full audit trail. **PII masking**\nand **tenant isolation** protect the case data, **prompt versioning** and **test suites** keep\nchanges controlled, and the platform is model agnostic with EU and UAE data residency, so the\nbank can choose a model and region its SAR confidentiality rules allow."},"faq":[{"question":"Can generative AI write a suspicious activity report?","answer":"It can draft the narrative from the case file, but the investigator remains responsible for the decision to file and for every fact in it. Treat it as a drafting copilot with citations, not as an adjudicator."},{"question":"Do regulators allow AI drafted SAR narratives?","answer":"No rule we know of forbids them, and none removes the filer's accountability. In November 2025 the Hong Kong Monetary Authority announced workshops on using generative AI to compile suspicious transaction reports, and in 2018 FinCEN and the US federal banking agencies encouraged innovative approaches to AML compliance, including pilot programs."},{"question":"Who is already using AI to write investigation narratives?","answer":"Mostly fintech, crypto and payment firms so far, through vendor platforms. Unit21 reports that by automating alert narratives and dispositions its AI agents automate 57% of alert reviews at Nexo, and FIS announced in May 2026 that BMO and Amalgamated Bank are developing with its Financial Crimes AI Agent. Public results from large banks are still scarce."},{"question":"How do you stop the model inventing facts?","answer":"Compute every number with data queries, require a source reference for every statement, flag anything unsupported, and have the investigator verify the facts before filing."}],"related":["aml-alert-triage","mule-network-detection","regulatory-report-assembly","trade-finance-crime-screening","market-abuse-surveillance-triage"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added the FinCEN 30 day filing deadline, dated the guidance and the FAQ references, made the EU AI Act basis precise (Annex III point 6, Article 50(4)), updated the FinCEN guidance URL, corrected evidence years and descriptions; added seoTitle and metaDescription."},{"date":"2026-09-26","note":"Second fact check against sources: the problem section now cites FinCEN's guidance on why narrative quality matters, the Nexo wording follows the Unit21 case study exactly, the meta description no longer overstates the Nexo result, and the Blits.ai section describes human in the loop approval accurately."}],"slug":"suspicious-activity-report-drafting","url":"https://www.blits.ai/ai-use-cases/suspicious-activity-report-drafting","benchmarks":[],"indicativeValueResult":{"low":101250,"high":800000},"evidence":["finshark-lucinity-luci-investigations","fis-financial-crimes-ai-agent","nexo-unit21-ai-alert-narratives","uphold-unit21-ai-agent-pilot"]},{"title":"AI copilot for underwriting risk assessment","shortTitle":"Underwriting risk assessment copilot","seoTitle":"AI underwriting copilot for risk assessment","metaDescription":"An underwriting copilot drafts the risk narrative the underwriter signs. Zurich North America underwriters report saving 2 hours per submission on average.","definition":"A copilot that assembles everything relevant to a risk (the submission, loss history, internal guidelines, third party data and public information), highlights exposures and gaps against the insurer's underwriting guidelines and drafts the underwriting narrative or referral note, while the underwriter makes and signs every decision.","aliases":["underwriting copilot","underwriting assistant","AI underwriting workbench"],"industries":["insurance"],"functions":["underwriting","risk-management"],"patterns":["rag-knowledge-assistant","summarization","content-generation","agentic-workflow"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"underwriting","problem":"Once a submission is in appetite, the real work starts. A middle market or specialty underwriter\nreads hundreds of pages of operations descriptions, loss runs and supplementals, searches the web\nand internal systems for anything the broker did not mention, checks the risk against underwriting\nguidelines and writes a narrative that justifies the decision to a referral authority, an auditor or\na reinsurer.\n\nMuch of that time goes into finding and restating information rather than judging it, and the\nquality depends on how thorough each underwriter is on a busy day. Exposures buried in a website or\na court filing are easy to miss, narratives vary in structure, and experienced underwriters spend\nhours on write ups instead of broker relationships and complex risks.","problemStats":[{"statement":"A Cytora case study on Markel UK reports that, before the project, Markel's underwriters spent more than 30% of their time on low skill, low value tasks such as rekeying risk data into different systems and pulling third party data by hand.","sourceTitle":"Markel uses Cytora and achieves +100% productivity uplift to fuel growth","sourceUrl":"https://www.cytora.com/risk-flow-center/blog/case-study-markel-records-113-productivity-increase-in-its-underwriting-team-following-cytora-partnership","year":2023}],"howItWorks":"1. **Assemble the file.** The copilot gathers the extracted submission, prior policies and claims,\n   internal notes and approved external data for the insured into one view.\n2. **Research the insured.** An agent searches approved sources (company website, news, court and\n   regulatory records) and summarizes what is relevant to the line of business, with links.\n3. **Check against guidelines.** Retrieval over the insurer's underwriting guidelines and referral\n   rules flags where the risk falls outside authority, where information is missing and which\n   questions to ask the broker.\n4. **Draft the narrative.** The copilot writes a first draft of the underwriting narrative or\n   referral note in the insurer's format, citing the document and page behind each statement.\n5. **Underwriter decides.** The underwriter edits the draft, sets terms and price, and signs the\n   decision; edits and outcomes are logged to improve prompts, retrieval and guidelines.","valueDrivers":["employee-productivity","risk-reduction","speed","compliance"],"kpis":["time-saved-per-task","processing-time-reduction","employee-adoption","productivity-gain","accuracy"],"indicativeValue":{"referenceOrg":"A commercial insurer with 100 underwriters in middle market and specialty lines","inputs":[{"key":"underwriters","label":"Underwriters using the copilot","low":100,"high":100,"unit":"underwriters","note":"The reference insurer."},{"key":"filesPerUnderwriter","label":"Submissions fully assessed per underwriter per year","low":150,"high":250,"unit":"submissions per underwriter per year","note":"Editorial assumption for middle market and specialty lines. Replace with your own volumes."},{"key":"hoursSaved","label":"Hours saved per assessed submission","low":0.5,"high":1.5,"unit":"hours per submission","note":"Conservative against the benchmark on this page (Zurich North America underwriters report an average of 2 hours saved per submission)."},{"key":"costPerHour","label":"Fully loaded cost of an underwriter hour","low":70,"high":110,"unit":"USD per hour","note":"Editorial assumption. Replace with your own fully loaded cost."}],"formula":"underwriters * filesPerUnderwriter * hoursSaved * costPerHour","currency":"USD","period":"per year","resultLabel":"Underwriter time released","caveat":"Time released, not cash saved. It leaves out the value of better risk selection (exposures found that change a decision), more quotes per underwriter, platform and data costs, and the time underwriters spend checking drafts."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The models are capable; the hard parts are clean underwriting guidelines to retrieve from, approved data sources, a narrative format underwriters accept, and a way to measure whether the drafts are right. Adoption depends on underwriters trusting the output, so accuracy tracking matters as much as the build.","dataPrerequisites":["Current underwriting guidelines, referral rules and authority levels per line, in retrievable form","Extracted submission data (from an intake process) and policy and claims history per insured","A list of approved external sources and data vendors, with use conditions","Examples of good underwriting narratives to set the format and tone"],"integrations":["Underwriting workbench or CRM","Policy administration and claims systems (read only)","Third party data providers and web research tools","Document management for submissions and guidelines"]},"implementation":{"steps":[{"title":"Choose a line where narratives are long and consistent","detail":"Middle market property and casualty, cyber and professional lines work well: the files are big, the narrative format is standard and underwriters feel the pain."},{"title":"Clean the guidelines first","detail":"Put underwriting guidelines and referral rules into one current, owned source. The copilot can only be as right as the guidelines it retrieves."},{"title":"Define what a good draft is","detail":"Agree a checklist with senior underwriters (exposures covered, guideline breaches flagged, sources cited) and score a sample of drafts against it before rollout."},{"title":"Pilot in a few offices with feedback loops","detail":"Start with a small group, hold regular feedback sessions and publish accuracy results to the users. Zurich North America combined accuracy tracking with feedback sessions and, according to its vendor, expanded from four offices to dozens within six months."},{"title":"Instrument adoption and edits","detail":"Track how often drafts are used, how much they are edited and whether flagged exposures change decisions; low edit rates on bad drafts are a warning sign, not a success."}],"guardrails":["The underwriter signs every decision; the copilot never binds, prices or declines on its own","Every statement in a draft links to its source document or web page","Research limited to an approved list of external sources","Guideline retrieval restricted to the current approved version, with owners and review dates","No use of protected characteristics or proxies for them in any scoring the copilot surfaces"],"humanInTheLoop":"Underwriters review, edit and own every narrative and decision. Underwriting management reviews a monthly sample of drafts against the quality checklist, and referral authorities see the draft and the underwriter's edits together.","kpisToInstrument":["Time from assignment to underwriting decision, per line","Share of drafts used and the average edit distance","Exposures flagged by the copilot that changed the decision","Accuracy of drafts on a scored monthly sample","Weekly active underwriters as a share of licensed users"],"failureModes":[{"title":"Automation bias","detail":"Underwriters accept a fluent draft without checking it. Show sources inline, sample drafts for quality and make the underwriter confirm key facts."},{"title":"Stale guidelines","detail":"The copilot confidently applies a guideline that changed last quarter. Give each guideline an owner and a review date and retrieve only the current version."},{"title":"Research that drifts beyond approved sources","detail":"Web research pulls in unreliable or personal information about individuals. Restrict sources and log every page used."},{"title":"Time saved but decisions unchanged","detail":"If the copilot only speeds up writing, risk selection does not improve. Track flagged exposures and outcomes, not just minutes."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"For commercial property and casualty lines the copilot is not listed in Annex III. Used for risk assessment of natural persons in life or health insurance it falls under Annex III point 5(c) and is high risk, with risk management, data governance, logging and human oversight duties, and deployers must carry out a fundamental rights impact assessment under Article 27."},"regulations":["eu-ai-act","gdpr","dora","nist-ai-rmf","iso-42001","solvency-ii"],"guidance":[{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Published in August 2025 and addressed to national supervisors, it sets out how insurance sector legislation applies to AI systems that are not prohibited or high risk under the AI Act, including governance, fairness, explainability and human oversight."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(c) makes life and health insurance risk assessment and pricing of natural persons high risk."},{"title":"SB21-169, Protecting Consumers from Unfair Discrimination in Insurance Practices","issuer":"Colorado Division of Insurance","region":"north-america","url":"https://doi.colorado.gov/for-consumers/sb21-169-protecting-consumers-from-unfair-discrimination-in-insurance-practices","note":"Requires insurers to test external data, algorithms and predictive models for unfair discrimination; Regulation 10-1-1 sets governance requirements for life, private passenger auto and health benefit plan insurers."}],"controls":["AI inventory entry per line of business with an accountable underwriting owner","Documented approved source list and data use conditions","Retention of drafts, edits and final narratives for audit","Periodic bias and fairness testing where personal lines or individuals are in scope","Model provider risk assessment and exit plan under DORA"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the copilot is an **AI agent** inside the underwriter's tools (web, Microsoft Teams or\nan API into the workbench) with a **knowledge base** of underwriting guidelines and referral rules,\nsearched with hybrid retrieval so exact guideline references and meaning both match. **SQL\nknowledge bases** and **custom functions** give read access to policy and claims history, and the\nbuilt in **web search and browsing tools** handle research on the insured, with the tool execution\npolicy controlling which tools the agent may use.\n\nDrafting the narrative runs as an **agentic workflow** that assembles the file, researches the\ninsured and returns a draft with sources as **structured output** for the workbench. **Guardrails**\nand **PII masking** keep personal data out of prompts where it is not needed, **execution tracing**\nrecords which sources each draft used, and **test suites** with LLM graded rules score drafts\nagainst the underwriting checklist on every change. Models can be switched per agent, and data can\nstay in EU or UAE regions."},"faq":[{"question":"How much time does an underwriting copilot save?","answer":"Zurich North America's middle market underwriters report saving an average of 2 hours per submission with AI drafted narratives. For Generali GC&C, turnaround times for its distribution channels on cyber submissions were cut by 50%. Both figures come from the vendor's case studies, and the Zurich figure is self reported by underwriters rather than a time study."},{"question":"Does it replace underwriting judgment?","answer":"No. The deployments on this page keep the underwriter as the decision maker: AIG shows the underwriter analyzing the AI output before quoting, Skyward Specialty keeps underwriters in the loop to apply their judgment, and at Zurich North America underwriters start from an AI first draft. The value is in reading more, missing less and writing faster."},{"question":"Is an underwriting copilot high risk under the EU AI Act?","answer":"It depends on the line. Commercial lines are outside Annex III. Risk assessment of individuals for life or health insurance is high risk under point 5(c), which brings conformity, logging and human oversight obligations."}],"related":["commercial-underwriting-submission-triage","life-underwriting-medical-record-summarization","insurance-pricing-and-actuarial-copilot","credit-memo-drafting-agent","insurance-broker-and-agent-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from insurer filings, press releases and vendor case studies verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added Solvency II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: reworded the Markel problem statistic to match the Cytora case study, attributed the Zurich rollout to the vendor, added the Article 27 impact assessment to the AI Act basis, sharpened the EIOPA note and the FAQ answers, corrected the Skyward Specialty summary, and added an SEO title and meta description."}],"slug":"underwriting-risk-assessment-copilot","url":"https://www.blits.ai/ai-use-cases/underwriting-risk-assessment-copilot","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"generali-gcc-sixfold-cyber-underwriting","pooled":true}]},{"kpi":"employee-adoption","label":"Employee adoption","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"generali-gcc-sixfold-cyber-underwriting","pooled":true}]},{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":120,"min":120,"max":120,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"zurich-north-america-sixfold-underwriting-narratives","pooled":true}]}],"indicativeValueResult":{"low":525000,"high":4125000},"evidence":["accelerant-ai-underwriting-and-actuarial-tools","aig-underwriter-assistance","arch-capital-ai-underwriting-insights","bowhead-specialty-kalepa-underwriting-workbench","generali-gcc-sixfold-cyber-underwriting","hiscox-generative-ai-lead-underwriting","skyward-specialty-sixfold-ai-underwriting","zurich-north-america-sixfold-underwriting-narratives"]},{"title":"AI credit scoring with alternative data for thin file applicants","shortTitle":"Alternative data credit scoring","seoTitle":"Alternative data credit scoring with AI","metaDescription":"AI credit models add permissioned data such as bank cash flow to bureau files. In simulations Upstart reported to the CFPB, its model approved 27% more applicants.","definition":"A machine learning credit model that adds consumer permissioned alternative data, such as bank account cash flow, rent, utility and telco payments or ecosystem data, to credit bureau data, so a lender can assess applicants with thin or no credit files and return a decision with specific reasons.","aliases":["cash flow underwriting for consumers","thin file credit scoring","open banking credit score","machine learning credit scoring"],"industries":["banking","payments"],"functions":["lending-and-credit","underwriting","risk-management"],"patterns":["prediction-and-scoring","document-processing","conversational-agent"],"channels":["api","mobile-app","web-chat"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"lending","problem":"A bureau score needs a credit history. Young adults, migrants, gig workers, the self employed and\npeople who simply never borrowed have little or none, so a traditional scorecard either declines\nthem or prices them as if they were high risk. In the United States alone, the CFPB estimates that\n26 million adults have no credit history at a nationwide credit bureau and another 19 million have\none too thin or stale to score.\n\nThe information that would show whether these people can repay usually exists: salary and\nexpenses in their bank account, years of rent, utility and phone payments, or activity on a super\napp. Lenders could not use it at scale because it was unstructured, scattered and not permissioned\nfor credit. Open banking, consent frameworks and machine learning now make it usable, but they also\nbring new questions about fairness, explainability and privacy that a bureau scorecard never raised.","problemStats":[{"statement":"The CFPB estimates that 26 million Americans are credit invisible, with no credit history at a nationwide consumer reporting agency, and that another 19 million have a history that is stale or insufficient to produce a score under most scoring models.","sourceTitle":"An update on credit access and the Bureau's first No-Action Letter","sourceUrl":"https://www.consumerfinance.gov/about-us/blog/update-credit-access-and-no-action-letter/","year":2019}],"howItWorks":"1. **Ask for consent.** The applicant chooses to share additional data, such as a bank account\n   connection through open banking or data from an ecosystem partner, and is told what it is used\n   for.\n2. **Turn raw data into features.** Transactions are categorised into income, rent, essential\n   spend, debt payments and overdraft use; stability and trend features are computed over several\n   months.\n3. **Score alongside the bureau.** A machine learning model, or a cash flow score added to the\n   existing scorecard, estimates default risk using both bureau and alternative features.\n4. **Decide within policy.** A decision engine applies credit policy, affordability rules and\n   limits, approves, declines or refers the case, and sets line size and price.\n5. **Explain the outcome.** Each decline or unfavourable term carries the specific principal\n   reasons derived from the model, and the applicant can ask what would change the outcome.\n6. **Monitor.** Approval rates, default rates and outcomes by protected group are tracked per\n   segment, and the model is revalidated when data or populations drift.","valueDrivers":["inclusion-and-access","revenue-growth","risk-reduction","speed"],"kpis":["conversion-rate-uplift","automation-rate","processing-time-reduction","users-served"],"indicativeValue":{"referenceOrg":"A consumer lender receiving 200,000 personal loan applications a year","inputs":[{"key":"applications","label":"Applications per year","low":200000,"high":200000,"unit":"applications per year","note":"The reference lender."},{"key":"declineShare","label":"Share of applications declined today","low":0.3,"high":0.4,"unit":"fraction of applications","note":"Editorial assumption for a mainstream personal loan book. Replace with your own decline rate."},{"key":"rescuedShare","label":"Share of declines that alternative data turns into sound approvals","low":0.05,"high":0.15,"unit":"fraction of declines","note":"The high end is the Atlanticus figure, where the vendor reports that 15% of marginal declines (not all declines) could be approved profitably, so it is an upper bound; the low end allows for weaker data coverage and consent drop off. Editorial assumption, replace with your own."},{"key":"contributionPerLoan","label":"Net contribution per additional approved loan over its life","low":150,"high":400,"unit":"USD per loan","note":"Editorial assumption after expected credit losses and funding cost. Replace with your own."}],"formula":"applications * declineShare * rescuedShare * contributionPerLoan","currency":"USD","period":"per year","resultLabel":"Net contribution from additional approvals","caveat":"Leaves out the cost of data access, model development and validation, the consent drop off rate, and any change in losses on loans the bank would have approved anyway. The uplift must be proven on your own population with a holdout before it is counted."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The model is the easy part. The work is in consent flows that people complete, data coverage, model validation, fair lending testing, reason codes that stay specific and accurate, and integration with the existing decision engine and bureau.","dataPrerequisites":["Historical applications with outcomes (performance over at least 12 months) for model training and validation","Access to permissioned alternative data with a clear legal basis, such as open banking or partner data","Protected attribute data or accepted proxies for fairness testing, where the law allows","Documented credit policy and affordability rules"],"integrations":["Open banking or account aggregation provider","Credit bureau","Decision engine or loan origination system","Model monitoring and model inventory","Customer channels for consent and status updates"]},"implementation":{"steps":[{"title":"Pick the segment and the question","detail":"Start with one product and one segment, such as thin file applicants for a personal loan, and one question: can alternative data safely approve some of today's declines?"},{"title":"Backtest before you lend","detail":"Score past applicants with and without the new data and compare default rates at equal approval rates. The published VantageScore pilots at Patelco Credit Union and Michigan State University Federal Credit Union tested the score on existing portfolios before lending on it, and that is the right first step."},{"title":"Design consent as a product","detail":"The uplift only reaches customers who share their data. Explain the benefit in plain words, keep the connection step short and offer it at the point of decline or referral."},{"title":"Build reasons in from the start","detail":"Choose features and methods that let you state the specific principal reasons for every adverse action. Test the reasons with real applicants before launch."},{"title":"Test for fairness and validate independently","detail":"Run disparate impact analysis and search for less discriminatory alternatives, then put the model through the same independent validation as any credit model."},{"title":"Launch with a champion and challenger","detail":"Route a share of traffic to the new model, keep the old one as control, and widen only when loss rates on the new approvals are confirmed."}],"guardrails":["Alternative data only with explicit, recorded consent and a documented legal basis","No feature that acts as a proxy for a protected characteristic, checked by testing, not by assertion","Specific, accurate reasons for every decline or unfavourable change in terms","Credit policy and affordability rules stay deterministic and outside the model","A human owns referred cases and any appeal against a decision"],"humanInTheLoop":"Credit officers decide referred and borderline cases and handle appeals. Model risk and fair lending teams approve the model before launch and review performance by segment every quarter. The customer can ask for human review of an automated decision.","kpisToInstrument":["Approval rate on the target segment versus a control group","Default and loss rates of the incremental approvals over time","Consent completion rate at the data sharing step","Approval and pricing gaps across protected groups","Share of decisions returned automatically, and decision time"],"failureModes":[{"title":"Uplift that disappears in production","detail":"A backtest on past applicants does not match who actually consents. Measure with a live control group, not only on history."},{"title":"Proxy discrimination","detail":"Behavioural or device data can stand in for race, sex or age. Test outcomes by group and remove features that drive unjustified gaps."},{"title":"Reasons nobody can act on","detail":"Generic reasons such as \"failed to achieve a qualifying score\" or \"based on internal standards or policies\" do not meet the requirement for specific reasons and frustrate applicants. Map features to plain, specific reasons."},{"title":"Data that goes stale or disappears","detail":"A partner or aggregator changes coverage and the model degrades silently. Monitor input distributions and fall back to the bureau model."}]},"risk":{"euAiAct":{"tier":"high","basis":"Annex III point 5(b): AI systems intended to evaluate the creditworthiness of natural persons or establish their credit score are high risk, except systems used to detect financial fraud. Providers need risk management, data governance, logging and human oversight. Deployers must carry out a fundamental rights impact assessment before use (Article 27), and affected persons have a right to an explanation of individual decisions from the deployer (Article 86)."},"regulations":["eu-ai-act","gdpr","eba-loan-origination","us-sr-11-7","uk-consumer-duty","mas-ai-risk-management","nist-ai-rmf","us-ecoa-reg-b","us-fcra","pra-ss1-23"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) lists creditworthiness evaluation and credit scoring of natural persons as high risk."},{"title":"Consumer Financial Protection Circular 2022-03: adverse action notification requirements for credit decisions based on complex algorithms (issued June 2022, withdrawn May 2025)","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/","note":"Withdrawn by the CFPB on 12 May 2025 (90 FR 20084, FR Doc 2025-08286). The underlying requirement to give specific reasons comes from Regulation B itself (12 CFR 1002.9(b)(2)), which still applies."},{"title":"Guidelines on loan origination and monitoring","issuer":"European Banking Authority","region":"europe","url":"https://www.eba.europa.eu/regulation-and-policy/credit-risk/guidelines-on-loan-origination-and-monitoring","note":"Sets expectations for creditworthiness assessment, including the use of automated models, data quality and explainability."},{"title":"MAS consultation paper: Guidelines on Artificial Intelligence Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Proposed guidelines, issued for consultation in November 2025, setting supervisory expectations for AI inventories, risk materiality assessment, fairness, explainability and human oversight at financial institutions."}],"controls":["Model inventory entry with an accountable owner, validation report and approved use","Consent records linked to every decision that used alternative data","Fair lending testing before launch and on a fixed schedule, with a documented search for less discriminatory alternatives","Reason code library reviewed by compliance and tested for accuracy against model output","Drift and performance monitoring with a documented fallback to the bureau model"],"incidents":[{"title":"Incident 92: Apple Card's credit assessment algorithm allegedly discriminated against women","url":"https://incidentdatabase.ai/cite/92/","note":"Customers alleged that men received much higher credit limits than women with similar credit qualifications, and the complaints led to a regulatory investigation of the issuer's credit card practices. It shows why specific reasons and fairness testing must be ready before complaints arrive."}]},"blitsAi":{"howToBuild":"Blits.ai does not replace the credit model or the decision engine; it runs the conversations and\nthe workflow around them. An **AI agent** on the app, web chat or WhatsApp explains the product,\nasks for consent to share bank data in plain language, collects documents through **receive\nattachment** blocks and gives status updates. **Custom functions** call the lender's consent,\naggregation and decision APIs, and regulated steps such as consent and the declaration run as\ndeterministic **flows**.\n\nWhen a decision comes back, the agent explains it using only the reason codes the decision engine\nreturned, retrieved against approved wording in the **knowledge base**, and offers a route to a\nhuman through **human handover**. **Agentic workflows with human in the loop approval** handle\nreferred cases, preparing a summary for the credit officer. **PII masking** runs at the gateway,\n**test suites** check that explanations match the reason codes, and deployments can run in the\nEU or UAE region for data residency."},"faq":[{"question":"How much does alternative data increase approvals?","answer":"It depends on the population and the data. In simulations it reported to the CFPB, which the CFPB did not separately replicate, Upstart's model approved 27% more applicants than a hypothetical traditional model, with lower average APRs. In a 2025 open banking pilot, Patelco Credit Union saw 12% of subprime and 15% of near prime members move to a higher credit tier. Prove it on your own applicants with a control group."},{"question":"Is alternative data credit scoring high risk under the EU AI Act?","answer":"Yes, when it evaluates the creditworthiness of natural persons (Annex III point 5(b)). That brings obligations for risk management, data governance, human oversight and logging, deployers must assess the impact on fundamental rights, and affected people have a right to an explanation of the decision."},{"question":"Does using machine learning excuse a lender from giving specific decline reasons?","answer":"No. Regulation B requires a statement of reasons that is specific and indicates the principal reasons for the adverse action; saying the applicant failed to reach a qualifying score or cites the creditor's internal standards is not enough. Design the model and the reason codes together."},{"question":"Which alternative data is safest to start with?","answer":"Consumer permissioned bank transaction data is a common starting point, because it measures income and spending directly and the applicant chooses to share it; the Atlanticus and Patelco examples on this page both use it. Device, social and behavioural data carry higher privacy and proxy discrimination risk."}],"related":["sme-cash-flow-underwriting","adverse-action-explanations","application-and-identity-fraud-detection","conversational-loan-application-intake","model-risk-validation-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog and verified against the sources; the catalog's Patelco figures were corrected."},{"date":"2026-09-25","note":"Consolidation pass: added ECOA and Regulation B, Fair Credit Reporting Act, PRA SS1/23 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: replaced the unsourced emerging markets claim with the CFPB credit invisible estimate (added as a problem statistic), corrected the EU AI Act basis (Article 27 impact assessment, Article 86 right of affected persons), marked the MAS guidelines as a consultation, softened an unsupported FAQ superlative, recorded Upstart's 27% as Upstart's own claim, removed an unsupported Golden 1 statement, and added seoTitle and metaDescription."}],"slug":"alternative-data-credit-scoring","url":"https://www.blits.ai/ai-use-cases/alternative-data-credit-scoring","benchmarks":[],"indicativeValueResult":{"low":450000,"high":4800000},"evidence":["atlanticus-cash-flow-underwriting","golden-1-credit-union-ai-credit-scorecard","gxs-bank-alternative-data-flexiloan","patelco-credit-union-open-banking-score-pilot","upstart-alternative-data-credit-model"]},{"title":"AI decision support for airline operations control and disruption recovery","shortTitle":"Airline operations control","seoTitle":"AI airline disruption recovery and ops control","metaDescription":"AI proposes aircraft swaps and retimings to controllers. SWISS says rotation optimization saved over CHF 1M; American says HEAT averted nearly 1,000 cancellations.","definition":"Decision support in an airline's operations control center that watches the day's operation, predicts where weather, delays, crew limits or technical problems will break the plan, and proposes recovery options across aircraft, crew and passengers, such as retiming flights, swapping aircraft or holding a connection, with the cost and passenger impact of each, for controllers to approve.","aliases":["airline operations control AI","irregular operations recovery optimization","IROPS decision support","airline schedule recovery","AI for the airline operations center"],"industries":["travel-and-hospitality"],"functions":["operations"],"patterns":["prediction-and-scoring"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"An airline's plan for the day links every aircraft, crew and passenger connection. When a storm\ncloses a hub, an aircraft goes technical or a crew runs out of legal duty time, one change ripples\nthrough the network: the next flight has no aircraft, the crew is in the wrong city, passengers\nmiss connections and hotels fill up. Operations controllers have minutes to decide what to delay,\nswap or cancel.\n\nTraditionally they do it by experience, looking across separate systems for aircraft rotation,\ncrew, passengers and maintenance that were never built to optimize together. Decisions that are\ngood for one dimension, such as protecting the schedule, can be expensive in another, such as\npassenger compensation, crew overtime or noise charges. Passenger facing rebooking can soften\nthe damage, but the cost is set earlier, by the recovery plan the control center chooses.","problemStats":[],"howItWorks":"1. **Build one picture of the operation.** Aircraft rotations, crew pairings and legality,\n   passenger bookings and connections, maintenance status, airport and air traffic control\n   constraints and weather are replicated into one near real time data layer.\n2. **Predict trouble early.** Models forecast delays, missed connections and the effect of\n   forecast weather at hubs, hours ahead.\n3. **Generate recovery options.** Optimization searches for plans that retime, swap or cancel\n   flights and reassign crew, scoring each on cost, passenger impact, crew legality and\n   knock on effects.\n4. **Explain and propose.** The controller sees the recommended plan and its alternatives with\n   the trade offs in plain terms, including costs such as fuel, charges and passenger care.\n5. **Decide and execute.** Controllers accept, change or reject the plan; accepted changes flow\n   to the rotation, crew and passenger systems, which trigger rebooking and notifications.\n6. **Learn.** Outcomes of each event feed back into the models and cost functions.","valueDrivers":["cost-to-serve","customer-experience","speed","employee-productivity"],"kpis":["cost-savings","cost-reduction","productivity-gain","customer-satisfaction"],"indicativeValue":{"referenceOrg":"An airline operating 200,000 flights a year","inputs":[{"key":"flights","label":"Flights per year","low":200000,"high":200000,"unit":"flights per year","note":"The reference airline."},{"key":"disruptedShare","label":"Share of flights that need a recovery decision","low":0.05,"high":0.1,"unit":"fraction of flights","note":"Editorial assumption for a network airline, replace with your own irregular operations data."},{"key":"costPerDisruption","label":"Direct cost per disrupted flight","low":4000,"high":10000,"unit":"USD per disrupted flight","note":"Editorial assumption covering crew, passenger care and compensation, repositioning and charges. Replace with your own cost model."},{"key":"costReduction","label":"Reduction in disruption cost from better recovery plans","low":0.02,"high":0.03,"unit":"fraction of disruption cost","note":"Editorial assumption, replace with your own. Neither deployment on this page reports savings from disruption recovery alone. SWISS reports more than CHF 1 million saved in the first 14 weeks of its rotation optimization feature, from lower fuel use and avoided charges such as airport noise charges, so that figure does not set this range."}],"formula":"flights * disruptedShare * costPerDisruption * costReduction","currency":"USD","period":"per year","resultLabel":"Disruption cost avoided","caveat":"Counts only direct disruption cost. It leaves out savings from optimizing rotations on normal days, the revenue effect of fewer cancellations and missed connections, the cost of the data platform and optimization software, and the change in the control center's way of working."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The optimization is hard but available from specialist vendors and cloud providers. The larger effort is the data layer that joins rotation, crew, passenger and maintenance systems in near real time, cost functions the airline agrees on, and trust from controllers who are accountable for safe and legal operation.","dataPrerequisites":["Near real time aircraft rotation, crew pairing and passenger connection data","Crew legality rules (duty and rest limits) and qualifications in machine readable form","Agreed cost functions for delay, cancellation, passenger care, compensation and charges","Historical disruption events with the decisions taken and their outcomes"],"integrations":["Operations control and aircraft rotation system","Crew management and tracking system","Passenger service and reservation system","Maintenance and technical fleet systems","Weather, air traffic flow and airport data feeds"]},"implementation":{"steps":[{"title":"Start with one decision and one metric","detail":"Pick a frequent, contained decision such as aircraft swaps within a fleet or retiming at a hub before forecast weather, and agree how success is measured before building."},{"title":"Build the shared data layer","detail":"Replicate rotation, crew, passenger and maintenance data into one near real time view. This alone helps controllers, and every later optimization depends on it."},{"title":"Agree the cost functions","detail":"Put a price on delay minutes, missed connections, cancellations, crew overtime and charges, signed off by operations, finance and customer teams, so the optimizer and the controllers weigh options the same way."},{"title":"Recommend, do not execute","detail":"Show proposals with their trade offs and let controllers accept or reject them. Track the acceptance rate and the reasons for rejection as the main signal of fit."},{"title":"Connect to passenger recovery","detail":"Feed accepted plans straight into rebooking and customer notifications, so the passenger side starts as soon as the operational decision is taken."},{"title":"Extend to crew and network recovery","detail":"Add crew reassignment and multi hub recovery once the first use case is trusted, with crew legality checked by rules the optimizer cannot relax."}],"guardrails":["Crew duty and rest limits, qualifications and maintenance status are hard constraints the model cannot override","Controllers approve every plan before it changes the operation","Every proposal shows its cost assumptions and the alternatives considered","A manual fallback process is rehearsed for when the tool or its data feeds fail","Cost functions and model changes go through change control with operations sign off"],"humanInTheLoop":"Operations controllers and duty managers remain responsible for every decision. They approve, change or reject each proposal, and coordinators decide whether a tool such as a weather retiming plan is used at all for a given event. Crew schedulers confirm reassignments, and a review after each major disruption checks the tool's proposals against what happened.","kpisToInstrument":["Acceptance rate of proposals and reasons for rejection","Cancellations, delay minutes and missed connections per disruption event, against comparable events","Direct disruption cost per event (crew, passenger care, compensation, charges)","Time from disruption detection to an approved recovery plan","Controller workload and satisfaction with the tool"],"failureModes":[{"title":"Optimizing on stale data","detail":"A plan built on an out of date crew or aircraft position is worse than none. Monitor data freshness and block proposals when feeds lag."},{"title":"Plans the controllers do not trust","detail":"If proposals ignore constraints controllers know about, acceptance collapses. Capture rejection reasons and turn them into constraints."},{"title":"Cheapest plan, worst experience","detail":"Cost functions that underprice passenger impact produce plans that save money and lose customers. Review the weights with customer teams."},{"title":"No fallback when the system fails","detail":"A recovery tool that fails during the peak of a meltdown leaves controllers without their usual process. Keep and rehearse the manual process."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Recommending schedule, aircraft and passenger recovery plans to controllers is not listed in Annex III. Annex III point 4(b) covers AI used to make decisions affecting terms of work relationships, to allocate tasks based on individual behavior or personal traits or characteristics, or to monitor and evaluate the performance and behavior of workers. A design that reassigns individual crew members on such grounds, or that scores controllers or crew on their performance, falls in that category; one that works on flights, aircraft and crew legality and qualifications alone is less likely to, although assigning duties by qualification can still touch terms of work. The tool is not itself a safety component of an aircraft or other product regulated under Regulation (EU) 2018/1139, which Annex I Section B lists, so that route to high risk does not normally apply. Decisions that affect flight safety stay under aviation safety regulation and the airline's approved procedures; keep crew legality and maintenance limits as hard rules outside the model."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"EASA Artificial Intelligence Roadmap 2.0: A human-centric approach to AI in aviation","issuer":"European Union Aviation Safety Agency","region":"europe","url":"https://www.easa.europa.eu/en/document-library/general-publications/easa-artificial-intelligence-roadmap-20","note":"Outlines EASA's vision for the safety and ethical considerations of AI in aviation."}],"controls":["Documented cost functions and hard constraints with an accountable owner in operations","Audit log of every proposal, the data behind it and the controller's decision","Post event reviews of major disruptions comparing proposals with outcomes","Tested manual fallback and data feed monitoring"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai does not replace an operations research optimizer; it connects controllers and\ndownstream teams to one. An **AI agent** with **custom functions** can call the optimizer and\nthe operational systems' APIs, and read the current state from a **SQL knowledge base**, so a\ncontroller can ask in plain language which flights are at risk at a hub, what a proposed swap\ncosts or why a plan was recommended. A **knowledge base** with the airline's operations manual\nand procedures, retrieved with hybrid search, answers the rule questions.\n\n**Agentic workflows** with **human in the loop confirmation** can prepare a recovery proposal,\nwait for a controller to approve or reject it and then hand accepted changes to rebooking and\ncustomer notification, with an audit trail per run. Controllers and crew teams can use the\nassistant in **Microsoft Teams**, **monitors** check it daily against known questions, and the\nplatform is **model agnostic** with **EU and UAE data residency**."},"faq":[{"question":"How is this different from an AI rebooking agent?","answer":"A rebooking agent helps passengers after the airline has decided what to do. Operations control decision support shapes that decision: which flights to delay, swap or cancel and how to recover aircraft and crew. The two work best connected, so an approved plan triggers rebooking and notifications straight away."},{"question":"What results have airlines reported?","answer":"SWISS says the rotation optimization in its Operations Decision Support Suite saved more than CHF 1 million in its first three and a half months, and that controllers accept about nine in ten optimization runs. In 2023, American Airlines said its HEAT tool had prevented nearly 1,000 flight cancellations since its first use in April 2022; that is American's own figure, published without a baseline."},{"question":"Does the AI make the decisions?","answer":"No. At SWISS, operations controllers approve every change before it takes effect. At American, coordinators work with air traffic control and meteorologists to decide whether HEAT is used for a given storm, and Cranky Flier reports that HEAT starts from crew rest and availability. Beyond what these deployments disclose, a sound design keeps crew legality and maintenance limits as hard rules outside the model, because controllers remain accountable for safe and legal operation."}],"related":["flight-disruption-and-rebooking-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version with SWISS and Lufthansa Group OPSD and American Airlines HEAT, researched and checked against the sources. Editor pass added American's own cancellations figure from the archived newsroom page, recorded the SWISS savings again from SWISS's direct quote, removed an unsourced platform vendor and tightened the EU AI Act basis. Second editor pass separated what SWISS and American disclose about human decisions from the recommendation on hard rules, dated the HEAT figure to its 2023 newsroom article, lowered the high end of the savings range and aligned the EASA note with its landing page. Third editor pass removed the \"only public figure\" wording and the editorial extrapolation from the savings note. Fourth editor pass narrowed the savings note to the evidence on this page, lowered the savings range again, separated the SWISS figure from recovery in the meta description and softened the EU AI Act basis with a note on Regulation (EU) 2018/1139."}],"slug":"airline-operations-control-decision-support","url":"https://www.blits.ai/ai-use-cases/airline-operations-control-decision-support","benchmarks":[{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"CHF","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1000000,"min":1000000,"max":1000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"swiss-lufthansa-opsd-operations-decision-support","pooled":true}]}],"indicativeValueResult":{"low":800000,"high":6000000},"evidence":["american-airlines-heat-hub-disruption-tool","swiss-lufthansa-opsd-operations-decision-support"]},{"title":"AI demand forecasting and automated replenishment for retail","shortTitle":"Demand forecasting and replenishment","seoTitle":"AI demand forecasting for retail replenishment","metaDescription":"AI forecasts demand per store and item and turns it into stock orders. Albert Heijn makes almost 1 billion forecasts a day; Morrisons replaced manual replenishment.","definition":"Machine learning that forecasts demand for every item in every store or fulfillment center, day by day, from sales history, promotions, prices, weather and local events, and turns the forecast into automatic store and warehouse orders within limits set by planners, who handle the exceptions.","aliases":["AI demand forecasting for retail","automated store replenishment","machine learning replenishment","AI inventory forecasting","grocery demand planning with AI"],"industries":["retail-and-ecommerce"],"functions":["operations","analytics-and-reporting"],"patterns":["prediction-and-scoring","anomaly-detection"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"A supermarket decides every day how much of each product to send to each store. Order too little\nand the shelf is empty, the sale is lost and the customer may go elsewhere. Order too much and\nfresh food expires, ties up cash and ends up as markdowns or waste. With tens of thousands of\nproducts and hundreds of stores, that is millions of small decisions a day.\n\nTraditional replenishment ran on averages, rules of thumb and store staff counting shelves and\nkeying in orders. It copes badly with the things that move demand from one day to the next:\npromotions, price changes, weather, holidays, new products and one product cannibalizing another. The cost shows up twice,\nin empty shelves and in waste, and grocers have both commercial and public commitments to reduce\nfood waste.","problemStats":[],"howItWorks":"1. **Assemble the demand signal.** Sales history per item and location, prices, promotions,\n   planned events, weather forecasts, holidays and stock positions are loaded daily.\n2. **Forecast at the level decisions are made.** Machine learning models forecast demand per\n   item, store and day, days or weeks ahead, and learn effects such as weather on ice or\n   cannibalization between promoted soft drinks.\n3. **Turn forecasts into orders.** A replenishment engine converts the forecast into orders,\n   taking into account stock on hand, shelf life, pack sizes, delivery schedules and the service\n   level chosen for each product.\n4. **Flag exceptions.** Unusual forecasts, sudden sales changes and data gaps are flagged for a\n   planner instead of being ordered blindly.\n5. **Clear what is left.** Some retailers add dynamic markdowns, raising the discount during the\n   day for products close to their date; Albert Heijn does this with electronic shelf labels.\n6. **Learn every day.** Actual sales, waste and stockouts feed back into the next forecast.","valueDrivers":["cost-to-serve","revenue-growth","risk-reduction","employee-productivity"],"kpis":["forecast-accuracy","productivity-gain","cost-reduction","revenue-uplift"],"indicativeValue":{"referenceOrg":"A grocery chain with EUR 2 billion in annual sales","inputs":[{"key":"sales","label":"Annual sales","low":2000000000,"high":2000000000,"unit":"EUR per year","note":"The reference retailer."},{"key":"wasteShare","label":"Stock written off as waste or deep markdown, as a share of sales","low":0.02,"high":0.04,"unit":"fraction of sales","note":"Editorial assumption for a grocer with a large fresh range. Replace with your own shrink and waste figures."},{"key":"costRatio","label":"Cost of goods as a share of the sales value","low":0.7,"high":0.75,"unit":"fraction","note":"Editorial assumption, so that waste is valued at cost rather than at selling price."},{"key":"wasteReduction","label":"Reduction in waste from better forecasts","low":0.04,"high":0.1,"unit":"fraction of waste","note":"Editorial assumption, replace with your own. No source on this page measures the waste cut by machine learning forecasting itself. For orientation only: RELEX reports a 4% reduction in fresh spoilage value (fresh products only, a vendor claim) from One Stop's 2019 forecasting and replenishment rollout, before machine learning was added; after adding machine learning, One Stop reports higher ultra fresh availability with no corresponding rise in spoilage. Albert Heijn hopes to cut its total food waste by more than 10% with its forecasting, a target rather than a result; its separate Dynamic Markdown initiative saves 250,000 kilos of food a year, a markdown effect rather than a forecasting one."}],"formula":"sales * wasteShare * costRatio * wasteReduction","currency":"EUR","period":"per year","resultLabel":"Cost of stock no longer written off","caveat":"Counts only waste avoided at cost. It leaves out the sales won back from fewer empty shelves, the working capital released by lower stock, the store hours saved on manual ordering and the cost of the forecasting platform and the data work."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Forecasting models are mature and available from several vendors. The hard parts are clean, timely data from every store (sales, stock, deliveries, waste), promotion and price calendars the model can trust, and changing the operating model so store staff and planners stop overriding the system out of habit.","dataPrerequisites":["Several years of sales history per item and location, including promotions and prices","Accurate stock on hand, deliveries and waste recorded per store","Promotion, price and range change calendars known in advance","Supply constraints such as lead times, pack sizes and delivery schedules"],"integrations":["Point of sale and ecommerce order data","Inventory and store stock systems","Ordering, warehouse and supplier systems","Promotion, pricing and range planning tools","Weather and event data feeds"]},"implementation":{"steps":[{"title":"Pick the categories where error costs most","detail":"Start with fresh, short shelf life and weather driven categories, where both waste and empty shelves are expensive and the gain over rules of thumb is largest."},{"title":"Fix stock and waste data first","detail":"A forecast is only as good as the stock figure it starts from. Audit stock accuracy and make recording of waste and markdowns part of the store routine before trusting automatic orders."},{"title":"Run in parallel against a holdout","detail":"Compare model forecasts and proposed orders with current orders for several weeks, on forecast error, availability and waste per category, and agree the targets before switching."},{"title":"Automate with limits and exceptions","detail":"Let the system place orders within bounds (maximum change against last week, maximum stock cover) and send exceptions to planners with the reason the forecast changed."},{"title":"Bring stores along","detail":"Explain why an order looks the way it does and track overrides. Stores that keep overriding need either better data or better explanations, not permission to ignore the system."},{"title":"Add markdowns and allocation later","detail":"Once forecasts are stable, link them to dynamic markdowns for products close to their date and to allocation across stores and fulfillment centers."}],"guardrails":["Order limits per item and store, with larger changes held for a planner","Minimum service levels set per product so essentials are never cut to meet a waste target","Human review of forecasts for new products, major promotions and unusual events","Monitoring of data freshness, so a missing sales feed stops automatic ordering instead of producing zeros","Override tracking with reasons, reviewed weekly"],"humanInTheLoop":"Planners set the service levels, order limits and promotion inputs, review exceptions and approve forecasts for new lines and big events. Store managers can override orders, but every override is logged with a reason and reviewed, because overrides are the main way a good forecast loses its value.","kpisToInstrument":["Forecast error per category at the level orders are placed (item, store, day)","Shelf availability and lost sales estimates per category","Waste and markdowns as a share of sales, per category and store","Days of stock cover in stores and distribution centers","Share of orders placed without manual change, and override reasons"],"failureModes":[{"title":"Garbage stock data","detail":"Phantom stock (the system thinks it is there, the shelf is empty) stops reordering. Audit stock accuracy and let stores flag empty shelves quickly."},{"title":"Promotions the model did not know about","detail":"A promotion or price change missing from the calendar produces a stockout or a mountain of waste. Make the promotion calendar a governed input with deadlines."},{"title":"Override culture","detail":"Staff who distrust the system reorder by hand and the gain disappears. Explain orders, measure overrides and fix the data behind them."},{"title":"Optimizing waste at the expense of availability","detail":"A tight waste target empties shelves. Set service levels per product and measure both waste and availability."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Forecasting product demand and ordering stock is not listed in Annex III. It would become high risk under Annex III point 4(b) only if the same system allocated tasks to employees based on their individual behavior or personal traits, or monitored and evaluated their performance, for example scheduling store staff by individual productivity. GDPR applies only when loyalty or customer level data feeds the forecasts; item and store aggregates on their own are not personal data."},"regulations":["eu-ai-act","gdpr"],"guidance":[],"controls":["Documented model ownership, validation per category and monitoring of forecast error","Change control for model updates, order limits and service levels","Data quality monitoring on sales, stock and promotion feeds","Audit trail of automatic orders and manual overrides"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai does not replace the forecasting and replenishment engine; it makes that engine easier\nto use and to govern. The retailer's own database of forecasts, orders and stock (PostgreSQL, for\nexample) can be registered as a **SQL knowledge base** that an **AI agent** queries directly, so a planner or store manager can ask in plain language why an order for an\nitem looks the way it does, and get the drivers back from the data. **Custom functions** call\nthe replenishment system's API to read exceptions and to submit approved overrides.\n\nAn **agentic workflow** can run each morning over the exception list, summarize the unusual\nforecasts with their likely causes and put proposed changes in front of a planner through\n**human in the loop confirmation**, with an audit trail per run. Store teams can reach the same\nassistant in **Microsoft Teams**, and **monitors** check every day that the assistant still\nanswers a set of known questions correctly. The platform is **model agnostic** and offers **EU\nand UAE data residency**."},"faq":[{"question":"How accurate is AI demand forecasting in retail?","answer":"It depends on the category and the level of detail. RELEX reports that machine learning raised One Stop's forecast accuracy by 3.17 percentage points at product and week level within four months, and 1.82 points at product, store and week level. Accuracy tends to be lower at finer levels of detail, so measure it at the level where orders are placed."},{"question":"Can store orders really be automated?","answer":"Yes, within limits. Morrisons replaced its manual model of stock replenishment with Blue Yonder's AI powered demand forecasting and replenishment in its stores, and One Stop manages forecasting and replenishment in one RELEX system that draws store level forecasts automatically into replenishment planning. Keep planners in charge of the rules and the exceptions."},{"question":"Does it reduce food waste?","answer":"It can, but published figures are few, mostly vendor claims, and none isolates machine learning. RELEX reports a 4% cut in One Stop's fresh spoilage value from its 2019 forecasting and replenishment rollout, before machine learning was added; after adding machine learning, One Stop reports 8.5% higher availability of ultra fresh products with no corresponding rise in spoilage. Albert Heijn hopes forecasting will cut its total food waste by more than 10%, a target rather than a result. Microsoft's customer story reports that Albert Heijn's Dynamic Markdown initiative, which discounts products close to their date, now saves 250,000 kilos of food a year. Measure waste and availability together, because a tight waste target can empty shelves."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version with Albert Heijn, One Stop, Morrisons and Walmart deployments, researched and checked against the sources. The waste reduction input is an editorial assumption; the One Stop 4% (fresh spoilage, before machine learning) and the Albert Heijn 10% target are shown for orientation only. The editor pass removed the unsourced claim about what drives most forecast error, tightened the EU AI Act basis and the GDPR scope, and softened the unsourced accuracy remark in the FAQ."},{"date":"2026-09-27","note":"Fact checked against sources: credited the 250,000 kilos a year figure to Microsoft's customer story instead of implying Albert Heijn stated it, matching the evidence file."}],"slug":"retail-demand-forecasting-and-replenishment","url":"https://www.blits.ai/ai-use-cases/retail-demand-forecasting-and-replenishment","benchmarks":[],"indicativeValueResult":{"low":1120000,"high":6000000},"evidence":["albert-heijn-ai-demand-forecasting","morrisons-blue-yonder-automated-replenishment","one-stop-relex-machine-learning-forecasting","walmart-ai-supply-chain-forecasting"]},{"title":"AI document intelligence for unstructured forms and documents","shortTitle":"Intelligent document processing","seoTitle":"Intelligent document processing (IDP) with AI","metaDescription":"AI classifies scanned forms and PDFs, extracts fields with confidence scores and routes doubtful cases to staff. Deployments include Volvo Group, USCIS and Ancine.","definition":"AI that takes documents in any format, such as scanned forms, PDFs, photos, emails and handwritten notes, splits and classifies them, extracts the required fields with a confidence score, validates them against business rules and source systems, and sends only the uncertain cases to a person before the data enters the downstream process.","aliases":["IDP","document intelligence","AI form processing","AI data extraction from documents","document classification and extraction"],"industries":["cross-industry","government","automotive","manufacturing"],"functions":["operations","case-management","finance-and-accounting"],"patterns":["document-processing","computer-vision","classification-and-routing"],"channels":["api","email","internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"Many organizations still run on documents that were designed for people: application forms,\nclaims, certificates, supporting evidence, delivery notes, tax documents and letters, arriving as\nscans, phone photos, PDFs and email attachments. Staff open each one, work out what it is, retype\nthe fields into a system and check them against other records. It is slow, error prone and hard to\nscale when volumes spike, and the backlog delays decisions that matter to citizens and customers.\n\nEarlier OCR and template tools worked well for fixed layouts and struggled with anything else.\nCurrent document AI combines layout aware extraction, handwriting recognition and language models,\nso it can handle varied layouts, stamps, handwritten notes, tables across pages and several\nlanguages, as the Volvo Group deployment on this page shows. The design question\nis no longer whether AI can read the document but where it may act alone: straight through\nprocessing for confident, validated extractions, and human review for the rest, with every value\ntraceable to the place on the page it came from.","problemStats":[],"howItWorks":"1. **Ingest from every channel.** Uploads, scans, email attachments and portal submissions land in\n   one intake, where images are cleaned, rotated and split.\n2. **Classify and separate.** Each page or bundle is classified by document type (form, identity\n   document, certificate, statement, invoice) and split into individual documents.\n3. **Extract with confidence.** Fields, tables, checkboxes and signatures are extracted, with the\n   location on the page and a confidence score for each value; text in other languages can be\n   translated.\n4. **Validate.** Values are checked against business rules (formats, totals, dates) and against\n   source systems (the customer, case or supplier record).\n5. **Route by confidence.** Confident, valid documents flow straight into the downstream system;\n   the rest go to a reviewer who sees the page and the extracted value side by side.\n6. **Learn from corrections.** Reviewer corrections are logged to improve extraction and to show\n   which document types or sources cause errors.","valueDrivers":["cost-to-serve","speed","employee-productivity","risk-reduction"],"kpis":["automation-rate","accuracy","hours-saved","productivity-gain","processing-time-reduction"],"indicativeValue":{"referenceOrg":"An organization that processes 500,000 forms and supporting documents a year","inputs":[{"key":"documents","label":"Documents processed per year","low":500000,"high":500000,"unit":"documents per year","note":"The reference organization."},{"key":"minutesPerDocument","label":"Manual handling time per document today","low":4,"high":8,"unit":"minutes per document","note":"Editorial assumption for classifying, keying and checking a document. Google Cloud reports that Pupuk Indonesia's data extraction took 5 to 10 minutes before AI."},{"key":"timeSaved","label":"Share of handling time removed, including review of uncertain cases","low":0.5,"high":0.8,"unit":"fraction of handling time","note":"Editorial assumption; review of low confidence documents stays with people."},{"key":"costPerMinute","label":"Fully loaded processing staff cost","low":0.5,"high":0.8,"unit":"USD per minute","note":"Editorial assumption, replace with your own."}],"formula":"documents * minutesPerDocument * timeSaved * costPerMinute","currency":"USD","period":"per year","resultLabel":"Manual document handling cost avoided","caveat":"Counts only handling time. It leaves out faster decisions for customers and citizens, fewer keying errors, the cost of the platform and integration, and the reviewer capacity needed at peaks."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Extraction from common document types works out of the box. The effort goes into the long tail of layouts and poor scans, validation against source systems, confidence thresholds per field and a review interface that staff can work in quickly.","dataPrerequisites":["A catalogue of document types with volumes and the fields each process needs","A labelled sample per document type to measure field level accuracy","Business rules and reference data for validation","Records rules for how originals and extracted data are kept"],"integrations":["Scanning, mailroom, email and portal intake","Case management, ERP or line of business systems that receive the data","Reference data and customer or supplier master data for validation","Records and content management for the originals"]},"implementation":{"steps":[{"title":"Start with volume and pain","detail":"Pick the document types with the highest volume and clearest fields, and measure today's handling time and error rate so the baseline is real."},{"title":"Measure accuracy per field","detail":"Build a labelled test set per document type and measure accuracy per field, not per document. An average field accuracy of 95% can hide a date field that is wrong half the time."},{"title":"Set confidence thresholds per field","detail":"Decide per field what confidence and which validation checks allow straight through processing, and start conservatively with more human review."},{"title":"Design the review screen","detail":"Show the page region next to each extracted value and let reviewers correct with one click. Review speed decides most of the business case."},{"title":"Validate against systems of record","detail":"Check names, numbers and totals against the case, customer or supplier record before data is accepted, and flag mismatches rather than overwrite."},{"title":"Watch for drift","detail":"Track corrections by document type and source, and retest when forms, suppliers or scanning change."}],"guardrails":["Straight through processing only above field level confidence thresholds and after validation checks","Every extracted value linked to its location in the source document","Unrecognized pages and documents always go to a person","Personal data in documents processed and stored under the same controls as the source system","Extraction informs decisions; eligibility, benefit or credit decisions stay with the owning process and people"],"humanInTheLoop":"Reviewers handle every document below the confidence threshold or failing validation, and their corrections are logged. Process owners set thresholds and approve changes to them, and quality teams sample straight through documents regularly to confirm accuracy holds.","kpisToInstrument":["Field level accuracy per document type on a labelled sample","Straight through processing rate per document type","Reviewer time per document and correction rate","End to end time from receipt to data available in the downstream system","Errors found downstream that originated in extraction"],"failureModes":[{"title":"High average, weak critical field","detail":"Overall accuracy looks good while one field that drives decisions is often wrong. Measure and threshold per field."},{"title":"Silent errors in straight through processing","detail":"Confident but wrong values enter systems unchecked. Validate against source systems and sample straight through documents."},{"title":"The long tail stalls the program","detail":"Rare layouts consume the project. Route them to people and automate by volume."},{"title":"Extraction becomes the decision","detail":"A missing field triggers an automatic rejection of an application. Keep decisions in the owning process with human review."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Classifying documents and extracting data for a person or process to use is usually minimal risk. Even inside an Annex III area, a system that only performs a narrow procedural task, such as splitting and classifying documents, can fall outside the high risk category under Article 6(3); the provider must document that assessment and register the system (Article 6(4) and Article 49(2)). The picture changes when extraction materially influences decisions in Annex III areas, such as eligibility for public assistance benefits (point 5(a)), creditworthiness (point 5(b)) or asylum, visa and residence permit applications (point 7), where the whole system must be assessed as potentially high risk. The Article 6(3) exception never applies when the system performs profiling of natural persons."},"regulations":["eu-ai-act","gdpr","iso-42001"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Points 5 and 7 cover public benefits, credit, and migration and asylum decisions that document processing often feeds."},{"title":"Article 22 GDPR, automated individual decision making, including profiling","issuer":"European Union","region":"europe","url":"https://gdpr-info.eu/art-22-gdpr/","note":"Relevant when extraction results trigger automatic decisions with significant effects on people."}],"controls":["Documented accuracy per field and document type before go live and after changes","Threshold and routing rules under change control","Audit trail from each extracted value to the source document and reviewer action","Retention of originals and extracted data aligned with records rules","Assessment of whether downstream decisions fall in an Annex III area"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai document intake runs as an **agentic workflow**: documents arrive through a **receive\nattachment** block in a conversation or through the **REST API**, and the workflow is started\nfrom a flow or through the API. **Custom functions** call the OCR or document AI service you\nchoose through its REST API, and an **AI agent** with **structured output** classifies the\ndocument, normalizes the fields and marks values it is unsure of for review. Further custom\nfunctions validate the values against systems of record through REST or SQL queries and write\naccepted data to the case or ERP system (the integration catalog includes SAP, Salesforce and\nServiceNow).\n\n**Human in the loop** confirmation, with approve and reject controls and a threshold you\nconfigure, lets a reviewer check extracted data before it is written, and the workflow's **run\nhistory with a full audit trail** records each run. Automatic **PII masking** at the gateway\nprotects personal data, and **test suites** evaluate the agents and workflow against expected results\nafter each change. The platform is model agnostic, so the model behind each agent can be chosen\nand changed, and it can run in the EU or UAE region."},"faq":[{"question":"How accurate is AI document extraction?","answer":"It depends on the document type and the field, so measure it per field on your own documents. Google Cloud reports that Ancine, Brazil's cinema industry regulator, reached over 90% data extraction accuracy on digitized tax documents and a tenfold increase in analysts' daily processing capacity. Set confidence thresholds per field and keep people on the uncertain cases."},{"question":"What volumes and savings do organizations report?","answer":"Microsoft reports that Volvo Group's solution for invoices, credit notes and claims documents has saved 10,000 manual hours since launch, about 850 a month. Google Cloud reports that Pupuk Indonesia cut data extraction time from 5 to 10 minutes to 40 to 70 seconds, with one employee validating the results. USCIS splits and classifies I-539 applications so that adjudicators find each supporting document faster; it publishes no figures."},{"question":"How is this different from invoice processing or correspondence triage?","answer":"Invoice processing is one specialized use of document AI, with purchase order matching and posting. Correspondence triage is about routing incoming mail. This page covers the general capability for forms, applications and supporting documents in any process."}],"related":["supplier-invoice-processing","correspondence-triage-and-routing","commercial-underwriting-submission-triage","application-and-identity-fraud-detection","trade-document-examination","account-servicing-execution"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical with five evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Volvo Group claimant in the FAQ and summary to Microsoft, dated the Ancine and Volvo Group records, added the USCIS form source, added Article 6(4) to the EU AI Act basis, aligned the Blits.ai build notes with the feature inventory, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Aligned the Ancine summary with the source (analysts' daily processing capacity, subsidized projects), corrected the Pupuk Indonesia dating note, added the profiling caveat to the Article 6(3) exception, and kept the Blits.ai build notes to explicit inventory items."}],"slug":"intelligent-document-processing","url":"https://www.blits.ai/ai-use-cases/intelligent-document-processing","benchmarks":[{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"ancine-tax-document-extraction","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10000,"min":10000,"max":10000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"volvo-group-document-intelligence","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10,"min":10,"max":10,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"ancine-tax-document-extraction","pooled":true}]}],"indicativeValueResult":{"low":500000,"high":2560000},"evidence":["ancine-tax-document-extraction","pupuk-indonesia-document-processing","us-immigration-and-customs-enforcement-intelligent-document-processing","uscis-i-539-intelligent-document-processing","volvo-group-document-intelligence"]},{"title":"AI drafted explanations for credit declines and adverse actions","shortTitle":"Adverse action explanations","seoTitle":"AI adverse action notices for credit declines","metaDescription":"AI can draft credit decline notices from the model's own reason codes. Regulation B requires specific reasons, and EU law an intelligible account of the scoring.","definition":"An assistant that turns the reason codes of a credit model into an accurate, specific and readable explanation of a decline, reduced limit or repricing for the customer, and a matching internal rationale for the file, without adding any reason the model did not produce.","aliases":["adverse action notice drafting","credit decline explanation","decision explanation assistant"],"industries":["banking","payments"],"functions":["lending-and-credit","regulatory-compliance","customer-service"],"patterns":["content-generation","rag-knowledge-assistant","conversational-agent"],"channels":["email","mobile-app","web-chat","agent-desktop"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"lending","problem":"When a lender declines an application, cuts a credit line or offers worse terms, it often has to\ntell the customer why: in the US, Regulation B requires the principal reasons for a decline or other\nadverse action, and EU law gives people a right to an explanation of automated credit assessments.\nIn practice a notice can be little more than a list of checked reasons from a sample form (\"limited\ncredit experience\", \"excessive obligations in relation to income\") that says little to the customer\nand may not reflect what really drove the decision. Complex\nmachine learning models make this harder: a model can weigh many features, and the reasons printed\nmust still be the principal ones and must be accurate.\n\nThe rules are explicit. In the US, Regulation B requires a statement of reasons that is specific\nand indicates the principal reasons, and says that checking the closest reason on a sample form is\nnot enough when it is not the factor actually used. The CFPB's 2022 circular stating that this\napplies equally to complex algorithms was withdrawn in May 2025 together with many other CFPB\nguidance documents, but the adverse action notice requirements in 12 CFR 1002.9 and their official\ninterpretation are unchanged. In the EU, the Court of Justice ruled in February 2025 (Dun &\nBradstreet Austria) that a person subject to an automated credit assessment is entitled to an\nexplanation of the procedure and principles actually applied, and the EU AI Act adds a right to an\nexplanation for decisions based on high risk systems such as credit scoring. Vague or inaccurate\nexplanations leave customers without a clear next step and the lender exposed on compliance.","problemStats":[],"howItWorks":"1. **Take the decision record.** The assistant receives the decision, the reason codes and their\n   ranking from the decision engine, and the relevant customer and product data.\n2. **Retrieve approved wording.** For each reason code it retrieves the approved plain language\n   description and, where allowed, what the customer could do about it.\n3. **Draft within strict limits.** A generative model composes the notice and a short internal\n   rationale using only those reasons, in the customer's language and at an agreed reading level.\n4. **Check automatically.** A second pass verifies that every reason in the draft maps to a code\n   from the decision, that no code is missing, and that no prohibited content appears.\n5. **Review and send.** A reviewer approves the draft, or approved templates go out automatically\n   once quality is proven. Follow up questions go to an assistant limited to the same reasons, with\n   a route to a person.","valueDrivers":["compliance","customer-experience","employee-productivity"],"kpis":["accuracy","error-reduction","time-saved-per-task","handling-time-reduction"],"indicativeValue":{"referenceOrg":"A consumer lender issuing 100,000 adverse action notices a year","inputs":[{"key":"notices","label":"Adverse action notices per year","low":100000,"high":100000,"unit":"notices per year","note":"The reference lender."},{"key":"followUpShare","label":"Share of notices that lead to a question, complaint or reconsideration request","low":0.05,"high":0.1,"unit":"fraction of notices","note":"Editorial assumption. Replace with your own contact and complaint data."},{"key":"minutesSaved","label":"Minutes saved per follow up case with a clear explanation and a prepared rationale","low":20,"high":40,"unit":"minutes per case","note":"Editorial assumption for looking up the decision and writing a response. Replace with your own time study."},{"key":"hourlyCost","label":"Fully loaded cost of a lending operations hour","low":40,"high":60,"unit":"USD per hour","note":"Editorial assumption."}],"formula":"notices * followUpShare * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Operations effort saved on decline follow ups","caveat":"Counts handling effort only. The larger value, lower regulatory and fair lending risk and fewer repeat failed applications, is real but hard to price and is left out."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Drafting text is easy. The hard parts are reliable reason codes from the model, a reason library that compliance owns, and automated checks that make it impossible to state a reason the model did not produce.","dataPrerequisites":["Ranked reason codes for every adverse decision from the decision engine","An approved library of plain language descriptions per reason code, per language","Notice templates and regulatory content requirements per product and market","Samples of past notices and complaints for testing"],"integrations":["Decision engine or loan origination system","Document generation and correspondence system","Complaint and case management","Customer channels for delivery and follow up questions"]},"implementation":{"steps":[{"title":"Fix the reason codes first","detail":"Confirm with model risk that the model produces specific, ranked principal reasons that are accurate for each decision. The assistant cannot repair weak reasons."},{"title":"Build the reason library","detail":"For every code, compliance approves a customer description, an internal description and, where appropriate, a note on what the customer could change. Keep owners and review dates."},{"title":"Constrain the generation","detail":"Pass only the decision's codes and approved descriptions to the model and forbid any other reason. Use templates for the legally required parts of the notice."},{"title":"Verify every draft","detail":"Add an automated check that maps each stated reason back to a code and blocks the draft on any mismatch, omission or prohibited term."},{"title":"Review, measure, then automate","detail":"Start with human review of every draft, measure the error rate, and allow automatic sending per product only when the error rate is proven near zero."}],"guardrails":["The draft may only contain reasons present in the decision record","Every principal reason in the decision record must appear in the notice","No protected characteristic or proxy may appear as a reason","Legally required elements come from templates, not from generation","Every sent notice is stored with the decision record and the codes it was built from"],"humanInTheLoop":"Compliance owns the reason library and approves templates. Reviewers approve drafts until the measured error rate allows automatic sending, and a human answers any dispute or reconsideration request.","kpisToInstrument":["Share of drafts that pass the automated reason check first time","Error rate found in human review and in sampling after launch","Time to issue a notice after the decision","Complaints and reconsideration requests that mention unclear reasons"],"failureModes":[{"title":"Invented or softened reasons","detail":"A fluent model adds a plausible reason or blurs the real one. Block any reason that does not map to a code."},{"title":"Reasons that are accurate but useless","detail":"Codes that describe model internals mean nothing to customers. Invest in the reason library, not only the model."},{"title":"Drift between model and library","detail":"A model update adds or renames features and the library falls behind. Tie library review to model change control."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The drafting assistant does not assess creditworthiness, so on its own it is not the Annex III point 5(b) credit scoring system. It helps the lender meet the Article 86 right of affected people to a clear and meaningful explanation of decisions based on such a high risk system. If it is built into the scoring system it shares that system's high risk obligations; as a separate drafting tool its tier depends on its design and on how its output is reviewed. The follow up chat assistant must tell customers they are dealing with an AI system (Article 50)."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","eba-loan-origination","us-ecoa-reg-b","us-fcra"],"guidance":[{"title":"Regulation B, 12 CFR 1002.9: notifications, with official interpretation","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/regulations/1002/9/","note":"The statement of reasons must be specific and indicate the principal reasons, must describe the factors actually considered or scored, and may not leave out a factor that was a principal reason."},{"title":"Regulation B, Appendix C: sample notification forms","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/regulations/1002/c/","note":"Checking the closest reason on a sample form does not satisfy the notice requirement when it is not the factor actually used."},{"title":"Consumer Financial Protection Circular 2022-03: adverse action notification requirements for credit decisions based on complex algorithms","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/compliance/circulars/circular-2022-03-adverse-action-notification-requirements-in-connection-with-credit-decisions-based-on-complex-algorithms/","note":"Stated that the specific reasons requirement applies equally to complex or opaque models. Withdrawn by the CFPB in May 2025 together with Circular 2023-03 on sample forms; the notice requirements in 12 CFR 1002.9 it interprets are unchanged."},{"title":"Interpretive rules, policy statements, and advisory opinions; withdrawal","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.federalregister.gov/documents/2025/05/12/2025-08286/interpretive-rules-policy-statements-and-advisory-opinions-withdrawal","note":"The May 2025 Federal Register notice that withdrew many CFPB guidance documents, including Circulars 2022-03 and 2023-03 on adverse action notices."},{"title":"Innovation spotlight: providing adverse action notices when using AI/ML models","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/about-us/blog/innovation-spotlight-providing-adverse-action-notices-when-using-ai-ml-models/","note":"An earlier CFPB blog on how the adverse action rules apply to machine learning models. The page now carries a notice that it describes the requirements incompletely."},{"title":"Case C-203/22, Dun & Bradstreet Austria: automated credit assessment and the right to an explanation","issuer":"Court of Justice of the European Union","region":"europe","url":"https://curia.europa.eu/jcms/upload/docs/application/pdf/2025-02/cp250022en.pdf","note":"Under the GDPR the controller must describe the procedure and principles actually applied so the person can understand which personal data were used and how; handing over the algorithm is not enough."},{"title":"Article 86, right to explanation of individual decision making","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/86/","note":"People affected by a deployer's decision based on an Annex III high risk system (except critical infrastructure, point 2) that has legal or similarly significant adverse effects on them can ask for a clear and meaningful explanation."},{"title":"SR 26-2: Revised Guidance on Model Risk Management","issuer":"Board of Governors of the Federal Reserve System, OCC and FDIC","region":"north-america","url":"https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm","note":"Interagency model risk management guidance of April 2026 that supersedes and replaces SR 11-7. It applies to the credit model whose reason codes the notice is built from, and is expected to be most relevant to banking organizations with over $30 billion in total assets."}],"controls":["Reason library under compliance ownership with version history","Automated reason to code reconciliation on every draft","Sampling of sent notices against decision records, reported to compliance","Change control that links model releases to reason library review"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered by the decision engine through an API token.\nA **custom function** fetches the decision record and reason codes, the **knowledge base** holds\nthe approved reason library and notice templates, and an **agent** with **structured output**\ndrafts the notice and the internal rationale. A deterministic **custom function** then checks that\nevery stated reason maps to a code, and **human in the loop approval** holds the draft until a\nreviewer releases it.\n\nCustomers who reply or ask in chat reach an **agent** that answers only from the same decision\nrecord and hands over to a person through **human handover** for disputes. **Guardrails** block\nprohibited terms, **PII masking** protects applicant data, and **test suites** with LLM based\ngrading check each release against a set of real decision records. The platform is model agnostic,\nso the drafting model can be changed without rebuilding the workflow."},"faq":[{"question":"Can a generative model write adverse action notices?","answer":"It can draft them, but only from the reason codes the credit model produced and with an automated check that nothing was added or left out. The legally required parts should come from templates."},{"question":"Does using AI change the duty to explain credit decisions?","answer":"No. In the US, Regulation B requires specific reasons that describe the factors actually considered or scored, whatever the technology. The CFPB withdrew its 2022 circular on complex algorithms in May 2025, but the notice requirements in 12 CFR 1002.9 and their official interpretation are unchanged. In the EU, the Court of Justice has ruled that a person subject to an automated credit assessment must be told the procedure and principles actually applied, in an intelligible way, and the EU AI Act adds a right to an explanation for decisions based on high risk systems such as credit scoring."},{"question":"Which named lenders have this documented?","answer":"Wells Fargo Bank holds a patent, granted in October 2022, for an adverse action methodology that ranks the characteristics of a machine learning credit risk model to identify the principal reasons behind a denial. Trade press reported in August 2021 that a Wells Fargo team had begun deploying an explainability technique for its deep learning credit models. Discover Financial Services holds a related patent, granted in July 2024, that turns Shapley based explanations of a credit model into adverse action reason codes. Both are the lender's own patent filings rather than a disclosed error rate or notice volume, so they show that the methodology is real and built, not that it runs at full production scale."}],"related":["alternative-data-credit-scoring","sme-cash-flow-underwriting","outbound-notice-drafting","conversational-loan-application-intake","loan-restructuring-recommendations"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the editor after a targeted blocker check."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog with regulator guidance verified; no deployment evidence found yet, so status is draft."},{"date":"2026-09-25","note":"Consolidation pass: added ECOA and Regulation B, Fair Credit Reporting Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: CFPB Circulars 2022-03 and 2023-03 were withdrawn in May 2025, so the problem, FAQ and guidance now cite Regulation B itself and record the withdrawal; added the CJEU Dun & Bradstreet Austria ruling, UK GDPR and the Article 50 duty for the chat assistant; replaced unsourced example reasons with Appendix C wording; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: removed SR 11-7 (superseded by SR 26-2 in April 2026) and added SR 26-2 as guidance; limited the opening claim to Regulation B and EU law; corrected the CFPB withdrawal wording to many guidance documents; aligned the CJEU wording in the FAQ and meta description with the ruling; stated the Article 86 conditions. Status stays draft: no deployment evidence yet."},{"date":"2026-09-27","note":"Searched again for a named lender or vendor case study describing generative AI drafting the customer facing adverse action explanation itself (Zest AI, Upstart, GDS Link, Taktile and general press coverage). All results describe reason code generation, decisioning or explainability capability in general terms, not a named deployment of this specific drafting job with a verifiable outcome. Status stays draft; no evidence added."},{"date":"2026-09-27","note":"Added two evidence records found through patent search: Wells Fargo Bank's adverse action methodology patent (granted October 2022) and Discover Financial Services' Shapley based adverse action reason code patent (granted July 2024), both grade B. Updated the FAQ entry that previously said no named lender had been found. Status moved from draft to review; neither record discloses an outcome metric, so no metric was added and the value model is unchanged."}],"slug":"adverse-action-explanations","url":"https://www.blits.ai/ai-use-cases/adverse-action-explanations","benchmarks":[],"indicativeValueResult":{"low":66666.66666666667,"high":400000},"evidence":["discover-financial-services-adverse-action-reason-codes","wells-fargo-adverse-action-methodology"]},{"title":"AI drafting copilot for civil servants for correspondence, briefings and ministerial replies","shortTitle":"Civil servant drafting copilot","seoTitle":"AI drafting assistant for civil servants","metaDescription":"AI drafts replies, briefings and summaries that civil servants edit and clear. UK Cabinet Office Assist users report saving about 3 hours a week.","definition":"A generative AI assistant that drafts replies to correspondence from the public and elected representatives, briefings, submissions and summaries for civil servants, grounded in the department's approved lines, policy documents and case data, with the official editing and approving every word before it is sent or cleared.","aliases":["ministerial correspondence drafting AI","government briefing drafting assistant","civil service writing assistant","public sector document drafting copilot"],"industries":["government"],"functions":["citizen-services","knowledge-management","case-management"],"patterns":["content-generation","rag-knowledge-assistant","summarization"],"channels":["internal-tools","email","microsoft-teams"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","problem":"A large share of civil service time goes into writing: replies to letters and emails from the public\nand from members of parliament, ministerial correspondence, briefings for ministers and senior\nofficials, submissions, meeting notes and summaries of long documents. Much of it follows known\npatterns. A correspondence officer finds the current approved lines in a briefing pack, adapts them\nto the question and routes the draft for clearance; a policy official condenses a stack of papers\ninto a two page brief.\n\nThe work is slow and uneven. Finding the right, current line takes time, service level targets for\nreplies are missed when volumes spike, and quality depends on who drafts. Generative AI can produce a\ngood first draft in seconds, but in government a fluent draft that states a superseded policy, gets a\ncase fact wrong or sounds careless in a ministerial letter is a real problem, so the design has to\nkeep drafts grounded in approved sources and every word owned by an official.","problemStats":[],"howItWorks":"1. **Start from the request.** The official pastes or forwards the incoming letter, or selects the\n   documents to be summarised or briefed on.\n2. **Retrieve approved content.** The assistant searches the department's approved standard lines,\n   briefing packs, policy documents and, for casework, the relevant case record.\n3. **Draft in house style.** It produces a first draft in the right template (reply letter,\n   briefing, submission) with the tone set for the audience, and cites the sources it used.\n4. **Edit and clear.** The official checks facts against the sources, edits the draft and sends it\n   through the normal clearance and approval route; nothing is sent automatically.\n5. **Learn from edits.** Feedback and edits show which lines are missing or outdated, so owners\n   update the approved content rather than the prompt.","valueDrivers":["employee-productivity","speed","customer-experience"],"kpis":["productivity-gain","hours-saved","time-saved-per-task","users-served","response-time-reduction"],"indicativeValue":{"referenceOrg":"A ministry that sends 40,000 replies and briefings a year","inputs":[{"key":"drafts","label":"Replies, briefings and summaries drafted per year","low":20000,"high":60000,"unit":"documents per year","note":"Editorial assumption. For scale, the UK Department for Education says its correspondence teams handle about 1,000 external queries a month that need a reply.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dfe-correspondence-drafter"},{"key":"minutesSaved","label":"Minutes saved per document after human review","low":10,"high":20,"unit":"minutes per document","note":"Editorial assumption, replace with your own. The Department for Education expects its tool to cut drafting from about 30 minutes to about a minute (a calculation made before user testing), and user research for Cabinet Office Assist found users save about 3 hours a week; review, fact checking and clearance still take time.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dfe-correspondence-drafter"},{"key":"hourlyCost","label":"Fully loaded cost of an official's hour","low":40,"high":60,"unit":"EUR per hour","note":"Editorial assumption. Replace with your own staff cost."}],"formula":"drafts * minutesSaved / 60 * hourlyCost","currency":"EUR","period":"per year","resultLabel":"Drafting time released","caveat":"Time released, not cash saved, unless headcount or contractor spend changes. It leaves out the cost of licences and assurance, the value of faster replies to citizens and the risk cost of errors that slip through review."},"macroEstimates":[{"statement":"The Alan Turing Institute estimates that AI could support up to 41% of tasks across the UK public sector, according to the UK government.","sourceTitle":"Landmark government trial shows AI could save civil servants nearly 2 weeks a year","sourceUrl":"https://www.gov.uk/government/news/landmark-government-trial-shows-ai-could-save-civil-servants-nearly-2-weeks-a-year","year":2025}],"feasibility":{"complexity":"low","complexityNote":"A drafting assistant over approved documents is quick to build. The hard parts are organisational: keeping standard lines current and owned, agreeing what may be pasted in at which security classification, and fitting drafts into existing clearance workflows.","dataPrerequisites":["Current approved standard lines and briefing packs, each with an owner and review date","Templates and style guides for letters, briefings and submissions","For casework replies, read access to the case record under the right permissions"],"integrations":["Correspondence management or case management system","Document management (for example SharePoint) for approved content","Email and Microsoft Teams","Identity and access management for security classifications"]},"implementation":{"steps":[{"title":"Start with high volume correspondence on stable lines","detail":"Pick a correspondence stream where replies mostly reuse approved lines, as the Department for Education did with its briefing packs, before moving to ministerial submissions."},{"title":"Clean up the source of truth","detail":"Put every standard line and briefing pack under an owner with a review date, and remove superseded versions; the assistant is only as current as this library."},{"title":"Fit the existing clearance route","detail":"Keep drafts inside the current approval workflow, as the Crown Prosecution Service did with its multi layer review, rather than creating a side channel."},{"title":"Mark AI content and require citations","detail":"Show which text is generated and which source each statement came from, so reviewers can check quickly."},{"title":"Measure time and quality together","detail":"Track drafting time, clearance rework and complaints or corrections after sending, not only user satisfaction."}],"guardrails":["Drafts only; nothing is sent, published or cleared without an official's approval","Answers grounded in approved lines and case data, with refusal when the library has no answer","Security classification limits enforced on what can be entered and on the model endpoint used","Personal data in correspondence masked in logs and prompts","No political or policy positions beyond the approved lines"],"humanInTheLoop":"The official who owns the reply or briefing edits and approves it, and the normal clearance chain applies. Content owners keep the standard lines current, and a sample of sent replies is reviewed for accuracy and tone.","kpisToInstrument":["Drafting time per document, before and after","Share of drafts sent with minor edits only","Replies within service level targets","Corrections, complaints or follow ups caused by errors in replies","Active users as a share of those with access"],"failureModes":[{"title":"Superseded lines in fluent prose","detail":"The assistant confidently repeats an outdated policy position. Prevent it with owned, dated content and by removing old versions from the index."},{"title":"Rubber stamp review","detail":"Busy officials approve drafts without reading them properly. Show sources, sample sent replies and keep accountability with the named official."},{"title":"Sensitive data in the wrong tool","detail":"Staff paste classified or personal information into a tool not cleared for it. Provide a sanctioned tool at the right classification so they have no reason to use public ones."},{"title":"Tone that damages trust","detail":"Replies that sound generic or evasive to a distressed constituent or a member of parliament. Tune templates by audience and review high profile correspondence closely."},{"title":"A general chatbot overtaken by enterprise suites","detail":"i.AI withdrew its Redbox assistant at the end of 2025 after tools such as Microsoft Copilot offered similar functionality, and it reports that most interactions were general drafting and chat. Build where a bespoke tool adds value that a suite does not, such as approved lines, case data and clearance workflows."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"An internal drafting assistant that an official reviews is not listed in Annex III. Article 50(4) requires disclosure of AI generated text published to inform the public on matters of public interest, unless it has undergone human review and a person holds editorial responsibility, which this design provides. If the tool is used to evaluate eligibility for public assistance benefits or services rather than to draft, Annex III point 5(a) can apply."},"regulations":["eu-ai-act","gdpr","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Algorithmic Transparency Recording Standard hub","issuer":"UK government","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"UK departments publish transparency records for drafting tools such as Assist, the Department for Education's Correspondence Drafter and the now withdrawn Redbox."},{"title":"Public attitudes to the use of AI in DfT consultations and correspondence","issuer":"Department for Transport","region":"europe","url":"https://www.gov.uk/government/publications/public-attitudes-to-the-use-of-ai-in-dft-consultations-and-correspondence","note":"Research on how the public views AI support for drafting replies to their correspondence."}],"controls":["Approved content library with named owners and review dates","Transparency record and staff guidance on permitted use","Security classification controls on inputs and model endpoints","Audit log of drafts, edits and the approving official","Periodic quality sampling of sent correspondence"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** of approved standard lines and\nbriefing packs, retrieved with **hybrid search** so that exact policy wording is found as well as\nrelated content. Approved lines and briefing packs are uploaded to the **document library** with\nversion control, so owners can replace a line and revert if needed; SharePoint is also available\nas a ready made tool for the agent. **Custom functions** read the case record for\ncasework replies, and the agent is instructed to draft in the department's templates and name\nthe documents it drew on.\n\nOfficials use it in **Microsoft Teams**, the web chat or through the **email channel**. **PII\nmasking** runs at the gateway before text reaches a model, and **guardrails** stop the agent from\nanswering outside the approved lines. Flagged real conversations can be added to **test suites**,\nwhich the team runs before publishing a content or prompt change. The platform is **model\nagnostic** and offers EU and UAE data residency, so a\ndepartment can choose the model and region that fits its security classification."},"faq":[{"question":"How much time does AI drafting save civil servants?","answer":"Reported figures vary by task. User research for Cabinet Office Assist found that users save about 3 hours a week on average, and participants in the UK's cross government Copilot trial estimated 26 minutes a day across all tasks (a self reported figure). The Department for Education calculated, before user testing, that its correspondence tool would be 30 times quicker than a manual process of about 30 minutes."},{"question":"Can AI send replies to the public on its own?","answer":"It should not in this design. Every tool on this page drafts for an official who edits and approves, and replies go through the normal clearance route."},{"question":"Is this high risk under the EU AI Act?","answer":"Not as designed here. Internal drafting assistance is not listed in Annex III. If the tool is used to evaluate eligibility for essential public assistance benefits and services rather than to draft, Annex III point 5(a) can apply. Public bodies should still tell people how AI is used and keep a named official responsible for what is sent."}],"related":["correspondence-triage-and-routing","freedom-of-information-request-processing","public-consultation-response-analysis","outbound-notice-drafting","email-and-ticket-reply-drafting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version for the government vertical, with UK civil service evidence verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed the Department for Education 30 times metric (a projection made before user testing) and reworded the FAQ and value note; corrected the CPS summary; added AWS Bedrock to Assist and the DBT Redbox repository as a source; added the Annex III point 5(a) caveat; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Recorded that i.AI withdrew Redbox at the end of 2025 (evidence stage paused, lessons post added as a source); adoption stage set to emerging; SharePoint in howToBuild limited to a ready made tool; FAQ wording aligned with the sources and Annex III point 5(a); dropped NIST AI RMF from the regulations."}],"slug":"civil-servant-drafting-copilot","url":"https://www.blits.ai/ai-use-cases/civil-servant-drafting-copilot","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":1015,"min":30,"max":2000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"uk-cabinet-office-redbox-civil-service-assistant","pooled":true},{"id":"crown-prosecution-service-correspondence-drafting-tool","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":3,"min":3,"max":3,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"cabinet-office-assist-government-communications","pooled":true}]}],"indicativeValueResult":{"low":133333.33333333334,"high":1200000},"evidence":["cabinet-office-assist-government-communications","crown-prosecution-service-correspondence-drafting-tool","department-for-education-correspondence-drafter","uk-cabinet-office-redbox-civil-service-assistant","uk-government-m365-copilot-experiment"]},{"title":"AI drafting of clinical study reports and regulatory documents","shortTitle":"Clinical and regulatory document drafting","seoTitle":"AI clinical study report and regulatory writing","metaDescription":"Generative AI drafts clinical study reports from trial tables for medical writers to review. Merck halved draft errors; Novo Nordisk cut CSR writing time by 90%.","definition":"Generative AI that drafts clinical study reports and other regulated documents, such as protocols, patient materials and submission modules, from the trial's statistical tables, listings and figures and from approved template text, for medical writers to verify, edit and approve before anything is submitted to a regulator.","aliases":["AI medical writing","clinical study report automation","CSR generation with generative AI","regulatory submission drafting","AI authoring of regulatory documents"],"industries":["pharma-and-life-sciences"],"functions":["regulatory-compliance","operations"],"patterns":["content-generation","rag-knowledge-assistant","document-processing"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"A clinical study report describes the design, conduct and results of a trial in the structure\nregulators expect, and can run to 300 pages. Medical\nwriters assemble it from thousands of pages of tables, listings and figures, reconcile conflicting\nnumbers, and write narrative in precise regulatory language, followed by rounds of review. The\nsame pattern repeats for protocols, investigator brochures, safety narratives, patient materials\nand the modules of a marketing application.\n\nThe work sits on the critical path to approval, so weeks spent writing are weeks a medicine is not\nwith patients, and each report ties up experienced writers for a long time. Manual drafting is also\nprone to errors, and every error has to be found and corrected in review. Generative AI fits the task because much of the text restates structured\nresults in standard language, but the output must be exactly right: a transposed number or an\ninvented statement in a regulatory document is a serious quality failure.","problemStats":[{"statement":"Anthropic's case study reports that Novo Nordisk's staff writers averaged only 2.3 clinical study reports per year.","sourceTitle":"Novo Nordisk accelerates clinical documentation and drug development with Claude","sourceUrl":"https://claude.com/customers/novo-nordisk","year":2025}],"howItWorks":"1. **Ingest the study outputs.** Statistical tables, listings and figures, the protocol and the\n   statistical analysis plan are loaded and preprocessed, so each table can be read reliably.\n2. **Map content to the template.** Each section of the report template is linked to the tables and\n   approved boilerplate it needs, following the house structure and the ICH format.\n3. **Draft section by section.** A language model writes each section from its mapped sources, with\n   every number traced back to the table it came from and approved text reused where possible.\n4. **Check automatically.** Rules compare numbers in the text with the source tables, flag missing\n   sections and check terminology and style before a person sees the draft.\n5. **Review and approve.** Medical writers and study experts review, edit and interpret the results;\n   the document follows the normal quality control and approval process before submission.","valueDrivers":["speed","employee-productivity","compliance","cost-to-serve"],"kpis":["handling-time-reduction","processing-time-reduction","error-reduction","hours-saved","productivity-gain"],"indicativeValue":{"referenceOrg":"A drug developer that writes 20 clinical study reports a year","inputs":[{"key":"reports","label":"Clinical study reports per year","low":20,"high":20,"unit":"reports per year","note":"The reference developer."},{"key":"hoursPerDraft","label":"Hours to a fully human reviewed first draft today","low":120,"high":180,"unit":"hours per report","note":"The high end is Merck's reported average of 180 hours before its platform. The low end is an editorial assumption, replace both with your own time data."},{"key":"timeSaved","label":"Share of drafting time saved","low":0.3,"high":0.55,"unit":"fraction of drafting hours","note":"The high end is Merck's achieved result: Merck went from an average of 180 to 80 hours for a fully human reviewed first draft (about 55% less). The low end is an editorial assumption for a less mature platform. Novo Nordisk reports 90% less writing time, which excludes review and approval, so it is not used as the high end."},{"key":"costPerHour","label":"Fully loaded cost per medical writer hour","low":80,"high":150,"unit":"USD per hour","note":"Editorial assumption covering internal writers and agency rates. Replace with your own."}],"formula":"reports * hoursPerDraft * timeSaved * costPerHour","currency":"USD","period":"per year","resultLabel":"Medical writing effort released on first drafts","caveat":"Drafting effort only. It leaves out the value of earlier submissions, which for a marketed medicine can far exceed the writing cost, review and quality control time that remains, the cost of building and validating the platform, and other document types."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The hard parts are reliable extraction from complex statistical tables, traceability from every sentence to its source, validation of the system under quality management rules, and changing the writing process. Merck reports a team of more than 80 people across data science, AI and medical expertise; Novo Nordisk spent months integrating legacy systems for device protocols.","dataPrerequisites":["Statistical tables, listings and figures in a consistent, machine readable format","Approved templates and boilerplate text per document type, with owners","Past reports and reviewer comments to test against","A style guide and terminology list"],"integrations":["Statistical computing environment that produces the study outputs","Document management and regulatory information management systems","Quality management system for review, approval and electronic signatures","Submission publishing tools"]},"implementation":{"steps":[{"title":"Start with one document type and its most formulaic sections","detail":"Clinical study report results sections that restate tables are the usual starting point; leave interpretation, discussion and conclusions to writers at first."},{"title":"Make every number traceable","detail":"Store the link between each sentence and the table cell it uses, and check numbers automatically, so reviewers verify instead of recalculating."},{"title":"Reuse approved text","detail":"Put expert approved boilerplate and templates in a governed library with versions, and generate only what changes per study."},{"title":"Validate under your quality system","detail":"Treat the platform as a computerized system used in a regulated process: define its intended use, validate it on real studies, and control changes to prompts, models and templates."},{"title":"Redesign the review, then scale","detail":"Train writers to review generated drafts, measure hours, error rates and review cycles per report, and extend to further document types once quality is stable."}],"guardrails":["Every generated number checked automatically against its source table before human review","No document submitted without review and approval by qualified medical writers and study experts","Interpretation of results and conclusions written or approved by named experts","Unpublished trial data processed only in approved, access controlled environments","Change control and revalidation for prompts, models and templates"],"humanInTheLoop":"Medical writers own every document: they review each section against its sources, edit the narrative and interpretation, and take the draft through the normal quality control and approval workflow. Study clinicians and statisticians approve the interpretation of results.","kpisToInstrument":["Writer hours to a human reviewed first draft, per document type","Errors found in quality control per draft, by category (data, terminology, citations)","Review cycles and elapsed days from database lock to final report","Share of generated text kept after review","Regulator questions attributable to document quality"],"failureModes":[{"title":"Numbers that do not match the tables","detail":"A model transposes, rounds or invents a figure. Check every number against its source automatically and block drafts that fail."},{"title":"Plausible but wrong interpretation","detail":"The draft states a conclusion the data do not support. Keep interpretation and discussion with experts and mark generated interpretive text clearly."},{"title":"Faster drafts, same timeline","detail":"Drafting gets quicker but review and approval stay slow. Train reviewers to work with generated drafts, and measure review cycles and elapsed days, not only drafting hours. Merck reports that it revamped its operations and trained teams in the skills needed to oversee its platform."},{"title":"Unvalidated changes","detail":"A model or prompt update changes output quality unnoticed. Regression test on reference studies before every change."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Drafting regulated documents for expert review is not listed in Annex III and is not a practice prohibited by Article 5, so the tier turns on the sponsor's role under Article 50. A sponsor that deploys a third party drafting tool has no specific AI Act obligations beyond AI literacy: the Article 50(4) disclosure duty covers AI generated text published to inform the public on matters of public interest, which clinical study reports and regulatory submissions are not. For that sponsor the tier is minimal. A sponsor that builds its own generating system, as Merck (a proprietary platform) and Novo Nordisk (NovoScribe) did, is its provider under Article 50(2) and must mark the synthetic text in a machine readable format, unless the exemption for an assistive function for standard editing applies; drafting whole report sections goes beyond that exemption, so for that sponsor the tier is limited. Quality expectations come from medicines regulation and EMA guidance: the EMA reflection paper expects close human supervision and quality review when AI drafts medicinal product information documents, and makes the clinical trial sponsor, marketing authorisation applicant or holder, or manufacturer responsible for ensuring that models and data pipelines are fit for purpose and meet GxP standards and EMA guidelines."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001"],"guidance":[{"title":"ICH E3, Structure and Content of Clinical Study Reports","issuer":"International Council for Harmonisation","region":"global","url":"https://database.ich.org/sites/default/files/E3_Guideline.pdf","note":"Sets the structure and content that a clinical study report follows for submission to regulators, the format the drafting tool's templates map against."},{"title":"Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle","issuer":"European Medicines Agency","region":"europe","url":"https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle_en.pdf","note":"States that AI used for drafting, compiling or reviewing medicinal product information documents should be used under close human supervision, with quality review so that all generated text is factually and syntactically correct before submission."},{"title":"Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (draft guidance)","issuer":"US Food and Drug Administration","region":"north-america","url":"https://www.fda.gov/media/184830/download","note":"The January 2025 draft, still marked as draft guidance on the FDA site in September 2026, does not address AI used for operational efficiencies such as drafting or writing a regulatory submission when it does not affect patient safety, drug quality or the reliability of study results; its credibility framework applies when AI produces data that support regulatory decisions."},{"title":"Part 11, Electronic Records; Electronic Signatures, Scope and Application","issuer":"US Food and Drug Administration","region":"north-america","url":"https://www.fda.gov/regulatory-information/search-fda-guidance-documents/part-11-electronic-records-electronic-signatures-scope-and-application","note":"States that FDA interprets the scope of Part 11 narrowly, covering electronic records and signatures required by predicate rules or submitted to FDA, and announces enforcement discretion for the validation, audit trail, record retention and record copying requirements, and for all Part 11 requirements on systems operational before 20 August 1997 (legacy systems). Predicate rules stay fully enforced regardless."}],"controls":["Documented intended use, validation report and change control for the drafting platform","Traceability from generated text to source tables and approved templates","Automated numeric consistency checks with results kept in the document history","Access controls and logging for unpublished trial data","Quality control sampling of generated sections by experienced writers"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** run per document. Study outputs are registered as a\n**SQL knowledge base** or ingested as documents, approved templates and boilerplate live in a\n**knowledge base** with version control, and **hybrid retrieval** pulls the right approved text for\neach section. An **AI agent** with **structured output** drafts each section together with the\ntable references it used, and **custom functions** run the numeric checks against the source data\nand write the draft to the document system.\n\n**Human in the loop approval** keeps the release of every document with a medical writer, the\n**run history and audit trail** show what was generated from which inputs, and **test suites**\ncompare drafts with approved reference reports before any prompt or model change. The platform is\nmodel agnostic, so the sponsor can choose or switch models per section type."},"faq":[{"question":"How much faster does AI make clinical study reports?","answer":"Merck reports that first drafts now take three to four days instead of two to three weeks, and that the time to a fully human reviewed first draft fell from an average of 180 to 80 hours. Novo Nordisk says writing times on clinical study reports fell by 90%. Review and approval still take time, so measure the whole cycle, not only drafting."},{"question":"What do regulators expect for AI drafted submissions?","answer":"Responsibility stays with the company that uses the AI: the EMA reflection paper makes the clinical trial sponsor, marketing authorisation applicant or holder, or manufacturer responsible for ensuring that models and data pipelines are fit for purpose and meet GxP standards and EMA guidelines. For medicinal product information documents, the EMA asks for close human supervision and quality review so that all generated text is factually and syntactically correct. The FDA's January 2025 draft guidance on AI for regulatory decision making does not address AI used to draft a submission when it does not affect patient safety, drug quality or the reliability of study results."},{"question":"Does AI reduce errors or add them?","answer":"Both are possible. Merck reports 50% fewer errors in drafts across data, messaging, citations, terminology and typography, but language models can also produce plausible wrong numbers, which is why automatic checks against the source tables matter."}],"related":["adverse-event-case-intake","clinical-trial-patient-matching"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with evidence from Merck and Novo Nordisk, verified against the sources. Editor pass: EMA guidance limited to what the reflection paper says, unsourced attributions removed. Second editor pass: EU AI Act tier set to context dependent with the Article 50(2) marking duty for sponsors that build their own generator and Article 50(4) limited to public interest text; editorial inputs in the worked example labelled; problem statement and Part 11 note tightened."},{"date":"2026-09-27","note":"Review fix: rewrote the Part 11 note to include validation among the requirements under enforcement discretion and to state predicate rules stay fully enforced; added ICH E3 as a guidance entry to support the house structure and ICH format claim; removed the unsourced prevalence claim from the problem section's first sentence."}],"slug":"clinical-and-regulatory-document-drafting","url":"https://www.blits.ai/ai-use-cases/clinical-and-regulatory-document-drafting","benchmarks":[{"kpi":"error-reduction","label":"Error reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"merck-clinical-study-report-generation","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"novo-nordisk-novoscribe-clinical-documentation","pooled":true}]}],"indicativeValueResult":{"low":57600,"high":297000.00000000006},"evidence":["merck-clinical-study-report-generation","novo-nordisk-novoscribe-clinical-documentation"]},{"title":"AI early warning and covenant monitoring for loan portfolios","shortTitle":"Credit early warning and covenants","seoTitle":"AI credit early warning and covenant monitoring","metaDescription":"AI early warning flags weakening borrowers from covenant tests, account flows and peer data. PNC took OakNorth's monitoring software in 2020; SMBC announced a deal.","definition":"A monitoring system that tracks covenant tests and borrower reporting across a loan book, reads financials, filings and news, and combines them with payment and sector signals to flag borrowers whose credit is deteriorating, with the evidence and a suggested next step for the relationship manager.","aliases":["early warning system for credit","covenant monitoring","watchlist monitoring","credit portfolio monitoring"],"industries":["banking"],"functions":["risk-management","lending-and-credit"],"patterns":["anomaly-detection","document-processing","agentic-workflow","summarization"],"channels":["internal-tools","email"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"lending","problem":"Many commercial loan books are still monitored on a calendar. Borrowers send financial statements\nand compliance certificates on fixed dates, often months apart, and analysts rekey them, test\ncovenants in spreadsheets and update the watchlist. By the time a breach shows up in audited numbers\nthe problem can be months old, and the options for the bank and the borrower have narrowed.\n\nMeanwhile the signals were visible elsewhere: falling inflows in the operating account, late\nsupplier payments, a lost customer in the news, a sector downturn, a director resigning. They sit\nin different systems and no one has time to watch them for every borrower. The pandemic made the\ngap obvious, when historic financials said little about which businesses would survive.","problemStats":[],"howItWorks":"1. **Build the obligation calendar.** Covenants, reporting duties and test dates are extracted from\n   facility agreements and kept per borrower.\n2. **Ingest and spread.** Incoming financial statements and compliance certificates are read with\n   document AI, spread into the bank's template and covenants recalculated.\n3. **Watch continuous signals.** Account flows, utilisation, days past due, bureau and registry\n   changes, filings, news and sector indicators are monitored for each borrower and compared with\n   peers.\n4. **Score and explain.** Signals are combined into an early warning score; every alert lists the\n   signals that drove it and links to the evidence.\n5. **Propose an action.** The system drafts a short note with suggested next steps, such as a\n   client call, a covenant waiver discussion or a watchlist review, for the relationship manager.\n6. **Record the outcome.** The relationship manager accepts, changes or dismisses the alert, and\n   the reason is kept for the audit trail and to tune thresholds.","valueDrivers":["risk-reduction","employee-productivity","compliance"],"kpis":["detection-rate-improvement","false-positive-reduction","productivity-gain"],"indicativeValue":{"referenceOrg":"A bank with a USD 5 billion commercial loan book","inputs":[{"key":"book","label":"Commercial loan book","low":5000000000,"high":5000000000,"unit":"USD","note":"The reference bank."},{"key":"defaultRate","label":"Annual default rate","low":0.01,"high":0.02,"unit":"fraction of the book per year","note":"Editorial assumption for a commercial book through the cycle. Replace with your own."},{"key":"lossGivenDefault","label":"Loss given default","low":0.3,"high":0.45,"unit":"fraction of exposure","note":"Editorial assumption. Replace with your own workout data."},{"key":"lossAvoided","label":"Share of default losses avoided through earlier action","low":0.03,"high":0.08,"unit":"fraction of losses","note":"Editorial assumption. No public deployment on this page discloses a measured loss effect, so the range is deliberately small."}],"formula":"book * defaultRate * lossGivenDefault * lossAvoided","currency":"USD","period":"per year","resultLabel":"Credit losses avoided","caveat":"Leaves out analyst time saved on spreading and covenant testing, and the cost of data feeds and implementation. Loss avoidance depends on what the bank actually does with an alert; without a disciplined response process the value is close to zero."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Needs data from core banking, payments, loan administration, documents and external sources, a covenant model per facility, and a workflow that relationship managers actually use. Scores that influence credit decisions need model validation.","dataPrerequisites":["Facility agreements with covenant definitions, or a structured covenant register","Borrower financial statements and compliance certificates in digital form","Account, payment and utilisation data per borrower","External data such as filings, news, bureau, registry and sector indicators","History of past defaults and watchlist moves to calibrate thresholds"],"integrations":["Loan administration and core banking systems","Document management for borrower reporting","External data providers (news, filings, bureau, registries)","Credit workflow or CRM for alerts and actions"]},"implementation":{"steps":[{"title":"Start with covenant testing","detail":"Automate extraction of covenants and recalculation from submitted financials first. It saves analyst time immediately and creates the data backbone for everything else."},{"title":"Add internal behaviour signals","detail":"The bank already owns account inflows, utilisation and payment delays, and they update far more often than borrower financials. Calibrate thresholds on past defaults."},{"title":"Layer external signals carefully","detail":"Add news, filings and sector data per segment, and measure whether each source improves detection or only adds noise."},{"title":"Design the alert to be actionable","detail":"One page per alert: what changed, the evidence, peer context and a proposed next step. Make dismissal require a reason."},{"title":"Close the loop","detail":"Track what happened to every alert and every default that was not alerted, and review thresholds and signals each quarter with credit risk."}],"guardrails":["An alert prompts a human review; it never triggers a downgrade, limit cut or exit on its own","Every alert shows the signals behind it and links to the source evidence","Signal logic and thresholds are documented, versioned and in the model inventory","Relationship manager actions and dismissal reasons are recorded"],"humanInTheLoop":"Relationship managers and credit officers decide what to do with every alert. Credit committee owns watchlist changes, rating changes and restructuring decisions. Credit risk reviews alert quality and missed defaults every quarter.","kpisToInstrument":["Share of defaults that had an alert at least 90 days earlier","Alert precision (alerts that led to an action or a watchlist move)","Time from signal to relationship manager review","Analyst hours spent on spreading and covenant testing"],"failureModes":[{"title":"Alert fatigue","detail":"Too many weak signals and relationship managers stop reading. Measure precision per signal and cut the ones that do not help."},{"title":"Automatic consequences","detail":"An alert that silently lowers a limit or rating creates conduct and legal risk. Keep every consequence behind a human decision."},{"title":"Spreading errors","detail":"A misread line item produces a false covenant breach or hides a real one. Show the source page next to every extracted figure."},{"title":"Watching without acting","detail":"Early warning only pays if the bank has a response playbook for each alert type. Define it before launch."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Monitoring the credit of companies is not listed in Annex III. Where the same system evaluates the creditworthiness of natural persons, such as sole traders or personal guarantors, it falls under Annex III point 5(b) and is high risk; because that evaluation profiles natural persons, the Article 6(3) exemption does not apply."},"regulations":["eba-loan-origination","us-sr-11-7","eu-ai-act","mas-ai-risk-management","gdpr"],"guidance":[{"title":"Guidance to banks on non-performing loans","issuer":"European Central Bank","region":"europe","url":"https://www.bankingsupervision.europa.eu/ecb/pub/pdf/guidance_on_npl.en.pdf","note":"Sets supervisory expectations for early warning indicators, watchlists and early intervention on deteriorating exposures."},{"title":"Guidelines on loan origination and monitoring","issuer":"European Banking Authority","region":"europe","url":"https://www.eba.europa.eu/regulation-and-policy/credit-risk/guidelines-on-loan-origination-and-monitoring","note":"Includes expectations for ongoing credit monitoring and early warning indicators."}],"controls":["Documented signal catalogue with owners, thresholds and validation evidence","Model inventory entry for any score that influences credit decisions","Audit trail of alerts, reviews, actions and dismissal reasons","Quarterly back testing against defaults and watchlist moves"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the monitoring runs as **agentic tasks** and **agentic workflows**: scheduled checks\nper borrower call **custom functions** that read account and utilisation data from the bank's\nsystems, **SQL knowledge bases** hold the covenant register, and the **knowledge base** ingests\nfinancial statements and compliance certificates (PDF, XLSX) for extraction. An **agent** with\n**structured output** recalculates covenants from the extracted figures and drafts the alert note\nwith the evidence attached, and the **web search** tool can bring in recent news for a named\nborrower.\n\nAlerts go to the relationship manager by email or in **Microsoft Teams**, and any proposed action\nwaits for **human in the loop approval**. The **run history and audit trail** record every alert\nand decision, **test suites** check extraction and alert logic on known cases, and **monitors**\nrun scheduled health checks on the agents involved. Deployments can stay in the EU or UAE region."},"faq":[{"question":"How much earlier can AI flag a deteriorating borrower?","answer":"It depends on the signals. OakNorth's CIO described audited financials as lagging during the pandemic and pointed to alternative data that updates much more frequently; the bank's own account and payment data also changes far more often than financial statements. OakNorth said its monitoring often reveals a subset of loans where borrowers who are still current might well be heading for difficulty, and that PNC took it to understand the pandemic's impact across its loan portfolios. No bank on this page has published a measured lead time yet."},{"question":"Should an early warning alert change a credit limit automatically?","answer":"No. Treat an alert as a prompt for review. Automatic limit cuts or downgrades create conduct and legal risk and remove the judgment that restructuring decisions need."},{"question":"Where should a bank start?","answer":"With covenant extraction and testing from submitted financials, because it saves analyst time at once, then with the bank's own account and payment signals, which it already holds and which update far more often than financial statements."}],"related":["credit-memo-drafting-agent","loan-restructuring-recommendations","sme-cash-flow-underwriting","collections-and-hardship-agent","client-briefing-and-call-report-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: tightened the EU AI Act basis (Article 6(3)), removed unsourced claims that account signals are the strongest indicators, attributed the lead time point to OakNorth, corrected the monitors capability and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: restored OakNorth's hedge on still current borrowers, attributed the frequent data point to alternative data only, removed unsourced reporting frequencies and aligned the meta description with SMBC's announced stage."}],"slug":"credit-early-warning-monitoring","url":"https://www.blits.ai/ai-use-cases/credit-early-warning-monitoring","benchmarks":[],"indicativeValueResult":{"low":450000,"high":3600000},"evidence":["oaknorth-bank-continuous-credit-monitoring","pnc-oaknorth-portfolio-monitoring","sumitomo-mitsui-banking-corporation-oaknorth-credit-intelligence"]},{"title":"AI enterprise knowledge search for employees","shortTitle":"Enterprise knowledge search","seoTitle":"Enterprise knowledge search with generative AI","metaDescription":"An AI assistant gives employees cited answers from internal documents they may see. Morgan Stanley says 98% of its Financial Advisor teams adopted its assistant.","definition":"An assistant that lets any employee ask a question in plain language and get a synthesized answer from the organization's own policies, procedures, product manuals and research, with citations to the source documents and only from documents the employee is allowed to see.","aliases":["enterprise search assistant","internal knowledge assistant","policy and procedure assistant","employee copilot for internal knowledge"],"industries":["cross-industry","banking","wealth-and-asset-management","insurance","government","professional-services"],"functions":["knowledge-management","operations","customer-service"],"patterns":["rag-knowledge-assistant","conversational-agent","summarization"],"channels":["internal-tools","microsoft-teams","agent-desktop"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"mainstream","problem":"In a bank or insurer the knowledge that staff need to do their job correctly is scattered across\ncredit and risk policies, operating procedures, product manuals, compliance guidance and internal\nresearch: many documents in several systems, each with versions. Keyword search returns\na list of PDFs. Staff ask a colleague, use an outdated copy, or give a customer a wrong answer.\n\nRetrieval augmented assistants change the interaction: ask a question, get an answer with the\nparagraphs it came from. The hard problems are not the model. They are permissions (an answer\nmust never come from a document the person may not see), currency (the answer must come from\nthe version in force), and trust (people must be able to check the source quickly). Done well,\none governed retrieval layer can serve the service desk, HR, frontline and specialist assistants\ninstead of each building its own.","problemStats":[],"howItWorks":"1. **Ingest and index approved sources.** Policies, procedures and research are ingested from\n   the document management system, intranet and knowledge base, with owner, version and access\n   rights kept as metadata.\n2. **Retrieve with permissions.** When an employee asks, retrieval runs only over documents\n   their role and entitlements allow, combining semantic and keyword search.\n3. **Answer with citations.** The model answers only from the retrieved passages and cites each\n   document and section, so the employee can open the source.\n4. **Refuse when unsure.** If retrieval finds nothing relevant or sources conflict, the assistant\n   says so and points to the owner, rather than guessing.\n5. **Learn from gaps.** Unanswered and badly rated questions go to content owners, who fix or\n   write the missing document.","valueDrivers":["employee-productivity","speed","compliance","customer-experience"],"kpis":["search-time-reduction","handling-time-reduction","employee-adoption","interactions-handled","accuracy","productivity-gain"],"indicativeValue":{"referenceOrg":"A bank with 5,000 employees who regularly look up policies and procedures","inputs":[{"key":"employees","label":"Employees who search internal knowledge regularly","low":5000,"high":5000,"unit":"employees","note":"The reference organization."},{"key":"searchHoursPerWeek","label":"Hours per week each spends finding and reading internal information","low":2,"high":4,"unit":"hours per employee per week","note":"Editorial assumption. Replace with a time study of your own staff."},{"key":"timeSaved","label":"Share of that time saved","low":0.1,"high":0.25,"unit":"fraction of search time","note":"Editorial assumption, replace with a time study of your own. No source on this page reports a measured share of search time saved. For comparison, Google Cloud reports that information searches by less experienced SIGNAL IDUNA agents are 30% faster (read as speed, that is about 23% less time per search, since 1 / 1.3 is about 0.77), and that Wells Fargo's tool reduced the workflow for query resolution by about 20%, without saying whether that is time."},{"key":"workingWeeks","label":"Working weeks per year","low":45,"high":45,"unit":"weeks","note":"Editorial assumption."},{"key":"hourlyCost","label":"Fully loaded cost per hour","low":50,"high":80,"unit":"USD per hour","note":"Editorial assumption. Replace with your own blended cost."}],"formula":"employees * searchHoursPerWeek * workingWeeks * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Employee time released","caveat":"Released time, not cash. It leaves out the value of fewer wrong answers to customers and fewer policy breaches, which is often larger, and the cost of cleaning up and maintaining the content."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"A prototype over a folder of PDFs is quick to build. Production takes longer: connecting several document systems, carrying access rights into the index, handling versions and retirement, and building evaluation sets per domain so answer quality can be measured.","dataPrerequisites":["An inventory of authoritative sources with an owner and review date per document","Access rights per document or collection that can be carried into the index","A set of real questions per domain with expected answers, for evaluation"],"integrations":["Document management and intranet (for example SharePoint or Confluence)","Identity provider and entitlement data for permission aware retrieval","The channels where employees work (Teams, the agent desktop, the intranet)","Feedback routing to content owners"]},"implementation":{"steps":[{"title":"Start with one domain and its owners","detail":"Pick a domain with heavy lookup volume and willing owners, such as operations procedures or product terms. Clean its documents before indexing anything."},{"title":"Carry permissions into retrieval","detail":"Index access rights with every chunk and filter at query time. Test with accounts of different roles that restricted content never appears."},{"title":"Build the evaluation set","detail":"Collect a few hundred real questions with expected answers and sources, and run them on every change to content, retrieval settings or model."},{"title":"Make citations the product","detail":"Show the source passage next to the answer with a link. Staff trust and adopt tools whose answers they can check in seconds."},{"title":"Close the loop with content owners","detail":"Send unanswered questions and negative feedback to owners weekly, and retire documents that are superseded."},{"title":"Offer it as a shared layer","detail":"Expose the same governed retrieval to the other assistants (service desk, HR, contact centre) so permissions, residency and versions are enforced in one place."}],"guardrails":["Permission aware retrieval, tested with role based test accounts","Answers only from retrieved passages, with citations, and refusal when nothing relevant is found","Only the version in force is indexed; superseded documents are removed","Prompt injection defences for content from shared or external sources","Query logs protected and retained according to policy"],"humanInTheLoop":"Content owners are accountable for their documents and review flagged answers. Employees remain responsible for decisions they take on the basis of an answer, and high impact decisions (credit, compliance, customer remediation) still follow their documented approval steps.","kpisToInstrument":["Answer accuracy and citation correctness on the evaluation set, per domain","Share of questions answered versus refused","Weekly active users among target employees","Time to find information in a time study, before and after","Negative feedback and content gaps closed per month"],"failureModes":[{"title":"Oversharing through search","detail":"The assistant surfaces documents that were technically accessible but never meant to be widely read. Review permissions before indexing, not after an incident."},{"title":"Confident answers from old versions","detail":"Superseded policies stay in the index. Index only the version in force and track effective dates."},{"title":"Answers without sources","detail":"Staff cannot verify and either distrust the tool or trust it blindly. Always show the cited passage."},{"title":"Many point solutions","detail":"Every department builds its own index with its own permissions. Build one governed layer and reuse it."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1) requires that people who interact directly with an AI system are informed of it, unless this is obvious from the context, as it usually is for an internal assistant. The system would be high risk only if it were intended for an Annex III purpose, such as assessing the creditworthiness of natural persons (point 5(b)) or making decisions on or evaluating workers (point 4(b))."},"regulations":["eu-ai-act","gdpr","dora","iso-42001","nist-ai-rmf","mas-ai-risk-management"],"guidance":[{"title":"Artificial Intelligence (AI) Model Risk Management, information paper","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Good practices for AI and generative AI model risk management observed in a thematic review of banks in mid 2024, covering governance and oversight, risk management systems and processes, and development and deployment."},{"title":"OWASP Top 10 for LLM Applications","issuer":"OWASP Gen AI Security Project","region":"global","url":"https://genai.owasp.org/llm-top-10/","note":"The 2025 list covers prompt injection (including through retrieved content), sensitive information disclosure and vector and embedding weaknesses, the main security risks of retrieval assistants."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Providers must design AI systems that interact directly with people so that those people are informed they are interacting with AI, unless this is obvious from the context. Applies from 2 August 2026."}],"controls":["Inventory entry with an accountable owner and the list of indexed sources","Permission tests per role before each new source is added","Evaluation set runs on every change to content, retrieval or model","Document ownership and review dates enforced for indexed content","Monitoring of refusals, negative feedback and unusual query patterns"],"incidents":[{"title":"CVE-2025-32711: AI command injection in Microsoft 365 Copilot","url":"https://nvd.nist.gov/vuln/detail/CVE-2025-32711","note":"A vulnerability recorded by NVD in June 2025, not a reported breach: AI command injection in Microsoft 365 Copilot allowed an unauthorized attacker to disclose information over a network. It shows that an enterprise assistant can be made to disclose information through injected instructions."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** over a **knowledge base** with **hybrid retrieval** (dense\nvectors plus BM25), ingesting PDFs, Office documents, email files and crawled web pages,\nand a central document library with **version control** and revert, so owners can replace\nsuperseded documents. SQL knowledge bases add structured data where answers need numbers.\nSharePoint and OneDrive are available as ready made tools, and Confluence is in the integration\ncatalog.\n\nThe same agent serves **Microsoft Teams**, the intranet through the web widget, and other\nassistants through the **REST API**, so one retrieval layer backs several front doors.\n**Guardrails** block prompt injection and unsafe output, **PII masking** protects personal\ndata, and **role based access control** limits who manages which agents and content. **Test\nsuites** with LLM grading that uses knowledge base evidence run the evaluation set on every\nchange, and **analytics** with response feedback show untrained questions and unexpected answers\nfor content owners. The platform is model agnostic and\ncan run in the EU or UAE region."},"faq":[{"question":"How is this different from the search we already have?","answer":"Search returns documents; the assistant returns an answer with the paragraphs it came from. Google Cloud reports that Wells Fargo's retrieval tool for branch bankers reduced the workflow for query resolution by about 20%, and that information searches by less experienced SIGNAL IDUNA service agents are 30% faster."},{"question":"Will employees actually use it?","answer":"Two wealth managers have published usage figures. Morgan Stanley said in June 2024 that 98% of its Financial Advisor teams had adopted its AI @ Morgan Stanley Assistant. Bank of America reports more than 23 million interactions in 2024 with ask MERRILL and ask PRIVATE BANK, a volume figure that does not say what share of employees use the tools."},{"question":"How do we stop it from showing confidential documents?","answer":"Carry each document's access rights into the index and filter at query time, then test with accounts of different roles. Review what is technically accessible before you index it, because an assistant makes forgotten oversharing easy to find."}],"related":["wealth-advisor-knowledge-assistant","hr-and-policy-assistant","it-service-desk-resolution-agent","live-agent-assist","support-knowledge-article-generation","governed-text-to-sql-analytics"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog as an industry neutral page, with four evidence records verified against their sources. The catalog's Morgan Stanley and Discover figures could not be verified on reachable pages and were not used as metrics."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; made the EU AI Act basis precise (Article 50(1), Annex III points 4(b) and 5(b)); corrected guidance notes; added CVE-2025-32711 as an incident; removed an unsupported document count, a build time claim and an unsupported adoption claim in the FAQ; aligned the Blits.ai section with the feature inventory; tightened evidence summaries and added an archived copy of the Morgan Stanley release."},{"date":"2026-09-27","note":"Editor review: removed three metrics the quotes do not state as recorded (Wells Fargo \"workflow for query resolution\" as handling time, SIGNAL IDUNA \"30% faster\" as a 30% search time reduction, and a computed 88.9% escalation reduction); both figures stay in the evidence summaries as the source words them. Dated the Wells Fargo and SIGNAL IDUNA records to 2025 from Wayback snapshots of the Google Cloud list. Rewrote the metaDescription, the first two FAQ answers and the time saved note so none presents the Wells Fargo figure as time; added search time reduction to the KPIs; changed crawled intranet pages to crawled web pages in the Blits.ai section; clarified the Morgan Stanley OpenAI relationship."}],"slug":"enterprise-knowledge-search","url":"https://www.blits.ai/ai-use-cases/enterprise-knowledge-search","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":23000000,"min":23000000,"max":23000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-ask-merrill-and-ask-private-bank","pooled":true}]}],"indicativeValueResult":{"low":2250000,"high":18000000},"evidence":["bank-of-america-ask-merrill-and-ask-private-bank","morgan-stanley-ai-assistant-knowledge-search","signal-iduna-co-si-knowledge-assistant","wells-fargo-branch-policy-retrieval"]},{"title":"AI examination of trade documents under letters of credit and collections","shortTitle":"Trade document examination","seoTitle":"AI for letter of credit document checking","metaDescription":"AI reads letter of credit presentations, cross checks them against the credit, UCP 600 and ISBP, and lists discrepancies. RMB went live with Traydstream in 2022.","definition":"AI that reads the full document presentation under a letter of credit or collection (bill of lading, commercial invoice, packing list, certificates), extracts and cross checks the data, tests it against the instructions and the ICC rules (for letters of credit, the credit terms, UCP 600 and ISBP), and lists discrepancies by severity with the rule cited, so qualified examiners focus on the genuine exceptions.","aliases":["letter of credit document checking","automated LC examination","trade document discrepancy checking","documentary credit automation"],"industries":["banking"],"functions":["operations","regulatory-compliance"],"patterns":["document-processing","classification-and-routing","agentic-workflow","summarization"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"specialized-businesses","problem":"Documentary trade still runs on paper. For every presentation under a letter of credit, a trained\nexaminer reads each document, compares names, dates, quantities, amounts, ports and goods\ndescriptions across them and against the credit, and applies a large body of international\npractice to decide whether the presentation complies. A missed discrepancy can leave the bank\npaying against documents its client may refuse to reimburse; an unnecessary one delays the\nclient's money.\n\nRMB's head of trade describes the checking of numerous unstructured trade documents as manual\nand extremely time consuming, and Standard Bank Group's head of trade presents automation as a way\nto minimise repetitive tasks. Microsoft notes that traditional OCR and template based systems can\nstruggle when layouts change or data is missing. Modern document AI plus a rules engine can do the extraction and the mechanical\ncross checks. In our view the main gain is examiner time freed for the genuinely ambiguous cases,\nas long as the contractual judgement on those stays with a qualified examiner.","problemStats":[{"statement":"Microsoft, citing ICC United Kingdom, states that an average international trade shipment can involve up to 50 separate documents exchanged between as many as 30 different stakeholders.","sourceTitle":"Reimagining trade finance with AI: A collaborative proof of concept from Microsoft, ANZ, HSBC, and Lloyds","sourceUrl":"https://www.microsoft.com/en-us/microsoft-cloud/blog/financial-services/2026/04/20/reimagining-trade-finance-with-ai-a-collaborative-proof-of-concept-from-microsoft-anz-hsbc-and-lloyds/","year":2026}],"howItWorks":"1. **Ingest the presentation.** Scanned and digital documents are classified by type and read with\n   OCR and document AI, whatever their layout.\n2. **Extract and normalise.** Parties, amounts, currencies, dates, ports, goods descriptions,\n   marks and quantities are extracted and normalised.\n3. **Cross check.** Data is compared across documents and against the credit terms (the MT700\n   fields), for example amount and currency, shipment dates, and consistency of goods descriptions.\n4. **Apply the rules.** A rules engine mapped to UCP 600 and ISBP tests each finding, and a\n   language model helps with free text comparisons such as whether two goods descriptions conflict.\n5. **Report discrepancies.** Findings are listed by severity with the rule or credit clause cited;\n   clean presentations go to a lighter review, exceptions to a qualified examiner who decides.","valueDrivers":["speed","employee-productivity","risk-reduction","cost-to-serve"],"kpis":["processing-time-reduction","handling-time-reduction","accuracy","automation-rate","error-reduction"],"indicativeValue":{"referenceOrg":"A bank examining 20,000 documentary credit presentations a year","inputs":[{"key":"presentations","label":"Presentations examined per year","low":20000,"high":20000,"unit":"presentations per year","note":"The reference bank."},{"key":"hoursPerPresentation","label":"Examiner hours per presentation","low":1.5,"high":3,"unit":"hours per presentation","note":"Editorial assumption, replace with your own time study. Complex presentations take much longer."},{"key":"reduction","label":"Share of examiner time saved","low":0.3,"high":0.5,"unit":"fraction of time","note":"Editorial assumption, replace with your own pilot results; banks on this page have not published measured figures."},{"key":"hourlyCost","label":"Loaded cost of an examiner hour","low":60,"high":90,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."}],"formula":"presentations * hoursPerPresentation * reduction * hourlyCost","currency":"USD","period":"per year","resultLabel":"Examiner time released, valued at loaded cost","caveat":"Values examiner time only. It leaves out the cost of the platform and integration, faster payment for clients, fewer missed discrepancies and the value of scaling without hiring scarce specialists."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Document variety and quality are the main difficulty, followed by encoding UCP and ISBP practice as testable rules and integrating with the trade processing system and SWIFT messages.","dataPrerequisites":["A library of past presentations with the examiners' decisions, for testing","The bank's discrepancy taxonomy and severity rules","Current credit terms from the trade system in structured form"],"integrations":["Trade finance processing system","SWIFT messaging (MT700 and related)","Document scanning and management","Trade screening, so compliance checks run on the same extracted data"]},"implementation":{"steps":[{"title":"Build a gold standard set","detail":"Collect past presentations with the examiners' final findings, including disputed ones, to measure extraction accuracy and discrepancy recall before go live."},{"title":"Separate mechanical from judgement checks","detail":"Automate the mechanical checks (amounts, dates, names, consistency) first and route judgement calls, such as whether a variation is a discrepancy, to examiners with the evidence."},{"title":"Version the rules","detail":"Keep every rule mapped to its UCP, ISBP or credit clause source, versioned, so a past decision can be reproduced with the rules in force at the time."},{"title":"Run in shadow mode","detail":"Let the system check live presentations in parallel with examiners for a period and compare findings before changing the workflow."},{"title":"Share the extraction with compliance","detail":"Feed the extracted data to trade screening so compliance and examination work from one version of the facts."}],"guardrails":["A qualified examiner decides on every discrepancy and signs off every refusal notice","Every finding cites the document, field and rule or credit clause behind it","Rule versions and model versions are logged per presentation","Low confidence extraction is shown as such and routed to manual review"],"humanInTheLoop":"Examiners review every exception and decide on ambiguous discrepancies. Clean presentations get a lighter human review until error rates are proven low, and a sample of automatically cleared presentations is re examined every month.","kpisToInstrument":["Examination time per presentation, clean and with discrepancies","Discrepancy recall and precision against examiner findings","Extraction accuracy per field and document type","Share of presentations cleared with lighter review","Refusals later disputed or overturned"],"failureModes":[{"title":"Missed discrepancy on a clean looking presentation","detail":"The system misses a subtle inconsistency and the presentation is waved through. Sample cleared presentations and track recall on the gold standard set."},{"title":"Discrepancy noise","detail":"Too many trivial findings and examiners start ignoring them. Tune severity and suppress findings that practice treats as non discrepant."},{"title":"Poor scans","detail":"Low quality images break extraction. Detect image quality and route poor scans to manual review."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Checking trade documents for compliance with credit terms is not listed in Annex III and does not decide about natural persons. AI literacy duties under Article 4 apply, and the process falls under the bank's operational resilience and model governance."},"regulations":["eu-ai-act","dora","mas-ai-risk-management","apra-cps-230","iso-42001"],"guidance":[{"title":"ICC trade finance rules and standards","issuer":"International Chamber of Commerce","region":"global","url":"https://iccwbo.org/business-solutions/trade-finance/","note":"The ICC publishes UCP 600, ISBP and URC 522 (collections), the rule base the checks are mapped to; the examiner's judgement under them stays with people."},{"title":"MAS Guidelines for Artificial Intelligence (AI) Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Consultation paper of November 2025 proposing supervisory expectations for all financial institutions on AI inventories, risk materiality, evaluation and testing, human oversight and monitoring."}],"controls":["Model and rule inventory with owners, versions and validation results","Examiner sign off recorded per presentation","Monthly sampling of automatically cleared presentations","Retention of documents, findings and rule versions for the statutory period"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that the trade system triggers through an API token.\nA **custom function** fetches the presentation and the text from the bank's scanning and OCR\nservice through a REST call, an agent extracts the fields with **structured output**, and the\ndeterministic cross checks run as **custom functions** in the isolated code sandbox. A **knowledge base** with hybrid\nretrieval holds the bank's examination guidelines and discrepancy examples, which the agent uses\nfor free text comparisons and to cite the rule behind each finding.\n\nFindings go back to the examiner through the trade system via REST calls, and that action can be\nset to require **human in the loop confirmation** (approve and reject controls); the examiner\nstill decides every discrepancy. The per run **audit trail** keeps inputs, outputs and tool calls,\n**test suites** can replay the gold standard set against the workflow on each change, and\n**monitors** can run scheduled checks with known inputs against the extraction agent to catch drift. The platform is model agnostic, so each agent\ncan use a different model, and data can stay in the EU or UAE region."},"faq":[{"question":"Which banks use AI to check trade documents?","answer":"RMB (Rand Merchant Bank) announced in September 2022 that it had gone live on Traydstream's AI enabled trade finance platform, which automates trade document checking, and Stanbic Bank Uganda (Standard Bank Group) signed an agreement in 2021 to implement the same platform after months of trade document processing on it. On the corporate side, ANZ, HSBC and Lloyds built a proof of concept with Microsoft in which an AI agent in the company's ERP cross checks a letter of credit against invoice and shipping data before sending structured data to the bank."},{"question":"Does AI decide whether a presentation complies?","answer":"It should not decide ambiguous cases. It extracts, cross checks and lists discrepancies with the rule cited; a qualified examiner decides and signs off, because the bank stays responsible for honouring or refusing the presentation whatever tool it uses."},{"question":"How fast is automated checking?","answer":"The banks on this page have not published measured figures. The RMB and Traydstream announcements speak of faster processing and improved turnaround times without numbers, so measure examination time on your own presentations in shadow mode."}],"related":["trade-finance-crime-screening","intelligent-document-processing","correspondence-triage-and-routing","payment-investigations-and-exceptions"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version. The catalog's ICC and WTO survey and the four to ten hours claim could not be verified from a primary source (the study returned 403 and the time claim appears only in trade press), so neither is used as a number."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed the unsourced vendor time claim and the claim that examiners are ageing, limited UCP 600 and ISBP to letters of credit, corrected the MAS note and the Stanbic Bank Uganda stage (pilot, implementation agreed), and added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: attributed the time consuming quote to RMB's head of trade only and marked the point about ambiguous cases as editorial, cited RMB's own media release (archived copy), corrected the FAQ wording on Stanbic and the proof of concept, and aligned the Blits.ai build description with the feature inventory."},{"date":"2026-09-27","note":"Fact checked against sources again: attributed the OCR and template limitation to Microsoft, described RMB's go live as the Traydstream platform rather than a confirmed checking use, and credited the turnaround wording to both the RMB and Traydstream releases."}],"slug":"trade-document-examination","url":"https://www.blits.ai/ai-use-cases/trade-document-examination","benchmarks":[],"indicativeValueResult":{"low":540000,"high":2700000},"evidence":["anz-hsbc-lloyds-trade-finance-agent-proof-of-concept","rmb-automated-trade-document-checking","standard-bank-automated-trade-document-checking"]},{"title":"AI financial wellbeing coach in the banking app","shortTitle":"Financial wellbeing coach","seoTitle":"AI financial wellbeing coach for banking apps","metaDescription":"An in app AI coach that explains spending, forecasts bills and helps customers save. See how Bank of America's Erica and RBC's NOMI do it, with the risks.","definition":"An in app AI assistant that the customer opens to understand their own money: it uses the customer's transaction data to explain their spending, forecast upcoming bills and cash flow, set and track savings goals and answer money questions in plain language, staying on the guidance side of the line between guidance and regulated financial advice.","aliases":["AI money coach","personal finance assistant","financial health assistant","proactive insights"],"industries":["banking"],"functions":["customer-service","marketing"],"patterns":["conversational-agent","recommendation-and-personalization","prediction-and-scoring","agentic-workflow"],"channels":["mobile-app"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"Banks see a detailed picture of a customer's financial life in their transaction data. A customer\nwho is surprised by a direct debit, or who does not know how much they can safely save, is often\nleft to work it out from a list of transactions.\n\nForrester's 2025 review of the mobile apps of the four biggest Australian banks found that most\nstill fall short in helping customers improve their overall financial health, while leading banks\nincreasingly use AI powered insights to help customers stay on top of their finances. The\nopportunity is a coach that does the analysis for the customer, speaks up at the right moment and\ncan act on a simple instruction, without drifting into selling or unlicensed advice.","problemStats":[{"statement":"Forrester's 2025 review of Australian mobile banking apps found that most banks still fall short in helping customers improve their overall financial health.","sourceTitle":"Conversational AI And Anticipatory Insights, What's New In Australian Mobile Banking In 2025","sourceUrl":"https://www.forrester.com/blogs/conversational-ai-and-anticipatory-insights-whats-new-in-australian-mobile-banking-in-2025/","year":2025}],"howItWorks":"1. **Understand the customer's money.** Models categorise transactions, detect income, bills and\n   subscriptions, and forecast the balance over the coming days.\n2. **Speak up at the right moment.** Proactive insights flag a bill that will not be covered, a\n   subscription price rise or a month of unusual spending, with an action attached.\n3. **Answer in conversation.** The customer asks \"can I afford this trip\" or \"where did my money\n   go\" and gets an answer grounded in their own data and the bank's approved guidance content.\n4. **Act within limits.** On instruction it sets up a savings goal, a transfer into a savings pot\n   or a budget, using the same authenticated APIs as the app, and confirms before moving money.\n5. **Know the boundary.** Questions that need regulated advice (investments, pensions, debt\n   solutions) or show signs of financial difficulty go to a human or to the right service.","valueDrivers":["customer-experience","revenue-growth","inclusion-and-access"],"kpis":["users-served","interactions-handled","customer-satisfaction","nps-change","churn-reduction"],"indicativeValue":{"referenceOrg":"A retail bank with 1 million digitally active customers","inputs":[{"key":"customers","label":"Digitally active customers","low":1000000,"high":1000000,"unit":"customers","note":"The reference bank."},{"key":"engagedShare","label":"Share of customers who use the coach regularly","low":0.1,"high":0.25,"unit":"fraction of customers","note":"Editorial assumption. For scale, RBC reports more than 900,000 clients used NOMI Forecast in its first 19 months."},{"key":"churnPoints","label":"Reduction in annual attrition among engaged users","low":0.005,"high":0.01,"unit":"fraction of engaged customers per year","note":"Editorial assumption; no deployment on this page discloses a retention effect. Measure it with a control group."},{"key":"valuePerCustomer","label":"Annual revenue of a retained main bank customer","low":100,"high":200,"unit":"USD per customer per year","note":"Editorial assumption, replace with your own customer economics."}],"formula":"customers * engagedShare * churnPoints * valuePerCustomer","currency":"USD","period":"per year","resultLabel":"Revenue retained through lower attrition","caveat":"Retention value only, and the most uncertain input is the attrition effect. It leaves out fees customers avoid (a benefit to them, not the bank), deposit growth from savings features, contact centre calls avoided and the cost of building and running the coach."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Categorisation and forecasting models are well understood; the difficulty is quality (a wrong forecast destroys trust), the boundary with regulated advice, and making insights useful rather than noisy.","dataPrerequisites":["Transaction history with reliable merchant and category enrichment","Scheduled payments, direct debits and income patterns","Approved guidance content on budgeting, saving and financial difficulty","Customer consent and preferences for proactive messages"],"integrations":["Core banking and payments data (read) and savings and transfer APIs (write, with confirmation)","Transaction enrichment and categorisation service","Notification and in app messaging","Referral routes to advice, debt support and the contact centre"]},"implementation":{"steps":[{"title":"Start with forecasting and alerts","detail":"A reliable view of upcoming bills and the projected balance is useful on its own and needs no conversation. RBC's NOMI Forecast (a seven day view of upcoming payments) and Erica's alerts on where balances are trending over the next seven days at Bank of America are examples."},{"title":"Add conversation over the customer's own data","detail":"Let customers ask about their spending and plans, with answers computed from their data by deterministic functions and explained by the model, never estimated by it."},{"title":"Draw the advice line in writing","detail":"With compliance, list what the coach may say (facts, general guidance, the bank's own product features) and what triggers a referral (investment, pension or debt advice, signs of financial difficulty)."},{"title":"Let it act with confirmation","detail":"Add savings goals and transfers between the customer's own accounts, each confirmed by the customer, before anything more autonomous."},{"title":"Measure outcomes, not clicks","detail":"Track fees avoided, savings built and financial difficulty referrals against a control group, not just insight views."}],"guardrails":["Numbers come from deterministic calculations on the customer's data, never from the model's own arithmetic","A written boundary between guidance and regulated advice, with automatic referral when it is crossed","No sales messages disguised as coaching; product offers are labelled and follow marketing consent","Signs of financial difficulty trigger support routes, not product offers","Any money movement is confirmed by the customer and limited to their own accounts"],"humanInTheLoop":"Customers approve every action. Humans take over for regulated advice and financial difficulty. A conduct and quality team reviews samples of insights and conversations for accuracy, tone and any drift towards selling.","kpisToInstrument":["Regular users and repeat use of insights","Forecast accuracy on upcoming balances","Fees avoided and savings built by users versus a control group","Referrals to advice and financial difficulty support","Complaints and satisfaction about the coach"],"failureModes":[{"title":"The coach becomes a sales channel","detail":"Insights turn into product pushes and customers stop trusting them. Separate coaching from offers and review the mix."},{"title":"Wrong numbers","detail":"A misclassified income or a missed direct debit gives a wrong forecast. Compute with deterministic functions, show the basis and let customers correct it."},{"title":"Advice by accident","detail":"The assistant recommends an investment or a debt product in a way that counts as regulated advice. Enforce the boundary with guardrails and tests."},{"title":"Alert fatigue","detail":"Too many nudges and customers mute them all. Cap frequency and measure action rates per insight type."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The conversational assistant carries the Article 50 transparency duty: customers must be told they are interacting with an AI system. The system becomes high risk if it is used to evaluate the creditworthiness of natural persons or establish their credit score (Annex III point 5(b)). Article 5(1)(b) prohibits AI that exploits vulnerabilities due to a person's specific social or economic situation to materially distort their behaviour in a way that causes, or is reasonably likely to cause, significant harm."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","mas-ai-risk-management","iso-42001"],"guidance":[{"title":"Advice Guidance Boundary Review","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/firms/advice-guidance-boundary-review","note":"The joint HM Treasury and FCA review of the boundary between financial advice and other forms of support. Its targeted support rules, confirmed as final on 26 February 2026 and expected by the FCA to take effect from 6 April 2026, let firms that hold the new targeted support permission suggest options on pensions and retail investments to groups of customers with common characteristics, which bears on how far a coach may go."},{"title":"Article 5: Prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Bans manipulative techniques and the exploitation of vulnerabilities, including those due to a person's economic situation."}],"controls":["Inventory entry for the coach and its personalisation models with an accountable owner","Documented guidance and advice boundary signed off by compliance","Fairness monitoring of insights and referrals across customer segments","Accuracy monitoring of forecasts and categorisation","Consent management for proactive messages and use of data"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the coach is an **AI agent** whose numbers come from **custom functions** that\ncall the bank's data and forecasting services, or from a **SQL knowledge base** the agent can\nquery, so the model explains figures rather than calculating them. Guidance content sits in a\n**knowledge base** with hybrid retrieval. Actions such as creating a savings goal run as\n**flows** or **agentic workflows** with a tool execution policy and **human in the loop**\nconfirmation above a threshold.\n\nThe coach lives in the **mobile app** through the REST or WebSocket API channel, with\nrich cards for balances, charts and insight panels and optional **voice**. **Guardrails**\nenforce the advice boundary and block sales language in coaching answers, **PII masking**\nprotects transaction data, **test suites** check the boundary cases on every change, and\n**analytics** track use and satisfaction. The platform is model agnostic, with EU and UAE\ndata residency."},"faq":[{"question":"Is a financial wellbeing assistant giving financial advice?","answer":"It should not be. Explaining a customer's own data, general guidance and the bank's own features is guidance; recommending a specific investment or debt product for a customer's circumstances can be regulated advice. Write the boundary down and test it."},{"question":"Which banks run AI financial coaches today?","answer":"Bank of America's Erica has delivered more than 1.7 billion proactive, personalized insights (from a predefined set of responses, without generative AI), and RBC's NOMI forecasts cash flow and helps clients put money aside with Find and Save. Starling added smart tools to its agentic assistant in August 2026 and says it will release new ones every week for the rest of 2026."},{"question":"How do you prove it helps customers?","answer":"Compare users with a control group on fees avoided, savings built and financial difficulty outcomes, not on clicks. Engagement alone can hide a coach that mostly sells."}],"related":["proactive-outbound-engagement-agent","offers-and-rewards-agent","goal-based-financial-planning-assistant","collections-and-hardship-agent","account-and-card-servicing-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog and verified against the sources. The catalog's Forrester source is analyst commentary and is used only as context."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; made the EU AI Act basis precise (Article 50 duty, Annex III point 5(b), Article 5(1)(b)); updated the FCA boundary review note with the final targeted support rules; removed unsupported wording on budgeting tools, NOMI automated saving and the webview channel; corrected the Starling and RBC evidence summaries."},{"date":"2026-09-27","note":"Review fixes: attributed the problem statement to the single Forrester source and dropped the unsourced superlative; worded the Starling weekly tools as a stated plan; noted the FCA targeted support permission and start date; Erica now records 3 billion client interactions instead of mapping 1.7 billion insights to interactions, and is recorded as mobile app only per Bank of America's Erica page; RBC channels sourced from the NOMI page."}],"slug":"financial-wellbeing-coach","url":"https://www.blits.ai/ai-use-cases/financial-wellbeing-coach","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":3,"nUpTo":0,"median":12600000,"min":10000000,"max":3000000000,"byClaimant":{"organization":3,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-erica-virtual-assistant","pooled":true},{"id":"hyundai-card-ai-annual-statement","pooled":true},{"id":"rbc-nomi-personal-insights","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":25450000,"min":900000,"max":50000000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-erica-virtual-assistant","pooled":true},{"id":"rbc-nomi-personal-insights","pooled":true}]}],"indicativeValueResult":{"low":50000,"high":500000},"evidence":["bank-of-america-erica-virtual-assistant","commonwealth-bank-customer-engagement-engine","hyundai-card-ai-annual-statement","rbc-nomi-personal-insights","starling-assistant-agentic-financial-assistant","westpac-app-ai-nudges"]},{"title":"AI for AML transaction monitoring alert triage","shortTitle":"AML alert triage","seoTitle":"AI for AML transaction monitoring alert triage","metaDescription":"AI scores AML alerts, closes clear false positives and ranks the rest for investigators. HSBC reports 60% fewer false positive cases using machine learning.","definition":"Machine learning and AI agents that score anti money laundering alerts for genuine risk, close clear false positives with a written and stored rationale, and hand investigators the remaining alerts already enriched with the customer, counterparty and transaction context.","aliases":["AML alert scoring","transaction monitoring alert prioritisation","AML false positive reduction","level one AML investigation automation"],"industries":["banking","payments"],"functions":["financial-crime-compliance"],"patterns":["prediction-and-scoring","anomaly-detection","agentic-workflow","summarization"],"channels":["internal-tools","agent-desktop"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"middle-office","problem":"Transaction monitoring takes a large share of the effort in anti money laundering. Rule based\nscenarios (thresholds, rapid movement of funds, high risk geographies) are tuned to miss nothing,\nso they generate large volumes of alerts, and the great majority close as false positives after\nan analyst has pulled statements, looked up counterparties and written a note.\n\nThe cost is not only money. Investigators spend their time clearing noise, real laundering\nhides in the backlog, and the quality of the written rationale varies from analyst to analyst,\nwhich is exactly what supervisors test. Adding people does not scale with payment volumes on\ninstant rails.","problemStats":[{"statement":"The UN Office on Drugs and Crime reported in 2011 that criminals may have laundered around USD 1.6 trillion, or 2.7% of global GDP, in 2009, consistent with a 2 to 5% of global GDP range previously established by the International Monetary Fund.","sourceTitle":"UNODC estimates that criminals may have laundered US$ 1.6 trillion in 2009","sourceUrl":"https://www.unodc.org/unodc/en/press/releases/2011/October/unodc-estimates-that-criminals-may-have-laundered-usdollar-1.6-trillion-in-2009.html","year":2011},{"statement":"Google Cloud, citing industry reporting, stated in 2023 that more than 95% of system generated alerts turn out to be false positives in the first phase of review, and that about 98% never lead to a suspicious activity report.","sourceTitle":"Google Cloud Launches AI-Powered Anti Money Laundering Product for Financial Institutions","sourceUrl":"https://www.googlecloudpresscorner.com/2023-06-21-Google-Cloud-Launches-AI-Powered-Anti-Money-Laundering-Product-for-Financial-Institutions","year":2023},{"statement":"The Hong Kong Monetary Authority reported in November 2025 that 48 authorized institutions had assessed AI for transaction monitoring and that more than 30% of authorized institutions had already adopted it in their monitoring systems, with use cases concentrated in risk detection and alert prioritisation.","sourceTitle":"Supporting Artificial Intelligence Adoption in AML/CFT","sourceUrl":"https://brdr.hkma.gov.hk/eng/doc-ldg/docId/getPdf/20251118-3-EN/20251118-3-EN.pdf","year":2025}],"howItWorks":"1. **Score the alert.** A model trained on past alert outcomes, and increasingly on the full\n   customer and transaction picture rather than the rule hit alone, scores each alert for the\n   likelihood that it leads to a suspicious activity report.\n2. **Enrich it.** An agent gathers the evidence an analyst would: customer due diligence data,\n   expected activity, the transactions behind the alert, counterparties, screening results and\n   previous alerts or reports on the customer.\n3. **Draft the rationale.** For each alert the agent writes a structured narrative: what\n   triggered it, what the evidence shows and a proposed disposition, with every fact linked to\n   its source record.\n4. **Close or escalate under policy.** Alerts that meet approved low risk criteria are closed\n   with the stored rationale; the rest go to investigators, ranked by risk, with the evidence\n   pack attached.\n5. **Learn and assure.** Investigator decisions and quality assurance findings are fed back, and\n   a random sample of closed alerts is reviewed independently.","valueDrivers":["compliance","employee-productivity","cost-to-serve","risk-reduction"],"kpis":["false-positive-reduction","alert-volume-reduction","detection-rate-improvement","processing-time-reduction","automation-rate","productivity-gain"],"indicativeValue":{"referenceOrg":"A mid sized bank working 100,000 transaction monitoring alerts a year","inputs":[{"key":"alerts","label":"Transaction monitoring alerts per year","low":100000,"high":100000,"unit":"alerts per year","note":"The reference bank."},{"key":"hoursPerAlert","label":"Analyst hours per alert today","low":0.5,"high":1,"unit":"hours per alert","note":"Editorial assumption for level one review including documentation. Replace with your own time study."},{"key":"alertReduction","label":"Share of alert workload removed by scoring and auto closure","low":0.2,"high":0.4,"unit":"fraction of alerts","note":"Conservative against HSBC's reported 60% fewer false positive cases, because most banks keep rules in place alongside the model at first."},{"key":"costPerHour","label":"Fully loaded analyst cost per hour","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption. Replace with your own."}],"formula":"alerts * hoursPerAlert * alertReduction * costPerHour","currency":"USD","period":"per year","resultLabel":"Investigation capacity released","caveat":"Counts analyst time only. It leaves out the value of finding more genuine laundering, lower regulatory risk, the cost of model validation and data work, and the effort of running rules and models in parallel during the transition."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Scoring and drafting are proven, but auto closing an AML alert is a regulated decision. Expect model validation, a parallel run against the existing process and a conversation with the supervisor before any alert is closed without a human.","dataPrerequisites":["Several years of alerts with final dispositions, and which ones led to a report","Customer due diligence data, expected activity and risk rating","Transaction and counterparty data at the level of detail the investigator uses","Written investigation procedures and quality assurance standards"],"integrations":["Transaction monitoring system","AML case management","Customer due diligence and KYC records","Core banking and payment data","Screening results (sanctions, PEP, adverse media)"]},"implementation":{"steps":[{"title":"Measure the baseline","detail":"Record alert volumes, conversion to reports, handling time and quality assurance findings per scenario. Without this baseline no one can show the model is better, including to the supervisor."},{"title":"Build the evidence pack before the score","detail":"Start with an agent that enriches alerts and drafts rationales for investigators. It delivers time savings early, carries little regulatory risk and generates the labelled data for scoring."},{"title":"Validate the scoring model","detail":"Train on historical dispositions, test against a hold out period, and have model risk validate it, including a check that alerts later reported as suspicious would not have been closed."},{"title":"Run in parallel","detail":"Score and draft on live alerts while investigators still work everything, and compare decisions for at least one full quarter before closing anything automatically."},{"title":"Introduce governed auto closure","detail":"Close only the lowest risk band, per scenario, with a stored rationale and independent sampling, and report the results to the money laundering reporting officer."}],"guardrails":["No alert on a high risk customer, a sanctions nexus or a prior report is closed without a human","Every closure stores the rationale, the evidence and the model version used","Independent sampling of auto closed alerts, with a threshold that stops auto closure","Facts in drafted narratives link to source records, and unsupported claims are rejected","Model inventory entry, validation and ongoing performance monitoring"],"humanInTheLoop":"Investigators decide every escalated alert and every filing. The money laundering reporting officer approves the auto closure policy and receives sampling results; model risk validates the scoring model and every material change.","kpisToInstrument":["Alert volume and false positive rate per scenario, before and after","Conversion from alert to suspicious activity report","Investigator handling time per alert and per case","Error rate found in independent sampling of auto closed alerts","Share of reports that came from alerts the model ranked low"],"failureModes":[{"title":"Training on yesterday's decisions","detail":"A model trained on past dispositions learns past blind spots. Include alerts later reported through other routes and review the lowest scored band regularly."},{"title":"Defensive auto closure rates","detail":"Teams close too little automatically to show any benefit, or too much to satisfy a target. Set the band from validation results, not from a savings goal."},{"title":"Rationales that read well but prove nothing","detail":"Fluent narratives without evidence links do not survive an audit. Require citations to source records in every draft."},{"title":"Rules and models never reconciled","detail":"Running both forever doubles cost. Plan when rules are retired or retuned based on the parallel run."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"AML transaction monitoring is not listed in Annex III; point 5(b) covers creditworthiness and credit scoring and excludes systems used to detect financial fraud. The Article 5(1)(d) ban on predicting criminal offences from profiling alone does not apply to systems that support a human assessment already based on objective and verifiable facts linked to criminal activity, which is how alert triage should be designed. A decision to restrict an account taken solely by automated means would fall under GDPR Article 22 and national AML law, so consequential decisions need human review."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","dora","us-sr-11-7","mas-ai-risk-management","nist-ai-rmf","iso-42001","eu-amlr","us-bsa","mas-notice-626"],"guidance":[{"title":"Supporting Artificial Intelligence Adoption in AML/CFT","issuer":"Hong Kong Monetary Authority","region":"asia-pacific","url":"https://brdr.hkma.gov.hk/eng/doc-ldg/docId/getPdf/20251118-3-EN/20251118-3-EN.pdf","note":"Reports adoption of AI in transaction monitoring by Hong Kong authorized institutions and announces workshops on risk detection, alert prioritisation and generative AI for compiling suspicious transaction reports."},{"title":"Joint Statement Encouraging Innovative Industry Approaches to AML Compliance","issuer":"FinCEN and the US federal banking agencies","region":"north-america","url":"https://www.fincen.gov/news/news-releases/treasurys-fincen-and-federal-banking-agencies-issue-joint-statement-encouraging","note":"Innovative pilot programs should not in themselves subject banks to supervisory criticism, even if they prove unsuccessful."},{"title":"Principles for Using Artificial Intelligence and Machine Learning in Financial Crime Compliance","issuer":"Wolfsberg Group","region":"global","url":"https://wolfsberg-group.org/resources/202/93","note":"Industry principles from 2022 for AI in financial crime compliance: legitimate purpose, proportionate use, design and technical expertise, accountability and oversight, and openness and transparency."},{"title":"Notice 626 Prevention of Money Laundering and Countering the Financing of Terrorism, Banks","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/notices/notice-626","note":"Example of national AML requirements for ongoing monitoring that an AI triage process must still meet."}],"controls":["Auto closure policy approved by the money laundering reporting officer, per scenario","Stored rationale, evidence and model version for every closed alert","Independent sampling with an error threshold that suspends auto closure","Model validation, inventory entry and ongoing performance monitoring","Documented parallel run results available for supervisory review"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the enrichment and drafting run as an **agentic workflow**, triggered through the\nAPI for each alert. The agent calls **custom functions** that read the alert, the customer's\ndue diligence record and the transactions behind the hit, queries **SQL knowledge bases** for\nprior alerts, and follows the bank's investigation procedures from a **knowledge base** with\nhybrid retrieval. It returns **structured output**: a risk summary, the evidence with record\nreferences and a proposed disposition.\n\nClosures go through **human in the loop approval** until the bank's policy allows otherwise, and\nevery run has a full audit trail. **PII masking** limits the personal data that reaches the\nmodel, **test suites** replay historical alerts with known outcomes before every change, and\n**monitors** run scheduled checks against the agent and alert on failures. The scoring model itself can stay in\nthe bank's own analytics stack; the platform is model agnostic, with EU and UAE data residency."},"faq":[{"question":"How much can AI reduce AML false positives?","answer":"Published results vary widely with the starting point. HSBC says it now has 60% fewer false positive cases and finds two to four times more financial crime after moving to a machine learning approach built with Google Cloud, and Shift4 reports an 86% reduction after adding ThetaRay's AI transaction monitoring. Measure against your own baseline per scenario, because rule sets differ so much between institutions."},{"question":"Can an AML alert be closed without a human?","answer":"None of the regulators cited on this page prohibits it, but the bank remains accountable for every closure. The Wolfsberg principles ask institutions to validate AI regularly and hold them responsible for decisions that rely on it, whoever built the system. A cautious path starts with drafting and ranking, and closes only the lowest risk band automatically after a parallel run."},{"question":"Does this replace transaction monitoring rules?","answer":"Not necessarily. UOB's model complements its rules, which remain its first line of defence, while HSBC made a machine learning risk score its primary transaction monitoring system in key markets. The HKMA sees both paths in Hong Kong: some institutions replace rules with a holistic approach, others add AI use cases step by step. Retire rules only once a parallel run shows the model finds at least as much."}],"related":["suspicious-activity-report-drafting","mule-network-detection","dynamic-customer-risk-rating","sanctions-screening-adjudication","fraud-alert-triage","trade-finance-crime-screening"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added SEO title and meta description, a Google Cloud false positive statistic, the HSBC alert volume metric and the alert volume KPI; made the HKMA statistic, Wolfsberg note, EU AI Act basis (Annex III point 5(b), Article 5(1)(d)), monitor description and FAQ answers match their sources; removed an unsupported superlative."},{"date":"2026-09-27","note":"Review fixes: corrected the monitor description to run against the agent, not the workflow; replaced the undated UNODC estimate with a dated 2011 UNODC estimate consistent with the same range; dropped the Australia Post false positive claim from the FAQ, because Napier's case study attributes it to rule configuration and calibration rather than to AI."}],"slug":"aml-alert-triage","url":"https://www.blits.ai/ai-use-cases/aml-alert-triage","benchmarks":[{"kpi":"false-positive-reduction","label":"False positive reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":73,"min":60,"max":86,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"shift4-thetaray-aml-transaction-monitoring","pooled":true},{"id":"hsbc-dynamic-risk-assessment","pooled":true}]},{"kpi":"alert-volume-reduction","label":"Alert volume reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":60,"min":60,"max":60,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"hsbc-dynamic-risk-assessment","pooled":true}]},{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":57,"min":57,"max":57,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"nexo-unit21-ai-alert-narratives","pooled":true}]}],"indicativeValueResult":{"low":400000,"high":2800000},"evidence":["australia-post-napier-aml-monitoring","fis-financial-crimes-ai-agent","hsbc-dynamic-risk-assessment","nexo-unit21-ai-alert-narratives","ratepay-hawk-aml-screening","shift4-thetaray-aml-transaction-monitoring","uob-tookitaki-aml-alert-prioritisation","uphold-unit21-ai-agent-pilot"]},{"title":"AI for application and identity fraud detection","shortTitle":"Application and identity fraud","seoTitle":"Application and identity fraud detection with AI","metaDescription":"AI checks loan and account applications for forged documents and synthetic identities before approval. Close Brothers Motor Finance, BCU and CNG Holdings use it.","definition":"AI that checks incoming account and loan applications for forged or AI generated documents, synthetic and stolen identities, and coordinated application rings, by analysing documents, device and application data across the whole queue and cross checking against bureau and official sources.","aliases":["application fraud detection","synthetic identity detection","document fraud detection","fake pay slip and bank statement detection","first party fraud detection at onboarding"],"industries":["banking","payments","cross-industry","government","telecommunications"],"functions":["fraud-prevention","onboarding-and-kyc","lending-and-credit"],"patterns":["document-processing","anomaly-detection","computer-vision","prediction-and-scoring"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"Lenders and banks decide on applications from documents and data the applicant provides: pay\nslips, bank statements, tax forms, identity documents and a selfie. Editing tools and now\ngenerative AI make convincing forgeries cheap, and synthetic identities built from real and\ninvented data can pass individual checks and build a credit history before they default.\n\nManual document review is slow and inconsistent, and reviewers cannot see metadata manipulation\nor the pattern across many applications, such as the same device, the same template or similar\nemail addresses. Blunt rules, on the other hand, turn away genuine applicants and slow down the\ndigital onboarding that customers expect. The same problem exists outside banking wherever\ndocuments prove eligibility: insurance, telecom contracts, rentals and public benefits.","problemStats":[{"statement":"FinCEN reported an increase in suspicious activity reports describing suspected deepfake media, particularly fraudulent identity documents used to get past identity verification, and issued an alert with red flag indicators.","sourceTitle":"FinCEN Issues Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions","sourceUrl":"https://www.fincen.gov/news/news-releases/fincen-issues-alert-fraud-schemes-involving-deepfake-media-targeting-financial","year":2024},{"statement":"In a Feedzai survey of 562 fraud and financial crime professionals at financial institutions, 92% of the institutions said fraudsters use generative AI.","sourceTitle":"AI Fraud Trends 2025: Banks Fight Back","sourceUrl":"https://www.feedzai.com/pressrelease/ai-fraud-trends-2025/","year":2025}],"howItWorks":"1. **Read the documents.** Document AI extracts the data from pay slips, statements and identity\n   documents and checks each file for tampering: metadata, fonts, layout against known templates,\n   arithmetic that does not add up, and signs of generation.\n2. **Check the identity.** Identity data is verified against bureau and official sources, the\n   identity document is checked for authenticity, and a liveness check confirms the applicant is\n   present and matches the document.\n3. **Look across the queue.** Anomaly and graph models link applications that share devices, IP\n   addresses, contact details, employers or document templates, exposing rings and synthetic\n   identity farms.\n4. **Score and explain.** Signals combine into a fraud risk score with the reasons behind it,\n   separate from the credit decision.\n5. **Route, do not reject.** Clean applications flow straight through; flagged ones go to a\n   fraud analyst with the evidence, who decides whether to request more information, verify\n   directly with the employer or bank, or decline.","valueDrivers":["risk-reduction","speed","customer-experience","cost-to-serve"],"kpis":["fraud-loss-reduction","detection-rate-improvement","automation-rate","processing-time-reduction","false-positive-reduction"],"indicativeValue":{"referenceOrg":"A consumer lender processing 100,000 applications a year","inputs":[{"key":"applications","label":"Applications per year","low":100000,"high":100000,"unit":"applications per year","note":"The reference lender."},{"key":"fraudRate","label":"Share of applications that are fraudulent and would be approved today","low":0.002,"high":0.005,"unit":"fraction of applications","note":"Editorial assumption. Replace with your own confirmed application fraud rate."},{"key":"lossPerFraud","label":"Average loss per approved fraudulent application","low":3000,"high":8000,"unit":"USD per case","note":"Editorial assumption. Replace with your own charge off data."},{"key":"extraCaught","label":"Additional share of that fraud caught before approval","low":0.2,"high":0.4,"unit":"fraction of fraudulent applications","note":"Editorial assumption, far below the vendor reported result on this page (CNG Holdings saw a greater than 80% drop in third party fraud within 90 days). Replace with results from a back test on your own confirmed fraud."}],"formula":"applications * fraudRate * lossPerFraud * extraCaught","currency":"USD","period":"per year","resultLabel":"Application fraud losses avoided","caveat":"Leaves out manual review time saved, faster approval for genuine applicants, lost revenue from wrongly declined applicants, and the cost of data sources and the platform."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Document and identity checks are available as services and integrate through APIs. The work is in combining signals across the application queue, feeding back confirmed fraud, and keeping fraud flags separate from credit decisions.","dataPrerequisites":["Confirmed application fraud and synthetic identity cases linked to the original applications","Application, device and session data for every application","Document images or files in their original format, not only extracted fields","Access to bureau, identity verification and, where available, official data sources"],"integrations":["Loan origination or account opening system","Identity verification and liveness service","Credit bureau and fraud data sharing schemes","Device intelligence","Fraud case management"]},"implementation":{"steps":[{"title":"Collect the evidence you already have","detail":"Link past confirmed fraud and early defaults with no payments to their applications, and keep original document files. This is your test set."},{"title":"Add document forensics to the existing flow","detail":"Run document checks on every uploaded file in shadow mode, measure what they catch against the test set, and tune thresholds before they affect any applicant."},{"title":"Look across applications","detail":"Link applications by device, contact details, employer and document template to find rings that single application checks miss."},{"title":"Route flagged applications to people","detail":"Send flags to fraud analysts with the evidence, and give them fast ways to verify, such as open banking data or direct employer confirmation."},{"title":"Keep fraud and credit separate","detail":"Document that a fraud flag leads to investigation, not an automatic credit decline, and report false positive rates to the risk committee."}],"guardrails":["A fraud flag triggers investigation or verification, never an automatic decline on its own","Reasons for every flag stored and available to the analyst","False positive rates monitored across customer groups to avoid unfair outcomes","Liveness and biometric checks used only for one to one verification with consent","Confirmed outcomes fed back to keep models current as generation tools improve"],"humanInTheLoop":"Fraud analysts decide every flagged application. Credit decisions stay in the credit process. The fraud strategy owner approves thresholds, and the risk committee receives false positive and performance reports.","kpisToInstrument":["Application fraud losses and first payment defaults, normalised for volume","Share of confirmed fraud flagged before approval","Share of genuine applicants flagged, and time to clear them","Straight through approval rate for clean applications","Rings detected and applications linked per ring"],"failureModes":[{"title":"The arms race","detail":"Generation tools improve faster than template checks. Combine document forensics with data verification at the source and network signals."},{"title":"Fraud flag as a hidden decline","detail":"Flags quietly turn into declines, creating fair lending and adverse action exposure. Separate the processes and audit outcomes."},{"title":"Punishing thin files","detail":"Young people and newcomers look like synthetic identities. Test false positive rates on these groups and provide alternative verification."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 5(b) excludes AI used to detect financial fraud from the high risk credit scoring category, but a system that in effect decides on creditworthiness is high risk, and remote biometric identification is high risk under point 1(a), which excludes one to one biometric verification. When a public authority uses the model on claims for public benefits, point 5(a) can apply, because it covers AI used to grant, reduce, revoke or reclaim benefits and has no fraud exception. Keep fraud detection separate from the credit or eligibility decision and use biometrics only for one to one verification."},"regulations":["eu-ai-act","gdpr","eba-loan-origination","fatf-recommendations","us-sr-11-7","uk-consumer-duty","nist-ai-rmf","eu-amlr","us-bsa","us-fcra"],"guidance":[{"title":"Annex III: High-Risk AI Systems Referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) excludes fraud detection from high risk credit scoring; point 5(a) covers public benefits decisions; point 1(a) covers remote biometric identification, excluding one to one verification."},{"title":"Guidelines on the use of remote customer onboarding solutions","issuer":"European Banking Authority","region":"europe","url":"https://www.eba.europa.eu/regulation-and-policy/anti-money-laundering-and-countering-financing-terrorism/guidelines-use-remote-customer-onboarding-solutions","note":"Sets expectations for the reliability of remote onboarding solutions, including checks that identity documents are genuine and not tampered with, and strong and reliable algorithms for biometric matching."},{"title":"FinCEN Issues Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions","issuer":"Financial Crimes Enforcement Network (FinCEN)","region":"north-america","url":"https://www.fincen.gov/news/news-releases/fincen-issues-alert-fraud-schemes-involving-deepfake-media-targeting-financial","note":"Describes typologies and red flag indicators for deepfake media, particularly fraudulent identity documents used to get past identity verification, and reminds institutions of their reporting duties under the Bank Secrecy Act."}],"controls":["Documented separation between fraud flags and credit decisions","Stored reasons for every flag and analyst decision","Fairness monitoring of flag rates across customer groups","Consent and data protection impact assessment for biometric checks","Model inventory entry, validation and drift monitoring"],"incidents":[{"title":"Revealed: bias found in AI system used to detect UK benefits fraud","url":"https://www.theguardian.com/society/2024/dec/06/revealed-bias-found-in-ai-system-used-to-detect-uk-benefits","note":"The Guardian reported in December 2024 that an internal fairness assessment of the UK Department for Work and Pensions' machine learning model for Universal Credit advance claims found it selected people for fraud investigation at different rates by age, disability, marital status and nationality. Staff make the final decision, but the case shows why flag rates need fairness monitoring."}]},"blitsAi":{"howToBuild":"Document forensics and identity verification usually come from specialist services, called\nthrough **custom functions**. Blits.ai adds the orchestration and the human side: an **agentic\nworkflow**, triggered through the API for each application, collects the verification results,\ncompares the documents' extracted data with the application and bureau data, checks the\n**SQL knowledge base** of previous applications for shared details, and returns **structured\noutput** with a risk summary and reasons.\n\nFlagged applications go to an analyst through **human in the loop approval**. When the\napplicant needs to provide more, a **conversational agent** on **web chat**, **WhatsApp** or the\nlender's own app through the **API channel** asks for the specific document and accepts **file uploads**, with **PII masking**,\n**guardrails** and a **human handover** for anything sensitive. Every run keeps an audit trail,\nand the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"Can AI detect AI generated pay slips and bank statements?","answer":"Often, by combining signals: file metadata, template and font analysis, arithmetic consistency and, most reliably, checking the data against the source through open banking or the employer. No single check is enough as generation tools improve."},{"question":"Should a fraud flag decline the application?","answer":"No. A flag should lead to verification or investigation. Automatic declines based on fraud scores create fair lending and adverse action exposure and turn away genuine applicants."},{"question":"Is application fraud detection high risk under the EU AI Act?","answer":"Fraud detection is excluded from the high risk credit scoring category, but the design matters. If the fraud score in effect decides credit, if you use remote biometric identification rather than one to one verification, or if a public authority uses it to decide on benefits, it can become high risk."}],"related":["digital-onboarding-assistant","real-time-fraud-scoring","mule-network-detection","conversational-loan-application-intake","intelligent-document-processing","alternative-data-credit-scoring"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: industries now include government and telecommunications, where its evidence comes from; added EU Anti Money Laundering Regulation, Bank Secrecy Act, Fair Credit Reporting Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; sharpened the FinCEN and Feedzai statistics and the guidance notes to match their sources; EU AI Act basis now covers Annex III point 5(a) for public benefits; Blits.ai build text now names the API channel instead of a mobile app channel; corrected the CNG Holdings release date and the BCU and Close Brothers summaries."}],"slug":"application-and-identity-fraud-detection","url":"https://www.blits.ai/ai-use-cases/application-and-identity-fraud-detection","benchmarks":[{"kpi":"detection-rate-improvement","label":"Detection improvement","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2.5,"min":2.5,"max":2.5,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dwp-universal-credit-advances-fraud-model","pooled":true}]}],"indicativeValueResult":{"low":120000,"high":1600000},"evidence":["bcu-inscribe-document-fraud-detection","close-brothers-resistant-ai-document-fraud","cng-holdings-sas-identity-fraud","dwp-universal-credit-advances-fraud-model","payoneer-resistant-ai-document-forensics","telstra-quantium-fraud-indicator"]},{"title":"AI for back office account servicing execution","shortTitle":"Account servicing execution","seoTitle":"AI for back office servicing requests","metaDescription":"AI agents read servicing requests, check policy and entitlements, and prepare or make changes under dual control, with a value model for the effort saved.","definition":"AI that executes the servicing requests that land in operations queues, such as address and mandate changes, standing instructions, beneficiary updates, reissues, payoff and reference letters and loan maintenance, by reading the request, checking it against policy and entitlements, and preparing or making the change in core systems under dual control.","aliases":["servicing operations automation","back office servicing agent","service request fulfilment AI","operations queue automation"],"industries":["banking","insurance","wealth-and-asset-management"],"functions":["operations","lending-and-credit"],"patterns":["agentic-workflow","document-processing","classification-and-routing"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"back-office","problem":"Behind every servicing channel sits an operations team that does the actual change. A customer\nasks in the app, a branch fills in a form, a letter arrives, or a relationship manager emails a\nrequest: in each case a person in operations keys the new address, updates the direct debit\nmandate, sets up the standing order, orders the replacement cheque book, calculates the payoff\nfigure or changes the repayment date on a loan. The work is repetitive, spread across many\nscreens and subject to dual control, so queues build up when volumes peak.\n\nThis page is about that execution layer. The customer facing conversation (the chatbot or voice\nagent that takes the request and resolves simple ones instantly) is covered on the related page\n\"AI agent for account and card servicing\". Execution is what happens when the request needs documents, checks, calculations or changes in\nsystems that the front end cannot touch directly, whichever channel it came from.","problemStats":[],"howItWorks":"1. **Receive the request from any source.** A handover from the servicing agent, a branch or\n   portal form, a scanned letter or an email becomes one structured service request.\n2. **Read and complete it.** Document AI extracts the fields from attached forms and evidence\n   (proof of address, signed mandate, power of attorney) and flags what is missing.\n3. **Check policy and entitlement.** The agent retrieves the servicing procedure for the request\n   type and checks the requester's authority, signatures, cut off times and limits.\n4. **Prepare or execute.** For low risk changes inside set limits it makes the change through the\n   core system APIs; for changes with financial consequence it prepares the change and the\n   evidence for a second person to approve.\n5. **Confirm and close.** It generates the confirmation or letter (such as a payoff statement),\n   updates the case and tells the originating channel, so the front end can inform the customer.","valueDrivers":["cost-to-serve","speed","employee-productivity","risk-reduction"],"kpis":["automation-rate","handling-time-reduction","processing-time-reduction","error-reduction","hours-saved"],"indicativeValue":{"referenceOrg":"A retail bank whose operations teams complete 500,000 servicing requests a year","inputs":[{"key":"requests","label":"Servicing requests completed by operations per year","low":500000,"high":500000,"unit":"requests per year","note":"The reference bank. Replace with the volume from your operations workflow tool."},{"key":"minutesPerRequest","label":"Operator minutes per request today","low":8,"high":15,"unit":"minutes per request","note":"Editorial assumption including checks and dual control. Replace with your own time study."},{"key":"effortReduction","label":"Share of operator time the AI removes","low":0.3,"high":0.6,"unit":"fraction of time per request","note":"Editorial assumption; checkers still approve changes with financial consequence."},{"key":"costPerHour","label":"Fully loaded operations cost per hour","low":30,"high":55,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"requests * minutesPerRequest / 60 * effortReduction * costPerHour","currency":"USD","period":"per year","resultLabel":"Operations effort avoided","caveat":"Labour only. It leaves out faster turnaround for customers, fewer errors and rework, lower complaint volumes and the cost of the platform and core system integration."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Each request type touches different core systems, many of them without clean APIs, and every change with financial consequence needs entitlement checks and dual control that auditors accept. Start with a handful of request types and widen over time.","dataPrerequisites":["Procedures per request type, with required documents, checks and limits","Request volumes and handling times per type from the operations workflow tool","Signature and mandate records reachable for entitlement checks"],"integrations":["Core banking, card and loan servicing systems","Operations workflow or case management tool","Document capture for forms and evidence","Customer communications for confirmations and letters","The front end servicing agent and contact centre for handover and status"]},"implementation":{"steps":[{"title":"Pick request types by volume and reversibility","detail":"Start with high volume changes that are easy to reverse and have no direct financial effect, such as contact detail updates and statement preferences. Leave payment instructions and loan changes for a later wave."},{"title":"Write the procedure as rules the agent can check","detail":"For each request type list the required evidence, the entitlement check, the limits and the systems touched. Automation stalls when procedures are unclear."},{"title":"Run as a preparer first","detail":"Let the AI prepare every change with its evidence while operators still execute. Measure how often the preparation is right before letting it execute anything."},{"title":"Separate execute from approve","detail":"Allow the agent to execute within limits only for low risk types, and keep a second person on anything that changes where money goes or how much is owed."},{"title":"Connect the front end","detail":"Feed status back to the customer facing channels so customers and agents stop chasing operations for updates."}],"guardrails":["Entitlement and signature checks before any change, with the evidence stored on the case","Dual control on changes to payees, standing instructions, mandates and loan terms","Execution only through an allow list of core system actions with limits per request type","Changes to contact details trigger a notification to the old and the new contact, as a fraud control"],"humanInTheLoop":"Checkers approve every change with financial consequence and every exception the agent flags. Operations leads decide which request types the agent may execute on its own, and a quality team samples completed requests weekly.","kpisToInstrument":["Share of requests completed without manual keying, per request type","Median turnaround from receipt to completion","Errors found in quality sampling and by customers","Share of prepared changes that checkers reject","Operator minutes per request"],"failureModes":[{"title":"Account takeover through a servicing change","detail":"A fraudster changes contact details or a payee through a convincing request. Keep strong identity checks and notify both old and new contact details."},{"title":"Automation without clean procedures","detail":"The agent follows an outdated or ambiguous procedure consistently and at scale. Assign owners and review dates to every procedure it uses."},{"title":"Screen based integrations that break","detail":"Changes made through fragile screen automation fail silently after a system update. Prefer APIs and monitor completion against the core record."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The tier depends on how the system is built. It stays minimal when the agent only executes changes approved by a person and any letter comes from a fixed template, since executing servicing changes is not listed in Annex III. It moves to limited risk when the same system talks to customers directly (the Article 50 transparency duty, described on the customer facing servicing page) or when generative AI drafts the confirmation or letter text: the provider of that generative function, the bank if it builds the system, then carries the Article 50(2) duty to mark the generated content in a machine readable way, unless the output only gets an assistive role or standard editing that does not substantially alter the input data. An AI system used to evaluate the creditworthiness of natural persons, for example to decide on a loan restructuring, is high risk under Annex III point 5(b); keep that assessment outside this agent, which only executes the decided change."},"regulations":["eu-ai-act","gdpr","dora","uk-consumer-duty","apra-cps-230","pci-dss"],"guidance":[{"title":"Revisions to the principles for the sound management of operational risk","issuer":"Basel Committee on Banking Supervision","region":"global","url":"https://www.bis.org/bcbs/publ/d515.htm","note":"The 2021 revision updates the guidance on change management and on information and communication technology; its control and mitigation principles apply equally to changes that an automated agent makes."}],"controls":["Documented boundary of what the agent may execute on its own, approved by the accountable executive","Full log of every change, the evidence used, and the maker and checker","Change control and regression tests for every new request type","Reconciliation of executed changes against the core system record"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that is started through the API with an API token,\nor from a flow (for example one behind the email channel or the customer facing servicing\nagent) through a **trigger workflow block**. The agent reads\nthe request it is given, retrieves the procedure from the **knowledge base**, and calls\n**custom functions** for the entitlement check and the change in core systems, limited by a\n**tool execution policy**.\n\n**Human in the loop confirmation** holds any change above the configured threshold for a\nchecker, and every run keeps a **full audit trail**. A flow can start the workflow through a\ntrigger workflow block and wait for its result, or confirm completion to the customer by\n**email**. **Test suites** run sample requests against the workflow before a new request type\ngoes live, and **PII masking** at the gateway can mask or redact personal data the model does\nnot need."},"faq":[{"question":"How is this different from an account and card servicing chatbot?","answer":"The chatbot talks to the customer and resolves what it can instantly through front end APIs. This use case is the operations layer behind every channel: it executes the requests that need documents, checks, calculations or changes in core systems, including those that arrive by letter, branch form or handover from the chatbot."},{"question":"Which servicing changes can an AI execute without a second person?","answer":"Only low risk, reversible changes inside set limits, such as statement preferences or contact details with a notification to both old and new details. Changes to payees, standing instructions, mandates and loan terms keep dual control."},{"question":"What results have been published?","answer":"SS&C Blue Prism reports a 58% cut in processing time for judicial orders on accounts at Banco Supervielle (average response time from 12 to 5 minutes), and SS&C says its generative AI document agents process loan credit agreements 95% faster than by hand. Both are vendor case studies of processes next to servicing (court orders and loan document intake) rather than customer servicing changes themselves."}],"related":["account-and-card-servicing-agent","correspondence-triage-and-routing","outbound-notice-drafting","intelligent-document-processing","corporate-account-onboarding-orchestration"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added publication dates and the manual baseline to the evidence, corrected the evidence summaries to what the case studies say, narrowed the Basel guidance note to what the document states, added the Annex III point 5(b) boundary for loan restructuring, aligned the Blits.ai section with the feature inventory, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: named SS&C Blue Prism as the claimant of the Banco Supervielle figure, removed the unscoped figure from the meta description, labelled both case studies as adjacent processes, corrected the Annex III point 5(b) wording, added the Article 50(2) marking duty for generated letters, reworded unsourced generalisations as advice, and limited the Blits.ai section to listed features."}],"slug":"account-servicing-execution","url":"https://www.blits.ai/ai-use-cases/account-servicing-execution","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":76.5,"min":58,"max":95,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"ssc-technologies-generative-ai-document-agents","pooled":true},{"id":"banco-supervielle-judicial-notification-automation","pooled":true}]}],"indicativeValueResult":{"low":600000,"high":4125000},"evidence":["banco-supervielle-judicial-notification-automation","ssc-technologies-generative-ai-document-agents"]},{"title":"AI for benefit fraud and error detection in social security","shortTitle":"Benefit fraud and error detection","seoTitle":"AI for benefit fraud and error detection","metaDescription":"Risk models help benefits agencies pick whom to check. DWP says its model is 2.5 times more effective than random checks; Rotterdam stopped its model in 2022.","definition":"Risk models that help a social security or benefits agency decide which claims, payments and recipients to check for fraud or error, so that caseworkers verify the riskiest cases first, while every decision on entitlement stays with a person and the model is tested for fairness before and during use.","aliases":["welfare fraud detection","benefit fraud risk scoring","social security fraud and error analytics","improper payments detection"],"industries":["government"],"functions":["fraud-prevention","citizen-services","case-management"],"patterns":["prediction-and-scoring","anomaly-detection"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"assist","adoptionStage":"early-adopters","problem":"Benefits agencies pay very large sums to millions of people, and some of it goes to the wrong\nplace: organised fraud, individual misrepresentation, and honest mistakes by claimants or by the\nagency itself. Checking everyone is impossible and would delay payments to people who need them\nurgently, so agencies have to choose whom to check.\n\nThat choice is where automated risk scoring has done serious harm. The Dutch childcare benefits\nscandal, the SyRI welfare fraud system that a Dutch court stopped in 2020, Rotterdam's welfare risk\nmodel and Australia's Robodebt scheme each show one or more of the same failures: selection on\nnationality or on proxies for it, systems too opaque to check or challenge, and a burden of proof\nshifted onto people who were then treated as if they owed money or had committed fraud. Under the EU AI Act, systems that evaluate eligibility for public\nassistance, or grant, reduce, revoke or reclaim it, are high risk. The job is not only to catch\nfraud but to do so lawfully, proportionately and transparently.","problemStats":[{"statement":"The UK Department for Work and Pensions paid 1.3 million Universal Credit advances worth GBP 700 million in 2025 to 2026 and estimates the fraud and error in them at between GBP 20 million and GBP 90 million.","sourceTitle":"Effectiveness Assessment of Universal Credit Advances Model","sourceUrl":"https://www.gov.uk/government/publications/effectiveness-assessment-including-statistical-analysis-of-the-universal-credit-advances-machine-learning-model-1-april-2025-to-31-march-2026/effectiveness-assessment-of-universal-credit-advances-model","year":2026}],"howItWorks":"1. **Define the risk precisely.** A model targets one defined risk at one point in the process,\n   for example an advance request before payment or a specific change in circumstances, not\n   \"fraud\" in general.\n2. **Score at the point of decision.** The claim data is scored in real time or in batch. The\n   output is a referral for a check, never a decision on entitlement.\n3. **Blind and mix the referrals.** High risk referrals go to caseworkers together with a random\n   control group, and the caseworker is not told which is which, so the check is not biased by the\n   score.\n4. **Verify with the person.** A caseworker reviews the evidence and, where needed, asks the\n   claimant, then decides. Declines follow the normal notice, review and appeal route.\n5. **Measure effectiveness and fairness.** Confirmed fraud and error rates in model referrals are\n   compared with the random group, overall and by group, and the model is retrained or stopped when\n   it targets groups without a matching rate of confirmed findings.","valueDrivers":["risk-reduction","compliance","cost-to-serve"],"kpis":["detection-rate-improvement","false-positive-reduction","fraud-loss-reduction","cost-savings"],"indicativeValue":{"referenceOrg":"A national benefits agency paying EUR 2 billion a year in a benefit with known fraud and error risk","inputs":[{"key":"paymentsValue","label":"Value of payments in scope","low":1000000000,"high":3000000000,"unit":"EUR per year","note":"Editorial assumption. Replace with the payments the model will actually screen."},{"key":"lossRate","label":"Share of payments lost to fraud and error","low":0.01,"high":0.03,"unit":"fraction of payments","note":"Editorial assumption. Use your own official fraud and error statistics, which vary widely by benefit."},{"key":"preventedShare","label":"Share of those losses prevented by better targeted checks","low":0.05,"high":0.15,"unit":"fraction of losses","note":"Editorial assumption, deliberately modest. DWP reports its advances model is 2.5 times more effective than random selection, but a better hit rate prevents only part of the loss."}],"formula":"paymentsValue * lossRate * preventedShare","currency":"EUR","period":"per year","resultLabel":"Fraud and error losses prevented","caveat":"Gross losses prevented only. It leaves out the cost of the checks, the cost of wrongly delayed or refused payments to legitimate claimants, legal and reputational risk, and the cost of the governance a high risk system requires."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Technically this is a scoring model; in practice it is one of the most regulated and contested uses of AI in government. It needs a clear legal basis, a data protection impact assessment, an equality or fundamental rights assessment, published documentation and an evaluation design with a random control group.","dataPrerequisites":["Confirmed outcomes of past checks, including checks that found nothing","Claim data available at the decision point, with a legal basis for each item","Protected characteristic data or reliable proxies for fairness testing"],"integrations":["Benefit claim and payment systems","Caseworker case management","Notice, review and appeal processes","Data warehouse for monitoring and evaluation"]},"implementation":{"steps":[{"title":"Choose a narrow, well evidenced risk","detail":"Start with one payment type where fraud is documented, as DWP did with Universal Credit advances, rather than a general score of every recipient."},{"title":"Do the rights assessment first","detail":"Complete the data protection impact assessment and an equality or fundamental rights impact assessment before build, and publish a transparency record."},{"title":"Build the evaluation into the design","detail":"Keep a random control group of referrals and hide the source of each referral from caseworkers, so effectiveness and fairness can be measured honestly."},{"title":"Test disparities in referral and in outcome","detail":"For every group you can measure, compare how often the model refers people with how often those referrals are confirmed. DWP's assessment found groups referred more often without more confirmed fraud, and DWP retrained the model, which was still in testing when it reported."},{"title":"Define the stop rule","detail":"Agree in advance what finding would pause the model. Rotterdam stopped its welfare model when it concluded it could not currently build a model that fits its policy."}],"guardrails":["The model only refers cases for a check; it never decides, reduces or stops a payment","No protected characteristic, nationality or obvious proxy as a feature","Random control group and blinded referrals for every model in use","Published transparency record and effectiveness assessment at least yearly","Payment timeliness for legitimate claimants monitored as a harm measure"],"humanInTheLoop":"Caseworkers review every referral and make every entitlement decision, with the claimant able to explain their circumstances. Claimants keep the normal review and appeal rights. A senior owner signs off the yearly effectiveness and fairness assessment and can suspend the model.","kpisToInstrument":["Confirmed fraud and error rate in model referrals versus the random control group","Referral and outcome disparities by group","Payment delay for legitimate claimants who were referred","Share of declines overturned on review or appeal","Losses prevented per check"],"failureModes":[{"title":"Discrimination through proxies","detail":"Nationality is used directly, or language, neighbourhood or family status stand in for protected characteristics. The Dutch childcare benefits model used nationality as a risk factor, and journalists found Rotterdam's model scored on age, gender and language skills."},{"title":"Automation reversing the burden of proof","detail":"When a score or data match is treated as proof, people must disprove a debt. Robodebt raised debts from averaged income data without other evidence and put the onus on recipients to contradict them. A person must establish the facts."},{"title":"Secrecy that blocks accountability","detail":"Agencies that refuse to disclose how their models work cannot be challenged or corrected, as investigations in Sweden and Denmark have shown. Publish the documentation."},{"title":"Chasing small sums at high human cost","detail":"Aggressive thresholds generate many checks on honest claimants for little recovered money. Measure the harm side alongside the savings."}]},"risk":{"euAiAct":{"tier":"high","basis":"Annex III point 5(a): AI systems used by or on behalf of public authorities to evaluate the eligibility of natural persons for essential public assistance benefits and services, or to grant, reduce, revoke or reclaim them. A fundamental rights impact assessment (Article 27) is required before a public body deploys it. A design that scores people over time on their social behaviour or personal characteristics and leads to unrelated or disproportionate detrimental treatment would fall under the Article 5(1)(c) prohibition on social scoring."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Algorithmic Transparency Recording Standard hub","issuer":"UK government","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"UK central government bodies publish records of their algorithmic tools here, including DWP's record for its Universal Credit advances model."},{"title":"Algoritmeregister van de Nederlandse overheid","issuer":"Government of the Netherlands","region":"europe","url":"https://algoritmes.overheid.nl/nl","note":"Dutch government bodies register algorithms used in benefit control here, for example UWV's unemployment benefit risk scan and Rotterdam's welfare risk model, which is listed as no longer in use."}],"controls":["Fundamental rights and data protection impact assessments before deployment","Yearly published effectiveness and fairness assessment with a random control group","Blinded referrals and human decision on every case","Documented stop rule and senior owner with authority to suspend","Notice, review and appeal routes that do not depend on knowing the model exists"],"incidents":[{"title":"Amnesty International: Xenophobic machines, the Dutch childcare benefits scandal","url":"https://www.amnesty.org/en/documents/eur35/4686/2021/en/","note":"The Dutch tax authorities used an algorithmic system to create risk profiles of childcare benefit applicants, with nationality as one of its risk factors. Amnesty found this resulted in discrimination and racial profiling."},{"title":"District Court of The Hague: SyRI judgment (ECLI:NL:RBDHA:2020:865)","url":"https://uitspraken.rechtspraak.nl/inziendocument?id=ECLI:NL:RBDHA:2020:865","note":"The court ruled in February 2020 that the legislation for SyRI, a Dutch government system that linked data to flag welfare fraud risk, breached the right to private life because it was insufficiently transparent and verifiable."},{"title":"Royal Commission into the Robodebt Scheme: report volume 1 (archived)","url":"https://web.archive.org/web/20241231202325/https://robodebt.royalcommission.gov.au/system/files/2023-07/robodebt_report_volume_1.pdf","note":"Australia's Robodebt scheme raised welfare debts largely through income averaging, without other evidence, and placed the onus on recipients to contradict the result. The Royal Commission found the method neither produced accurate results nor complied with the income calculation provisions of the Social Security Act 1991. In 2020 the government decided to refund debts raised wholly or partly through averaging that had been repaid, and to reduce unpaid ones to zero, and in November 2020 it settled a class action. The report was presented on 7 July 2023. It was automation rather than machine learning, but the lesson on burden of proof applies."},{"title":"Lighthouse Reports: Suspicion Machines (Rotterdam)","url":"https://www.lighthousereports.com/investigation/suspicion-machines/","note":"Reconstruction of Rotterdam's welfare fraud model, which took 315 inputs including age, gender and language skills; the investigation found it discriminated by ethnicity, age, gender and parenthood."},{"title":"Lighthouse Reports: Sweden's Suspicion Machine","url":"https://www.lighthousereports.com/investigation/swedens-suspicion-machine/","note":"Analysis of the Swedish Social Insurance Agency's fraud prediction algorithm for temporary child support found it disproportionately flagged women, migrants, low income earners and people without a university education."},{"title":"Amnesty International: Denmark's AI powered welfare system fuels mass surveillance","url":"https://www.amnesty.org/en/latest/news/2024/11/denmark-ai-powered-welfare-system-fuels-mass-surveillance-and-risks-discriminating-against-marginalized-groups-report/","note":"Amnesty's 2024 report on the fraud control algorithms of Denmark's welfare authority, Udbetaling Danmark, warns of mass surveillance and discrimination against marginalised groups, including through inputs on \"foreign affiliation\". The authority refused full access to the code and data."}]},"blitsAi":{"howToBuild":"Blits.ai is not a fraud scoring engine and should not be used to decide entitlement. Where it\nhelps is around the model: an **agentic workflow** can take a referral, gather the claim facts\nthrough **custom functions**, prepare a neutral case summary for the caseworker and draft the\ninformation request to the claimant, with **human in the loop approval** before anything is\nsent. A **knowledge base** with hybrid retrieval over the benefit rules and the evidence policy\nhelps caseworkers apply the rules consistently.\n\n**PII masking** at the gateway, a full audit trail per run and **test suites** that check that\nsummaries stay neutral and factual support the documentation a high risk use needs. The platform\nis **model agnostic** and offers EU and UAE data residency for claimant data."},"faq":[{"question":"Is benefit fraud detection high risk under the EU AI Act?","answer":"Yes, when the system evaluates eligibility for public assistance or is used to grant, reduce, revoke or reclaim benefits (Annex III point 5(a)). Public bodies must also carry out a fundamental rights impact assessment before use."},{"question":"Does it work?","answer":"It can improve targeting. DWP reports its Universal Credit advances model is 2.5 times more effective than random selection in 2025 to 2026, with a median payment delay of one day for approved referrals. The same assessment found disparities by nationality, age and couple status, which is why the random control group and yearly assessment matter."},{"question":"What went wrong in the Dutch, Australian and Rotterdam cases?","answer":"Selection used nationality or proxies such as language skills, people had to disprove a score or data match, and systems were hard to scrutinise. Rotterdam stopped its model in 2022, a Dutch court struck down the SyRI legislation in 2020, and debts Robodebt raised through income averaging were refunded or reduced to zero before a Royal Commission reported in 2023."}],"related":["tax-compliance-risk-scoring","benefits-eligibility-and-application-assistant","inspection-prioritization","application-and-identity-fraud-detection","ai-model-inventory"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version for the government vertical, with UK, Dutch and US evidence, a stopped deployment and the main known incidents verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Robodebt, Dutch childcare benefits and Denmark incident notes and the matching FAQ and failure mode text to what the sources state, softened the problem framing, added the Article 5 social scoring boundary, fixed the DWP disparity summary (couples were not referred more often), added the UWV random sample control and Accenture as Rotterdam integrator, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Corrected the Robodebt incident to the report order (refunds for repaid debts and zeroing of unpaid ones decided in 2020, class action settled in November 2020) and pointed it at report volume 1; aligned the FAQ and the Rotterdam stop rule wording with the sources."}],"slug":"benefit-fraud-and-error-detection","url":"https://www.blits.ai/ai-use-cases/benefit-fraud-and-error-detection","benchmarks":[{"kpi":"detection-rate-improvement","label":"Detection improvement","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2.5,"min":2.5,"max":2.5,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dwp-universal-credit-advances-fraud-model","pooled":true}]}],"indicativeValueResult":{"low":500000,"high":13500000},"evidence":["cms-prescription-drug-cost-anomaly-detection","dwp-universal-credit-advances-fraud-model","gemeente-rotterdam-welfare-reassessment-risk-model","us-treasury-payment-integrity-machine-learning","uwv-unemployment-benefit-risk-scan"]},{"title":"AI for building HVAC and energy optimization","shortTitle":"Building energy optimization","seoTitle":"AI HVAC optimization for building energy","metaDescription":"BrainBox AI reports a 15.8% cut in HVAC related electricity at Cammeby's International and 10% HVAC energy savings at Loyola's Schreiber Center.","definition":"AI that continuously predicts a building's heating, cooling and ventilation needs and adjusts setpoints, equipment sequencing and start times in real time through the existing building management system, instead of following fixed schedules, so the building uses less energy without a person retuning it by hand.","aliases":["autonomous HVAC optimization","AI building energy management","smart building energy control","automated emissions reduction"],"industries":["real-estate","cross-industry"],"functions":["operations"],"patterns":["prediction-and-scoring"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"autonomous","adoptionStage":"early-adopters","segment":"hvac-optimization","problem":"Space heating is the single largest energy end use in US commercial buildings, at about 32% of\ntotal energy use in 2018, with ventilation adding roughly another 10% (US EIA). A building that\nruns this equipment on fixed schedules and setpoints does not adapt as occupancy, weather and\nelectricity prices change hour to hour. According to BrainBox AI, rising energy costs and\nregulations led Cammeby's International, a real estate investment company, to look at AI for its\n32 storey office property in New York City's financial district.\n\nFacility teams can tune a building management system by hand, but re tuning every zone\ncontinuously as conditions change takes staff time. Without that continuous attention, a building\ncan end up running wider, more conservative margins than a continuously optimized building would\nneed, and can miss the chance to shift load to the hours when the local electricity grid is\ncleanest.","problemStats":[{"statement":"Space heating was about 32% of total US commercial building energy use in 2018, the largest single end use; ventilation and lighting were next at about 10% each (US EIA).","sourceTitle":"Use of energy in commercial buildings","sourceUrl":"https://www.eia.gov/energyexplained/use-of-energy/commercial-buildings.php","year":2018}],"howItWorks":"1. **Connect to the existing building management system.** An edge device or cloud connection\n   reads live data points, such as temperatures, valve positions and equipment status, typically\n   over the BACnet protocol, without replacing hardware.\n2. **Learn the building's thermal behaviour.** The AI models how each zone responds to outside\n   weather, occupancy and equipment changes, specific to that building rather than a generic\n   template.\n3. **Predict and act ahead of time.** The system forecasts the next hours of demand and adjusts\n   valve positions, fan speeds and equipment sequencing continuously, cooling or heating the\n   building before the need arrives.\n4. **Shift load where a grid signal is available.** With a marginal emissions or price signal, the\n   AI pre cools the building during the hours when the grid is cleanest, then lets the\n   temperature drift during the hours when the grid relies more on fossil fuels, so the\n   building's thermal mass carries the load instead of the equipment.\n5. **Report through the existing tools.** Facility engineers watch the AI's decisions through\n   customised graphics inside their own building management system, rather than a separate\n   interface, and can adjust or override a setpoint if needed.","valueDrivers":["compliance"],"kpis":["energy-savings","cost-savings"],"indicativeValue":{"referenceOrg":"An office building with 300,000 square feet of AI controlled HVAC","inputs":[{"key":"squareFeet","label":"Square feet under AI control","low":150000,"high":500000,"unit":"square feet","note":"Editorial assumption. Cammeby's International's controlled space was 251,104 square feet (BrainBox AI), within this range."},{"key":"hvacCostPerSqFt","label":"HVAC energy cost per square foot per year","low":0.75,"high":1.5,"unit":"USD per square foot per year","note":"Editorial assumption. BrainBox AI's Cammeby's International figures ($42,951 saved from a 15.8% reduction in HVAC related electricity, across 251,104 square feet over an 11 month period) imply a total HVAC electricity cost of about $271,800 over that period ($42,951 / 0.158), or roughly $1.08 per square foot over the 11 months and about $1.18 per square foot annualised, assuming the dollar saving is proportional to the 15.8% consumption reduction; replace with your own utility spend."},{"key":"savingsShare","label":"Share of HVAC energy cost saved","low":0.05,"high":0.1,"unit":"fraction of HVAC energy cost","note":"At or below the two reported results on this page (a 15.8% reduction in HVAC related electricity consumption at Cammeby's International and a 10% savings in HVAC related energy at Loyola University's Schreiber Center, both BrainBox AI), to allow for a different building's baseline and scope of HVAC energy cost."}],"formula":"squareFeet * hvacCostPerSqFt * savingsShare","currency":"USD","period":"per year","resultLabel":"Annual HVAC energy cost avoided","caveat":"HVAC energy cost avoided only. It leaves out the software subscription, any emissions credit or compliance benefit, and the risk that an aggressive deployment without a hard comfort band produces occupant complaints."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The AI needs a working connection to the building's existing controls, commonly BACnet, and a controls contractor's cooperation to expose the right points. In a typical rollout, most of the effort is integration and change management with the building's engineering team, rather than the AI modelling itself.","dataPrerequisites":["Live read and write access to the building management system's HVAC data points","A record of the HVAC equipment and how zones map to it","At least a few weeks of historical operating data to calibrate the thermal model"],"integrations":["Building management system, typically over BACnet or an equivalent protocol","Weather forecast data","Optionally, a grid emissions or price signal for load shifting"]},"implementation":{"steps":[{"title":"Start with the highest value system","detail":"At Cammeby's International, the building's own team set out to target savings particularly in its chilled water loop, according to BrainBox AI's case study. Naming the equipment that carries the most load up front gives an early, defensible result to point to, even though the rollout itself went floor by floor across all the connected equipment, air handling units, variable air valves, fans and both the hot and chilled water systems, rather than one system at a time."},{"title":"Work with the controls contractor, not around them","detail":"Program the AI's decisions into graphics the facility team already uses inside its own building management system. At Cammeby's International, BrainBox AI rolled out floor by floor in close cooperation with the building's Chief Engineer, so daily operations were not disrupted."},{"title":"Measure a real baseline first","detail":"Establish a measured baseline period before claiming a saving, so the figure holds up against normal seasonal swings. BrainBox AI reported Cammeby's International's 15.8% reduction over an 11 month results period in 2023, without disclosing how the baseline itself was built."},{"title":"Add a grid or price signal once the basics work","detail":"Once core optimization is stable, a marginal emissions or price signal, as Loyola University's Schreiber Center used with WattTime, lets the building shift load to cleaner or cheaper hours by pre cooling ahead of time and drifting during the dirtier hours."},{"title":"Keep occupant comfort as a hard constraint","detail":"Set temperature and ventilation bands the AI may not cross regardless of the savings opportunity, and track comfort complaints alongside the energy numbers."}],"guardrails":["Hard temperature and ventilation bands the AI cannot exceed, enforced in the building management system itself, not only in the AI's own logic","No changes to life safety, fire, or code required ventilation and pressure sequences","A facility engineer can see and override every AI decision through the existing building management system graphics"],"humanInTheLoop":"Facility engineers watch the AI's live decisions through their own building management system graphics and can override any setpoint, and a reliability or sustainability team reviews measured savings against the baseline on a schedule.","kpisToInstrument":["HVAC related energy consumption and cost, measured against a comparable prior period rather than a single year over year comparison","Occupant comfort complaints during and after rollout","Emissions avoided when a grid signal drives load shifting"],"failureModes":[{"title":"Savings measured against the wrong baseline","detail":"Weather and occupancy swing year to year; a saving claimed against a single prior year, rather than a normalised baseline, will not survive scrutiny."},{"title":"Comfort complaints erase trust in the program","detail":"Aggressive pre cooling or setpoint drift without a hard comfort band produces complaints that get the whole program switched off; keep comfort bands non negotiable from day one."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Optimizing HVAC equipment is not listed in Annex III. Point 2 covers AI safety components in the management and operation of critical infrastructure such as the supply of water, gas, heating or electricity, and a building's own HVAC controller does not manage infrastructure at that level. The system also does not decide credit, employment, access to essential services or another high risk use. Under Article 6(1), an AI system that is a safety component of a product covered by Annex I harmonisation legislation, such as the Machinery Regulation, and that needs third party conformity assessment, would be high risk regardless of Annex III; this is why the guardrails on this page keep the AI out of fire, life safety and code required ventilation sequences rather than letting it override them."},"regulations":["eu-ai-act"],"guidance":[],"controls":["A documented safe operating envelope, covering temperature, humidity and ventilation rate, that the AI cannot leave, enforced independently of the AI's own decisions","A logged history of every setpoint or sequence change the AI made, so a disputed comfort complaint or an unusual energy month can be traced to a specific decision"],"incidents":[]},"blitsAi":{"howToBuild":"The HVAC optimization model itself, the part that predicts a building's thermal response and\nwrites setpoints back to the building management system, runs on the building automation\nvendor's own control platform; Blits.ai is not a building management system. What fits on\nBlits.ai is the layer facility teams use to understand and govern it: an SQL knowledge base over\nthe building's point data and energy logs lets an agent answer a facility manager's question\nabout why a zone ran outside its normal range or what changed overnight, in plain language\ninstead of a raw trend graph.\n\nAgentic tasks can watch for a condition, such as a zone repeatedly hitting its comfort band\nlimit, or query an SQL knowledge base on a schedule to compare the optimization model's own\npredictions against actual consumption, and draft a ticket for the facility engineer with the\nrelevant data attached, with human in the loop confirmation before anything escalates to the\ncontrols contractor. Monitors run scheduled checks that the facility agent still answers\ncorrectly, and the platform's EU and UAE data residency options suit a portfolio that spans\nregions."},"faq":[{"question":"How much energy does AI HVAC optimization actually save?","answer":"BrainBox AI reports an 11 month measured 15.8% reduction in HVAC related electricity consumption at Cammeby's International's office tower in Manhattan, saving $42,951 and avoiding 37.14 tonnes of CO2 equivalent. At Loyola University's Schreiber Center in Chicago, a year long project found a 10% savings in HVAC related energy and 10% lower HVAC related CO2e emissions. Results vary by building age, climate and how the baseline is measured."},{"question":"Does it need new hardware?","answer":"Both deployments on this page connected to the building's existing management system over BACnet, but Cammeby's International needed a new edge device: BrainBox AI's case study says the solution was deployed through its own edge device, communicating over BACnet IP. At Loyola University's Schreiber Center, BrainBox AI says it integrated its system with the building's existing HVAC controls via BACnet, with no edge device mentioned."},{"question":"Can it help a building use more renewable energy?","answer":"Indirectly. Loyola University's Schreiber Center paired HVAC optimization with WattTime's marginal emissions signal to shift some cooling to the hours when the local grid relies more on renewables. BrainBox AI reports a 15% average reduction in HVAC emissions during marginal emissions events, including the following four hour drift period."},{"question":"Is this high risk under the EU AI Act?","answer":"Usually not. HVAC optimization is not listed in Annex III's high risk categories, including point 2 on critical infrastructure for water, gas, heating or electricity supply, since a building's own HVAC controller does not manage infrastructure at that level, and it is not a safety component of a product under Annex I harmonisation legislation such as the Machinery Regulation. If it did perform that kind of safety function, Article 6(1) could make it high risk regardless of Annex III, which is why a well designed deployment keeps it out of fire, life safety and code required ventilation sequences."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched for the real estate and energy scope, with BrainBox AI's Cammeby's International and Loyola University Schreiber Center case studies checked against the primary sources."},{"date":"2026-09-28","note":"Editorial review: corrected the hardware, KPI and baseline claims against the primary sources, added a cited problem stat (US EIA), narrowed the indicative value ranges to the evidence, and corrected the Blits.ai monitors and analytics description against FEATURE_INVENTORY.md."},{"date":"2026-09-28","note":"Adversarial review fixes: corrected the EIA problem stat (space heating is the largest end use, not \"after lighting\"; ventilation ties with lighting), fixed the EU AI Act basis and FAQ to cite Article 6(1) and Annex I product safety rather than only critical infrastructure, corrected the pre cooling and drift mechanism (drift happens in the dirtier hours, not the cleanest), recomputed the HVAC cost per square foot note ($1.08 over 11 months, $1.18 annualised) and widened \"Below\" to \"At or below\" in the savingsShare note, softened the unsourced \"most buildings\" claim, corrected the attribution of the chilled water loop savings target to Cammeby's own team, and fixed the \"integrated directly\" and \"without hurting comfort\" overstatements against the sources."},{"date":"2026-09-28","note":"Round 2 review fixes: corrected the pre cooling and drift mechanism in the Loyola evidence record to match the fix already made on this page, and reworded the fixed schedule and conservative margin claims in the problem section as conditional framing since no source supports them as a factual claim about building stock."}],"slug":"building-energy-optimization","url":"https://www.blits.ai/ai-use-cases/building-energy-optimization","benchmarks":[{"kpi":"energy-savings","label":"Energy savings","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":12.9,"min":10,"max":15.8,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"cammebys-international-brainbox-ai-hvac","pooled":true},{"id":"loyola-university-schreiber-center-brainbox-ai","pooled":true}]},{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":42951,"min":42951,"max":42951,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"cammebys-international-brainbox-ai-hvac","pooled":true}]}],"indicativeValueResult":{"low":5625,"high":75000},"evidence":["cammebys-international-brainbox-ai-hvac","loyola-university-schreiber-center-brainbox-ai"]},{"title":"AI for business onboarding (KYB) and beneficial ownership discovery","shortTitle":"Business onboarding and UBO","seoTitle":"AI for KYB onboarding and UBO discovery","metaDescription":"AI agents build the KYB file for corporate clients and trace ownership to the ultimate beneficial owners. Google Cloud reports 30 times productivity at M-DAQ Global.","definition":"An AI agent that builds the know your business (KYB) due diligence file for a new or reviewed corporate client, before any account is opened: it collects registry, incorporation and ownership documents, resolves the entity across sources, maps the ownership chain through holding companies, nominees and trusts to the ultimate beneficial owners, screens the entity and its owners, and presents a risk scored case for a compliance analyst to decide.","aliases":["KYB automation","UBO discovery","corporate KYC agent","beneficial ownership mapping"],"industries":["banking","payments","capital-markets"],"functions":["onboarding-and-kyc","financial-crime-compliance"],"patterns":["document-processing","agentic-workflow","classification-and-routing","summarization"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"specialized-businesses","problem":"Onboarding a company is slower and harder than onboarding a person. The bank has to establish who\nthe company is, who controls it and who ultimately owns it, which means pulling filings from\nregistries in several countries, reading articles of association and trust deeds, following\nownership through layers of holding companies, and screening every entity and person found. Much\nof the data is in PDFs, in other languages, or missing from registries with no digital access.\n\nAnalysts can spend more of their time building the file than judging it, clients receive\nrepeated document requests, and cases wait while documents are chased. Opaque structures are exactly where the\nfinancial crime risk sits, so shortcuts are not an option. AI can build the file and the\nownership graph quickly and consistently, as long as every link in the chain stays traceable to a\nsource and a person makes the decision.","problemStats":[],"howItWorks":"1. **Fetch.** From the company name or registration number, the agent pulls registry records,\n   filings and ownership data from the available registers and data providers.\n2. **Read the documents.** Document AI extracts officers, shareholders, percentages and control\n   rights from articles, share registers, trust deeds and client submissions.\n3. **Resolve and map.** Entity resolution links the same company or person across sources, and a\n   graph of ownership and control is built up to the ultimate beneficial owners, with the\n   calculated effective ownership per person.\n4. **Screen.** Every entity and person is screened for sanctions, politically exposed persons and\n   adverse media, and gaps (no registry, missing documents) are listed.\n5. **Score and summarise.** The case is risk scored against the bank's policy and summarised in\n   plain language, with every finding linked to its source.\n6. **Decide.** A compliance analyst reviews, overrides where needed and decides; low risk cases\n   that meet the policy still get a person's lighter review before the decision is confirmed.","valueDrivers":["speed","compliance","cost-to-serve","risk-reduction"],"kpis":["processing-time-reduction","productivity-gain","automation-rate","cycle-time-days","accuracy"],"indicativeValue":{"referenceOrg":"A bank onboarding 3,000 business clients a year","inputs":[{"key":"cases","label":"Business onboarding cases per year","low":3000,"high":3000,"unit":"cases per year","note":"The reference bank."},{"key":"hoursPerCase","label":"Analyst hours to build and review a KYB file","low":4,"high":8,"unit":"hours per case","note":"Editorial assumption, replace with your own time study. Complex structures take far longer."},{"key":"reduction","label":"Share of analyst time saved","low":0.3,"high":0.5,"unit":"fraction of time","note":"Editorial assumption, replace with your own pilot results."},{"key":"hourlyCost","label":"Loaded cost of an analyst hour","low":50,"high":80,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."}],"formula":"cases * hoursPerCase * reduction * hourlyCost","currency":"USD","period":"per year","resultLabel":"Analyst time released, valued at loaded cost","caveat":"Values analyst time only. It leaves out data subscription costs, revenue from clients onboarded sooner, fewer clients lost to slow onboarding, and the reduced risk of missing a hidden owner."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Registry coverage and quality vary widely by country, ownership calculations through circular or layered structures are tricky, and the output feeds a regulated decision that examiners review.","dataPrerequisites":["Access to company registries and commercial ownership data for the bank's markets","The bank's KYB policy, risk model and beneficial ownership thresholds per jurisdiction","Screening lists for sanctions, politically exposed persons and adverse media"],"integrations":["Client lifecycle management or onboarding case system","Registry and data provider APIs","Screening engine","Client portal for document requests","CRM"]},"implementation":{"steps":[{"title":"Codify the policy first","detail":"Write down, per jurisdiction and client type, which documents are required, the ownership threshold, and what makes a case low, medium or high risk, before any automation."},{"title":"Build the file, not the decision","detail":"Start with the agent assembling the case file and ownership graph for analysts, and measure time and quality, before letting any case move forward with lighter review."},{"title":"Show confidence and gaps","detail":"Make the agent state where registry data is missing or weak, and route those cases to manual research instead of filling the gap."},{"title":"Keep the chain examinable","detail":"Store for every ownership link the document or record it came from, so an examiner can follow the chain from the client to each beneficial owner."},{"title":"Tune on real cases","detail":"Compare agent built files with analyst built files on a sample of past cases, including complex structures, and fix systematic misses before scaling."}],"guardrails":["Every ownership link and screening hit is traceable to a source document or record","Gaps and low confidence are surfaced, never silently filled","A compliance officer decides onboarding, exits and enhanced due diligence","Low risk cases that meet written criteria still get lighter review, with a person confirming the onboarding decision and quality assurance sampling the outcomes","Personal data of owners and directors is used only for the compliance purpose and retained per policy"],"humanInTheLoop":"Analysts review every case above the low risk threshold and can override any finding. Low risk cases still get a lighter review, and a compliance officer decides onboarding and exits in every case. Quality assurance samples cases that passed with lighter review, and the model owner reviews screening and scoring performance.","kpisToInstrument":["Elapsed days from application to decision, by risk level","Analyst hours per case, by structure complexity","Share of cases with complete files on first review","Ownership errors and missed owners found in quality assurance","Document requests sent to clients per case"],"failureModes":[{"title":"False completeness","detail":"The graph looks complete because a registry returned nothing about a layer. Show coverage per jurisdiction and flag layers without data."},{"title":"Wrong entity match","detail":"Two companies or people with similar names are merged. Use identifiers where available and send low confidence matches to an analyst."},{"title":"Opaque scoring","detail":"Analysts cannot explain why a case scored high. Show the factors behind each score and keep the model documented and validated."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Customer due diligence on legal entities is not listed in Annex III, and an internal analyst tool usually carries no Article 50 transparency duty, so the system is usually minimal risk. The design decides the rest: biometric verification that only confirms a director is who they claim to be is excluded from Annex III point 1(a), but remote biometric identification (one to many matching) is high risk, and so is any use of the output to assess the creditworthiness of the natural persons involved (point 5(b)). GDPR applies to the personal data of owners and directors throughout. Keep biometric and credit steps in separately assessed components."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","mas-ai-risk-management","dora","iso-42001","eu-amlr","us-bsa","mas-notice-626"],"guidance":[{"title":"Regulation (EU) 2024/1624 on the prevention of the use of the financial system for money laundering or terrorist financing","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1624/oj","note":"The EU Anti Money Laundering Regulation applies from 10 July 2027. It sets the customer due diligence and beneficial ownership rules the case file must meet, including the ownership and control tests. Article 76(5) requires meaningful human intervention in every decision to enter, refuse or maintain a business relationship with a customer, and in every decision to raise or lower the customer due diligence measures applied, so no such decision can be left to straight through automation."},{"title":"CDD Final Rule","issuer":"FinCEN","region":"north-america","url":"https://www.fincen.gov/resources/statutes-and-regulations/cdd-final-rule","note":"The Bank Secrecy Act rule that requires US banks and other covered institutions to identify and verify the beneficial owners of legal entity customers when those companies open accounts. FinCEN ruling FIN-2026-R001 grants exceptive relief from repeating this at each new account opening."},{"title":"MAS Guidelines for Artificial Intelligence (AI) Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Consultation paper of 13 November 2025 proposing supervisory expectations on AI oversight, AI inventories, risk materiality, human oversight, testing and monitoring, including for AI agents. Relevant where AI output supports financial crime decisions."}],"controls":["Written KYB policy per jurisdiction that the agent is configured and tested against","Source record for every ownership link and screening result, retained with the case","Model inventory, validation and monitoring of entity resolution and risk scoring","Quality assurance sampling, with higher sampling for straight through cases"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow**: an agent loop that calls registry and data provider\nAPIs through **custom functions** or **MCP**, reads the client documents received through a\nreceive attachment step or uploaded to the document library, and uses a **knowledge base** with the bank's KYB\npolicy per jurisdiction. **Structured output**\nproduces the ownership table, the list of gaps and the case summary in a fixed format for the\nonboarding system.\n\nThe **tool execution policy** limits which systems the agent may call, and **human in the loop\napproval** is configured so the analyst approves before anything is written back. The per run **audit trail** records each source consulted, **PII masking**\nprotects personal data in prompts, **guardrails** check the output, and **test suites** replay past cases with known\noutcomes. Data can stay in the EU or UAE region, and the model is chosen per agent."},"faq":[{"question":"Can AI decide whether to onboard a company?","answer":"It should not. AI can build the file, map ownership and score risk, but a compliance officer decides onboarding and exits, and a person confirms every decision. Low risk cases that meet written criteria only get a lighter review, with sampling."},{"question":"How does AI find hidden beneficial owners?","answer":"By extracting owners from filings and documents, linking the same entities and people across sources, and calculating effective ownership through every layer. It still depends on registry coverage, so gaps must be flagged rather than guessed."},{"question":"Are there published results?","answer":"Two vendor case studies name a real deployment. Google Cloud lists M-DAQ Global, a fintech group in foreign exchange and cross border payments, whose KYB compliance system on Vertex AI improves productivity by 30 times. Kyndryl reports that a proof of concept it ran with Incore Bank, a Swiss bank whose customers are other banks, financial intermediaries and corporates, and Google Cloud reached up to 99 percent accuracy extracting data from onboarding documents, with a further, unmeasured claim that onboarding time could fall from months to days. Both are vendor claims, one from a fintech and one from a bank, and neither has independent oversight behind it. Treat such figures as a starting point and test them in a pilot."}],"related":["corporate-account-onboarding-orchestration","perpetual-kyc","pep-and-adverse-media-screening","sanctions-screening-adjudication","dynamic-customer-risk-rating","digital-onboarding-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the editor after a targeted blocker check."},{"date":"2026-09-25","note":"First version, kept as draft because only one public deployment with a named organization was found; the catalog cited vendor pages only."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: rewrote the EU AI Act basis (Annex III points 1(a) and 5(b), no Article 50 duty), corrected the M-DAQ Global FAQ to the vendor claim as stated, removed the unsupported 'weeks' wait time, added the FinCEN CDD Final Rule and detail to the AMLR and MAS guidance notes, aligned the Blits.ai build text with the feature inventory, and added seoTitle and metaDescription. Kept as draft: one public deployment."},{"date":"2026-09-27","note":"Editor review: lowered the adoption stage to emerging (one public deployment, a fintech), rewrote the published results FAQ to state that no named bank result was found, softened the Article 50 statement, added FinCEN ruling FIN-2026-R001 to the CDD Final Rule note, and aligned the document intake and guardrail wording in the Blits.ai build text with the feature inventory."},{"date":"2026-09-27","note":"Review fix: rewrote the AMLR note to state the Article 76(5) meaningful human intervention requirement, replaced straight through treatment with lighter human review that a person confirms, and aligned howItWorks step 6, humanInTheLoop and FAQ 1 with that wording."},{"date":"2026-09-27","note":"Added a second public deployment: Incore Bank's proof of concept with Kyndryl and Google Cloud on business customer KYC, which reached up to 99 percent accuracy extracting data from onboarding documents. Rewrote the published results FAQ to cover both vendor claims. Moved status to review; the page now has two grade C public evidence records."}],"slug":"business-onboarding-and-ubo-discovery","url":"https://www.blits.ai/ai-use-cases/business-onboarding-and-ubo-discovery","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":25,"min":25,"max":25,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bny-eliza-onboarding-research","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":20,"min":20,"max":20,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bny-eliza-onboarding-research","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":30,"min":30,"max":30,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"m-daq-global-kyb-onboarding","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"incore-bank-agentic-kyc-onboarding","pooled":false}]}],"indicativeValueResult":{"low":180000,"high":960000},"evidence":["bny-eliza-onboarding-research","incore-bank-agentic-kyc-onboarding","m-daq-global-kyb-onboarding"]},{"title":"AI for cash application and remittance matching","shortTitle":"Cash application and remittance matching","seoTitle":"AI cash application for accounts receivable","metaDescription":"AI matches payments to invoices for accounts receivable teams. HighRadius reports 98% auto applied at Keurig Dr Pepper and 96% at ResMed as automation results.","definition":"AI that reads remittance advices in many formats, matches incoming customer payments to open receivable invoices, proposes deduction and short pay reason codes from prior resolutions, and routes only the genuine exceptions to a cash application analyst, so the accounts receivable sub ledger clears itself for the clean majority of payments.","aliases":["cash application automation","AI accounts receivable matching","remittance processing","automated payment matching"],"industries":["cross-industry","manufacturing","retail-and-ecommerce","healthcare","logistics-and-transportation"],"functions":["finance-and-accounting"],"patterns":["document-processing","agentic-workflow","anomaly-detection","classification-and-routing"],"channels":["email","internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"back-office","problem":"Every company that sells on credit has to turn an incoming payment into a cleared invoice, and\nthe payment rarely arrives with clean instructions. A wire lands with a reference number that\ndoes not match any invoice, a check comes with a remittance advice stapled to a delivery note, an\nERP portal payment bundles twelve invoices into one line, and a short paid invoice gives no reason\nat all. Rule based matching engines clear the exact, one to one payments; everything else becomes\na growing pile of unapplied cash that a credit or accounts receivable analyst has to open,\ninterpret and apply by hand, invoice by invoice.\n\nThe cost shows up twice. Analyst time goes into repetitive lookup and data entry instead of\ngenuine exceptions, and unapplied or misapplied cash distorts the accounts receivable ageing\nreport, triggers unnecessary collections calls to customers who already paid, and pushes up days\nsales outstanding, which is money the company has effectively already earned but cannot yet use.\nMachine learning changes what a matching engine can clear on its own: a learned matching layer can\npropose one to many and many to many matches, tolerate partial references and short pays within a\ntolerance, and suggest a deduction reason code from how similar cases were resolved before,\nleaving people to judge the cases that are genuinely new.","problemStats":[],"howItWorks":"1. **Capture every remittance.** Emails, customer portal downloads, EDI 820 files and scanned\n   lockbox images arrive in one pipeline; document AI reads the unstructured ones and extracts\n   payer, amount, currency and any invoice references.\n2. **Match beyond the rules.** Deterministic rules clear exact one to one matches first. A learned\n   matching layer then proposes one to many and many to many matches using fuzzy references,\n   amounts within tolerance and payment history, each with a confidence score.\n3. **Propose a reason for what does not match.** For short pays and deductions the agent suggests\n   a reason code and the likely open item, drawn from how the team resolved similar cases before,\n   instead of leaving a blank exception for someone to start from scratch.\n4. **Auto apply within limits.** Matches above the confidence and value threshold post\n   automatically to the sub ledger; everything else goes to an analyst queue with the payment,\n   remittance and candidate invoices shown together.\n5. **Feed collections and credit.** Genuine deductions and disputes that need a decision are\n   handed to the deductions or collections team with the evidence attached, so they do not sit\n   unresolved inside the cash application queue.\n6. **Learn under control.** Analyst decisions feed back as candidate matching rules or reason code\n   suggestions, which an accounts receivable manager approves before they change what posts\n   automatically.","valueDrivers":["cost-to-serve","speed","employee-productivity","risk-reduction"],"kpis":["automation-rate","hours-saved","cost-savings","handling-time-reduction","cycle-time-days"],"indicativeValue":{"referenceOrg":"A manufacturer that processes 200,000 customer payments a year","inputs":[{"key":"paymentsPerYear","label":"Customer payments processed per year","low":200000,"high":200000,"unit":"payments per year","note":"The reference company. Replace with your own payment volume."},{"key":"manualMinutes","label":"Minutes an analyst spends manually applying a payment","low":6,"high":12,"unit":"minutes per payment","note":"Editorial assumption, replace with your own time study."},{"key":"autoApplyShare","label":"Share of payments applied without a person once the AI is added","low":0.75,"high":0.95,"unit":"fraction of payments","note":"Total share including exact matches a rules engine already clears, conservative against the evidence on this page (HighRadius reported 98% of payments auto applied at Keurig Dr Pepper and a 96% cash posting hit rate at ResMed, both totals that include rules based matching, not the AI increment alone). Treat both vendor reported figures as an upper bound, not a typical first year result."},{"key":"ruleBaselineShare","label":"Share of payments existing rules already clear before adding AI","low":0.3,"high":0.5,"unit":"fraction of payments","note":"Editorial assumption, replace with your own baseline auto match rate. Deterministic one to one rules typically clear a meaningful share of exact matches on their own; the AI should be credited only for what it adds on top of this baseline."},{"key":"costPerHour","label":"Fully loaded cost of an accounts receivable or credit analyst","low":30,"high":50,"unit":"USD per hour","note":"Editorial assumption for a blended onshore and offshore accounts receivable team."}],"formula":"paymentsPerYear * (autoApplyShare - ruleBaselineShare) * manualMinutes / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Manual cash application effort avoided by the AI, over what rules already clear","caveat":"Labour only, and only the increment over what existing rule based matching already clears before any AI is added. It leaves out the working capital value of a lower days sales outstanding, deductions recovered, fewer unnecessary collections calls to customers who already paid, and the cost of the platform and the integration work."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The matching logic is well understood; the work is in the data. Remittance formats and channels vary by customer, references are inconsistent, and short pays need documented tolerances that the credit team agrees before any threshold is automated.","dataPrerequisites":["Twelve months of history of payments matched to invoices, with the resolution chosen","Clean customer and invoice master data, including known third party payers","Documented tolerances and write off thresholds for short pays and deductions"],"integrations":["ERP or accounts receivable sub ledger (open invoices, customer master)","Bank statement and lockbox feeds (BAI2, MT940, camt.053)","Email and customer portal for remittance advices","Deductions and claims management system","Collections platform for unresolved items"]},"implementation":{"steps":[{"title":"Start with your highest volume payment channel","detail":"Rank payment channels (ACH, wire, card, check lockbox) by manual cash application volume and pick the one with the most legible remittance data first, so the model has something to learn from quickly."},{"title":"Standardize remittance capture","detail":"Route every channel, including email attachments and portal downloads, into one pipeline and let document AI extract payer, amount and any invoice references before matching runs."},{"title":"Baseline what the current rules already clear","detail":"Measure the existing auto match rate so the AI is credited only for the increment, and fix obvious master data problems (duplicate customers, stale bank details) before tuning a model."},{"title":"Run in shadow mode","detail":"Let the AI propose matches and reason codes next to the analysts for several cycles, compare its proposals with what they actually did, and only then raise the auto apply threshold."},{"title":"Automate posting within limits","detail":"Auto apply only matches above an agreed confidence and value threshold; everything else, and anything that would change a customer's bank details, goes to a person."},{"title":"Close the loop into deductions and collections","detail":"Route unresolved short pays and disputes to the team that owns them with the evidence attached, so cash application does not become the place where deductions go to wait."}],"guardrails":["Auto apply only above an agreed confidence threshold and below an agreed value threshold; anything else goes to a person","Customer bank account and master data changes verified out of band, never inferred from a remittance","Write offs and reason code changes above an approval limit enforced by the ERP, not the model","Every automated posting keeps a link back to the source remittance and the matching logic used"],"humanInTheLoop":"Cash application analysts confirm or correct proposed matches below the confidence threshold and any deduction reason code. An accounts receivable or credit manager approves write offs above a set amount and reviews a sample of auto applied postings every month for drift.","kpisToInstrument":["Auto apply rate by payment channel","Unapplied cash balance and its age","Minutes per manually applied payment","Deduction and short pay resolution time","Days sales outstanding trend after go live"],"failureModes":[{"title":"Confident but wrong postings","detail":"A matching model trained on a messy history applies a payment to the wrong invoice with high confidence. Set a hard confidence floor below which nothing posts automatically, and sample auto applied postings weekly."},{"title":"Auto apply that hides real deductions","detail":"Short pays get force matched to keep the ageing report clean instead of being flagged as pricing or delivery disputes. Track resolution reason and root cause, not only the match rate."},{"title":"Customer master drift","detail":"New subsidiaries, factoring arrangements or third party payers (where the payer is not the invoiced customer) block matching that used to work. Keep the customer master current, including known third party payers."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Matching a company's own incoming payments to its own open invoices is a back office finance operation. It is not listed in Annex III and does not decide a natural person's creditworthiness or eligibility for a service, so it is minimal risk and the AI literacy duty of Article 4 applies. Using deduction or payment behaviour to score an individual sole trader's creditworthiness would need a fresh risk assessment."},"regulations":["eu-ai-act","gdpr"],"guidance":[],"controls":["Auto apply confidence and value thresholds set and periodically reviewed by the controller","Full audit trail from remittance to posted journal entry, including the confidence score used","Out of band verification of any customer bank account or master data change","Monthly sample review of auto applied postings by the accounts receivable or credit manager"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai an inbound remittance email on the **email channel** starts a dialog flow whose\n**trigger workflow** block starts the **agentic workflow**; a bank or lockbox file can start\nthe same workflow through an **API token** instead. Inside the workflow an **AI agent** with\n**structured output** extracts payer, amount and invoice references from the remittance text,\nand **custom functions** look up open items in the ERP or accounts\nreceivable sub ledger (for example through the SAP, Oracle NetSuite or Microsoft Dynamics 365\nconnections in the integration catalog) to build the proposed match. Deduction reason codes,\ntolerances and write off policy sit in the **knowledge base** with hybrid retrieval, so the\nagent can explain why it proposed a match or a code.\n\nAn approval step, built with flow or custom function logic rather than a separate built in\nfeature, holds any match above an agreed value threshold, or below an agreed confidence\nthreshold, for a cash application analyst to approve or reject before a custom function posts\nthe approved match to the ledger. Every run keeps a\n**full audit trail**, **test suites** replay a labelled set of past remittances before any\nchange goes live, and **monitors**\nrun scheduled checks against the agent and alert on failure. The platform is **model agnostic**\nand can run in the **EU or UAE data residency** region a customer needs."},"faq":[{"question":"How accurate is AI cash application?","answer":"It depends heavily on how legible the remittance data is. HighRadius reported that Keurig Dr Pepper auto applied 98% of payments and that ResMed reached a 96% cash posting hit rate across its business units, but neither source describes machine learning or AI matching specifically: both are automation results that could include a large share of rule based matching. Treat both as an upper bound rather than a typical result: plan for a lower rate on a first deployment, especially on channels with heavy check or manual remittance volume."},{"question":"What is the difference between cash application and bank or ledger reconciliation?","answer":"Cash application matches a company's incoming customer payments to its own open receivable invoices, so the accounts receivable sub ledger clears. Bank and ledger reconciliation matches a bank's or fund's own statements, settlement files and general ledger to each other; see AI for ledger and payment reconciliation for that job."},{"question":"What should stay with a person?","answer":"Short pays and deductions that need a judgment call on the reason, any write off above the approval limit, and anything that touches a customer's bank details or master data."},{"question":"Does AI cash application reduce days sales outstanding?","answer":"Faster, more accurate application clears invoices sooner and stops collections calls to customers who already paid, both of which help days sales outstanding. HighRadius reported that ResMed's days sales outstanding fell by about 33 days within 10 months; that case study covers ResMed's broader Customer-to-Cash Receivables Management suite (deductions and payment options included), so it is unclear how much of the reduction is attributable to cash application alone, and it is not one of the benchmarked KPIs on this page. Treat a days sales outstanding improvement as a likely secondary effect to measure on your own data, not a guaranteed one."}],"related":["ledger-and-payment-reconciliation","supplier-invoice-processing","treasury-cash-flow-forecasting"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version. Researched as a distinct job from AI for ledger and payment reconciliation (which covers a bank's own nostro, vostro and settlement matching): this page covers a company's own accounts receivable cash application. Evidence from HighRadius customer stories for Keurig Dr Pepper and ResMed, quotes checked against the live pages with usecases:source."},{"date":"2026-09-28","note":"Editorial review. Corrected the evidence year for both records to 2023 (the first Wayback capture of each source page), dropped the invented \"2026 activity\" justification, added archivedUrl to both sources, removed \"agentic\" from the Keurig Dr Pepper title and summary (the source only says \"cash application software\"), named Dr Pepper Snapple Group as the entity on the source page, changed the cost savings period to \"in the first year\" with a note that this is the vendor's framing, removed the cycle time in days metric from the ResMed record because it does not describe a days sales outstanding reduction and moved the days sales outstanding figure into the ResMed summary and verification note instead, reported the missing KPI as a taxonomy gap in the run report, broadened the ResMed summary to the Customer-to-Cash Receivables Management suite, removed the unsupported \"mature ERP integration and years of matching history\" qualifier from the FAQ and the indicative value note, and corrected the blitsAi build description to use the trigger workflow block from a dialog flow or an API token (not a workflow directly triggered by a channel) and to drop the unsupported \"lockbox scans\" (image or vision) extraction claim."},{"date":"2026-09-28","note":"Second editorial review, after an adversarial check. Corrected the evidence year for both records to 2021: the live pages' JSON-LD `datePublished` and `article:published_time` meta tags give 2021-02-12 (Keurig Dr Pepper) and 2021-02-23 (ResMed), checked directly in the page HTML, which disproves the earlier note that the pages give no publication date; the 2023 date was only the first Wayback capture of the renamed URL. Checked every quoted sentence against the live pages again (`dateModified` on both is 2025-05-28) and all still match word for word. Removed the Keurig Dr Pepper cost savings metric: its \"approximately\" and \"in the first year\" qualifiers were not supported by the quote or the page, and the $2.5M figure mixes the effect of bringing outsourced processing in house with the effect of the software, which the page's own \"annual run rate\" framing confirms; the figure is kept in the evidence summary only. Remapped the ResMed 50% time saving metric from the handling time reduction KPI to productivity gain, because the quote is about the share of an analyst's time saved on one sub task (data aggregation), not average handling time per case; added the productivity gain KPI to the page's kpis list. Softened the ResMed DSO framing in the evidence summary, the verification note and the FAQ from an assertion that the 33 day reduction excludes cash application to \"unclear how much is attributable to cash application alone\", and removed the unsupported \"within 10 months of go live\" wording (the source says only \"within 10 months\"). Rewrote the metaDescription to name HighRadius as the claimant for both the 98% and 96% figures instead of implying Keurig Dr Pepper and ResMed reported them themselves, and dropped the unsupported \"clearing most receivables automatically\" claim. Rewrote the blitsAi human in the loop paragraph to describe approve or reject controls above a configurable value threshold, matching FEATURE_INVENTORY.md and the `confirmationThresholdMinor` amount threshold in the agentic task service code, and to describe any confidence based routing as flow or custom function logic rather than a built in feature; also dropped \"and its attachments\" from the remittance extraction sentence, since the inventory lists attachment ingestion only for the knowledge base email address, not for parsing attachments in a triggered workflow. Added a ruleBaselineShare input to the indicative value formula so the AI is credited only for the increment over what existing rules already clear, in line with the page's own implementation advice; the worked range drops from about $450K to $1.9M to about $270K to $900K a year for the reference manufacturer."},{"date":"2026-09-28","note":"Third editorial review, after a further adversarial check. Removed \"AI\" from the ResMed evidence title and dropped a remittance capture sentence in its summary that belonged to the Keurig Dr Pepper record, not ResMed; checked again against the live page, which does not mention AI or machine learning. Reworded the metaDescription and the \"How accurate is AI cash application?\" FAQ answer to present the 98% and 96% figures as automation results, since neither source describes machine learning or an AI matching method specifically, and added the same caveat to the Keurig Dr Pepper evidence summary about its \"Saved in One Year with AI\" stat box caption. Removed the ResMed 50% time saving figure as a productivity gain metric and dropped productivity gain from the page's kpis list, since the quote covers one sub task (data aggregation) and no listed KPI fits it without overstating the claim; the figure stays in the evidence summary. Rewrote the blitsAi human in the loop paragraph: the platform's threshold based confirmation gate is built for agentic tasks tied to a payment amount, not a custom function posting inside a triggered agentic workflow, so the approval step is now described as flow or custom function logic rather than the confirmation gate feature."}],"slug":"cash-application-and-remittance-matching","url":"https://www.blits.ai/ai-use-cases/cash-application-and-remittance-matching","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":97,"min":96,"max":98,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"keurig-dr-pepper-cash-application-automation","pooled":true},{"id":"resmed-cash-application-automation","pooled":true}]}],"indicativeValueResult":{"low":270000,"high":899999.9999999998},"evidence":["keurig-dr-pepper-cash-application-automation","resmed-cash-application-automation"]},{"title":"AI for chargeback and representment operations","shortTitle":"Chargeback and representment","seoTitle":"AI chargeback and representment automation","metaDescription":"AI maps card disputes to reason codes, assembles evidence and drafts representment responses. Stripe says GitHub Sponsors saves 20 hours a month on average.","definition":"AI that runs the dispute engine room for issuers, acquirers and merchants: it maps each dispute to the network reason code, gathers the matching evidence, assembles a network compliant chargeback or representment package, drafts the rebuttal, tracks every deadline and processes pre dispute alerts so a refund can be issued before a chargeback lands.","aliases":["chargeback automation","representment automation","dispute operations AI","card dispute back office"],"industries":["payments","banking","retail-and-ecommerce"],"functions":["operations","fraud-prevention","customer-service"],"patterns":["agentic-workflow","document-processing","content-generation","classification-and-routing"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"back-office","problem":"Card disputes are a volume problem wrapped in a rulebook. Each network defines its own reason\ncodes, evidence requirements and deadlines, and revises them regularly. For every\ndispute an analyst at the issuer, the acquirer or the merchant has to find the transaction, pull\nauthorisation and 3DS records, delivery or usage evidence and prior correspondence, decide\nwhether to accept or fight, and write the case in the format the network expects, before the\nwindow closes.\n\nMuch of that work is low value but unforgiving. A missed deadline usually means the case is lost,\nweak packages lose winnable cases, and first party misuse (a customer disputing a purchase they\nmade) is hard to separate from genuine fraud. Dispute volumes keep rising, so teams that work by\nhand grow with them. The customer facing side, where a cardholder first reports \"I do not\nrecognise this charge\", is a separate job; this page is about what happens after the case is\nopened.","problemStats":[{"statement":"Visa reports that it processed 106 million disputes globally in 2025, a 35% increase since 2019.","sourceTitle":"Visa Unveils New Services to Modernize Dispute Resolution Process","sourceUrl":"https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22261.html","year":2026}],"howItWorks":"1. **Ingest and classify.** New disputes, retrieval requests and pre dispute alerts arrive from\n   the network systems. The agent maps each one to the reason code and the applicable rule set.\n2. **Gather the evidence.** It pulls authorisation and clearing data, 3DS and device records,\n   delivery and usage logs, refund history and correspondence, and checks them against the\n   evidence the reason code requires.\n3. **Recommend accept or fight.** It estimates the chance of winning from the evidence and past\n   outcomes, and recommends refunding, accepting or contesting, with reasons.\n4. **Assemble and draft.** For contested cases it builds the package in the network's format and\n   drafts the rebuttal narrative from the evidence.\n5. **Submit and track.** An analyst approves the package where required, the system submits it,\n   tracks each deadline through pre arbitration and arbitration, and records the outcome for\n   learning.","valueDrivers":["cost-to-serve","risk-reduction","speed","employee-productivity"],"kpis":["automation-rate","handling-time-reduction","hours-saved","cost-reduction","interactions-handled"],"indicativeValue":{"referenceOrg":"A card issuer or acquirer working 200,000 disputes a year","inputs":[{"key":"disputes","label":"Disputes worked per year","low":200000,"high":200000,"unit":"disputes per year","note":"The reference organization. Replace with your own dispute volume."},{"key":"minutesPerDispute","label":"Analyst minutes per dispute today","low":15,"high":30,"unit":"minutes per dispute","note":"Editorial assumption covering evidence gathering, decision and package. Replace with your own time study."},{"key":"effortReduction","label":"Share of analyst time the AI removes","low":0.2,"high":0.3,"unit":"fraction of time per dispute","note":"Editorial assumption, capped at the one handle time figure on this page (a vendor reports nearly 30% lower handle time per assignment); analysts still approve contested cases and write offs. Replace with your own."},{"key":"costPerHour","label":"Fully loaded dispute analyst cost per hour","low":30,"high":55,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"disputes * minutesPerDispute / 60 * effortReduction * costPerHour","currency":"USD","period":"per year","resultLabel":"Dispute handling effort avoided","caveat":"Labour only. It leaves out recovered losses from better packages and fewer missed deadlines, fees avoided through pre dispute refunds, and the cost of the platform and network integrations."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The networks already provide structured dispute systems and APIs. The work is gathering evidence from many internal systems and keeping the rule logic current with network releases.","dataPrerequisites":["Dispute history with reason codes, evidence submitted and outcomes","Current network rules and evidence requirements per reason code","Access to authorisation, clearing, 3DS, device and delivery data"],"integrations":["Card network dispute systems, such as Visa Resolve Online","Card management and transaction processing systems","Fraud and authentication platforms","Merchant order, delivery and refund systems (acquirer and merchant side)","Pre dispute alert services"]},"implementation":{"steps":[{"title":"Segment disputes by reason code and value","detail":"Find the reason codes with the most volume and the most avoidable losses. Low value disputes are often best refunded automatically; high value ones need the best packages."},{"title":"Build the evidence map","detail":"For each reason code list the evidence that wins, where it lives and how to fetch it. This map is the core asset, with or without AI."},{"title":"Automate assembly before decisions","detail":"Let the AI gather evidence and build packages while analysts decide, then measure package completeness and win rates."},{"title":"Add recommendations with thresholds","detail":"Allow automatic accept or refund below a value threshold and on alerts, and keep analyst approval for contested and high value cases."},{"title":"Keep the rules current","detail":"Assign an owner to network rule releases and run regression tests on the rule logic before each release date."}],"guardrails":["Rebuttals use only evidence from the case; the model never invents facts or documents","Rules and deadlines come from a maintained rule set, not from the model's memory","Write offs and contested cases above the threshold need analyst approval","Cardholder data masked in prompts and logs in line with PCI DSS"],"humanInTheLoop":"Analysts approve contested cases, high value decisions and write offs, and handle suspected fraud or hardship. A quality team samples automatic accepts and refunds each week, and the rule owner signs off changes when networks publish new rules.","kpisToInstrument":["Win rate on contested cases, by reason code","Deadlines missed","Analyst minutes per dispute","Share of disputes resolved by pre dispute alerts or automatic refund","Net losses from disputes as a share of sales or volume"],"failureModes":[{"title":"Outdated rules","detail":"A network rule change invalidates the evidence logic and cases start losing. Version the rule set and test it on every release."},{"title":"Fighting everything","detail":"Automation makes it cheap to contest, so weak cases are fought and fees rise. Use win probability and value thresholds."},{"title":"Invented evidence","detail":"A drafted rebuttal describes evidence that is not in the package. Link every statement to an attached document."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Dispute processing between issuers, acquirers and merchants is not listed in Annex III. It is not an evaluation of creditworthiness or credit scoring under Annex III point 5(b), and because cardholders do not interact with the system directly, the Article 50(1) transparency duty for AI that talks to people does not apply. Article 50(2) marking of generated text is a duty of the provider of the AI system that generates it, which includes an institution that builds its own dispute drafting agent and puts it into service under its own name. A drafted rebuttal built from attached case evidence performs an assistive function for standard editing of that evidence and does not substantially alter the underlying input, so it falls under the Article 50(2) exception and does not need machine readable marking. With that point checked, the tier stays minimal. A customer facing intake agent is assessed separately."},"regulations":["eu-ai-act","pci-dss","gdpr","dora","apra-cps-230"],"guidance":[{"title":"Visa Core Rules and Visa Product and Service Rules","issuer":"Visa","region":"global","url":"https://usa.visa.com/content/dam/VCOM/download/about-visa/visa-rules-public.pdf","note":"The public rulebook that defines dispute conditions, evidence and time limits for Visa transactions."}],"controls":["Versioned rule set per network with an owner and release testing","Deadline tracking with alerts well before each network cut off","Audit trail of evidence, decision, approver and outcome for every dispute","Documented value thresholds for automatic accept and refund"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered through the API when a new dispute arrives\nfrom the network systems. **Custom functions** fetch transaction, authentication and delivery\nevidence, the network rules and evidence maps sit in the **knowledge base** with hybrid\nretrieval, and an **AI agent** with **structured output** builds the package and drafts the\nrebuttal. **Agentic tasks** recheck deadlines on a schedule.\n\n**Human in the loop approval** holds contested and high value cases for an analyst. The custom\nfunctions return masked card data, so the agent sees only what each step needs. Each run keeps a\n**full audit trail**, and **test suites** replay past disputes with known outcomes before rule or\nprompt changes go live. A customer facing agent on the same platform can open the dispute and\nshow its status to the cardholder; in that chat, card numbers the cardholder types are detected\nand tokenized at the gateway."},"faq":[{"question":"How are chargeback operations different from dispute intake?","answer":"Intake is the customer facing moment when a cardholder reports a charge and the case is opened. Chargeback and representment is the back office work that follows: reason codes, evidence, network packages, deadlines and arbitration, on the issuer, acquirer and merchant side."},{"question":"What results have been published?","answer":"Mostly vendor figures. Stripe reports that GitHub Sponsors, using its Smart Disputes product to generate and submit dispute evidence, spends 20 hours a month less on disputes on average, and the dispute platform vendor Quavo reports that institutions using its product cut average handle time per assignment by nearly 30%. Visa announced AI dispute tools for issuers, acquirers and merchants in April 2026, some generally available and some in pilot, without outcome figures."},{"question":"Should the AI decide to write off a dispute?","answer":"Below an agreed value threshold, automatic accept or refund can make sense, because fighting a small dispute can cost more than it recovers. Above it, and for contested cases, an analyst should approve, because the network rules and the economics of each case differ."}],"related":["card-dispute-and-chargeback-intake","payment-investigations-and-exceptions","ledger-and-payment-reconciliation","fee-and-interest-leakage-detection","order-status-and-returns-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added SEO title and meta description, removed an unsupported claim about how often network rules change, tightened the EU AI Act basis, lowered the effort reduction range to match the evidence, named Quavo and the Visa tool status in the FAQ, and corrected the GitHub Sponsors metric qualifier."},{"date":"2026-09-27","note":"Review fixes: limited the card number tokenization claim to the customer facing chat, capped the effort reduction range at the cited handle time figure, swapped the recovery KPI for hours saved, kept only the sourced network dispute system, and extended the EU AI Act basis to Article 50(2) and 50(4)."}],"slug":"chargeback-and-representment","url":"https://www.blits.ai/ai-use-cases/chargeback-and-representment","benchmarks":[{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":20,"min":20,"max":20,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"github-sponsors-stripe-smart-disputes","pooled":true}]}],"indicativeValueResult":{"low":300000,"high":1650000},"evidence":["github-sponsors-stripe-smart-disputes","visa-dispute-resolution-services"]},{"title":"AI for claims triage and straight through processing","shortTitle":"Claims triage and STP","seoTitle":"AI claims triage and straight through processing","metaDescription":"AI reads each new insurance claim, routes it to the right handler and settles simple claims automatically. Lemonade reports roughly 55% of claims fully automated.","definition":"AI that reads each new insurance claim and its documents, scores its complexity, cover questions, fraud and recovery signals, sends it to the right handling path and handler, and settles simple, low risk claims end to end within set limits without a person touching them.","aliases":["claims straight through processing","touchless claims","claims segmentation","automated claims handling","claims complexity scoring"],"industries":["insurance"],"functions":["claims","operations"],"patterns":["classification-and-routing","prediction-and-scoring","document-processing","agentic-workflow","summarization"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"claims","problem":"Where new claims are routed with a handful of rules and a queue, a simple glass claim\nand a complex injury claim can wait in the same line, experienced handlers spend time on claims\nthat need no judgment, and claims can reach a handler without the authority to settle them. Every claim is read\nfrom scratch: emails, estimates, invoices, medical reports and photos, and complex claims can carry\nthousands of pages of expert evidence.\n\nThe cost shows up as slow settlement of easy claims, and as leakage\non hard claims that did not reach the right specialist early enough. Straight through processing\nhas been an ambition for years, but rules based automation stalls on unstructured documents, so\nmany simple claims still pass through a handler.","problemStats":[],"howItWorks":"1. **Read the claim.** The AI extracts the facts from the intake record and every attached\n   document (estimates, invoices, reports, photos) and writes a short claim summary.\n2. **Check cover and completeness.** It compares the loss with the policy wording and limits and\n   lists what is missing before anyone starts work.\n3. **Score and segment.** Models score complexity, expected severity, fraud risk and recovery\n   potential, and a rule table turns the scores into a handling path.\n4. **Settle the simple ones.** Claims inside strict limits (claim type, amount, clean fraud score,\n   confirmed cover) are approved and paid automatically, or declined only where a rule is\n   unambiguous and the decision is explained.\n5. **Route the rest.** Other claims go to the handler with the right skill, authority and capacity,\n   with the summary, open questions and suggested next steps attached, and the AI keeps checking\n   as new information arrives.","valueDrivers":["cost-to-serve","speed","customer-experience","employee-productivity","risk-reduction"],"kpis":["automation-rate","processing-time-reduction","handling-time-reduction","productivity-gain","accuracy","cycle-time-days"],"indicativeValue":{"referenceOrg":"A personal lines insurer that settles 200,000 claims a year","inputs":[{"key":"claims","label":"Claims per year","low":200000,"high":200000,"unit":"claims per year","note":"The reference insurer."},{"key":"stpShare","label":"Share of claims newly settled straight through","low":0.1,"high":0.3,"unit":"fraction of claims","note":"Conservative against the evidence on this page (Lemonade reports roughly 55% of claims automated end to end), because an incumbent starts from legacy systems and a broader mix of claim types."},{"key":"costPerSimpleClaim","label":"Internal handling cost of a simple claim today","low":40,"high":100,"unit":"USD per claim","note":"Editorial assumption covering handler time, checks and payment; replace with your own claims expense data."},{"key":"otherClaimsMinutesSaved","label":"Handler minutes saved on each remaining claim by the summary and routing","low":5,"high":15,"unit":"minutes per claim","note":"Editorial assumption; replace with a time study."},{"key":"costPerHour","label":"Fully loaded cost per handler hour","low":35,"high":55,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"claims * stpShare * costPerSimpleClaim + claims * (1 - stpShare) * otherClaimsMinutesSaved / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Claims handling expense avoided","caveat":"Handling expense only. It leaves out leakage reduction from better routing, customer retention from faster settlement, the risk of paying claims that a person would have questioned, and the cost of the platform, models and integration."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Routing on a summary is achievable quickly; paying claims without a person is not. Straight through settlement needs reliable document extraction, machine readable policy wording and limits, fraud screening in the same flow, payment integration and a governance model that claims, compliance and actuarial teams accept.","dataPrerequisites":["Historical claims with handling path, outcome, severity and leakage findings to train and test the scores","Policy wordings, limits and excesses in a form the system can check against","Handler skills, authority levels and capacity for routing","Labelled documents per claim type for extraction testing"],"integrations":["Claims management system for claim data, reserves, tasks and payments","Policy administration system for cover, limits and excess","Document intake and extraction for estimates, invoices and reports","Fraud detection and subrogation models as inputs to the routing rules","Payment system for automated settlement"]},"implementation":{"steps":[{"title":"Map today's handling paths and their cost","detail":"Take a year of claims and group them by type, severity and handling path. The simple, high volume groups are the straight through candidates; the complex ones are where routing and summaries pay off."},{"title":"Summaries and routing before settlement","detail":"Start with claim summaries and routing suggestions that handlers accept or correct. It builds the labelled data and trust you need before any claim is paid without a person."},{"title":"Define the straight through envelope","detail":"Write down per claim type the maximum amount, required documents, cover checks and fraud score threshold for automatic settlement, and have claims, compliance and actuarial sign it off."},{"title":"Put fraud and recovery checks in the same flow","detail":"Every claim on the automatic path must pass the fraud and subrogation checks first; speed must not open a door for fraud or leave recoveries on the table."},{"title":"Run in shadow mode, then widen","detail":"Let the system decide in parallel with handlers for a period, compare outcomes claim by claim, and switch on automatic settlement one claim type at a time."},{"title":"Audit a sample for ever","detail":"Keep reviewing a random sample of automatically settled claims, and watch leakage and complaint trends per claim type."}],"guardrails":["Automatic settlement only inside a signed off envelope per claim type (amount, documents, cover and fraud score)","Automatic declines only where a rule is unambiguous, with an explanation and a route to a person","Every automated decision logged with the inputs, scores, rules and model versions used","Fraud and recovery screening before any automatic payment","Handlers can override any routing or decision, and overrides feed back into testing"],"humanInTheLoop":"Handlers work every claim outside the envelope and can reopen any automated decision. A quality team reviews a random sample of straight through settlements every week, and claims leadership approves each change to the envelope, the rules or the models.","kpisToInstrument":["Share of claims settled straight through, per claim type","Time from first notice of loss to payment for straight through and routed claims","Share of routing suggestions handlers accept, and reasons for overrides","Leakage found in audits of automated settlements","Complaints and reopened claims after automated decisions"],"failureModes":[{"title":"Speed opens a door for fraud","detail":"Automatic payment is exactly what fraudsters look for. Keep fraud scoring in the same flow and limit amounts per claim type."},{"title":"Extraction errors become payment errors","detail":"A misread invoice total is paid automatically. Cross check extracted amounts against estimates and limits, and hold claims with low extraction confidence."},{"title":"Automated declines without recourse","detail":"Customers receive a decline they cannot challenge. Explain every decline, avoid automatic declines where cover is a judgment, and give a route to a person."},{"title":"Routing that ignores capacity","detail":"The best handler for every complex claim ends up with a queue of hundreds. Route on skill, authority and current workload together."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Claims handling as such is not listed in Annex III. The same system becomes high risk when it is also used for risk assessment and pricing of natural persons in life and health insurance (point 5(c)), or when it is used by or on behalf of a public authority to grant, reduce, revoke or reclaim essential public assistance benefits and services, including healthcare services (point 5(a)). Otherwise the tier is minimal, so the design and the operator decide. Decisions on claims based solely on automated processing are also subject to Article 22 of the GDPR and the UK GDPR."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","solvency-ii","dora","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Addressed to national supervisors (August 2025); clarifies how existing insurance sector legislation applies to AI systems, with a risk based and proportionate approach to governance, fairness, explainability and human oversight."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Lists life and health insurance pricing and risk assessment and public benefit eligibility as high risk, which matters when triage is used for those lines or schemes."}],"controls":["Signed off straight through envelope per claim type, under change control","Decision log with inputs, scores, rules and model versions for every automated settlement","Weekly random audit of automated settlements with leakage and fairness checks","Model validation and monitoring for drift in complexity, fraud and severity scores","Customer route to a person for every automated decision"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered through the API when a claim is opened. An\n**AI agent** reads the claim and its documents through the **knowledge base** document ingestion\n(PDF, images and Outlook email files), writes a summary with **structured output**, and calls **custom\nfunctions** that fetch cover and limits from the policy system, pull fraud and severity scores, and\nwrite the routing decision back to the claims system.\n\n**Human in the loop approval** holds any payment above the configured threshold for a handler, and\nthe **tool execution policy** limits which systems the agent may change. Every run has a full\n**audit trail**, **test suites** grade summaries and routing against labelled historical claims on\neach change, and **monitors** run scheduled checks against the agent and alert by email or webhook\nwhen an expectation fails. The platform is model agnostic and can run in EU or UAE regions for\ndata residency."},"faq":[{"question":"What share of claims can be settled straight through?","answer":"It depends on the lines of business and the systems. Lemonade reports that roughly 55% of its claims were automated from start to finish at the end of 2025, and Travelers says more than half of its claims are eligible for straight through digital processing. Starting with a narrow set of simple claim types and widening from there limits the risk."},{"question":"Should an AI decline claims automatically?","answer":"Rarely. Lemonade already pays or declines simple claims automatically within seconds; automatic declines are safe only where a rule is unambiguous, and the customer must get an explanation and a route to a person. Where cover is a matter of judgment, the AI should prepare the file and a handler should decide."},{"question":"Where does triage pay off if straight through settlement is not yet possible?","answer":"In handler time and routing quality. Sedgwick gives examiners priorities, claim forecasts and next step guidance inside its claims systems. At Hiscox, a senior technical claims underwriter says that with Microsoft 365 Copilot, identifying and recording the key information of a new claim now takes him as little as 10 minutes instead of up to an hour."}],"related":["claims-first-notice-of-loss-agent","claims-fraud-detection","photo-based-damage-assessment","subrogation-opportunity-detection","health-prior-authorization-and-claims-adjudication","outbound-notice-drafting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the insurance vertical with evidence from Lemonade, Travelers, Sedgwick, Hiscox, Tokio Marine & Nichido Fire, Admiral Seguros and a US travel insurer, verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; softened two unsourced statements in the problem; made the EU AI Act basis follow the Annex III wording and added UK GDPR and Solvency II; tightened the EIOPA note, the Hiscox FAQ answer and the monitors description; corrected the Sedgwick and Tokio Marine summaries."}],"slug":"claims-triage-and-straight-through-processing","url":"https://www.blits.ai/ai-use-cases/claims-triage-and-straight-through-processing","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":1,"median":55,"min":55,"max":55,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"allianz-partners-ai-claims-turnaround","pooled":false},{"id":"lemonade-ai-jim-claims-automation","pooled":true}]},{"kpi":"cycle-time-days","label":"Cycle time","unit":"days","aggregate":false,"higherIsBetter":false,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"allianz-partners-ai-claims-turnaround","pooled":false}]}],"indicativeValueResult":{"low":1325000,"high":7925000},"evidence":["admiral-seguros-ai-vehicle-damage-estimates","allianz-partners-ai-claims-turnaround","hiscox-copilot-claims-handling","lemonade-ai-jim-claims-automation","sedgwick-sidekick-agent-claims-guidance","tokio-marine-nichido-shift-claims-review","travelers-generative-ai-fnol-voice-agent"]},{"title":"AI for commercial underwriting submission intake and triage","shortTitle":"Underwriting submission triage","seoTitle":"AI for commercial underwriting submission triage","metaDescription":"AI reads commercial broker submissions, extracts the data, checks appetite and ranks each risk. AIG prepares submissions for underwriter review within one day.","definition":"AI that reads incoming broker submissions for commercial insurance (emails, applications, schedules of values, loss runs and supplements), extracts the risk data into a structured record, checks clearance and appetite, enriches the risk with internal and third party data and ranks it, so underwriters open a complete, prioritized file instead of an inbox.","aliases":["submission ingestion and clearance","broker submission triage","underwriting intake automation"],"industries":["insurance"],"functions":["underwriting","operations"],"patterns":["document-processing","classification-and-routing","prediction-and-scoring","agentic-workflow"],"channels":["email","api","internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"underwriting","problem":"Commercial and specialty insurers receive far more broker submissions than their underwriters can\nread. Each one arrives as an email with attachments in different formats: an application, a\nschedule of locations, several years of loss runs, financials and supplementals. Before anyone can\njudge the risk, someone has to clear it (is it a duplicate, is another broker already on it), check\nit against appetite, rekey the data into the rating and policy systems and pull third party data.\n\nThat work is slow and falls on expensive people. Senior underwriters triage their own inboxes, so\nin appetite business waits behind business the insurer will decline, broker turnaround slips and\nsome submissions are never looked at. Insurers treat speed as part of what they sell to brokers:\nKinsale presents service as a competitive advantage to investors and reports an average\nsubmission clearance time of 9 minutes.","problemStats":[{"statement":"Hiscox says extracting key data from email submissions is a manual process that can typically take up to three days in today's insurance operating model.","sourceTitle":"Hiscox and Google Cloud Collaborate on AI in lead underwriting for the London Market","sourceUrl":"https://www.hiscoxgroup.com/news/press-releases/2023/12-12-23","year":2023},{"statement":"Paragon Insurance Group's CTO told Kalepa that one of its programs, receiving around 50,000 submissions a year, was only looking at about 30% of its submissions.","sourceTitle":"How Paragon Doubled Its Quote-to-Bind Rate and Achieved 99% Submission Accuracy with Kalepa","sourceUrl":"https://www.kalepa.com/case-studies/paragon-doubled-quote-to-bind-rate","year":2026}],"howItWorks":"1. **Ingest.** Submissions arrive in a shared mailbox, a broker portal or a placing platform. The\n   system splits the email and attachments, classifies each document (application, loss run,\n   schedule, financials) and reads scanned and native files.\n2. **Extract to a schema.** A model extracts the fields the underwriting workbench needs (insured,\n   address, class of business, revenue, limits, loss history) into a fixed schema, with a\n   confidence score and a pointer to the page each value came from.\n3. **Clear and classify.** The record is matched against existing accounts and open submissions to\n   catch duplicates and broker conflicts, and the business is classified into the insurer's\n   industry codes.\n4. **Check appetite and enrich.** Rules and models compare the risk to the written appetite and\n   add internal history and approved third party data (firmographics, hazard scores, news).\n5. **Rank and route.** Each submission gets a priority (fit, likelihood to bind, broker\n   importance) and goes to the right underwriter or team; clear declines are drafted for a human\n   to confirm.\n6. **Underwriter decides.** The underwriter opens a decision ready file, corrects any extracted\n   field (the correction is logged and fed back) and makes every quote or decline decision.","valueDrivers":["speed","employee-productivity","revenue-growth","cost-to-serve"],"kpis":["processing-time-reduction","productivity-gain","accuracy","conversion-rate-uplift","interactions-handled","cycle-time-days","automation-rate"],"indicativeValue":{"referenceOrg":"A commercial insurer receiving 40,000 broker submissions a year","inputs":[{"key":"submissions","label":"Broker submissions per year","low":40000,"high":40000,"unit":"submissions per year","note":"The reference insurer."},{"key":"minutesPerSubmission","label":"Manual intake and triage time per submission","low":20,"high":40,"unit":"minutes per submission","note":"Editorial assumption for reading, clearance, appetite check and rekeying. Replace with a time study of your own intake."},{"key":"timeSavedShare","label":"Share of intake time the AI removes","low":0.4,"high":0.7,"unit":"fraction of intake time","note":"Editorial assumption. The evidence on this page measures elapsed time rather than effort (Sixfold reports a 50% cut in turnaround for Generali GC&C cyber submissions; AIG says submissions are prepared for underwriter review within one day, against a submission to quote process of about three to four weeks), so replace this with a time study from your own pilot."},{"key":"costPerHour","label":"Fully loaded cost of an underwriting or assistant hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption. Replace with your own fully loaded cost."}],"formula":"submissions * minutesPerSubmission / 60 * timeSavedShare * costPerHour","currency":"USD","period":"per year","resultLabel":"Underwriting intake capacity released","caveat":"Capacity released, not cash saved, unless headcount or outsourcing changes. It leaves out the usually larger effect of quoting more in appetite business faster (higher bind ratios), the cost of the platform and integration, and data licences for enrichment."},"macroEstimates":[{"statement":"Evident reports that the 30 insurers in its AI Index for Insurance announced 37 new AI use cases in the second quarter of 2026, that underwriting and pricing was the fastest growing application area, and that insurers prioritized tasks that structure and route information, such as submission intake and triage, over pricing and portfolio management.","sourceTitle":"Evident: Insurance Use Case Trends Q2 2026","sourceUrl":"https://evidentinsights.com/insights/insurance-use-case-trends-q2-2026","year":2026}],"feasibility":{"complexity":"medium","complexityNote":"Extraction from messy broker documents now works well enough to use; the work is in the schema, the clearance logic, the appetite rules and the write back into the underwriting workbench and policy system. Reported rollouts are measured in weeks per team: Sixfold's case study gives about six weeks for AXIS, Skyward Specialty reports an average deployment timeline of 8 to 10 weeks, and Paragon had its full submission inbox running through Kalepa within 90 days.","dataPrerequisites":["A written appetite per line of business, precise enough to encode as rules","A target data schema for each line (the fields the rating and policy systems need)","A labelled sample of past submissions with the correct extracted values and outcomes (quoted, declined, bound)","Account and broker master data for clearance and duplicate checks"],"integrations":["Submission mailboxes, broker portals or placing platforms","Underwriting workbench or CRM where the file is opened","Rating engine and policy administration system","Approved third party data providers (firmographics, hazard and property data)","Document storage for the original submission and extraction audit trail"]},"implementation":{"steps":[{"title":"Start with one line and its inbox","detail":"Pick a high volume line with a clear appetite (small commercial property, cyber, E&S casualty) and measure today's baseline: submissions per week, share reviewed, time to first response, quote and bind ratios."},{"title":"Fix the schema before the model","detail":"Agree with underwriters which fields matter and how they are defined. Extraction quality is judged against this schema, so vague fields produce endless disputes about accuracy."},{"title":"Measure field accuracy on real submissions","detail":"Run the extractor on a few hundred historical submissions and compare with the values underwriters actually used. Set a confidence threshold per field below which a human must confirm."},{"title":"Encode appetite and clearance as reviewable rules","detail":"Keep hard rules (excluded classes, territories, limits) deterministic and visible to underwriting management; use models only for ranking and propensity, not for silent declines."},{"title":"Put the output where underwriters work","detail":"Deliver the enriched, ranked file inside the existing workbench with a link to the source page for every value. A separate screen gets ignored."},{"title":"Close the loop","detail":"Log every correction underwriters make and every quote, decline and bind outcome, and use them to retrain extraction and ranking each month."}],"guardrails":["No automatic declines at launch; declines are drafted by the system and confirmed by an underwriter","Every extracted value carries its source page and a confidence score","Appetite rules owned and signed off by underwriting management, with version control","Only approved third party data sources for enrichment, with licence and use conditions recorded","Personal data in submissions (named individuals, claimant details) masked before it reaches a model that does not need it"],"humanInTheLoop":"Underwriters make every quote, decline and referral decision and confirm low confidence fields. Underwriting management owns the appetite rules and reviews a weekly sample of auto ranked and declined submissions to check that good business is not being buried.","kpisToInstrument":["Share of submissions reviewed within one business day, before and after","Field level extraction accuracy on a weekly audited sample","Time from submission receipt to first broker response","Quote ratio and bind ratio for top ranked versus lower ranked submissions","Underwriter corrections per submission"],"failureModes":[{"title":"Accuracy measured on the wrong thing","detail":"A single overall accuracy number hides that the fields that drive pricing (revenue, total insured value, loss history) are the ones that fail. Measure accuracy per critical field."},{"title":"Ranking that encodes yesterday's appetite","detail":"A propensity model trained on past binds keeps favouring classes the insurer is trying to exit. Retrain after every appetite change and let underwriting override the ranking."},{"title":"Silent declines and broker damage","detail":"Automated declines sent without a human check lose broker trust quickly when one is wrong. Keep a person on every decline until the error rate is proven low."},{"title":"Workbench integration left for later","detail":"If extracted data still has to be copied into the policy system, the time saving disappears. Plan the write back from the start."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Intake and triage for commercial insurance is not listed in Annex III, which covers risk assessment and pricing of natural persons in life and health insurance. It moves up to high risk only if the same pipeline is used to assess or price life or health cover for individuals."},"regulations":["eu-ai-act","gdpr","dora","nist-ai-rmf","iso-42001","solvency-ii"],"guidance":[{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Sets out how existing insurance legislation (governance, risk management, data quality, human oversight) applies to AI systems that are not high risk under the AI Act."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(c) lists life and health insurance risk assessment and pricing of natural persons; commercial lines triage is outside it."}],"controls":["AI inventory entry per line with an accountable underwriting owner","Documented appetite rules with version history and sign off","Audit trail linking each ranked or declined submission to the data and rule versions used","Weekly quality sample of extraction and triage outcomes","Third party and model provider risk assessment under DORA for hosted AI services"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this runs as an **agentic workflow**, started through the **REST API** from the\nbroker portal or mailbox integration, or from a flow on the inbound **email channel** with a\ntrigger workflow block. The platform's document ingestion reads PDF, DOCX, XLSX and Outlook .msg\nfiles, and an **agent with structured output** extracts the fields into the insurer's schema. **Custom functions** call the clearance service, the appetite rules\nand approved data providers over REST, and a **SQL knowledge base** gives the agent read access to\naccount and loss history.\n\nDeclines and referrals go through **human in the loop approval** above a configurable threshold,\nso an underwriter confirms before anything reaches the broker. **PII masking** at the gateway,\nwith custom masking patterns per bot, masks identifiers such as email addresses, phone numbers\nand account or policy numbers before they reach a model. Every run has a full **audit trail and\ntrace**, and **test suites** replay a labelled set of past submissions before each change goes\nlive to track field accuracy. The\nplatform is model agnostic, so extraction and ranking can use different models, and it can run in\nEU or UAE data residency regions."},"faq":[{"question":"How accurate is AI extraction from broker submissions?","answer":"Good enough to use when measured per field and checked on low confidence values. Paragon's CTO says its submission extraction is around 98 to 99% accurate with Kalepa, better than its former manual team. That is a single figure from one deployment; accuracy can differ between applications, schedules and scanned loss runs, so measure it per field and per document type."},{"question":"Should the AI decline submissions on its own?","answer":"Not at first. AIG and Hiscox describe AI that extracts, enriches or prices while an underwriter reviews the output before anything goes to the broker, and Skyward Specialty says its Sixfold platform ranks and assesses submissions while keeping underwriters in the loop. Hard appetite rules can draft declines, but a person should confirm them until the error rate is proven."},{"question":"What does it do for brokers?","answer":"Faster answers. Sixfold reports that Generali GC&C cut turnaround for its distribution channels on cyber submissions by 50%, and AIG says its assistant prepares submissions for underwriter review within one day instead of a process it described as taking weeks."},{"question":"Is submission triage high risk under the EU AI Act?","answer":"For commercial lines, no: Annex III covers life and health insurance risk assessment and pricing for natural persons. Insurance supervisors still expect governance, data quality and human oversight under existing rules, as EIOPA's 2025 opinion on AI governance sets out."}],"related":["underwriting-risk-assessment-copilot","intelligent-document-processing","correspondence-triage-and-routing","insurance-renewal-and-retention","insurance-broker-and-agent-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from insurer filings, press releases and vendor case studies verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added Solvency II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Paragon problem statistic and its year; replaced unsourced claims on broker behaviour, rollout time and per document accuracy with cited facts; attributed the Generali GC&C turnaround to Sixfold; aligned the Blits.ai build notes with the feature inventory; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: limited the PII masking claim to what the gateway masks; aligned the Hiscox, Skyward Specialty and AIG wording with their sources; softened the Skyward Specialty human review claim; corrected the Markel turnaround to an SLA."}],"slug":"commercial-underwriting-submission-triage","url":"https://www.blits.ai/ai-use-cases/commercial-underwriting-submission-triage","benchmarks":[{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":98,"min":98,"max":98,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"paragon-kalepa-submission-triage","pooled":true}]},{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2,"min":2,"max":2,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"paragon-kalepa-submission-triage","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"generali-gcc-sixfold-cyber-underwriting","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":15000,"min":15000,"max":15000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"axis-sixfold-submission-classification","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":113,"min":113,"max":113,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"markel-cytora-submission-triage","pooled":true}]},{"kpi":"cycle-time-days","label":"Cycle time","unit":"days","aggregate":false,"higherIsBetter":false,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"aig-underwriter-assistance","pooled":false}]}],"indicativeValueResult":{"low":320000.00000000006,"high":1866666.6666666667},"evidence":["aig-underwriter-assistance","axis-sixfold-submission-classification","cna-ai-underwriting-triage","generali-gcc-sixfold-cyber-underwriting","hiscox-generative-ai-lead-underwriting","kinsale-ai-submission-routing","markel-cytora-submission-triage","paragon-kalepa-submission-triage","skyward-specialty-sixfold-ai-underwriting"]},{"title":"AI for complaints root cause and systemic issue analysis","shortTitle":"Complaints root cause analysis","seoTitle":"AI for complaints root cause analysis","metaDescription":"AI groups every complaint into themes and traces themes to likely causes for analysts to validate. The Federal Reserve Board clusters complaints; CMS pilots it.","definition":"AI that reads the free text of complaints across all channels, clusters them into themes, separates systemic causes from one off events, links each theme to the product, process or control behind it and routes the insight to the owner who can fix it, with a human validating every root cause and every remediation.","aliases":["complaint theme analysis","systemic issue detection","complaints insight analytics","voice of the customer root cause"],"industries":["cross-industry","banking","insurance","payments","telecommunications","government"],"functions":["regulatory-compliance","customer-service","analytics-and-reporting"],"patterns":["classification-and-routing","summarization","agentic-workflow","rag-knowledge-assistant"],"channels":["internal-tools"],"audience":"back-office","autonomy":"copilot","adoptionStage":"emerging","segment":"second-line","problem":"Every complaint is handled one by one, but the reason it happened is rarely unique. The same\nunclear letter, broken app journey or misapplied fee produces hundreds of complaints, spread\nacross phone notes, emails, chat logs and ombudsman referrals, each coded slightly differently\nby a different handler. Complaint categories are built for handling and reporting volumes, not\nfor finding causes, so management information shows how many complaints arrived, not why.\n\nRegulators expect more. UK rules require firms to identify and remedy recurring or systemic\nproblems, and the FCA's review of complaints and root cause analysis at 40 firms found that\nfirms did not always measure whether their fixes worked, and that some complaints reports appeared\nto be prepared for operational purposes such as resourcing without also looking at how to improve\ncustomer outcomes. Manual root cause work at this volume often relies on samples, and the theme that matters most\ncan be the one nobody sampled.","problemStats":[{"statement":"The FCA's thematic review of complaints and root cause analysis at 40 firms found that firms did not always measure the impact of the changes they made after finding a root cause.","sourceTitle":"Complaints and root cause analysis: good practice and areas for improvement","sourceUrl":"https://www.fca.org.uk/publications/good-and-poor-practice/complaints-and-root-cause-analysis-good-practice-and-areas-improvement","year":2024}],"howItWorks":"1. **Collect every complaint.** Complaint records, call and chat transcripts, emails and\n   ombudsman cases land in one store with product, channel, outcome and customer segment.\n2. **Normalise and deduplicate.** The AI summarises each complaint in a standard form, removes\n   duplicates about the same event and tags vulnerability signals.\n3. **Cluster into themes.** Complaints are grouped by what actually went wrong, not by the code\n   a handler picked, and each theme gets a plain language description with example cases.\n4. **Trace to a cause.** Retrieval over process maps, product terms, change logs and the control\n   library links each theme to the likely process, product change or control failure, and flags\n   themes that grow, spread across products or hit vulnerable customers.\n5. **Validate and assign.** A root cause analyst validates the theme and the cause on a sample,\n   then assigns an owner in the business, not the complaints team.\n6. **Track the fix.** Actions, owners and dates are tracked, and the AI measures whether the\n   theme shrinks after the fix, which is the evidence regulators ask for.","valueDrivers":["compliance","customer-experience","risk-reduction","cost-to-serve"],"kpis":["processing-time-reduction","productivity-gain","accuracy","interactions-handled"],"indicativeValue":{"referenceOrg":"A retail bank receiving 100,000 complaints a year","inputs":[{"key":"complaints","label":"Complaints received per year","low":100000,"high":100000,"unit":"complaints per year","note":"The reference bank. Replace with your own volume."},{"key":"systemicShare","label":"Share of complaints linked to a fixable systemic cause","low":0.1,"high":0.2,"unit":"fraction of complaints","note":"Editorial assumption. Replace with the share your own root cause work attributes to recurring causes."},{"key":"avoided","label":"Share of those complaints avoided after the cause is fixed","low":0.2,"high":0.4,"unit":"fraction of systemic complaints","note":"Editorial assumption, replace with your own. No public benchmark exists yet."},{"key":"costPerComplaint","label":"Fully loaded cost of handling one complaint","low":150,"high":300,"unit":"USD per complaint","note":"Editorial assumption covering handling time and review, excluding redress. Replace with your own."}],"formula":"complaints * systemicShare * avoided * costPerComplaint","currency":"USD","period":"per year","resultLabel":"Complaint handling cost avoided by fixing systemic causes","caveat":"Handling cost only. It leaves out redress and remediation programmes avoided, ombudsman fees, the customer and conduct benefit and the cost of the fixes themselves, and it assumes the business acts on the insight."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Clustering text is straightforward; linking themes to causes needs process, product and change data, and the value depends on a governance route that makes business owners act.","dataPrerequisites":["Complaint records with free text, product, channel, outcome and redress","Call and chat transcripts or notes linked to the complaint","Process maps, product terms, change and incident logs, and the control library","A small set of complaints with expert validated root causes for testing"],"integrations":["Complaints and case management system","Contact centre transcripts and conversation analytics","Change management, incident and control library (GRC) systems","Business intelligence and conduct risk reporting"]},"implementation":{"steps":[{"title":"Build the evidence base first","detail":"Bring complaint text, transcripts and outcomes together for at least twelve months, with vulnerability and redress fields, before any modelling. Themes need history to show trend."},{"title":"Let themes emerge, then fix the taxonomy","detail":"Run clustering on the full population, have root cause analysts name and merge themes, and publish a stable theme list. Keep complaint handling codes separate."},{"title":"Link themes to owners","detail":"Map each theme to a process, product and control owner using the control library and change log, so insight goes to the person who can fix it, not back to complaints."},{"title":"Validate on samples","detail":"For every theme, analysts review a sample of complaints against the AI description and record agreement. Themes below the agreement threshold are not reported."},{"title":"Close the loop","detail":"Track actions to closure and measure the theme's volume and severity after each fix, and report both to the conduct or risk committee."}],"guardrails":["A human root cause analyst validates every theme and cause before it is reported or actioned","Every theme links to example complaints so a reviewer can check it","Personal data masked in prompts; outputs report themes, not individual customers","Vulnerability and detriment themes always escalate, regardless of volume","Theme definitions and model changes are versioned so trends stay comparable"],"humanInTheLoop":"Root cause analysts validate themes and causes, business owners decide the remediation, and the conduct or risk committee reviews progress. The AI never decides redress or closes an individual complaint.","kpisToInstrument":["Share of complaints read by the analysis (target the full population, not a sample)","Analyst agreement rate with AI themes and causes on samples","Time from a theme emerging to an owner being assigned","Complaint volume per theme before and after remediation","Share of remediation actions with measured impact"],"failureModes":[{"title":"Themes that mirror the handling codes","detail":"The model learns the existing categories and finds nothing new. Cluster on the free text and compare with codes, rather than training on codes."},{"title":"Insight with no owner","detail":"Reports go back to the complaints team and nothing changes. Route every validated theme to a named business owner with a date."},{"title":"Plausible but wrong causes","detail":"The AI attributes a theme to a recent change because it is in the retrieved context. Require evidence and analyst validation before a cause is reported."},{"title":"Volume blindness","detail":"Small but severe themes, such as harm to vulnerable customers, are ranked low. Weight by severity and vulnerability, not volume alone."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Analysing complaints in aggregate to find causes is not listed in Annex III, is not a practice prohibited by Article 5 and does not decide on individuals. It does not interact with the public, so the disclosure duty in Article 50(1) does not apply; the machine readable marking of generated text in Article 50(2) is a duty of the provider of the generative model or system that writes the summaries. If the same system decided individual complaint outcomes or redress, or its themes were used to evaluate the performance of individual complaint handlers (Annex III point 4), that design would need its own assessment."},"regulations":["uk-consumer-duty","gdpr","uk-gdpr","eu-ai-act","iso-42001"],"guidance":[{"title":"DISP 1.3 Complaints handling rules","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.handbook.fca.org.uk/handbook/DISP/1/3.html","note":"Firms must put management controls in place to identify and remedy any recurring or systemic problems found in complaints."},{"title":"Complaints and root cause analysis: good practice and areas for improvement","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publications/good-and-poor-practice/complaints-and-root-cause-analysis-good-practice-and-areas-improvement","note":"Thematic review of 40 firms with examples of good practice, including involving the owner of the process in root cause work and measuring the impact of fixes."},{"title":"RG 271 Internal dispute resolution","issuer":"Australian Securities and Investments Commission","region":"asia-pacific","url":"https://asic.gov.au/regulatory-resources/find-a-document/regulatory-guides/rg-271-internal-dispute-resolution/","note":"ASIC's standards for internal dispute resolution by Australian financial firms. Enforceable paragraph RG 271.120 requires firms to regularly analyse complaint data sets to identify systemic issues and escalate them for investigation and action."}],"controls":["Documented method for theme detection and cause attribution, owned by the complaints or conduct function","Sample based validation with recorded agreement rates per theme","Action log from theme to owner to fix to measured impact","Regular reporting of themes and actions to senior management and the board","Data protection impact assessment for the use of complaint text and transcripts"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that runs on a schedule over new complaints. It reads\ncomplaint records and transcripts through **custom functions** and a **SQL knowledge base**,\nsummarises each case into **structured output**, and groups cases into themes. A **knowledge\nbase** with hybrid retrieval over process documents, product terms and the control library lets\nthe agent propose the likely cause from the retrieved documents.\n\n**Human in the loop approval** holds each proposed theme and cause until a root cause analyst\napproves it, and only approved themes are pushed to the owner's case system. Complaint extracts\nare masked or redacted before ingestion, as a design step in the customer's own pipeline, **test\nsuites** evaluate theme and cause quality against expert labelled samples, and per run audit trails show how\nevery reported theme was built. The platform is model agnostic and supports EU and UAE data\nresidency."},"faq":[{"question":"How is this different from AI complaint handling?","answer":"Complaint handling works one complaint at a time: recognise, investigate, respond. Root cause analysis works across all complaints to find the recurring causes behind them and get them fixed, which is what rules such as the FCA's DISP 1.3 require on top of good handling."},{"question":"Can AI find root causes on its own?","answer":"It can read every complaint rather than a sample, cluster themes and propose likely causes. A human analyst should validate each cause, because a plausible explanation from retrieved context is not proof, and a business owner has to decide the fix."},{"question":"Who already does this?","answer":"The public records on this page come from US agencies. The Federal Reserve Board has applied topic modelling to consumer complaint narratives from the CFPB database since 2019, and the Federal Trade Commission has classified the complaints it receives and grouped duplicates with machine learning since 2019. Of these, only the US Centers for Medicare and Medicaid Services goes as far as root causes, and it is still a pilot that finds root causes and trends in complaint cases for expert validation."}],"related":["complaints-handling-agent","customer-feedback-analysis","call-quality-and-compliance-monitoring","continuous-controls-testing","supervisory-exam-response-assembly"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against FCA publications and US federal AI inventories."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: dated the FCA statistic to its 2024 publication, aligned the FCA resourcing finding with its wording, sharpened the RG 271 note to paragraph RG 271.120, added UK GDPR, added the FTC example to the FAQ, removed an unsupported speed claim, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: adoption stage set to emerging because the only root cause deployment on the page is a pilot, removed Blits.ai claims not in the feature inventory (citations, tests on every change), softened the meta description and the sampling claim, aligned the FCA finding with its wording, and extended the EU AI Act basis to Article 50(2) and Annex III point 4."}],"slug":"complaints-root-cause-analysis","url":"https://www.blits.ai/ai-use-cases/complaints-root-cause-analysis","benchmarks":[],"indicativeValueResult":{"low":300000,"high":2400000},"evidence":["cms-complaint-root-cause-analysis","federal-reserve-board-consumer-complaints-explorer","ftc-consumer-complaint-classification-and-grouping"]},{"title":"AI for continuous controls testing and control self assessment","shortTitle":"Continuous controls testing","seoTitle":"AI for continuous controls testing and RCSA","metaDescription":"AI for continuous controls testing. The US Interior Department uses AI tools that produce audit ready records for over 29,000 financial assistance actions a year.","definition":"AI that moves control testing from periodic samples to continuous, full population assurance: it collects evidence from source systems, maps each artefact to the control it supports, tests every transaction or record against the control's rule, flags exceptions for a human to judge and prepares the risk and control self assessment from incident and loss data for the business to review.","aliases":["continuous controls monitoring","AI control testing","automated RCSA","full population control testing"],"industries":["cross-industry","banking","insurance","capital-markets","government"],"functions":["risk-management","regulatory-compliance","operations"],"patterns":["agentic-workflow","document-processing","anomaly-detection","classification-and-routing"],"channels":["internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"second-line","problem":"Periodic control testing checks a sample of items at a point in time, with evidence gathered\nfrom control owners by email and screenshots. The first line collects the evidence, the second\nline reviews it, and the result describes the control as it was when the sample was drawn. A\ncontrol that fails soon after a test can go unnoticed until the next cycle, and is found only if\nthe failing items happen to be in the sample.\n\nRisk and control self assessments have the same weakness. Anaptyss, a services vendor, describes\nthem as heavily manual, with control narratives that differ across teams and business units. Rules\nraise the bar at the same time: APRA CPS 230 requires regulated entities to regularly monitor,\nreview and test controls for design and operating effectiveness and to report the results to\nsenior management.","problemStats":[{"statement":"Anaptyss, a services vendor, states that risk and control self assessments in banks often take three to six weeks to complete, with some complex assessments extending beyond eight weeks.","sourceTitle":"From 20 Days to 5: The Operational Economics of AI-Led RCSA Execution","sourceUrl":"https://www.anaptyss.com/blog/from-20-days-to-5-the-operational-economics-of-ai-led-rcsa-execution/","year":2026}],"howItWorks":"1. **Codify the control.** Each automatable control gets a testable rule (for example: every\n   payment above a threshold has a second approver who is not the initiator) and a list of the\n   evidence that proves it.\n2. **Collect evidence continuously.** Agents pull records, logs, approvals and documents from\n   source systems on a schedule, instead of asking control owners for screenshots.\n3. **Read the artefacts.** Document AI reads policies, sign off records, reconciliations and\n   reports and maps each artefact to the control and the period it covers.\n4. **Test the whole population.** Every item is tested against the rule; exceptions are grouped\n   and explained with the evidence attached.\n5. **Human judgement on exceptions.** A tester or control owner reviews each exception, decides\n   whether it is a control failure and records the reason.\n6. **Prepare the RCSA.** The AI drafts control narratives and proposed ratings from test\n   results, incidents and losses; the business reviews, challenges and owns the final assessment.","valueDrivers":["compliance","risk-reduction","employee-productivity","speed"],"kpis":["automation-rate","processing-time-reduction","hours-saved","interactions-handled","error-reduction"],"indicativeValue":{"referenceOrg":"A bank that tests 2,000 key controls a year","inputs":[{"key":"controls","label":"Key controls tested per year","low":2000,"high":2000,"unit":"controls per year","note":"The reference bank. Replace with your own control library."},{"key":"hoursPerTest","label":"Hours per control test (evidence collection, testing, review)","low":10,"high":20,"unit":"hours per control test","note":"Editorial assumption across first and second line effort. Replace with your own time records."},{"key":"automatable","label":"Share of controls suitable for automated testing","low":0.2,"high":0.4,"unit":"fraction of controls","note":"Editorial assumption; system based controls automate first, judgement based controls stay manual."},{"key":"effortSaved","label":"Share of test effort saved on those controls","low":0.4,"high":0.7,"unit":"fraction of test effort","note":"Editorial assumption, replace with your own pilot results. No public bank benchmark was found."},{"key":"hourlyCost","label":"Fully loaded cost of a control tester","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"controls * hoursPerTest * automatable * effortSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Control testing effort released","caveat":"Effort only. It leaves out the build and integration cost, the value of finding failures months earlier across the full population, and the saving in audit and regulatory findings."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Testing logic is simple once a control is codified. The work is in rewriting controls so they are testable, reaching evidence in dozens of source systems and agreeing with audit and the regulator that automated tests are reliable.","dataPrerequisites":["A control library with owners, objectives and testable attributes","Access to transaction, approval, access and configuration data in source systems","Incident, loss and issue data linked to controls","Prior test results to compare automated and manual outcomes"],"integrations":["Governance, risk and compliance (GRC) platform with the control library and issues","Core banking, payments, identity and access management and ticketing systems","Document stores for policies, sign offs and reconciliations","Incident and operational loss databases"]},"implementation":{"steps":[{"title":"Pick controls that are data rich and rule based","detail":"Start with access, approval, reconciliation and segregation of duties controls whose evidence already sits in systems. Leave judgement heavy controls for later."},{"title":"Rewrite each control as a test","detail":"For every control in scope, write the rule, the population, the evidence and the exception definition, and have the control owner and second line sign it off."},{"title":"Run parallel with manual testing","detail":"For one or two cycles test both ways, compare results and explain every difference before retiring the manual test. Share the comparison with internal audit."},{"title":"Industrialise exception handling","detail":"Route exceptions to owners with evidence, reasons and due dates, and link confirmed failures to issues and the RCSA."},{"title":"Draft, never auto approve, the RCSA","detail":"Use AI to prepare narratives and proposed ratings from data, and require the business to review and change them, so the assessment stays theirs."}],"guardrails":["Every exception and every proposed rating is reviewed and decided by a human","Test rules are versioned and signed off by the control owner and second line","The agent reads source systems with read only access through an allow list","Every test run stores its population, evidence and result for audit","Automated tests are recalibrated when the underlying process or system changes"],"humanInTheLoop":"Control owners and testers adjudicate every exception, second line approves test designs, the business owns the self assessment ratings and internal audit reviews the reliability of the automated testing.","kpisToInstrument":["Share of key controls tested on the full population","Time from control failure to detection","Exceptions raised, confirmed as failures and closed on time","Hours of evidence collection and testing per control, before and after","Differences between automated and manual test outcomes during parallel runs"],"failureModes":[{"title":"Testing the wrong population","detail":"The agent tests the data it can reach, not the full population the control covers. Reconcile populations to source totals every run."},{"title":"Rubber stamped RCSA","detail":"The business accepts AI drafted ratings without challenge and the self assessment loses its point. Require recorded review and track how often drafts change."},{"title":"Silent test decay","detail":"A system change breaks the extraction and the test passes on empty data. Alert on volume anomalies and zero populations."},{"title":"Exception floods","detail":"A poorly specified rule produces thousands of exceptions that nobody reviews. Tune on parallel runs and group exceptions by cause."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Testing controls over transactions and systems is not an Annex III use. Controls that monitor and evaluate individual employees' behaviour, such as trading or access conduct, can fall under Annex III point 4(b), so the design decides the tier."},"regulations":["eu-ai-act","dora","apra-cps-230","gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Revisions to the principles for the sound management of operational risk","issuer":"Basel Committee on Banking Supervision","region":"global","url":"https://www.bis.org/bcbs/publ/d515.htm","note":"The Basel Committee's principles for operational risk management and the control environment, revised in 2021 with updated guidance on change management and ICT."},{"title":"Operational risk management (CPS 230)","issuer":"Australian Prudential Regulation Authority","region":"asia-pacific","url":"https://www.apra.gov.au/operational-risk-management","note":"Australia's cross industry operational risk standard for APRA regulated entities, covering operational risk controls, critical operations and material service providers."},{"title":"Article 4, AI literacy","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/4/","note":"Providers and deployers must take measures to support the AI literacy of staff who use and oversee AI systems (the amended wording asks them to support it rather than ensure a sufficient level). Here that means training testers on the limits of automated results."}],"controls":["Inventory entry for the testing AI with an owner, scope and validation status","Signed off test specifications per control, under change control","Population reconciliation and completeness checks on every run","Parallel run evidence retained before a manual test is retired","Internal audit review of the reliability of automated testing"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** per control family, triggered on a schedule or\nthrough the API. **Custom functions** (REST calls and SQL queries) pull populations and evidence\nfrom source systems with read only access, a **SQL knowledge base** lets the agent query test\ndata directly, and the **knowledge base** ingests policies, sign offs and reports in PDF, Word,\nExcel and email formats so the agent can map artefacts to controls. Results come back as\n**structured output** that the GRC platform can store.\n\n**Agentic tasks** with scheduled rechecks keep testing continuous, the **tool execution policy**\nfixes which tools each agent may use, and **human in the loop confirmation** holds agentic\nactions above a configurable threshold for an approve or reject decision. Exceptions and\nproposed RCSA ratings go to testers and control owners as structured results, and people make\nthe decision. The per run audit trail and downloadable\nrun data give auditors the evidence, and **test suites** and **monitors** catch prompt or model\nchanges that would alter test outcomes. The platform is model agnostic with EU and UAE data\nresidency."},"faq":[{"question":"Does continuous controls testing replace control testers?","answer":"It changes their work. Evidence collection and routine testing move to automation, and testers spend their time on exceptions, control design and judgement based controls that cannot be codified."},{"question":"Which controls should be automated first?","answer":"Controls whose evidence already sits in systems and whose rule can be written precisely: access reviews, approvals and limits, reconciliations and segregation of duties. Judgement heavy controls stay manual or AI assisted rather than automated."},{"question":"Is there public evidence this works?","answer":"Public bank figures are scarce. The US Department of the Interior reports AI tools in production since 2024 that run internal controls testing and produce audit ready records for more than 29,000 financial assistance actions a year. The FDIC and the Pension Benefit Guaranty Corporation list similar tools as pre deployment in the 2025 federal AI inventory. Vendor claims about faster RCSA cycles exist but are not tied to named banks."}],"related":["internal-audit-copilot","policy-drafting-and-gap-analysis","complaints-root-cause-analysis","vendor-due-diligence","ai-model-inventory","regulatory-horizon-scanning"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against public sources. The catalog's RCSA claim is kept only as an attributed vendor statement."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources. Rewrote the problem section without unsourced sample sizes and adoption claims, tied the regulatory point to CPS 230 paragraph 29, aligned the Blits.ai human in the loop wording with the feature inventory, named the FDIC and PBGC entries in the FAQ, corrected the Interior, FDIC and PBGC evidence summaries, and added an SEO title and description."},{"date":"2026-09-27","note":"Second fact check. Removed an unsourced claim about incidents and losses from the Anaptyss attribution and an unsourced testing frequency, rewrote the SEO description to match the Interior inventory entry, set the adoption stage to emerging (one production deployment, two pre deployment), added the threshold to the human in the loop wording and corrected the Interior and FDIC evidence records."}],"slug":"continuous-controls-testing","url":"https://www.blits.ai/ai-use-cases/continuous-controls-testing","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":29000,"min":29000,"max":29000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"interior-department-grants-internal-controls-testing","pooled":true}]}],"indicativeValueResult":{"low":96000,"high":1120000},"evidence":["fdic-transactional-data-monitoring","interior-department-grants-internal-controls-testing","pbgc-security-and-privacy-control-assessment"]},{"title":"AI for court and case file summarization","shortTitle":"Case file summarization","seoTitle":"AI court and case file summarization","metaDescription":"AI summarizes court filings, case files and recorded evidence for staff to check. See how the UK CPS and Brazil's Supreme Court use it, and the EU AI Act rules.","definition":"AI that condenses court filings, case files, evidence recordings and earlier decisions into structured summaries, chronologies and draft case reports with references to the source pages, so that judges, prosecutors, tribunal staff and government lawyers find what matters faster, while the person responsible reads the underlying material and makes every legal judgment.","aliases":["case file summarisation AI","court filing summaries","judicial case report drafting","legal case bundle summarization"],"industries":["government"],"functions":["legal","case-management"],"patterns":["summarization","document-processing","rag-knowledge-assistant"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Courts, tribunals, prosecutors and government legal teams work through very large files. A single\nimmigration filing can run to hundreds of pages that the parties have not organised; a prosecution\nmay rest on long recorded interviews that prosecutors have to watch while taking notes manually;\nBrazil's Federal Supreme Court drafts case reports and headnotes for the appeals it decides. Finding,\nordering and summarising this material is slow, manual work that comes before the legal analysis starts.\n\nSummaries are an obvious help and an obvious risk. A summary that omits a key fact, misattributes a\nstatement or invents a citation can distort a decision about someone's liberty, residence or\nrights, and the High Court of England and Wales has already dealt with fictitious case citations,\nsuspected to come from generative AI, put before it.\nThe design has to keep every summary traceable to the record and every judgment with a person.","problemStats":[],"howItWorks":"1. **Ingest the file.** Filings, bundles, evidence recordings and earlier decisions are uploaded in\n   the secure environment; scans are converted to text and recordings are transcribed with time\n   stamps.\n2. **Organise.** Documents are classified by type (application, evidence, submission, decision),\n   tabbed and deduplicated, so the record has a navigable structure.\n3. **Summarise with references.** The model produces a summary in an agreed template (parties, key\n   facts, chronology, issues, relief sought) where every statement points to the page or time stamp\n   it came from.\n4. **Draft routine documents.** For high volume work it drafts standard documents such as case\n   reports or headnotes for a staff member to review and adapt.\n5. **Verify and decide.** The judge, prosecutor or lawyer checks the summary against the record and\n   makes every decision; feedback on errors improves the templates.","valueDrivers":["employee-productivity","speed"],"kpis":["time-saved-per-task","processing-time-reduction","accuracy","hours-saved"],"indicativeValue":{"referenceOrg":"A tribunal or prosecution service that reviews 20,000 case files a year","inputs":[{"key":"files","label":"Case files reviewed per year","low":10000,"high":30000,"unit":"case files per year","note":"Editorial assumption. For scale, the Crown Prosecution Service expects to use its video evidence tool on about 11,000 cases a year.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/the-crown-prosecution-service-beam-notes"},{"key":"hoursSaved","label":"Staff hours saved per file after checking","low":0.5,"high":1.5,"unit":"hours per file","note":"Editorial assumption. No organization on this page publishes a measured saving; the person must still read the underlying material."},{"key":"hourlyCost","label":"Fully loaded cost of a lawyer's or case officer's hour","low":50,"high":90,"unit":"EUR per hour","note":"Editorial assumption. Replace with your own staff cost."}],"formula":"files * hoursSaved * hourlyCost","currency":"EUR","period":"per year","resultLabel":"Legal staff time released","caveat":"Time released, not cash saved. It leaves out the value of shorter backlogs and faster decisions for the people involved, the cost of secure infrastructure and assurance, and the cost of an error that reaches a decision."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Summarisation itself is mature. The work is in secure hosting for highly sensitive material, templates agreed with the judiciary or legal profession, reliable source references, and a culture where summaries support rather than replace reading the record.","dataPrerequisites":["Digital case files, or scans that can be converted to text reliably","Templates for the summaries and standard documents, agreed with users","A set of files with expert summaries to evaluate against"],"integrations":["Case management system of the court, tribunal or prosecution service","Document and evidence stores, including audio and video evidence","Identity and access management with need to know permissions per case"]},"implementation":{"steps":[{"title":"Start with staff facing preparation, not decisions","detail":"Begin with summaries that help staff navigate a file, such as the Crown Prosecution Service's summaries of video interviews, or the summaries of earlier objection advice that Amsterdam has registered for its lawyers' internal case library."},{"title":"Agree templates with the users","detail":"Co design the summary structure with judges, prosecutors or lawyers, as the CPS did with its supplier, so the output fits the way they work."},{"title":"Require references for every statement","detail":"Make each summary point link to the page or time stamp in the record, so checking is quick and omissions show."},{"title":"Evaluate against expert summaries","detail":"Measure omissions and factual errors on a sample of files summarised by experienced staff before scaling."},{"title":"Train users on limits","detail":"Mandate training before access and make clear that the summary never replaces reading or watching the evidence, as the CPS requires."}],"guardrails":["The person responsible reads or watches the underlying material; the summary is an aid","Every summary statement is referenced to the record","No generation of legal authorities; case law is retrieved from trusted databases and verified","Hosting and retention appropriate to the sensitivity of the case material","AI generated content labelled as such in the case file"],"humanInTheLoop":"Judges, prosecutors and lawyers review every summary against the record and make every decision. Summaries are never placed in the case record or shared with parties without review, and users can flag and correct errors.","kpisToInstrument":["Time to prepare a case for review, before and after","Omission and error rate on a sampled set of summaries","Share of summaries rated usable without major edits","Backlog and time to decision","Errors reported by users, and time to fix templates"],"failureModes":[{"title":"Omitted facts","detail":"A summary that leaves out a key fact or a contradiction in the evidence can steer the reader. Require references and sample for omissions."},{"title":"Invented authorities","detail":"Generative models can produce plausible but fictitious case law. Only cite from trusted legal databases and verify every authority."},{"title":"Over reliance","detail":"Under time pressure staff read the summary instead of the record. Make reading the source part of the process and audit it."},{"title":"Transcription errors in names and details","detail":"Speech recognition misspells names or mishears details in recorded evidence. Let users correct transcripts and check key details against the recording."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 8(a) makes AI high risk when it is intended to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts. Tools for prosecutors fall under point 6(c) if they evaluate the reliability of evidence, and tools that assist the examination of asylum, visa or residence applications fall under point 7(c). Under Article 6(3) a system that only performs a narrow procedural task or a preparatory task, such as organising a file or transcribing and summarising it for the person who decides, may not be high risk, but the provider must document that assessment (Article 6(4)). Summaries of internal legal advice for government lawyers, as Amsterdam plans, are generally outside Annex III."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Artificial Intelligence (AI) judicial guidance (October 2025)","issuer":"Courts and Tribunals Judiciary of England and Wales","region":"europe","url":"https://www.judiciary.uk/guidance-and-resources/artificial-intelligence-ai-judicial-guidance-october-2025/","note":"Guidance for judicial office holders on confidentiality, hallucinations and bias; it lists summarising large bodies of text as a potentially useful task, provided the summary is checked for accuracy."},{"title":"EU AI Act Annex III, high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 8(a) covers AI used by or for judicial authorities to research and interpret facts and law; points 6(c) and 7(c) cover evaluating evidence in criminal cases and examining asylum, visa and residence applications."}],"controls":["Documented Article 6(3) assessment or high risk conformity work for court facing tools","Transparency record and user guidance for every tool","Mandatory user training before access","Sampled quality review of summaries against the record","Retention and access controls matched to case sensitivity"],"incidents":[{"title":"High Court of England and Wales: Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin)","url":"https://www.judiciary.uk/judgments/ayinde-v-london-borough-of-haringey-and-al-haroun-v-qatar-national-bank/","note":"The Divisional Court dealt with two cases in which fictitious or inaccurate case citations, suspected to come from generative AI, were put before the court, and warned the legal profession about its duty to verify authorities."}]},"blitsAi":{"howToBuild":"On Blits.ai case files are ingested into a secure **knowledge base** (PDF, DOCX, images and Outlook\nemail files), and **self hosted transcription with speaker diarization** turns recorded interviews into\ntime stamped transcripts on Blits.ai infrastructure. An **agentic workflow** classifies the\ndocuments, produces a summary in the agreed template with references to pages and time stamps, and\nwaits for **human in the loop approval** before anything is saved to the case system through\n**custom functions**.\n\nAn **output guardrail** with an admin authored policy blocks responses that present case law as\nauthority, and **role based access control** with per bot roles limits who can use the tool and\nview its logs. **PII masking** at the gateway can mask configured patterns, such as emails, phone\nnumbers and IBANs, in user messages, and **test suites** grade summaries against expert summaries\nbefore each change goes live. The platform is **model agnostic** and offers EU and UAE data\nresidency for sensitive case material, with a model choice that fits the material's sensitivity."},"faq":[{"question":"Do courts use AI to summarise case files?","answer":"Some do, with staff in control. Brazil's Federal Supreme Court uses its Maria platform to draft case reports and headnotes for staff to review, the UK Crown Prosecution Service summarises video interviews with Beam Notes, and the US immigration courts have listed filing summaries as a planned use."},{"question":"Is AI summarisation for judges high risk under the EU AI Act?","answer":"It can be. Annex III point 8(a) covers AI that assists judicial authorities in researching and interpreting facts and the law. A purely preparatory tool, such as one that organises and summarises a file, may fall under the Article 6(3) exception, but the provider has to document that assessment."},{"question":"What is the biggest risk?","answer":"That a summary or a generated citation is trusted without checking. English courts have already dealt with fictitious authorities put before them. Keep summaries referenced to the record and verify every authority."}],"related":["freedom-of-information-request-processing","civil-servant-drafting-copilot","enterprise-knowledge-search","public-service-translation","meeting-summarization-and-action-items"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version for the government vertical, with Brazilian, UK, Dutch and US public sector evidence verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsourced figures from the problem, added Annex III points 6(c) and 7(c) to the EU AI Act basis, corrected the grade of the Supreme Court evidence to B and its vendor, qualified the Amsterdam controls as planned, aligned the Blits.ai build with the feature inventory, added SEO title and description."},{"date":"2026-09-27","note":"Review fixes: recorded the Amsterdam tool as announced (its register entry says it still has to be developed), tied the Supreme Court claim to the source, removed an unsupported grounding guardrail claim from the Blits.ai build, dropped the inclusion value driver and reworded the SEO description."},{"date":"2026-09-27","note":"Review fixes: the Blits.ai build paragraph no longer claims the gateway PII masking reaches ingested case files, describes the output guardrail as blocking rather than flagging responses for review, and narrows role based access control to per bot roles instead of per case access."}],"slug":"court-and-case-file-summarization","url":"https://www.blits.ai/ai-use-cases/court-and-case-file-summarization","benchmarks":[],"indicativeValueResult":{"low":250000,"high":4050000},"evidence":["crown-prosecution-service-beam-notes-video-evidence-summaries","doj-eoir-immigration-filing-summaries","gemeente-amsterdam-objection-advice-summaries","supremo-tribunal-federal-maria-case-reports"]},{"title":"AI for creating employee training and eLearning content","shortTitle":"Training content creation","seoTitle":"AI for employee training content creation","metaDescription":"Generative AI drafts course outlines, quizzes, narration and avatar videos for staff training, as used by Zoom and the Veterans Benefits Administration.","definition":"Generative AI that helps learning and development teams turn source material such as procedures, product documentation and policies into training: course outlines, lesson text, quizzes, narration, avatar videos and translations, which instructional designers and subject matter experts review before publishing.","aliases":["AI course creation","AI eLearning authoring","AI training video generation","AI instructional design assistant","generative AI for learning and development"],"industries":["cross-industry","government","technology","manufacturing"],"functions":["human-resources","knowledge-management"],"patterns":["content-generation","translation","summarization"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Training content is expensive to make and quickly out of date. A single eLearning module takes an\ninstructional designer many hours of scripting, storyboarding, question writing and production,\nand subject matter experts lose days recording themselves: a senior instructional designer at Zoom\ndescribes experts and designers spending an entire day recording to get about 15 minutes of video. When the product, the\nprocedure or the regulation changes, the video has to be reshot, so outdated training stays in\ncirculation.\n\nNew systems, compliance topics and multilingual workforces all need material, often in several\nlanguages and in accessible formats: Carlsberg, for example, used to hire a second agency to\ntranslate each eLearning. Public bodies feel the pressure too. The Veterans Benefits\nAdministration uses an AI assistant to reduce instructor and instructional design burden amid\ndecreased hiring abilities, and to cut classroom time, and the IRS turned to AI voices when return\nto office mandates meant it could no longer record narration with employees.","problemStats":[],"howItWorks":"1. **Start from approved sources.** The designer uploads the procedure, product documentation or\n   policy the course must teach, and defines the audience and learning objectives.\n2. **Draft the structure.** The AI proposes an outline, lesson text and knowledge checks (quizzes\n   and scenarios) mapped to the objectives.\n3. **Produce media.** Narration is generated from the script with synthetic voices, and short\n   videos can use AI avatars instead of filmed presenters; images and sounds are generated or\n   selected.\n4. **Localise.** Text, narration and subtitles are translated into the languages each market\n   needs, as Carlsberg does for supply chain training.\n5. **Review and publish.** A subject matter expert checks accuracy, the designer assembles the\n   module in the authoring tool, and the course is published to the learning platform.\n6. **Update cheaply.** When the source changes, the script is edited and the media regenerated\n   instead of reshot.","valueDrivers":["employee-productivity","speed","cost-to-serve","inclusion-and-access"],"kpis":["processing-time-reduction","productivity-gain","hours-saved","cost-savings"],"indicativeValue":{"referenceOrg":"A learning and development team that builds or updates 200 eLearning modules a year","inputs":[{"key":"modules","label":"Modules built or substantially updated per year","low":200,"high":200,"unit":"modules per year","note":"The reference team. Replace with your own volumes."},{"key":"hoursPerModule","label":"Design and production hours per module today","low":40,"high":80,"unit":"hours per module","note":"Editorial assumption covering scripting, questions, media and assembly; replace with your own."},{"key":"timeSaved","label":"Share of those hours saved with AI drafting and media generation","low":0.2,"high":0.4,"unit":"fraction of hours","note":"Conservative against the vendor reported 90% time savings on video creation at Zoom on this page, because video is only part of a module and expert review time does not shrink."},{"key":"hourlyCost","label":"Blended instructional designer and expert hour","low":50,"high":90,"unit":"USD per hour","note":"Editorial assumption."}],"formula":"modules * hoursPerModule * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Course production time released","caveat":"Counts internal production time only. It leaves out translation savings, the value of training that is current instead of outdated, licence costs and the effect on learning outcomes. Synthesia reports at least €30,000 a year in avoided agency fees at Carlsberg, which the model does not include. Synthesia's reported $1,000 to $1,500 a month per employee saving at Zoom is a vendor estimate of internal production efficiency, which overlaps with what this model already counts, so it is not additional value. None of the deployments on this page has published effects on learning outcomes."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Authoring tools such as Articulate 360 and video tools such as Synthesia and TechSmith Camtasia already include these features, as the deployments on this page show, and the source material usually exists. The effort goes into review workflows, keeping content tied to approved sources, consent for any real person's likeness or voice, and accessibility.","dataPrerequisites":["Approved and current source material for each course","Learning objectives and assessment standards","Brand, tone and terminology guidelines, including approved translations","Written consent for any avatar or voice modelled on a real employee"],"integrations":["eLearning authoring tools","Learning management system","Document management or knowledge base holding the source material","Translation memory and terminology tools"]},"implementation":{"steps":[{"title":"Pick content that changes often","detail":"Start where reshooting and rewriting hurt most: product training, system training and procedures. Zoom uses it for sales enablement, where videos had to be recorded again whenever the training material changed."},{"title":"Ground drafts in approved sources","detail":"Generate from the procedure or documentation, not from the model's general knowledge, and keep a link from each lesson to its source so updates can be traced."},{"title":"Keep experts in the review, not the recording","detail":"Move subject matter experts from recording to reviewing scripts and quizzes. That is where their time is best spent and where errors are caught."},{"title":"Set rules for synthetic media","detail":"Decide which avatars and voices may be used, get written consent for any real person's likeness, label AI generated media, and check captions and transcripts for accessibility."},{"title":"Measure learning, not only production","detail":"Track completion, assessment results and learner feedback against earlier versions, so faster production does not come at the cost of learning."}],"guardrails":["Every course reviewed and approved by a named subject matter expert before publishing","Drafts generated from approved sources, with a link to the source version","AI generated video and audio labelled as such to learners","No avatar or cloned voice of a real person without written consent","Captions, transcripts and accessible formats checked before release"],"humanInTheLoop":"Instructional designers direct and assemble the course, subject matter experts approve the content, and learning owners sign off assessments, especially any that count towards certification or role decisions.","kpisToInstrument":["Hours from request to published module, before and after","Expert review hours per module","Errors found after publication per module","Assessment scores and completion rates versus earlier versions","Age of content (time since last update) across the catalogue"],"failureModes":[{"title":"Confident errors in the content","detail":"The model fills gaps with plausible but wrong details. Generate from sources only and require expert sign off."},{"title":"More content, not better learning","detail":"Production gets cheaper and the catalogue grows, but nobody checks whether people learn. Track assessment results and retire unused content."},{"title":"Consent and likeness problems","detail":"An employee's face or voice is reused after they leave or without clear agreement. Keep consent records and prefer stock avatars."},{"title":"Uncanny or disengaging media","detail":"Avatar videos that feel artificial lose learners' attention. Test with learners and mix formats."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Generating training content is not listed in Annex III. Providers of tools that generate synthetic audio, image, video or text content must mark the output as AI generated (with an exception for assistive editing that does not substantially alter the source), and deployers must disclose deep fakes, such as an avatar or voice that resembles a real person and would falsely appear authentic (Article 50(2) and (4), with the definition in Article 3(60)). If the same system evaluates learning outcomes or decides access to training that affects a person's work, Annex III point 3 (education and vocational training) and point 4 (employment) must be checked, and those parts can be high risk."},"regulations":["eu-ai-act","gdpr","iso-42001"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Machine readable marking of synthetic audio, image, video and text content by providers, and disclosure of deep fakes (image, audio and video) by deployers."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 3 covers AI that evaluates learning outcomes or decides access in education and vocational training; point 4 covers employment decisions."},{"title":"Section508.gov","issuer":"US General Services Administration","region":"north-america","url":"https://www.section508.gov/","note":"Accessibility requirements for information and communication technology at US federal agencies, which cover their training content; cited as a goal in the US Marshals Service pilot. Organizations elsewhere follow their own accessibility rules."}],"controls":["Content ownership and review dates for every course","Register of avatars and voices in use, with consent records","Labelling policy for AI generated media","Accessibility check in the publishing workflow"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai is not an eLearning authoring tool, but the drafting work behind a course can run on\nit. An **agentic workflow** takes the approved source documents from the **knowledge base**,\ndrafts an outline, lesson text and quiz questions against the stated objectives as **structured\noutput**, and waits for **human in the loop** approval by the subject matter expert before\nanything is exported, with a full audit trail per run. **Text to speech** across many providers,\nwith custom voices, produces narration audio, and **machine translation** and **multi language**\nsupport cover localisation.\n\nThe same knowledge base can power an **AI agent** that answers learners' follow up questions\nabout the course material in **Microsoft Teams** or web chat, or through a **digital human**\nwhere a face helps. The platform is **model agnostic**, and **test suites** with\nexpert written questions check that drafts and answers stay faithful to the source material."},"faq":[{"question":"How much faster is course production with AI?","answer":"For individual steps, much faster: Synthesia reports 90% time savings on training video creation by Zoom's instructional designers. For a whole module the saving is smaller, because experts still review the content. None of the deployments on this page has published effects on learning outcomes."},{"question":"Who uses it in the public sector?","answer":"The Veterans Benefits Administration uses an AI assistant in its authoring tool to build eLearning for claims processors, the IRS generates course narration with AI voices, and the US Marshals Service is piloting AI presenters and narration for training videos."},{"question":"Do we need to tell learners that a video uses an AI avatar?","answer":"Under the EU AI Act, providers must mark synthetic media as AI generated and deployers must disclose deep fakes, such as an avatar or voice that resembles a real person. Labelling all AI generated media is the simple policy."}],"related":["conversation-roleplay-training","employee-onboarding-assistant","support-knowledge-article-generation","audio-and-video-transcription-and-captioning"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched and written with evidence from the Veterans Benefits Administration, the IRS, the US Marshals Service, Zoom and Carlsberg. Editor review added the vendor figures for Carlsberg and Zoom, aligned the VBA wording with its inventory entry and corrected the value caveat. Second review credited the Carlsberg and Zoom cost figures to Synthesia, scoped the Zoom hours figure to what the source states, and scoped the tooling claim to the tools on this page."}],"slug":"training-content-generation","url":"https://www.blits.ai/ai-use-cases/training-content-generation","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"zoom-ai-video-sales-training","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":15,"min":15,"max":15,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"zoom-ai-video-sales-training","pooled":true}]},{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"zoom-ai-video-sales-training","pooled":false}]}],"indicativeValueResult":{"low":80000,"high":576000},"evidence":["carlsberg-ai-video-supply-chain-training","irs-ai-voiceover-elearning","us-marshals-service-ai-training-video-pilot","vba-articulate-ai-assistant-elearning","zoom-ai-video-sales-training"]},{"title":"AI for customs classification and declaration preparation","shortTitle":"Customs classification and declarations","seoTitle":"AI customs classification and HS code automation","metaDescription":"AI proposes HS codes and drafts customs declarations for a broker to check. UPS says it cleared 90% of dutiable US parcels without manual work in September 2025.","definition":"AI that reads what is being shipped (the commercial invoice, the product data and sometimes a photo), proposes the tariff classification code with its reasoning and a confidence score, drafts the customs declaration with value, origin and parties, and sends only uncertain or high risk entries to a licensed customs specialist before filing.","aliases":["AI HS code classification","automated tariff classification","AI customs brokerage","customs declaration automation","AI customs entry preparation"],"industries":["logistics-and-transportation","retail-and-ecommerce","cross-industry"],"functions":["operations","regulatory-compliance"],"patterns":["classification-and-routing","document-processing","agentic-workflow","computer-vision"],"channels":["api","internal-tools","email"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Every item that crosses a border needs a tariff classification code. The first six digits come\nfrom the World Customs Organization's Harmonized System; countries then add their own digits, and\none wrong digit can change the duty rate, trigger a licence requirement or hold the shipment. The\nrules for choosing a code are precise, national tariff nomenclatures change every year and the\nHarmonized System itself every five years, but the input is usually a vague commercial\ndescription (\"parts\", \"gift\", \"cotton top\") typed by a shipper who is not a customs expert.\n\nCustoms brokers, carriers and cross border retailers have handled this with specialist teams\nthat look up codes, retype invoice data into the declaration system and chase shippers for\nmissing facts. That model breaks when volumes jump. After the United States ended its de minimis\nexemption in 2025, low value parcels that had entered with minimal data suddenly needed a formal,\ndutiable clearance, and UPS reported a tenfold increase in daily customs entries. Errors are costly in both directions: an underpaid duty can be reassessed years\nlater with penalties, and an incomplete declaration holds the parcel at the border while the\ncustomer waits.","problemStats":[{"statement":"UPS told investors that after the removal of the de minimis exemption for US imports it saw a tenfold increase in daily customs entries.","sourceTitle":"United Parcel Service Inc (UPS) Q3 2025 Earnings Call Transcript","sourceUrl":"https://equibles.com/stocks/ups/calls/2025-q3","year":2025},{"statement":"Swiss Post, citing an analysis by Deloitte, puts the average error rate of manual customs classification at 10 to 30 percent, depending on product complexity.","sourceTitle":"AI-based customs tariff classification for international e-commerce","sourceUrl":"https://international.post.ch/en/blog/ai-based-tariff-classification-for-e-commerce","year":2026}],"howItWorks":"1. **Collect the facts.** The system reads the commercial invoice, packing list and product\n   master data, or the shipper's own description at booking. Where the shipper is a consumer,\n   some carriers ask for a photo and propose the description from it, as DHL Express does.\n2. **Propose a code.** A model trained on past declarations and binding rulings, combined with\n   retrieval over the tariff nomenclature and explanatory notes, proposes the most likely code\n   for the destination country, with the reasoning and a confidence score.\n3. **Validate the entry.** Rules check the code against the declared material and use, the\n   value against the invoice, the country of origin, licence and sanctions requirements, and any\n   additional duties that apply to that code and origin.\n4. **Draft the declaration.** The system fills the declaration or entry in the format of the\n   customs system, with duties and taxes calculated.\n5. **Route by confidence.** High confidence, low risk entries go to filing; low confidence\n   codes, high value goods, controlled items and missing data go to a licensed specialist, who\n   can also ask the shipper for more information.\n6. **Learn from corrections.** Specialist corrections and customs queries flow back into the\n   reference data, so the same product is classified the same way next time.","valueDrivers":["cost-to-serve","speed","compliance","customer-experience"],"kpis":["automation-rate","handling-time-reduction","processing-time-reduction","cost-reduction","accuracy","error-reduction","interactions-handled"],"indicativeValue":{"referenceOrg":"A customs broker or cross border retailer filing 200,000 declarations a year","inputs":[{"key":"declarations","label":"Declarations or entries per year","low":200000,"high":200000,"unit":"declarations per year","note":"The reference organization."},{"key":"minutesPerDeclaration","label":"Specialist minutes per declaration today","low":10,"high":30,"unit":"minutes per declaration","note":"Editorial assumption covering classification lookups and data entry for a mixed parcel and freight workload. Swiss Post puts manual classification at 15 to 45 minutes per item; repeat items with a known code take far less, hence the lower range. Replace with your own time study."},{"key":"timeSavedShare","label":"Share of specialist time the AI removes","low":0.3,"high":0.6,"unit":"fraction of time","note":"Conservative against the benchmarks on this page (Digicust reports a 90% reduction in processing time per clearance at ZLS; UPS says its brokerage technology cleared 90% of dutiable parcels without manual intervention in September 2025), because the specialist still reviews the uncertain entries."},{"key":"costPerHour","label":"Fully loaded cost of a customs specialist","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption, replace with your own cost."}],"formula":"declarations * minutesPerDeclaration / 60 * timeSavedShare * costPerHour","currency":"USD","period":"per year","resultLabel":"Specialist time released","caveat":"Values only the specialist time released. It leaves out the cost of the platform and the integration, the value of fewer border holds and faster delivery, and the avoided cost of duty reassessments and penalties, which can be larger but is hard to predict."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Proposing a code from a good description is well understood. The work is in the data (clean product master data, past declarations with the final code), the integration with the customs filing system and the broker's workflow, and in drawing the line between what may be filed automatically and what a licensed person must review.","dataPrerequisites":["Historical declarations with the final, accepted classification codes","Product master data with materials, use and composition","Current tariff nomenclature, explanatory notes and binding rulings for each destination","Lists of controlled, licensed and sanctioned goods and additional duty measures"],"integrations":["Customs filing system or national single window (for example ICS2 in the EU or ACE in the US)","Order, booking or transport management system that supplies shipment data","Product information or ERP system with the item master","Duty and tax calculation engine","Case tool for specialist review and shipper queries"]},"implementation":{"steps":[{"title":"Measure the baseline","detail":"Take a sample of recent entries and record the time per entry, the share with a later correction or customs query, and where the delays sit (classification, missing data, retyping). This is the benchmark the AI has to beat."},{"title":"Build the reference set","detail":"Assemble past declarations with the final accepted code, the binding rulings you hold and the product master. Remove codes that customs later corrected, or the model learns the mistakes."},{"title":"Run in shadow mode","detail":"Let the AI propose codes and draft entries next to the specialists for several weeks. Compare at the level that matters (six digits and the full national code) per product family, and set the confidence threshold per family, not one number for everything."},{"title":"Automate the confident, low risk share","detail":"File automatically only where the confidence is above the threshold and the goods are not controlled, high value or subject to additional duties. Everything else goes to a specialist with the AI's reasoning attached."},{"title":"Close the loop with shippers","detail":"Ask for missing facts at the moment of booking, in plain language, instead of after the parcel is held. A photo or a guided question is cheaper than a hold."},{"title":"Monitor and recalibrate","detail":"Track corrections, customs queries and post clearance audits per code and retrain before each annual nomenclature change."}],"guardrails":["Automatic filing only above a confidence threshold set per product family, and never for controlled, sanctioned or licensed goods","Every proposed code carries its reasoning and the rule or ruling it relies on, kept with the entry for audit","The licensed broker or declarant remains responsible and signs off the rules and thresholds","Declared value and origin are checked against the invoice and never changed by the model","Personal data of consignees is limited to what the declaration requires and masked in logs"],"humanInTheLoop":"Licensed customs specialists review every low confidence code, every controlled or high value item and every entry customs queries. They own the classification rules, approve threshold changes and sample automatically filed entries every week, because a classification error can repeat across thousands of shipments before an audit finds it.","kpisToInstrument":["Share of entries filed without manual intervention","Classification accuracy at six digits and at the full national code, on a weekly checked sample","Customs holds, queries and post clearance corrections per thousand entries","Specialist minutes per entry and time from arrival to release","Duty reassessments and penalties over time"],"failureModes":[{"title":"Confidently wrong at scale","detail":"A systematic error on one product family repeats on every shipment until an audit finds it. Sample automatically filed entries and watch corrections per code."},{"title":"Garbage descriptions in, garbage codes out","detail":"The model cannot classify \"gift\" or \"parts\" reliably. Push back to the shipper for facts instead of guessing, and treat vague descriptions as low confidence."},{"title":"Stale tariff data","detail":"Nomenclature changes and new additional duties make yesterday's correct code wrong. Update the reference data on the effective date and rerun affected products."},{"title":"Automation that hides accountability","detail":"The declarant is liable whatever the tool did. Keep the reasoning, the data used and the approver with every entry."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The classification and declaration work in the back office is minimal risk: classifying goods and preparing customs declarations is not listed in Annex III, which covers border control only where AI assesses natural persons (point 7). If the design adds a shipper facing assistant that asks for missing information in chat or by email, that assistant is limited risk and carries the transparency duty of Article 50 (people must know they are dealing with AI). The obligations that matter most come from customs law: the declarant stays responsible for the accuracy of the declaration whatever tool prepared it."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"How Artificial Intelligence (AI) can help Customs in automating HS Classification","issuer":"World Customs Organization","region":"global","url":"https://www.wcoomd.org/en/media/newsroom/2022/april/how-ai-can-help-customs-in-automating-hs-classification.aspx","note":"Describes the WCO BACUDA model that recommends HS codes from commercial descriptions with probabilities, positioned as support for customs officials, with a course and demonstration tool for member administrations."},{"title":"What Every Member of the Trade Community Should Know About: Tariff Classification","issuer":"U.S. Customs and Border Protection","region":"north-america","url":"https://www.cbp.gov/document/publications/tariff-classification","note":"Informed compliance publication on how goods are classified and on the importer's duty of reasonable care, which applies whether a person or a system proposed the code."}],"controls":["Documented classification rules and confidence thresholds per product family, approved by a licensed specialist","Audit trail per entry with the proposed code, reasoning, data used and approver","Weekly sampling of automatically filed entries and monitoring of customs queries and corrections","Change control on reference data around each nomenclature update","Controlled goods and sanctions checks that the AI cannot bypass"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that takes the shipment data through the REST API,\nuses an **AI agent** with **structured output** to propose the classification code, its\nreasoning and a confidence score, and grounds that proposal in a **knowledge base** holding the\ntariff nomenclature, explanatory notes and the broker's own rulings, retrieved with hybrid\nsearch. Past declarations and the product master sit in a **SQL knowledge base** the agent can\nquery, and **custom functions** call the duty engine and the filing system.\n\nThe workflow routes entries with a low confidence score or controlled goods to a filing step\nthat waits for **human in the loop confirmation**, where a specialist approves or rejects it,\nand every run keeps a full **audit trail**. Where a shipper has to supply missing facts, a\nseparate agent on **web chat, WhatsApp or email** can collect them. **Test suites** can run a\nlabelled set of past entries against the workflow before a change goes live, the platform is\n**model agnostic**, and **EU and UAE data residency** keeps declaration data in region."},"faq":[{"question":"Can AI classify goods for customs without a human?","answer":"For standard goods with good descriptions, often yes, within thresholds a licensed specialist sets. UPS says its brokerage technology, which it enhanced with agentic AI, cleared 90% of US parcels that needed a dutiable clearance without manual intervention in September 2025 (about 21% in March 2025). That figure covers the whole entry, not classification alone, the other 10% needed people, and the declarant stays legally responsible for every code."},{"question":"How much time does AI save a customs broker?","answer":"It depends on how much of the work is retyping and looking up codes. Digicust reports that processing time per extensive clearance at the German customs agent ZLS fell from three to four hours to 10 to 15 minutes. Plan more conservatively for mixed or complex goods."},{"question":"Does it help consumers and small shippers too?","answer":"Yes, at the moment of booking. DHL Express lets shippers photograph an item and proposes a customs compliant description that they can edit, live in eight markets since May 2026, to avoid holds caused by vague descriptions."}],"related":["intelligent-document-processing","parcel-tracking-and-delivery-exception-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version with UPS, DHL Express and ZLS deployments, researched and checked against the sources. Editor pass corrected the ZLS year and cost qualifier, dropped an unsourced staffing claim and scoped the EU AI Act tier."}],"slug":"customs-classification-and-declaration","url":"https://www.blits.ai/ai-use-cases/customs-classification-and-declaration","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ups-agentic-ai-customs-brokerage","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"zls-digicust-customs-declaration-automation","pooled":true}]},{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"zls-digicust-customs-declaration-automation","pooled":false}]}],"indicativeValueResult":{"low":400000,"high":4200000},"evidence":["dhl-express-ai-item-identification-customs","ups-agentic-ai-customs-brokerage","zls-digicust-customs-declaration-automation"]},{"title":"AI for drafting customer letters and outbound notices","shortTitle":"Outbound notice drafting","seoTitle":"AI drafting of customer letters and notices","metaDescription":"AI drafts complaint replies, arrears notices and claim letters from case data and approved clauses for staff to approve. Evidence from SS&C GIDS and Acentra Health.","definition":"AI that drafts the letters and notices operations must send at scale, such as arrears notices, decline letters, complaint responses, servicing confirmations and product change notices, from case data and approved templates and clauses, in the customer's language, for a person to approve where the notice is regulated.","aliases":["AI letter drafting","letter and notice generation","generative AI customer correspondence","notice generation","claims correspondence drafting","settlement and claim decision letters","appeal determination letters"],"industries":["cross-industry","banking","insurance","government","healthcare","wealth-and-asset-management"],"functions":["operations","customer-service","collections-and-recovery","regulatory-compliance","claims"],"patterns":["content-generation","rag-knowledge-assistant","translation"],"channels":["email","internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"back-office","problem":"Banks, insurers and public bodies send large volumes of letters. The high volume ones come from fixed\ntemplates, but a long tail does not fit a template cleanly: a complaint response that has to\naddress the customer's specific points, an arrears letter that must reflect an agreed payment\nplan, a decline letter with the right reasons, a notice in the customer's own language. Staff\nwrite these by hand, copying clauses from a library and data from several systems, which is slow\nand produces uneven quality.\n\nThe quality matters because many of these letters are regulated. In the UK, a final response to\na complaint must meet content and timing rules, in the US an adverse action notice on a credit\napplication must state the specific reasons, or tell the applicant they can get them within 30\ndays, and the FCA Consumer Duty expects communications that customers are likely to understand.\nA free writing model is not acceptable here; the safe gain comes from drafting inside approved\nwording.","problemStats":[],"howItWorks":"1. **Start from the case.** The trigger is a case event (complaint investigated, arrears stage\n   reached, application declined, product changed) with its structured data.\n2. **Select the approved template.** Rules, not the model, choose the template and the mandatory\n   clauses for the notice type and jurisdiction.\n3. **Draft the variable parts.** The AI writes the case specific paragraphs (the summary of the\n   complaint and findings, the payment plan terms, the plain language explanation) using only\n   facts from the case and wording retrieved from the approved clause library.\n4. **Check before a human sees it.** Automated checks compare every figure and date in the draft\n   with the case data, confirm mandatory clauses are present, and score readability.\n5. **Approve, send and store.** A person approves regulated notice types; approved letters are\n   sent through the customer's preferred channel and stored with their version and data.","valueDrivers":["employee-productivity","compliance","customer-experience","speed"],"kpis":["time-saved-per-task","handling-time-reduction","processing-time-reduction","productivity-gain","error-reduction","hours-saved"],"indicativeValue":{"referenceOrg":"A bank or insurer whose staff write 150,000 non standard letters a year","inputs":[{"key":"letters","label":"Letters written or heavily edited by hand per year","low":150000,"high":150000,"unit":"letters per year","note":"The reference organization. Count only letters that do not go out from a fixed template."},{"key":"minutesSaved","label":"Minutes saved per letter","low":3,"high":15,"unit":"minutes per letter","note":"The low end is Acentra Health's reported fall from about six to three minutes per appeal letter; the high end is an editorial assumption for complaint responses written from scratch. Replace with a time study.","sourceUrl":"https://www.microsoft.com/en/customers/story/19280-acentra-health-azure"},{"key":"costPerHour","label":"Fully loaded cost per hour of the writing staff","low":35,"high":60,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"letters * minutesSaved / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Drafting effort avoided","caveat":"Drafting time only. It leaves out the review time that remains, the value of fewer complaint escalations and ombudsman referrals, and the cost of the platform and template work."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The model work is modest. Most effort goes into a clean, approved clause library, the mapping of case data into drafts, and agreeing with compliance which notice types need human approval.","dataPrerequisites":["Approved templates and clause library per notice type, language and jurisdiction","Structured case data for each trigger (complaint findings, arrears status, decision reasons)","A sample of good historical letters per type as a quality reference"],"integrations":["Case management and complaint systems","Collections, lending and servicing systems for case data","Customer communications management platform for layout and dispatch","Document archive for versioned storage"]},"implementation":{"steps":[{"title":"Pick the notice types that are hand written today","detail":"Fully templated letters need no AI. Start where staff write free text, such as complaint responses and bespoke arrears letters."},{"title":"Clean up the clause library","detail":"Give every clause an owner, a version and a legal approval date, and remove duplicates. The model can only be as safe as the wording it is allowed to use."},{"title":"Draft next to people first","detail":"Let staff start from the AI draft and measure edit distance and time per letter by type before changing any approval rule."},{"title":"Automate the checks","detail":"Build deterministic checks for figures, dates, names and mandatory clauses so reviewers focus on tone and judgment instead of proofreading."},{"title":"Decide approval by notice type","detail":"Keep human approval on regulated and adverse notices, and consider sampling instead of full review only for low risk confirmations with a stable quality record."}],"guardrails":["Template and mandatory clauses are chosen by rules, never by the model","Every figure, date and name in the draft is checked against the case data before review","Regulated and adverse notices always have a named human approver","Every sent letter is stored with its version, template and source data"],"humanInTheLoop":"Case handlers approve every regulated or adverse notice and edit drafts where needed. Compliance owns the clause library and samples approved letters monthly, including those in other languages.","kpisToInstrument":["Minutes per letter from draft to approval, by notice type","Share of drafts approved without material edits","Factual errors caught by the automated checks and by reviewers","Readability score of sent letters","Complaints about letters and repeat contacts after a notice"],"failureModes":[{"title":"A plausible but wrong fact","detail":"The draft states an amount or date that is not in the case. Block sending until the automated fact check passes."},{"title":"Reviewers stop reading","detail":"After weeks of good drafts, approval turns into a click. Track review time and seed test drafts with known errors."},{"title":"Tone that fails vulnerable customers","detail":"A correct letter that is cold or confusing. Include vulnerability flags in the case data and test drafts with plain language checks."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Drafting letters for human approval is not listed in Annex III. The decision the letter communicates may come from a separate high risk system, such as credit scoring (Annex III point 5(b)) or a public body's eligibility decision on benefits (point 5(a)); the drafting tool does not make that decision. Article 50(2) requires the provider of an AI system that generates text to mark the output as artificially generated, which puts this on the limited risk (transparency) tier; this includes an organization that builds its own drafting tool. Article 50(2) does not apply where the AI has only an assistive function for standard editing and does not substantially alter the input data or the semantics of the output."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","us-ecoa-reg-b","hipaa","iso-42001"],"guidance":[{"title":"PRIN 2A.5 Consumer Duty, consumer understanding outcome","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.handbook.fca.org.uk/handbook/PRIN/2A/5.html","note":"Communications must be likely to be understood by the customers they are aimed at, which applies to AI drafted letters too."},{"title":"DISP 1.6 Complaints time limit rules","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.handbook.fca.org.uk/handbook/DISP/1/6.html","note":"Sets the deadline for a final response (eight weeks for most complaints, 15 business days for payment services and e money complaints) and what it must contain, including the Financial Ombudsman Service referral rights. A drafted complaint response must meet these rules."},{"title":"Regulation B, 12 CFR 1002.9 Notifications","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/regulations/1002/9/","note":"Adverse action notices on credit applications must state the specific reasons, or tell the applicant they can get them within 30 days; a drafting tool must use the reasons the credit decision produced."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Article 50(2) requires providers of systems that generate text to mark the output as artificially generated."}],"controls":["Clause library under change control with legal approval dates","Automated fact and clause checks logged for every draft","Named approver recorded for every regulated notice","Versioned storage of every sent letter for the statutory retention period"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered by a case event through an API token. An\n**AI agent** retrieves approved clauses from the **knowledge base** with hybrid retrieval and\ndrafts the variable parts with **structured output**. **Custom functions** fetch the case data,\nchoose the template and the mandatory clauses by rule, and run the fact checks.\n\n**Human in the loop confirmation** holds the send step for regulated notice types until a case\nhandler approves it, and approved letters go out through the **email channel** or, through a\ncustom function, the organization's own communications platform.\n**Guardrails** check output against policy, **prompt versioning** keeps a record of what\nchanged, and **test suites** grade drafts against reference letters before any change. Multi\nlanguage support and machine translation cover notices in the customer's language."},"faq":[{"question":"Is it safe to let generative AI write regulated customer letters?","answer":"Only inside approved wording and with a human approver for regulated and adverse notices. The template and mandatory clauses are chosen by rules, the model drafts only the case specific paragraphs from case data, and automated checks compare every figure and date with the case before a person reviews it."},{"question":"Where does AI drafting save the most time?","answer":"In the letters staff write by hand today, such as complaint outcome letters and bespoke customer letters. SS&C Blue Prism reports that SS&C GIDS produces customer letters with an in house language model three times faster than with the manual process, and that complaint cycle times fell by 25% after AI agents took over steps including drafting the closing letter for an employee to check."},{"question":"Does this work for insurance claims letters?","answer":"Yes, claim updates and decision letters follow the same pattern. Microsoft reports that a Hiscox claims underwriter uses Microsoft 365 Copilot to pull the progress of a claim from several emails and compose an update to a broker or customer. In Medicare appeals, Microsoft reports that Acentra Health cut nurse time per appeal determination letter by approximately 50% with its MedScribe drafting tool."},{"question":"Does the EU AI Act make letter drafting high risk?","answer":"No, it is limited risk. Drafting letters is not listed in Annex III. The decision the letter communicates, such as a credit decline, may come from a high risk system, which is governed separately. The provider of the drafting system must mark generated text under Article 50(2), unless the AI only performs an assistive function for standard editing."}],"related":["correspondence-triage-and-routing","complaints-handling-agent","adverse-action-explanations","civil-servant-drafting-copilot","health-prior-authorization-and-claims-adjudication"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Insurance vertical added claims correspondence and settlement letters to the scope, with evidence from Acentra Health, Progressive and Hiscox, and added healthcare to the industries."},{"date":"2026-09-25","note":"Consolidation pass: industries now include wealth and asset management, where its evidence comes from; added ECOA and Regulation B to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed the unsourced letter volume, tied the regulatory statements to DISP 1.6, Regulation B and the Consumer Duty (two guidance entries added), attributed the SS&C figures to SS&C Blue Prism, replaced the Progressive example with Hiscox, lowered the minutes saved per letter to match Acentra Health, added UK GDPR and HIPAA, and added the SEO title and description."}],"slug":"outbound-notice-drafting","url":"https://www.blits.ai/ai-use-cases/outbound-notice-drafting","benchmarks":[{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":5638,"min":276,"max":11000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"acentra-health-medscribe-appeal-letters","pooled":true},{"id":"hrsa-ai-audit-resolution-assistant","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":25,"min":25,"max":25,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"ssc-gids-complaint-closing-letters","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"acentra-health-medscribe-appeal-letters","pooled":true}]}],"indicativeValueResult":{"low":262500,"high":2250000},"evidence":["acentra-health-medscribe-appeal-letters","hiscox-copilot-claims-handling","hrsa-ai-audit-resolution-assistant","ssc-gids-complaint-closing-letters","ssc-gids-llm-customer-letters"]},{"title":"AI for eDiscovery and disclosure document review","shortTitle":"eDiscovery document review","seoTitle":"AI for eDiscovery and disclosure document review","metaDescription":"AI ranks and codes millions of documents for relevance and privilege so lawyers review fewer. The UK Serious Fraud Office and the US FTC use it in investigations.","definition":"AI that sorts, prioritises and codes large collections of emails, chats and files for relevance, issues and legal privilege in litigation, investigations and regulatory requests, so that lawyers review the documents most likely to matter and can show the court how the rest were handled.","aliases":["technology assisted review","TAR","predictive coding","AI document review","generative AI eDiscovery","privilege review AI","AI disclosure review"],"industries":["cross-industry","professional-services","government"],"functions":["legal","case-management"],"patterns":["classification-and-routing","document-processing","summarization"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"mainstream","problem":"Every lawsuit, investigation and regulatory information request starts with a collection of\nelectronic material that nobody can read in full. A single civil matter can involve more than\n300,000 documents, and some Serious Fraud Office cases start with up to 60 million. Lawyers\nstill have to find what is relevant, what supports or undermines each side, and what is\nprivileged, under deadlines set by a court or a regulator.\n\nKeyword searches were the first answer, and they are weak: they miss documents that use other\nwords, return large volumes of noise, and give a false sense of completeness. Linear review by\ncontract lawyers is slow and expensive, and the US Department of Justice lists errors and speed\ndelays in exclusively manual review of voluminous electronic information as the problem its\neLitigation tools address. In criminal\ncases the stakes are higher still. The UK Serious Fraud Office offered no evidence in its G4S case\nafter ten years, with disclosure featuring as a core reason, and its inspectorate found that a\nmisunderstanding of how searches worked in its older review system compounded the problems.","problemStats":[{"statement":"HM Crown Prosecution Service Inspectorate reports that the Serious Fraud Office roughly estimates that managing and handling disclosure takes 25% of its operational budget and 40% of its staff capacity.","sourceTitle":"Serious Fraud Office: Disclosure. An inspection of the handling and management of disclosure in the Serious Fraud Office","sourceUrl":"https://cdn.websitebuilder.service.justice.gov.uk/uploads/sites/24/2024/08/SFO-Disclosure-Report-2.pdf","year":2024},{"statement":"HM Crown Prosecution Service Inspectorate reports that some Serious Fraud Office cases start with up to 60 million documents.","sourceTitle":"Serious Fraud Office: Disclosure. An inspection of the handling and management of disclosure in the Serious Fraud Office","sourceUrl":"https://cdn.websitebuilder.service.justice.gov.uk/uploads/sites/24/2024/08/SFO-Disclosure-Report-2.pdf","year":2024}],"howItWorks":"1. **Collect and process.** Material from mailboxes, devices, chat tools and file shares is\n   loaded into a review platform, text is extracted, duplicates and near duplicates are removed\n   and email threads are grouped.\n2. **Explore early.** Clustering, timelines and concept search show what the collection is\n   about, so the team can exclude irrelevant data and agree search parameters before paid review\n   starts.\n3. **Rank or code.** Either a classifier learns from reviewers' decisions and keeps ranking the\n   remaining documents by likely relevance (active learning), or a large language model reads each\n   document against written review instructions and returns a relevance call, the issues it\n   touches and a short rationale.\n4. **Screen for privilege and sensitivity.** A separate pass flags likely privileged documents,\n   personal data and material that needs redaction, for lawyers to confirm.\n5. **Validate.** Random samples from the documents classed as relevant and not relevant are\n   reviewed by people to estimate precision, recall and elusion, and the method and results are\n   recorded so they can be explained to the other side or the court.\n6. **Review and produce.** Lawyers review the prioritised set, decide what to disclose, withhold\n   or redact, and the platform keeps the audit trail.","valueDrivers":["cost-to-serve","speed","employee-productivity","compliance","risk-reduction"],"kpis":["processing-time-reduction","hours-saved","cost-savings","interactions-handled"],"indicativeValue":{"referenceOrg":"A litigation or investigations team that reviews 1 million documents a year","inputs":[{"key":"documents","label":"Documents collected for review per year after deduplication","low":1000000,"high":1000000,"unit":"documents per year","note":"The reference team. Replace with your own review volumes."},{"key":"hoursPerThousand","label":"Reviewer hours per 1,000 documents in a linear first pass review","low":15,"high":25,"unit":"hours per 1,000 documents","note":"Editorial assumption, replace with your own review rates."},{"key":"reviewAvoided","label":"Share of first pass review hours avoided by prioritisation or AI coding","low":0.3,"high":0.6,"unit":"fraction of review hours","note":"Conservative against the vendor reported 85% reduction in review time for one Purpose Legal matter on this page, because validation sampling, privilege review and quality control still need people."},{"key":"hourlyCost","label":"Fully loaded cost of a reviewer hour","low":50,"high":100,"unit":"USD per hour","note":"Editorial assumption covering contract reviewers and supervising associates."}],"formula":"documents / 1000 * hoursPerThousand * reviewAvoided * hourlyCost","currency":"USD","period":"per year","resultLabel":"First pass review cost avoided","caveat":"Covers first pass review labour only. It leaves out platform and model costs, processing and hosting fees, the cost of validation and of any dispute about the method, and the value of meeting deadlines that manual review would miss."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The technology is mature and widely available inside review platforms. The work is in the review protocol, the validation method, privilege handling and agreeing the approach with the other side, the court or the regulator, and in training case teams to use it properly.","dataPrerequisites":["Processed collections with extracted text, metadata and deduplication","Written review instructions per issue, with examples of relevant and not relevant documents","A seed or control set coded by lawyers who know the case","A sampling plan for validation with target recall agreed in advance"],"integrations":["eDiscovery or review platform (processing, review, production)","Legal hold and collection tools for mailboxes, devices and chat","Matter management and privilege log tooling","Secure hosting in the jurisdiction the data must stay in"]},"implementation":{"steps":[{"title":"Agree the protocol before the model","detail":"Write the review questions, issue definitions and privilege rules first, and decide how success will be measured (recall target, sample sizes). Where the other side or a court will scrutinise the process, share the approach early rather than defend it later."},{"title":"Pilot on a coded sample","detail":"Run the classifier or the language model on a few hundred documents that lawyers have already coded, compare, and refine the instructions until the disagreements are understood. Relativity reports that Purpose Legal reached a workable prompt for ten issues after three iterations on a sample of fewer than 500 documents."},{"title":"Run, rank and route","detail":"Apply the model to the full population, send the highest ranked documents to human review first, and keep a separate privilege and personal data pass."},{"title":"Validate with statistics, not impressions","detail":"Draw random samples from both the relevant and the not relevant sets, have people review them blind, estimate recall and elusion, and document every step so it can be explained."},{"title":"Train the case team","detail":"Make sure the people running the review understand what the tool does and does not do. The inspectorate warned of a risk that some SFO staff are not confident using the new platform, found that a number were not using it to its full potential and that many saw the training as inadequate, and found that misunderstandings about search had contributed to earlier failures."}],"guardrails":["No document is withheld as privileged or produced without a lawyer's decision","Validation sampling with a recall estimate before any review is declared complete","Written record of search terms, model instructions, versions and sampling results","Separate handling of privileged and personal data, with redaction checked by a person","Data stays in the hosting region and is not used to train a vendor's general models"],"humanInTheLoop":"Lawyers write the review instructions, code the training or validation samples, decide every privilege call and every production, and sign off the validation results. The AI decides only the order and the first view of relevance.","kpisToInstrument":["Estimated recall and elusion rate from validation samples, per matter","Reviewer hours per 1,000 documents, before and after","Share of the collection reviewed by people","Days from collection to production","Privilege clawback requests and disclosure errors found later"],"failureModes":[{"title":"False confidence in completeness","detail":"Teams treat search or model output as if it found everything. HMCPSI warned that searching millions of documents is not an exact science. Measure recall and say what it is."},{"title":"Processing gaps upstream","detail":"Documents that were never extracted or indexed properly cannot be found by any model. Check processing exceptions, container files and encoding before trusting results."},{"title":"Instructions that drift from the case","detail":"The issues change as the case develops but the model instructions do not. Version the instructions and rerun validation when they change."},{"title":"Unexplainable method","detail":"The team cannot describe how documents were excluded when challenged. Keep the protocol, versions and samples as part of the matter record."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Document review for a party in civil litigation or an internal investigation is not listed in Annex III, so it is usually minimal risk. It becomes high risk where a law enforcement authority uses AI to evaluate the reliability of evidence in the investigation or prosecution of criminal offences (Annex III point 6(c)), or where a judicial authority uses it to research and interpret facts and law (point 8(a)). Prosecutors and investigators should classify each use against those points."},"regulations":["eu-ai-act","gdpr","uk-gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Attorney General's Guidelines on Disclosure (2024)","issuer":"UK Attorney General's Office","region":"europe","url":"https://www.gov.uk/government/publications/attorney-generals-guidelines-on-disclosure","note":"Where examining every item of seized material would be disproportionate, disclosure officers can apply search techniques under Annex A, and must record the reasons for their approach in writing."},{"title":"Disclosure Review Working Group considering simplification of Practice Direction 57AD","issuer":"Courts and Tribunals Judiciary (England and Wales)","region":"europe","url":"https://www.judiciary.uk/disclosure-review-working-group-considering-simplification-of-practice-direction-57ad/","note":"A judiciary led group is reviewing civil disclosure rules in the Business and Property Courts, including the use of technology assisted review and AI."},{"title":"Serious Fraud Office: Disclosure, an inspection report","issuer":"HM Crown Prosecution Service Inspectorate","region":"europe","url":"https://cdn.websitebuilder.service.justice.gov.uk/uploads/sites/24/2024/08/SFO-Disclosure-Report-2.pdf","note":"Describes how a prosecutor uses an AI enabled review platform, the limits of search, and the need to train staff and quality assure review."}],"controls":["Documented review protocol and validation plan for every matter","Named lawyer accountable for the review method and its explanation","Audit trail of model versions, instructions, coding decisions and samples","Data processing agreements and hosting location that match the data's origin","Quality assurance batches reviewed by a second person"],"incidents":[{"title":"SFO drops decade-long probe and prosecution and announces further issue with legacy disclosure system","url":"https://www.lawgazette.co.uk/news/sfo-investigates-troubling-disclosure-issues/5125889.article","note":"The Law Gazette reports that the Serious Fraud Office found a further problem in its legacy Autonomy review system, in how some digital container files were expanded, so that some items may not have been available for review in about 20 long running cases; an earlier review covered 66 historical conviction cases. Not an AI model error, but a reminder that review results are only as complete as the processing underneath them."}]},"blitsAi":{"howToBuild":"Blits.ai is not a review platform, so the review itself stays in the organization's eDiscovery\ntool. What can be built on Blits.ai is the layer around it: an **agentic workflow** that takes a\nbatch of extracted document text through the **REST API**, asks an **agent** with a versioned\nprompt and **structured output** for a relevance call, the issues touched and a rationale per\ndocument, and writes the result back through **custom functions**, with a full **audit trail**\nper run and **human in the loop** approval before any bulk coding is applied.\n\n**PII masking** at the gateway limits the personal data the model sees, **EU or UAE data\nresidency** keeps the data in region, and the platform is **model agnostic**, so the team can compare models on the same instructions.\n**Test suites** with lawyer coded documents act as a regression set when the instructions change,\nand a **knowledge base** with the review protocol and issue definitions lets reviewers ask\nquestions about the matter's rules while they work."},"faq":[{"question":"Can we rely on AI review when the other side or the court will scrutinise it?","answer":"Only with a method you can explain. In England and Wales the Attorney General's disclosure guidelines let investigators use search techniques when reviewing everything would be disproportionate, provided they record their reasons, and the judiciary is reviewing civil disclosure rules with technology assisted review and AI in mind. Agreed instructions, statistical validation and a written record are what make the result defensible."},{"question":"How much review time does it save?","answer":"It depends on the collection and the recall target. Relativity reports that Purpose Legal cut project time by 85% (4,000 review hours) on one 300,000 document matter, measured against a contract review that would have taken multiple weeks and was never run. The UK Serious Fraud Office used its pilot tool to screen about 30 million documents for privilege in the Rolls-Royce case, work that independent barristers had done by hand. Plan on people still reviewing the prioritised set, the privilege calls and the validation samples."},{"question":"What is the difference between active learning and generative AI review?","answer":"Active learning trains a classifier on reviewers' decisions and keeps reranking the rest; the FTC uses Relativity Active Learning to predict pertinent documents. Generative AI review has a language model read each document against written instructions and explain its call. Both need the same validation."},{"question":"Is AI document review high risk under the EU AI Act?","answer":"For a company or law firm reviewing documents in civil litigation, usually not. It is high risk when law enforcement uses it to evaluate the reliability of evidence in criminal cases, or when a judicial authority uses it to research and apply the law."}],"related":["court-and-case-file-summarization","freedom-of-information-request-processing","intelligent-document-processing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the discovery workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched and written with evidence from the UK Serious Fraud Office and its inspectorate, the US Department of Justice, the FTC and Purpose Legal."},{"date":"2026-09-27","note":"Editorial review fixes, with the SFO throughput capability claim replaced by its Rolls-Royce screening figure, Purpose Legal figures attributed to Relativity and set against their counterfactual, and the problem statement sourced."}],"slug":"ediscovery-and-disclosure-document-review","url":"https://www.blits.ai/ai-use-cases/ediscovery-and-disclosure-document-review","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":15150000,"min":300000,"max":30000000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"serious-fraud-office-ai-document-review","pooled":true},{"id":"purpose-legal-generative-ai-issues-review","pooled":true}]},{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":70000,"min":70000,"max":70000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"purpose-legal-generative-ai-issues-review","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":85,"min":85,"max":85,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"purpose-legal-generative-ai-issues-review","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":4000,"min":4000,"max":4000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"purpose-legal-generative-ai-issues-review","pooled":true}]}],"indicativeValueResult":{"low":225000,"high":1500000},"evidence":["doj-elitigation-ai-document-review","ftc-relativity-active-learning-review","purpose-legal-generative-ai-issues-review","serious-fraud-office-ai-document-review"]},{"title":"AI for fee and interest leakage detection","shortTitle":"Fee and interest leakage","seoTitle":"AI for bank fee and interest leakage detection","metaDescription":"How AI can help banks recompute fees, interest and FX margins against contract terms and flag overcharges and undercharges for correction and remediation.","definition":"An independent verification layer that recomputes what each fee, FX margin, spread and interest charge should have been under the contract and pricing tables, compares it with what was actually billed, and surfaces overcharges and undercharges account by account for correction, customer remediation and revenue recovery.","aliases":["fee leakage detection","revenue assurance for banks","interest and fee recalculation","pricing error detection","income leakage detection"],"industries":["banking","payments","cross-industry"],"functions":["finance-and-accounting","product-and-pricing","regulatory-compliance","operations"],"patterns":["anomaly-detection","agentic-workflow","rag-knowledge-assistant"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"copilot","adoptionStage":"emerging","segment":"back-office","problem":"Banks charge through many systems: core banking, card platforms, loan servicing, trade finance,\nFX and payments engines, each with its own pricing tables, waivers and exceptions. Over time the\nconfigured prices drift from what contracts, product terms and negotiated deals say. Some\ncustomers are overcharged, which is a conduct risk that can end in remediation programmes and fines.\nOthers are undercharged, which can leak revenue across many transactions.\n\nSome of these errors come to light only when a customer complains, an audit samples the right\naccounts or a regulator investigates, sometimes years after the error began. Recomputing every\ncharge independently was too expensive to do by hand. Cheap compute, contract reading with\nlanguage models and anomaly detection on fee lines can make continuous checking practical.","problemStats":[{"statement":"The CFPB ordered Wells Fargo in 2022 to pay more than USD 2 billion in redress across more than 16 million consumer accounts, for harms that included fees and interest improperly charged on auto and mortgage loans and incorrect charges on deposit accounts.","sourceTitle":"CFPB Orders Wells Fargo to Pay $3.7 Billion for Widespread Mismanagement of Auto Loans, Mortgages, and Deposit Accounts","sourceUrl":"https://www.consumerfinance.gov/archive/newsroom/cfpb-orders-wells-fargo-to-pay-37-billion-for-widespread-mismanagement-of-auto-loans-mortgages-and-deposit-accounts/","year":2022}],"howItWorks":"1. **Build the price book.** Contract terms, product disclosure documents, negotiated pricing and\n   waivers are read and turned into a structured, versioned price book. Language models help\n   extract terms from contracts; people approve every entry.\n2. **Recompute independently.** For each account and period, the engine recomputes what should\n   have been charged (fees, interest accruals, FX margins, spreads) from the price book and the\n   transaction data, outside the billing systems.\n3. **Compare and detect.** It compares expected with actual charges line by line and runs anomaly\n   detection on fee lines to catch patterns the rules miss, such as a waiver that never expired.\n4. **Explain.** For each discrepancy it produces an explanation (which term, which system, since\n   when, how many accounts) so the product owner can decide quickly.\n5. **Correct and remediate.** Confirmed errors go to the owners of the billing configuration for\n   a fix, and to a remediation process that refunds customers or recovers undercharges, with\n   human approval.","valueDrivers":["risk-reduction","compliance","revenue-growth","customer-experience"],"kpis":["detection-rate-improvement","error-reduction","cost-savings","accuracy"],"indicativeValue":{"referenceOrg":"A bank with USD 500 million in annual fee and commission income","inputs":[{"key":"feeIncome","label":"Annual fee and commission income","low":500000000,"high":500000000,"unit":"USD per year","note":"The reference bank. Replace with your own fee and commission income."},{"key":"leakageRate","label":"Share of fee income lost to undercharging","low":0.002,"high":0.01,"unit":"fraction of fee income","note":"Editorial assumption, replace with the results of a sample recomputation on your own accounts. No verified public benchmark was found."},{"key":"recoveryShare","label":"Share of leakage found and fixed going forward","low":0.3,"high":0.6,"unit":"fraction of leakage","note":"Editorial assumption; some leakage is found but deliberately left in place, for example commercial waivers."}],"formula":"feeIncome * leakageRate * recoveryShare","currency":"USD","period":"per year","resultLabel":"Fee income recovered from undercharging","caveat":"Undercharging only. It leaves out the benefit of finding overcharges early (smaller remediation programmes, fewer penalties), and the cost of building the price book, the platform and the remediation process."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The recompute logic must reproduce interest and fee conventions exactly (day count, rounding, tiering, value dates), and the contract terms are scattered across documents and systems. The AI helps with reading and detecting; the hard part is a trusted, versioned price book.","dataPrerequisites":["Contracts, product terms and negotiated pricing, with their effective dates","Transaction, balance and charge data per account from each billing system","Waiver and exception records with approvals and expiry dates","Past remediation cases as labelled examples"],"integrations":["Core banking, card, loan servicing and payments systems (read only)","Contract and document management","Pricing and deal management tools","Remediation and complaints case management","General ledger for recovered income"]},"implementation":{"steps":[{"title":"Start with one product and a sample","detail":"Pick a product with complex pricing and many accounts, such as business accounts or trade finance, and recompute a sample by hand and by machine to prove the logic."},{"title":"Build the price book with owners","detail":"Extract terms with AI assistance, but have product owners approve every entry and its effective dates. The price book becomes a control in its own right."},{"title":"Run continuously, report monthly","detail":"Recompute every account each cycle and report discrepancies by root cause, not only by account, so configuration errors are fixed once."},{"title":"Connect to remediation","detail":"Agree with compliance how confirmed overcharges become remediation cases, and how customers are contacted and refunded."},{"title":"Extend to more systems","detail":"Add products and systems one by one, reusing the price book structure and the recompute engine."}],"guardrails":["The recompute logic is deterministic, versioned and documented; the model never calculates charges","Every price book entry has an owner, a source document and an effective date","Corrections to customer accounts need human approval and are logged","Overcharges are always escalated to remediation, never netted against undercharges"],"humanInTheLoop":"Product owners approve the price book and decide on each class of discrepancy. Remediation teams approve refunds and customer contact, and finance approves recovery of undercharged income. Internal audit reviews the recompute logic periodically.","kpisToInstrument":["Discrepancies found per product and root cause","Value of overcharges refunded and undercharges recovered","Time from error start to detection","Share of discrepancies confirmed as real on review","Remediation cases opened from the control versus from complaints"],"failureModes":[{"title":"False alarms from convention mismatches","detail":"The recompute uses a different day count or rounding than the billing system, so every account looks wrong. Validate conventions per product before scaling."},{"title":"Findings without owners","detail":"Discrepancies pile up because no one owns the fix. Assign each product and system an accountable owner before switching on."},{"title":"Quietly keeping overcharges","detail":"Commercial pressure favours recovering undercharges over refunding overcharges. Make overcharge remediation a mandatory, audited path."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Verifying charges against contracts is not listed in Annex III and is not a practice prohibited by Article 5. The system is internal, so the Article 50(1) duty to tell people they are dealing with AI does not arise; the Article 50(2) duty to mark generated text, such as the discrepancy explanations, falls on the provider of the generative model or system. It supports, but does not take, decisions about individual customers; remediation decisions stay with people."},"regulations":["eu-ai-act","uk-consumer-duty","gdpr","apra-cps-230","us-sr-11-7"],"guidance":[{"title":"RG 277 Consumer remediation","issuer":"Australian Securities and Investments Commission","region":"asia-pacific","url":"https://asic.gov.au/regulatory-resources/find-a-document/regulatory-guides/rg-277-consumer-remediation/","note":"ASIC guidance, issued in 2022, for financial services and credit licensees on running consumer remediation, including identifying affected customers and returning money, for example after wrong fees or charges."},{"title":"FCA Consumer Duty","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/firms/consumer-duty/about","note":"The FCA's overview of the Duty, including the price and value outcome and its fair value assessments; charges above the agreed terms work against fair value."}],"controls":["Price book under change control with owners and effective dates","Versioned recompute logic with documented conventions and independent validation","Log of every discrepancy, decision, correction and refund","The control itself inventoried and monitored, with coverage reported by product"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the scheduled part runs as an **agentic workflow**: **custom functions** run the\ndeterministic recompute in isolated custom code or call the bank's own engine, **SQL knowledge\nbases** give read access to charge and transaction data, and an **AI agent** with **structured\noutput** explains each discrepancy and groups them by root cause. Contract terms and product\ndocuments sit in the **knowledge base** with hybrid retrieval, so the agent can retrieve the\nclause behind each explanation and include it in its structured output.\n\n**Human in the loop approval** holds every correction or refund for the product owner, each run\nkeeps a **full audit trail**, and **agentic tasks** recheck open discrepancies on a schedule\nuntil they are resolved. The platform is model agnostic and runs in EU or UAE regions where\ndata must stay local."},"faq":[{"question":"Why not rely on the billing systems to charge correctly?","answer":"Because configuration can drift from contracts over years, across many systems, and errors can be found only through complaints, audits or regulators. Enforcement cases such as the CFPB's 2022 order against Wells Fargo, which covered fees and interest improperly charged on loans, show how large the consequences can become."},{"question":"Does the AI calculate the correct charges?","answer":"No. The recompute logic is deterministic and versioned. AI helps read contracts into the price book, detect unusual fee patterns and explain discrepancies, but the numbers come from rules that product owners approve."},{"question":"Are there public examples of banks doing this with AI?","answer":"Few, and they disclose little. State Bank of India's 2019-20 annual report lists models to identify income leakage among the machine learning models its Analytics Department built in house, and an article by an SBI chief manager reports processing and facility fees recovered in two fiscal years. Revenue assurance and pricing vendors describe similar work, but the case studies we checked either do not name the bank or do not say the detection uses AI."}],"related":["ledger-and-payment-reconciliation","complaints-root-cause-analysis","chargeback-and-representment","continuous-controls-testing"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog. Kept in draft because no named public deployment with verified results was found."},{"date":"2026-09-25","note":"Consolidation pass: industries now include government, where its evidence comes from; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; corrected the ASIC refund statistic (will refund, over 920,000 customers); CMS example described as an initiated project, not in use; softened unsupported generalisations; clarified the EU AI Act basis and the guidance notes. Kept in draft: one public evidence record."},{"date":"2026-09-27","note":"Review fixes: metaDescription now describes the use case without implying adoption or a prevented outcome; EU AI Act basis separates Article 50(1) from 50(2); the knowledge base no longer claims guaranteed citations; removed the ASIC low income fee statistic (a different harm than charges that break contract terms); FCA guidance links the Consumer Duty overview page; softened unsourced generalisations."},{"date":"2026-09-27","note":"Evidence review: the CMS drug cost anomaly detection record did not genuinely fit a bank fee and interest use case, so it now sits under benefit fraud and error detection instead; dropped the government industry tag that depended on it. A further search for named banks or financial institutions with a verified AI deployment (newsrooms, annual reports, regulator publications, vendor customer stories) found none; the page stays in draft with no evidence records."},{"date":"2026-09-28","note":"Evidence search: added the first public record, State Bank of India (grade B, annual report 2019-20: in house machine learning models to identify income leakage, with fee recoveries reported by an SBI chief manager); added the alias income leakage detection; rewrote the FAQ on public examples; moved to review."}],"slug":"fee-and-interest-leakage-detection","url":"https://www.blits.ai/ai-use-cases/fee-and-interest-leakage-detection","benchmarks":[],"indicativeValueResult":{"low":300000,"high":3000000},"evidence":["state-bank-of-india-income-leakage-models"]},{"title":"AI for freedom of information request processing","shortTitle":"Freedom of information requests","seoTitle":"AI for FOI and FOIA request processing","metaDescription":"AI groups similar FOI requests, deduplicates records and proposes redactions for officers to review, as at the US FDA, the Interior Department and North Holland.","definition":"AI that helps a public body handle freedom of information and open government requests: logging and clarifying requests, spotting duplicates, searching and deduplicating the records in scope, proposing redactions with the exemption that applies, and drafting the response letter, with an FOI officer deciding what is released.","aliases":["FOIA processing AI","AI redaction for FOI requests","open government request handling","public records request automation"],"industries":["government"],"functions":["citizen-services","legal","case-management"],"patterns":["document-processing","classification-and-routing","content-generation"],"channels":["internal-tools","email"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Freedom of information laws give everyone the right to ask for government records, and request\nvolumes keep rising. Each request means finding every relevant record across email, file shares and\ncase systems, removing duplicates, reading everything, redacting personal data and exempt material,\nand explaining the decision within a statutory deadline. Large requests can involve very large\ndocument sets, and similar requests can reach several offices of the same government body.\n\nThe work is mostly manual and legal in nature, so backlogs grow and deadlines are missed, which\nundermines the transparency the law is meant to deliver. Errors cut both ways: over redaction\nwithholds information the public is entitled to, and under redaction leaks personal data.","problemStats":[{"statement":"US federal agencies received a record 1,707,197 FOIA requests in fiscal year 2025, 13.7% more than the year before, and ended the year with 339,671 backlogged requests, a 27% increase.","sourceTitle":"2025 Annual FOIA Report Summary","sourceUrl":"https://www.justice.gov/oip/media/1450791/dl?inline","year":2026}],"howItWorks":"1. **Log and clarify.** Incoming requests are read, logged with key fields and compared with open and\n   past requests, so similar requests are grouped and answered consistently. Unclear requests get a\n   drafted clarification question.\n2. **Collect and cull.** Records gathered under the search plan are made searchable (including text\n   recognition for scans), deduplicated and grouped by topic so reviewers see what is relevant first.\n3. **Propose redactions.** Named entity recognition and trained models mark personal data and other\n   candidate redactions, each with the proposed exemption code.\n4. **Review and decide.** FOI officers and lawyers accept, change or reject every proposed\n   redaction and decide what is released, with a second reviewer for doubtful cases.\n5. **Draft the response.** The response letter, including the exemptions relied on and appeal\n   rights, is drafted from the decision record for the officer to finalise.","valueDrivers":["employee-productivity","speed","compliance","inclusion-and-access"],"kpis":["processing-time-reduction","hours-saved","cycle-time-days","accuracy"],"indicativeValue":{"referenceOrg":"A government department that receives 5,000 information requests a year","inputs":[{"key":"requests","label":"Requests per year","low":3000,"high":7000,"unit":"requests per year","note":"Editorial assumption. Replace with your own request log."},{"key":"hoursPerRequest","label":"Staff hours per request today","low":5,"high":12,"unit":"hours per request","note":"Editorial assumption covering search, review, redaction and response. Large requests take far longer."},{"key":"savedShare","label":"Share of those hours saved","low":0.15,"high":0.3,"unit":"fraction of hours","note":"Editorial assumption. No agency on this page publishes a measured saving; legal review remains human work."},{"key":"hourlyCost","label":"Fully loaded cost of an FOI officer's hour","low":40,"high":60,"unit":"EUR per hour","note":"Editorial assumption. Replace with your own staff cost."}],"formula":"requests * hoursPerRequest * savedShare * hourlyCost","currency":"EUR","period":"per year","resultLabel":"FOI officer time released","caveat":"Time released, not cash saved. It leaves out licence and assurance costs, the value of meeting statutory deadlines, and the cost of a redaction error, which can be far larger than the saving."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Document review and redaction platforms are established products (the Province of North Holland has used one since 2021); the effort is in connecting to the places records live, tuning redaction to the agency's exemptions and building a review workflow that lawyers trust.","dataPrerequisites":["Request log with past requests, decisions and exemptions applied","Access to the record stores searched for requests (email, file shares, case systems)","Written redaction rules per exemption, with examples"],"integrations":["FOI case management or request tracking system","eDiscovery or document review platform","Email and records management systems","Public disclosure log or reading room for published responses"]},"implementation":{"steps":[{"title":"Start with deduplication and grouping","detail":"Grouping similar requests and exact duplicate documents is a lower risk place to start, because no redaction or release decision is automated. Check near duplicate removal with care, since versions that differ can both be responsive. The Department of the Interior's Office of the Solicitor has used request similarity and clustering tools since 2023 to coordinate answers to similar requests."},{"title":"Introduce proposed redactions with full review","detail":"Let the tool propose redactions with the redaction code for each, as FDA's FRED tool does, while officers still review and approve every proposal. Measure how often proposals are changed."},{"title":"Tune rules to your exemptions","detail":"Configure generic patterns (phone numbers, national identifiers) and individual rules for names, as the Province of North Holland does."},{"title":"Draft the response letter last","detail":"Once decisions are recorded per document, generate the response letter from them, so the letter matches what was actually decided."}],"guardrails":["Every redaction and release decision is taken by an FOI officer, never by the tool","Second review for documents where the officer is in doubt","Redaction burned into the released file, with the original kept securely","Personal data in request logs and prompts masked and retained only as the law allows","Search plan documented, so the scope of records is defensible"],"humanInTheLoop":"FOI officers and lawyers own the search plan, review every proposed redaction and decide what is released. A second reviewer or team lead checks doubtful documents. Requesters keep their complaint and appeal rights.","kpisToInstrument":["Median days from request to response and share within the statutory deadline","Backlog of open requests","Share of proposed redactions changed by reviewers","Redaction errors found after release","Appeals upheld against over redaction"],"failureModes":[{"title":"Missed personal data","detail":"A name in an image, a signature or an unusual format is not recognised and is released. Reviewers must check every page, and scanned material needs extra care."},{"title":"Over redaction by default","detail":"Accepting every proposal withholds information that should be public. Track changed proposals and appeals."},{"title":"Incomplete search","detail":"AI speeds up review but does not fix a search that missed a record store. Keep the search plan explicit."},{"title":"Redaction that can be undone","detail":"Visual boxes over text that can still be copied from the file. Use tools that remove the underlying text."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Tools that support staff in searching, deduplicating and proposing redactions are not listed in Annex III (point 5(a) covers eligibility for public assistance benefits and services, not access to documents), and every release decision stays with an officer. A public facing request assistant that talks to requesters would carry the Article 50(1) transparency duty."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Algoritmeregister van de Nederlandse overheid","issuer":"Government of the Netherlands","region":"europe","url":"https://algoritmes.overheid.nl/nl","note":"Dutch public bodies register the tools they use to support open government requests, including how staff check proposed redactions."},{"title":"US Department of Justice, Office of Information Policy: annual FOIA report summaries","issuer":"U.S. Department of Justice","region":"north-america","url":"https://www.justice.gov/oip/reports-1","note":"Government wide request volumes, processing times and backlogs, the baseline against which improvements should be measured."}],"controls":["Documented redaction rules per exemption and a review workflow with sign off","Audit trail of proposed and final redactions per document","Quality sampling of released documents for missed personal data","Access controls on the record sets gathered for each request","Register or inventory entry for each AI tool used"],"incidents":[]},"blitsAi":{"howToBuild":"Redaction itself is best done in a specialist document review platform, called through **custom\nfunctions**. Blits.ai adds the request handling around it: an **agentic workflow** that reads a new\nrequest from the **email channel**, logs it, compares it with past requests stored in a **SQL\nknowledge base**, drafts a clarification question where needed and, once officers have recorded\ntheir decisions, drafts the response letter, with **human in the loop approval** at each step.\n\nA **knowledge base** with hybrid retrieval over the FOI law, exemption guidance and past decision\nletters helps officers apply exemptions consistently. **PII masking** at the gateway keeps\nrequester data out of prompts and logs, every run keeps a full audit trail, and EU and UAE data\nresidency keeps records in region."},"faq":[{"question":"Can AI redact documents for FOI requests?","answer":"It can propose redactions. FDA's FRED tool marks the text it recommends redacting with a redaction code, and the Province of North Holland uses general rules that recognise phone and citizen service numbers, while names need their own individual rules. In both cases staff review the proposals and decide; accountability stays with the officer."},{"question":"Where does AI help most in FOI work?","answer":"In the volume steps: grouping similar requests, deduplicating and sorting records, and proposing redactions of personal data. Judgment on exemptions and the public interest remains human work."},{"question":"Is AI for FOI processing high risk under the EU AI Act?","answer":"No. Handling requests for access to documents is not listed in Annex III, and good practice keeps every release decision with an officer. If you add a chatbot that talks to requesters, it must tell them they are dealing with AI (Article 50)."}],"related":["correspondence-triage-and-routing","civil-servant-drafting-copilot","court-and-case-file-summarization","enterprise-knowledge-search","intelligent-document-processing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version for the government vertical, with US federal and Dutch evidence verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Interior Department dates (August and November 2023), the FRED redaction code wording and an unsupported staff review claim for the Justice Department, softened unsupported statements in the problem, steps and feasibility note, made the EU AI Act basis precise, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: North Holland name rules now follow the register wording (individual rules, not rules per request), dropped the web chat channel so channels match the minimal tier, reworded the EU AI Act answer and softened the page volume and deduplication statements."}],"slug":"freedom-of-information-request-processing","url":"https://www.blits.ai/ai-use-cases/freedom-of-information-request-processing","benchmarks":[],"indicativeValueResult":{"low":90000,"high":1512000},"evidence":["doj-foia-production-tools","fda-foia-redaction-tool","provincie-noord-holland-woo-request-support","us-department-of-the-interior-foia-request-similarity-tools"]},{"title":"AI for health insurance prior authorization and claims adjudication support","shortTitle":"Health prior authorization and adjudication","seoTitle":"AI prior authorization and health claims review","metaDescription":"AI checks health claims and prior authorization files against policy and clinical criteria; people make every denial. ICICI Lombard cut time per claim by over 50%.","definition":"AI that reads prior authorization requests, medical claims and appeals with their clinical and billing documents, extracts diagnoses, treatments and costs, checks them against the policy and published clinical criteria, and prepares a summary and recommendation for a clinician or adjudicator, who makes every adverse decision.","aliases":["AI prior authorization","utilization management AI","health claims adjudication copilot","medical claims automation","appeals letter drafting for health plans"],"industries":["insurance","healthcare"],"functions":["claims","case-management","operations"],"patterns":["document-processing","summarization","rag-knowledge-assistant","classification-and-routing","content-generation"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"claims","problem":"Health insurers and the administrators that work for them review large volumes of paper heavy cases.\nAn inpatient health claim arrives with a discharge summary, lab reports and bills; at ICICI Lombard\nthat meant reading 20 or more pages per claim. A prior authorization request brings the clinical\nnotes that justify the planned treatment. In both cases an adjudicator or nurse must check the file\nagainst the policy terms and clinical criteria. Much of the time goes into reading and retyping, not\ninto the judgment the role exists for.\n\nThe stakes are high on both sides. Slow reviews delay care and frustrate providers, while loose\nreviews let waste, abuse and billing errors through. Automation also carries a clear public risk:\ninvestigations and lawsuits in the United States have accused insurers of using algorithms to deny\ncare with little human review. Any AI in this process must make reviews faster and more consistent\nwithout taking the clinical decision away from people.","problemStats":[{"statement":"The US Centers for Medicare & Medicaid Services states that waste contributes to up to 25% of health care spending in the United States, and targets its WISeR prior authorization model at services with a history of waste, fraud and abuse.","sourceTitle":"WISeR (Wasteful and Inappropriate Service Reduction) Model","sourceUrl":"https://www.cms.gov/priorities/innovation/innovation-models/wiser","year":2025}],"howItWorks":"1. **Take in the request or claim.** Documents arrive through portals, electronic submissions,\n   email or scans; the AI classifies them and links them to the member and the case.\n2. **Extract and structure.** It extracts diagnosis, clinical presentation, history, treatment,\n   procedures, codes and billed amounts, with confidence levels.\n3. **Check against policy and criteria.** It compares the case with the member's cover and with the\n   published clinical guidelines the insurer uses, and lists what matches, what is missing and what\n   conflicts.\n4. **Prepare the file.** The reviewer receives a summary with the evidence behind each point and a\n   suggested outcome; clean, low risk cases that meet all criteria can be approved automatically.\n5. **Decide and communicate.** A clinician or adjudicator makes every denial or reduction, and the AI\n   drafts the plain language decision or appeal letter for approval.","valueDrivers":["speed","employee-productivity","cost-to-serve","compliance","customer-experience"],"kpis":["handling-time-reduction","processing-time-reduction","automation-rate","accuracy","hours-saved","interactions-handled"],"indicativeValue":{"referenceOrg":"A health insurer that reviews 400,000 claims and authorization requests by hand each year","inputs":[{"key":"cases","label":"Claims and authorization requests reviewed by a person per year","low":400000,"high":400000,"unit":"cases per year","note":"The reference insurer."},{"key":"minutesSaved","label":"Reviewer minutes saved per case","low":4,"high":12,"unit":"minutes per case","note":"Editorial assumption. The evidence on this page reports percentages, not minutes, for full case reviews (Microsoft reports that ICICI Lombard cut the time to process a single health claim by over 50%), and Acentra Health saved three minutes on the letter step alone (from six to three minutes per appeal letter). Replace with a time study."},{"key":"costPerHour","label":"Fully loaded cost per reviewer hour","low":40,"high":80,"unit":"USD per hour","note":"Editorial assumption; nurses and clinical reviewers cost more than claims processors. Replace with your own."}],"formula":"cases * minutesSaved / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Reviewer time released","caveat":"Review effort only. It leaves out faster decisions for members and providers, the effect of more consistent reviews on waste and appeals, the cost of clinical content and validation, and the cost of the platform."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Extraction from medical documents is proven, but the process handles sensitive health data, works under strict decision deadlines and appeal rules, and needs clinical criteria in a form the system can apply. Clinical, legal and compliance teams must own the design, and every adverse decision stays with a licensed reviewer.","dataPrerequisites":["Policy wordings, benefit rules and the clinical criteria or guidelines the insurer applies","Historical cases with documents, decisions and appeal outcomes","Code sets and provider data for billing checks","Approved letter templates for approvals, denials and appeal outcomes"],"integrations":["Claims adjudication and utilization management systems","Provider portals and electronic prior authorization interfaces","Document capture and storage for clinical records","Member and provider communication channels for decisions and letters"]},"implementation":{"steps":[{"title":"Start with summaries for reviewers","detail":"Let the AI structure documents and write a case summary that the reviewer checks. It saves time immediately and changes no decision rights."},{"title":"Encode the criteria with clinicians","detail":"Work with medical directors to turn the clinical criteria and benefit rules into checks the system can apply, each linked to its source document and version."},{"title":"Automate approvals only","detail":"If automation goes further, let it approve clean cases that meet every criterion. Denials, reductions and partial approvals always go to a licensed reviewer. In the EU, an automatic approval is still a solely automated decision based on health data: GDPR Article 22(4) allows it only with the member's explicit consent or on grounds of substantial public interest in law, so check that basis before you switch it on."},{"title":"Draft decision letters in plain language","detail":"Use the AI to turn the reviewer's rationale into a clear letter for the member and provider, approved by the reviewer before it is sent."},{"title":"Monitor outcomes by group and by reviewer","detail":"Track approval, denial and overturn rates on appeal across member groups, conditions and reviewers, and investigate any pattern the AI may have introduced."}],"guardrails":["No denial, reduction or termination of care without a licensed clinician or adjudicator reviewing the case","Every recommendation cites the clinical criterion or policy clause it relies on","Automatic decisions limited to approvals of cases that meet every criterion","Health data processed under HIPAA or GDPR special category rules, masked in logs and prompts","Overturn rates on appeal monitored for AI supported decisions"],"humanInTheLoop":"Licensed clinicians and adjudicators make every adverse decision and approve every letter before it is sent. Medical directors own the criteria encoded in the system, and a quality team samples AI supported approvals and summaries every week.","kpisToInstrument":["Reviewer minutes per case, before and after","Time from request to decision, against regulatory deadlines","Share of cases approved automatically and share sent to review","Overturn rate on appeal for AI supported decisions","Agreement between AI summaries and reviewer findings on a weekly sample"],"failureModes":[{"title":"Rubber stamp review","detail":"Reviewers approve the AI's suggested denial without reading the file. Measure review time per case, audit samples and never let the system propose a denial without the evidence attached."},{"title":"Criteria that drift from medicine","detail":"Encoded criteria fall behind updated guidelines. Give every criterion an owner, a version and a review date."},{"title":"Extraction errors in clinical detail","detail":"A missed comorbidity or wrong code changes the outcome. Show confidence per field and require the reviewer to confirm key facts."},{"title":"Unequal outcomes for vulnerable members","detail":"The system denies or delays care more often for elderly, disabled or chronically ill members, for example because their files are longer and less standard. Monitor outcomes by group and fix the cause."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 5(a) makes AI high risk when it is used by or on behalf of public authorities to evaluate eligibility for essential public assistance benefits and services, including healthcare services, or to grant, reduce or revoke them, which can cover statutory health schemes run by or for public bodies. Point 5(c) covers risk assessment and pricing in life and health insurance, not claim review. A copilot for a private insurer's claim review, where people decide, is usually outside Annex III; for public schemes, Article 6(3) may exempt a system that only performs a preparatory task, unless it profiles natural persons. GDPR rules on health data (Article 9) and on solely automated decisions (Article 22) apply in every case."},"regulations":["eu-ai-act","gdpr","hipaa","solvency-ii","nist-ai-rmf","iso-42001"],"guidance":[{"title":"CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F)","issuer":"Centers for Medicare & Medicaid Services","region":"north-america","url":"https://www.cms.gov/newsroom/fact-sheets/cms-interoperability-and-prior-authorization-final-rule-cms-0057-f","note":"Requires impacted payers, such as Medicare Advantage organizations and state Medicaid programs, to send prior authorization decisions within 72 hours for urgent and seven calendar days for standard requests, to give a specific reason for denials from 2026, and to offer a Prior Authorization API."},{"title":"WISeR (Wasteful and Inappropriate Service Reduction) Model","issuer":"Centers for Medicare & Medicaid Services","region":"north-america","url":"https://www.cms.gov/priorities/innovation/innovation-models/wiser","note":"Shows the US government's design for AI assisted prior authorization, in which every recommendation for non payment is decided by a licensed clinician."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(a) on public assistance benefits and services, including healthcare services, and point 5(c) on life and health insurance determine when this use is high risk in the EU."},{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Addressed to national supervisors, it clarifies how existing insurance legislation applies to the governance and risk management of AI systems used by insurers, following a risk based and proportionate approach."}],"controls":["Documented decision rights, where AI may recommend and approve within criteria and only people deny or reduce","Versioned clinical criteria and benefit rules with clinical owners","Audit trail of documents, extracted facts, criteria applied, reviewer and decision per case","Outcome and appeal overturn monitoring by member group and condition","Privacy and security controls for health data, including access logging"],"incidents":[{"title":"How Cigna Saves Millions by Having Its Doctors Reject Claims Without Reading Them","url":"https://www.propublica.org/article/cigna-pxdx-medical-health-insurance-rejection-claims","note":"ProPublica reported that Cigna doctors signed off payment denials flagged by its PXDX review process in batches, spending an average of 1.2 seconds per case according to company documents. Cigna said it was incorrect that the process lets its doctors reject claims without examining them. A warning about human review that becomes a formality."},{"title":"UnitedHealth sued over use of algorithm in Medicare Advantage plans","url":"https://www.statnews.com/2023/11/14/unitedhealth-class-action-lawsuit-algorithm-medicare-advantage/","note":"A class action alleged that UnitedHealth and its subsidiary NaviHealth used an algorithm, nH Predict, to deny rehabilitation care to seriously ill Medicare Advantage patients, and claimed a 90% error rate based on the share of denials reversed on appeal. UnitedHealth said the tool is not used to make coverage determinations."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered through the API for each request, claim or\nappeal. The case documents are passed to the workflow, an **AI agent** extracts the clinical and\nbilling facts with **structured output**, and **hybrid retrieval** over a **knowledge base** that\nholds the policy wordings and clinical criteria provides the citations for each check. **Custom\nfunctions** read member and benefit data from the adjudication system and write back the summary.\n\n**Human in the loop approval** keeps every adverse decision with a licensed reviewer, and the agent\ndrafts the decision or appeal letter for the reviewer to approve. **PII masking** keeps health data\nout of model prompts where possible, the **audit trail** records every run, and **test suites**\ngrade summaries against reviewed cases. EU and UAE data residency keeps health data in region."},"faq":[{"question":"Can AI deny prior authorization requests or claims?","answer":"It should not. The US government's own WISeR model uses AI to assist prior authorization reviews, but every recommendation for non payment is decided by an appropriately licensed clinician. The safe design lets AI summarise, check criteria and approve clean cases, and leaves every denial or reduction to a person."},{"question":"What time savings do health insurers report?","answer":"Microsoft reports that ICICI Lombard's claims copilot cut the time to process a single health claim by over 50%, and that Acentra Health halved nurse time per Medicare appeal letter, saving 11,000 nursing hours. Sprout.ai reports that AdvanceCare settles some routine claims in 60 seconds."},{"question":"What went wrong in the publicised cases of algorithmic denials?","answer":"ProPublica reported that Cigna doctors signed off denials in batches, at an average of 1.2 seconds per case, and a class action alleged that UnitedHealth used an algorithm to deny rehabilitation care and that most appealed denials were reversed. Both companies disputed these accounts. The lesson is to measure real human review (time per case, overturn rates on appeal) and to keep denials out of automation."}],"related":["claims-triage-and-straight-through-processing","life-underwriting-medical-record-summarization","outbound-notice-drafting","correspondence-triage-and-routing","medical-coding-automation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the insurance vertical with evidence from ICICI Lombard, AdvanceCare, Acentra Health, Manulife and the CMS WISeR model, verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription, a cited CMS problem statistic, Solvency II, a more precise EU AI Act basis and guidance notes, UnitedHealth's response to the lawsuit, an honest note on the minutes saved assumption, and corrected AdvanceCare and WISeR evidence details."},{"date":"2026-09-27","note":"Editorial review: meta description now matches the automatic approval design, the 20 page figure is limited to claims and attributed to ICICI Lombard, Cigna's response is quoted precisely, a GDPR Article 22(4) note was added to automatic approvals, and the Acentra Health approval rating is no longer counted as accuracy."}],"slug":"health-prior-authorization-and-claims-adjudication","url":"https://www.blits.ai/ai-use-cases/health-prior-authorization-and-claims-adjudication","benchmarks":[{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"acentra-health-medscribe-appeal-letters","pooled":true},{"id":"icici-lombard-health-claims-copilot","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":98,"min":98,"max":98,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"advancecare-health-claims-automation","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":11000,"min":11000,"max":11000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"acentra-health-medscribe-appeal-letters","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1000000,"min":1000000,"max":1000000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"advancecare-health-claims-automation","pooled":true}]}],"indicativeValueResult":{"low":1066666.6666666667,"high":6400000},"evidence":["acentra-health-medscribe-appeal-letters","advancecare-health-claims-automation","cms-wiser-prior-authorization-model","icici-lombard-health-claims-copilot","manulife-health-dental-claims-document-ai"]},{"title":"AI for immigration and visa applications, from applicant questions to case preparation","shortTitle":"Immigration and visa application assistant","seoTitle":"AI for visa and immigration applications","metaDescription":"USCIS uses AI to extract form data and tag evidence; the Home Office routes visitor visa applications by rules and profiles. Officers decide every case.","definition":"AI that helps applicants understand immigration and visa requirements and submit complete applications, and helps immigration staff prepare cases by extracting form data, classifying evidence, routing applications and supporting interviews, while every grant or refusal is decided by an officer against the immigration rules.","aliases":["visa chatbot","immigration virtual assistant","visa application processing AI","consular assistant"],"industries":["government"],"functions":["citizen-services","case-management","operations"],"patterns":["conversational-agent","document-processing","classification-and-routing","translation"],"channels":["web-chat","voice","internal-tools","api"],"audience":"customer-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Immigration systems combine high volume, complex rules and applicants who often write in a second\nlanguage, from abroad, under time pressure. Applicants struggle to understand which route applies\nand which evidence to send; contact centres answer the same routine questions again and again;\nofficers sift through case files that can hold hundreds of pages of scanned evidence; and\nscanned forms must be validated against form specific business rules.\n\nIt is also one of the most sensitive places to use algorithms. Decisions shape people's lives,\nand the UK withdrew a visa streaming algorithm in 2020 after a legal challenge alleging\nnationality based discrimination. AI here has to make applicants better informed and officers\nfaster without shifting the decision to a model.","problemStats":[],"howItWorks":"1. **Inform the applicant.** An assistant answers questions about routes, fees, documents and\n   processes from official guidance, in the applicant's language, and points case specific\n   questions to online status or the contact centre.\n2. **Capture the application digitally.** Scanned or uploaded forms are read by document AI,\n   fields are extracted and checked against form rules, and gaps are flagged before the case\n   reaches an officer.\n3. **Organise the evidence.** Each page of supporting evidence is classified (passport, marriage\n   certificate, bank statement) so officers can jump to what they need.\n4. **Route by complexity.** Rules on declared answers send straightforward cases to one queue and\n   complex ones to more senior officers; officers can reroute. Any risk profiles or watch lists\n   added to the rules need their own equality review.\n5. **Support the interview.** Real time transcription and translation help officers and applicants\n   understand each other, with a transcript kept when enabled.\n6. **Decide.** An officer assesses the case against the rules and records the decision and reasons.","valueDrivers":["speed","employee-productivity","inclusion-and-access","customer-experience"],"kpis":["interactions-handled","processing-time-reduction","accuracy","time-saved-per-task","containment-rate"],"indicativeValue":{"referenceOrg":"An immigration authority that processes 1 million visa applications a year","inputs":[{"key":"applications","label":"Applications processed per year","low":1000000,"high":1000000,"unit":"applications per year","note":"The reference authority."},{"key":"minutesSaved","label":"Officer minutes saved per application on form entry and evidence navigation","low":3,"high":8,"unit":"minutes per application","note":"Editorial assumption. No public benchmark states time saved per case yet; replace with your own time study."},{"key":"hourlyCost","label":"Fully loaded officer cost per hour","low":40,"high":60,"unit":"USD per hour","note":"Editorial assumption. Replace with your own cost."}],"formula":"applications * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Officer time released, valued at cost","caveat":"Values officer time on data entry and evidence navigation only. It leaves out contact centre savings from applicant assistants, faster decisions for applicants and the cost of the systems, and assumes released time goes to backlog rather than headcount reduction."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Applicant facing information is moderate; anything inside the case touches high risk decisions, legacy case systems and strict evidence rules. Governance (impact assessments, equality duties, audit trails) is as much work as the technology.","dataPrerequisites":["Official guidance per route, with owners and effective dates","Form definitions and business rules for each application type","Labelled examples of evidence document types for classification","Historic routing and outcome data for testing, handled under strict access controls"],"integrations":["Online application portal and document upload","Case management system for immigration decisions","Identity and biometrics systems (read only for the AI components)","Contact centre and appointment systems","Interpretation and translation services"]},"implementation":{"steps":[{"title":"Map every AI output to the decision it touches","detail":"For each component (applicant assistant, form extraction, evidence tagging, routing, interpretation) state whether it informs the applicant, prepares the case or affects the decision, and govern it accordingly."},{"title":"Start with applicant information and form capture","detail":"Answering general questions and turning paper forms into data (as USCIS did with PDF Intake for myUSCIS) reduce work without touching the decision."},{"title":"Keep routing rules simple and published","detail":"If you route by complexity, use transparent rules on declared answers, let officers reroute, and publish a record of how routing works. The Home Office published a transparency record for its visitor visa routing and lets officers reroute cases, but its rules also use risk profiles and bulk data tables that the record does not set out, and allow direct nationality based discrimination under a Ministerial Authorisation; decide deliberately whether you want either."},{"title":"Treat interpretation as assistive","detail":"Use AI interpretation to support, not replace, certified interpreters where the law requires them, and keep transcripts for review. The State Department describes its citizen services pilot as assistive only, and its consular pilots can produce a time stamped transcript."},{"title":"Test for bias across nationalities and languages","detail":"Measure error and routing rates by nationality, language and other protected characteristics before launch and at intervals."}],"guardrails":["No AI output grants, refuses or recommends a decision; officers decide against the rules","Routing and prioritisation rules documented, explainable and reroutable by officers","Equality impact assessment and monitoring by nationality and language","Applicant assistants answer only from official guidance and never predict the outcome of a case","Translations and extracted data shown next to the original for officer checking"],"humanInTheLoop":"Officers decide every application and can override any routing or extraction. Supervisors review samples of AI prepared cases and interpretation transcripts; policy owners approve changes to guidance and routing rules; complaints and appeals are handled by people.","kpisToInstrument":["Time from submission to decision per route","Extraction accuracy per form field and evidence classification accuracy on a checked sample","Share of cases rerouted by officers after automated routing","Routing and error rates by nationality and language","Contact volume on status and document questions"],"failureModes":[{"title":"Discriminatory routing","detail":"Nationality or proxies for it drive a case into a slower or stricter lane. The UK visa streaming tool was withdrawn in 2020 after a legal challenge. Exclude protected characteristics and their proxies from routing rules and risk profiles, and monitor outcomes."},{"title":"Translation errors in the record","detail":"Machine translation errors in asylum applications or visa interviews can change the meaning of an applicant's account. Keep transcripts, allow correction and use certified interpreters where required."},{"title":"Silent extraction errors","detail":"A misread date or name propagates through the case. Show extracted data next to the source and flag low confidence fields."},{"title":"Applicants read predictions into answers","detail":"An assistant that sounds certain about eligibility creates false expectations. Answer only from guidance and state that officers decide."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 7(c) makes AI high risk when it assists public authorities in examining applications for asylum, visas or residence permits, including assessing the reliability of evidence. Applicant facing information assistants that give general guidance fall under the Article 50 transparency duties (limited risk). Evidence classification, routing and interview support used in the examination are likely high risk, unless the provider documents under Article 6(3) that a component only performs a narrow procedural or preparatory task. That exception never applies to a system that profiles natural persons, which matters for routing on personal attributes or risk profiles."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 7 lists migration, asylum and border control management uses, including the examination of visa and residence permit applications."},{"title":"Article 27, fundamental rights impact assessment for high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/27/","note":"Public bodies must assess fundamental rights impacts before deploying high risk AI."},{"title":"Directive on Automated Decision-Making","issuer":"Treasury Board of Canada Secretariat","region":"north-america","url":"https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32592","note":"Sets algorithmic impact assessment, notice, testing and human involvement requirements for automated decision systems used by Canadian federal institutions."},{"title":"M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust","issuer":"US Office of Management and Budget","region":"north-america","url":"https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf","note":"Minimum risk management practices for high impact AI in US federal agencies; the federal AI use case inventory flags the State Department's visa interview pilot as high impact."}],"controls":["Public register entry for each AI component, with its role in the decision","Fundamental rights and equality impact assessments before launch","Audit trail linking every AI output to the officer's decision","Right to an interpreter and to correct the record","Periodic independent review of routing and extraction outcomes"],"incidents":[{"title":"Home Office says it will abandon its racist visa algorithm, after we sued them","url":"https://www.foxglove.org.uk/2020/08/04/home-office-says-it-will-abandon-its-racist-visa-algorithm-after-we-sued-them/","note":"The UK Home Office withdrew its visa application streaming tool in 2020 after a legal challenge alleging nationality based discrimination."},{"title":"Lost in AI translation: growing reliance on language apps jeopardizes some asylum applications","url":"https://www.theguardian.com/us-news/2023/sep/07/asylum-seekers-ai-translation-apps","note":"The Guardian reported cases where machine translation errors affected US asylum applications."}]},"blitsAi":{"howToBuild":"On Blits.ai the applicant side is an **AI agent** grounded in a **knowledge base** of official\nimmigration guidance, crawled from the authority's website and retrieved with **hybrid search**,\nwith **language detection and translation** so applicants can ask in their own language.\n**Guardrails** stop it from predicting case outcomes, **PII masking** removes passport and case\nnumbers before text reaches a model, and **human handover** routes case questions to the contact centre.\n\nOn the case side, an **agentic workflow** triggered through the API takes the text of uploaded\nforms and evidence (from the authority's own OCR service, called through **custom functions**),\nextracts form fields, classifies evidence pages and writes the result to the case system, with\n**human in the loop confirmation** so an officer approves agentic actions before they are carried out. Every run has a full audit trail.\nRouting rules stay deterministic in a **flow**. **Test suites** check extraction and answers by\nlanguage, the platform is **model agnostic**, and **EU and UAE data residency** keeps data in region."},"faq":[{"question":"Is AI deciding visa applications?","answer":"Not in the examples on this page. The Home Office routing tool uses rules, risk profiles and bulk data tables to set a complexity label and does not decide applications; USCIS uses AI to extract form data and tag evidence; the State Department's interview tool translates. Officers decide, and the EU AI Act lists AI that assists the examination of visa applications as high risk in Annex III."},{"question":"What volumes are involved?","answer":"Large. The Home Office has routed visitor visa applications under Appendix V with its tool since April 2023, and it received around 2.5 million visit applications in 2023. At that scale even small error rates affect many people."},{"question":"Can AI interpretation replace interpreters at visa interviews?","answer":"Not in the public examples. The State Department is piloting live AI interpretation at the visa interview window and flags it as high impact in the federal inventory, and its citizen services pilot is explicitly assistive, not a replacement for certified interpreters where they are required."}],"related":["public-service-translation","citizen-information-assistant","benefits-eligibility-and-application-assistant","permit-and-licence-application-processing","intelligent-document-processing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from the US federal AI inventory, UK transparency records and agency pages, with every source checked."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsupported claims from the problem (share of officer time, backlogs, appeal rights), added the Article 6(3) nuance to the EU AI Act basis, corrected guidance notes, tightened evidence summaries and the case side build description; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Corrected the Home Office routing description: it also uses risk profiles and bulk data tables and allows nationality based direct discrimination under a Ministerial Authorisation, has applied since 25 April 2023, and the 2.5 million figure is applications received, not routed (metric removed); aligned the EU AI Act FAQ with Annex III and the profiling limit on Article 6(3); fixed evidence channels, years and summaries."}],"slug":"immigration-and-visa-application-assistant","url":"https://www.blits.ai/ai-use-cases/immigration-and-visa-application-assistant","benchmarks":[],"indicativeValueResult":{"low":2000000,"high":8000000.000000001},"evidence":["home-office-visit-visa-complexity-routing","us-department-of-state-consular-ai-interpretation","us-immigration-and-customs-enforcement-sevp-voice-assistant","uscis-myuscis-pdf-intake-and-evidence-classification"]},{"title":"AI for inbound correspondence triage and routing","shortTitle":"Correspondence triage and routing","seoTitle":"AI correspondence triage for banks and insurers","metaDescription":"AI sorts inbound letters and emails and routes each to the right team. Travelers reached 91% accuracy in testing; the VA ingests up to 40,000 packets a day.","definition":"AI that sorts inbound correspondence before anyone answers it: it takes every inbound letter, email, upload and secure message into one intake, identifies what it is, extracts the key fields, links it to the right customer and account, sets priority and routes it to the right team or workflow, replacing the manual sorting desk.","aliases":["intelligent mailroom","email triage AI","inbound document classification","digital mailroom automation"],"industries":["cross-industry","banking","insurance","government"],"functions":["operations","customer-service","case-management"],"patterns":["classification-and-routing","document-processing","summarization"],"channels":["email","internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"mainstream","segment":"back-office","problem":"Banks, insurers and public bodies still receive a large share of their work as unstructured\ncorrespondence: scanned letters, emails with attachments, portal uploads and secure messages.\nSomeone has to open each item, decide what it is (a complaint, a power of attorney, a\nbereavement notice, a change of address, a payment instruction), find the customer and send it\nto the right queue. That sorting step adds delay before anyone starts the real work.\n\nManual sorting is also where risk hides. A complaint filed as a general enquiry can miss its\nregulatory deadline, a bereavement letter can sit in the wrong queue, and a fraud warning can be\nread days late. Rule based keyword routing helps with obvious cases but struggles with free text\nand mixed documents.","problemStats":[],"howItWorks":"1. **One intake.** Post is scanned; emails, uploads and secure messages land in the same queue\n   with their attachments.\n2. **Classify.** The AI identifies the document or request type, the language and any urgency\n   signal (complaint, vulnerability, fraud, legal deadline), with a confidence score.\n3. **Extract and link.** It extracts the key fields (names, account numbers, dates, amounts,\n   reference numbers) and matches the item to the customer and account in the core systems.\n4. **Route or trigger.** It sets the priority and service level, routes the item to the right\n   team, or starts the downstream workflow directly (for example an address change or a\n   bereavement case), with a short summary for the receiving officer.\n5. **Fall back to people.** Low confidence items, unmatched customers and anything sensitive go to\n   a human review queue, and every correction becomes training data.","valueDrivers":["cost-to-serve","speed","compliance","employee-productivity"],"kpis":["automation-rate","processing-time-reduction","accuracy","hours-saved","interactions-handled"],"indicativeValue":{"referenceOrg":"A bank or insurer that receives 1 million inbound letters and emails a year","inputs":[{"key":"items","label":"Inbound items per year","low":1000000,"high":1000000,"unit":"items per year","note":"The reference organization. Replace with your own mailroom and mailbox volume."},{"key":"minutesPerItem","label":"Minutes to read, classify, index and route one item manually","low":2,"high":4,"unit":"minutes per item","note":"Editorial assumption, replace with your own time study."},{"key":"automatedShare","label":"Share of items routed without human touch","low":0.5,"high":0.8,"unit":"fraction of items","note":"Editorial assumption, replace with your own. None of the evidence on this page reports a share of correspondence routed with no human touch; measure it on your own labelled sample before relying on it."},{"key":"costPerHour","label":"Fully loaded cost per hour of intake staff","low":30,"high":50,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"items * minutesPerItem / 60 * automatedShare * costPerHour","currency":"USD","period":"per year","resultLabel":"Manual sorting effort avoided","caveat":"Sorting labour only. It leaves out the value of faster downstream handling, fewer missed complaint deadlines and the cost of scanning, the platform and integration."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Classification and extraction are mature. The effort is in the taxonomy of request types, the customer matching against core data and the connections to every downstream queue.","dataPrerequisites":["A labelled sample of historical correspondence per request type","An agreed taxonomy of request types with owner teams and service levels","Customer and account data reachable for matching"],"integrations":["Scanning and mailroom capture","Shared mailboxes and secure messaging","Core banking or policy administration for customer matching","Case management and workflow tools for routing","Complaint management system"]},"implementation":{"steps":[{"title":"Build the request taxonomy with the receiving teams","detail":"List the request types, their owner team, priority and service level. Keep it short at first; a catch all class routed to people is better than twenty rare classes."},{"title":"Label a real sample","detail":"Label a few thousand recent items, including the messy ones (handwritten, multi topic, forwarded chains), and use them as the test set for every model change."},{"title":"Start with classification and summaries only","detail":"Let the AI propose the class and a summary while people still route. Measure accuracy per class before switching routing on."},{"title":"Route automatically per class above a confidence threshold","detail":"Switch on automatic routing class by class, with a review queue below the threshold and mandatory human review for complaints, vulnerability and legal documents."},{"title":"Trigger downstream workflows","detail":"For simple requests, start the fulfilment workflow directly from the extracted fields instead of dropping a task in a queue."}],"guardrails":["Complaints, vulnerability signals and fraud warnings are always flagged and never auto closed","Items below the confidence threshold go to a human review queue","Customer matching requires at least two strong identifiers before an item is linked","Personal data in documents is masked in logs and model prompts"],"humanInTheLoop":"People review low confidence items and every item flagged as a complaint, vulnerability or legal matter. A quality team samples automatically routed items weekly and feeds corrections back into the labelled set.","kpisToInstrument":["Classification accuracy per class on a weekly sample","Share of items routed without human touch","Time from receipt to arrival in the right queue","Misrouted items reported by receiving teams","Complaints identified at intake versus later"],"failureModes":[{"title":"A complaint routed as an enquiry","detail":"The regulatory clock runs while the item sits in the wrong queue. Treat complaint detection as its own high recall check."},{"title":"Linked to the wrong customer","detail":"A document attached to the wrong account is a data breach. Require strong identifiers and route ambiguous matches to people."},{"title":"Taxonomy drift","detail":"New products and campaigns create request types the model never saw. Review the catch all class monthly."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"It depends on where the system runs. Classifying and routing a bank's or insurer's correspondence is not a use listed in Annex III, so it is minimal risk: the AI literacy duty of Article 4 applies, and the Article 50 duty to tell people they are dealing with AI does not, because the system does not interact with the sender. Used by or for a public authority in a benefits process covered by Annex III point 5(a), the provider can treat it as not high risk only while it performs a narrow procedural or preparatory task under Article 6(3); the provider must then document that assessment before it goes live (Article 6(4)) and register the system in the EU database (Article 49(2)). If the system evaluates eligibility for benefits or profiles the people who write in, it is high risk, so those judgements stay with people."},"regulations":["eu-ai-act","gdpr","uk-gdpr","dora","uk-consumer-duty","apra-cps-230"],"guidance":[{"title":"DISP 1.6 Complaints time limit rules","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.handbook.fca.org.uk/handbook/DISP/1/6.html","note":"The response deadlines (eight weeks for most complaints, 15 business days for payment services and electronic money complaints) run from the firm's receipt of the complaint, so intake must recognise complaints wherever they arrive."}],"controls":["Misclassification rate tracked as a model health metric, per class","Log of every classification, extracted field and routing decision","Separate high recall check for complaints and vulnerability","Retention and access controls on scanned documents in line with the records policy"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the **email channel** and the REST API take inbound items into an **agentic\nworkflow**. An **AI agent** with **structured output** returns the class, the extracted fields\nand a summary for each item, and **custom functions** look up the customer and create the task\nor case in the downstream system. Scanned post enters through the same API once the capture\nsystem has turned it into text.\n\nRules in **flows** force complaints and vulnerability signals to a person, a **human review step**\nin the workflow takes low confidence items, and **PII\nmasking** keeps account and card numbers out of model prompts. **Test suites** run the labelled\nset on every change, and the workflow run history keeps an audit trail of every classification.\nData can stay in the EU or UAE region."},"faq":[{"question":"How accurate is AI at classifying inbound correspondence?","answer":"Good enough to route most items, not all. In a post written by AWS and Travelers staff, the Travelers policy service email classifier reached 68% accuracy in initial testing and 91% after prompt engineering and condensed categories; the ground truth set had over 4,000 labelled emails in 13 classes. These are test results, not production figures, which is why the design keeps a human review queue below a confidence threshold."},{"question":"Does this work for scanned paper as well as email?","answer":"Yes, once the capture step has turned the scan into text. The US Department of Veterans Affairs reports that its Mail Automation Services platform combines OCR, handwriting recognition and language processing, ingests 25,000 to 40,000 packets a day, mostly from its Centralized Mail Portal, and sends each submission to the right business line."},{"question":"What should never be routed automatically?","answer":"Complaints, signs of customer vulnerability, fraud warnings and legal documents such as court orders and powers of attorney should always be flagged for a person, even when the classifier is confident."}],"related":["email-and-ticket-reply-drafting","intelligent-document-processing","complaints-handling-agent","account-servicing-execution","outbound-notice-drafting","payment-investigations-and-exceptions"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed an unsourced delay claim, sharpened the EU AI Act basis (Articles 4, 6(3) and 50), added UK GDPR, corrected the Travelers attribution, the Loadsure and Encova summaries and source dates, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: EU AI Act tier set to context dependent (Articles 4, 6(3), 6(4), 49(2) and 50), the Encova 99% recorded as accuracy instead of automation rate, the automated share recorded as a plain editorial assumption, the Travelers result framed as a test result, the VA claimant note and IBM role corrected, and the human review wording aligned with the platform."},{"date":"2026-09-27","note":"Review fixes: the Encova metric reverted to automation rate in the organization's own words (a success rate of documents processed through without issues), since the vendor's separate accuracy headline does not match the taxonomy's accuracy definition of correct on a checked sample."}],"slug":"correspondence-triage-and-routing","url":"https://www.blits.ai/ai-use-cases/correspondence-triage-and-routing","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":62500,"min":25000,"max":100000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"master-trust-bank-of-japan-financial-document-capture","pooled":true},{"id":"us-department-of-veterans-affairs-mail-automation-services","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":91,"min":91,"max":91,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"travelers-policy-service-email-classification","pooled":true}]},{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":99,"min":99,"max":99,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"encova-insurance-document-intake-automation","pooled":true}]}],"indicativeValueResult":{"low":500000.00000000006,"high":2666666.666666667},"evidence":["ecclesia-group-claims-correspondence-routing","encova-insurance-document-intake-automation","loadsure-claims-document-classification","master-trust-bank-of-japan-financial-document-capture","travelers-policy-service-email-classification","us-department-of-veterans-affairs-mail-automation-services"]},{"title":"AI for insurance claims fraud detection","shortTitle":"Claims fraud detection","seoTitle":"Insurance claims fraud detection with AI","metaDescription":"AI scores insurance claims for fraud at first notice of loss. Shift reports AXA Switzerland analysed over 1 million claims and stopped over EUR 12 million in fraud.","definition":"AI that scores every insurance claim for fraud from first notice of loss onwards, combining claim, policy, document, image and network data to find suspicious claims, organised rings and inflated losses, and sends each alert with its reasons to a claims handler or special investigations unit for review.","aliases":["insurance fraud detection","claims fraud scoring","SIU referral scoring","fraud network analytics for claims"],"industries":["insurance"],"functions":["claims","fraud-prevention"],"patterns":["anomaly-detection","prediction-and-scoring","document-processing","computer-vision","classification-and-routing"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"assist","adoptionStage":"mainstream","segment":"claims","problem":"Fraud hides among honest claims. Most suspicious claims look normal when viewed alone: a slightly\ninflated invoice, a staged accident with credible witnesses, a repair shop or clinic that appears\nin too many claims, a photo reused from another insurer. Traditional detection relies on business\nrules and on handlers noticing something odd, which produces many false alerts and misses the\norganised schemes that span claims and insurers.\n\nThe pressure grows as insurers speed claims up. Faster payment and straight through processing are\nwhat customers want, and exactly what fraudsters exploit. Special investigations units are small,\nso the quality of each referral matters more than the number of alerts.","problemStats":[{"statement":"The Coalition Against Insurance Fraud states that insurance fraud steals at least USD 308.6 billion every year from American consumers and that fraud occurs in about 10% of property and casualty insurance losses.","sourceTitle":"Insurance Fraud Statistics: $308.6B Stolen Every Year","sourceUrl":"https://insurancefraud.org/fraud-stats/","year":2026}],"howItWorks":"1. **Score at first notice of loss.** Each new claim is scored in real time so honest claims can\n   go straight to processing and suspicious ones are held before payment.\n2. **Combine many signals.** Models use claim and policy history, the text of notes and\n   documents, images, and external data such as industry databases and public records.\n3. **Look across claims.** Network analysis links people, vehicles, addresses, repairers and\n   providers across claims and, through industry initiatives, across insurers.\n4. **Explain every alert.** Each alert states the scenario and the facts behind it, so a handler\n   or investigator can decide quickly whether to refer, investigate or clear it.\n5. **Keep watching.** The score is recalculated as new information arrives, and investigation\n   outcomes feed back into the models.","valueDrivers":["risk-reduction","cost-to-serve","speed","employee-productivity"],"kpis":["fraud-losses-prevented","fraud-loss-reduction","detection-rate-improvement","false-positive-reduction","interactions-handled","cost-savings","accuracy"],"indicativeValue":{"referenceOrg":"A property and casualty insurer paying USD 1 billion in claims a year","inputs":[{"key":"claimsPaid","label":"Claims paid per year","low":1000000000,"high":1000000000,"unit":"USD per year","note":"The reference insurer."},{"key":"fraudShare","label":"Share of claims losses affected by fraud","low":0.05,"high":0.1,"unit":"fraction of claims paid","note":"The high value follows the Coalition Against Insurance Fraud statement (a US figure) that fraud occurs in about 10% of property and casualty losses; the low value is an editorial assumption.","sourceUrl":"https://insurancefraud.org/fraud-stats/"},{"key":"additionalStopped","label":"Share of that fraud additionally stopped thanks to AI detection","low":0.02,"high":0.06,"unit":"fraction of fraudulent losses","note":"Editorial assumption; replace with results from a controlled pilot on your own book."}],"formula":"claimsPaid * fraudShare * additionalStopped","currency":"USD","period":"per year","resultLabel":"Additional fraudulent payments avoided","caveat":"Avoided fraudulent payments only. It leaves out investigator time saved by better referrals, the faster payment of honest claims, the deterrent effect, the cost of investigations and the cost of the platform. The share of fraud stopped varies widely by line and market."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Scoring itself is well established, often through specialised vendors. The effort is in data: joining claims, policy, payment, document and external data, labelling past investigation outcomes, and fitting alerts into the handler and investigator workflow without flooding it.","dataPrerequisites":["Claims, policy and payment history with investigation outcomes as labels","Claim notes, documents and images linked to each claim","External data such as industry fraud databases, public records and sanctions lists where lawful","Scenario definitions agreed with the special investigations unit"],"integrations":["Claims management system for real time scoring and holds on payment","Special investigations unit case management","Industry fraud data sharing schemes and databases","Document and image analysis services","Subrogation and triage models in the same claims flow"]},"implementation":{"steps":[{"title":"Agree what a good referral is","detail":"With the investigators, define the scenarios that matter per line of business and what an accepted referral looks like; that becomes the target and the main quality measure."},{"title":"Back test on closed claims","detail":"Score several years of closed claims and compare alerts with known fraud and with the current rules, at the same number of alerts investigators can handle."},{"title":"Score at first notice of loss","detail":"Move scoring to the start of the claim so honest claims are not slowed down and suspicious ones are held before payment."},{"title":"Put reasons in front of people","detail":"Show each alert with its scenario and facts inside the handler's screen and track the decision taken, so feedback is captured for every alert."},{"title":"Join industry data sharing","detail":"Organised fraud crosses insurers. Industry schemes that share claims data under clear legal bases find rings that no insurer sees alone."}],"guardrails":["The AI raises alerts; people decide on refusal, investigation and any report to authorities","Every alert carries its scenario and supporting facts, and alerts without reasons are not actioned","Protected characteristics and close proxies excluded from features, with fairness testing across customer groups","Honest customers are not delayed beyond a set time by a pending alert without a human decision","Data sharing with other insurers only under documented legal bases and agreements"],"humanInTheLoop":"Handlers and investigators review every alert and decide on the next step; no claim is refused on a score alone. Investigation outcomes are recorded and feed the models, and the special investigations unit approves changes to scenarios and thresholds.","kpisToInstrument":["Share of alerts accepted for investigation and share confirmed as fraud","Fraud stopped per period, compared with the pre AI baseline on the same lines","Alerts per investigator and time to decision per alert","Payment delay caused to claims that turned out to be honest","Complaints and appeals linked to fraud holds"],"failureModes":[{"title":"Alert floods","detail":"Too many low quality alerts teach handlers to ignore them. Tune to investigator capacity and measure acceptance, not volume."},{"title":"Bias against groups of customers","detail":"Features such as postcode that stand in for ethnicity or age treat honest customers as suspects. Test outcomes across groups and remove proxies."},{"title":"Honest customers punished for speed","detail":"Suspicious claims are held but nobody looks at them, so honest customers wait. Set a service level for every held claim."},{"title":"Models that fall behind fraudsters","detail":"Schemes change quickly, for example with AI generated documents and images. Retrain on recent outcomes and add image and document integrity checks."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Claims fraud detection by an insurer is not listed in Annex III, and point 5(b) explicitly excludes AI systems used to detect financial fraud from the credit scoring category. Point 5(c) covers only risk assessment and pricing in life and health insurance, so a fraud model becomes high risk when it also feeds those decisions, or when it is used by or on behalf of a public authority to grant, reduce, revoke or reclaim public assistance benefits (point 5(a)). Profiling and automated decisions remain subject to GDPR, including Article 22 where a claim is refused on a decision based solely on automated processing."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Calls for fairness, data governance, explainability and human oversight of AI systems used by insurers, proportionate to their impact on customers."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Shows which insurance and public benefit uses are high risk and that fraud detection is carved out of the credit scoring category."}],"controls":["Documented scenarios and thresholds, approved by the special investigations unit","Fairness testing of alert rates and outcomes across customer groups","Audit log of every alert, its reasons and the human decision taken","Service level for claims on hold because of a fraud alert","Legal basis and data sharing agreements for external and industry data"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai is not a fraud scoring engine; insurers use their own models or a specialised vendor and\nconnect them through **custom functions**. Blits.ai adds the investigation workflow around the\nscore: an **agentic workflow** that, when an alert fires, gathers the claim file, documents and\nearlier claims through **custom functions** and **SQL knowledge bases**, and has an **AI agent**\nwrite a structured alert summary with the reasons and open questions for the investigator.\n\n**Human in the loop approval** steps put holds, referrals and letters to the customer in front of a\nperson before they happen, the **audit trail** records every step of each run, and a **tool\nexecution policy** limits which functions the agent may call. **Test suites** check summaries\nagainst known cases, and the platform runs in EU or UAE regions where claims data must stay local."},"faq":[{"question":"How much fraud can AI detection stop?","answer":"Among the deployments on this page, the only quantified result is a cumulative total, not a rate: Shift Technology reports that AXA Switzerland has analysed more than 1 million claims with its real time detection and stopped over EUR 12 million in fraud. Lemonade says only that its fraud system has helped it avoid millions of dollars of potential losses. Results depend on the line, the market and how well alerts fit the investigators' workflow."},{"question":"Why score at first notice of loss rather than later in the claim?","answer":"Early scoring lets honest claims move straight to processing while suspicious ones are held before money leaves the insurer. AXA Switzerland chose detection at first notice of loss for exactly that reason, and reruns the models whenever new claim data arrives."},{"question":"Can insurers detect fraud rings that span several companies?","answer":"Yes, through industry data sharing. Member insurers of the General Insurance Association of Singapore analyse travel and motor claims together to find links between people, providers and claims that look genuine to each insurer alone."}],"related":["claims-triage-and-straight-through-processing","claims-first-notice-of-loss-agent","photo-based-damage-assessment","subrogation-opportunity-detection","application-and-identity-fraud-detection"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the insurance vertical with evidence from AXA Switzerland, Assurant, the General Insurance Association of Singapore, Lemonade and Tokio Marine & Nichido Fire, verified against the sources."},{"date":"2026-09-26","note":"Fact checked against sources. Made the EU AI Act basis precise (Annex III points 5(a), 5(b) and 5(c)), added UK GDPR, added the fraud losses prevented KPI used by the AXA Switzerland metric, clarified a failure mode, an FAQ question and the human approval wording, added source dates to two evidence records, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Replaced the PII masking claim in the Blits.ai section with the tool execution policy, since gateway masking does not cover data pulled in through functions, and corrected the FAQ to say that only the fraud stopped figure is a cumulative total."}],"slug":"claims-fraud-detection","url":"https://www.blits.ai/ai-use-cases/claims-fraud-detection","benchmarks":[{"kpi":"fraud-losses-prevented","label":"Fraud losses prevented","unit":"currency","currency":"EUR","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":12000000,"min":12000000,"max":12000000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"axa-switzerland-claims-fraud-detection","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1000000,"min":1000000,"max":1000000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"axa-switzerland-claims-fraud-detection","pooled":true}]}],"indicativeValueResult":{"low":1000000,"high":6000000},"evidence":["assurant-claims-fraud-detection","axa-switzerland-claims-fraud-detection","gia-singapore-industry-fraud-analytics","lemonade-ai-jim-claims-automation","tokio-marine-nichido-shift-claims-review"]},{"title":"AI for insurance renewal processing and customer retention","shortTitle":"Renewal and retention","seoTitle":"AI for insurance renewals and policy retention","metaDescription":"AI can speed up insurance renewals and flag customers likely to lapse, while prices stay under pricing rules. Hiscox began its Gemini underwriting model on renewals.","definition":"AI that prepares and runs the renewal cycle: it digitizes renewal submissions and changes in risk for underwriters, flags policies at risk of lapsing or leaving, prepares the renewal conversation and answers customers' renewal questions, while renewal prices stay governed by the insurer's pricing rules and fair value obligations.","aliases":["renewal automation","lapse and churn prevention","policy retention assistant"],"industries":["insurance"],"functions":["underwriting","customer-service","sales"],"patterns":["prediction-and-scoring","document-processing","conversational-agent","recommendation-and-personalization"],"channels":["email","web-chat","voice","internal-tools"],"audience":"back-office","autonomy":"copilot","adoptionStage":"emerging","segment":"distribution","problem":"In general insurance, most policies come up for renewal every year, so renewals decide how much of\nthe book an insurer keeps. In commercial lines, renewal submissions arrive in the same unstructured formats as new business,\nand underwriters must spot what changed in the risk before the renewal date, which is hard to do for\nevery account when volumes peak. In personal lines, customers can leave quietly at renewal when a\npremium rises, and renewal peaks put pressure on contact centres. Life and protection policies\nusually stay in force while premiums are paid, so there the risk is a lapse after a missed or\nfailed payment rather than a renewal decision.\n\nRetention is also regulated. In the UK, home and motor insurers may not offer renewing customers a\nprice above the equivalent new business price, and in the EU, AI used for risk assessment and\npricing of individuals in life and health insurance is high risk under the AI Act. So the\nopportunity is in better renewal processing, earlier outreach and clearer explanations, not in using\nAI to find customers who will tolerate higher prices.","problemStats":[],"howItWorks":"1. **Digitize the renewal.** Renewal submissions and updated schedules are read and compared with\n   the expiring policy, and changes in exposure are highlighted for the underwriter.\n2. **Triage the renewal book.** Low complexity renewals within appetite are prepared for automated\n   or light touch processing under existing rules; complex or deteriorating risks go to an\n   underwriter early.\n3. **Spot lapse risk.** A model flags customers likely to lapse or leave (failed payments,\n   engagement changes, large premium changes) so service teams can reach out in time.\n4. **Prepare the conversation.** The assistant drafts renewal explanations and answers customers'\n   questions about what changed and why, from the renewal documents.\n5. **Keep price decisions governed.** Renewal prices come from the insurer's pricing rules, checked\n   against fair value and renewal pricing rules; the AI does not set them.","valueDrivers":["revenue-growth","employee-productivity","customer-experience","compliance"],"kpis":["churn-reduction","automation-rate","processing-time-reduction","revenue-uplift","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A personal lines insurer with 500,000 policies up for renewal each year","inputs":[{"key":"renewals","label":"Policies up for renewal per year","low":500000,"high":500000,"unit":"policies per year","note":"The reference insurer."},{"key":"lapseRate","label":"Baseline share of policies not renewed","low":0.15,"high":0.2,"unit":"fraction of renewals","note":"Editorial assumption. Replace with your own lapse and cancellation data."},{"key":"relativeReduction","label":"Relative reduction in avoidable lapses","low":0.02,"high":0.05,"unit":"fraction of lapses","note":"Editorial assumption; no insurer on this page publishes a measured retention effect. Measure it against a control group."},{"key":"averagePremium","label":"Average annual premium","low":400,"high":600,"unit":"USD per policy","note":"Editorial assumption. Replace with your own average premium."}],"formula":"renewals * lapseRate * relativeReduction * averagePremium","currency":"USD","period":"per year","resultLabel":"Premium retained","caveat":"Premium retained, not profit, and not all lapses are worth preventing. It leaves out underwriting time saved on renewal processing, the cost of outreach and the platform, and any customer who would have renewed anyway."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Renewal digitization reuses submission intake technology; lapse models are standard machine learning. The difficult part is governance: keeping retention work separate from price optimization that regulators restrict, and proving the effect with a control group.","dataPrerequisites":["Expiring policy data and renewal submissions per account","Payment, contact and engagement history for lapse prediction","Renewal pricing rules and fair value assessments","Consent and preference data for outreach"],"integrations":["Policy administration and renewal processing","Underwriting workbench","Billing and payment systems","CRM and contact centre for outreach and handover"]},"implementation":{"steps":[{"title":"Separate the two problems","detail":"Treat commercial renewal processing (underwriting efficiency) and personal lines retention (customer outreach) as separate projects with different owners and controls."},{"title":"Start with renewal intake in commercial lines","detail":"Reuse submission extraction on renewal documents and show underwriters what changed against the expiring terms, starting with lines that have high renewal volumes."},{"title":"Build lapse prediction with outreach, not pricing","detail":"Use lapse scores only to decide who gets a service call, a payment reminder or a policy review, never to adjust the renewal price."},{"title":"Test with a control group","detail":"Hold out a random share of flagged customers to measure the true retention effect before scaling outreach."},{"title":"Review for fair value","detail":"Have pricing and compliance confirm that nothing in the retention process changes renewal price by tenure or propensity to shop around where rules forbid it."}],"guardrails":["Lapse and churn scores never feed renewal pricing","Renewal prices only from governed pricing rules, checked against renewal pricing requirements","Outreach limited to customers who consented to contact, with vulnerability checks","Automated renewals only for low complexity risks within written rules","Explanations of premium changes drawn from the actual renewal documents"],"humanInTheLoop":"Underwriters review every renewal outside the automated cohort and every material change in risk. Service staff make retention calls with AI prepared context, and pricing and compliance own the rules that separate retention outreach from price setting.","kpisToInstrument":["Share of renewals reviewed before the renewal date","Retention rate for flagged customers versus a random control group","Underwriting time per renewal","Complaints about renewal prices and communications","Fair value and renewal pricing test results"],"failureModes":[{"title":"Retention models become price optimization","detail":"Scores that predict who will not shop around drift into pricing decisions. Keep the systems and teams separate and audit the data flows."},{"title":"Rolled over risks nobody looked at","detail":"Automated renewal processes a risk that changed materially. Compare against the expiring policy and route changes to an underwriter."},{"title":"Retention effect that was never there","detail":"Outreach goes to customers who would have renewed anyway. Measure against a control group."},{"title":"Pushy outreach to vulnerable customers","detail":"Retention calls pressure customers in financial difficulty. Check vulnerability signals first."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Renewal intake for commercial lines and outreach are not listed in Annex III. Renewal risk assessment or pricing for life or health insurance of natural persons is high risk under point 5(c), and so is a lapse score that feeds those decisions; a lapse score used only to decide who gets a service call is not listed. Customer facing renewal assistants carry the Article 50(1) duty to tell people they are interacting with an AI system, unless that is obvious from the context."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","eu-idd"],"guidance":[{"title":"PS21/5: General insurance pricing practices market study, feedback to CP20/19 and final rules","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publication/policy/ps21-5.pdf","note":"Final rules requiring firms to offer UK home and motor insurance renewal prices no greater than the equivalent new business price, however the price is set."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(c) covers AI systems for risk assessment and pricing in relation to natural persons in life and health insurance, which includes those decisions when they are made at renewal."},{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Published August 2025 for national supervisors. Asks insurers to treat customers fairly across the AI lifecycle and value chain, monitor outcomes with fairness and non discrimination metrics, and points to EIOPA's 2023 Supervisory Statement on Differential Pricing Practices."}],"controls":["Documented separation between retention scoring and pricing","Control group design and results for every retention programme","Fair value and renewal pricing tests signed off by pricing and compliance","Vulnerability screening before outreach","Audit trail of automated renewals and the rules applied"],"incidents":[{"title":"Suckers List: How Allstate's Secret Auto Insurance Algorithm Squeezes Big Spenders","url":"https://themarkup.org/allstates-algorithm/2020/02/25/car-insurance-suckers-list","note":"The Markup and Consumer Reports analysed a \"retention model\" Allstate proposed in Maryland and found it would have charged big spenders more rather than following risk; Maryland rejected it as discriminatory and Allstate called the reporting inaccurate. It shows why retention signals must stay out of pricing."}]},"blitsAi":{"howToBuild":"On Blits.ai commercial renewal intake is an **agentic workflow** that reads renewal documents from\nthe **email channel** or an API, compares them with the expiring policy retrieved through a **SQL\nknowledge base** or **custom functions**, and returns a change summary as **structured output** for\nthe underwriter. **Human in the loop confirmation** keeps every non routine renewal with an\nunderwriter, who approves or rejects it.\n\nFor personal lines, **agentic tasks** can watch for conditions such as a failed payment before\nrenewal and send an outbound message through the **email channel**, with consent checks in the\nflow. Customers who reply or get in touch on **web chat, WhatsApp, SMS or voice** reach a renewal\nassistant that retrieves their renewal documents through **custom functions** and answers what\nchanged and why, with **human handover** to retention staff. The platform does not set prices:\nthe **tool execution policy** keeps pricing tools out of the agent's reach. Each workflow's\n**run data** can be downloaded, so you can compare treated and control groups in your own\nanalysis outside the platform."},"faq":[{"question":"Can AI improve insurance retention without breaking pricing rules?","answer":"Yes, if it works on service rather than price: earlier outreach, fixing payment problems, explaining changes clearly and processing renewals on time. In the UK, renewal prices for home and motor may not exceed the equivalent new business price, so retention models must stay out of pricing."},{"question":"Are there published results?","answer":"Few so far. Hiscox's generative AI lead underwriting model started with renewals of existing sabotage and terrorism risks, Nsure.com's copilot helps customers review renewal offers, and Microsoft's customer story on Zurich says user feedback shows its sales copilot improves sales and retention ratios, but none publish a measured retention effect."},{"question":"Is AI in renewals high risk under the EU AI Act?","answer":"Only where it assesses risk or sets prices for life or health insurance of individuals, which is high risk under Annex III point 5(c). Commercial renewal processing and service outreach are not listed, but a customer facing renewal assistant must tell people they are talking to an AI unless that is obvious from the context (Article 50(1))."}],"related":["insurance-policy-servicing-agent","commercial-underwriting-submission-triage","insurance-pricing-and-actuarial-copilot","churn-prediction-and-retention-offers","insurance-broker-and-agent-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from insurer press releases, vendor case studies and regulatory sources verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added Insurance Distribution Directive to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: narrowed the UK renewal pricing rule to home and motor, removed unsourced claims from the problem, attributed the Zurich retention remark to Microsoft, sharpened the EU AI Act basis and guidance notes, added UK GDPR, aligned the Blits.ai build with the feature inventory, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: limited Blits.ai outreach to outbound email (the only outbound channel in the feature inventory), credited tool restriction to the tool execution policy only, said control group comparison happens outside the platform, narrowed renewal to general insurance with life described as lapse, added the Article 50(1) obvious from context exception, softened the meta description."}],"slug":"insurance-renewal-and-retention","url":"https://www.blits.ai/ai-use-cases/insurance-renewal-and-retention","benchmarks":[],"indicativeValueResult":{"low":600000,"high":3000000},"evidence":["hiscox-generative-ai-lead-underwriting","nsure-friendly-john-copilot","zurich-copilot-for-sales-crm-updates"]},{"title":"AI for IT incident triage and root cause analysis (AIOps)","shortTitle":"AIOps incident triage","seoTitle":"AI incident triage and root cause analysis","metaDescription":"AI groups alerts into one incident, routes it and ranks likely root causes for an engineer to confirm. Meta reports 42% top five accuracy in backtests.","definition":"AI that turns a flood of monitoring alerts into one probable incident, routes it to the right team, proposes likely root causes and remediation from runbooks and past incidents, and drafts the stakeholder updates and the post incident review, while an engineer authorizes every change.","aliases":["AIOps","incident management copilot","SRE copilot","AI root cause analysis","alert correlation"],"industries":["cross-industry","banking","technology","telecommunications","payments"],"functions":["it-and-engineering","operations","risk-management"],"patterns":["anomaly-detection","classification-and-routing","summarization","rag-knowledge-assistant","agentic-workflow"],"channels":["internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"When a critical system degrades, the monitoring estate fires a burst of alerts across\napplications, infrastructure and dependent services. Engineers spend the first part of the\nincident working out which alerts belong together, which team owns the problem and what changed,\nwhile business and customer teams ask for updates. Mizuho and IBM described the pattern plainly:\nwhen an error is detected, operators receive an influx of messages and reports, which makes it\nhard to pinpoint the cause and delays recovery.\n\nTime to recover is not only a cost question. For regulated firms an outage of a critical service\ncan become a reportable event. DORA requires EU financial entities to detect, manage, record and\nclassify ICT related incidents, to report major ones and to review them afterwards. NIS2 sets\nincident reporting duties for essential and important entities, including telecom operators. APRA\nCPS 230 expects Australian regulated entities to keep critical operations within tolerance levels\nthrough severe disruptions and to notify APRA of tolerance breaches. The knowledge that shortens an incident\n(runbooks, past post incident reviews, recent changes) exists, but it is scattered and nobody has\ntime to search it at 3 a.m.","problemStats":[],"howItWorks":"1. **Correlate.** Alerts, logs, traces and change events are grouped into one probable incident\n   using topology and timing, so responders see one problem instead of a stream of separate alerts.\n2. **Route.** A classifier or a set of team agents decides which team owns the incident, based\n   on service ownership and past incidents, and pages them with the correlated evidence.\n3. **Suggest causes.** The AI ranks recent changes and known failure patterns as likely root\n   causes and retrieves the matching runbook steps and similar past incidents, with links so the\n   engineer can verify each suggestion.\n4. **Propose, do not execute.** Remediation is proposed as a concrete command or change; a\n   policy layer checks it, and an engineer confirms it before anything runs.\n5. **Communicate.** The AI drafts status updates for stakeholders at a set cadence from the\n   incident timeline, for the incident commander to approve.\n6. **Learn.** After recovery it drafts the post incident review from the timeline, chat and\n   changes, and proposes follow up actions and runbook updates.","valueDrivers":["speed","risk-reduction","employee-productivity","customer-experience"],"kpis":["mttr-reduction","accuracy","detection-rate-improvement","time-saved-per-task"],"indicativeValue":{"referenceOrg":"A bank with 120 major IT incidents a year on customer facing services","inputs":[{"key":"incidents","label":"Major incidents per year","low":120,"high":120,"unit":"incidents per year","note":"The reference organization. Replace with your own count of priority 1 and 2 incidents."},{"key":"hoursToRestore","label":"Average hours from detection to mitigation","low":2,"high":4,"unit":"hours per incident","note":"Editorial assumption. Replace with your own mean time to restore."},{"key":"reduction","label":"Reduction in time to mitigate","low":0.1,"high":0.25,"unit":"fraction of time","note":"Conservative against the evidence on this page (Microsoft reports a 38% time to mitigate reduction for one team; TD Bank's vendor reports 20% faster response), because those are the best early results."},{"key":"responders","label":"Engineers engaged per incident","low":4,"high":8,"unit":"engineers","note":"Editorial assumption including the incident commander and service owners."},{"key":"hourlyCost","label":"Fully loaded engineer hour","low":80,"high":120,"unit":"USD per hour","note":"Editorial assumption. Replace with your own rate."},{"key":"reviewHoursSaved","label":"Hours saved drafting each post incident review and status updates","low":2,"high":5,"unit":"hours per incident","note":"Editorial assumption; the review still needs the owning team's analysis."}],"formula":"incidents * (hoursToRestore * reduction * responders + reviewHoursSaved) * hourlyCost","currency":"USD","period":"per year","resultLabel":"Engineering time released during and after major incidents","caveat":"Engineering time only. It leaves out the largest effect, the revenue, customer harm and regulatory exposure avoided by restoring service faster, which depends on the service and is best estimated per critical business service against its impact tolerance."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Summaries and drafted updates are quick wins. Correlation and root cause suggestions depend on a service map, clean change records and labeled past incidents; without them, suggestions are generic. Automated remediation is a separate, higher risk step that most firms defer.","dataPrerequisites":["A current service catalog with owners and dependencies","Alerts, logs and traces from the observability platform","Change records and deployment history linked to services","Past incidents and post incident reviews with the confirmed root cause","Runbooks with an owner and a review date"],"integrations":["Observability and monitoring platforms","IT service management tool for incidents, problems and changes","Paging and on call scheduling","ChatOps in Microsoft Teams or Slack","Code repositories and deployment pipelines for change history"]},"implementation":{"steps":[{"title":"Start with the write ups, not the fixes","detail":"The lowest risk value is drafting stakeholder updates and post incident reviews from the incident timeline. It earns trust with engineers and builds the labeled incident history the later steps need."},{"title":"Clean the service map and change feed","detail":"Correlation and routing are only as good as ownership data and change records. Fix the services with the most incidents first."},{"title":"Add root cause suggestions with evidence","detail":"Show a short ranked list, each item with the change, the log excerpt or the past incident that supports it. Meta narrows its suggestions to the top five code changes, measures the ranker by backtesting on historical investigations and holds back low confidence answers."},{"title":"Measure on history before going live","detail":"Replay past incidents and measure how often the true cause was in the suggestions and how often routing picked the right team. Publish the number to the engineers who will use it."},{"title":"Keep remediation behind confirmation","detail":"Let the AI propose commands, pass them through a policy layer (no global restarts at peak, two person approval for high impact changes) and require an engineer to confirm. Google's SRE tooling logs what the AI proposed and what the human approved."},{"title":"Close the loop","detail":"Feed confirmed root causes and runbook fixes from each review back into the knowledge base, so the next incident starts with better context."}],"guardrails":["No change to production without explicit confirmation by an authorized engineer","A policy layer that blocks or escalates high impact commands regardless of what the AI proposes","Every suggestion shows its evidence (change, log, past incident) so it can be verified in seconds","Low confidence suggestions are held back rather than shown","Full logging of AI proposals, human decisions and timestamps into the incident timeline","Secrets and customer data masked before logs reach a model"],"humanInTheLoop":"The incident commander owns the incident, decides on customer communication and approves every status update. Engineers authorize every remediation. The owning team signs off the post incident review and its actions; the AI drafts, it does not conclude.","kpisToInstrument":["Mean time to detect, to engage the right team and to mitigate, before and after, per service","Share of incidents where the confirmed root cause was among the AI suggestions","Routing accuracy (incidents that stayed with the first team paged)","Alerts per incident after correlation","Time from resolution to a published post incident review"],"failureModes":[{"title":"Plausible but wrong root cause","detail":"An engineer anchors on a confident suggestion and loses time. Show evidence per suggestion, hold back low confidence ones and track the hit rate openly."},{"title":"Automation that amplifies the outage","detail":"An automated fix runs against the wrong target or at the wrong time. Keep confirmation and a policy layer in front of every mutation, and rehearse in game days."},{"title":"Stale runbooks retrieved with authority","detail":"The AI surfaces an outdated procedure. Give every runbook an owner and a review date and prefer recent post incident reviews."},{"title":"Noise moved, not removed","detail":"Correlation merges unrelated alerts or hides a second incident. Let responders split incidents easily and review correlation quality weekly."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"An internal tool that supports engineers on IT incidents; it is not a use listed in Annex III and makes no decisions about people. Annex III point 2 covers AI used as a safety component in the management and operation of critical digital infrastructure, and recital 55 limits safety components to systems that directly protect the physical integrity of that infrastructure or the health and safety of persons and property. A triage copilot that proposes causes and fixes to engineers does not normally do that, but operators of critical digital infrastructure (cloud, data centers, telecom networks) should confirm this for their own design."},"regulations":["dora","apra-cps-230","nist-ai-rmf","iso-42001","gdpr","nis2"],"guidance":[{"title":"Digital Operational Resilience Act (Regulation (EU) 2022/2554)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2022/2554/oj","note":"Financial entities must detect, manage, record and classify ICT related incidents (Articles 17 and 18), report major ones (Article 19) and review major incidents afterwards (Article 13); AI drafted timelines and reviews become part of that record."},{"title":"Operational risk management (CPS 230)","issuer":"Australian Prudential Regulation Authority","region":"asia-pacific","url":"https://www.apra.gov.au/operational-risk-management","note":"Regulated entities must keep critical operations within tolerance levels through severe disruptions and notify APRA of operational risk incidents likely to have a material impact and of disruptions to a critical operation outside tolerance, which is the measure AIOps value should be judged against."}],"controls":["Written limits on what the AI may propose and what always needs a named approver","Incident timeline that records AI suggestions, human decisions and timestamps","Backtest results per release of the model or prompts before engineers rely on it","Periodic review of suggestion hit rate and routing accuracy by the SRE or operations lead","Inventory entry for the AIOps system with an owner and a review date"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered **via API** by the incident management or\nmonitoring tool when a major incident opens. The workflow's **custom functions** (REST calls)\nfetch the correlated alerts, recent changes and service ownership, and an **AI agent** with a\n**knowledge base** of runbooks and past post incident reviews, retrieved with **hybrid retrieval**,\nproposes likely causes with links to the evidence. Any action above a set threshold goes\nthrough **human in the loop confirmation**, and the **tool execution policy** limits which tools\nthe agent may call at all.\n\nResponders talk to the agent in **Microsoft Teams** or Slack, where it drafts status updates and,\nafter recovery, the post incident review. **Run history with a full audit trail** keeps every\nsuggestion and approval for the incident record, **PII masking** at the gateway masks personal\ndata in what responders type, and **test suites** replay past incidents before each change goes\nlive. Logs and alerts fetched by custom functions do not pass the gateway, so mask secrets and\ncustomer data at the source or inside the custom function before they reach the agent. The\nplatform is model agnostic, so the operations team can choose the model per agent."},"faq":[{"question":"How accurate is AI root cause analysis today?","answer":"It helps, but it is not an oracle. Meta reports that in backtesting 42% of investigations had the root cause in the top five suggested code changes, and Microsoft reports 90% accuracy for its team triage agents in early results. Treat suggestions as a ranked shortlist with evidence, not an answer."},{"question":"Should the AI fix incidents on its own?","answer":"Not at first, and not for high impact changes. Keep the AI proposing and an engineer confirming, with a policy layer in between; Google's SRE tooling forces a confirmation step and logs what the AI proposed and what the human approved. Automate only narrow, reversible fixes after a track record."},{"question":"Where does the value show up first?","answer":"In routing and write ups. Getting the incident to the right team faster and drafting updates and post incident reviews saves time on every incident, while root cause suggestions improve as the labeled incident history grows."},{"question":"Does this matter to regulators?","answer":"Yes for financial firms. DORA in the EU sets rules for managing, recording and reporting ICT related incidents, and CPS 230 in Australia expects critical operations to stay within tolerance levels, so the AI's suggestions and the human decisions belong in the incident record."}],"related":["it-service-desk-resolution-agent","network-fault-triage-copilot","developer-coding-assistant","ai-model-inventory","support-knowledge-article-generation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog, made industry neutral and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added NIS2 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: tightened the DORA, NIS2 and CPS 230 statements to the legal text, clarified the EU AI Act basis (Annex III point 2, recital 55), attributed the Mizuho quote to Mizuho and IBM, corrected the TD Bank, Microsoft and Mizuho evidence details; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: tracked time to repair instead of cycle time so the Microsoft time to mitigate result shows in the benchmarks; dated the TD Bank claim to 2024 from the archived customer story; narrowed the PII masking statement to the gateway; removed the unsourced alert count and aligned the CPS 230 wording and spelling."}],"slug":"aiops-incident-triage","url":"https://www.blits.ai/ai-use-cases/aiops-incident-triage","benchmarks":[{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":3,"nUpTo":0,"median":90,"min":42,"max":98,"byClaimant":{"organization":2,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"mizuho-generative-ai-event-detection","pooled":true},{"id":"microsoft-azure-triangle-incident-triage","pooled":true},{"id":"meta-ai-assisted-root-cause-analysis","pooled":true}]},{"kpi":"mttr-reduction","label":"Time to repair reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":29,"min":20,"max":38,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"microsoft-azure-triangle-incident-triage","pooled":true},{"id":"td-bank-aiops-observability","pooled":true}]},{"kpi":"detection-rate-improvement","label":"Detection improvement","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":25,"min":25,"max":25,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"td-bank-aiops-observability","pooled":true}]}],"indicativeValueResult":{"low":26880,"high":187200},"evidence":["google-sre-gemini-cli-incident-response","meta-ai-assisted-root-cause-analysis","microsoft-azure-triangle-incident-triage","mizuho-generative-ai-event-detection","td-bank-aiops-observability"]},{"title":"AI for ledger and payment reconciliation","shortTitle":"Ledger and payment reconciliation","seoTitle":"AI reconciliation for bank ledgers and payments","metaDescription":"AI matches statements, settlement files and ledger entries and routes the breaks it cannot clear to an operator. See where banks and funds already run it.","definition":"AI that matches entries across nostro and vostro statements, card and scheme settlement files, the general ledger and suspense accounts, proposes matches and clearing journals, and routes only the genuine breaks to an operator with a plain language explanation.","aliases":["AI reconciliation","nostro reconciliation automation","intelligent transaction matching","exception management for reconciliations"],"industries":["banking","payments","capital-markets","cross-industry","wealth-and-asset-management","government"],"functions":["finance-and-accounting","operations"],"patterns":["agentic-workflow","anomaly-detection","document-processing"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"back-office","problem":"Every bank reconciles its own books against the outside world many times a day: nostro\nstatements against expected cash flows, card and scheme settlement files against authorised\ntransactions, clearing and suspense accounts against the general ledger. Rule based matching\nengines handle the clean cases, but timing differences, partial references, split and bulked\npayments, bank charges and FX conversions leave a steady stream of breaks that people clear by\nhand, often in spreadsheets.\n\nThose breaks are where the cost and the risk sit. Items ageing in suspense accounts can distort\nthe balance sheet, tie up capital and liquidity, hide fraud and lead to audit findings. Writing a new\nmatching rule for every new pattern is slow, so operations teams grow with volume instead of\nstaying flat.","problemStats":[],"howItWorks":"1. **Ingest every feed.** Statements (MT940, camt.053), settlement files, ledger extracts and\n   remittance advices arrive in one pipeline; document AI reads the unstructured ones.\n2. **Match beyond the rules.** Deterministic rules clear exact matches first. A learned matching\n   layer then proposes one to one, one to many and many to many matches using fuzzy references,\n   amounts within tolerance, value dates and FX, each with a confidence score.\n3. **Explain and propose.** For each proposed match or break the agent states why (for example\n   \"bank charge of 15 EUR deducted by the correspondent\") and drafts the clearing journal or\n   the adjustment.\n4. **Route the exceptions.** Low confidence items and anything above a materiality threshold go\n   to an operator queue with the evidence attached. Retrieval over prior resolutions suggests how\n   similar breaks were cleared before.\n5. **Learn under control.** Operator decisions feed back as candidate rules or training data,\n   which a reconciliation owner approves before they change production matching.","valueDrivers":["cost-to-serve","risk-reduction","speed","employee-productivity"],"kpis":["automation-rate","processing-time-reduction","hours-saved","error-reduction","productivity-gain"],"indicativeValue":{"referenceOrg":"A mid sized bank that clears 300,000 reconciliation breaks by hand each year","inputs":[{"key":"manualItems","label":"Breaks cleared manually per year","low":300000,"high":300000,"unit":"items per year","note":"The reference bank. Replace with the exception count from your reconciliation platform."},{"key":"automatedShare","label":"Share of those breaks the AI clears or pre clears","low":0.3,"high":0.6,"unit":"fraction of manual items","note":"Editorial assumption; no verified public benchmark for AI match rates was found. Replace with a pilot result on your own data."},{"key":"minutesPerItem","label":"Minutes an operator spends per break","low":6,"high":12,"unit":"minutes per item","note":"Editorial assumption, replace with your own time study."},{"key":"costPerHour","label":"Fully loaded operations cost per hour","low":35,"high":60,"unit":"USD per hour","note":"Editorial assumption for a blended onshore and offshore operations team."}],"formula":"manualItems * automatedShare * minutesPerItem / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Manual reconciliation effort avoided","caveat":"Labour only. It leaves out the value of fewer aged items in suspense (capital, liquidity and fraud exposure), fewer audit findings, and the cost of the platform and the integration work."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The matching logic is well understood; the work is data. Feeds arrive in many formats and timings, references are inconsistent across systems, and every automated journal needs a control design that auditors accept.","dataPrerequisites":["Twelve months of history of matched items and cleared breaks, with the resolution chosen","Clean static data for accounts, counterparties and correspondent banks","Documented tolerances and materiality thresholds per reconciliation"],"integrations":["Reconciliation platform or matching engine","General ledger and subledgers","SWIFT or bank statement feeds (MT940, MT950, camt.053)","Card scheme and acquirer settlement files","Workflow or case tool for exception queues"]},"implementation":{"steps":[{"title":"Start with the reconciliations that hurt most","detail":"Rank reconciliations by manual breaks per month and by aged value in suspense. Pick two or three with high volume and clear ownership, such as a card settlement or a busy nostro."},{"title":"Baseline the current rules","detail":"Measure what the existing engine already matches so the AI is credited only for the increment. Many teams find quick wins by fixing static data before any model is involved."},{"title":"Run in shadow mode","detail":"Let the AI propose matches and journals next to the operators for several cycles and compare. Only promote a match type to automatic once its precision on your data is proven."},{"title":"Set materiality and maker checker thresholds","detail":"Agree with finance and audit which proposals may post automatically and which need a second person, by amount, account type and confidence."},{"title":"Close the learning loop","detail":"Capture why operators accept or reject a proposal and review those patterns monthly with the reconciliation owner before they become new rules."}],"guardrails":["No automatic posting above the materiality threshold; those journals need a maker and a checker","Every match and journal carries the evidence and confidence it was based on","Tolerances and thresholds are configuration owned by finance, not learned by the model","Unmatched items are never forced to clear; low confidence always goes to a person"],"humanInTheLoop":"Operators own the exception queue and approve every journal above the threshold. The reconciliation owner approves new match types and tolerances, and internal audit samples automated matches each quarter.","kpisToInstrument":["Auto match rate per reconciliation, on top of the rule engine baseline","Precision of automated matches from a monthly sample","Number and value of items aged over 30 days in suspense","Operator minutes per break","Audit findings related to reconciliations"],"failureModes":[{"title":"False matches that hide a real break","detail":"A plausible but wrong match clears an item that should have been investigated. Sample automated matches and keep tight tolerances on amount."},{"title":"Drift after an upstream change","detail":"A new file format or reference convention quietly lowers match quality. Monitor match rate per feed and alert on sudden drops."},{"title":"Credit for work the rules already did","detail":"Benefits are overstated because the baseline was not measured. Report the increment over the existing engine."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Matching entries between internal financial records is not a use listed in Annex III and is not a practice prohibited by Article 5. Operators knowingly use an internal AI tool, so no Article 50(1) disclosure is needed. If a generative model drafts the explanations or journals, the provider of that system may have to mark its output as AI generated under Article 50(2). The AI literacy duty of Article 4 applies to the bank as deployer."},"regulations":["eu-ai-act","dora","apra-cps-230","iso-42001"],"guidance":[{"title":"Principles for effective risk data aggregation and risk reporting (BCBS 239)","issuer":"Basel Committee on Banking Supervision","region":"global","url":"https://www.bis.org/publ/bcbs239.htm","note":"Principle 3 expects risk data to be reconciled with the bank's sources, including accounting data where appropriate, and aggregated on a largely automated basis."},{"title":"SS1/23 Model risk management principles for banks","issuer":"Bank of England, Prudential Regulation Authority","region":"europe","url":"https://www.bankofengland.co.uk/prudential-regulation/publication/2023/may/model-risk-management-principles-for-banks-ss","note":"Applies to UK banks, building societies and PRA designated investment firms with internal model approval for regulatory capital. A learned matching model that suggests or posts journals falls under its model definition and belongs in the model inventory with validation and monitoring."},{"title":"SR 26-2 Revised Guidance on Model Risk Management","issuer":"Board of Governors of the Federal Reserve System, OCC and FDIC","region":"north-america","url":"https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm","note":"Issued on 17 April 2026, it supersedes and replaces SR 11-7 and asks for a risk based approach tailored to each bank's model risk profile, size and complexity. The letter says it is most relevant to banking organizations with over $30 billion in total assets regulated by the Federal Reserve. A US bank that treats its learned matching model as a model under its policy validates and monitors it on this basis."}],"controls":["Model inventory entry with an owner, validation and drift monitoring per reconciliation","Maker checker on journals above the materiality threshold","Immutable log of every proposal, the evidence used and who approved it","Quarterly sample of automated matches reviewed by an independent team"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that runs on a schedule or is triggered through the\nAPI when a new file arrives. **Custom functions** read the feeds and the ledger through REST calls\nand SQL queries, **SQL knowledge bases** let the agent query reconciliation tables directly, the\nagent proposes matches and journals within a **tool execution policy**, and **human in the loop\napproval** holds any journal above the configured threshold until an operator approves or\nrejects it.\n\nPrior break resolutions and reconciliation procedures sit in the **knowledge base** with hybrid\nretrieval, so the explanation for each exception cites how similar items were cleared. Each run\nkeeps a **full audit trail**, **test suites** replay known breaks against the workflow before any\nchange goes live, and **monitors** run scheduled checks with email or webhook alerts when an\nexpectation fails. The platform is model agnostic and can run in EU or UAE regions for data\nresidency."},"faq":[{"question":"How is AI reconciliation different from a rule based matching engine?","answer":"Rules clear exact and near exact matches and should stay. AI adds matching on partial references, amounts within tolerance, one to many and many to many combinations, and a plain language explanation of each break, so fewer items reach an operator and those that do arrive with a suggested resolution."},{"question":"Can the AI post clearing journals on its own?","answer":"Only within limits agreed with finance and audit. A common design lets low value, high confidence matches post automatically and keeps a maker and a checker on anything above a materiality threshold, with every proposal logged with its evidence."},{"question":"Who is already using AI for reconciliation?","answer":"Public evidence is still thin on measured results. National Bank of Greece in Cyprus consolidated four reconciliation systems onto an AI enabled platform in 2026, and Comrade Trustee Services in Papua New Guinea went live on the same vendor platform, which says processing time fell from up to eight hours to under five minutes. A 2024 US federal AI inventory entry listed a World Food Programme machine learning tool for reconciling cash transfers at the implementation and assessment stage. (Ginnie Mae uses machine learning to find anomalies and exceptions in subledger transaction data before financial reporting, which is a related data quality use, not reconciliation matching.)"}],"related":["payment-investigations-and-exceptions","fee-and-interest-leakage-detection","supplier-invoice-processing","chargeback-and-representment","regulatory-report-assembly","treasury-cash-flow-forecasting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: industries now include wealth and asset management and government, where its evidence comes from; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed the unsupported list of four systems from the National Bank of Greece record, tightened the Ginnie Mae and World Food Programme summaries, made the EU AI Act basis cite Articles 4, 5 and 50, corrected the BCBS 239 and SS1/23 notes, aligned the Blits.ai build with the feature inventory, added Comrade Trustee Services to the FAQ, and added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: replaced the withdrawn SR 11-7 with its successor SR 26-2 as guidance, removed Ginnie Mae from the meta description because its source describes anomaly detection rather than reconciliation, dated the World Food Programme claim to its 2024 inventory entry, softened the EU AI Act Article 50 basis, and added the 2025 federal inventory and Ernst & Young to the Ginnie Mae record."},{"date":"2026-09-27","note":"Review fixes: removed the absolute \"flags only the real breaks\" claim and the unsupported ties to National Bank of Greece Cyprus and Comrade Trustee Services from the meta description, and removed Ginnie Mae from the \"who is already using AI for reconciliation\" FAQ answer."}],"slug":"ledger-and-payment-reconciliation","url":"https://www.blits.ai/ai-use-cases/ledger-and-payment-reconciliation","benchmarks":[],"indicativeValueResult":{"low":315000,"high":2160000},"evidence":["comrade-trustee-services-ai-reconciliation","ginnie-mae-subledger-data-quality-machine-learning","national-bank-of-greece-cyprus-ai-reconciliation","world-food-programme-darts-cash-transfer-reconciliation"]},{"title":"AI for legacy code modernization","shortTitle":"Legacy code modernization","seoTitle":"AI for COBOL and legacy code modernization","metaDescription":"AI explains legacy code and drafts specifications so engineers can modernize with less risk. Morgan Stanley says its tool saved about 280,000 hours in five months.","definition":"AI that reads legacy code such as COBOL, PL/I or old Java, explains what each program does, maps its data flows and dependencies, drafts the equivalent modern code or specification, and generates the regression tests needed to prove the new system behaves like the old one.","aliases":["COBOL modernization with AI","mainframe modernization","legacy code translation","code migration agent"],"industries":["cross-industry","banking","capital-markets","automotive","technology"],"functions":["it-and-engineering"],"patterns":["code-generation","summarization","agentic-workflow"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Core systems in many large organizations still run on code written decades ago, often in COBOL or\nproprietary languages, with little documentation. The evidence on this page shows both: a global\nbank whose foundational systems were designed decades ago, at a time when COBOL talent is becoming\nscarce, and Toyota Motor Europe, whose applications in a proprietary language depend on developers\nwho are retiring. Every change is slow and risky, and a full rewrite is hard to plan because nobody\ncan say with confidence what the old system actually does.\n\nThe hard part of modernization was never typing the new code. It is comprehension (what does\nthis program do, which rules are buried in it, what depends on it) and proof (does the new\nversion behave the same on real data). Three of the four deployments on this page use AI for\ncomprehension: Morgan Stanley and Toyota Motor Europe turn code into readable specifications and\ndocumentation, and at the GFT bank AI also generated test scenarios and made the converted code more\nreadable, while deterministic tools did the conversion. Amazon went furthest, using a code\ntransformation agent to help migrate applications to a newer Java version.","problemStats":[],"howItWorks":"1. **Inventory and dependency mapping.** Tools parse the estate to find programs, copybooks,\n   jobs, data stores and the calls between them, so work can be split into modules.\n2. **Explain the code.** A model generates technical and business documentation per program:\n   what it does, its inputs and outputs, and the business rules and conditions it applies.\n3. **Review by the remaining experts.** Subject matter experts check a sample of the\n   documentation against the code and correct it. Their corrections improve the next batch.\n4. **Convert or rewrite.** Deterministic converters or engineers produce the modern code from\n   the specification, with AI assistance for readability and idiomatic structure.\n5. **Prove equivalence.** AI generates regression tests and test data from the documented rules;\n   old and new systems run in parallel on production like data until the differences are\n   explained.\n6. **Cut over in controlled steps.** Each module moves through the normal change process, with\n   traceability from the legacy module to the new service.","valueDrivers":["speed","cost-to-serve","risk-reduction","employee-productivity"],"kpis":["hours-saved","cost-savings","productivity-gain","processing-time-reduction"],"indicativeValue":{"referenceOrg":"A bank with 5 million lines of legacy code in scope for modernization","inputs":[{"key":"linesOfCode","label":"Lines of legacy code in scope","low":5000000,"high":5000000,"unit":"lines of code","note":"The reference organization."},{"key":"hoursPerThousandLines","label":"Analysis, documentation and test design hours per thousand lines, done manually","low":10,"high":20,"unit":"hours per thousand lines","note":"Editorial assumption. Replace with the estimate from your own modernization plan."},{"key":"aiReduction","label":"Share of that effort the AI removes","low":0.3,"high":0.5,"unit":"fraction of effort","note":"Editorial assumption, conservative against the evidence on this page (Morgan Stanley reports roughly 280,000 hours saved on nine million lines, about 31 hours per thousand lines)."},{"key":"hourlyCost","label":"Blended cost of an engineering hour","low":80,"high":120,"unit":"USD per hour","note":"Editorial assumption. Replace with your own rates, including specialist contractors."}],"formula":"linesOfCode / 1000 * hoursPerThousandLines * aiReduction * hourlyCost","currency":"USD","period":"over the program","resultLabel":"Analysis and test design effort avoided","caveat":"Covers comprehension, documentation and test design only. It leaves out conversion, parallel running, infrastructure and license savings after decommissioning, and the risk reduction of having documented systems, which is often the larger benefit."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The model is rarely the hardest part. Modernization programs touch the most critical systems, need parallel running and, in regulated firms, early engagement with supervisors, and depend on experts who are scarce. AI shortens comprehension and testing; it does not remove the program.","dataPrerequisites":["Complete source code, including copybooks, job control and configuration","Access to the remaining experts for review of generated documentation","Production like test data, masked where it contains personal data","An agreed target architecture and coding standards"],"integrations":["Source repositories and mainframe code management","Static analysis and dependency mapping tools","Model access in a tenant with zero retention and suitable data location","Test automation and parallel run comparison tooling"]},"implementation":{"steps":[{"title":"Start with comprehension, not conversion","detail":"Pick one module and have the AI document it. Let the experts grade the documentation. This tells you quickly how reliable the model is on your code and languages."},{"title":"Split the estate into modules","detail":"Use dependency mapping to find boundaries where a module can move on its own, and sequence the program by business risk and dependency."},{"title":"Choose the conversion approach per module","detail":"Deterministic conversion keeps logic identical but produces unidiomatic code; rewriting from a specification gives better code but more risk. Many programs combine both."},{"title":"Generate and run the tests","detail":"Turn the documented rules into regression tests and compare old and new outputs on the same data. Differences are investigated, not waved through."},{"title":"Keep traceability","detail":"Link every new service back to the legacy programs and documented rules it replaces, for auditors and for the next change."},{"title":"Retire the old code","detail":"Plan decommissioning from the start. Savings only arrive when the legacy runtime is switched off."}],"guardrails":["No direct cutover from AI output; every module passes testing and parallel running","Expert review of generated documentation before it is used as a specification","Code processed only in a tenant with zero retention and agreed data location","Traceability from each legacy module to its replacement","Change approval by the owners of the business process, not only by IT"],"humanInTheLoop":"Experts validate the documentation, engineers own the new code, and business owners sign off on behavioral equivalence after parallel running. Human sign off at the cutover gate is not optional.","kpisToInstrument":["Documentation accuracy on expert reviewed samples","Hours per module for analysis and test design, before and after","Differences found in parallel running and their root causes","Defects after cutover per module","Legacy capacity decommissioned"],"failureModes":[{"title":"Plausible but wrong documentation","detail":"The model describes what similar code usually does, not what this code does. Expert sampling and generated tests against real behavior catch it."},{"title":"Converting dead code","detail":"Large parts of old estates are unused. Measure what runs before converting everything."},{"title":"Losing the business rules","detail":"Rules hidden in data or job control are missed when only programs are analyzed. Include the whole runtime in scope."},{"title":"Big bang cutover","detail":"A full switch without parallel running turns small differences into incidents. Move module by module."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Tools that analyze, document and translate code are not prohibited practices under Article 5 and are not listed in Annex III, so no high risk obligations apply to the tooling. Engineers and analysts know they are working with an AI tool, including when they query the documentation through a chat assistant, so the Article 50 disclosure duty has no practical effect for the deploying organization. What remains is AI literacy for the staff who use it (Article 4). If the system being modernized is itself an AI system in an Annex III area (for example creditworthiness assessment, point 5(b)), its new version still has to meet the high risk requirements."},"regulations":["eu-ai-act","gdpr","dora","iso-42001","apra-cps-230"],"guidance":[{"title":"Guidelines on Risk Management Practices, Technology Risk","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/guidelines/technology-risk-management-guidelines","note":"Supervisory expectations for technology risk governance, system development, testing and change management at financial institutions in Singapore."},{"title":"Guidelines for secure AI system development","issuer":"UK National Cyber Security Centre","region":"europe","url":"https://www.ncsc.gov.uk/collection/guidelines-secure-ai-system-development","note":"Security guidance for organizations that build AI systems, relevant to in house pipelines and agents that process and generate code."}],"controls":["Program level risk assessment with the AI tooling in scope","Third party risk review of model providers and integrators","Test evidence and parallel run results retained per module","Traceability records from legacy to new components","Independent review of cutover readiness for critical systems"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai does not convert code. It fits around a modernization program as the knowledge\nlayer: generated program documentation and specifications go into a **knowledge base** with\nversion control and hybrid retrieval, and an **AI agent** lets engineers and business analysts\nask what a legacy program does, which rules it applies and what depends on it.\n\n**Agentic workflows** can draft documentation module by module through **custom functions**,\nwith **human in the loop confirmation** and a full audit trail per run. Approved documents are\nthen uploaded to the **knowledge base** (document library with version control). **Test suites**\ncheck the agent's answers against expert approved question sets. The platform is **model\nagnostic**, so the model can be chosen per task, and it can run in the EU or UAE region."},"faq":[{"question":"Can AI convert COBOL to Java on its own?","answer":"Not reliably enough for core systems. Most deployments on this page use AI for comprehension and documentation, and humans or deterministic tools produce the new code. Morgan Stanley's DevGen.AI turns legacy code into English specifications that developers rewrite, because the firm says the tool does not yet write new code as well as a human, and GFT describes a bank where deterministic tools did the conversion while generative AI produced documentation and test scenarios, the only deployment here that reports AI generated tests. Amazon's agent did help upgrade applications from Java 8 or 11 to Java 17, a version upgrade rather than a change of language."},{"question":"How much time does it save?","answer":"Morgan Stanley says DevGen.AI worked through nine million lines in five months and saved about 280,000 developer hours. Amazon reports that its code transformation agent helped migrate tens of thousands of production applications to Java 17 and estimates that this saved more than 4,500 years of development work. Toyota Motor Europe's documentation proof of concept gives no time figure, and savings on mainframe estates depend heavily on how much expert review the output needs."},{"question":"What should stay with humans?","answer":"Validation of business rules, the decision to cut over, and sign off on test and parallel run results. The retiring experts are most valuable as reviewers of AI generated documentation."}],"related":["developer-coding-assistant","developer-api-integration-assistant","enterprise-knowledge-search","aiops-incident-triage"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog as an industry neutral page, with four evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed an unsourced claim about rewrite overruns and a source citation feature that is not in the platform inventory, corrected the EU AI Act basis (Article 5, Annex III, Article 50), clarified the guidance notes and FAQ wording, raised Toyota Motor Europe to production after its reported rollout, dropped COBOL from the Morgan Stanley summary, added source dates, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Second fact check: the problem section and FAQ now describe what each deployment used AI for (test generation appears only at the GFT bank, and Amazon's agent assisted version upgrades), tied the legacy code claim to the evidence and removed insurance and government, which had no evidence, corrected the Amazon record's title and summary, moved the GFT record from production to pilot, rewrote the EU AI Act basis around Article 4, added GDPR and unified the spelling."},{"date":"2026-09-27","note":"Review pass: reworded the Blits.ai build section to describe documentation drafting with human in the loop confirmation and upload to the knowledge base, since the feature inventory has no path from a workflow into the knowledge base without a manual upload step, and removed the MAS AI Risk Management regulation from risk.regulations because the only cited source is a consultation paper, not a finalized guideline."}],"slug":"legacy-code-modernization","url":"https://www.blits.ai/ai-use-cases/legacy-code-modernization","benchmarks":[{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":260000000,"min":260000000,"max":260000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"amazon-java-upgrade-code-transformation","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"google-legacy-code-migration","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":280000,"min":280000,"max":280000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"morgan-stanley-devgen-ai","pooled":true}]}],"indicativeValueResult":{"low":1200000,"high":6000000},"evidence":["airbnb-test-migration-llm","amazon-java-upgrade-code-transformation","google-legacy-code-migration","morgan-stanley-devgen-ai","toyota-motor-europe-legacy-code-documentation"]},{"title":"AI for market abuse surveillance alert triage","shortTitle":"Market abuse surveillance","seoTitle":"AI trade surveillance for market abuse alerts","metaDescription":"AI gathers the evidence behind each market abuse alert and explains its trigger. In a Nasdaq proof of concept, analysts estimated 33% less investigation time.","definition":"AI that helps surveillance analysts triage market abuse and conduct alerts, such as spoofing, layering, wash trades, ramping and insider dealing, by gathering the trade, order, news and communications context, explaining in plain language what triggered each alert and drafting the investigation narrative for the analyst to disposition.","aliases":["trade surveillance AI","market abuse alert triage","communications surveillance triage","insider dealing detection"],"industries":["capital-markets","banking","wealth-and-asset-management"],"functions":["regulatory-compliance","financial-crime-compliance"],"patterns":["anomaly-detection","agentic-workflow","summarization","classification-and-routing"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"second-line","problem":"Banks, brokers and trading venues must detect and report suspicious orders and transactions.\nMost surveillance still runs on rules: an order that exceeds a size, moves a price or matches a\npattern raises an alert, and an analyst has to reconstruct what happened before deciding whether\nit is worth a closer look. That reconstruction is the expensive part: pulling the order book,\nrelated trades, the issuer's filings, news around the event and, for conduct cases, the trader's\nemails and chats.\n\nRules produce large volumes of false positives, and subtle manipulation that spans venues,\ninstruments or time zones does not match a single rule. Supervisors also expect firms to prove\nthat their surveillance works: in Market Watch 79 the FCA described alert scenarios that failed\nunnoticed, in one case for over three years, because of faulty alert logic or data that was never\ningested. The result is a team that spends much of its time closing noise, with the hard cases\ngetting less attention than they deserve.","problemStats":[{"statement":"1LoD's 2026 Surveillance Benchmarking Survey found that 89% of banks want AI enhanced trade surveillance but only 11% have it, and that 78% want generative AI assistants for analysts while 7% have deployed one.","sourceTitle":"Banks want AI surveillance but lack the data to run it","sourceUrl":"https://fintech.global/2026/09/21/banks-want-ai-surveillance-but-lack-the-data-to-run-it/","year":2026},{"statement":"The same 1LoD survey reports that 93% of banks rate false positives a meaningful drag on surveillance and 52% call them a high challenge, which the survey attributes to fragmented data capture and ageing platforms upstream of the alert stage.","sourceTitle":"Banks want AI surveillance but lack the data to run it","sourceUrl":"https://fintech.global/2026/09/21/banks-want-ai-surveillance-but-lack-the-data-to-run-it/","year":2026}],"howItWorks":"1. **Alert in.** The existing surveillance system (rules or models) raises an alert on an order\n   pattern, a trade ahead of a price move or a flagged message.\n2. **Assemble the context.** An agent pulls the relevant orders and trades, the instrument's\n   price and volume around the event, the issuer's filings and news, the trader's history and\n   prior alerts, and for conduct cases the linked communications.\n3. **Explain the trigger.** The AI states in plain language which behaviour set off the alert\n   and which facts make it more or less suspicious, with a link to each underlying record.\n4. **Score and route.** Alerts are ranked by likely risk; clear false positives are proposed\n   for closure with a reason, and the rest go to an analyst queue.\n5. **Draft the case.** For alerts that go further, the AI drafts the investigation narrative\n   and, where needed, the first version of a suspicious transaction and order report.\n6. **Human disposition.** An analyst reviews, edits and decides every alert. Their decisions\n   and reasons are logged and feed back into tuning.","valueDrivers":["compliance","employee-productivity","risk-reduction","speed"],"kpis":["handling-time-reduction","false-positive-reduction","detection-rate-improvement","productivity-gain"],"indicativeValue":{"referenceOrg":"A bank with a markets business raising 40,000 surveillance alerts a year","inputs":[{"key":"alerts","label":"Surveillance alerts reviewed per year","low":40000,"high":40000,"unit":"alerts per year","note":"The reference bank. Replace with your own alert volume across trade and communications surveillance."},{"key":"hoursPerAlert","label":"Analyst hours per alert at first review","low":0.5,"high":1,"unit":"hours per alert","note":"Editorial assumption for gathering evidence and writing the first assessment. Replace with your own time study."},{"key":"timeSaved","label":"Share of review time saved","low":0.15,"high":0.3,"unit":"fraction of review time","note":"Conservative against the benchmark on this page (in Nasdaq's proof of concept testing, surveillance analysts estimated a 33% reduction in investigation time)."},{"key":"hourlyCost","label":"Fully loaded cost of a surveillance analyst","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"alerts * hoursPerAlert * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Analyst time released from first line alert review","caveat":"Counts analyst time only. It leaves out the cost of the AI and data work, any change in the number of alerts, and the value of detecting abuse that rules miss, which is the larger prize but hard to price."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The AI layer is the easier part. The hard work is complete, reconciled trade, order and communications data with a clear chain of custody, and model governance that a supervisor will accept for a system that clears alerts.","dataPrerequisites":["Complete order and trade records across venues and asset classes, reconciled to source","Communications records (email, chat, voice transcripts) held in original form and linked to traders","Reference data, issuer filings and a news feed with timestamps","A labelled history of past alert dispositions with reasons"],"integrations":["Existing trade and communications surveillance platforms (alert source)","Order management and execution systems, market data","Communications archive and voice recording platform","Case management for investigations and suspicious transaction and order reports","Model inventory and model risk management tooling"]},"implementation":{"steps":[{"title":"Start with explanation, not auto closure","detail":"First give analysts a plain language summary and an evidence pack for every alert. Measure review time and analyst agreement before letting the system propose closures."},{"title":"Fix the data before the model","detail":"Reconcile order and trade feeds to source and check that every business line and venue is actually monitored. An AI layer on incomplete data reprocesses the same gaps faster."},{"title":"Define the closure policy","detail":"Write down which alert types may be proposed for closure, the evidence required, the sampling rate for quality checks and who signs off the policy."},{"title":"Validate like any surveillance model","detail":"Put the triage model through model validation: back testing on past alerts including known true cases, stability over time and a documented explanation of its logic."},{"title":"Add communications and cross product views","detail":"Once trade triage is trusted, link trade alerts to communications and to related instruments, where rules alone miss the most."}],"guardrails":["A human analyst dispositions every alert; the AI proposes, it never closes on its own","Every explanation links to the underlying orders, trades and messages it relies on","Random quality sampling of alerts the AI proposed to close, with results reported to compliance","Access to communications data limited by role and logged, with personal data minimised in prompts","Change control and regression tests on known true positive cases for every model or prompt change"],"humanInTheLoop":"Surveillance analysts own every disposition and every escalation to a suspicious transaction and order report. Second line compliance approves the closure policy and reviews quality samples, and model validation signs off the triage model before it goes live and after changes.","kpisToInstrument":["Median review time per alert, by alert type","Share of alerts proposed for closure and the analyst agreement rate","True positives found per period, including cases the rules did not flag first","Quality sample findings on closed alerts","Share of alerts with a complete evidence pack"],"failureModes":[{"title":"Confident explanations on missing data","detail":"The AI writes a fluent rationale while part of the order flow was never ingested. Check data completeness per venue and show gaps in the evidence pack."},{"title":"Automation bias","detail":"Analysts accept the suggested disposition without reading the evidence. Track agreement rates, rotate blind reviews and sample closures."},{"title":"Tuning away real abuse","detail":"Optimising for fewer alerts lowers detection. Always back test on known true cases and report detection alongside false positives."},{"title":"Unexplainable to the supervisor","detail":"A model that cannot show why it cleared an alert fails regulatory scrutiny. Keep the reasoning and evidence for every alert."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Surveillance of orders and transactions as such is not listed in Annex III. Where the system monitors and evaluates the behaviour of the firm's own staff, in their communications or their trading, it can fall under Annex III point 4(b) (AI used to monitor and evaluate the performance and behaviour of persons in work relationships), so the tier depends on whether the system scores individual employees. Inferring employees' emotions from biometric data such as voice recordings is prohibited in the workplace under Article 5(1)(f)."},"regulations":["eu-ai-act","gdpr","us-sr-11-7","mas-ai-risk-management","iso-42001","eu-mar","mifid-ii"],"guidance":[{"title":"Market Abuse Regulation (EU) No 596/2014","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2014/596/oj","note":"Article 16 requires firms that arrange or execute transactions to have effective arrangements, systems and procedures to detect and report suspicious orders and transactions."},{"title":"Commission Delegated Regulation (EU) 2016/957","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg_del/2016/957/oj","note":"Technical standards for detecting and reporting suspicious orders and transactions, including the duty to keep for five years the analysis of each examined order or transaction and the reasons for submitting or not submitting a STOR."},{"title":"Market Watch 79","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publications/newsletters/market-watch-79","note":"FCA examples of surveillance failures caused by data ingestion and alert logic issues, and its 2023 peer review of how 9 investment banks test automated surveillance models under UK MAR."},{"title":"Artificial Intelligence (AI) Model Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Good practices MAS observed in its 2024 thematic review of banks' AI and generative AI model risk management, covering governance, oversight, development and deployment; relevant when validating a triage model."}],"controls":["Surveillance model and triage AI registered in the model inventory with an owner and validation status","Documented closure policy approved by compliance, with sampling of AI assisted closures","Data completeness checks per venue, asset class and communications channel","Full audit trail of alert, evidence, AI output, analyst decision and reason","Periodic back testing against known true positive cases"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered for each alert from the surveillance\nplatform through the API. **Custom functions** call the order, trade and market data services,\na **SQL knowledge base** exposes the alert and disposition history, and a **knowledge base**\nwith hybrid retrieval holds the surveillance procedures and typologies the explanation must\nfollow. The agent returns **structured output** (trigger, evidence links, risk rank, draft\nnarrative) into the case management system.\n\n**Human in the loop approval** keeps every disposition with the analyst, the **tool execution\npolicy** limits what the agent may read or write, and **PII masking** keeps personal data out of\nprompts where it is not needed. **Test suites** replay past alerts, including known true cases,\non every prompt or model change, and the per run audit trail and execution tracing give model\nvalidation and supervisors the full record. The platform is model agnostic and can run in EU or\nUAE data residency regions."},"faq":[{"question":"Can AI close market abuse alerts on its own?","answer":"It should not. The defensible pattern is that the AI assembles evidence, explains the trigger and proposes a disposition, and a named analyst decides. In the EU, Commission Delegated Regulation 2016/957 requires firms to keep, for five years, the analysis of every examined order or transaction and the reasons for reporting it or not, so each closure needs a documented rationale."},{"question":"How much analyst time does AI triage save?","answer":"Public figures are still few and early. Nasdaq reported that surveillance analysts estimated a 33% reduction in investigation time during proof of concept testing of its generative AI feature. Treat that as an estimate from a pilot and measure your own review times per alert type."},{"question":"Is AI surveillance high risk under the EU AI Act?","answer":"Surveillance of client orders and transactions generally is not. Monitoring and evaluating the behaviour of the firm's own employees, in their communications or their trading, can fall under Annex III point 4(b), so a design that scores individual staff needs the high risk controls."},{"question":"What is the biggest obstacle?","answer":"Data. In 1LoD's 2026 Surveillance Benchmarking Survey, 71% of answers on what most hinders surveillance pointed to fragmented, non standardised or poor quality data, and the report says this leaves many AI projects stuck at proof of concept."}],"related":["aml-alert-triage","call-quality-and-compliance-monitoring","suspicious-activity-report-drafting","model-risk-validation-copilot","continuous-controls-testing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against public sources, including the Nasdaq release and US federal AI inventories."},{"date":"2026-09-25","note":"Consolidation pass: added EU Market Abuse Regulation, MiFID II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: tightened the FCA Market Watch 79 and 1LoD survey wording, widened the EU AI Act basis to staff trading and Article 5(1)(f), added Delegated Regulation 2016/957 as guidance and FAQ basis, corrected the MAS paper title, added SEO title and description, and corrected details in the Japan Exchange Group, Deutsche Bank and CFTC records."}],"slug":"market-abuse-surveillance-triage","url":"https://www.blits.ai/ai-use-cases/market-abuse-surveillance-triage","benchmarks":[{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":33,"min":33,"max":33,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"nasdaq-market-surveillance-generative-ai","pooled":true}]}],"indicativeValueResult":{"low":180000,"high":1200000},"evidence":["cftc-spoofing-detection-pilot","deutsche-bank-agentic-trade-surveillance","japan-exchange-group-ai-market-surveillance","nasdaq-market-surveillance-generative-ai","sec-single-event-insider-trading-analysis"]},{"title":"AI for merchant underwriting and risk monitoring","shortTitle":"Merchant underwriting and monitoring","seoTitle":"AI merchant underwriting and risk monitoring","metaDescription":"AI screens merchant websites and ranks risky merchants for review. Airwallex reports 50% fewer false positives; Coris reports about 89% fewer daily reviews at Weave.","definition":"AI that helps acquirers, payment facilitators and software platforms with embedded payments decide which merchants to accept and on what terms, by checking what a business really sells and how risky it is at onboarding, and then watches every active merchant for changes in behaviour, ranking the few that need an analyst so fraud, prohibited trade and credit losses are caught early.","aliases":["merchant risk AI","merchant onboarding underwriting automation","merchant website screening","merchant portfolio monitoring","PayFac risk automation"],"industries":["payments","technology","banking"],"functions":["onboarding-and-kyc","fraud-prevention","risk-management"],"patterns":["prediction-and-scoring","anomaly-detection","classification-and-routing","summarization","agentic-workflow"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Every merchant accepted for card payments brings risk with it. The merchant may sell something\nthe card schemes or the law prohibit, never ship the goods, launder transactions for another\nbusiness or go out of business with open refunds and chargebacks. Visa requires the acquirer,\nthe financial institution with the direct relationship with the merchant, to run compliance\nchecks before a merchant can accept Visa payments, and the acquirer removes merchants engaged in\nillegal commerce that cannot comply with Visa's rules and applicable law. Visa's Acquirer Monitoring\nProgram, effective April 2025, also consolidates earlier fraud and dispute programs and sets\nfraud thresholds and enumeration criteria for acquirers and their merchants.\n\nThe traditional answer is manual. At onboarding an analyst checks the business registration,\nreads the website, looks up reviews and decides; after onboarding a small team works through\nexports of balances, refunds and chargebacks in spreadsheets. It does not scale: platforms with\nembedded payments sign up thousands of small merchants, keyword screening of websites produces\nmany false alarms, and only a small sample of the portfolio is ever looked at. Coris reports that\nbefore automation only about 3% of Weave's merchants received any manual scrutiny. That leaves\nroom for what Visa describes as merchants who fraudulently conceal the true nature of their\nbusinesses to avoid its compliance requirements.","problemStats":[],"howItWorks":"1. **Collect the application and the footprint.** The system takes the application data and\n   enriches it with business registry records, the merchant's website, online reviews, adverse\n   media and, where relevant, credit data.\n2. **Understand what the merchant sells.** A language model reads the website and product\n   descriptions in context, distinguishes allowed goods from prohibited ones and checks that\n   the declared business and merchant category match what is actually offered.\n3. **Score and decide at onboarding.** A risk score combines identity, footprint and category\n   signals. Low risk merchants are approved automatically within policy; the rest go to an\n   analyst with a summary of the evidence and a proposed decision, such as approve, approve with\n   a reserve or payout delay, request documents, or decline.\n4. **Monitor every active merchant.** Models watch processing behaviour (volume spikes, refunds,\n   chargebacks, negative balances, concentration of card numbers, changes to the website) and\n   score each merchant daily, so analysts start with the riskiest accounts instead of a random\n   sample.\n5. **Act and record.** Analysts decide on payout holds, reserves, outreach or offboarding, with\n   the agent drafting the merchant message and the case note, and every decision is kept for the\n   acquirer, the sponsor bank and the card schemes.","valueDrivers":["risk-reduction","employee-productivity","speed","compliance","cost-to-serve"],"kpis":["alert-volume-reduction","false-positive-reduction","detection-rate-improvement","handling-time-reduction","automation-rate","fraud-losses-prevented"],"indicativeValue":{"referenceOrg":"A software platform with embedded payments onboarding 20,000 merchants a year","inputs":[{"key":"merchants","label":"New merchant applications per year","low":20000,"high":20000,"unit":"applications per year","note":"The reference platform. Replace with your own application volume."},{"key":"minutesPerReview","label":"Analyst minutes per manual review today","low":20,"high":30,"unit":"minutes per application","note":"Editorial assumption for web searches, registry checks and a decision. Replace with your own time study."},{"key":"reviewReduction","label":"Share of manual reviews the AI removes","low":0.4,"high":0.5,"unit":"fraction of reviews","note":"The high end matches Airwallex's early result of 50 percent fewer websites needing a manual check at onboarding. Coris reports 70% fewer manual reviews at Tekmetric and about 89% fewer daily reviews at Weave, but those vendor figures include ongoing monitoring, so they are not used as the ceiling."},{"key":"costPerHour","label":"Fully loaded risk analyst cost per hour","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"merchants * minutesPerReview / 60 * reviewReduction * costPerHour","currency":"USD","period":"per year","resultLabel":"Onboarding review effort avoided","caveat":"Onboarding labour only. It leaves out the fraud and credit losses prevented by catching bad merchants earlier, the effort saved in ongoing monitoring, faster activation of good merchants, fewer breaches of card scheme thresholds, and the cost of the data sources, the models and the integration."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Scoring and website review can start on application data and public sources with little integration. The harder part is ongoing monitoring, which needs processing data from every acquirer or processor the platform uses, and a clear risk policy that says which score leads to which action.","dataPrerequisites":["A written risk appetite with prohibited and restricted categories and the action per risk level","Historical merchant applications with outcomes (approved, declined, later terminated, losses)","Processing data per merchant (volumes, refunds, chargebacks, balances, payouts)","Access to business registry, credit and adverse media data sources"],"integrations":["Onboarding or application system","Payment processor or acquirer APIs and dashboards (for example connected account data)","Payout and reserve controls","Case management or CRM for analyst reviews and merchant outreach","Messaging or ticketing tool for merchant communication"]},"implementation":{"steps":[{"title":"Write the policy before the model","detail":"Agree the prohibited and restricted categories, the risk levels and the action for each (approve, reserve, delay payouts, request documents, decline) with the sponsor bank or acquirer. The AI applies the policy; it does not invent it."},{"title":"Start with website and category screening","detail":"Let the model read the merchant's website and product descriptions and compare them with the declared business. Keyword rules on websites produce many false alarms, and Airwallex reports that its generative AI website scanner halved them in early results."},{"title":"Run in parallel before you switch","detail":"Score new applications and the existing portfolio alongside the current manual process for a few weeks, compare decisions and losses, and tune thresholds before the AI queue becomes the primary one."},{"title":"Extend to the whole portfolio","detail":"Monitor every active merchant daily, not a sample, and rank the queue by risk so analysts start with the top decile. Add automated actions such as payout holds only for clear, high confidence signals, with an analyst review the same day."},{"title":"Feed outcomes back","detail":"Label every decision with its outcome (losses, chargebacks, terminations) and use the labels to retrain and to prune rules that only add noise."}],"guardrails":["Declines, terminations and reserves above a set level always need an analyst decision","Every score comes with the evidence behind it, so the analyst and the merchant can be told why","Website and document content is treated as data, never as instructions to the model","Sole traders' personal data is limited to what the risk policy needs, and masked in prompts and logs","Regular checks that approval and decline rates do not differ unfairly between comparable merchant groups"],"humanInTheLoop":"Analysts decide every decline, termination and material reserve or payout hold, and review a sample of automatic approvals each week. The risk policy owner approves every change to categories, thresholds or automated actions, and the sponsor bank or acquirer can audit the decisions.","kpisToInstrument":["Share of applications approved automatically and time from application to approval","Manual reviews per week and share of reviews that lead to an action","False positive rate of website and category screening","Losses from merchants terminated for fraud or credit, per 1,000 merchants onboarded","Chargeback and fraud ratios against the card scheme thresholds"],"failureModes":[{"title":"Fraudsters who look good on paper","detail":"Generated websites, fake reviews and borrowed business identities pass a one time check. Keep monitoring after onboarding and watch for website and behaviour changes."},{"title":"Automatic declines of good small businesses","detail":"Thin footprints and unusual categories push legitimate merchants into declines. Route uncertain cases to an analyst and track appeals."},{"title":"Rules and models that fight each other","detail":"Rules added on top of the score without review double the alert volume. Review the combined queue and remove rules that do not change decisions."},{"title":"A queue that nobody trusts","detail":"Analysts ignore scores they cannot explain. Show the drivers and the evidence with every score."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Assessing businesses and detecting fraud is not an Annex III use as such, and Annex III point 5(b) excludes systems used to detect financial fraud. If the system evaluates the creditworthiness of a natural person, for example a sole trader applying to accept payments, it can fall under Annex III point 5(b), which covers evaluating the creditworthiness of natural persons or establishing their credit score, and be high risk. Keep credit assessment of individuals separate or treat it as a high risk system."},"regulations":["eu-ai-act","gdpr","pci-dss","eu-amlr","fatf-recommendations","us-bsa","dora"],"guidance":[{"title":"Introducing the Visa Acquirer Monitoring Program","issuer":"Visa","region":"global","url":"https://corporate.visa.com/en/sites/visa-perspectives/security-trust/introducing-visa-acquirer-monitoring-program.html","note":"Visa's acquirer program, effective April 2025, consolidates earlier fraud and dispute programs and sets fraud thresholds and enumeration criteria for acquirers and their merchants."},{"title":"Visa Network Integrity","issuer":"Visa","region":"global","url":"https://corporate.visa.com/en/about-visa/visa-network-integrity.html","note":"Describes how Visa monitors acquirers in high risk categories and how acquirers must remove merchants engaged in illegal commerce."}],"controls":["Documented risk policy with prohibited categories, thresholds and approved automated actions","Model inventory entry, validation and drift monitoring for the scoring models","Audit trail of every score, the evidence shown and the decision taken, per merchant","Periodic fairness and outcome review of approvals, declines and terminations","Oversight by the sponsor bank or acquirer, including access to decisions and samples"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the merchant review is an **agentic workflow** that the onboarding system starts\nthrough the API for each application, and that runs on a schedule for portfolio checks. The\nagent uses built in **web page browsing** and **web search** to read the merchant's website and\nfootprint, **custom functions** to call registry, processor and internal systems through REST\nor SQL, and a **knowledge base** with hybrid retrieval that holds the risk policy and the\nprohibited category list. It returns a decision proposal with the evidence as **structured\noutput**.\n\n**Human in the loop approval** holds declines, terminations and payout holds above a set level\nfor an analyst, and the **tool execution policy** limits which actions the agent may take on its\nown. **Guardrails** and **PII masking** protect sole traders' personal data, every run has a\n**full audit trail**, and **test suites** evaluate the workflow against labelled historical applications before any\nchange to the policy or the model goes live. The platform is model agnostic, so the risk team can\npick the model per agent."},"faq":[{"question":"What does generative AI add to merchant underwriting?","answer":"Reading websites and documents in context. Airwallex says its generative AI website scanner can better distinguish a retailer selling a military style jacket from a merchant selling prohibited military goods, and that early results from its own internal analysis show 50 percent fewer false positives on average than its earlier rules based model."},{"question":"Is monitoring after onboarding really needed?","answer":"Yes. A check at signup only sees what the merchant chooses to show, and Visa says some merchants conceal the true nature of their businesses to avoid compliance requirements. Coris reports that Weave went from manual scrutiny of about 3% of merchants to automated checks across nearly all merchants with meaningful volume, while cutting daily reviews by about 89%. Visa says that, using its AI tools and machine learning models, it saw a fivefold increase in acquirer remediation and terminations for merchant noncompliance between 2020 and 2024."},{"question":"Can the AI decline merchants on its own?","answer":"It should not decline or terminate on its own. Let it approve clear low risk cases within policy and send everything else, with the evidence, to an analyst. If the assessment covers the creditworthiness of a sole trader, check whether it is high risk under the EU AI Act."}],"related":["business-onboarding-and-ubo-discovery","chargeback-and-representment","real-time-fraud-scoring"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Unpublished by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Editor fixes after the skeptic review: removed the unsourced claim about scheme fines and liability of payment facilitators and platforms, replaced it with Visa's stated acquirer duties, removed an unsourced superlative from the implementation steps, and lowered the high end of the review reduction to 0.5."},{"date":"2026-09-27","note":"First version, written in a discover run for payments with four evidence records from Airwallex, Visa, Weave and Tekmetric, quotes checked against the sources."}],"slug":"merchant-underwriting-and-risk-monitoring","url":"https://www.blits.ai/ai-use-cases/merchant-underwriting-and-risk-monitoring","benchmarks":[{"kpi":"alert-volume-reduction","label":"Alert volume reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":79.5,"min":70,"max":89,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"weave-coris-merchant-risk-monitoring","pooled":true},{"id":"tekmetric-coris-merchant-risk-automation","pooled":true}]},{"kpi":"detection-rate-improvement","label":"Detection improvement","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":5,"min":5,"max":5,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"visa-network-integrity-merchant-monitoring","pooled":true}]},{"kpi":"false-positive-reduction","label":"False positive reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"airwallex-generative-ai-website-screening","pooled":true}]}],"indicativeValueResult":{"low":106666.66666666669,"high":350000},"evidence":["airwallex-generative-ai-website-screening","tekmetric-coris-merchant-risk-automation","visa-network-integrity-merchant-monitoring","weave-coris-merchant-risk-monitoring"]},{"title":"AI for mobile network planning and capacity optimization","shortTitle":"Network planning and capacity","seoTitle":"AI for mobile network capacity planning","metaDescription":"AI forecasts where mobile networks will congest and tunes radio settings. See Nokia's deployments for NTT DOCOMO and stc, plus Vodafone's trials.","definition":"Machine learning that forecasts where and when a mobile network will run out of capacity, recommends where to add cells, spectrum or hardware, and continuously tunes radio parameters so existing capacity carries more traffic, with planners approving investments and major changes.","aliases":["AI network planning","capacity planning for mobile networks","cognitive SON","AI RF optimization","self optimizing networks"],"industries":["telecommunications"],"functions":["network-operations","analytics-and-reporting"],"patterns":["prediction-and-scoring","recommendation-and-personalization","anomaly-detection","agentic-workflow"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"network","problem":"Mobile data traffic does not grow evenly across a network. Some cells can congest at busy hours\nwhile others are rarely loaded, and new housing, offices and events can move demand around faster\nthan annual planning cycles follow. A new site or carrier in the wrong place ties up investment\nwhile customers a few streets away still see slow speeds.\n\nBetween investments, radio engineers tune thousands of parameters (antenna tilts, power, handover\nand load balancing settings) to squeeze more out of the existing network. That work is slow and\nmanual. Vodafone describes a trial in which a machine learning algorithm found optimal voice over\nLTE settings for 450 cells in four hours, a task that would have taken an engineer around two and a\nhalf months by hand. With 4G, 5G and several vendors in one network, manual tuning gets harder\nstill.","problemStats":[],"howItWorks":"1. **Forecast demand.** Models forecast traffic per cell and area from history, subscriber growth,\n   device mix and planned developments, and flag where congestion will appear.\n2. **Estimate capacity.** The system estimates how much more traffic each cell can carry with its\n   current configuration and where the limit is (spectrum, hardware, backhaul, interference).\n3. **Optimise before building.** Self optimizing network functions tune parameters such as tilt,\n   power and load balancing so neighbouring cells share load. They can act ahead of demand: in a\n   Vodafone trial in Ireland, algorithms predicted where 3G traffic would peak in the next hour so\n   the network could rebalance load in advance.\n4. **Recommend investments.** Where optimisation is not enough, the system simulates candidate\n   sites, carriers or hardware upgrades and ranks them by traffic served per unit of spend.\n5. **Close the loop.** Planners approve investments; after each change the system compares the\n   measured effect with its forecast and recalibrates.","valueDrivers":["cost-to-serve","customer-experience","employee-productivity","speed"],"kpis":["productivity-gain","processing-time-reduction","cost-savings","cost-reduction"],"indicativeValue":{"referenceOrg":"A mobile operator with a capacity driven radio investment budget of USD 200 million a year","inputs":[{"key":"capacityCapex","label":"Annual capacity driven radio investment","low":200000000,"high":200000000,"unit":"USD per year","note":"The reference operator."},{"key":"capexEfficiency","label":"Share of capacity investment avoided or deferred through better targeting and optimisation","low":0.02,"high":0.06,"unit":"fraction of capacity investment","note":"Editorial assumption. None of the evidence on this page publishes a verified investment saving; replace with your own post investment reviews."}],"formula":"capacityCapex * capexEfficiency","currency":"USD","period":"per year","resultLabel":"Capacity investment avoided or deferred","caveat":"Investment only, and deferral is not the same as saving. It leaves out engineering time released, the revenue and churn effect of fewer congested cells, and the cost of the planning platform and data."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Forecasting and optimisation need clean, granular performance data from every vendor and a trusted digital view of the network. Automated parameter changes touch live customers, so they need strong guardrails and radio engineering buy in.","dataPrerequisites":["Traffic, quality and utilisation counters per cell and carrier, with at least a year of history","Site, antenna and configuration inventory that matches the live network","Subscriber and device growth data by area","Planned developments, events and competitor coverage where available"],"integrations":["Radio network management and configuration per vendor","Self organizing network (SON) platform","Planning and propagation tools","Network data lake or analytics platform","Capital planning and project tracking systems"]},"implementation":{"steps":[{"title":"Clean the network view","detail":"Reconcile inventory with the live configuration. Forecasts and simulations on a wrong site database lead to wrong investments."},{"title":"Start with forecasting and ranking","detail":"Use models to rank congested and soon to be congested cells, and let planners compare the ranking with their own judgement for a planning cycle."},{"title":"Automate reversible optimisation","detail":"Let SON functions change parameters within bounds and with automatic rollback, starting in one cluster with a control area."},{"title":"Link to the investment process","detail":"Feed the ranked recommendations into capital planning, and review every investment afterwards against the forecast it was based on."},{"title":"Extend across vendors and technologies","detail":"Move from single vendor tools to a view that covers 4G, 5G and every vendor, so optimisation in one layer does not hurt another."}],"guardrails":["Parameter changes only within engineering defined bounds, with automatic rollback on quality loss","Investments above a set value always approved by planners and finance","Exclusion of critical sites and emergency coverage from automated changes","Every automated change logged with its reason and measured effect"],"humanInTheLoop":"Radio planners and engineers own investment decisions and the bounds for automated optimisation. They review recommendations each planning cycle, approve changes outside the bounds, and use post investment reviews to decide how much to trust the forecasts.","kpisToInstrument":["Share of congested cells, by hour and area","Forecast accuracy of traffic and congestion per planning cycle","Throughput and quality before and after each optimisation or investment","Engineering hours per optimisation task","Capacity investment per unit of traffic carried"],"failureModes":[{"title":"Optimising the average, hurting the edge","detail":"Changes improve cell averages while users at cell edges or indoors lose service. Monitor distributions, not only averages."},{"title":"Forecasts built on the past only","detail":"Models miss new housing, venues or competitor moves. Add planners' local knowledge and external data."},{"title":"Oscillating parameters","detail":"Several automated functions fight each other and parameters flip back and forth. Coordinate SON functions and damp changes."},{"title":"Trusting a vendor's black box","detail":"Recommendations cannot be explained to finance or engineers. Require the reasoning and data behind each recommendation."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Forecasting demand, ranking congested cells and recommending investments is normally minimal risk. Under Article 6(2), Annex III point 2 lists AI systems intended as safety components in the management and operation of critical digital infrastructure as high risk, and Recital 55 ties this to the digital infrastructure in the Annex to Directive (EU) 2022/2557, which includes providers of public electronic communications networks. Recital 55 defines such safety components as systems that directly protect the physical integrity of the infrastructure or the health and safety of persons and property and that are not necessary for the system to function. Closed loop parameter optimisation on the live radio network is high risk only when it serves in that role, for example a loop whose purpose is to protect emergency call availability, so each automated loop should be assessed against point 2 and the outcome documented. Loops that only optimise performance or capacity are usually not safety components."},"regulations":["eu-ai-act","nist-ai-rmf","iso-42001","nis2"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 2 is the relevant test if automated optimisation becomes a safety component of network operation."},{"title":"Recital 55, safety components of critical infrastructure","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/recital/55/","note":"Explains which critical digital infrastructure is meant and what counts as a safety component, and excludes components used solely for cybersecurity."}],"controls":["Documented bounds for automated parameter changes with an accountable owner","Post investment review comparing forecast and measured traffic","Audit log of automated configuration changes","Model validation of forecasts before each planning cycle"],"incidents":[]},"blitsAi":{"howToBuild":"Forecasting and SON optimisation run in specialist radio tools. Blits.ai adds the planning\nassistant around them: an **AI agent** with a **SQL knowledge base** over traffic, congestion and\ninvestment tables lets planners ask questions in plain language (\"which cells in this region will\ncongest by next summer, and what did we spend there last year?\"), and a **knowledge base** holds\nplanning guidelines and vendor documentation.\n\n**Agentic workflows** can assemble a recommendation pack for a planning cycle through **custom\nfunctions** that call the forecasting and planning APIs, with **human in the loop approval**\nbefore anything reaches capital planning. Answers are logged with full traces, **test suites**\ncheck answer quality, and the platform is model agnostic."},"faq":[{"question":"What does AI add to self optimizing networks?","answer":"Classic SON functions apply rules that engineers set. AI based SON uses models to choose and time parameter changes autonomously. Nokia reported in 2024 that its MantaRay Cognitive SON, deployed in stc's commercial network in Saudi Arabia, processed more than 10,000 actions in a high traffic period and raised the utilisation of loaded cells by about 30 percent."},{"question":"Can AI decide where to build new sites?","answer":"It can rank candidate locations. In 2022 Nokia deployed its AI capacity planning software for NTT DOCOMO to predict the capacity of 4G cells and simulate the best candidate locations for 5G cells and radio hardware, and DOCOMO said it expected the software to help its network capacity design work. The investment decision should stay with planners and finance, who can weigh cost, coverage and local knowledge."},{"question":"How fast does machine learning tune a network compared with engineers?","answer":"In a Vodafone Germany trial with Huawei, a machine learning algorithm found optimal voice over LTE settings for 450 cells in four hours, which Vodafone says would have taken an engineer around two and a half months."}],"related":["ran-energy-optimization","autonomous-network-operations","predictive-network-maintenance","network-fault-triage-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched for the telecommunications vertical and verified against operator and vendor sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact check against sources: added SEO title and description, removed an unsourced budget claim and response time, sourced the stc figures, clarified the EU AI Act basis with Recital 55 and added NIS2."},{"date":"2026-09-27","note":"Review fixes: EU AI Act tier set to context dependent in line with the other network automation pages, dated the Nokia deployments and softened unsourced background claims."},{"date":"2026-09-27","note":"Fact checked against sources: all four owned sources and their dates confirmed; added the Recital 55 condition that safety components are not necessary for the system to function; removed unsourced language tags from three evidence records."},{"date":"2026-09-27","note":"Review fixes: reworded the closed loop example around protection rather than effect, in line with the autonomous network operations page; cleared the unsourced language tag on the Deutsche Telekom RAN Guardian and MINDR evidence record."}],"slug":"network-planning-and-capacity-optimization","url":"https://www.blits.ai/ai-use-cases/network-planning-and-capacity-optimization","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":95,"min":95,"max":95,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"deutsche-telekom-ran-guardian-and-mindr-agents","pooled":true}]}],"indicativeValueResult":{"low":4000000,"high":12000000},"evidence":["deutsche-telekom-ran-guardian-and-mindr-agents","ntt-docomo-nokia-ai-capacity-planning","stc-nokia-cognitive-son","telefonica-espana-network-analytics-optimization","vodafone-machine-learning-son-trials"]},{"title":"AI for money mule account and network detection","shortTitle":"Mule network detection","seoTitle":"AI money mule detection for banks","metaDescription":"AI finds money mule accounts through the links between them. RBI's MuleHunter.AI reached 21 Indian banks in 2025; five Australian banks share receiving account data.","definition":"Graph and behavioural machine learning that finds money mule accounts and the networks around them, such as circular flows, layering chains and clusters of newly linked accounts, and supports investigators in tracing scam proceeds and restricting accounts before the money is gone.","aliases":["money mule detection","mule account detection","scam proceeds tracing","network analytics for financial crime"],"industries":["banking","payments"],"functions":["fraud-prevention","financial-crime-compliance"],"patterns":["anomaly-detection","prediction-and-scoring","agentic-workflow","summarization"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"copilot","adoptionStage":"early-adopters","segment":"middle-office","problem":"Scams and most fraud need a place to land the money. Mule accounts, opened by fraudsters or run by\nrecruited account holders, receive the proceeds and move them on quickly, for example to other\nbanks, crypto exchanges, cash machines or remittance services. By the time the victim reports, the\nmoney has often left.\n\nA single bank looking at one account at a time sees little: a new account with some incoming\ntransfers. The pattern only shows in the network, such as many senders who are scam victims, a\nfan in and fan out shape, shared devices and addresses, or accounts that were opened in a burst.\nRegulators are also shifting scam losses onto firms. Under the UK reimbursement rules for\nauthorised push payment scams, the sending and receiving firms split the cost of reimbursing\nvictims equally, so for a receiving bank detecting mules is now a loss and compliance issue, not\nonly a crime prevention one.","problemStats":[{"statement":"The US Federal Trade Commission reports that in 2024 consumers reported losing more money to scams paid by bank transfer or cryptocurrency than through all other payment methods combined.","sourceTitle":"New FTC Data Show a Big Jump in Reported Losses to Fraud to $12.5 Billion in 2024","sourceUrl":"https://www.ftc.gov/news-events/news/press-releases/2025/03/new-ftc-data-show-big-jump-reported-losses-fraud-125-billion-2024","year":2025}],"howItWorks":"1. **Build the graph.** Accounts, customers, devices, IP addresses, addresses, phone numbers and\n   payment flows become nodes and edges, updated continuously from onboarding and payment data.\n2. **Score accounts and communities.** Behavioural models score each account for mule like\n   activity (rapid in and out, pass through balances, sudden change after dormancy), and graph\n   algorithms find clusters and chains that share signals.\n3. **Bring in external signals.** Scam reports from other banks, confirmation of payee\n   mismatches, industry or central bank mule lists and law enforcement requests enrich the scores.\n4. **Assemble the case.** An agent drafts a fund flow timeline and case narrative for each\n   cluster: who received what from whom, where it went next, and which signals link the accounts.\n5. **Decide and act.** An investigator decides on restrictions, exits, recall requests to peer\n   banks and reporting, and the decision and reason are recorded against every account touched.","valueDrivers":["risk-reduction","compliance","speed"],"kpis":["detection-rate-improvement","fraud-loss-reduction","false-positive-reduction","processing-time-reduction","interactions-handled"],"indicativeValue":{"referenceOrg":"A retail bank receiving scam proceeds in 2,000 reported cases a year","inputs":[{"key":"cases","label":"Scam cases per year where the bank received the funds","low":2000,"high":2000,"unit":"cases per year","note":"The reference bank. Replace with your own count of inbound scam reports."},{"key":"averageLoss","label":"Average amount received per case","low":1500,"high":4000,"unit":"USD per case","note":"Editorial assumption. Replace with your own data."},{"key":"extraRecovery","label":"Additional share of funds frozen or recovered through earlier detection","low":0.05,"high":0.15,"unit":"fraction of funds","note":"Editorial assumption. Public deployments rarely disclose recovery rates, so keep this low until you have your own results."}],"formula":"cases * averageLoss * extraRecovery","currency":"USD","period":"per year","resultLabel":"Scam proceeds frozen or recovered","caveat":"Covers recovered funds only. It leaves out reimbursement liabilities avoided under schemes that share losses with the receiving bank, the investigation time saved, regulatory benefits, and the cost of false restrictions on genuine customers."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Graph analytics needs clean entity resolution across customers, devices and counterparties, and the best signals come from outside the bank. Restricting accounts is high impact, so the decision process and customer remediation need as much work as the model.","dataPrerequisites":["Payment flows with counterparty identifiers, including inbound instant payments","Onboarding data, device and session data linked to accounts","Confirmed mule and scam case outcomes for training and evaluation","Access to industry or central bank mule intelligence where it exists"],"integrations":["Payments hub and core banking system","Onboarding and identity verification systems","Fraud and AML case management","Industry data sharing schemes and peer bank recall processes","Account restriction and exit workflows"]},"implementation":{"steps":[{"title":"Start from confirmed cases","detail":"Collect every confirmed mule account and inbound scam case from the last two years and map the signals they shared. This becomes the training set and the benchmark."},{"title":"Resolve entities before modelling","detail":"Link customers, accounts, devices and contact details reliably; poor entity resolution produces false networks that lead to wrong restrictions."},{"title":"Combine rules, behaviour and graph","detail":"Start with known typologies as rules, add behavioural scoring, then graph features and community detection, and measure the lift each layer adds on the benchmark set."},{"title":"Give investigators the network view","detail":"Provide a visual network and a drafted fund flow narrative per cluster, so investigators can act on a whole network at once instead of account by account."},{"title":"Connect to the outside","detail":"Join industry intelligence sharing and agree recall and freeze procedures with peer banks, so detection turns into recovered funds."}],"guardrails":["Account restrictions and exits only by a trained investigator, with the reason recorded","Fast review and remediation route for customers restricted in error","Graph links shown with the evidence behind them, never as an unexplained score","Regular testing for disparate impact across customer groups, for example by age, nationality and student status","Data sharing with peer banks only under the legal gateway that allows it"],"humanInTheLoop":"The models and agent find and assemble; investigators decide. Every restriction, exit, recall request and report is a documented human decision, and a second line reviews samples of both actioned and dismissed clusters.","kpisToInstrument":["Mule accounts identified per month and the share confirmed on investigation","Time from first inbound scam payment to restriction","Value of funds frozen or recovered","Share of restricted customers released after review","Inbound scam reports from peer banks per million accounts"],"failureModes":[{"title":"Networks built on bad links","detail":"Shared addresses in student housing or shared devices in families create false clusters. Weight links by strength and require investigator review of the evidence."},{"title":"Detection without recovery","detail":"Mules are found after the money has moved on. Measure time to restriction, not only detection counts, and connect to recall processes."},{"title":"Targeting the recruited, missing the organisers","detail":"Restricting individual mules without mapping the network leaves the organisers active. Work at cluster level and share intelligence."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Detecting mule accounts is fraud and AML detection by a private firm, which Annex III does not list; point 5(b) explicitly excludes systems used to detect financial fraud from the credit scoring category. Restricting an account based solely on an automated score can be a decision with similarly significant effects under GDPR Article 22, so keep a human decision and a route to challenge."},"regulations":["eu-ai-act","gdpr","uk-gdpr","fatf-recommendations","uk-consumer-duty","dora","mas-ai-risk-management","us-sr-11-7","eu-amlr","us-bsa","uk-psr-app-reimbursement"],"guidance":[{"title":"APP scams","issuer":"Payment Systems Regulator","region":"europe","url":"https://www.psr.org.uk/our-work/app-scams/","note":"UK reimbursement for authorised push payment scams over Faster Payments and CHAPS is split 50:50 between sending and receiving firms, which puts mule detection on the receiving bank's balance sheet."},{"title":"Guidelines on Shared Responsibility Framework","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/guidelines/guidelines-on-shared-responsibility-framework","note":"Singapore's framework, in force since 16 December 2024, that assigns anti phishing duties to financial institutions and telcos and requires payouts to scam victims where those duties are breached."},{"title":"COSMIC, Collaborative Sharing of ML/TF Information and Cases","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/anti-money-laundering/cosmic","note":"Platform launched by MAS with six major banks in April 2024 for sharing red flag information on customers across institutions. It currently covers misuse of legal persons, trade finance and proliferation financing, not retail mule accounts."},{"title":"Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) Committee report","issuer":"Reserve Bank of India","region":"asia-pacific","url":"https://www.rbi.org.in/Scripts/PublicationReportDetails.aspx?UrlPage=&ID=1306","note":"India's framework for responsible AI in the financial sector (August 2025), which names fraud detection as a high stakes use and recommends that AI models are validated and tested periodically, including for drift and bias."}],"controls":["Documented decision process for restrictions and exits, with recorded reasons","Customer remediation route with a service level for review","Model inventory entry, validation and fairness testing","Legal basis documented for every external data sharing arrangement","Audit trail linking each restriction to the network evidence and the investigator"],"incidents":[]},"blitsAi":{"howToBuild":"The graph and scoring models run in the bank's analytics platform. Blits.ai adds the\ninvestigation layer: an **agentic workflow**, triggered through the API for each flagged\ncluster, calls **custom functions** and **SQL knowledge bases** to pull the flows, accounts and\nprior cases, and drafts a fund flow timeline and case narrative as **structured output**, with\neach statement tied to the underlying records.\n\nRestrictions, recall requests and exits go through **human in the loop approval**, with a full\naudit trail per run. Where the bank wants to check a flagged customer, a **conversational\nagent** in the bank's app (through the **API channel**), on **WhatsApp** or on **voice** can ask about the purpose of recent payments and\nhand over to a specialist. **PII masking**, **guardrails** and **test suites** apply throughout,\nand the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"Why does mule detection need graph analytics?","answer":"A mule account often looks ordinary on its own. The signal is in the links: many victims paying in, money leaving quickly to the same onward accounts, and devices or contact details shared across accounts opened around the same time."},{"question":"Can banks detect mules together?","answer":"Increasingly yes. In Australia, five large banks joined BioCatch Trust Australia in November 2024 to share intelligence on receiving accounts before a payment leaves. In India, the central bank's innovation hub offers MuleHunter.AI to banks; the Governor said in October 2025 that it had scaled to 21 banks."},{"question":"Should an AI model freeze accounts automatically?","answer":"No. Freezing or exiting an account is high impact for the customer and often irreversible in practice. Let the model find and prioritise, and let a trained investigator decide with the evidence in front of them."}],"related":["real-time-fraud-scoring","scam-payment-interception","aml-alert-triage","fraud-alert-triage","suspicious-activity-report-drafting","application-and-identity-fraud-detection"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Blocker fix: corrected the BioCatch Trust Australia summary to say the currency is not stated instead of assuming AUD; reworded the BigPay evidence summary so it credits BigPay's result to analyst built rules and shows the AI feature as a forward looking layer, not the cause; removed the mule detection link from the Pay dot UK and Visa pilot evidence, since its own release does not mention money mules."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added EU Anti Money Laundering Regulation, Bank Secrecy Act, UK APP scam reimbursement rules to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: softened unsourced timing claims, tied the receiving bank liability to the UK 50:50 APP rule, corrected the FTC wording, the COSMIC and FREE-AI guidance notes and the GDPR basis, replaced the COSMIC FAQ example with BioCatch Trust Australia, added UK GDPR, SEO title and meta description."},{"date":"2026-09-26","note":"Second fact check against sources: removed the BigPay 90 percent figure from the fraud loss reduction benchmark (it measures mule activity, not losses), added the BioCatch launch release naming the five banks, corrected source titles and dates for the RBI speech and MediaNama, tightened the problem wording, the FREE-AI note and the meta description."},{"date":"2026-09-26","note":"Third fact check against sources: all quotes, guidance notes and FAQ facts confirmed; BigPay evidence summary now describes the rules based approach precisely and carries its publication date, and the MuleHunter.AI metric period is dated to October 2025."}],"slug":"mule-network-detection","url":"https://www.blits.ai/ai-use-cases/mule-network-detection","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":180000000,"min":180000000,"max":180000000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"biocatch-trust-australia-mule-intelligence","pooled":true}]}],"indicativeValueResult":{"low":150000,"high":1200000},"evidence":["bigpay-feedzai-mule-detection","biocatch-trust-australia-mule-intelligence","reserve-bank-innovation-hub-mulehunter-ai"]},{"title":"AI for non emergency service requests and 311 routing","shortTitle":"Non emergency service request routing","seoTitle":"AI for 311 and non emergency service requests","metaDescription":"AI agents answer 311 and non emergency calls, create service cases and route urgent ones. Prepared reports 73% of Galt PD call volume handled before a dispatcher.","definition":"An AI agent on a city's 311 style phone, chat and messaging channels that answers routine municipal questions, takes service requests such as potholes, missed collections or broken street lights with the right location and details, creates the case in the work order system and routes anything urgent or complex to the right team.","aliases":["311 chatbot","municipal service request bot","non emergency line AI","city service request assistant"],"industries":["government"],"functions":["citizen-services","customer-service","case-management"],"patterns":["conversational-agent","voice-agent","classification-and-routing","agentic-workflow"],"channels":["voice","web-chat","whatsapp","mobile-app","sms"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Local government runs dozens of services, and residents contact it about all of them through the\nsame few channels: a 311 or general number and a website. Most contacts are\nroutine (collection days, opening hours, a pothole, a noisy neighbour), but each one needs a person\nto listen, find the right department, type the location and create a case. At the same time\npolice and 911 centres spend dispatcher time on non emergency calls to their ten digit lines,\npulling attention from real emergencies.\n\nMenus and web forms do not fix this: residents do not know which department owns their problem,\nand forms lose the detail (exact location, photos) that crews need.","problemStats":[],"howItWorks":"1. **Listen and classify.** The agent asks what the resident needs and classifies it into the\n   city's service catalogue, or detects that it is actually an emergency and transfers immediately.\n2. **Answer information questions.** Collection days, bylaws or opening hours are answered from\n   the city's content and data, including address specific answers from GIS.\n3. **Capture the request.** For a service request, the agent collects the location (address, map\n   pin or photo), description and contact details, and checks for duplicates nearby.\n4. **Create and route the case.** It creates the case in the work order or CRM system with the\n   right category and priority and tells the resident the reference number.\n5. **Keep the resident updated.** Status updates go back on the same channel until the case closes.\n6. **Hand over.** Complex, sensitive or vulnerable cases go to a contact centre agent with the\n   conversation attached.","valueDrivers":["cost-to-serve","customer-experience","speed","inclusion-and-access"],"kpis":["contact-deflection","interactions-handled","accuracy","response-time-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A city of 500,000 residents with 400,000 contacts a year to its 311 service","inputs":[{"key":"contacts","label":"311 contacts per year across phone, chat and messaging","low":400000,"high":400000,"unit":"contacts per year","note":"The reference city."},{"key":"automatable","label":"Share of contacts that are routine questions or simple service requests","low":0.5,"high":0.7,"unit":"fraction of contacts","note":"Editorial assumption. Prepared's Galt case study says more than 73% of calls to that police department are non emergency, which is a different mix; replace with your own data."},{"key":"containment","label":"Share of those the agent completes without an agent","low":0.25,"high":0.5,"unit":"fraction of automatable contacts","note":"Editorial assumption, conservative against the contact deflection benchmark on this page."},{"key":"costPerContact","label":"Cost of an agent handled contact","low":4,"high":7,"unit":"USD per contact","note":"Editorial assumption. Replace with your own fully loaded cost."}],"formula":"contacts * automatable * containment * costPerContact","currency":"USD","period":"per year","resultLabel":"Agent handled contact cost avoided","caveat":"Gross contact cost only. It leaves out better case data for crews, fewer duplicate reports, dispatcher time protected on police lines and the cost of integration and running the agent."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering questions is simple; creating good cases needs integration with the work order or CRM system, a clean service catalogue, location capture and duplicate checks, and a safe transfer path for emergencies.","dataPrerequisites":["The city's service request catalogue with categories, owners and priorities","Municipal content and address level data (collection schedules, zoning, service areas)","Historic 311 cases to train and test classification"],"integrations":["311 CRM or work order system (case creation and status)","GIS and address lookup","Telephony, web chat and messaging channels","Contact centre platform for handover and emergency transfer"]},"implementation":{"steps":[{"title":"Clean the service catalogue first","detail":"Agree the categories, required fields and owning team for each request type; the agent can only route as well as the catalogue allows."},{"title":"Start with the top ten requests","detail":"Launch with the highest volume information questions and two or three simple request types, such as missed collections and potholes."},{"title":"Build the emergency exit","detail":"Define phrases and signals that trigger an immediate transfer to 911 or the dispatcher, and test them with real transcripts, as Galt's agent transfers genuine emergencies at once."},{"title":"Capture location well","detail":"Use address validation, map pins or photos (TAMM's assistant helps fill in a report from a photo) so crews can find the problem the first time."},{"title":"Close the loop","detail":"Send status updates and closure notices on the channel the resident used, and measure repeat reports."}],"guardrails":["Immediate transfer on any sign of emergency, with a tested phrase list and a low threshold","Case creation only in defined categories with required fields validated","Answers only from city content and data, with refusal outside scope","Personal data minimised and masked in logs","Clear AI disclosure and a way to reach a person during office hours"],"humanInTheLoop":"Contact centre agents take handovers, complaints and sensitive cases; supervisors review samples of contained conversations and misrouted cases every week; department owners approve changes to categories and priorities.","kpisToInstrument":["Containment per request type and share of calls transferred as emergencies","Routing accuracy (cases moved to another department after creation)","Share of cases with a valid location and complete required fields","Time from report to case creation and to resolution","Duplicate reports and repeat contacts on the same issue"],"failureModes":[{"title":"A missed emergency","detail":"A caller on a non emergency line describes an emergency in vague words. Keep the transfer threshold low and review every transferred and non transferred call with emergency keywords."},{"title":"Wrong department, lost case","detail":"Misrouted cases bounce between teams. Measure reassignment and fix the catalogue."},{"title":"Accuracy only where content exists","detail":"Barnet's transparency record gives the Ami model an average of 90% correct answers, but only for queries it has content for (a model figure, not a measured pilot result); questions outside that content fail. Track unanswered questions, not only accuracy."},{"title":"Reports without follow up","detail":"Residents report, hear nothing and call again. Send status updates."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A 311 assistant must disclose that it is AI (Article 50). It is not high risk while it only informs and creates service cases. If it evaluates or classifies emergency calls or sets dispatch priority for police, fire or medical services, it falls under Annex III point 5(d) and becomes high risk."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Residents must be informed that they are interacting with an AI system."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(d) covers classification of emergency calls and dispatch priority, the boundary a non emergency line must not cross silently."},{"title":"Algorithmic Transparency Recording Standard Hub","issuer":"Government Digital Service","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"Recommended for local government in the UK; Barnet and Newcastle councils publish records."}],"controls":["AI disclosure and a published description of what the agent can and cannot do","Tested emergency transfer path, reviewed after every change","Access control and retention limits on conversation logs and location data","Weekly review of misrouted cases with department owners","Accessibility testing for voice and chat channels"],"incidents":[{"title":"NYC's AI chatbot tells businesses to break the law","url":"https://themarkup.org/news/2024/03/29/nycs-ai-chatbot-tells-businesses-to-break-the-law","note":"In tests by The Markup, New York City's business chatbot told landlords and business owners that illegal practices, such as refusing housing vouchers or taking workers' tips, were allowed; the same risk applies to municipal information on 311 channels."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** grounded in a **knowledge base** of municipal content, with\n**custom functions** that look up address data and create cases in the city's CRM or work order\nsystem through REST calls. Service request journeys run as **flows** with slot filling for\nlocation and details, validation blocks and a receive attachment block for photos; a condition\non emergency language triggers an immediate **redirect call** or **agent handover**.\n\nThe same agent answers on **voice, web chat, WhatsApp and SMS**, with streaming speech\nrecognition and call transfer on the phone and language detection for multilingual residents.\n**Guardrails** and **PII masking** protect residents' data, **test suites** run conversation\nsets that include emergency phrases before each release, and **analytics** show top intents,\nflow statistics and unanswered questions per channel."},"faq":[{"question":"How much of a non emergency line can AI handle?","answer":"It varies by call mix. Prepared, the vendor, reports that 73% of Galt Police Department's call volume is now handled before it reaches a dispatcher, in a department where more than 73% of calls are non emergency. Microsoft reports that Kelowna's assistant answers 80% of snowplow calls correctly, so measure by request type rather than for the line as a whole."},{"question":"Does AI on a non emergency line fall under the EU AI Act high risk rules?","answer":"Not while it only answers questions and creates service cases; the Article 50 duty to disclose that residents are talking to AI still applies. It becomes high risk under Annex III point 5(d) if it evaluates or classifies emergency calls or sets dispatch priority for emergency services, so keep emergency detection as a simple transfer to a human dispatcher who decides, and document that design choice."},{"question":"What volumes do city assistants handle?","answer":"Google reports more than 30,000 conversations a month on Rio de Janeiro's 1746 chatbot, and Microsoft reports more than 20,000 conversations for Montgomery County's Monty 2.0 since its beta. Start small and scale with the catalogue."}],"related":["citizen-information-assistant","emergency-call-triage-support","public-service-translation","correspondence-triage-and-routing","first-line-contact-centre-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from UK transparency records, city and vendor case studies, with every quote verified against the source."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed the Barnet 90% accuracy metric (a model level average in a pre pilot record, not a Barnet result), attributed the Galt, Rio de Janeiro and Montgomery County figures to their vendors, corrected the NYC incident note and the EU AI Act FAQ, added seoTitle and metaDescription."}],"slug":"non-emergency-service-request-routing","url":"https://www.blits.ai/ai-use-cases/non-emergency-service-request-routing","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":25000,"min":20000,"max":30000,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"rio-de-janeiro-1746-citizen-service-chatbot","pooled":true},{"id":"montgomery-county-monty-chatbot","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":80,"min":80,"max":80,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"city-of-kelowna-311-ai-assistant","pooled":true}]},{"kpi":"contact-deflection","label":"Contact deflection","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":73,"min":73,"max":73,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"galt-police-department-non-emergency-call-triage","pooled":true}]},{"kpi":"customer-satisfaction","label":"Customer satisfaction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"montgomery-county-monty-chatbot","pooled":true}]}],"indicativeValueResult":{"low":200000,"high":980000},"evidence":["abu-dhabi-tamm-ai-assistant","barnet-council-ami-chatbot","city-of-kelowna-311-ai-assistant","galt-police-department-non-emergency-call-triage","montgomery-county-monty-chatbot","newcastle-city-council-contact-centre-ai","rio-de-janeiro-1746-citizen-service-chatbot"]},{"title":"AI for payment investigations and exceptions","shortTitle":"Payment investigations and exceptions","seoTitle":"AI for payment exceptions and investigations","metaDescription":"AI sorts failed payments, drafts ISO 20022 investigation messages and chases banks. BNY says a digital employee handles over 10% of its payment repair issues.","definition":"AI that works the payments that fall out of straight through processing: it reads the failure, repairs or enriches the message, drafts the ISO 20022 or SWIFT investigation, chases the counterparty bank and proposes a return, recall or correction, while an operator approves anything that moves money.","aliases":["payment exceptions and investigations automation","payment repair AI","SWIFT case management automation","E&I automation"],"industries":["banking","payments"],"functions":["operations","customer-service"],"patterns":["agentic-workflow","document-processing","classification-and-routing","content-generation"],"channels":["internal-tools","api","email"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"back-office","problem":"A payment that flows straight through needs no manual work, but every payment that falls out\ndoes. A missing or malformed field, a name that does not match the account, a sanctions hit, a\nduplicate or a customer asking \"where is my payment\" each opens a case. Operators then read MT\nand MX messages, look up the payment in several systems, write free text queries to\ncorrespondent banks and wait for an answer that may arrive by message, email or not at all.\n\nMuch of this work is reading and writing: free format messages such as the MT199, emails and\ncustomer queries, while the customer keeps asking for news. ISO 20022 defines structured\nmessages for exceptions and investigations, such as the interbank payment cancellation request\n(camt.056) and its response (camt.029), the payment status request (pacs.028) and the\ninvestigation request and response (camt.110 and camt.111). The Committee on Payments and Market\nInfrastructures recommends that payment system operators and participants align with its\n[harmonised ISO 20022 data requirements](https://www.bis.org/cpmi/publ/d230.htm) for cross\nborder payments before the end of 2027, and expects correct account data to mean fewer\nexceptions and investigations. Structured data makes cases easier to classify; AI does the\nreading, drafting and chasing that remain.\n\nBanks have started with the repair step. In a\n[Microsoft customer story](https://www.microsoft.com/en/customers/story/27322-bny-microsoft-365-copilot),\nBNY says a digital employee in its Eliza platform repairs missing or incomplete payment\ninstructions and handles over ten percent of its payment repair issues around the world.","problemStats":[],"howItWorks":"1. **Classify the exception.** The agent reads the rejected or held payment, the error codes and\n   any inbound query (camt.056 recall, camt.110 investigation request, camt.029 or camt.111\n   response, gpi tracker status, free text MT199 or email) and assigns the case type.\n2. **Gather the facts.** It pulls the payment's lifecycle from the payment hub, the tracker and\n   the customer record, and retrieves the relevant scheme rules and internal procedures.\n3. **Repair or propose.** For repairable errors it proposes the corrected fields (for example a\n   BIC derived from the IBAN) with its reasoning. For investigations it proposes the next action:\n   request information, recall, return or beneficiary correction.\n4. **Draft the messages.** It drafts the structured investigation message to the counterparty and\n   the plain language update to the customer or the relationship manager.\n5. **Approve and chase.** An operator approves any repair, recall or credit adjustment. The agent\n   then sends, tracks deadlines, chases unanswered queries and closes the case with a full trail.","valueDrivers":["cost-to-serve","speed","customer-experience","risk-reduction"],"kpis":["automation-rate","processing-time-reduction","handling-time-reduction","cycle-time-days","interactions-handled"],"indicativeValue":{"referenceOrg":"A regional bank handling 100,000 payment exception and investigation cases a year","inputs":[{"key":"cases","label":"Exception and investigation cases per year","low":100000,"high":100000,"unit":"cases per year","note":"The reference bank. Replace with your own case volume."},{"key":"minutesPerCase","label":"Operator minutes per case today","low":20,"high":45,"unit":"minutes per case","note":"Editorial assumption covering reading, lookups, drafting and follow up. Replace with your own time study."},{"key":"effortReduction","label":"Share of operator time the AI removes","low":0.25,"high":0.5,"unit":"fraction of time per case","note":"Editorial assumption; the AI drafts and gathers, a human still approves money movement."},{"key":"costPerHour","label":"Fully loaded operations cost per hour","low":35,"high":60,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"cases * minutesPerCase / 60 * effortReduction * costPerHour","currency":"USD","period":"per year","resultLabel":"Investigation effort avoided","caveat":"Labour only. It leaves out fewer customer chasers, lower compensation and claim costs from faster resolution, and the cost of the platform and the integration with the payment hub."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Cases span the payment hub, sanctions filtering, the SWIFT interface, tracker data and customer systems, and every recall or repair moves or redirects money. Structured ISO 20022 case messages help, but many counterparties still answer in free text.","dataPrerequisites":["Case history with case type, actions taken and outcome","Payment lifecycle data from the payment hub and tracker","Scheme rulebooks and internal procedures for each case type","Standing settlement instructions and correspondent bank static data"],"integrations":["Payment hub or payment engine (repair queues, returns)","SWIFT interface, gpi tracker and case management service","Sanctions and fraud filtering systems (read only)","Case management or CRM for the customer side","Email and secure messaging for counterparties that do not use structured messages"]},"implementation":{"steps":[{"title":"Map the case types","detail":"Take a quarter of cases and group them by type (repair, unable to apply, claim non receipt, recall, fee query, duplicate). Volume and handling time per type decide where to start."},{"title":"Automate the reading and the gathering first","detail":"Before any drafting, let the AI classify cases and assemble the facts into the case file. This targets the lookup time first and is low risk, because nothing leaves the bank."},{"title":"Add drafting with templates","detail":"Draft structured messages from the case data and free text only where the counterparty requires it, with operators approving every outgoing message at first."},{"title":"Introduce straight through handling per case type","detail":"Once a case type shows stable quality, allow the agent to send information requests and chasers on its own, while repairs, recalls and credits keep maker checker approval."},{"title":"Close the loop with the customer","detail":"Connect the case status to the customer channel so the front office and the customer see the same state without calling operations."}],"guardrails":["Maker checker approval on every repair, recall, return or credit adjustment","The agent never overrides or clears a sanctions or fraud hit","Outgoing messages validated against the ISO 20022 schema before sending","Customer updates use approved wording and never speculate on the outcome"],"humanInTheLoop":"Operators approve every action that moves or redirects money and own cases that involve fraud, sanctions or a complaint. Team leads review a sample of automated chasers and closures weekly.","kpisToInstrument":["Cases resolved without manual lookup, per case type","Median days to resolution per case type","Operator minutes per case","Share of outgoing messages rejected or queried by counterparties","Customer chasers per case"],"failureModes":[{"title":"Wrong repair sent at scale","detail":"A plausible but wrong field correction sends money to the wrong place. Keep human approval on repairs and validate against static data."},{"title":"Automated chasing that annoys counterparties","detail":"Duplicate or badly timed chasers damage correspondent relationships. Respect agreed response windows and deduplicate."},{"title":"Case file and customer story drift apart","detail":"The customer is told something the case does not support. Generate customer updates only from the case status."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Handling payment exceptions is not a use listed in Annex III and is not a prohibited practice under Article 5. If the agent interacts directly with customers, for example in a chat about the case, Article 50(1) requires that they are told they are interacting with an AI system."},"regulations":["eu-ai-act","gdpr","dora","fatf-recommendations","apra-cps-230","eu-psd2"],"guidance":[{"title":"Harmonised ISO 20022 data requirements for enhancing cross border payments (updated report)","issuer":"Committee on Payments and Market Infrastructures","region":"global","url":"https://www.bis.org/cpmi/publ/d230.htm","note":"Data requirements developed with the Payments Market Practice Group, first published in October 2023 and updated in February 2026 with a separate technical annex. They are not regulatory requirements, but the CPMI encourages adoption by the end of 2027. The report lists the ISO 20022 return and investigation messages in its core message set and says the unique end to end transaction reference simplifies exception and investigation handling and enables its automation."}],"controls":["Maker checker on money movement, with the approver recorded on the case","Full case trail of inputs, drafts, approvals and messages for audit and complaints","Schema validation of every outgoing ISO 20022 message","Monthly quality sample per case type"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that the payment hub or case system starts through\nthe API, with scheduled runs for chasers.\n**Custom functions** call the payment hub, the tracker and the case system through REST, the\n**knowledge base** holds scheme rules and procedures with hybrid retrieval, and the agent drafts\nmessages with **structured output** so they can be validated before sending. With **human in\nthe loop approval** configured for money movement, every repair, recall, return or credit waits\nfor an operator's confirmation before the workflow continues, and a **tool execution policy**\nlimits which tools the agent may call on its own.\n\nThe **email channel** handles counterparties that still reply in free text, and the same case\nstatus can feed a customer facing agent. **Guardrails** and **PII masking** protect customer\ndata in prompts, every run has a **full audit trail**, and **test suites** built from historical\ncases run before a new case type goes live."},"faq":[{"question":"Does ISO 20022 remove the need for AI in payment investigations?","answer":"No, it makes AI more useful. ISO 20022 defines structured exception and investigation messages, which make cases easier to classify and automate when both banks use them, but counterparties can still reply late or in free text, and someone still has to gather the facts, decide on the next step and keep the customer informed."},{"question":"Can an AI agent recall or repair a payment on its own?","answer":"It should not. The agent can propose the repair or recall with its reasoning and draft the message, but a person approves any action that moves or redirects money, and sanctions or fraud hits are never cleared by the agent."},{"question":"What results have banks published?","answer":"Very few so far. In a Microsoft customer story, BNY says a digital employee in its Eliza platform repairs missing or incomplete payment instructions and handles over ten percent of its payment repair issues worldwide. J.P. Morgan's AI payment validation screening works one step earlier, on preventing exceptions rather than investigating them. BNY also reports faster handling of client transaction inquiries in the same story, but does not say how many of those are payment investigations rather than other transaction queries."}],"related":["ledger-and-payment-reconciliation","chargeback-and-representment","sanctions-screening-adjudication","corporate-client-servicing-assistant","correspondence-triage-and-routing","scam-payment-interception"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the editor after a targeted blocker check."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog. Kept in draft because only one public AI deployment with verified results was found."},{"date":"2026-09-25","note":"Consolidation pass: added PSD2 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: the J.P. Morgan quote is confirmed; the problem text now cites the CPMI harmonised data requirements instead of unsourced history; the CPMI guidance note, the EU AI Act basis (Article 5, Article 50(1)) and the approval wording in the Blits.ai section were made precise; the FAQ now dates the J.P. Morgan result; added seoTitle and metaDescription. Kept in draft: one public evidence record."},{"date":"2026-09-26","note":"Second fact check: the CPMI guidance now points to the February 2026 updated report; removed unsourced wording in the problem text and the playbook; the Blits.ai section now names only API and scheduled workflow triggers; added error reduction to the KPIs to match the J.P. Morgan metric; the J.P. Morgan record is now production stage with no assumed vendor, channel or language. Kept in draft: one public evidence record."},{"date":"2026-09-27","note":"Review fixes: adoption stage lowered to emerging, since no source shows several production deployments; the meta description no longer cites the J.P. Morgan figure, which measures exception prevention at validation, not investigation; removed error reduction from the KPIs and the metric from the J.P. Morgan record, which stays as context only; the FAQ now names BNY's payment repair digital employee (Microsoft customer story); added pacs.028, camt.110 and camt.111 from the CPMI updated report; reworded the approval and test suite lines in the Blits.ai section. Kept in draft until an in scope evidence record is linked."},{"date":"2026-09-27","note":"Fact checked against sources: the J.P. Morgan quote, date and forum were rechecked on the live page; the BNY payment repair statement in the FAQ was confirmed on the Microsoft customer story and is now cited inline in the problem text, since no BNY record is linked to this page; the ISO 20022 message names and the 2027 wording now follow the CPMI updated report (26 February 2026); the CPMI guidance note now says what the report says about exception handling; the meta description names BNY; the feasibility note no longer calls every repair a regulated action; the Blits.ai section now says approval is configured for money movement and names the tool execution policy. Kept in draft: one public evidence record."}],"slug":"payment-investigations-and-exceptions","url":"https://www.blits.ai/ai-use-cases/payment-investigations-and-exceptions","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10,"min":10,"max":10,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bny-eliza-payment-repair-digital-employee","pooled":true}]}],"indicativeValueResult":{"low":291666.6666666667,"high":2250000},"evidence":["bny-eliza-payment-repair-digital-employee","jpmorgan-payment-validation-screening"]},{"title":"AI for PEP and adverse media screening","shortTitle":"PEP and adverse media screening","seoTitle":"AI for PEP and adverse media screening","metaDescription":"AI checks news and records for PEP links and adverse media, sets aside namesakes and cites sources. WorkFusion reports 95% fewer false positives at Scotiabank.","definition":"AI that continuously scans news, court records, registries and other open sources in many languages for negative information and political exposure linked to customers, counterparties and beneficial owners, discards look alikes, and summarises credible risk for the analyst with the sources attached.","aliases":["adverse media screening","negative news screening","politically exposed person screening","AI powered media monitoring for KYC"],"industries":["banking","payments","wealth-and-asset-management"],"functions":["financial-crime-compliance","onboarding-and-kyc"],"patterns":["rag-knowledge-assistant","summarization","classification-and-routing","translation"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"middle-office","problem":"Due diligence requires banks to know whether a customer, a director or a beneficial owner is a\npolitically exposed person or has been linked to crime, corruption or other serious wrongdoing.\nCurated databases cover only part of the world's news, and keyword searches on the open web\nreturn pages of irrelevant hits: people with the same name, old stories, opinion pieces.\n\nAnalysts read article after article to rule out namesakes, often in languages they do not speak,\nand the result is inconsistent. Real risk gets missed in the noise, while onboarding and periodic\nreviews slow down. The quality of the written conclusion, why a hit was or was not relevant, is\nwhat auditors check, and it often varies between analysts.","problemStats":[],"howItWorks":"1. **Search broadly.** For each subject the system queries curated risk databases, news\n   archives, court and regulatory records and the open web, in the languages that match the\n   subject's footprint.\n2. **Disambiguate.** Entity resolution compares each article's person or company with the\n   subject's known attributes (age, location, occupation, associated companies) and discards\n   look alikes with a stated reason.\n3. **Classify the risk.** Relevant articles are classified by risk category (fraud, corruption,\n   sanctions evasion, organised crime) and by credibility and recency of the source.\n4. **Summarise with citations.** The system writes a short summary of the credible findings,\n   translated where needed, with a link to every source article.\n5. **Analyst decides.** The analyst confirms relevance and source reliability, records the\n   disposition and decides whether it changes the customer's risk rating; monitoring continues\n   between reviews.","valueDrivers":["compliance","employee-productivity","speed","risk-reduction"],"kpis":["false-positive-reduction","alert-volume-reduction","handling-time-reduction","processing-time-reduction","time-saved-per-task","productivity-gain","accuracy"],"indicativeValue":{"referenceOrg":"A bank running 40,000 adverse media reviews a year across onboarding and periodic reviews","inputs":[{"key":"reviews","label":"Adverse media reviews per year","low":40000,"high":40000,"unit":"reviews per year","note":"The reference bank."},{"key":"minutesPerReview","label":"Analyst minutes per review today","low":15,"high":40,"unit":"minutes per review","note":"Editorial assumption. Replace with your own time study."},{"key":"timeSaved","label":"Share of review time saved","low":0.3,"high":0.5,"unit":"fraction of review time","note":"Conservative against the benchmark on this page (Xapien reports that Save the Children cut donor due diligence review times by over 60% with its AI due diligence tool). Replace with results from your own pilot."},{"key":"costPerHour","label":"Fully loaded analyst cost per hour","low":35,"high":60,"unit":"USD per hour","note":"Editorial assumption. Replace with your own."}],"formula":"reviews * minutesPerReview / 60 * timeSaved * costPerHour","currency":"USD","period":"per year","resultLabel":"Analyst capacity released","caveat":"Counts analyst time only. It leaves out faster onboarding, risk found that manual searches missed, data licence costs and the cost of the platform."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Retrieval and summarisation are mature. The hard parts are reliable disambiguation of common names, licensed access to news content, and keeping the analyst accountable for the conclusion.","dataPrerequisites":["Subject attributes for disambiguation (date of birth, nationality, addresses, related companies)","Licensed news and risk data sources, plus rules for which open web sources count as credible","The bank's adverse media risk taxonomy and materiality criteria","Historical dispositions to measure false positive rates"],"integrations":["KYC and customer due diligence system","Screening engine and PEP database","News and risk data providers","Case management and customer risk rating"]},"implementation":{"steps":[{"title":"Define what counts as adverse","detail":"Write down the risk categories, how old a story may be, and which sources count as credible, with examples. The model can only be as consistent as the policy."},{"title":"Get disambiguation right","detail":"Measure how often the system wrongly matches or wrongly discards a namesake on a labelled sample, per language and naming culture, before analysts rely on it."},{"title":"Summaries with sources, never without","detail":"Require a link to every source in the summary and reject any claim that is not supported by a retrieved article."},{"title":"Pilot on periodic reviews","detail":"Start with periodic reviews of existing customers, where time pressure is lower, then extend to onboarding and continuous monitoring."}],"guardrails":["Every finding links to its source; unsupported statements are rejected","Adverse media changes a risk rating only after an analyst confirms relevance and reliability","Bias testing across names, nationalities and languages for both false hits and misses","Source articles and dispositions retained for audit","Licence terms respected for every news source"],"humanInTheLoop":"The system searches, filters and summarises; the analyst decides whether a finding is about the subject, whether it is credible and whether it matters. Changes to risk rating or relationship decisions stay with named people.","kpisToInstrument":["Hits per subject presented to analysts, before and after","Share of analyst overturned discards and matches on a labelled sample","Review time per subject","Material findings per thousand reviews","Miss rate on a known test set of adverse subjects"],"failureModes":[{"title":"Namesake contamination","detail":"A common name links a customer to someone else's crimes. Require multiple matching attributes and show them in the summary."},{"title":"Language and culture bias","detail":"Disambiguation works well for some naming conventions and badly for others. Measure performance per language and naming culture."},{"title":"Summary replaces reading","detail":"Analysts stop opening sources. Sample decisions against the source articles and keep the analyst's conclusion in their own words."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Adverse media and PEP screening for due diligence is not listed in Annex III. It processes personal data, including data about alleged offences, so GDPR Article 10 and national AML law govern what may be collected and how long it is kept."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","mas-ai-risk-management","cbuae-ai-guidance","nist-ai-rmf","eu-amlr","us-bsa","mas-notice-626"],"guidance":[{"title":"Principles for Using Artificial Intelligence and Machine Learning in Financial Crime Compliance","issuer":"Wolfsberg Group","region":"global","url":"https://wolfsberg-group.org/resources/202/93","note":"Industry principles for legitimate, proportionate and transparent use of AI in financial crime compliance."},{"title":"Notice 626 Prevention of Money Laundering and Countering the Financing of Terrorism, Banks","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/notices/notice-626","note":"Example of national rules on customer due diligence, PEP checks and ongoing monitoring."}],"controls":["Written adverse media policy with risk categories, recency and source credibility rules","Source links and analyst disposition retained for every finding","Bias and accuracy testing per language and naming culture","Human decision on any change to a risk rating or relationship"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** with the **web search** and **web page browsing**\ntools, plus **custom functions** that call the bank's licensed news and risk data providers. The\nagent searches in the subject's languages, uses **language translation** where needed, compares\neach article with the subject's attributes and returns **structured output**: relevant findings,\ndiscarded look alikes with reasons, and a summary with a source link for every statement.\n\nThe bank's adverse media policy sits in a **knowledge base** with hybrid retrieval, so the agent\napplies the same categories and credibility rules every time. Changes to a risk rating go\nthrough **human in the loop approval**, runs keep a full audit trail, **test suites** hold known\nadverse subjects and namesakes, and the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"Is adverse media proof of risk?","answer":"No. It is an input to a risk decision. The analyst checks that the article is about the customer, that the source is credible and that the allegation is material before it changes a rating."},{"question":"How does AI reduce adverse media false positives?","answer":"Mostly through disambiguation: comparing ages, locations, occupations and related companies in the article with what the bank knows about the customer, and discarding namesakes with a stated reason. Measure it on a labelled sample in every language you screen."},{"question":"What about bias against certain names?","answer":"It is a real risk. Names that are common in a community, or that are transliterated from another script in several ways, can produce more false hits and so more manual scrutiny for some customers. Test false hit and miss rates per naming culture and language, and fix the gaps before scaling."}],"related":["sanctions-screening-adjudication","perpetual-kyc","dynamic-customer-risk-rating","business-onboarding-and-ubo-discovery","source-of-wealth-diligence","vendor-due-diligence"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: fixed the claimant of the Save the Children figure (Xapien), removed an unsupported superlative, added the KPIs the evidence reports (alert volume, handling time, cycle time), corrected the Santander UK and Banxa summaries and the Scotiabank languages, added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Second fact check: removed the Banxa 43.5% match reduction from the metrics because ComplyAdvantage attributes it to screening configuration, not AI, and corrected its year to 2023 from the page metadata; set Deutsche Bank to 2020 and English only, and scoped its handling time range to the whole KYC programme; moved the Save the Children figure to handling time and aligned its summary with the case study."},{"date":"2026-09-27","note":"Third fact check: all quotes rechecked on the live pages; added the publication dates from page metadata to the Scotiabank, Deutsche Bank and Santander UK sources, tightened the Scotiabank summary to what the story states, and softened an unsourced FAQ claim about name bias."},{"date":"2026-09-27","note":"Fourth fact check: removed the Santander UK 12 to 2 days cycle time metric because the source credits the whole digital onboarding proposition, not adverse media screening, and never describes the screening as AI (the same issue that already removed Banxa's figure); removed the Banxa record, which supported no AI use case; dropped the cycle time days KPI from the KPI list since no remaining evidence supports it."}],"slug":"pep-and-adverse-media-screening","url":"https://www.blits.ai/ai-use-cases/pep-and-adverse-media-screening","benchmarks":[{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":1,"median":60,"min":60,"max":60,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"save-the-children-xapien-donor-due-diligence","pooled":true},{"id":"deutsche-bank-workfusion-screening-automation","pooled":false}]},{"kpi":"false-positive-reduction","label":"False positive reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":95,"min":95,"max":95,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"scotiabank-workfusion-adverse-media-monitoring","pooled":true}]}],"indicativeValueResult":{"low":105000,"high":800000},"evidence":["deutsche-bank-workfusion-screening-automation","hsbc-silent-eight-screening-automation","mashreq-silent-eight-alert-adjudication","ocbc-helios-agentic-due-diligence","santander-uk-complyadvantage-onboarding-screening","save-the-children-xapien-donor-due-diligence","scotiabank-workfusion-adverse-media-monitoring"]},{"title":"AI for permit and licence application processing","shortTitle":"Permit and licence application processing","seoTitle":"AI for permit and planning application checks","metaDescription":"AI checks permit and planning applications for missing items and drafts officer reports. Leeds reports 85%+ accuracy in initial testing, and officers still decide.","definition":"AI that helps applicants submit complete permit and licence applications and helps officers process them, by answering questions about requirements, checking applications for missing or inconsistent information, pulling the relevant policies, history and constraints, and drafting reports, while the grant or refusal stays with a named officer or a published rule.","aliases":["AI permit review","planning application validation AI","licensing assistant","building permit AI"],"industries":["government"],"functions":["citizen-services","case-management","regulatory-compliance"],"patterns":["document-processing","agentic-workflow","rag-knowledge-assistant","conversational-agent"],"channels":["web-chat","internal-tools","api"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","problem":"Permits and licences (planning and building permits, environmental permits, trade and wildlife\nlicences, trade marks) sit on the critical path of housing, infrastructure and business. Many\napplications arrive incomplete or wrong, so officers spend their first pass on validation:\nchecking documents against a checklist, chasing missing plans and fees, and searching several\nsystems for site history, policies and constraints before they can assess the merits. Applicants\noften learn that something was missing only after the application has been filed; at the UK\nIntellectual Property Office, trade mark applications that missed essential criteria used to be\nrejected automatically.\n\nMuch of this is document and rules work that AI can prepare. The decision itself weighs policy,\nobjections and local judgment, and applicants have appeal rights, so the officer has to stay the\ndecision maker and be able to show how the AI's input was used.","problemStats":[],"howItWorks":"1. **Help the applicant before submission.** An assistant explains which permit is needed and\n   what to submit, and pre checks (such as a trade mark similarity search) show likely problems\n   before the application is filed.\n2. **Validate on arrival.** Documents are read and checked against a checklist tailored to the\n   application type; missing or inconsistent items are flagged with the reason and source.\n3. **Assemble the context.** The system retrieves site history, policies and constraints from GIS\n   and document stores and proposes which apply.\n4. **Triage.** Simple applications that meet published rules (for example a categorical exclusion\n   in environmental review) are separated from those that need detailed assessment.\n5. **Draft the report.** The system drafts sections of the officer's report and letters to the\n   applicant for the officer to edit.\n6. **Decide and record.** The officer decides; an audit log shows what the AI suggested and what\n   the officer kept.","valueDrivers":["speed","employee-productivity","customer-experience","compliance"],"kpis":["processing-time-reduction","accuracy","time-saved-per-task","interactions-handled","cycle-time-days"],"indicativeValue":{"referenceOrg":"A local planning authority that receives 5,000 applications a year","inputs":[{"key":"applications","label":"Applications received per year","low":5000,"high":5000,"unit":"applications per year","note":"The reference authority."},{"key":"hoursSaved","label":"Officer hours saved per application on validation, research and report drafting","low":0.5,"high":1.5,"unit":"hours per application","note":"Editorial assumption. Leeds reports only a design aim (30% faster determination), not a measured saving; replace with your own time study."},{"key":"hourlyCost","label":"Fully loaded planning officer cost per hour","low":40,"high":60,"unit":"USD per hour","note":"Editorial assumption. Replace with your own cost."}],"formula":"applications * hoursSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Officer time released, valued at cost","caveat":"Values officer time only. It leaves out faster decisions for applicants, fewer invalid applications, fewer appeals from errors and the cost of the tool; released time in stretched planning teams usually goes to backlog."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Document extraction and checklist validation are mature; the effort is in encoding local validation requirements, connecting GIS and planning history, and building an audit trail that stands up in an appeal.","dataPrerequisites":["Validation checklists and requirements per application type","Local policies, constraints and planning history in searchable form","GIS layers for sites and constraints","A sample of past applications with validation outcomes for testing"],"integrations":["Permitting or planning case management system","Document management and applicant portal","GIS and address data","Correspondence and fee systems"]},"implementation":{"steps":[{"title":"Start with validation of simple applications","detail":"Pick the simplest, highest volume type (the Leeds pilot starts with householder applications and about ten officers) and automate the validation checklist first."},{"title":"Show sources for every suggestion","detail":"Every flag, policy reference and draft sentence needs its source and reasoning, so the officer can accept or reject it quickly and defend the decision later."},{"title":"Log what officers keep","detail":"Record each suggestion and the officer's action; this is your audit trail and your accuracy measure."},{"title":"Move checks upstream","detail":"Offer applicants the same checks before they submit, as the UK IPO does with its trade mark pre check, to cut invalid applications at source."},{"title":"Automate decisions only for simple published rules","detail":"Automated outcomes fit narrow permissions with clear rules and a way to appeal: West Berkshire rejects larger bin applications that miss its policy criteria automatically and sends the rest to its waste team. Keep discretionary permits with officers."}],"guardrails":["No automated recommendation to approve or refuse a discretionary permit","Sources and reasoning shown with every suggestion; officers accept or reject each one","Audit log of AI suggestions and officer actions for every case","Personal data redacted before documents are sent to external models","Fully automated decisions only for rule based permissions, with notice and a route to challenge"],"humanInTheLoop":"Validation and planning officers review every suggestion and make every decision. Team leaders review accuracy of suggestions each month; policy owners approve checklist and rule changes; appeals are handled entirely by people with access to the audit log.","kpisToInstrument":["Share of AI validation flags accepted by officers, by application type","Time from receipt to valid application and to decision","Share of applications returned as invalid, before and after applicant pre checks","Appeals and complaints citing errors in validation or reports","Officer time per application from time studies"],"failureModes":[{"title":"Rubber stamping","detail":"Officers accept suggestions without reading them. Measure acceptance rates, sample cases and show sources prominently."},{"title":"Local rules missed","detail":"A generic model misses a local policy or constraint. Retrieve from the authority's own policy index and test with local cases."},{"title":"Pilot never scales","detail":"A tool that saves time on householder cases stalls on complex ones. Add application types only when accuracy on each is proven."},{"title":"Automation without recourse","detail":"Automated decisions on permissions without a way to challenge can breach data protection rules on solely automated decisions. Provide notice and human review."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Permit and licence decisions are not listed as such in Annex III, so officer decision support is usually minimal risk, and an assistant that talks to applicants carries the Article 50 transparency duty. The exceptions are permits in an Annex III area: examining applications for visas and residence permits (point 7) and evaluating eligibility for essential public assistance benefits and services (point 5(a)) are high risk. Solely automated decisions with legal or similarly significant effects on a person fall under GDPR Article 22 whatever the tier."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"AI Playbook for the UK Government","issuer":"UK Government","region":"europe","url":"https://www.gov.uk/government/publications/ai-playbook-for-the-uk-government","note":"Guidance for civil servants and government organisations; principle 4 asks for meaningful human control at the right stages."},{"title":"Algorithmic Transparency Recording Standard Hub","issuer":"Government Digital Service","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"Leeds, the IPO and West Berkshire publish records for the tools on this page."},{"title":"Directive on Automated Decision-Making","issuer":"Treasury Board of Canada Secretariat","region":"north-america","url":"https://www.tbs-sct.canada.ca/pol/doc-eng.aspx?id=32592","note":"A model for impact assessment, notice and human intervention in automated administrative decisions."}],"controls":["Transparency record for each tool, stating it does not decide discretionary permits","Data protection impact assessment covering applicant documents sent to models","Audit trail retained for the appeal period","Monthly accuracy review by application type","Notice to applicants when a decision is automated, with a route to human review"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the officer side is an **agentic workflow** that reads the uploaded application\ndocuments (PDF, Office files, images), checks them against the checklist for that application\ntype, retrieves relevant policies and history from a **knowledge base** with **hybrid retrieval**\nand from **SQL knowledge bases** for structured records, and drafts findings. **Human in the\nloop approval** means nothing is written back to the permitting system through **custom\nfunctions** until the officer approves, and every run has a full **audit trail**.\n\nThe applicant side is an **AI agent** on the council website or portal that explains\nrequirements from official guidance and answers in the applicant's language. **PII masking**\nredacts personal data before text reaches a model, the platform is **model agnostic**, and\n**test suites** replay past applications with known validation outcomes on every change."},"faq":[{"question":"Does AI decide planning or permit applications?","answer":"Not in the public examples for discretionary permits. Leeds' Xylo Core gives no approve or refuse recommendation and officers accept or reject every suggestion; the US Fish and Wildlife Service states no permit decision rests on AI alone. Narrow rule based permissions can be partly automated: West Berkshire rejects larger bin applications that miss its policy criteria automatically, with a route to appeal, and sends the rest to staff."},{"question":"How accurate is AI validation?","answer":"Early figures only. Leeds reports initial accuracy of 85% or more across its validation and policy prompts, with a target of 99% or more, starting with householder applications. Measure acceptance of suggestions on your own cases."},{"question":"Where should an authority start?","answer":"With validation of the simplest, highest volume application type, and with checks applicants can run before submitting. The UK IPO says its trade mark pre check supports about 20% of filings."}],"related":["citizen-information-assistant","inspection-prioritization","intelligent-document-processing","benefits-eligibility-and-application-assistant","immigration-and-visa-application-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Unpublished by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from UK transparency records, the US federal AI inventory and city case studies, with every quote verified against the source."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected West Berkshire (only rejections are automated), the Leeds accuracy target and pilot tense, the EU AI Act basis (Annex III points 5(a) and 7, Article 50), the IPO pre check wording and the AI Playbook note; added seoTitle and metaDescription."},{"date":"2026-09-26","note":"Fact checked again against all sources (GOV.UK transparency records and dates, federal inventory entries DOI-0011, DOI-0012 and USDA-175, guidance pages): all claims supported; renamed hybrid search to hybrid retrieval to match the platform feature inventory."}],"slug":"permit-and-licence-application-processing","url":"https://www.blits.ai/ai-use-cases/permit-and-licence-application-processing","benchmarks":[{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":85,"min":85,"max":85,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"leeds-city-council-xylo-planning-validation","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":42,"min":42,"max":42,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"west-berkshire-council-larger-bin-application-assessment","pooled":true}]}],"indicativeValueResult":{"low":100000,"high":450000},"evidence":["intellectual-property-office-trade-mark-pre-application-check","leeds-city-council-xylo-planning-validation","us-fish-and-wildlife-service-epermits-assistant","usda-environmental-permitting-review","west-berkshire-council-larger-bin-application-assessment"]},{"title":"AI for perpetual KYC and event driven customer due diligence","shortTitle":"Perpetual KYC","seoTitle":"AI for perpetual KYC and event driven reviews","metaDescription":"AI keeps KYC files current by reviewing on trigger events, not the calendar. JPMorgan Chase reports a 40% lower KYC unit cost since 2022 from AI and technology.","definition":"AI that keeps each customer's due diligence file current by replacing calendar driven KYC reviews with continuous, event driven refreshes: it watches for trigger events such as a change of ownership, address, behaviour or a new adverse finding, refreshes the file automatically where it can, and involves an analyst only when something material has changed. The risk rating itself and the first file for a new business client are separate use cases.","aliases":["pKYC","continuous KYC","event driven KYC review","KYC refresh automation","periodic review automation"],"industries":["banking","payments","wealth-and-asset-management"],"functions":["onboarding-and-kyc","financial-crime-compliance"],"patterns":["agentic-workflow","document-processing","rag-knowledge-assistant","summarization"],"channels":["internal-tools","email","mobile-app"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"middle-office","problem":"Banks review each customer's due diligence file on a fixed cycle; Fenergo describes periodic KYC\nas checks at set intervals, typically every year or every two years. Each review means re\ncollecting documents, checking registries and ownership structures, rescreening and writing a\nconclusion, and for corporate clients it can take weeks of back and forth.\n\nA review where nothing material has changed costs much the same effort as one that finds\nsomething, and it annoys customers with repeated requests for information the bank already has.\nAt the same time, a real change, such as a new beneficial owner or a sudden shift in activity,\ncan go unnoticed until the next scheduled review. When reviews fall behind schedule, the backlog\nof overdue files becomes a compliance risk in its own right.","problemStats":[{"statement":"A Fenergo study found that more than half of financial institutions spend between 61 and 150 days on client KYC reviews, at an average cost of USD 2,200 per review.","sourceTitle":"Ongoing Customer Due Diligence with Perpetual KYC","sourceUrl":"https://resources.fenergo.com/blogs/perpetual-kyc-pkyc","year":2026}],"howItWorks":"1. **Watch for triggers.** The system monitors company registries, ownership changes, screening\n   results, adverse media, transaction behaviour, contact detail changes and document expiry\n   for every customer.\n2. **Assess materiality.** Each event is classified against the bank's trigger policy: ignore,\n   refresh automatically, or open a review.\n3. **Refresh straight through.** Low risk changes (a new registry filing that confirms existing\n   data, a renewed identity document) update the file automatically, with the source recorded.\n4. **Prepare the review.** Material events open a review with the file already assembled: what\n   changed, current documents, registry extracts, screening results and a drafted summary with\n   sources.\n5. **Reach out only when needed.** When information is missing, the customer gets one targeted\n   request through the channel they use, and the analyst concludes and signs off the review.","valueDrivers":["compliance","cost-to-serve","customer-experience","risk-reduction"],"kpis":["automation-rate","processing-time-reduction","cost-reduction","productivity-gain","cycle-time-days"],"indicativeValue":{"referenceOrg":"A bank with 20,000 corporate and business clients under periodic review","inputs":[{"key":"reviews","label":"Periodic KYC reviews per year","low":6000,"high":8000,"unit":"reviews per year","note":"Editorial assumption for a mix of one, three and five year review cycles on 20,000 clients."},{"key":"costPerReview","label":"Cost of a periodic review","low":800,"high":2200,"unit":"USD per review","note":"The high value is the average cost per client KYC review reported in a Fenergo study cited on this page; the low value is an editorial assumption for simpler files."},{"key":"avoidedShare","label":"Share of review effort avoided by straight through refresh and prepared files","low":0.2,"high":0.4,"unit":"fraction of review cost","note":"Conservative against the benchmarks on this page (JPMorgan Chase reports a 40% reduction in KYC unit cost since 2022 from AI and technology). Replace with results from your own pilot."}],"formula":"reviews * costPerReview * avoidedShare","currency":"USD","period":"per year","resultLabel":"KYC review cost avoided","caveat":"Leaves out the value of detecting material changes earlier, the effect on customer experience and attrition, backlog reduction, data costs and the cost of building the trigger monitoring."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Needs reliable external data feeds, a trigger policy agreed with compliance and sometimes the regulator, and a customer lifecycle system that can take automated updates with an audit trail. AML rules such as MAS Notice 626 still expect regular account reviews, so the design needs a lighter backstop cycle alongside the triggers.","dataPrerequisites":["Structured KYC files with data lineage per attribute","Company registry, ownership and document expiry data feeds","Screening, adverse media and transaction monitoring signals per customer","A written trigger policy that defines material events"],"integrations":["Client lifecycle management or KYC platform","Company registries and data providers","Screening and transaction monitoring systems","Customer channels for information requests (portal, app, email)","Document management"]},"implementation":{"steps":[{"title":"Write the trigger policy","detail":"Agree with compliance which events matter, for which customer types, and what each should cause. Check whether your regulator still expects fixed review cycles and design around it."},{"title":"Fix the file first","detail":"Structure KYC data by attribute with its source and date. Event driven review is impossible when the file is a folder of PDFs."},{"title":"Automate the assembly of reviews","detail":"Start by preparing scheduled reviews automatically (registry extracts, screening, draft summary). This saves time before any policy change."},{"title":"Switch on triggers for one segment","detail":"Enable event driven reviews for one segment, run them in parallel with the calendar, and compare what each approach finds."},{"title":"Extend straight through refresh","detail":"Allow automatic updates for low risk, well sourced changes and measure the error rate on a sample before widening."}],"guardrails":["Material changes and every risk rating change need analyst sign off","Automatic updates only from approved sources, with source and date recorded per attribute","Logged reason why each trigger did or did not open a review","Customer outreach limited to information the bank does not already hold","A fallback to scheduled review for customers whose data feeds are incomplete"],"humanInTheLoop":"Analysts sign off every review opened by a material trigger and every change in risk rating. Compliance owns the trigger policy and reviews samples of events that were ignored or refreshed automatically.","kpisToInstrument":["Share of trigger events refreshed straight through, and the error rate in sampling","Time from trigger to completed review","Overdue review backlog","Customer outreach requests per review","Material findings per review, compared with calendar reviews"],"failureModes":[{"title":"Trigger storms","detail":"Noisy feeds open thousands of trivial reviews. Tune materiality and measure the share of triggered reviews with a finding."},{"title":"Silent gaps","detail":"A customer with no data feed never triggers and never gets reviewed. Keep a backstop review cycle."},{"title":"Regulatory mismatch","detail":"The regulator still expects fixed cycles. Agree the approach and document it before switching off calendar reviews."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Keeping customer due diligence files current is not listed in Annex III, so a back office system that assembles reviews for an analyst to decide is usually minimal risk. The design decides the rest: a conversational agent that asks customers for missing information must tell them they are interacting with an AI system (Article 50(1)); biometric verification that only confirms a person is who they claim to be is excluded from Annex III point 1(a), while remote biometric identification is high risk; and Article 5(1)(d) prohibits assessing the risk that a person will commit a criminal offence based solely on profiling, so behavioural triggers should open a review for a human rather than score the customer. GDPR applies to the collection and retention of KYC data, including Article 22 if an automated refresh leads to a decision with legal or similarly significant effect, such as closing an account."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","mas-ai-risk-management","cbuae-ai-guidance","nist-ai-rmf","eu-amlr","us-bsa","mas-notice-626"],"guidance":[{"title":"Guidelines on the use of remote customer onboarding solutions","issuer":"European Banking Authority","region":"europe","url":"https://www.eba.europa.eu/regulation-and-policy/anti-money-laundering-and-countering-financing-terrorism/guidelines-use-remote-customer-onboarding-solutions","note":"Sets expectations for remote identity verification and data collection when customers are onboarded remotely; the guidelines cover onboarding, but they are a useful reference when KYC data is refreshed through remote channels."},{"title":"Notice 626 Prevention of Money Laundering and Countering the Financing of Terrorism, Banks","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/notices/notice-626","note":"Singapore's AML and CFT rules for banks, covering customer due diligence, regular account reviews and the monitoring and reporting of suspicious transactions."},{"title":"Principles for Using Artificial Intelligence and Machine Learning in Financial Crime Compliance","issuer":"Wolfsberg Group","region":"global","url":"https://wolfsberg-group.org/resources/202/93","note":"Industry principles from 2022 for the accountable use of AI and machine learning in financial crime compliance programmes, covering legitimate purpose, proportionate use, design and technical expertise, accountability and oversight, and openness and transparency."}],"controls":["Approved trigger policy under change control","Attribute level source and date for every automated update","Sampling of automatically refreshed files and ignored triggers","Backstop review cycle for customers without reliable data feeds","Audit trail of every review conclusion and sign off"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai trigger handling runs as **agentic tasks** (\"when this condition is met, do this\")\nand **agentic workflows** triggered on a schedule or through the API. The agent calls **custom\nfunctions** for registries, screening and the KYC platform, reads documents from the **knowledge\nbase** (PDF, Office files and Outlook email files), compares them with the file and drafts a\nreview summary with sources as **structured output**. Updates to the file and risk rating\nchanges go through **human in the loop approval**, and every run keeps a full audit trail.\n\nWhen the bank needs something from the client, a **conversational agent** on **email or\nWhatsApp**, or in the bank's own app through the **REST or WebSocket API channel**, asks for exactly the missing document or confirmation and accepts\n**attachments**, with **PII masking** and a **human handover** to the relationship team.\n**Test suites** and **monitors** keep the workflow honest, and the platform is model agnostic\nwith EU and UAE data residency."},"faq":[{"question":"Does perpetual KYC replace periodic reviews?","answer":"Partly. Event driven reviews catch change sooner and let the bank skip work where nothing changed, but AML rules such as MAS Notice 626 still expect regular account reviews. A sound design keeps a lighter backstop cycle alongside the triggers, agreed with the regulator."},{"question":"What triggers a review in perpetual KYC?","answer":"Typical triggers are changes of ownership or directors, new adverse media or sanctions results, unusual transaction behaviour, changes of address or country, and expiring documents. The bank's written trigger policy defines which ones matter for which customer types."},{"question":"Where does AI help most?","answer":"In assembling the file and judging materiality: reading registry filings and documents, comparing them with what the bank holds, and drafting a summary so that the analyst decides rather than collects. JPMorgan Chase reports a 40% reduction in KYC unit cost since 2022 from AI and technology, and Nasdaq Verafin describes an agent that automates a bank's periodic enhanced due diligence review process, closing low risk cases itself and escalating the rest."}],"related":["dynamic-customer-risk-rating","pep-and-adverse-media-screening","business-onboarding-and-ubo-discovery","sanctions-screening-adjudication","digital-onboarding-assistant","source-of-wealth-diligence"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsourced review cycle, overdue backlog and industry practice claims from the problem, feasibility note and FAQ; corrected the Fenergo statistic year to 2026 (page date); changed the EU AI Act tier to context dependent (Article 50, Annex III point 1(a), Article 5(1)(d)); corrected the EBA, MAS and Wolfsberg guidance notes; aligned the Blits.ai channels with the feature inventory; added seoTitle and metaDescription."}],"slug":"perpetual-kyc","url":"https://www.blits.ai/ai-use-cases/perpetual-kyc","benchmarks":[{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40,"min":40,"max":40,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"jpmorgan-chase-kyc-unit-cost","pooled":true}]}],"indicativeValueResult":{"low":960000,"high":7040000},"evidence":["deutsche-bank-workfusion-screening-automation","fnbo-verafin-agentic-edd-and-sanctions","jpmorgan-chase-kyc-unit-cost","ocbc-helios-agentic-due-diligence","origin-bank-verafin-agentic-edd-reviews"]},{"title":"AI for pharmacovigilance adverse event case intake","shortTitle":"Adverse event case intake","seoTitle":"AI for adverse event case intake in drug safety","metaDescription":"AI reads adverse event reports, checks validity and seriousness and extracts case data for safety staff. Pfizer and the FDA have deployed it for case intake.","definition":"AI that takes in adverse event reports about medicines, vaccines and devices from calls, emails, forms, literature and partner files, decides whether each is a valid case, flags seriousness, extracts and codes the case data into the safety database format, and routes it to drug safety professionals, who review medical content and regulatory reporting.","aliases":["pharmacovigilance automation","ICSR intake automation","AI case processing for drug safety","adverse event report extraction","adverse event reporting chatbot"],"industries":["pharma-and-life-sciences","government"],"functions":["regulatory-compliance","case-management","operations"],"patterns":["document-processing","classification-and-routing","conversational-agent"],"channels":["internal-tools","email","web-chat","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Marketing authorisation holders must collect, assess and report adverse events, and the volumes\nkeep rising: reports arrive from patients, doctors, clinical trial sites, call centres, social media,\nliterature and business partners, in every format and language. Each report has to be checked for\nthe minimum criteria of a valid case, triaged for seriousness because serious cases have short legal\nreporting deadlines, entered into the safety database and coded before medical review.\n\nMost of that intake work is repetitive data entry done by trained staff, while the scarce skill in\npharmacovigilance is medical judgment: causality, signal detection and benefit risk. Regulators face\nthe same flood from the other side, receiving individual case safety reports and sponsor safety\nreports that must be extracted and loaded before anyone can analyse them.","problemStats":[{"statement":"Pfizer reports that its Worldwide Safety organization processed approximately 1.4 million adverse events globally in 2019.","sourceTitle":"AI in Drug Safety: Building the Elusive 'Loch Ness Monster' of Reporting Tools","sourceUrl":"https://www.pfizer.com/news/articles/ai-drug-safety-building-elusive-%E2%80%98loch-ness-monster%E2%80%99-reporting-tools","year":2020}],"howItWorks":"1. **Collect reports from every channel.** Emails, scanned forms, call centre notes, partner files,\n   literature hits and web form or chatbot submissions arrive in one intake queue.\n2. **Check validity and triage.** The AI checks the four minimum criteria (an identifiable patient\n   and reporter, a suspect product and an adverse event), detects duplicates and flags seriousness,\n   such as fatal or life threatening outcomes, so the reporting clock is visible from day zero.\n3. **Extract and code.** It extracts patient, product, event, dates and narrative into the safety\n   database structure (E2B), suggests MedDRA terms and product dictionary matches, and shows a\n   confidence level per field.\n4. **Route for review.** Safety professionals check the extraction, especially low confidence fields\n   and serious cases, and complete medical assessment, follow up and regulatory reporting.\n5. **Help reporters report.** On the public side, a conversational assistant can guide patients and\n   professionals to the right form, ask for missing information and submit the report.","valueDrivers":["compliance","speed","employee-productivity","cost-to-serve","risk-reduction"],"kpis":["automation-rate","handling-time-reduction","processing-time-reduction","accuracy","interactions-handled"],"indicativeValue":{"referenceOrg":"A drug company that takes in 100,000 adverse event reports a year","inputs":[{"key":"cases","label":"Adverse event reports taken in per year","low":100000,"high":100000,"unit":"reports per year","note":"The reference company."},{"key":"intakeMinutes","label":"Intake minutes per report today (triage, data entry and coding)","low":20,"high":40,"unit":"minutes per report","note":"Editorial assumption. Replace with a time study of your own intake step."},{"key":"effortSaved","label":"Share of intake effort the AI removes","low":0.2,"high":0.5,"unit":"fraction of intake minutes","note":"Editorial assumption; the evidence on this page publishes no measured intake savings. Validate in a pilot before relying on it."},{"key":"costPerHour","label":"Fully loaded cost per case processing hour","low":40,"high":80,"unit":"USD per hour","note":"Editorial assumption covering internal staff and outsourced case processing. Replace with your own."}],"formula":"cases * intakeMinutes / 60 * effortSaved * costPerHour","currency":"USD","period":"per year","resultLabel":"Case intake effort released","caveat":"Intake effort only. It leaves out medical review, follow up and submission work, the value of fewer late reports, validation and inspection readiness costs, and the platform cost. No organization on this page has published measured savings, so treat the range as a hypothesis."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Extraction from mixed documents is proven, but pharmacovigilance is a regulated, inspected process with legal reporting deadlines. The system must be validated, integrated with the safety database, and designed so that no serious case is missed or delayed.","dataPrerequisites":["Historical cases with source documents and the final database entries for testing","MedDRA and product dictionaries under licence and version control","Written case processing conventions and validity rules","Intake channel inventory with volumes and formats"],"integrations":["Safety database that accepts E2B formatted cases","Email, call centre, web form and partner exchange intake channels","Literature monitoring sources","Quality management system for deviations and corrective actions"]},"implementation":{"steps":[{"title":"Map intake by channel and volume","detail":"List every source of reports, its format and volume, and start with high volume, well structured sources such as email forms and partner files."},{"title":"Automate validity and seriousness triage first","detail":"Deciding whether a report is a valid case and whether it is fatal or life threatening is the first phase of Pfizer's tool; it protects deadlines and gives reviewers a prioritised queue."},{"title":"Extract with confidence scores","detail":"Extract fields into the E2B structure with a confidence per field, and send low confidence fields and all serious cases to a person for verification."},{"title":"Validate as a computerized system","detail":"Define the intended use, test against historical cases, document performance per field and case type, and put models, prompts and dictionaries under change control."},{"title":"Monitor and extend","detail":"Track missed cases, extraction accuracy and timeliness weekly, and add channels and languages one at a time."}],"guardrails":["Every case the AI marks invalid or non serious is sampled by a person; no report is discarded unreviewed","Serious and fatal cases always routed to a safety professional with the reporting deadline shown","Medical assessment, causality and regulatory submission decisions stay with qualified staff","Personal and health data of patients and reporters masked in logs and prompts","Models, prompts and coding dictionaries under validation and change control"],"humanInTheLoop":"Drug safety professionals verify low confidence extractions and every serious case, perform medical review and causality assessment, and decide what is reported to regulators. The qualified person for pharmacovigilance owns the process, and quality staff sample cases the AI screened out.","kpisToInstrument":["Share of reports processed without manual data entry","Field level extraction accuracy on a weekly sample","Time from receipt to case creation, and share of expedited reports submitted on time","Missed or wrongly invalidated cases found in quality samples","Duplicate cases detected"],"failureModes":[{"title":"A serious case screened out","detail":"The AI marks a valid serious case as invalid or non serious and the deadline is missed. Sample every negative decision and track misses as deviations."},{"title":"Silent extraction errors","detail":"A wrong dose, date or product enters the database and distorts signal detection. Show confidence per field and verify key fields on every case."},{"title":"Validation that ends at go live","detail":"Performance drifts as new products, languages and sources arrive. Monitor accuracy continuously and revalidate after changes."},{"title":"A chatbot that discourages reporting","detail":"A public reporting assistant that is hard to use lowers reporting. Keep a direct form and a human route and test with real reporters."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Internal intake, extraction and coding for review by safety staff is not listed in Annex III and is usually minimal risk. A public facing reporting assistant must tell people they are talking to an AI under Article 50. The main obligations come from pharmacovigilance law and good pharmacovigilance practices, which require validated, inspectable processes, and from GDPR rules on health data."},"regulations":["eu-ai-act","gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle","issuer":"European Medicines Agency","region":"europe","url":"https://www.ema.europa.eu/en/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle_en.pdf","note":"Covers AI in pharmacovigilance, including adverse event report management and signal detection, in line with good pharmacovigilance practices, and expects marketing authorisation holders to validate, monitor and document these tools."},{"title":"Good pharmacovigilance practices (GVP)","issuer":"European Medicines Agency","region":"europe","url":"https://www.ema.europa.eu/en/human-regulatory-overview/post-authorisation/pharmacovigilance-post-authorisation/good-pharmacovigilance-practices-gvp","note":"The EU rules for collecting, managing and submitting reports of suspected adverse reactions, which an automated intake process must still meet."},{"title":"Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (draft guidance)","issuer":"US Food and Drug Administration","region":"north-america","url":"https://www.fda.gov/media/184830/download","note":"Proposes a risk based credibility assessment for AI models that produce information or data used to support regulatory decisions about safety, effectiveness or quality."}],"controls":["Validation package with intended use, test results per case type and field, and acceptance criteria","Audit trail from source document to database entry, including AI suggestions and human changes","Deviation and corrective action process for missed or late cases","Periodic sampling of AI negative decisions (invalid, non serious, duplicate)","Data protection controls for patient and reporter data, including transfers to vendors"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai intake runs as an **agentic workflow** per report. Reports arrive through the **email\nchannel**, partner systems trigger the workflow via **API tokens**, and a **flow** with a **receive\nattachment** block collects files that reporters upload. An **AI agent** with **structured output**\nextracts the case fields from the report text, checks the validity criteria and flags seriousness. **Custom functions** look up products and write the case to the safety\ndatabase, and a **knowledge base** holds the case processing conventions the agent follows.\n\nFor reporters, a **web chat** or **WhatsApp** agent with **flows** asks for the minimum information,\nin several languages, and hands over to a person when needed. **Human in the loop approval** keeps\nserious cases and low confidence extractions with safety staff, **PII masking** protects patient\ndata, the **audit trail** records every run, and **test suites** grade extraction against\nhistorical cases before each change."},"faq":[{"question":"Which parts of pharmacovigilance can AI automate safely?","answer":"Intake steps with clear rules: checking whether a report is a valid case, flagging seriousness, detecting duplicates, extracting data and suggesting codes. Pfizer's first phase covers basic intake decisions such as validity and whether a case is fatal or life threatening. Medical assessment, causality and reporting decisions stay with qualified people."},{"question":"Do regulators use AI on adverse event reports too?","answer":"Yes. The US FDA lists a deployed tool that extracts data from sponsors' IND safety reports and loads it into its adverse event database, and a chatbot that helps people submit adverse event and product problem reports through its Safety Reporting Portal."},{"question":"Why are there so few published results?","answer":"Most deployments are described by companies and vendors without measured before and after data, and for older programmes, such as Bayer's 2018 agreement with Genpact, we found no published results. Validate on your own historical cases and publish internally what the system misses, not only what it saves."}],"related":["clinical-and-regulatory-document-drafting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with evidence from Pfizer, the US FDA (two federal inventory entries) and Bayer. Editor pass rewrote the Bayer record from the literal Genpact release text and limited claims to what the sources state."}],"slug":"adverse-event-case-intake","url":"https://www.blits.ai/ai-use-cases/adverse-event-case-intake","benchmarks":[],"indicativeValueResult":{"low":266666.66666666674,"high":2666666.666666667},"evidence":["bayer-genpact-pharmacovigilance-ai","fda-ind-safety-report-extraction","fda-safety-reporting-portal-chatbot","pfizer-adverse-event-case-intake"]},{"title":"AI for photo based damage assessment in insurance claims","shortTitle":"Photo damage assessment","seoTitle":"AI photo damage assessment for insurance claims","metaDescription":"AI assesses damage from claim photos and drafts or checks the repair estimate. Tractable reports 90% of Admiral Seguros estimates in 2021 needed no human appraiser.","definition":"Computer vision that assesses damage from photos or video of a vehicle or property taken by the policyholder, a repairer or an adjuster, identifies the damaged parts and the repair or replace decision, produces or checks the repair estimate, and flags total losses and inconsistencies for a person to review.","aliases":["AI damage estimation","computer vision claims assessment","photo estimating","virtual vehicle inspection","AI estimate review"],"industries":["insurance"],"functions":["claims"],"patterns":["computer-vision","prediction-and-scoring","agentic-workflow"],"channels":["mobile-app","web-chat","api"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"claims","problem":"For motor and many property claims, the size of the loss is decided by an expert looking at the\ndamage: a field appraiser drives to the car or house, or a desk assessor reviews a repairer's\nestimate and photos. Appraiser capacity is limited, the visit can add days to the claim, and after\na hail storm, flood or other large event the number of claims surges and the backlog grows.\nCustomers wait without a car or with a damaged home, and repairers wait for approval before they\ncan start.\n\nDesk review has its own problem: assessors check large numbers of estimates line by line, and the\nconsistency of the decision depends on who looks at it. Inflated or duplicated items can slip\nthrough, while honest estimates wait in the same queue.","problemStats":[],"howItWorks":"1. **Capture the images.** The policyholder receives a link at first notice of loss and is guided\n   to take the right photos, or the repairer uploads them with the estimate.\n2. **Check the images.** The model confirms the images show the insured vehicle or property, are\n   usable and have not been reused from another claim.\n3. **Assess the damage.** Computer vision identifies the damaged parts, the severity and the repair\n   or replace decision per part, and estimates labour and parts cost from repair data.\n4. **Decide the path.** A clear, low value case gets an estimate or cash settlement offer within\n   minutes; a likely total loss goes to the total loss process; everything else goes to an\n   assessor with the AI findings.\n5. **Review estimates from repairers.** For repairer estimates, the AI compares each line with the\n   photos and flags items that are not supported, so assessors review exceptions only.","valueDrivers":["speed","cost-to-serve","customer-experience","risk-reduction"],"kpis":["automation-rate","interactions-handled","processing-time-reduction","cycle-time-days","accuracy","cost-reduction"],"indicativeValue":{"referenceOrg":"A motor insurer handling 100,000 repairable vehicle damage claims a year","inputs":[{"key":"claims","label":"Repairable vehicle damage claims per year","low":100000,"high":100000,"unit":"claims per year","note":"The reference insurer."},{"key":"eligibleShare","label":"Share of claims assessed from photos instead of a physical inspection","low":0.3,"high":0.6,"unit":"fraction of claims","note":"Conservative against the evidence on this page (Tractable reports that 90% of Admiral Seguros claim estimates were processed without human appraisers in 2021), because eligibility depends on the claim mix and customer uptake."},{"key":"inspectionCost","label":"Cost of a physical or detailed desk assessment avoided","low":60,"high":150,"unit":"USD per claim","note":"Editorial assumption covering appraiser time and travel; replace with your own appraisal costs."}],"formula":"claims * eligibleShare * inspectionCost","currency":"USD","period":"per year","resultLabel":"Appraisal cost avoided","caveat":"Appraisal cost only. It leaves out shorter rental car periods, better estimate accuracy and leakage control, customer satisfaction, catastrophe surge capacity and the cost of the service and the integration."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Specialised vendor models exist for vehicle damage, and the motor deployments on this page use one (Tractable) rather than a model built in house; Tokio Marine & Nichido Fire uses Shift Technology's claims platform to review damage photos and estimates. The work is in the customer photo journey, integration with estimating and claims systems, repair cost data for the local market, and the rules that decide when a claim may settle on the AI estimate.","dataPrerequisites":["Local repair cost, labour rate and parts data, or a vendor that holds it","Historical claims with photos and final repair costs to calibrate and test","Rules for photo settlement per claim type, value and customer situation"],"integrations":["Claims management system for the claim, reserve and payment","Customer photo capture link or app, sent at first notice of loss","Estimating platform and repairer network systems","Total loss valuation process","Fraud detection for image reuse and manipulation checks"]},"implementation":{"steps":[{"title":"Choose the journey","detail":"Decide whether the AI starts with the policyholder's photos at first notice of loss, with repairer estimates, or both. Repairer estimate review can be the faster first step, because repairers already send photos with every estimate."},{"title":"Calibrate on your own claims","detail":"Run the model over a few thousand closed claims with known repair costs and compare estimates, repair or replace decisions and total loss calls with what happened."},{"title":"Design the photo capture for customers","detail":"Guided capture with examples and live checks for blur and angle decides how many claims can be assessed at all. Test it with real customers, not staff."},{"title":"Set settlement rules","detail":"Define the value limits, claim types and confidence thresholds under which an AI estimate may be offered or approved without an assessor."},{"title":"Close the loop with repair outcomes","detail":"Feed final invoices, supplements and reinspections back into monitoring so estimate accuracy is measured on real repairs."}],"guardrails":["Cash settlement offers only inside value limits and confidence thresholds per claim type","The customer can always ask for a human assessment","Image checks for reuse, manipulation and the wrong vehicle or property before any offer","Supplements and reinspection results monitored against AI estimates per repairer","Total loss and injury indications always go to a person"],"humanInTheLoop":"Assessors review every claim outside the settlement rules and every flagged estimate line, and sample AI settled claims each week. Customers who disagree with an estimate get a human assessment.","kpisToInstrument":["Share of claims assessed from photos, per claim type","Days from first notice of loss to estimate and to repair authorisation","Difference between AI estimate and final repair cost, including supplements","Share of customers who complete the photo journey","Complaints and disputes about AI based estimates"],"failureModes":[{"title":"Customers do not finish the photo journey","detail":"Poor guidance leads to unusable photos and a fallback inspection, which is slower than before. Invest in guided capture and measure completion."},{"title":"Estimates that miss hidden damage","detail":"Photos show the outside, not the structural damage behind it. Allow supplements and track them against AI estimates."},{"title":"Low offers that damage trust","detail":"A fast but low settlement offer creates complaints and regulatory risk. Monitor disputes and offer a human assessment by default."},{"title":"Reused or manipulated images","detail":"Fraudsters submit old or edited photos. Check image metadata, similarity across claims and signs of manipulation before any payment."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Assessing damage to vehicles or property for property and casualty claims is not listed in Annex III, which covers insurance only for risk assessment and pricing of natural persons in life and health insurance (point 5(c)). Article 50(1) transparency duties apply when the customer interacts directly with the AI, for example a guided photo journey that returns an AI estimate or offer, or a chat agent. A purely internal repairer estimate review with no customer interaction is minimal. A settlement or refusal decided solely by automated processing can fall under GDPR Article 22."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","iso-42001"],"guidance":[{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Addressed to national supervisors (August 2025). Sets out how insurers should govern AI systems across the value chain, including claims, with measures proportionate to their risk and impact on customers, such as data governance, explainability and human oversight."}],"controls":["Documented settlement rules and thresholds under change control","Estimate accuracy monitored against final repair costs per model version","Customer disclosure that AI assesses the photos, with a right to a human assessment","Image integrity checks logged per claim"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai does not supply the damage recognition model itself; the insurer uses a specialised\ncomputer vision vendor or its own model and connects it through **custom functions** (REST\ncalls). Blits.ai carries the customer journey around it: an **AI agent** on **web chat, WhatsApp\nor the insurer's own app through the API channel** that guides the policyholder through the\nphotos with **receive attachment** blocks, explains the estimate and next steps, and hands over\nto an assessor.\n\n**Guardrails** keep the agent from promising amounts outside the settlement rules, **human in the\nloop approval** in an **agentic workflow** holds offers above a threshold for an assessor, and\n**human handover** passes the case with photos and findings. **Test suites** evaluate the\nconversation around the photo journey before each release, **monitors** run scheduled health\nchecks, and **analytics** with flow statistics show how many customers complete the photo\njourney."},"faq":[{"question":"How many claims can be assessed from photos without an appraiser?","answer":"It depends on the claim mix and on how many customers finish the photo journey. Tractable reports that 90% of Admiral Seguros claim estimates in 2021 were processed without human appraisers, with 98% of claims completed in less than 15 minutes, and that 70 to 75% of customers who receive the link complete the claim. Complex damage, injuries and likely total losses are best kept with a person."},{"question":"Is this only for customers' own photos?","answer":"No. Photo AI can also serve repairers and assessors who review estimates: Covéa's partner repairers receive instant assessments based on photos of the damage, and at Tokio Marine & Nichido Fire the AI highlights points to check for consistency across estimates, damage photos and claim statements."},{"question":"Does photo AI increase fraud risk?","answer":"It changes it. When the estimate rests on photos alone, reused, edited or AI generated images become a risk to manage, so image integrity checks and similarity search across claims should run before any payment."}],"related":["claims-first-notice-of-loss-agent","claims-triage-and-straight-through-processing","claims-fraud-detection","subrogation-opportunity-detection"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the insurance vertical with evidence from Admiral Seguros, Covéa, PZU, Foyer and Tokio Marine & Nichido Fire, verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsourced generalisations from the problem, feasibility and FAQ text, added the Admiral Seguros completion rate to the FAQ, made the EU AI Act basis and EIOPA note more precise, aligned the Blits.ai channels with the feature inventory, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: corrected which vendors the deployments use, limited the Blits.ai test suite claim to what the platform supports, clarified the repairer FAQ, and set the Covéa deployment to production because the source does not show it covers a large share of claims."},{"date":"2026-09-27","note":"Fact checked against sources again: all quotes, dates and the EIOPA note confirmed; removed an unsourced claim that image fraud is the main risk from the FAQ, and removed the API channel from the Covéa record because the source does not state it."}],"slug":"photo-based-damage-assessment","url":"https://www.blits.ai/ai-use-cases/photo-based-damage-assessment","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":86000,"min":12000,"max":160000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"covea-ai-photo-damage-assessment","pooled":true},{"id":"admiral-seguros-ai-vehicle-damage-estimates","pooled":true}]}],"indicativeValueResult":{"low":1800000,"high":9000000},"evidence":["admiral-seguros-ai-vehicle-damage-estimates","covea-ai-photo-damage-assessment","foyer-ai-motor-claims-photos","pzu-ai-car-damage-assessment","tokio-marine-nichido-shift-claims-review"]},{"title":"AI for policy drafting and policy gap analysis","shortTitle":"Policy drafting and gaps","seoTitle":"AI for policy drafting and gap analysis","metaDescription":"AI finds the policies a new rule touches, flags gaps and drafts redlines for owner approval. See how HHS ACF flags documents for review, and what it is worth.","definition":"An assistant that takes a new or changed obligation, finds every internal policy, standard and procedure it touches, flags clauses that now conflict or are silent, and drafts the updated wording in house style as a redline for the policy owner to approve.","aliases":["AI policy writing assistant","policy gap analysis AI","regulatory change policy update","policy redlining assistant"],"industries":["cross-industry","banking","insurance","capital-markets","government"],"functions":["regulatory-compliance","legal","knowledge-management"],"patterns":["rag-knowledge-assistant","content-generation","document-processing","summarization"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"second-line","problem":"A bank's policy estate is large and layered: group policies, standards, procedures and desk\ninstructions, written by different teams over many years. When a rule changes, someone has to\nfind every document it touches, work out which clauses now conflict or say nothing, and rewrite\nthem consistently. That work is mostly reading and cross referencing, done under deadline by\nspecialists whose time is better spent on judgement.\n\nThe result is predictable: documents that contradict each other, procedures that lag the policy\nthey implement, and wording that differs from team to team. When a supervisor asks how a policy\nimplements a rule, the trail from obligation to clause is often reconstructed after the fact.","problemStats":[],"howItWorks":"1. **Take the obligation.** Start from an obligation already mapped by horizon scanning or legal:\n   the new rule text, its effective date and the business it applies to.\n2. **Find what it touches.** Retrieval over the policy estate returns every policy, standard and\n   procedure that covers the topic, with the relevant clauses.\n3. **Flag gaps and conflicts.** The assistant compares each clause with the obligation and marks\n   it as compliant, conflicting, silent or unclear, quoting both texts.\n4. **Draft the change.** For each gap it drafts replacement or new wording using the approved\n   template and style guide, as a redline, never inventing requirements beyond the source.\n5. **Owner review.** The policy owner edits and approves; legal or compliance signs off where\n   required.\n6. **Keep the trail.** The link from obligation to clause, the drafts and the approvals are\n   stored with the version history.","valueDrivers":["compliance","employee-productivity","speed","risk-reduction"],"kpis":["processing-time-reduction","time-saved-per-task","productivity-gain","accuracy"],"indicativeValue":{"referenceOrg":"A bank that updates 300 policy and procedure documents a year after regulatory change","inputs":[{"key":"documents","label":"Policy and procedure documents updated per year","low":300,"high":300,"unit":"documents per year","note":"The reference bank. Replace with your own volume."},{"key":"hoursPerDocument","label":"Specialist hours per document update (analysis and drafting)","low":8,"high":16,"unit":"hours per document","note":"Editorial assumption, replace with your own time records."},{"key":"timeSaved","label":"Share of analysis and drafting time saved","low":0.2,"high":0.4,"unit":"fraction of time","note":"Editorial assumption. No public measured benchmark for policy drafting was found; keep this conservative."},{"key":"hourlyCost","label":"Fully loaded cost of a policy or compliance specialist","low":80,"high":150,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"documents * hoursPerDocument * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Specialist time released from policy updates","caveat":"Time only. It leaves out the value of fewer inconsistencies and findings, faster implementation of new rules, and the cost of building and maintaining the policy knowledge base."},"macroEstimates":[{"statement":"In the Bank of England and FCA 2024 survey of UK financial firms, an additional 32% of respondents expected to use AI for regulatory compliance and reporting over the next three years.","sourceTitle":"Artificial intelligence in UK financial services 2024","sourceUrl":"https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024","year":2024}],"feasibility":{"complexity":"medium","complexityNote":"Retrieval and drafting are mature. The effort is in a clean, versioned policy estate with owners, a mapped obligation library and a template the drafts must follow.","dataPrerequisites":["A current, versioned policy and procedure library with owners","An obligation library or the new rule texts with effective dates","Approved templates and a style guide","Past redlines and approvals to test the assistant against"],"integrations":["Policy management or document management system","Regulatory change or obligation management tool","Workflow for review and approval","Collaboration tools where owners edit drafts"]},"implementation":{"steps":[{"title":"Clean the estate","detail":"Retire duplicates, assign an owner and review date to every document and fix the version history. Retrieval over a messy estate finds the wrong clause confidently."},{"title":"Start with gap analysis, then drafting","detail":"First use the assistant to find affected clauses and classify gaps, and measure how many it misses against expert review. Only then let it draft wording."},{"title":"Constrain the drafting","detail":"Give the model the template, the style guide and the obligation text, and require every drafted sentence to cite the obligation it implements. Anything without a source is removed."},{"title":"Test on past changes","detail":"Replay previous regulatory changes where the final policy text is known and compare the assistant's gaps and drafts with what the experts did."},{"title":"Embed in the approval workflow","detail":"Deliver drafts as redlines into the existing review tool, record who approved what, and keep the obligation to clause link after publication."}],"guardrails":["Every drafted clause cites the obligation or source text it implements","Drafts follow the approved template; the assistant cannot publish or approve","The policy owner approves every change; legal or compliance signs off where required","Retrieval is limited to current, approved versions of documents","Version history and approval trail retained for supervisors and audit"],"humanInTheLoop":"The policy owner reviews and approves every change and owns the interpretation of the rule. Compliance or legal signs off material changes. The assistant drafts and flags; it never decides that a policy is compliant.","kpisToInstrument":["Time from a rule's publication to approved policy updates","Recall of affected clauses against expert review on sampled changes","Share of drafted text accepted without material edits","Inconsistencies found between policy and procedure in periodic reviews","Audit and supervisory findings on policy coverage"],"failureModes":[{"title":"Invented obligations","detail":"The model adds requirements that are not in the rule. Require a citation per sentence and strip anything unsupported."},{"title":"Missed documents","detail":"A procedure outside the indexed estate never shows up, so the gap persists. Measure recall and keep the estate complete."},{"title":"Style over substance","detail":"Polished drafts get approved without checking the interpretation. Owners must confirm the interpretation separately from the wording."},{"title":"Stale sources","detail":"Retrieval returns a superseded version. Index only current approved versions and show the version in every citation."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Drafting internal policy text for human approval is not an Annex III use and has no direct effect on individuals. The Article 4 AI literacy measures still apply to the staff who use it."},"regulations":["eu-ai-act","gdpr","iso-42001","nist-ai-rmf","mas-ai-risk-management"],"guidance":[{"title":"Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence","note":"A 2024 companion to the AI RMF that lists risks of generative AI, including confabulation, with suggested actions that apply directly to drafting assistants."},{"title":"Article 4, AI literacy","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/4/","note":"Providers and deployers must take measures to support the AI literacy of their staff and others who operate and use AI systems on their behalf."},{"title":"Artificial Intelligence (AI) Model Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Good practices for AI and generative AI model risk management observed in a 2024 thematic review of banks, covering governance, oversight, development and deployment."}],"controls":["Inventory entry for the assistant with an owner, scope and approved sources","Citation requirement and automated check for unsupported text","Approval workflow with recorded sign off per change","Periodic recall testing against expert gap analysis","Retention of the obligation to clause trail and version history"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** over a **knowledge base** that holds the policy estate and\nthe obligation texts, with document version control and hybrid retrieval (vector and BM25) so\nexact clause wording is found as well as related concepts. **Prompt versioning** holds the house\ntemplate and style rules, and **structured output** returns each affected clause with its gap\nclassification, the cited obligation and a drafted redline.\n\nFor larger changes an **agentic workflow** walks the whole estate for one obligation, and **human\nin the loop confirmation** holds each drafted change for approval or rejection. Output\n**guardrails** check drafts against an admin authored policy, such as rejecting text without a\ncited source, **test suites** replay past regulatory changes against expert outcomes before a\nprompt or model change goes live, and the workflow run history and audit trail keep the trail\nfrom obligation to draft. The platform is model agnostic,\nwith EU and UAE data residency."},"faq":[{"question":"Can AI write our policies?","answer":"It can find what a new rule touches and draft consistent wording quickly, but the policy owner must approve every change and own the interpretation. The safe design constrains drafting to approved templates and requires a citation for every clause."},{"question":"Where do public deployments stand?","answer":"Mostly early. The Administration for Children and Families, part of the US Department of Health and Human Services, reported using AI since March 2025 to flag grants and position descriptions that may need revision under new directives, with staff making the final assessments. Policy drafting assistants reported by HRSA (initiated in 2024) and the FDIC (retired by 2025) never reported reaching production."},{"question":"How is this different from regulatory horizon scanning?","answer":"Horizon scanning finds new rules and maps them to obligations. Policy drafting and gap analysis starts from a mapped obligation and changes the bank's own documents to implement it."}],"related":["regulatory-horizon-scanning","continuous-controls-testing","supervisory-exam-response-assembly","hr-and-policy-assistant","internal-audit-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against public sources. The catalog's Thomson Reuters source returned 404 and is not used."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: made the Bank of England and FCA survey statement precise, corrected the MAS paper title and note, reworded the FAQ on deployments and the Blits.ai build description to match sources, and added an SEO title and meta description."},{"date":"2026-09-26","note":"Second fact check against sources: all evidence, the survey statement and guidance links confirmed; set the adoption stage to emerging (one public production deployment), aligned the Article 4 note with the current text on the cited page and made the meta description name only the deployed example."},{"date":"2026-09-26","note":"Third fact check against sources: all three inventory records, the Bank of England and FCA survey figure and the guidance links confirmed; gave the NIST profile its exact title and described it more precisely, and aligned the Article 4 note with the current text on the cited page."}],"slug":"policy-drafting-and-gap-analysis","url":"https://www.blits.ai/ai-use-cases/policy-drafting-and-gap-analysis","benchmarks":[],"indicativeValueResult":{"low":38400,"high":288000},"evidence":["fdic-plain-language-policy-assistant","hhs-acf-directive-alignment-document-review","hrsa-policy-document-drafting-assistant"]},{"title":"AI for predictive network maintenance in telecom","shortTitle":"Predictive network maintenance","seoTitle":"AI predictive maintenance for telecom networks","metaDescription":"AI spots early signs of network failure so operators fix faults before customers notice. Telstra's SmartFix performed 2.5 million proactive actions in FY25.","definition":"Machine learning that spots the early signs of network failure, such as degrading cells, faulty customer equipment, ageing hardware or planned digging near fibre, and triggers a preventive fix, a remote reset or a targeted intervention before customers lose service.","aliases":["predictive maintenance for telecom networks","proactive network assurance","network failure prediction","silent cell detection"],"industries":["telecommunications"],"functions":["network-operations","field-service","operations"],"patterns":["anomaly-detection","prediction-and-scoring","agentic-workflow"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"network","problem":"Much network maintenance is reactive or calendar based. Faults are found when an alarm fires\nor when customers call, and field teams replace equipment on a schedule whether it needs it or\nnot. Many failures give warning signs first: a cell whose throughput slowly degrades, a modem\nthat keeps dropping its connection, a router with rising error counts, a battery that no longer\nholds its charge. Those signals sit in performance data that nobody has time to watch.\n\nSome outages have nothing to do with the equipment itself. Verizon notes that every year thousands\nof fiber lines are damaged by accidental cuts during construction and excavation, which can affect\ncustomers' connectivity for anything from a few hours to several days. A fault that reaches the\ncustomer can cost a support call, a technician visit and some goodwill. The opportunity is to act\non the warning signs early enough to fix the problem remotely, during a planned window, or before\nthe digger arrives.","problemStats":[],"howItWorks":"1. **Gather the signals.** Performance counters, alarms, device telemetry from customer equipment,\n   environmental and power data from sites, and external data such as dig requests or weather.\n2. **Score the risk.** Models learn the patterns that preceded past failures and score each cell,\n   line, device or site for the probability of failure or degradation in the coming days.\n3. **Classify the likely cause.** For each at risk element the system proposes the probable root\n   cause (hardware, configuration, interference, power, external damage) so the right fix is chosen.\n4. **Act at the right level.** Low risk fixes such as a remote reset, a configuration rollback or\n   a customer equipment reboot run automatically within limits; hardware swaps and site visits are\n   scheduled as planned work; external risks trigger outreach, such as contacting an excavator.\n5. **Learn from the outcome.** Every prevented and every missed failure is fed back to retrain the\n   models and tune the thresholds.","valueDrivers":["customer-experience","cost-to-serve","risk-reduction","speed"],"kpis":["interactions-handled","detection-rate-improvement","cost-savings","processing-time-reduction"],"indicativeValue":{"referenceOrg":"A national fixed and mobile operator with about 40,000 customer affecting network faults a year","inputs":[{"key":"faults","label":"Customer affecting network faults per year","low":20000,"high":60000,"unit":"faults per year","note":"Editorial assumption for a national operator. Replace with your own fault volume."},{"key":"preventableShare","label":"Share of faults prevented or fixed before customers are affected","low":0.1,"high":0.25,"unit":"fraction of faults","note":"Editorial assumption, not calibrated by any source on this page. Telstra's 2.5 million SmartFix proactive actions in FY25 show the scale of such programs but say nothing about the share of faults that can be prevented; the share of your own faults that show warning signs is the number to measure first."},{"key":"costPerFault","label":"Cost of a customer affecting fault","low":300,"high":800,"unit":"USD per fault","note":"Editorial assumption covering repair, technician visits and customer contacts. Replace with your own fully loaded cost."}],"formula":"faults * preventableShare * costPerFault","currency":"USD","period":"per year","resultLabel":"Fault handling cost avoided","caveat":"Direct fault cost only. It leaves out avoided service level penalties, churn and complaint handling, the extra cost of preventive work on false alarms, and the cost of the data platform."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Predicting failures needs years of clean history that links alarms and performance data to the faults that followed. Many operators have the data but not the labels, and acting on predictions means changing how field and NOC work is planned.","dataPrerequisites":["Historical performance and alarm data per network element, kept for at least a year","Fault and repair records with root cause codes that can be joined to that data","Telemetry from customer premises equipment where fixed access is in scope","Site power, battery and environmental data","External data sources where relevant, such as dig request notifications"],"integrations":["Performance and fault management systems","Network inventory and topology","Workforce management and field scheduling","Device management platforms for customer equipment","Trouble ticketing and customer notification systems"]},"implementation":{"steps":[{"title":"Pick a failure mode with a clear payoff","detail":"Start with one failure that is frequent, costly and preceded by measurable signals, such as degrading cells, unstable customer equipment or fibre damage from digging."},{"title":"Build the labelled history","detail":"Join past faults to the data that preceded them. This is usually most of the work and the main reason projects stall."},{"title":"Decide the action before the model","detail":"Agree what happens when the score is high: an automated reset, a planned visit or a call to a third party. A prediction nobody acts on is only a report."},{"title":"Automate only the safe fixes","detail":"Let the system run reversible remote actions within limits and route everything else to engineers and field planners with the evidence attached."},{"title":"Measure prevented and missed failures","detail":"Track both, because a model that raises many alarms can look busy while missing the failures that matter."}],"guardrails":["Automatic actions limited to reversible, low impact fixes with rollback and rate limits","Hardware swaps and site visits approved by planners, not triggered directly by a score","Maintenance windows and change freezes respected by every automated action","Human review of any action that affects many customers at once"],"humanInTheLoop":"Engineers set the thresholds and the list of automated fixes, planners approve preventive visits, and the NOC can pause automation at any time. A sample of automated actions is reviewed every week against what actually happened to the element afterwards.","kpisToInstrument":["Faults prevented, measured against a comparable control group of elements","Precision of predictions, as the share of flagged elements that really degraded","Customer contacts and technician visits per thousand customers","Mean time between failures for the targeted element types","Automated actions that had to be rolled back"],"failureModes":[{"title":"Predictions without actions","detail":"The model scores risk accurately but nobody owns the follow up. Tie every score band to a named action and owner."},{"title":"Preventive work on healthy equipment","detail":"Too many false alarms send technicians to sites that were fine. Track precision and the cost of each preventive visit."},{"title":"Automation that causes outages","detail":"An automated reset during peak hours takes down more customers than the fault would have. Use windows, rate limits and rollback."},{"title":"Drift after network change","detail":"New equipment and software releases change what normal looks like. Retrain and revalidate after major upgrades."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Scoring failure risk and planning maintenance is normally minimal risk. Annex III point 2 lists AI systems intended as safety components in the management and operation of critical digital infrastructure as high risk, and public electronic communications networks fall within that infrastructure. Recital 55 limits safety components to systems that directly protect the physical integrity of the infrastructure or the health and safety of persons and property, and excludes components used solely for cybersecurity. An operator whose automated actions meet that test must treat the system as high risk."},"regulations":["eu-ai-act","nist-ai-rmf","iso-42001","nis2"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 2 covers AI systems intended as safety components in the management and operation of critical digital infrastructure."},{"title":"Recital 55, safety components of critical infrastructure","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/recital/55/","note":"Explains which critical digital infrastructure is meant and what counts as a safety component, and excludes components used solely for cybersecurity."},{"title":"Directive (EU) 2022/2557 on the resilience of critical entities","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2022/2557/oj","note":"Point 8 of the Annex (digital infrastructure) lists providers of public electronic communications networks, the infrastructure that Recital 55 of the AI Act points to for Annex III point 2."},{"title":"NIS2 Directive, securing network and information systems","issuer":"European Commission","region":"europe","url":"https://digital-strategy.ec.europa.eu/en/policies/nis2-directive","note":"NIS2 covers providers of public electronic communications networks and services, with risk management and incident reporting duties for their network and information systems, which include the automation that acts on the network."}],"controls":["Documented list of automated actions with owners, limits and rollback procedures","Change control for model thresholds and automation rules","Audit trail linking each prediction to the action taken and the outcome","Periodic validation of model precision and recall on recent failures"],"incidents":[]},"blitsAi":{"howToBuild":"Prediction models usually run in the operator's own data platform; Blits.ai adds the layer that\nacts on them. **Agentic tasks** watch for high risk scores (\"when a cell's risk passes the\nthreshold, check recent changes and propose a fix\") and **agentic workflows** call **custom\nfunctions** against the operator's device management, ticketing and scheduling APIs, with **human\nin the loop approval** above a configurable impact threshold and a **tool execution policy** that\nallows only reversible actions without approval.\n\nEngineers and planners query the results through an **AI agent** with a **SQL knowledge base**\nover prediction tables held in a supported SQL database and a **knowledge base** of runbooks. Where\na fault will affect customers, an agentic workflow sends the notice through the outbound\n**email** channel or a **custom function** that calls the operator's SMS or messaging gateway, and\ncustomers who reply or call reach the same agent on the **SMS**, **WhatsApp**, **email** or\n**voice** channel. **Test suites**, **monitors** and full run audit trails keep the automation\nreviewable, and the platform is model agnostic."},"faq":[{"question":"What does predictive maintenance look like at scale in a telecom operator?","answer":"Telstra reports that its SmartFix system performed 2.5 million proactive actions in FY25, fixing many issues before customers noticed and preventing nearly 1 million support calls. Nokia reported in December 2022 that KDDI monitors its 4G and 5G radio network around the clock with Nokia's AVA PDDR, which detects silent cell degradations that raise no alarm and hands them to KDDI's automatic recovery system."},{"question":"Is it only about network equipment?","answer":"No. Verizon applies machine learning to more than ten million 811 dig requests a year to identify high risk excavations near its fiber, and takes preventive steps such as extra communication with the excavator. The same approach can target customer premises equipment, site power and batteries."},{"question":"Where should an operator start?","answer":"With one frequent, costly failure mode that has measurable warning signs and an agreed action, and with the labelled history that links past faults to the data that preceded them."}],"related":["network-fault-triage-copilot","field-technician-copilot-and-dispatch","autonomous-network-operations","network-outage-communication-agent","device-and-connectivity-troubleshooting-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched for the telecommunications vertical and verified against operator and vendor sources."},{"date":"2026-09-25","note":"Consolidation pass: added NIS2 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; corrected the EU AI Act basis and added Recital 55 as guidance; made the NIS2 note precise; aligned the Verizon and KDDI statements in the problem and FAQ with the sources; added a Wayback copy for the KDDI release."},{"date":"2026-09-27","note":"Review fixes: limited Blits.ai customer notifications to SMS, WhatsApp and email (voice stays inbound); added the CER Directive Annex as guidance for the critical infrastructure claim; clarified that the Telstra figure does not calibrate the preventable share; dated the KDDI FAQ statement; removed languages and channels on evidence records that the sources do not state."},{"date":"2026-09-27","note":"Fact checked again against the live Telstra, Verizon and Nokia pages, the AI Act texts and the NIS2 page: softened two unsourced statements in the problem section (share of reactive maintenance, cost of a fault) and made the NIS2 note precise; evidence unchanged."},{"date":"2026-09-27","note":"Review fix: rewrote the customer notification sentence so only inventory capabilities are claimed, outbound email or a custom function to the operator's own SMS or messaging gateway, with SMS, WhatsApp, email and voice as reply channels."}],"slug":"predictive-network-maintenance","url":"https://www.blits.ai/ai-use-cases/predictive-network-maintenance","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":6250000,"min":2500000,"max":10000000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"verizon-fiber-cut-prevention","pooled":true},{"id":"telstra-smartfix-proactive-network-fixes","pooled":true}]}],"indicativeValueResult":{"low":600000,"high":12000000},"evidence":["kddi-nokia-performance-degradation-detection","orange-augtera-noc-alarm-correlation","telefonica-espana-network-analytics-optimization","telstra-smartfix-proactive-network-fixes","verizon-fiber-cut-prevention","vodafone-nokia-network-anomaly-detection"]},{"title":"AI for public consultation response analysis","shortTitle":"Consultation response analysis","seoTitle":"AI analysis of public consultation responses","metaDescription":"AI proposes themes and maps every consultation response for human review. The UK Department for Transport reports over 92% raw agreement with human coders.","definition":"AI that reads every free text response to a public consultation or rulemaking comment period, proposes themes, maps each response to the themes that officials have validated, flags duplicates, campaign letters and responses that need special attention, and produces counts and summaries for the analysts who write the government's response.","aliases":["consultation analysis AI","public comment analysis","AI thematic analysis of consultation responses","rulemaking comment review"],"industries":["government"],"functions":["citizen-services","analytics-and-reporting"],"patterns":["summarization","classification-and-routing","content-generation"],"channels":["internal-tools"],"audience":"back-office","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Governments ask the public for views before they change policy or make rules, and the answers\narrive as free text: a few hundred responses to a technical consultation, or tens of thousands\nwhen an issue catches public attention. Every response has to be read, coded against a set of\nthemes and counted, so that officials can show what people said and how it shaped the decision.\nDone by hand this can take months and, according to the UK Department for Transport, typically\nconsumes over half of the consultation budget; the UK government notes that the work is often\noutsourced to contractors.\n\nSpeed is not the only problem. Coding is subjective, so two analysts can put the same response\nunder different themes, and for very large consultations teams sometimes analyse a sample instead\nof every response. Mass campaigns and duplicate letters distort counts, and fake submissions have been\nused to manufacture the appearance of public support. Whatever tool is used, the government has\nto be able to show that every voice was heard and that the analysis was fair.","problemStats":[{"statement":"Across the 500 consultations it runs each year, the UK government estimates that an AI tool could save officials around 75,000 days of analysis a year, work that costs GBP 20 million in staffing costs.","sourceTitle":"Government-built \"Humphrey\" AI tool reviews responses to consultation for first time, in bid to save millions","sourceUrl":"https://www.gov.uk/government/news/government-built-humphrey-ai-tool-reviews-responses-to-consultation-for-first-time-in-bid-to-save-millions","year":2025},{"statement":"The UK Department for Transport alone runs around 55 consultations a year, each generating free text that requires thematic analysis.","sourceTitle":"AI Consultation Analysis Tool v1.0 evaluation","sourceUrl":"https://assets.publishing.service.gov.uk/media/696f6641011505255b2d4203/ai-consultation-analysis-tool-CAT-evaluation.pdf","year":2025}],"howItWorks":"1. **Load and clean the responses.** Responses from the consultation platform, email and\n   regulations.gov style dockets are loaded per question, with personal data masked and exact and\n   near duplicates (campaign letters) grouped so they are counted but read once.\n2. **Propose themes.** A language model, or an ensemble of models, reads the responses to each\n   open question and proposes a set of themes, including rare but important points, with example\n   responses for each.\n3. **Validate the theme set with people.** Analysts read a random sample of responses, merge,\n   split, rename and add themes. Only the validated theme set is used from here on.\n4. **Map every response.** The model assigns each response to one or more validated themes and\n   records its stance (agree, disagree, neutral) where the question asks for one. Responses that\n   are off topic, abusive or that raise safeguarding concerns are flagged for a human.\n5. **Check and report.** Analysts review a sample of the mapping, correct errors and use the\n   counts, summaries and representative quotes (always verified against the original response)\n   to write the consultation response.","valueDrivers":["employee-productivity","speed","cost-to-serve","compliance"],"kpis":["hours-saved","cost-reduction","cost-savings","accuracy","interactions-handled","processing-time-reduction"],"indicativeValue":{"referenceOrg":"A national ministry that runs 20 public consultations a year with substantial free text","inputs":[{"key":"consultations","label":"Consultations with free text analysis per year","low":10,"high":30,"unit":"consultations per year","note":"Editorial assumption. The UK Department for Transport runs around 55 a year; replace with your own portfolio."},{"key":"costPerConsultation","label":"Cost of analysing and reporting one medium sized consultation","low":50000,"high":100000,"unit":"EUR per consultation","note":"Editorial assumption, informed by the Department for Transport's estimate of GBP 80,000 to 100,000 for a medium sized consultation (about 15,000 responses). Replace with your own staff or contractor cost.","sourceUrl":"https://assets.publishing.service.gov.uk/media/696f6641011505255b2d4203/ai-consultation-analysis-tool-CAT-evaluation.pdf"},{"key":"savingShare","label":"Share of that cost saved","low":0.3,"high":0.5,"unit":"fraction of cost","note":"Conservative against the Department for Transport's modelled estimate of 50 to 70 percent for a notional medium sized consultation, because theme review, synthesis and report writing remain human work and small consultations save less.","sourceUrl":"https://assets.publishing.service.gov.uk/media/696f6641011505255b2d4203/ai-consultation-analysis-tool-CAT-evaluation.pdf"}],"formula":"consultations * costPerConsultation * savingShare","currency":"EUR","period":"per year","resultLabel":"Consultation analysis cost avoided","caveat":"Gross analysis cost avoided only. It leaves out the cost of running and assuring the tool, the value of faster policy decisions, and the option of analysing every response in consultations that are sampled today."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"The data is text the government already holds and no transaction systems are touched. The work is in method, not integration: a defensible human review step, an evaluation against human coded samples and a bias check across respondent groups.","dataPrerequisites":["Exported responses per question, with respondent type (individual, organisation) where collected","A few previously human coded consultations to evaluate against","A policy on how personal data in responses is masked and retained"],"integrations":["Consultation platform or regulations.gov style docket export","Email inbox for responses submitted outside the platform","Analysis workspace or dashboard for analysts"]},"implementation":{"steps":[{"title":"Evaluate on consultations you have already coded","detail":"Run the tool on two or three past consultations that humans coded, blind, and compare theme recall and mapping agreement with the human result before you use it live. Publish the method, as the UK Department for Transport did."},{"title":"Design the human theme review","detail":"Decide how many responses analysts read per question before they sign off the theme set, and write it down. The Department for Transport reports 1 to 5 hours per 100 responses for this step."},{"title":"Run the first live consultation in parallel","detail":"On the first live use, let analysts also code every response by hand, as the Scottish Government did with Consult, and measure where the two disagree."},{"title":"Handle campaigns and duplicates explicitly","detail":"Group identical and near identical responses, count them, and report organised campaigns separately so that one template letter does not read as thousands of independent views."},{"title":"Check for bias across respondent groups","detail":"Compare mapping accuracy for responses from different groups (for example by writing style, language or respondent type) and act on any gap before the method is used at scale."}],"guardrails":["Only human validated themes are used for mapping and counts","Every quote in the report is copied from the original response, never from a model summary","Personal data is masked before responses reach a model and in stored outputs","The tool never decides policy or weights responses; it organises them for analysts","Duplicate and campaign detection is reported, not used to discard responses"],"humanInTheLoop":"Analysts own the theme framework, review a random sample of the mapping for every question and write the response. Policy officials see the underlying responses behind every theme. Responses flagged for safeguarding or abuse go to a named person.","kpisToInstrument":["Theme recall and mapping agreement against a human coded sample, per question","Analyst hours per 1,000 responses, before and after","Days from consultation close to published response","Share of mapped responses changed by analysts during review","Accuracy differences across respondent groups"],"failureModes":[{"title":"Missing the rare but important point","detail":"Models favour frequent themes and can miss a single expert response that changes the policy. Ask for rare themes explicitly and have analysts read a random sample."},{"title":"Hallucinated or softened quotes","detail":"A summary that paraphrases respondents can put words in their mouths. Pull quotes from the source text only."},{"title":"Counting as if consultations were polls","detail":"Consultation respondents are self selected; reporting that a percentage of respondents felt something invites misreading. Report counts with context, as the Department for Transport cautions."},{"title":"Campaigns and fake submissions distorting results","detail":"Mass template letters or fabricated submissions can swamp genuine views. Detect duplicates and unusual submission patterns and report them openly."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Organising and summarising consultation responses for analysts does not decide on individuals and is not listed in Annex III, so no high risk obligations apply. If AI generated text is published to inform the public on matters of public interest without human review and editorial responsibility, Article 50(4) requires disclosure."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Algorithmic Transparency Recording Standard hub","issuer":"Government Digital Service","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"The standard is mandatory for all UK government departments; the Department for Science, Innovation and Technology has published a record for Consult describing its purpose, data, human oversight and risks."},{"title":"DSIT: Consult (algorithmic transparency record)","issuer":"Department for Science, Innovation and Technology","region":"europe","url":"https://www.gov.uk/algorithmic-transparency-records/dsit-consult","note":"The published transparency record for the UK government's Consult tool, an example of what to disclose about a consultation analysis tool."},{"title":"Public attitudes to the use of AI in DfT consultations and correspondence","issuer":"Department for Transport","region":"europe","url":"https://www.gov.uk/government/publications/public-attitudes-to-the-use-of-ai-in-dft-consultations-and-correspondence","note":"Research on what the public expects before AI is used to analyse their consultation responses; it shaped the design of the Department for Transport's tool."}],"controls":["Published method and evaluation before live use, including a transparency record","Documented human theme review with a sample size rule","Audit trail from each reported count back to the underlying responses","Bias testing across respondent groups on every major model or prompt change","Retention and masking rules for personal data in responses"],"incidents":[{"title":"New York Attorney General: millions of fake comments in the FCC's 2017 net neutrality proceeding","url":"https://ag.ny.gov/press-release/2021/attorney-general-james-issues-report-detailing-millions-fake-comments-revealing","note":"The investigation found that nearly 18 million of the more than 22 million comments were fake, including 9.3 million using fictitious identities, most of them submitted by one person using automated software. Comment analysis must detect campaigns and fabricated submissions rather than count them as public opinion."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that runs over the exported responses. Responses are\nuploaded as CSV or spreadsheet files into the **document library**, **PII masking** at the\ngateway masks personal data, and an **agent** with **structured output** proposes themes per\nquestion. A **human in the loop** confirmation step lets an analyst approve or reject the\nproposed theme set before the workflow maps every response to the approved themes. Loaded into a database registered as a\n**SQL knowledge base**, the results can be queried by an agent when analysts ask about counts\nand themes.\n\n**Test suites** run previously human coded responses against the workflow and grade its theme\nassignments, so every prompt or model change is checked before live use. Every run keeps a full\naudit trail and downloadable run data. The platform is **model agnostic** and offers EU and UAE\ndata residency, so a ministry can keep responses in region and choose the models it uses."},"faq":[{"question":"How accurate is AI at coding consultation responses?","answer":"Close to human coders when humans validate the themes. The UK Department for Transport reports over 92% raw agreement between its tool and human coders (an F1 score of 0.75 for theme mapping in its blind evaluation), and the UK government reported an F1 score of 0.76 on Consult's first live consultation. Human review of the theme set remains necessary; without it, the Department for Transport's tool found about 75% of the human themes."},{"question":"Can AI replace the analysts?","answer":"No. It replaces most of the reading and tagging, while analysts decide the themes, check the mapping and write the response. The Department for Transport estimates, in a model rather than a measured result, savings of 50 to 70% of the cost of a medium sized consultation, with review, synthesis and report writing remaining human work."},{"question":"Should respondents be told that AI analyses their responses?","answer":"Yes. Say so in the consultation document and privacy notice, publish the method and keep a human accountable for the analysis. In the UK the Algorithmic Transparency Recording Standard is mandatory for government departments, and the Department for Science, Innovation and Technology has published a record for Consult."}],"related":["customer-feedback-analysis","civil-servant-drafting-copilot","complaints-root-cause-analysis","freedom-of-information-request-processing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version for the government vertical, with UK and US public sector evidence verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: reworded the UK 75,000 days statistic as the government's savings estimate, aligned the sampling claim with the DfT evaluation, corrected the transparency hub issuer, tightened the incident note and the Blits.ai build description to listed capabilities, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: removed the modelled 50 to 70 percent cost estimate from the Department for Transport metrics, sourced the transparency record claim to the DSIT Consult record and the mandatory ATRS policy, added the blind F1 score for balance, and marked the DfT savings figures as cumulative and modelled."},{"date":"2026-09-27","note":"Fact checked against sources: restated the 75,000 days estimate as per year, sourced the manual cost and outsourcing statement to the DfT evaluation and the DSIT press release, clarified the blind F1 score as theme mapping, and corrected two evidence publishers."}],"slug":"public-consultation-response-analysis","url":"https://www.blits.ai/ai-use-cases/public-consultation-response-analysis","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":101000,"min":2000,"max":200000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"uk-department-for-transport-consultation-analysis-tool","pooled":true},{"id":"uk-incubator-for-ai-consult-consultation-analysis","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":92,"min":92,"max":92,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"uk-department-for-transport-consultation-analysis-tool","pooled":true}]},{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"GBP","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":500000,"min":500000,"max":500000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"uk-department-for-transport-consultation-analysis-tool","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":15000,"min":15000,"max":15000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"uk-department-for-transport-consultation-analysis-tool","pooled":true}]}],"indicativeValueResult":{"low":150000,"high":1500000},"evidence":["cdc-public-comment-stance-analysis","federal-reserve-board-public-comment-review-system","uk-department-for-transport-consultation-analysis-tool","uk-incubator-for-ai-consult-consultation-analysis","us-department-of-transportation-public-comment-analyzer"]},{"title":"AI for radio access network energy optimization","shortTitle":"RAN energy optimization","seoTitle":"AI for RAN energy optimization and cell sleep","metaDescription":"AI puts idle radio carriers to sleep in quiet hours. Telefónica saved up to 8% of a 5G test site's energy, and BT Group runs cell sleep on over 19,500 EE sites.","definition":"Machine learning that predicts traffic per cell and puts radio carriers, cells and hardware components into sleep modes when demand is low, then wakes them before users notice, so a mobile network uses less electricity without losing coverage or quality.","aliases":["AI energy saving for mobile networks","cell sleep optimization","RAN power saving","green RAN AI"],"industries":["telecommunications"],"functions":["network-operations"],"patterns":["prediction-and-scoring"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"autonomous","adoptionStage":"early-adopters","segment":"network","problem":"For operators such as BT Group and Orange, the network uses most of the energy. BT Group says its\nnetworks account for around 89 per cent of its total energy consumption, and Orange puts IT and\nnetworks at around 85% of its energy requirements. In a mobile network, part of that power keeps radio capacity switched\non in periods when it is not needed, such as quiet nights: this is the capacity that cell sleep\nfeatures switch off.\n\nVendors ship power saving features (carrier shutdown, micro sleep, deep sleep), but switching them\non with fixed schedules leaves savings on the table in quiet cells and risks quality in busy ones.\nThe settings differ per cell, traffic patterns change with events, holidays and new sites, and\nnobody can tune tens of thousands of cells by hand. Energy prices and net zero targets make it a\ncost and climate priority: Orange stepped up its energy saving measures during the 2022 energy\ncrisis, and BT Group calls network energy efficiency integral to its net zero ambition.","problemStats":[{"statement":"BT Group states that its networks account for around 89 per cent of its total energy consumption.","sourceTitle":"BT Group rolls-out energy-saving ‘cell sleep’ technology to EE mobile sites nationwide","sourceUrl":"https://newsroom.bt.com/bt-group-rolls-out-energy-saving-cell-sleep-technology-to-ee-mobile-sites-nationwide/","year":2024},{"statement":"Orange states that IT and networks represent around 85% of the group's energy requirements.","sourceTitle":"Orange steps up efforts to reduce energy consumption across Europe","sourceUrl":"https://newsroom.orange.com/orange-steps-up-efforts-to-reduce-energy-consumption-across-europe/","year":2022}],"howItWorks":"1. **Learn each cell's rhythm.** Models forecast traffic per cell and carrier from history,\n   calendar effects and local events.\n2. **Choose the saving action.** For each forecast quiet period the system picks the deepest\n   power saving mode that the cell can use safely: switching off capacity carriers, micro sleep,\n   deep sleep of radio units or shutdown of idle components.\n3. **Protect quality.** Coverage layers stay on, neighbouring cells absorb the remaining traffic,\n   and the system watches live load so sleeping capacity wakes within seconds if demand rises.\n4. **Tune per cell.** Thresholds are adjusted per cell from measured quality and savings, rather\n   than one setting for the whole network.\n5. **Report.** Energy saved, quality indicators and wake up events are reported per site and\n   cluster, so engineers can see where savings cost quality and adjust.","valueDrivers":["cost-to-serve"],"kpis":["energy-savings","cost-reduction","cost-savings"],"indicativeValue":{"referenceOrg":"A mobile operator with 10,000 radio sites","inputs":[{"key":"sites","label":"Radio sites in scope","low":10000,"high":10000,"unit":"sites","note":"The reference operator."},{"key":"kwhPerSite","label":"Electricity use per radio site per year","low":30000,"high":50000,"unit":"kWh per site per year","note":"Editorial assumption for a multi band macro site, including radio, power and cooling equipment. Replace with metered data."},{"key":"savingShare","label":"Share of site electricity saved by AI driven sleep modes","low":0.005,"high":0.024,"unit":"fraction of site energy","note":"BT Group expects 4.5m kWh a year across EE's more than 19,500 sites, about 0.5 to 0.8% of the assumed site consumption estate wide, with a per site ceiling of up to 2 kWh a day (about 730 kWh a year, 1.5 to 2.4% of the assumed site consumption). The range spans that estate wide expectation up to the per site ceiling. Telefónica reports savings of up to 8% of a 5G site's 24 hour consumption, but in a single site test, and Nokia expects planned network energy cost savings of 8 to 10% for Safaricom; both are above this range and are not used to set it."},{"key":"pricePerKwh","label":"Electricity price","low":0.1,"high":0.25,"unit":"USD per kWh","note":"Editorial assumption. Replace with your own contracted price."}],"formula":"sites * kwhPerSite * savingShare * pricePerKwh","currency":"USD","period":"per year","resultLabel":"Radio network electricity cost avoided","caveat":"Site electricity cost only. It leaves out carbon value, longer battery backup during grid outages, software licence and integration costs, and any quality impact that has to be compensated."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The power saving features usually exist in the radio equipment already. The work is in reliable per cell traffic data, safe automation across vendors and convincing radio engineers that quality will hold.","dataPrerequisites":["Traffic and quality counters per cell and carrier at 15 minute or finer granularity","Site energy metering, ideally per site rather than estimated","Configuration and capability data for the power saving features per vendor and software release","Calendar of events and planned works"],"integrations":["Radio network management and configuration systems per vendor","Performance management and network data platform","Energy metering and site management systems","Self organizing network (SON) platform where one exists"]},"implementation":{"steps":[{"title":"Measure the baseline","detail":"Meter energy per site and record quality indicators before any change, so savings and quality impact can be proven rather than estimated."},{"title":"Switch on vendor features with guardrails","detail":"Activate the available sleep features with conservative thresholds in a cluster, and compare with a control cluster."},{"title":"Add prediction and per cell tuning","detail":"Replace fixed schedules with traffic forecasts and per cell thresholds, and let the system tune them from measured quality."},{"title":"Scale by cluster, not by country","detail":"Roll out region by region with quality checks at each step, and keep special sites (hospitals, stadiums, transport hubs) under manual rules."},{"title":"Report savings finance can trust","detail":"Agree the measurement method with finance and sustainability teams up front, so the savings count in budgets and emissions reporting."}],"guardrails":["Coverage layers and emergency service capability are never switched off","Automatic wake up on load thresholds, with a maximum wake up time per mode","Exclusion lists for critical sites and events, maintained by radio engineering","Quality key performance indicators monitored per cell, with automatic rollback when they degrade"],"humanInTheLoop":"Radio engineers set the guardrails, exclusion lists and quality thresholds, and review weekly reports of savings against quality per cluster. The system acts on its own within those limits, because sleep and wake decisions across thousands of cells happen too often to approve one by one.","kpisToInstrument":["Energy saved per site and per cluster against a metered baseline or control group","Accessibility, retainability and throughput per cell during sleep periods","Number and duration of wake up events","Customer complaints about coverage in optimized areas"],"failureModes":[{"title":"Savings that exist only in the model","detail":"Savings are estimated from switch off time rather than metered. Use metered energy and control clusters."},{"title":"Quality loss at the edges","detail":"Neighbouring cells cannot absorb the traffic and users at the cell edge lose throughput. Monitor edge quality, not just averages."},{"title":"Events the forecast did not know","detail":"A match, a concert or an emergency brings traffic to a sleeping area. Feed event calendars and keep fast wake up paths."},{"title":"Vendor lock in of the optimizer","detail":"A rollout can end up covering only one vendor's part of the network (O2 Telefónica Germany uses Nokia's software on the Nokia part of its radio network, and Indosat Ooredoo Hutchison's rollout covers its Nokia radio footprint in four regions). Check multi vendor support and plan for a view across vendors where networks are mixed."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Optimizing energy use is normally minimal risk. Under Article 6(2), Annex III point 2 lists AI systems intended as safety components in the management and operation of critical digital infrastructure as high risk, and public electronic communications networks fall under that infrastructure. Recital 55 limits safety components to systems that directly protect the infrastructure or the health and safety of persons, so an optimizer is not high risk by default, but a design in which it could affect emergency service availability should be assessed against point 2."},"regulations":["eu-ai-act","nis2","nist-ai-rmf","iso-42001"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 2 covers AI systems intended as safety components in the management and operation of critical digital infrastructure."},{"title":"Recital 55, safety components of critical infrastructure","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/recital/55/","note":"Defines safety components as systems that directly protect the physical integrity of critical infrastructure or the health and safety of persons and property, and refers to the digital infrastructure listed in point 8 of the Annex to Directive (EU) 2022/2557."},{"title":"Directive (EU) 2022/2557 on the resilience of critical entities","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2022/2557/oj","note":"Point 8 of the Annex (digital infrastructure) lists providers of public electronic communications networks, which is why mobile radio networks count as critical digital infrastructure for Annex III point 2."}],"controls":["Documented guardrails and exclusion lists with an accountable radio engineering owner","Change control for thresholds and new power saving modes","Monitoring of quality indicators with automatic rollback","Metered measurement method agreed with finance and sustainability reporting"],"incidents":[]},"blitsAi":{"howToBuild":"The sleep decisions themselves run in the radio vendors' software or a SON platform, close to the\nnetwork. Blits.ai fits around that loop: an **AI agent** with a **SQL knowledge base** over energy\nand quality tables lets engineers and sustainability teams ask where savings and quality moved,\nand a **knowledge base** holds the vendor feature documentation.\n\nAn **agentic task** or a scheduled **agentic workflow** can watch for quality drops in optimized\nclusters and propose a threshold change or an exclusion through **custom functions**, with\n**human in the loop approval** before anything is applied. Workflows can run on a schedule to\nproduce the daily savings and quality summary, **monitors** run recurring health checks on the\nagents themselves, and the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"How much energy can AI save in a radio network?","answer":"The results on this page are site or trial level. Telefónica reports that Ericsson's Radio Deep Sleep Mode, supported by AI and machine learning, saved up to 8% of a 5G site's 24 hour consumption and up to 26% in low traffic hours in a Madrid test. The operators and vendors cited here give network wide figures only as expectations: Nokia expects its software to deliver planned network energy cost savings of 8 to 10% across about 30,000 Safaricom cells, and BT Group expects up to 2 kWh per site per day."},{"question":"Is this different from the power saving features vendors already ship?","answer":"The features are the same; the difference is when and where they are used. BT Group puts capacity carriers to sleep based on quiet periods predicted for each site through machine learning, instead of one fixed schedule for the whole network."},{"question":"Does it hurt network quality?","answer":"It should not if coverage layers stay on, capacity wakes within seconds and quality is monitored per cell with automatic rollback. BT Group says its sleeping carriers wake within seconds without interruption to customers, and Telefónica's platforms periodically review quality so as not to affect network performance or user experience."}],"related":["network-planning-and-capacity-optimization","autonomous-network-operations","predictive-network-maintenance"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched for the telecommunications vertical and verified against operator and vendor sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources (Nokia pages through Wayback copies). Sourced the problem framing, rescoped the value model to total site electricity with a 2 to 8% saving range, replaced an unsupported FAQ claim with the published Safaricom and BT Group expectations, corrected the vendor lock in failure mode, sharpened the EU AI Act basis with recital 55 and Directive (EU) 2022/2557, removed the agentic workflow pattern and unrelated value drivers, added the energy savings KPI, fixed the monitors description, and added the SEO title and description."},{"date":"2026-09-27","note":"Review fixes: limited the FAQ claim about published results to the evidence on this page, attributed the Safaricom 8 to 10% expectation to Nokia, lowered the saving range to 1.5 to 5%, sourced the problem framing, corrected the vendor scope failure mode, dropped the anomaly detection pattern, aligned the agentic wording with the platform features and made the spelling consistent."},{"date":"2026-09-27","note":"Fact checked against sources: rechecked all five evidence pages (the three Nokia releases on the live pages), the Orange and BT Group problem statistics, Annex III, recital 55 and the annex to Directive (EU) 2022/2557. Added the NIS2 Directive, which covers providers of public electronic communications networks, completed the recital 55 note and made the Indosat Ooredoo Hutchison scope in the vendor lock in failure mode precise."},{"date":"2026-09-27","note":"Review fixes: lowered the saving range to 0.5 to 2.4% so it spans BT Group's estate wide expectation (4.5m kWh a year across more than 19,500 sites) up to its per site ceiling, instead of a range that in fact sat above the estate wide figure. Added the estate wide figure to the note and removed the inaccurate \"kept below the figures on this page\" and \"in line with BT Group's expectation\" wording."}],"slug":"ran-energy-optimization","url":"https://www.blits.ai/ai-use-cases/ran-energy-optimization","benchmarks":[{"kpi":"energy-savings","label":"Energy savings","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"telefonica-ai-radio-power-saving-features","pooled":false}]}],"indicativeValueResult":{"low":150000,"high":3000000},"evidence":["bt-ee-cell-sleep-energy-saving","indosat-ooredoo-hutchison-nokia-energy-efficiency","o2-telefonica-germany-nokia-energy-saas","safaricom-nokia-ava-energy-efficiency","telefonica-ai-radio-power-saving-features"]},{"title":"AI for recruitment screening and interview scheduling","shortTitle":"Recruitment screening and scheduling","seoTitle":"AI recruitment screening and interview scheduling","metaDescription":"AI answers candidates, prequalifies them and books interviews, with cases from Chipotle and Gojob. In the EU, AI that screens CVs is high risk under the AI Act.","definition":"AI that answers candidates' questions, collects applications in conversation, schedules interviews and, where the organization chooses, assesses applications against the job requirements for a recruiter, who makes every selection decision. In the EU, the screening part is a high risk AI system under Annex III point 4 of the AI Act.","aliases":["AI recruiting assistant","conversational hiring assistant","AI resume screening","CV screening with AI","automated interview scheduling","candidate prequalification chatbot"],"industries":["cross-industry","government","travel-and-hospitality","professional-services"],"functions":["human-resources"],"patterns":["conversational-agent","classification-and-routing","prediction-and-scoring","agentic-workflow"],"channels":["sms","whatsapp","web-chat","mobile-app","email"],"audience":"customer-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"High volume hiring is an administrative marathon. A restaurant chain, retailer, staffing agency or\npublic body receives large numbers of applications for similar roles, and recruiters and hiring\nmanagers spend their time chasing candidates for basic information, answering the same questions\nand trading emails to find interview slots. At the temporary work agency Gojob, a recruiter had to\ncontact between 50 and 80 candidates to fill a single vacancy. Candidates who wait too long may\naccept another offer. For specialist roles the bottleneck is reading: at the consultancy Trace3, working\nthrough résumés could take HR managers up to three weeks. US Immigration and Customs Enforcement\nnames bias and variability in how HR specialists evaluate resumes as problems its screening tool\nis meant to reduce.\n\nAI helps on both fronts, but the two halves carry very different risk. Answering questions,\ncollecting information and booking interviews is logistics. Ranking, filtering or scoring people\ndecides who gets a chance at a job, and it has a documented history of encoding bias, from\nAmazon's abandoned experimental ranking tool to recruitment tools, audited by the UK ICO, that let\nrecruiters filter out candidates with certain protected characteristics. In the EU, AI that analyses and filters applications or evaluates\ncandidates is high risk under the AI Act. Design the system so the logistics run fast and the\njudgment stays demonstrably human.","problemStats":[],"howItWorks":"1. **Engage the candidate.** A conversational assistant on the careers site, SMS or WhatsApp\n   answers questions about the role, pay, shifts and process, in the candidate's language, and\n   says clearly that it is an AI.\n2. **Collect the application.** It gathers the information the job actually requires (right to\n   work, availability, location, required licences or qualifications) and nothing more.\n3. **Apply knockout rules transparently.** Objective, job related minimum requirements that the\n   organization has documented are checked; candidates who do not meet them are told why and how\n   to ask for a human review.\n4. **Summarize, and only if chosen, assess.** For roles where the organization decides to use it,\n   the AI compares the application with the documented job requirements and shows the recruiter\n   matching and missing experience with evidence, without a pass or fail decision.\n5. **Schedule.** The assistant offers interview slots from the hiring manager's calendar, books,\n   reschedules and sends reminders.\n6. **Humans decide.** Recruiters and hiring managers review, interview and select; the offer\n   is sent only for candidates they choose.","valueDrivers":["speed","employee-productivity","customer-experience","inclusion-and-access"],"kpis":["cycle-time-days","interactions-handled","processing-time-reduction","productivity-gain"],"indicativeValue":{"referenceOrg":"An employer that hires 5,000 people a year into high volume roles","inputs":[{"key":"hires","label":"Hires per year","low":5000,"high":5000,"unit":"hires per year","note":"The reference organization."},{"key":"adminHoursPerHire","label":"Recruiter and manager admin hours per hire (chasing, scheduling, answering questions)","low":2,"high":4,"unit":"hours per hire","note":"Editorial assumption for high volume hiring. Replace with your own time study."},{"key":"automatedShare","label":"Share of that admin the assistant takes over","low":0.4,"high":0.7,"unit":"fraction of admin hours","note":"Editorial assumption; screening decisions and interviews are excluded and stay with people."},{"key":"hourlyCost","label":"Blended hourly cost of recruiters and hiring managers","low":35,"high":60,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"hires * adminHoursPerHire * automatedShare * hourlyCost","currency":"USD","period":"per year","resultLabel":"Recruiting admin time released","caveat":"Counts only administrative time around scheduling and candidate communication. It leaves out the value of filling roles faster (fewer lost candidates, less overtime), the cost of the platform, and the compliance cost of a high risk system if screening is included."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Scheduling and candidate questions are medium effort: integration with the applicant tracking system and calendars. Screening is high effort because it is a high risk system in the EU and regulated elsewhere: documented job criteria, bias testing, human oversight, logging, candidate notices and, for those who build it, the provider obligations of the AI Act.","dataPrerequisites":["Documented, job related requirements per role, approved by HR and legal","Approved answers on pay, benefits, process and policies for candidate questions","Hiring manager calendars and interview formats","Historical outcome data by stage, to test for adverse impact before and after launch"],"integrations":["Applicant tracking system (for example Workday, SAP SuccessFactors or a specialist ATS)","Calendar systems for interview scheduling","Careers site, SMS and WhatsApp channels","HR information system for offers and onboarding handover"]},"implementation":{"steps":[{"title":"Separate logistics from judgment in the design","detail":"Write down which steps the AI performs (answer, collect, schedule) and which it only informs (assessment). Classify each under the AI Act and local law before building, and document the reasoning."},{"title":"Define criteria that are job related","detail":"Knockout questions and assessment criteria must be objective, necessary for the role and approved by HR and legal. Remove proxies for protected characteristics such as names, photos, postcodes and gaps explained by care or illness."},{"title":"Build the candidate experience first","detail":"Launch question answering and scheduling, with AI disclosure and a route to a person, and measure drop off and time to interview before adding any assessment."},{"title":"If you add screening, test for adverse impact","detail":"Before launch and on a schedule, compare selection rates across groups on real outcomes, with an independent review. Keep a written record of tests, results and fixes."},{"title":"Make human review real","detail":"Recruiters see the evidence behind any assessment, can override it easily, and are trained on the tool's limits. Track how often they disagree; zero disagreement is a warning sign."},{"title":"Tell candidates and workers","detail":"Inform candidates that AI is used and how to request human review, and inform workers' representatives where required, before the system is used."}],"guardrails":["The AI never rejects or selects a candidate on its own; outcomes are decided and recorded by a person","Only documented, job related criteria are assessed; no inference of protected characteristics, personality or emotion","No analysis of facial expressions or voice to judge candidates","Candidates are told they are dealing with AI and can ask for a human","Adverse impact testing before launch and at regular intervals, with results kept","Security basics on the candidate data store (unique credentials, multifactor authentication, access logs)"],"humanInTheLoop":"Recruiters and hiring managers make every screening, interview and hiring decision and can see and override any AI assessment. HR owns the criteria, reviews adverse impact results with legal, and handles every candidate request for human review. A named owner is accountable for the system's oversight, logs and incidents.","kpisToInstrument":["Time from application to interview and to offer","Candidate drop off rate by stage and channel","Selection rates by group at each stage (adverse impact ratio)","Share of AI assessments that recruiters override, and why","Candidate requests for human review and their outcomes"],"failureModes":[{"title":"Bias learned from history","detail":"A model trained on past hires reproduces who was hired before. Amazon scrapped an experimental CV ranking tool that, according to press reports collected in the AI Incident Database, showed bias against women. Use documented criteria, not past decisions, and test outcomes."},{"title":"Automation bias in review","detail":"Recruiters accept the ranking without reading. Show evidence rather than scores, measure overrides and train reviewers."},{"title":"Screening creep","detail":"A scheduling bot starts filtering on free text answers without a risk assessment. Treat any new decision logic as a change that needs review."},{"title":"Candidate data breach","detail":"A hiring chatbot platform exposes applicants' chats and contact details through weak security. Apply the same security controls as any system holding personal data at scale."}]},"risk":{"euAiAct":{"tier":"high","basis":"Annex III point 4(a) lists AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications and to evaluate candidates. Screening, ranking and scoring applications is therefore high risk. A component limited to a narrow procedural task, such as booking interview slots or answering process questions, can fall outside the high risk category under Article 6(3), but only if it does not materially influence the outcome and does not profile people, and that assessment must be documented (Article 6(4)). Deployers of the high risk part must follow the instructions for use, assign competent human oversight, keep logs, inform workers' representatives and inform candidates that a high risk system is used (Article 26). An organization that builds its own screening system becomes its provider, with conformity assessment duties. The chatbot part also carries the Article 50 disclosure duty."},"regulations":["eu-ai-act","gdpr","uk-gdpr","iso-42001","nist-ai-rmf","nyc-local-law-144"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(a) covers recruitment and selection, including analysing and filtering applications and evaluating candidates."},{"title":"Article 26, obligations of deployers of high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/26/","note":"Human oversight, logs, informing workers' representatives and informing the people affected. The AI Act Explorer lists 2 December 2027 as the application date for Annex III systems."},{"title":"Article 5, prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Point 1(f) prohibits AI systems that infer the emotions of a person in the areas of workplace and education institutions, except for medical or safety reasons. The European Commission's guidelines read \"workplace\" as including candidates in the selection and hiring process."},{"title":"Guidelines on prohibited artificial intelligence practices established by the AI Act","issuer":"European Commission","region":"europe","url":"https://digital-strategy.ec.europa.eu/en/library/commission-publishes-guidelines-prohibited-artificial-intelligence-ai-practices-defined-ai-act","note":"The guidelines say the notion of workplace in Article 5(1)(f) also applies to candidates during selection and hiring, and give the use of emotion recognition during the recruitment process as an example of a prohibited practice. The guidelines are not binding; the Court of Justice gives the authoritative interpretation."},{"title":"ICO intervention into AI recruitment tools leads to better data protection for job seekers","issuer":"UK Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2024/11/ico-intervention-into-ai-recruitment-tools-leads-to-better-data-protection-for-job-seekers/","note":"Audits of AI recruitment tool providers led to almost 300 recommendations, after some tools allowed filtering by protected characteristics or inferred gender and ethnicity from names."},{"title":"Automated employment decision tools (Local Law 144)","issuer":"New York City Department of Consumer and Worker Protection","region":"north-america","url":"https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page","note":"Employers and employment agencies in New York City may use an automated employment decision tool only if it had a bias audit within one year of its use, a summary of the results is public and candidates or employees received notices. The independence requirement for the auditor is set out in the linked DCWP rule."}],"controls":["AI Act classification of each component, with the Article 6(3) assessment documented where a derogation is used","Data protection impact assessment and a lawful basis for any automated assessment, with GDPR Article 22 safeguards","Bias audit before launch and at least yearly, with results retained","Logging of assessments, human decisions and overrides for at least the legally required period","Candidate notices, a human review route and a named owner for oversight and incidents","Supplier due diligence covering security testing, model documentation and instructions for use"],"incidents":[{"title":"Amazon's experimental hiring tool allegedly displayed gender bias in candidate rankings","url":"https://incidentdatabase.ai/cite/37/","note":"AI Incident Database entry 37. Amazon scrapped an experimental tool for ranking CVs that, according to the press reports collected there, showed bias against women."},{"title":"McDonald's AI hiring bot exposed millions of applicants' data to hackers who tried the password 123456","url":"https://www.wired.com/story/mcdonalds-ai-hiring-chat-bot-paradoxai/","note":"Security researchers accessed the back end of the McHire chatbot platform run by Paradox.ai through the guessable password 123456 on an old test administrator account that appeared to lack multifactor authentication, then changed applicant ID numbers to see other applicants' chats and contact details. Paradox confirmed the findings and said no third party other than the researchers accessed the account."}]},"blitsAi":{"howToBuild":"On Blits.ai the candidate facing part is an **AI agent** on the **web chat**, **WhatsApp** and\n**SMS** channels, answering from a **knowledge base** of approved role and process content, in the\ncandidate's language with **multi language** support, and disclosing that it is an AI. A **flow**\ncollects the required application fields with **slot filling** and **sensitive data flags**, and\n**custom functions** write them to the applicant tracking system and book interview slots in the\norganization's calendar system (the integration catalog includes Workday and Google Workspace,\nand Office 365 is available as a ready made tool).\n\nIf the organization chooses to add an assessment, it runs as an **agentic workflow** that produces\na **structured output** summary of matching and missing experience against documented criteria,\nwith **human in the loop** approval so no status changes without a recruiter, and a full **audit\ntrail** per run. **Guardrails** block questions about protected characteristics, **PII masking**\nand the **GDPR toolkit** handle consent, retention and removal requests, and **test suites** can\nreplay matched candidate profiles to check for inconsistent treatment. The organization remains\nthe deployer, and the provider of any screening logic it builds, under the AI Act."},"faq":[{"question":"Is AI CV screening high risk under the EU AI Act?","answer":"Yes. Annex III point 4(a) lists AI used to analyse and filter job applications and to evaluate candidates. That brings risk management, data governance, logging, human oversight and, for deployers, duties to inform workers' representatives and candidates. Interview scheduling alone can fall outside the high risk category if it is a narrow procedural task that does not influence who is selected, but that assessment must be documented."},{"question":"What results do employers report?","answer":"Mostly speed. Microsoft reports that Gojob's assistant prequalifies a candidate within 15 minutes and has held 1.5 million exchanges; Chipotle began rolling out a conversational hiring assistant to more than 3,500 restaurants in 2024 and expects time to hire to fall by as much as 75%, but its announcement reports no measured result. Public evidence on fairness outcomes is scarce, which is a reason to measure it yourself."},{"question":"Can the AI reject candidates automatically?","answer":"It should not. Beyond the AI Act, GDPR Article 22 restricts decisions based solely on automated processing with significant effects, and Recital 71 names online recruiting without any human intervention as an example. Keep objective knockout rules transparent, tell candidates how to ask for a human review and let a person make the decision."},{"question":"What does a government deployment look like?","answer":"US Immigration and Customs Enforcement uses a GPT-4 based tool that scores resumes against job requirements and shows related and missing experience to HR specialists. The department classifies it as high impact and lists its impact assessment, monitoring and appeal process as still in progress, which shows how much governance work sits behind such a tool."}],"related":["employee-onboarding-assistant","hr-and-policy-assistant","conversation-roleplay-training","outbound-reminder-and-confirmation-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical with four evidence records and two incidents verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: added NYC Local Law 144 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: replaced the unsourced CV volume in the problem with cited Gojob and Trace3 figures; aligned the ICO, Amazon and McHire wording with their sources; described Chipotle's rollout as phased; removed an unsupported calendar integration claim; added UK GDPR; added seoTitle and metaDescription."},{"date":"2026-09-26","note":"Second fact check: attributed the resume review variability point to the ICE inventory entry; removed an unsourced application volume; corrected the Amazon and McHire descriptions to match their sources; cited GDPR Recital 71 in the FAQ."},{"date":"2026-09-27","note":"Review fixes: repaired a broken sentence in the results FAQ; aligned the NYC Local Law 144 and Article 5 guidance notes with their sources; made the meta description attribute each practice less broadly."},{"date":"2026-09-27","note":"Third fact check: corrected the Article 5 note, since the Commission guidelines apply the workplace emotion recognition ban to job candidates, and added those guidelines as guidance; softened an unsourced claim about candidates lost while waiting. All four evidence records, both incidents and the other guidance re checked with no changes."}],"slug":"recruitment-screening-and-interview-scheduling","url":"https://www.blits.ai/ai-use-cases/recruitment-screening-and-interview-scheduling","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"mastercard-ai-interview-scheduling","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1500000,"min":1500000,"max":1500000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"gojob-aglae-candidate-prequalification","pooled":true}]},{"kpi":"cycle-time-days","label":"Cycle time","unit":"minutes","aggregate":false,"higherIsBetter":false,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"gojob-aglae-candidate-prequalification","pooled":false}]}],"indicativeValueResult":{"low":140000,"high":840000},"evidence":["chipotle-ava-cado-conversational-hiring","gojob-aglae-candidate-prequalification","ice-ai-assisted-resume-screening","mastercard-ai-interview-scheduling","trace3-microsoft-copilot-recruiting-and-meetings"]},{"title":"AI for regulatory report assembly","shortTitle":"Regulatory report assembly","seoTitle":"AI for regulatory report assembly and validation","metaDescription":"AI assembles, validates and reconciles regulatory returns and drafts variance commentary for officer sign off. The Federal Reserve Board and NCUA run similar checks.","definition":"AI that assembles periodic and data driven regulatory filings and returns, such as prudential and statistical returns, threshold and transaction reports and disclosure packs, by pulling data into the regulator's schema, validating it, reconciling figures to source, explaining movements against prior periods and drafting commentary, before a named officer reviews and submits. Narratives for individual suspicious activity cases are a separate use case.","aliases":["AI regulatory reporting","regulatory return preparation","XBRL return automation","goAML report assembly"],"industries":["banking","insurance","capital-markets","payments"],"functions":["regulatory-compliance","finance-and-accounting","financial-crime-compliance"],"patterns":["agentic-workflow","anomaly-detection","content-generation","summarization"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"back-office","problem":"Banks file a steady stream of regulatory reports: prudential returns on capital, liquidity and\nlarge exposures, statistical returns to the central bank, threshold and cross border transaction\nreports to the financial intelligence unit, and public disclosures. Much of the effort goes into\ngathering data from many systems, formatting it into the regulator's schema (the EBA data point\nmodel and XBRL taxonomies for EU prudential returns, the goAML reporting format where the\nfinancial intelligence unit runs UNODC's goAML system), reconciling it to the ledger and\nexplaining why numbers moved, rather than into judgment.\n\nErrors are costly, because a wrong or late filing has to be corrected and explained to the\nauthority that received it. The Basel Committee's BCBS 239 principles ask banks to aggregate risk\ndata on a largely automated basis, reconcile it with source and accounting data, and control and\ndocument any manual processes and spreadsheets they still rely on. Where those manual steps sit\non top of the reporting platform, every period end becomes a scramble. The narrative of a\nsuspicious activity report is a separate job with its own page; this page covers the assembly\nand quality of the filing itself.","problemStats":[],"howItWorks":"1. **Collect the data.** The agent pulls ledger, risk, customer and transaction data for the\n   period from the reporting data warehouse and source systems.\n2. **Map and validate.** It fills the regulator's schema field by field and runs the official\n   validation and business rules, explaining every failure in plain language.\n3. **Reconcile.** It reconciles totals to the general ledger and to related returns, and flags\n   breaks with the likely cause.\n4. **Explain movements.** It compares every material line with prior periods and drafts the\n   variance commentary from the underlying drivers, citing the data behind each statement.\n5. **Review and submit.** A named officer reviews the pack, resolves open points, signs and\n   submits. The system records what was compiled, changed and approved.","valueDrivers":["compliance","employee-productivity","speed","risk-reduction"],"kpis":["processing-time-reduction","productivity-gain","error-reduction","hours-saved","time-saved-per-task"],"indicativeValue":{"referenceOrg":"A bank with a regulatory reporting team of 40 people","inputs":[{"key":"fte","label":"Regulatory reporting staff","low":40,"high":40,"unit":"full time employees","note":"The reference bank. Replace with your own team size."},{"key":"hoursPerFte","label":"Working hours per employee per year","low":1600,"high":1700,"unit":"hours per year","note":"Editorial assumption."},{"key":"assemblyShare","label":"Share of time spent gathering, formatting, reconciling and explaining","low":0.4,"high":0.6,"unit":"fraction of working time","note":"Editorial assumption; replace with your own activity analysis."},{"key":"effortReduction","label":"Share of that work the AI removes","low":0.15,"high":0.35,"unit":"fraction of assembly time","note":"Editorial assumption, deliberately cautious because public evidence with measured results is thin."},{"key":"costPerHour","label":"Fully loaded cost per hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"fte * hoursPerFte * assemblyShare * effortReduction * costPerHour","currency":"USD","period":"per year","resultLabel":"Reporting effort released","caveat":"Labour only. It leaves out fewer resubmissions and supervisory findings, and the cost of the platform, data lineage work and model validation, which are often larger than the model cost."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The formats are well defined, but the data sits in many systems of varying quality, every figure must be traceable, and the reporting officer stays accountable for every number and word.","dataPrerequisites":["Data lineage from source systems to each reported field","The regulator's schema, taxonomy and validation rules for each report","Prior period filings, adjustments and review comments"],"integrations":["Regulatory reporting platform or data warehouse","General ledger, risk engines and customer data","Transaction and payments data for transaction reports","The regulator's submission portal"]},"implementation":{"steps":[{"title":"Choose one report family","detail":"Start with a single return or a high volume transaction report with a stable schema, not the full reporting estate."},{"title":"Trace every field to source","detail":"Document where each field comes from and how it is transformed. Gaps in lineage are the main blocker and are worth fixing regardless of AI."},{"title":"Automate validation and reconciliation first","detail":"Plain language explanations of validation failures and reconciliation breaks save time with the least model risk, because the checks themselves stay deterministic."},{"title":"Add variance commentary with citations","detail":"Draft commentary only from the data, each statement linked to the figures behind it, and measure how much reviewers change."},{"title":"Validate and monitor","detail":"Put any drafting or anomaly model in the model inventory, test it on past periods and monitor edit rates and resubmissions after go live."}],"guardrails":["Nothing is filed without a named officer's review and submission","Every number comes from the source of record; the model never generates figures","Every file passes the regulator's schema and validation rules before review","Commentary cites the data behind each statement and is blocked if it cannot"],"humanInTheLoop":"The reporting officer reviews every return and report, decides on adjustments, and submits. Finance and risk owners confirm variance explanations for their lines, and a second line team samples filed reports each period.","kpisToInstrument":["Days from period end to submission","Validation failures at first run and at submission","Share of commentary text changed by reviewers","Resubmissions and restatements","Manual adjustments outside the reporting platform"],"failureModes":[{"title":"Commentary that asserts more than the data","detail":"The draft explains a movement with a plausible but wrong driver. Require citations and have line owners confirm."},{"title":"Automation bias in review","detail":"Reviewers accept packs because they look complete. Track edit rates and seed known errors in quality checks."},{"title":"Silent data drift","detail":"A source system change alters a field's meaning. Keep lineage and reconciliation checks in every run."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Not an Article 5 practice and not listed in Annex III: the system prepares filings for authorities and makes no decision on the credit, insurance, employment or access to services of a natural person. It is an internal tool whose users know they are working with AI, and drafted text that ends up in public disclosures passes human review under a named person's editorial responsibility, which takes it outside the Article 50(4) deployer disclosure duty. The system still drafts variance commentary and plain language explanations of validation failures from underlying data, rather than lightly editing existing text, so the assistive function for standard editing exception does not fit. The bank that builds or operates the system is then the provider and carries the Article 50(2) duty to mark that generated text in a machine readable way as artificially generated, which has applied since 2 August 2026. The AI literacy duty of Article 4 also applies."},"regulations":["eu-ai-act","dora","fatf-recommendations","us-sr-11-7","apra-cps-230","us-bsa","solvency-ii"],"guidance":[{"title":"Principles for effective risk data aggregation and risk reporting (BCBS 239)","issuer":"Basel Committee on Banking Supervision","region":"global","url":"https://www.bis.org/publ/bcbs239.htm","note":"Written for group risk reporting, and the Committee notes banks may also apply it to supervisory reporting. Expects accurate, complete and timely risk data, aggregated on a largely automated basis and reconciled with source and accounting data."},{"title":"Reporting frameworks","issuer":"European Banking Authority","region":"europe","url":"https://www.eba.europa.eu/risk-and-data-analysis/reporting-frameworks","note":"The EU supervisory reporting taxonomies and validation rules that returns must pass."},{"title":"goAML","issuer":"United Nations Office on Drugs and Crime","region":"global","url":"https://www.unodc.org/unodc/en/global-it-products/goaml.html","note":"UNODC software built for financial intelligence units to receive, process and analyse the reports financial institutions file. Where a unit runs goAML, its reporting format is the target for the transaction reports a bank assembles."}],"controls":["Named officer sign off recorded for every submission","Full record of the data compiled, the draft, the edits and the approval","Model inventory entry and validation for any drafting or anomaly model","Reconciliation of every return to the ledger kept as evidence"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that runs at period end or on a schedule. **Custom\nfunctions** and **SQL knowledge bases** pull the reporting data, deterministic checks run as\ncustom code, and an **AI agent** with **structured output** explains validation failures and\ndrafts variance commentary. Reporting instructions and prior review comments sit in the\n**knowledge base** with hybrid retrieval.\n\n**Human in the loop approval** holds the pack for the reporting officer; the platform does not\nsubmit on its own. Each run keeps a **full audit trail**, **prompt versioning** records how the\ndrafting instructions changed, and **test suites** grade commentary against past periods. The\nplatform is model agnostic and runs in EU or UAE regions where data must stay local."},"faq":[{"question":"Can AI file regulatory reports on its own?","answer":"No. A named officer reviews and submits every report. AI gathers and maps the data, explains validation failures and drafts variance commentary, but the numbers come from the systems of record and accountability stays with the reporting officer."},{"question":"How is this different from drafting suspicious activity reports?","answer":"Suspicious activity report drafting is about writing the investigation narrative, which has its own page. This use case covers assembling, validating and explaining the filing and the prudential and statistical returns around it."},{"question":"Who uses machine learning on regulatory report data today?","answer":"Supervisors do. The US National Credit Union Administration uses machine learning to list potential outliers in each credit union's Call Report data, and the Federal Reserve Board gives its analysts model predicted values to compare with what each firm reported. Banks can run the same kind of checks before they submit. Public, measured results from banks using AI for their own returns are still rare."}],"related":["supervisory-exam-response-assembly","ledger-and-payment-reconciliation","governed-text-to-sql-analytics","suspicious-activity-report-drafting","regulatory-horizon-scanning"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added Bank Secrecy Act, Solvency II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed the unsourced goAML XML claim, tied the manual process and reconciliation points to BCBS 239 and narrowed its guidance note to what the principles cover, corrected the EU AI Act basis (Article 4 and Article 50(4) instead of 'no natural person interacts'), aligned the Federal Reserve Board FAQ wording and both evidence summaries with the inventory entries, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Added the Article 50(2) provider marking duty to the EU AI Act basis and moved the tier to limited; softened the unsourced 'most of the effort' claim to 'much of the effort' to match the assemblyShare range."}],"slug":"regulatory-report-assembly","url":"https://www.blits.ai/ai-use-cases/regulatory-report-assembly","benchmarks":[],"indicativeValueResult":{"low":230400,"high":1428000},"evidence":["federal-reserve-board-regulatory-data-analysis","ncua-call-report-machine-learning-validation"]},{"title":"AI for RFP, tender and sales proposal response drafting","shortTitle":"RFP and proposal drafting","seoTitle":"AI for RFP responses and proposal writing","metaDescription":"AI finds approved answers and drafts RFP and proposal responses for bid teams to review. Responsive says Microsoft sellers used AI answers 200,000+ times.","definition":"AI that helps sales and bid teams answer requests for proposal, tenders, security questionnaires and sales proposals: it breaks the request into questions and requirements, retrieves approved answers and past proposals, drafts the response and a compliance matrix, and routes open points to subject matter experts, with a proposal manager reviewing everything before submission.","aliases":["AI RFP response software","AI proposal writing","AI bid writing","tender response automation","security questionnaire automation","RFx response management"],"industries":["cross-industry","professional-services","technology"],"functions":["sales","knowledge-management"],"patterns":["content-generation","rag-knowledge-assistant","document-processing"],"channels":["internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"For many business to business sellers, the formal response is the sale: a public tender, a\nrequest for proposal with hundreds of questions, a due diligence or security questionnaire from a\ncustomer's procurement team. Each one arrives with a deadline, a mandatory format and questions\nthat have mostly been answered before, somewhere, by someone. Proposal teams spend their time\nfinding the latest approved answer, chasing experts for the rest and reformatting, and sellers\nwho are not bid professionals struggle to write a structured proposal at all.\n\nThe volume keeps rising. Loopio's 2026 benchmark of more than 1,500 companies reports that\nresponse teams now submit an average of 166 responses a year. The work does not scale by adding\nwriters, and a wrong or outdated answer can become a contractual commitment.","problemStats":[{"statement":"Loopio's 2026 RFP trends report, based on more than 1,500 companies, finds that response teams submit an average of 166 RFP responses a year.","sourceTitle":"RFP Report: 2026 Trends & Benchmarks","sourceUrl":"https://loopio.com/trends-report/","year":2026}],"howItWorks":"1. **Read the request.** The AI parses the RFP, tender or questionnaire (PDF, Word, spreadsheet\n   or portal export) into individual questions, mandatory requirements, evaluation criteria and\n   deadlines, and builds a compliance matrix.\n2. **Find approved answers.** For each question it retrieves the best matching answers from a\n   curated answer library and past winning proposals, with the owner and the date each answer was\n   last approved.\n3. **Draft the response.** It drafts answers tailored to the buyer's context and wording, marks\n   which parts come from approved content and which are new, and flags questions it cannot\n   answer with confidence.\n4. **Route to experts.** Security, legal, pricing and technical questions without an approved\n   answer go to the right subject matter expert, and their approved answers flow back into the\n   library.\n5. **Review and submit.** The proposal manager reviews the whole response, checks commitments\n   and pricing, and submits it; the final version and the outcome are stored for the next bid.","valueDrivers":["employee-productivity","revenue-growth","speed","compliance"],"kpis":["processing-time-reduction","handling-time-reduction","time-saved-per-task","hours-saved","users-served","interactions-handled","cost-savings"],"indicativeValue":{"referenceOrg":"A business to business company that submits 150 RFP and questionnaire responses a year","inputs":[{"key":"responses","label":"Responses submitted per year","low":100,"high":200,"unit":"responses per year","note":"Around the Loopio 2026 benchmark of 166 responses a year cited on this page. Replace with your own volume."},{"key":"hoursPerResponse","label":"Team hours per response","low":25,"high":40,"unit":"hours per response","note":"Editorial assumption covering writers, reviewers and subject matter experts. Replace with a time study of your own bids."},{"key":"timeReduction","label":"Share of response time saved","low":0.3,"high":0.6,"unit":"fraction of hours","note":"Conservative against the evidence on this page (GroupeActive reports 75% less drafting time in a pilot, and Microsoft reports in its customer story that Copilot cut ICG's proposal response time by 80%), because both are small firms and review time does not shrink as much."},{"key":"hourlyCost","label":"Fully loaded cost per hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption for a blended proposal, sales and expert team."}],"formula":"responses * hoursPerResponse * timeReduction * hourlyCost","currency":"USD","period":"per year","resultLabel":"Proposal team time released","caveat":"Time value only. It leaves out the cost of building and curating the answer library, the software, and the upside that matters most: more bids answered and a higher win rate, which none of the organizations on this page has quantified."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"The technology is retrieval and drafting over documents the organization already owns. The real work is curating an answer library with owners and review dates, and agreeing who approves security, legal and pricing answers.","dataPrerequisites":["A curated library of approved answers, with an owner and a review date per answer","Past proposals and their outcomes, cleaned of client confidential details where needed","Current product, security, certification and company fact sheets","A list of statements that need legal or pricing approval every time"],"integrations":["Document storage such as SharePoint or Google Drive","CRM for the opportunity, the account and the outcome","Collaboration tools such as Microsoft Teams for expert routing","Procurement portals or email for receiving and submitting responses"]},"implementation":{"steps":[{"title":"Curate the answer library before the model","detail":"Harvest answers from the last two years of bids, deduplicate them, and give every answer an owner and a review date. Microsoft's library was built as a curated, verified source so the AI output could be trusted."},{"title":"Start with questionnaires","detail":"Security, due diligence and vendor questionnaires are repetitive and scored on accuracy, so they show value fastest. Move to narrative proposals once the library is trusted."},{"title":"Show the source of every answer","detail":"Mark each drafted answer as approved content, adapted content or new text, with a link to its source, so reviewers spend their time on what is new."},{"title":"Route the gaps to experts","detail":"Send unanswered or low confidence questions to named experts with a deadline, and feed their approved answers back into the library so the next bid starts further ahead."},{"title":"Close the loop with outcomes","detail":"Store the submitted version with the win or loss and the buyer's feedback, and retire answers that keep losing or keep being rewritten."}],"guardrails":["Answers drafted only from the approved library and named source documents, with citations","Human approval for every statement about security controls, certifications, pricing, liability and service levels","No client confidential information from past bids reused in a proposal for another client","Disclosure of AI use when a buyer asks for it, as UK central government buyers may under PPN 017"],"humanInTheLoop":"The proposal manager owns the response and reviews every answer before submission. Subject matter experts approve new or changed answers in their area, and legal and pricing sign off commitments. The AI never submits a response.","kpisToInstrument":["Hours per response and elapsed days from receipt to submission","Share of answers drafted from approved content without edits","Number of bids answered per quarter and bids declined for lack of capacity","Win rate on comparable bids, before and after","Answers flagged as outdated or wrong in review"],"failureModes":[{"title":"Outdated answers","detail":"The AI confidently reuses a security or certification answer that is no longer true, and it becomes a contractual commitment. Give answers review dates and let the draft show them."},{"title":"Generic proposals","detail":"Drafts read the same for every buyer and lose on evaluation criteria. Make the buyer's own requirements and scoring the structure of the draft."},{"title":"Confidentiality leaks between clients","detail":"Content from one client's proposal ends up in another's. Separate client specific material from reusable answers in the library."},{"title":"Nobody curates the library","detail":"Without owners, the library fills with duplicates and the AI surfaces the wrong version. Budget time for knowledge managers, as Microsoft's proposal team does according to Responsive."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Drafting bid responses for staff to review is not listed in Annex III, and the buyer receives the seller's own document rather than interacting with an AI system, so the high risk tier and the Article 50(1) duty towards the buyer do not apply. Staff who chat with the agent must know it is an AI system, which an internal tool labelled as an AI assistant meets by design. Article 50(2) does apply to the drafting itself: the provider of a system that generates text must mark its output in a machine readable format as artificially generated, whether or not a person reviews the draft, unless the system only performs an assistive function for standard editing. A seller that uses a third party drafting tool relies on that tool's provider for the marking; a seller that builds its own agent that generates proposal text, as GroupeActive did with Witivio on Copilot Studio, can be the provider and then carries the duty itself. AI literacy under Article 4 applies in both cases, and the seller remains responsible for every statement in the submitted response."},"regulations":["eu-ai-act","gdpr","iso-42001"],"guidance":[{"title":"Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Official text on EUR-Lex. Article 50(2) requires providers of AI systems that generate text to mark the output as artificially generated in a machine readable format, with an exception for systems that only perform an assistive function for standard editing. Article 50(1) covers systems that interact directly with people, here the staff who use the agent."},{"title":"PPN 017: Improving transparency of AI use in procurement","issuer":"UK Cabinet Office","region":"europe","url":"https://www.gov.uk/government/publications/ppn-017-improving-transparency-of-ai-use-in-procurement","note":"Published 17 February 2025. Does not prohibit suppliers from using AI to write bids, but lets UK central government departments, their executive agencies and non departmental public bodies ask suppliers to disclose it and do extra due diligence, because AI can introduce misleading statements through hallucination. Other public sector buyers may choose to apply it. For procurements commenced, or contracts awarded, before 24 February 2025, the earlier PPN 02/24 applies."}],"controls":["An answer library with an owner, an approval status and a review date per answer","Mandatory human approval of security, legal, pricing and service level statements","Access controls that keep client confidential bid content out of other clients' drafts","A record of which answers were AI drafted and who approved them, per submitted bid"],"incidents":[{"title":"Incident 1193: Purportedly Taxpayer-Funded Deloitte Report for Australian Government Contains Alleged AI-Generated Citations and Fabricated Legal Quote","url":"https://incidentdatabase.ai/cite/1193/","note":"A consultancy report for Australia's Department of Employment and Workplace Relations contained nonexistent academic references and a misquoted court judgment; the firm acknowledged using generative AI, issued a corrected version and partially refunded the fee. It was a client deliverable rather than a bid, but it shows what unchecked AI drafted content costs when it reaches a client under the firm's name."}]},"blitsAi":{"howToBuild":"On Blits.ai the answer library, past proposals and fact sheets go into a **knowledge base**\n(PDF, DOCX, PPTX, XLSX and crawled web pages, with version control) searched with **hybrid\nretrieval**. An **agentic workflow**, triggered by API or from the team's chat, reads the\nincoming RFP, splits it into questions with **structured output**, drafts each answer with\ncitations to its sources, and uses the **file generation** tool to return the draft and a\ncompliance matrix.\n\nA **human in the loop** confirmation on the workflow action that returns or sends the draft lets\na reviewer approve or reject it first, and the draft can be reviewed with the team in\n**Microsoft Teams**. **Custom functions** read the opportunity from the CRM (Salesforce is in the\nintegration catalog, and HubSpot and Dynamics 365 are ready made tools) and write the outcome\nback. An **output guardrail** with an admin written policy checks answers against the approved\ncontent, **test suites** check the agent against known questions with approved answers, and the platform\nis model agnostic, with EU and UAE data residency for bid content that must stay in region."},"faq":[{"question":"How much time does AI save on RFP responses?","answer":"The public figures come mostly from small teams: GroupeActive reports that, in a pilot, its members save an average of 75% of drafting time on sales proposals, and Microsoft reports in its customer story that Copilot cut ICG's proposal response time by 80%. At scale, Responsive reports that Microsoft's sellers save 20 minutes per search for proposal content. Expect less on complex bids, where expert review dominates."},{"question":"Can AI write a whole tender response on its own?","answer":"It can draft most of it from approved content, but a proposal manager should review everything, and security, legal and pricing statements need an expert's approval because they become commitments. Some public buyers may ask suppliers to disclose AI use in their bids: UK central government guidance (PPN 017) gives buyers example disclosure questions for this."},{"question":"What matters more, the AI tool or the content library?","answer":"The library. AI retrieval and drafting are only as good as the approved answers behind them, which is why, as Responsive describes it, Microsoft relies on the proposal team's knowledge managers and on technical experts across the company to keep more than 18,000 question and answer pairs current."}],"related":["business-connectivity-quoting-and-service-assistant","enterprise-knowledge-search"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with evidence from Microsoft, GroupeActive, ICG and Verdantas checked against the sources. Editor pass cited PPN 017 as the current UK guidance and attributed vendor figures to the vendor."},{"date":"2026-09-27","note":"Second editor pass: attributed the Microsoft figure in the meta description to Responsive, set the EU AI Act tier to limited with the Article 50(2) marking duty and cited EUR-Lex, labelled the GroupeActive figure as a pilot, aligned the Blits.ai section with the feature inventory, corrected who curates Microsoft's library, dropped the unsupported email channel and added AI Incident Database entry 1193."}],"slug":"rfp-and-proposal-response-drafting","url":"https://www.blits.ai/ai-use-cases/rfp-and-proposal-response-drafting","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":80,"min":80,"max":80,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"icg-copilot-proposal-drafting","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":75,"min":75,"max":75,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"groupeactive-gaia-propale-proposals","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":93000,"min":93000,"max":93000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"microsoft-responsive-proposal-resource-library","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":200000,"min":200000,"max":200000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"microsoft-responsive-proposal-resource-library","pooled":true}]},{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":20,"min":20,"max":20,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"microsoft-responsive-proposal-resource-library","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":18000,"min":18000,"max":18000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"microsoft-responsive-proposal-resource-library","pooled":true}]}],"indicativeValueResult":{"low":45000,"high":480000},"evidence":["groupeactive-gaia-propale-proposals","icg-copilot-proposal-drafting","microsoft-responsive-proposal-resource-library","verdantas-copilot-studio-proposal-agent"]},{"title":"AI for risk based inspection prioritization in food safety, workplace and environmental regulation","shortTitle":"Inspection prioritization","seoTitle":"AI risk scoring for inspection prioritization","metaDescription":"Regulators such as DVSA, the Dutch NVWA and the Netherlands Labour Authority use risk models to pick which sites to inspect first; staff still choose the visits.","definition":"Models that predict which premises, operators or activities are most likely to be non compliant, so that inspectors in food safety, workplace safety, environmental and other regulation spend their visits where the risk is highest, ideally with inspectors choosing the visits and random inspections testing the model.","aliases":["risk based inspection targeting","inspection targeting AI","predictive inspections","regulatory inspection risk scoring"],"industries":["government"],"functions":["risk-management","case-management","regulatory-compliance"],"patterns":["prediction-and-scoring","anomaly-detection"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","problem":"Regulators and local authorities oversee far more restaurants, farms, workplaces, garages, care\nhomes and industrial sites than their inspectors can visit. Visits are traditionally scheduled by\nfixed frequencies, time since the last visit, the type of premises and complaints received. That\nspends scarce inspector time on operators that are almost always compliant, lets backlogs build\nup when new businesses register faster than they can be visited, and finds problems late at\noperators whose risk has changed since their last rating.\n\nRegulators hold useful signals: past inspection results, notifications, complaints, whistleblowing,\nregistration data and, for some sectors, detailed transaction data such as MOT test records. Using\nthem to rank the next visits is the promise of predictive targeting. The risk is that a model\nlearns from past inspection choices, keeps sending inspectors to the same kind of operator and\nlabels businesses before anyone has looked.","problemStats":[],"howItWorks":"1. **Pick the decision.** The model supports one choice, such as which new food businesses to\n   inspect first or which notified asbestos removals to visit, within an existing inspection\n   programme.\n2. **Learn from outcomes.** A supervised model (such as gradient boosting or a random forest, or\n   the best of several techniques on a held out test set) is trained on past inspection results,\n   or an outlier model flags operators whose behaviour deviates from peers when labelled outcomes\n   are scarce. Some regulators add a rules layer, for example on the age of the last rating.\n3. **Score the population.** Every operator in scope gets a risk score or a red, amber or green\n   rating, refreshed monthly or when new data arrives, including operators never inspected before.\n4. **Inspectors decide.** Officers see the score alongside other information and local knowledge\n   and choose the visits; the score is never a finding and never replaces the inspection.\n5. **Keep a random sample and feed back.** A share of visits stays random or complaint driven, and\n   their results measure whether the model finds more non compliance than the old approach and\n   whether it keeps skipping certain operators.","valueDrivers":["risk-reduction","employee-productivity","compliance"],"kpis":["detection-rate-improvement","users-served","interactions-handled","accuracy"],"indicativeValue":{"referenceOrg":"A national or regional inspectorate that carries out 20,000 inspections a year","inputs":[{"key":"inspections","label":"Inspections per year","low":15000,"high":25000,"unit":"inspections per year","note":"Editorial assumption. Replace with your own programme."},{"key":"redirectedShare","label":"Share of visits moved from low risk to higher risk operators","low":0.05,"high":0.15,"unit":"fraction of inspections","note":"Editorial assumption. Most programmes keep fixed frequency, complaint and random visits, so only part of the programme can be redirected."},{"key":"costPerInspection","label":"Fully loaded cost of one inspection","low":300,"high":600,"unit":"EUR per inspection","note":"Editorial assumption covering preparation, travel, visit and reporting time. Replace with your own cost."}],"formula":"inspections * redirectedShare * costPerInspection","currency":"EUR","period":"per year","resultLabel":"Inspection capacity redirected to higher risk operators","caveat":"This values the inspection capacity that is redirected, not the public health, safety or environmental harm avoided, which is the real benefit but is rarely measured. It leaves out the cost of building, validating and monitoring the model."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Regulators usually hold the inspection history, so the data for a first model is often at hand. The effort goes into data quality across local authorities or regions, a random inspection programme to evaluate against, and guidance that stops the score from being treated as a verdict.","dataPrerequisites":["Inspection history with outcomes, including inspections that found no breach","A register of the operators in scope, including new registrations","Notifications, complaints and other signals with a documented legal basis"],"integrations":["Inspection case management or regulatory platform","Operator register and notification systems","Dashboard, map or list for inspection planners"]},"implementation":{"steps":[{"title":"Start with a backlog or a clearly bounded programme","detail":"The Food Standards Agency piloted its model on food businesses awaiting their first inspection (the transparency record is now marked retired); the Netherlands Labour Authority scores notified asbestos removals. A bounded population makes the benefit measurable."},{"title":"Keep random and complaint driven inspections","detail":"Reserve part of the programme for random visits. The NVWA uses random inspections to test whether the model finds more problems than it would otherwise, and the Netherlands Labour Authority uses the results of its random inspections to improve its model."},{"title":"Show the score next to other information","detail":"Present the rating in the tools inspectors already use, with the main drivers. DVSA shows its monthly red, amber and green ratings in a Power BI app next to other data on each site, and the Care Quality Commission shows its risk category and the main data drivers on its regulatory platform."},{"title":"Write usage guidance","detail":"State what the score may and may not be used for: prioritizing visits, not judging an operator or deciding enforcement."},{"title":"Monitor coverage and drift","detail":"Compare the operators the model selects with the whole population every cycle, and retrain when the sector or the data changes."}],"guardrails":["The score only prioritizes visits; findings and enforcement rest on evidence from the inspection","Inspectors can override the ranking with local knowledge, and overrides are recorded","A random inspection sample runs alongside model selection","Scores stay internal and are not published or shared with the operator, so a score is never read as a finding about that operator","Personal data, such as businesses run from home addresses, is minimised and protected"],"humanInTheLoop":"Inspection planners and inspectors decide which visits to make and carry out every inspection. Regulators review the model's selections against the population and the random sample each cycle and can pause it when it stops predicting or keeps selecting the same type of operator.","kpisToInstrument":["Non compliance rate found in model selected visits versus random visits","Share of the operator population never selected over a full cycle","Backlog of operators awaiting a first inspection","Share of scores overridden by inspectors, with reasons","Time spent on building visit lists"],"failureModes":[{"title":"Self confirming targeting","detail":"The model learns from where inspectors went before and keeps sending them back. Random visits and a check of population coverage break the loop."},{"title":"Prejudging the operator","detail":"An inspector who sees a red rating may look harder or rate lower. Guidance and, where possible, keeping the score out of the inspection itself reduce this bias."},{"title":"Automation bias or distrust","detail":"Inspectors either follow the ranking blindly or ignore it. Training, visible drivers and feedback on outcomes keep use balanced."},{"title":"Stale data for new operators","detail":"New businesses have no history, so predictions lean on area or type data that can encode socioeconomic bias. Test performance for this group separately."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Prioritizing inspections of businesses and premises is not a use listed in Annex III, so such a system is usually not high risk. The assessment changes when it scores natural persons, such as individual licensed professionals or sole traders, and the inspectorate acts as a law enforcement authority: assessing the risk that a person offends, or profiling persons in the detection or investigation of criminal offences, is high risk under Annex III point 6 (d) and (e), and predicting that a person will commit a criminal offence based solely on profiling is prohibited by Article 5(1)(d). GDPR applies wherever sole traders, home based businesses or named professionals are scored."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Algorithmic Transparency Recording Standard hub","issuer":"UK government","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"The Food Standards Agency, DVSA and the Care Quality Commission publish transparency records for their inspection prioritization tools."},{"title":"Algoritmeregister van de Nederlandse overheid","issuer":"Government of the Netherlands","region":"europe","url":"https://algoritmes.overheid.nl/nl","note":"Dutch inspectorates, including the NVWA and the Netherlands Labour Authority, register their risk models with purpose, method and human oversight."}],"controls":["Transparency record per model, published before use","Written usage guidance for inspectors and planners","Random inspection sample and yearly effectiveness review","Fairness and coverage checks across operator types and areas","Change control and revalidation when the model is retrained"],"incidents":[]},"blitsAi":{"howToBuild":"The risk model itself usually lives in the regulator's analytics environment. Blits.ai adds the\ninspector facing layer: an **AI agent** in **Microsoft Teams** or an internal web chat that answers\n\"why is this site rated red?\" from the model drivers and the site's history through **custom\nfunctions** and a **SQL knowledge base**, and a **knowledge base** with hybrid retrieval over the\ninspection manual and guidance.\n\nAn **agentic workflow** can prepare a visit pack (history, open issues, recent notifications)\nfor each site on the list, with **human in the loop approval** before anything goes to the\noperator. **Monitors** check the agent's answers on a schedule, **PII masking** protects personal\ndata about sole traders, and EU and UAE data residency keeps regulatory data in region."},"faq":[{"question":"Which regulators use AI to choose inspections?","answer":"Models in production on this page include DVSA's outlier model for MOT testing stations, the Netherlands Labour Authority's random forest for asbestos removal jobs, the Dutch NVWA's supervised machine learning model for pig farm welfare and the US EPA's classical machine learning model for hazardous waste generators. The UK Food Standards Agency piloted a LightGBM model with local authorities for food hygiene inspections (the transparency record is now marked retired). The Care Quality Commission combines its sector risk model scores with rating rules in a rules based risk categorisation, and says its machine learning model to prioritize care home inspections is still in development."},{"question":"Does predictive targeting replace routine inspections?","answer":"Not in the UK and Dutch cases on this page: there the score only informs which visits or reviews come first, and Care Quality Commission teams review the category and may then make a site visit or another form of review. The Netherlands Labour Authority, for example, in principle still inspects every certified asbestos removal company at least once every three years, and also follows up reports and selects some jobs at random. The US EPA inventory entry says only that its scores support inspections, classifies the use as high impact and does not describe how staff use the scores."},{"question":"Is it high risk under the EU AI Act?","answer":"Usually not when it targets businesses and premises, because inspection targeting is not listed in Annex III. It needs a closer look when it scores individuals, such as sole traders or licensed professionals, for a regulator that also investigates criminal offences: that can fall under Annex III point 6, and prediction based solely on profiling is banned by Article 5."}],"related":["tax-compliance-risk-scoring","benefit-fraud-and-error-detection","permit-and-licence-application-processing","continuous-controls-testing","governed-text-to-sql-analytics"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version for the government vertical, with UK, Dutch and US regulator evidence verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Netherlands Labour Authority random inspection claim and the three year cycle (in principle), removed logistic regression from the methods, rewrote the EU AI Act basis around Annex III point 6 and Article 5(1)(d), added source dates to the evidence, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: the FAQ now separates production models (DVSA, Netherlands Labour Authority, NVWA, EPA) from the Food Standards Agency pilot and the rules based Care Quality Commission categorisation, limits the \"score only informs\" claim to the UK and Dutch cases, and softens the definition on human selection and random inspections."},{"date":"2026-09-27","note":"Review fixes: the Food Standards Agency record now notes that GOV.UK marks its transparency record retired, with the FAQ and implementation step updated to match."}],"slug":"inspection-prioritization","url":"https://www.blits.ai/ai-use-cases/inspection-prioritization","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":46000,"min":46000,"max":46000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"care-quality-commission-inspection-risk-categorisation","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":150,"min":150,"max":150,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dvsa-mot-garage-risk-rating","pooled":true}]}],"indicativeValueResult":{"low":225000,"high":2250000},"evidence":["care-quality-commission-inspection-risk-categorisation","dvsa-mot-garage-risk-rating","epa-hazardous-waste-generator-inspection-risk-scoring","food-standards-agency-food-hygiene-inspection-prioritisation","nederlandse-arbeidsinspectie-asbestos-removal-risk-model","nvwa-pig-welfare-compliance-model"]},{"title":"AI for sanctions screening alert adjudication","shortTitle":"Sanctions screening adjudication","seoTitle":"AI for sanctions screening alert adjudication","metaDescription":"AI resolves sanctions and watchlist name matches, closing clear false positives and escalating real hits. UOB cut name screening false positives by 60% in a pilot.","definition":"AI that works the alerts raised when customer, counterparty or payment names match sanctions and watchlists: it resolves fuzzy matches across transliterations, aliases and naming conventions, clears clear non matches with a documented reason, and escalates true or uncertain hits with the evidence attached.","aliases":["name screening alert adjudication","sanctions false positive reduction","payment screening alert review","watchlist match resolution"],"industries":["banking","payments"],"functions":["financial-crime-compliance"],"patterns":["classification-and-routing","prediction-and-scoring","agentic-workflow"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"middle-office","problem":"Missing a sanctioned party can bring enforcement action and large fines, so screening engines\nare tuned to match generously. A customer named Mohammed Ali, a ship with a common name or a\ncompany whose address contains a sanctioned city all create alerts, and most of them turn out to\nbe false positives. Each one needs an analyst to compare dates of birth, nationalities,\nidentifiers and context against the list entry.\n\nOn payments the pressure is time. An instant or cross border payment held for a name match\nbreaks the settlement promise to the customer, and a backlog on a busy day means delayed payroll\nor trade payments. On onboarding, screening alerts slow account opening. Meanwhile list updates\nafter a new sanctions package can raise alert volumes sharply overnight.","problemStats":[],"howItWorks":"1. **Parse the record.** Names, dates, addresses, identifiers and free text are extracted from\n   the customer record or the payment message, including structured ISO 20022 fields.\n2. **Resolve the match.** Entity resolution compares the record with the list entry across\n   transliterations, aliases, name order and cultural naming patterns, and weighs secondary\n   identifiers such as date of birth, nationality and registration numbers.\n3. **Score and explain.** A model estimates whether the alert is a true match and lists the\n   factors that support or contradict it, in words an analyst and an auditor can follow.\n4. **Decide under policy.** Alerts that meet approved criteria for a clear non match are closed\n   with that explanation stored. Possible and likely true matches go to an analyst, ranked by\n   risk, with the evidence side by side.\n5. **Assure.** A sample of automated closures is reviewed by a second analyst, and every list\n   update triggers regression tests on known true and false matches.","valueDrivers":["compliance","speed","employee-productivity","customer-experience"],"kpis":["false-positive-reduction","automation-rate","processing-time-reduction","accuracy","productivity-gain"],"indicativeValue":{"referenceOrg":"A bank screening payments and customers with 300,000 name screening alerts a year","inputs":[{"key":"alerts","label":"Name and payment screening alerts per year","low":300000,"high":300000,"unit":"alerts per year","note":"The reference bank."},{"key":"minutesPerAlert","label":"Analyst minutes per alert today","low":3,"high":8,"unit":"minutes per alert","note":"Editorial assumption for level one review. Replace with your own time study."},{"key":"autoClosed","label":"Share of alerts closed automatically as clear non matches","low":0.3,"high":0.6,"unit":"fraction of alerts","note":"Editorial assumption, kept at or below UOB's name screening pilot result on this page (60% fewer false positives on individual name alerts). Replace with results from your own parallel run."},{"key":"costPerHour","label":"Fully loaded analyst cost per hour","low":35,"high":60,"unit":"USD per hour","note":"Editorial assumption. Replace with your own."}],"formula":"alerts * minutesPerAlert / 60 * autoClosed * costPerHour","currency":"USD","period":"per year","resultLabel":"Screening analyst capacity released","caveat":"Counts analyst time only. It leaves out faster payment release and onboarding, lower penalty risk, and the cost of validation, list management and the platform."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The matching problem is well understood, but a missed sanctions hit can lead to regulatory fines, as the Starling Bank case on this page shows. Validation, list management, change control and explainability carry more of the effort than the model.","dataPrerequisites":["Historical screening alerts with final decisions and reasons","Clean customer reference data with secondary identifiers (dates of birth, nationalities, registration numbers)","Current sanctions and watchlists with version history","Payment messages with structured party fields where available"],"integrations":["Screening engine for customers and payments","Payment hub and message queues for held payments","Onboarding and customer data systems","Case management for escalated alerts","List management and watchlist data providers"]},"implementation":{"steps":[{"title":"Tune the engine first","detail":"Before adding AI, fix data quality and fuzzy matching thresholds in the screening engine, and measure alert volume per list and per source. Some false positives are cheaper to prevent than to adjudicate."},{"title":"Build a labelled test set","detail":"Assemble historical alerts with final decisions plus known true matches and synthetic hard cases (aliases, transliterations, partial names) that every model version must pass."},{"title":"Run as a recommender","detail":"Show the model's recommendation and explanation to analysts for a full quarter and measure agreement, especially on the alerts analysts escalated."},{"title":"Automate only clear non matches","detail":"Approve auto closure for the band where secondary identifiers clearly contradict the list entry, and never for alerts where the model is uncertain."},{"title":"Govern list and model changes together","detail":"Re run the regression test set on every list update, model change and threshold change, and keep the results as evidence for auditors."}],"guardrails":["The model may close clear non matches but never a possible or confirmed true match","Every closure stores the explanation, the list version and the model version","Regression tests on known true matches run before every list, model or threshold change","Second analyst sampling of automated closures with a hard stop on errors","Screening coverage monitored so that no customer or payment skips screening when the AI service is down"],"humanInTheLoop":"Analysts decide every possible or likely true match and every payment rejection or asset freeze. A second line samples automated closures, and the sanctions compliance officer approves auto closure criteria and every change to them.","kpisToInstrument":["Alert volume and share closed automatically, per list and source","Error rate in second analyst sampling of automated closures","Time payments are held for screening","Agreement between model recommendation and analyst decision","Regression test pass rate on known true matches"],"failureModes":[{"title":"A true hit closed automatically","detail":"The one failure that matters. Prevent it with conservative auto closure bands, regression tests on known matches and second analyst sampling."},{"title":"List update floods","detail":"A new sanctions package can raise alert volumes sharply overnight, and the model has not seen the new entries. Keep capacity plans and re test on every list update."},{"title":"Unexplainable decisions","detail":"A score without reasons cannot be defended to an examiner. Require the model to state the identifiers that support or contradict the match."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Sanctions screening by banks and payment firms is not listed in Annex III: point 5 covers credit scoring and life and health insurance pricing, and point 6 covers AI used by or on behalf of law enforcement authorities. It is not a prohibited practice under Article 5, and as an internal tool it carries no Article 50 transparency duty. It still processes personal data at scale, so GDPR applies, and decisions that block a payment or freeze assets remain human decisions."},"regulations":["eu-ai-act","gdpr","uk-gdpr","fatf-recommendations","dora","us-sr-11-7","mas-ai-risk-management","cbuae-ai-guidance","nist-ai-rmf","eu-amlr","mas-notice-626"],"guidance":[{"title":"A Framework for OFAC Compliance Commitments","issuer":"US Office of Foreign Assets Control","region":"north-america","url":"https://ofac.treasury.gov/media/16331/download?inline","note":"Lists sanctions screening software or filter faults among the root causes of apparent violations, which is why screening changes need testing and oversight."},{"title":"Principles for Using Artificial Intelligence and Machine Learning in Financial Crime Compliance","issuer":"Wolfsberg Group","region":"global","url":"https://wolfsberg-group.org/resources/202/93","note":"The 2022 principles of a group of global banks for using AI and machine learning in financial crime compliance: legitimate purpose, proportionate use, design and technical expertise, accountability and oversight, and openness and transparency."}],"controls":["Written auto closure criteria approved by the sanctions compliance officer","Stored explanation, list version and model version for every decision","Regression test set of known true matches, run on every change","Second analyst sampling with a hard stop on errors","Model inventory entry with validation and change control"],"incidents":[{"title":"FCA fines Starling Bank for failings in its financial crime systems and controls","url":"https://www.fca.org.uk/news/press-releases/fca-fines-starling-bank-failings-financial-crime-systems-and-controls","note":"The FCA fined Starling Bank about GBP 29 million in 2024. From 2017 its automated screening system had screened customers against only a fraction of the full sanctions list, which shows why screening coverage and list completeness must be tested independently of any model."}]},"blitsAi":{"howToBuild":"The screening engine stays in place. On Blits.ai the adjudication runs as an **agentic\nworkflow**, triggered through the API for each alert: the agent reads the alert and the list\nentry, calls **custom functions** to fetch secondary identifiers from customer and payment\nsystems, and applies the bank's adjudication procedures from a **knowledge base** with hybrid\nretrieval. It returns **structured output** with a recommendation and the identifiers that\nsupport or contradict the match.\n\nClosures and escalations follow **human in the loop approval** with a configurable threshold,\nand every run keeps a full audit trail. **Test suites** hold known true matches and hard cases\nand run before every change, **monitors** check the service on a schedule, and the platform is\nmodel agnostic with EU and UAE data residency."},"faq":[{"question":"Can AI clear sanctions alerts automatically?","answer":"Banks let AI close clear non matches under written criteria, with sampling and regression tests. It should never close a possible or likely true match, because a missed sanctioned party can bring enforcement action and large fines, as the FCA's GBP 29 million fine on Starling Bank for sanctions screening failings shows."},{"question":"What makes sanctions screening alerts so noisy?","answer":"Engines match generously to avoid misses, names are common and transliterated in many ways, and customer records often lack the secondary identifiers that would rule a match out. Better reference data reduces noise before any AI is added."},{"question":"Which banks use AI for screening adjudication?","answer":"Standard Chartered, HSBC and Mashreq have announced screening automation with Silent Eight, and UOB reported a 60 per cent reduction in false positives on individual name screening alerts in a six month pilot with Tookitaki. AJ Bell says it cut customer screening alert volume by 82 per cent with ComplyAdvantage. Most announcements disclose no production results, so insist on your own parallel run."}],"related":["pep-and-adverse-media-screening","aml-alert-triage","trade-finance-crime-screening","payment-investigations-and-exceptions","perpetual-kyc","business-onboarding-and-ubo-discovery"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added EU Anti Money Laundering Regulation, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources. Removed the Ratepay 99% false positive figure because Hawk's AI is only planned there, corrected evidence years and dates (AJ Bell and Ratepay 2025, Standard Chartered 2018-07-09), fixed the Wolfsberg guidance note, replaced unsourced claims (strict liability, \"almost all\" false alerts, \"most heavily penalised\"), expanded the EU AI Act basis, added UK GDPR and added seoTitle and metaDescription."}],"slug":"sanctions-screening-adjudication","url":"https://www.blits.ai/ai-use-cases/sanctions-screening-adjudication","benchmarks":[{"kpi":"false-positive-reduction","label":"False positive reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":60,"min":60,"max":60,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"uob-tookitaki-name-screening-pilot","pooled":true}]}],"indicativeValueResult":{"low":157500,"high":1440000},"evidence":["aj-bell-complyadvantage-customer-screening","fnbo-verafin-agentic-edd-and-sanctions","hsbc-silent-eight-screening-automation","mashreq-silent-eight-alert-adjudication","ratepay-hawk-aml-screening","standard-chartered-silent-eight-screening","uob-tookitaki-name-screening-pilot"]},{"title":"AI for security alert triage and investigation in the SOC","shortTitle":"Security alert triage","seoTitle":"AI SOC alert triage and phishing investigation","metaDescription":"AI triage of SOC alerts and reported phishing. St. Luke's says its agent saves nearly 200 hours monthly; Google Cloud reports 97% less triage time at Human Managed.","definition":"An AI agent in the security operations centre that picks up each new alert or user reported phishing email, gathers the evidence from the SIEM, endpoint, identity and threat intelligence tools, gives a verdict with its reasoning and a draft incident summary, and closes clear false positives while an analyst approves every containment action.","aliases":["AI SOC analyst","agentic SOC","security alert triage agent","phishing triage agent","reported phishing email analysis","AI incident investigation copilot"],"industries":["cross-industry","healthcare","technology","government","professional-services"],"functions":["security-operations","it-and-engineering"],"patterns":["agentic-workflow","classification-and-routing","summarization","rag-knowledge-assistant"],"channels":["internal-tools","microsoft-teams","api"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"A security operations centre lives on alerts: from the SIEM, the endpoint and email protection\ntools, identity systems and cloud platforms, plus the suspicious emails that staff report with a\nbutton in their mail client. Many of these alerts are false positives (St. Luke's University Health\nNetwork describes finding the true threats \"amidst a sea of false positives\"), but each one has to\nbe opened, enriched and judged, because the one real intrusion hides among them.\n\nThe judgment is not the slow part; the legwork is. An analyst checks the sender and the links,\nlooks up the IP address and the file hash, pulls the sign in history of the user, compares the\nevent with earlier incidents and writes up what they found. Before AI, that legwork took up to 20\nminutes per event at SEP2 and up to 30 minutes per alert at Human Managed, and 10 to 77 minutes per\nidentity investigation at Avanade. Most alerts take far less, but at St. Luke's triaging hundreds of\nalerts a day still took hours every day. At that volume the queue wins:\nalerts wait, analysts burn out and real threats are found late.\n\nSecurity teams also struggle with consistency. No two analysts document an investigation the same\nway, which makes handovers between shifts, escalations to forensics and reports to leadership\nslower than they need to be.","problemStats":[],"howItWorks":"1. **Pick up the alert.** The agent is triggered by a new alert in the SIEM or extended detection\n   and response (XDR) console, or by an email a user reported as suspicious.\n2. **Gather the evidence.** It queries the connected tools the way an analyst would: email\n   headers, URLs and attachments, endpoint and firewall logs, identity sign ins, threat\n   intelligence on indicators, and similar past incidents.\n3. **Reason to a verdict.** It classifies the alert (malicious, suspicious, benign, false\n   positive) with a confidence level and writes down the evidence and the reasoning behind it, so\n   an analyst can check the verdict in a minute instead of rebuilding it.\n4. **Act within limits.** Clear false positives, such as a reported marketing email, are closed\n   with the reasoning attached. Anything malicious or uncertain is escalated with a draft incident\n   summary; containment actions such as isolating a device or disabling an account are proposed,\n   and an analyst approves them.\n5. **Learn from feedback.** Analysts correct wrong verdicts in plain language, and the agent's\n   instructions, allow list and examples are updated, so the same mistake is not repeated.","valueDrivers":["employee-productivity","risk-reduction","speed","compliance"],"kpis":["handling-time-reduction","mttr-reduction","productivity-gain","hours-saved","alert-volume-reduction","false-positive-reduction","accuracy"],"indicativeValue":{"referenceOrg":"An enterprise security operations centre that triages 50,000 to 100,000 alerts and reported emails a year","inputs":[{"key":"alertsPerYear","label":"Alerts and reported emails triaged by analysts per year","low":50000,"high":100000,"unit":"alerts per year","note":"Editorial assumption for a large enterprise SOC. Replace with the count from your SIEM or case management system."},{"key":"minutesPerAlert","label":"Average analyst minutes per alert before AI","low":3,"high":10,"unit":"minutes per alert","note":"An average across the alert mix, well below the upper bounds on this page (up to 20 minutes at SEP2 and up to 30 minutes at Human Managed for a single alert), because most alerts are closed quickly. St. Luke's says triaging hundreds of alerts a day took hours, and its nearly 200 hours saved a month across thousands of closed false positives implies a few minutes per reported email. Even the high case (about 10,000 hours a year) is roughly four times St. Luke's 2,400 hours a year from phishing alone, for a SOC that covers every alert type."},{"key":"timeSavedShare","label":"Share of triage time the agent saves","low":0.3,"high":0.6,"unit":"fraction of triage time","note":"Conservative against the evidence on this page, because a first deployment covers only part of the alert types. Google Cloud reports that manual triage at Human Managed was reduced by more than 60%. TÜV SÜD reports analysis about 60% to 70% faster; if faster means more analyses per hour, that is roughly 40% less time per analysis, and if it means 60% to 70% less time, it sits above this range."},{"key":"costPerHour","label":"Fully loaded cost of a SOC analyst hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost or your managed security provider's rate."}],"formula":"alertsPerYear * minutesPerAlert / 60 * timeSavedShare * costPerHour","currency":"USD","period":"per year","resultLabel":"Analyst triage time released, valued at cost","caveat":"Counts analyst time released on triage only. It leaves out the cost of the AI and the security tools it calls, the integration work, the value of finding real incidents sooner and the cost of a wrong verdict, which the guardrails on this page are meant to keep rare."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The model work is modest; the integrations and the trust are the work. The agent needs read access to many security tools, a clear policy on what it may close on its own, and weeks of side by side running before analysts stop double checking every verdict.","dataPrerequisites":["Alerts and reported emails available through the SIEM, XDR or email security API","Read access to logs from endpoints, identity, email, firewalls and cloud platforms","Threat intelligence feeds for indicators such as IP addresses, domains and file hashes","Past incidents with their outcomes, to calibrate verdicts and write examples","Written triage runbooks per alert type"],"integrations":["SIEM and XDR platform (alerts, queries, incident updates)","Email security gateway and the user reporting mailbox","Identity provider for sign in history and account actions","Endpoint detection and response for device context and isolation","Security orchestration (SOAR) or case management for tickets and playbooks","Collaboration tool such as Microsoft Teams for analyst notifications"]},"implementation":{"steps":[{"title":"Start with the noisiest, best understood alert type","detail":"User reported phishing is the usual first choice: high volume, mostly benign, and the evidence (headers, links, attachments, sender history) is well defined. Measure the current volume, handling time and false positive share before you begin."},{"title":"Write the triage runbook down","detail":"Turn the senior analysts' practice into explicit steps and decision rules per alert type, including which evidence decides the verdict. The agent follows this, and it is what you audit against later."},{"title":"Connect read only first","detail":"Give the agent read access to the tools it needs and nothing else. Let it produce verdicts and summaries next to the analysts for several weeks without closing anything."},{"title":"Compare verdicts and calibrate","detail":"Compare the agent's verdicts with the analysts' on the same alerts, per alert type. Only where agreement is consistently high, and the misses are understood, allow the agent to close false positives on its own."},{"title":"Keep humans on containment","detail":"Isolating devices, disabling accounts, blocking senders and deleting emails stay proposed actions that an analyst approves, with the agent's evidence attached, until the organization has a documented reason to automate a specific action."},{"title":"Extend alert by alert","detail":"Add alert types one at a time (identity, endpoint, data loss prevention), each with its own runbook, test set and agreement threshold, and keep sampling closed alerts every week. Alert types that watch individual employees, such as data loss prevention and insider risk, need a fresh legal and works council review first."}],"guardrails":["The agent closes only the alert types and verdicts it has been cleared for; everything else goes to an analyst","Containment and account actions require analyst approval, logged with the evidence behind them","Content from emails, attachments and web pages is treated as untrusted input, so instructions hidden in a phishing email cannot steer the agent","Read only, least privilege service accounts for every connected tool","Every verdict carries its evidence and reasoning, stored with the incident record"],"humanInTheLoop":"Analysts own every containment decision and every escalation. They review a random sample of alerts the agent closed each week, correct wrong verdicts, and approve each new alert type or automated action before it goes live. The agent prepares; the analyst decides.","kpisToInstrument":["Mean time to triage per alert type, before and after","Share of alerts closed by the agent, and the share of those later reopened","Agreement rate between agent and analyst verdicts on a weekly sample","Missed true positives found in sampling or later investigations","Analyst hours spent on triage versus threat hunting and response"],"failureModes":[{"title":"A confident false negative","detail":"The agent closes a real phishing email or intrusion as benign. Limit autonomous closure to well calibrated alert types, sample closed alerts every week and treat every miss as an incident with a root cause."},{"title":"Prompt injection through the evidence","detail":"Attackers write text into emails or files that tells the model to mark them safe. Separate instructions from evidence, and test the agent with adversarial samples before and after every change."},{"title":"Automation without the runbook","detail":"The agent is switched on without written triage rules, so nobody can say whether a verdict was right. Write the runbook first; it is also what auditors and new analysts need."},{"title":"Analysts stop looking","detail":"Once the agent is usually right, reviews become rubber stamps. Keep a structured sample review and rotate who does it."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Triage of phishing, endpoint, network and cloud alerts for an organization's own cyber defence is not listed in Annex III. Recital 55 of the AI Act says that components intended to be used solely for cybersecurity purposes should not qualify as safety components, so the agent does not fall under Annex III point 2 (critical infrastructure), and for this scope the tier is minimal. The design changes that when the agent triages identity, data loss prevention, insider risk or user behaviour alerts in a way that scores or monitors individual employees: monitoring and evaluating the behaviour of persons in a work relationship falls under Annex III point 4(b), so that scope needs its own high risk assessment before it goes live. The Article 50(1) duty to disclose AI interaction does not apply because it is obvious to a reasonably well informed analyst that they are working with an AI agent. An operator that lets AI act autonomously on network or operational technology controls should assess that design separately, and reading employees' emails and sign in data remains subject to data protection law."},"regulations":["eu-ai-act","gdpr","nis2","dora","nist-ai-rmf","iso-42001","hipaa","pci-dss"],"guidance":[{"title":"NIST SP 800-61 Rev. 3, Incident Response Recommendations and Considerations for Cybersecurity Risk Management: A CSF 2.0 Community Profile","issuer":"NIST","region":"north-america","url":"https://csrc.nist.gov/pubs/sp/800/61/r3/final","note":"The US reference for incident detection, analysis and response, mapped to the NIST Cybersecurity Framework 2.0; a useful baseline for the runbooks an agent follows."},{"title":"OWASP Top 10 for LLM Applications","issuer":"OWASP","region":"global","url":"https://genai.owasp.org/llm-top-10/","note":"Lists prompt injection and excessive agency among the main risks of LLM applications, both directly relevant when an agent reads attacker controlled emails and can act in security tools."}],"controls":["Inventory entry for the agent with an accountable owner, the alert types in scope and the actions it may take","Written triage runbook per alert type, versioned with the agent's instructions","Audit trail of every verdict, the evidence used and every action proposed or taken","Weekly sample review of closed alerts with tracked agreement and misses","Adversarial test set (prompt injection, look alike domains, novel lures) run on every change"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered through an API token or on a schedule. Its\n**agent loop** calls the **custom functions** attached to the workflow, REST calls that query the\norganization's SIEM, email security, identity and endpoint tools. **MCP** servers expose further\ntools the agent can call, and retrieval over a **knowledge base** with hybrid retrieval surfaces\nrunbooks and past incident write ups, so the verdict cites the rule it followed. A further\nfunction can write the verdict and a draft incident summary back to the SIEM or case management\ntool.\n\nA **tool execution policy** controls which of these the agent may use, and that scoping is the\nmain defence against prompt injection in the evidence. **Guardrails** check conversations, not\nwhat a workflow reads, so an instruction hidden in an email is limited by what the agent is able\nto call rather than detected. Link only read only functions for evidence, and keep containment\n(isolating a device, disabling an account) out of the workflow: the agent proposes it and an\nanalyst carries it out. Every run keeps a **full audit trail** with downloadable run data.\n**Test suites** can replay labelled alerts, including adversarial samples, against the workflow\nwith deterministic or LLM based grading, and the platform is model agnostic, with EU and UAE data\nresidency for regions that require it."},"faq":[{"question":"How much analyst time does AI alert triage save?","answer":"It depends on the alert mix and on how much the agent may close on its own. St. Luke's University Health Network says its triage agent saves nearly 200 hours a month on reported phishing, Google Cloud reports that manual triage at Human Managed was reduced by more than 60%, and TÜV SÜD reports analysis about 60% to 70% faster. These are vendor case studies, so measure your own baseline before you plan on similar numbers."},{"question":"Should the AI close alerts by itself?","answer":"Only for alert types where its verdicts have matched your analysts' for weeks, and usually only for false positives such as benign reported emails. St. Luke's describes the agent closing thousands of false positive alerts, with analysts focusing on the real threats it surfaces. Containment actions such as isolating a device should stay with an analyst."},{"question":"Can attackers trick a triage agent?","answer":"Yes, that is the specific risk here: the agent reads text the attacker wrote. Treat email and file content as untrusted data, keep the agent's permissions read only, and test it with prompt injection and novel phishing samples on every change."},{"question":"Is phishing triage a separate use case from SOC alert triage?","answer":"No, it is a common first step in the same job: user reported phishing is high in volume and its evidence is well defined. Microsoft has renamed its Phishing Triage Agent the Security Alert Triage Agent, and the US Federal Housing Finance Agency has automated responses to user reported suspicious emails since 2021."}],"related":["aiops-incident-triage","it-service-desk-resolution-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with seven deployment records checked against their sources (Microsoft and Google Cloud case studies and the 2025 US federal AI use case inventory). Editor pass the same day corrected before AI triage times, credited Human Managed figures to Google Cloud and grounded the EU AI Act basis in Recital 55."},{"date":"2026-09-27","note":"Second editor pass. EU AI Act tier set to context dependent with an Annex III point 4(b) split for alerts that monitor employees and the Article 50 reasoning corrected; minutes per alert lowered to 3 to 10 and checked against St. Luke's hours; the Blits.ai build no longer claims guardrails, PII masking, a tool execution policy or confirmation steps inside the workflow, and drops Slack."}],"slug":"security-alert-triage-and-investigation","url":"https://www.blits.ai/ai-use-cases/security-alert-triage-and-investigation","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":3,"nUpTo":0,"median":60,"min":60,"max":70,"byClaimant":{"organization":1,"vendor":2,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"avanade-security-copilot-identity-investigations","pooled":true},{"id":"human-managed-google-secops-alert-triage","pooled":true},{"id":"tuv-sud-security-copilot-threat-analysis","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":97,"min":97,"max":97,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"human-managed-google-secops-alert-triage","pooled":true}]},{"kpi":"hours-saved","label":"Hours saved","unit":"hours","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":200,"min":200,"max":200,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"st-lukes-security-alert-triage-agent","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":20,"min":20,"max":20,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"sep2-gemini-security-triage-agents","pooled":true}]},{"kpi":"mttr-reduction","label":"Time to repair reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"avanade-security-copilot-identity-investigations","pooled":false}]}],"indicativeValueResult":{"low":45000,"high":1000000},"evidence":["avanade-security-copilot-identity-investigations","fhfa-phishing-email-identification","human-managed-google-secops-alert-triage","sep2-gemini-security-triage-agents","st-lukes-security-alert-triage-agent","tuv-sud-security-copilot-threat-analysis","us-ice-soc-compromised-email-detector"]},{"title":"AI for settlement fail prediction and post trade exception management","shortTitle":"Settlement fail prediction","seoTitle":"AI settlement fail prediction for T+1","metaDescription":"AI scores which securities trades will fail to settle and works the exceptions early. Clearstream flags at risk instructions up to four business days in advance.","definition":"AI that scores each pending securities settlement instruction for its likelihood of failing, names the probable cause (unmatched instruction, wrong settlement details, lack of securities or cash), and helps operations teams work the exceptions and counterparty queries before the intended settlement date, so fewer trades fail and fewer late settlement penalties are paid.","aliases":["settlement fail prediction","post trade exception management AI","trade settlement exceptions automation","CSDR penalty prediction","T+1 settlement fail prevention"],"industries":["capital-markets","banking","wealth-and-asset-management"],"functions":["operations","risk-management"],"patterns":["prediction-and-scoring","classification-and-routing","agentic-workflow","content-generation"],"channels":["internal-tools","api","email"],"audience":"back-office","autonomy":"copilot","adoptionStage":"early-adopters","segment":"back-office","problem":"A securities trade settles only when both sides have sent matching instructions, the seller\nhas the securities and the buyer has the cash, all by the intended settlement date. When any of\nthese is missing the trade fails. Fails tie up liquidity and collateral, create credit exposure\nbetween counterparties and, in the EU, trigger cash penalties under the settlement discipline\nregime of the Central Securities Depositories Regulation (CSDR), which applies since February\n2022. BNY Mellon has described how on a typical day about two percent of US Treasury\ntransactions fail to settle.\n\nMany operations teams still find a fail after it has happened: they work through a pending\nand failing report in the morning, look up each trade in several systems and chase\ncounterparties and custodians by email and phone. The window to fix a problem is getting\nshorter. The United States moved most broker dealer transactions to settlement one business day\nafter the trade date (T+1) in May 2024, and the EU plans to follow in October 2027, which leaves\nhours rather than days to spot a mismatch and correct it. The step change is to predict which\ninstructions are at risk while there is still time to act, and to take the reading, looking up\nand chasing out of the exception queue.","problemStats":[{"statement":"BNY Mellon states that on a typical day approximately two percent of US Treasury transactions fail to settle.","sourceTitle":"BNY Mellon and Google Cloud Collaborate to Help Transform U.S. Treasury Market Settlement and Clearance Process","sourceUrl":"https://www.bny.com/corporate/global/en/about-us/newsroom/company-news/bny-mellon-and-google-cloud-collaborate-to-help-transform-us-treasury-market-settlement-and-clearance-process.html","year":2021},{"statement":"Markets Media reports that, according to Euroclear's settlement efficiency analysis, late matching fails represent around 25% of fails.","sourceTitle":"Euroclear Updates EasyFocus+ to Ease T+1 Transition","sourceUrl":"https://www.marketsmedia.com/euroclear-updates-easyfocus-to-ease-t1-transition/","year":2025},{"statement":"ESMA reports that in June 2024 settlement fails across all EEA CSDs ranged from about 2.5% of the number of settlement instructions for sovereign bonds to about 20% for ETFs.","sourceTitle":"Final Report on Technical Advice on CSDR Penalty Mechanism","sourceUrl":"https://www.esma.europa.eu/sites/default/files/2024-11/ESMA74-2119945925-2059_Final_Report_on_Technical_Advice_on_CSDR_Penalty_Mechanism.pdf","year":2024}],"howItWorks":"1. **Collect the instruction lifecycle.** The system reads every pending instruction with its\n   matching status, counterparty, instrument, market, place of settlement, amount and the\n   securities and cash positions behind it, from the firm's own books and from the status\n   messages of the central securities depository (CSD) or custodian.\n2. **Score the fail risk.** A model trained on historical settlement outcomes estimates the\n   probability that each instruction settles on time and ranks the drivers, such as an\n   unmatched instruction, a counterparty with a poor record, an illiquid bond or a short\n   position. Some CSDs now offer this score to their participants as a service.\n3. **Estimate the cost.** For each instruction at risk it estimates the exposure: the late\n   settlement penalty, the funding cost and the client impact, so the team works the most\n   expensive problems first.\n4. **Work the exception.** For the top of the queue an AI agent gathers the facts (both\n   instructions side by side, standing settlement instructions, inventory, previous\n   correspondence), proposes the likely fix and drafts the query to the counterparty or\n   custodian in the channel they use.\n5. **Decide and act.** An operator approves any amended instruction, securities borrow,\n   partial settlement or cash movement. The agent sends approved queries, tracks answers,\n   chases before the cutoff and records the root cause so the same break does not return.","valueDrivers":["risk-reduction","cost-to-serve","employee-productivity","speed"],"kpis":["handling-time-reduction","accuracy","error-reduction","cost-reduction","automation-rate"],"indicativeValue":{"referenceOrg":"A broker dealer settling 2 million securities instructions a year in EU markets","inputs":[{"key":"instructions","label":"Settlement instructions per year","low":2000000,"high":2000000,"unit":"instructions per year","note":"The reference firm. Replace with your own instruction volume."},{"key":"failRate","label":"Share of instructions that fail today","low":0.025,"high":0.05,"unit":"fraction of instructions","note":"The low value follows ESMA's final report on the CSDR penalty mechanism (November 2024), where fails across all EEA CSDs in June 2024 were about 2.5 percent of the number of settlement instructions for sovereign bonds, the lowest asset class, and about 20 percent for ETFs. The high value of five percent is an editorial assumption for a mixed book, not a sourced figure. Replace both with your own fail rate."},{"key":"preventedShare","label":"Share of fails prevented by acting on the prediction","low":0.1,"high":0.25,"unit":"fraction of fails","note":"Editorial assumption, deliberately conservative. No deploying organization on this page has published a measured reduction in fails."},{"key":"costPerFail","label":"Cost of one fail","low":50,"high":150,"unit":"EUR per fail","note":"Editorial assumption covering cash penalties, funding cost and operations handling time. Replace with your own penalty and handling data."}],"formula":"instructions * failRate * preventedShare * costPerFail","currency":"EUR","period":"per year","resultLabel":"Fail cost avoided","caveat":"Counts only fails that are prevented. It leaves out the operator time saved on fails that still happen, penalties received from counterparties, the capital and liquidity effect of fewer open fails, client impact, and the cost of the data, the models and the integration."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The prediction needs clean, joined history of instructions, matching statuses and outcomes across every CSD and custodian the firm uses, and the model falls under model risk management. Using the score a CSD already provides lowers the entry cost; the exception work still needs integration with the settlement system, inventory and counterparty communication.","dataPrerequisites":["At least a year of instruction history with matching status changes and the final settlement outcome","Standing settlement instructions and counterparty static data","Securities inventory and cash positions per account and depot","Penalty reports from each CSD, to price the cost of a fail","Past exception cases with root cause and the correspondence that resolved them"],"integrations":["Settlement or post trade processing system (pending and failing instructions)","CSD and custodian status feeds, and their prediction services where offered","Inventory, securities lending and collateral systems","Email and post trade query platforms used with counterparties","Case or exception management tool for the operations queue"]},"implementation":{"steps":[{"title":"Measure the fails you have","detail":"Take six to twelve months of fails and penalties and group them by cause, market, counterparty and asset class. Late matching, wrong settlement details and lack of securities usually need different fixes, and the size of each group decides where to start."},{"title":"Use the CSD score before you build your own","detail":"Where your CSD or custodian offers a settlement prediction, feed its score and drivers into the operations queue first. Build an internal model only for flows the service does not see, such as your own internal settlements and positions."},{"title":"Rank the queue by cost, not by age","detail":"Combine the fail probability with the penalty, funding and client impact so operators start with the instructions that matter most, and show the drivers next to each score."},{"title":"Automate the gathering and the first query","detail":"Let the agent assemble both instructions, the standing settlement instructions and the inventory, propose the fix and draft the counterparty query. Operators approve every outgoing message at first, then allow routine information requests to go out directly. From that point, tell recipients in each message that it was written and sent by an AI agent."},{"title":"Close the loop on root causes","detail":"Record the confirmed cause of every fail and every prevented fail, feed it back into the model and fix the static data or counterparty set up that caused it."}],"guardrails":["Maker checker approval on every amended instruction, borrow, partial settlement or cash movement","The agent never changes standing settlement instructions or static data on its own","Every score shows its drivers, so operators can see why an instruction is flagged","Outgoing counterparty queries use approved templates and contain only the data needed for the trade","Model monitoring for drift, with a documented fallback to the manual pending report"],"humanInTheLoop":"Operators decide on every action that changes an instruction or moves securities or cash, and own escalations to the front office and clients. Team leads review a weekly sample of predicted and actual fails, and model owners validate the prediction model on a schedule.","kpisToInstrument":["Settlement efficiency by value and volume, before and after, per market","Late settlement penalties paid per month, net of penalties received","Precision and recall of the fail prediction on a holdout period","Median time from flag to resolution for instructions at risk","Operator minutes per exception and share of queries sent without manual drafting"],"failureModes":[{"title":"A score nobody acts on","detail":"The prediction lands in a dashboard outside the operations queue and changes nothing. Put the score and the drivers inside the tool operators already work in."},{"title":"Too many flags","detail":"A model tuned for recall floods the team with instructions that would have settled anyway. Tune the threshold on cost and track precision per market."},{"title":"Stale static data behind the prediction","detail":"Wrong standing settlement instructions cause fails the model cannot explain. Treat reference data fixes as part of the programme."},{"title":"A model that learns from a different cycle","detail":"A model trained under T+2 behaviour misjudges risk after a move to T+1. Retrain and revalidate around every change in settlement cycle or market practice."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Predicting settlement fails and handling post trade exceptions between professional market participants is not a use listed in Annex III and is not a prohibited practice under Article 5, so the tier depends on how the agent communicates. While an operator reviews and sends every message, the system is minimal risk: the messages are the firm's own correspondence and the firm as deployer owes AI literacy for staff (Article 4). Once the agent sends queries or chasers to counterparty or custodian staff itself, as the playbook recommends for routine information requests, it interacts directly with natural persons and Article 50(1) requires telling the recipients they are dealing with an AI system. In both designs the provider of the text generating system must mark its output as AI generated in a machine readable format under Article 50(2). Model risk and operational resilience controls apply on top."},"regulations":["eu-ai-act","dora","us-sr-11-7","iso-42001"],"guidance":[{"title":"Regulation (EU) No 909/2014 on improving securities settlement in the European Union and on central securities depositories","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2014/909/oj","note":"The CSDR, whose settlement discipline measures include cash penalties for participants that fail to deliver securities or cash by the intended settlement date."},{"title":"ESMA finalises its advice on the CSDR Penalty Mechanism","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/press-news/esma-news/esma-finalises-its-advice-csdr-penalty-mechanism","note":"ESMA's November 2024 technical advice proposes a moderate increase of penalty rates and says the penalty mechanism has improved settlement efficiency since February 2022."},{"title":"SEC Finalizes Rules to Reduce Risks in Clearance and Settlement","issuer":"US Securities and Exchange Commission","region":"north-america","url":"https://www.sec.gov/newsroom/press-releases/2023-29","note":"The SEC shortened the standard settlement cycle for most broker dealer transactions from two business days after the trade date to one (T+1)."}],"controls":["Model inventory entry and periodic validation for any internal fail prediction model","Documented fallback to the manual pending and failing report if the model or feed is unavailable","Full trail of scores, proposed fixes, approvals and counterparty messages per instruction","Monthly review of penalties and fails against the prediction, per market and counterparty","Third party risk assessment for CSD or vendor prediction services under DORA"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the prediction itself comes from the CSD or custodian service or from the firm's own\nmodel; the platform does the exception work around it. An **agentic workflow**, started through\nthe API by the settlement system or on a schedule before each cutoff, uses **custom functions**\nto read the at risk instructions, the matching status, standing settlement instructions and\ninventory through REST or SQL, and a **knowledge base** with hybrid retrieval holds market\nrules, CSD procedures and internal runbooks. The agent proposes the fix and drafts the\ncounterparty query with **structured output**.\n\n**Human in the loop approval** holds every amended instruction, borrow or cash movement until an\noperator approves it, and the inbound and outbound **email channel** can carry approved\ncounterparty queries and their replies. Every run keeps a **full audit trail**, **guardrails**\ncheck outgoing text, and **test suites** run the agent and workflow against sample cases before\na new market or exception type goes live. The platform is model\nagnostic and can run in EU or UAE regions for data residency."},"faq":[{"question":"Can AI really predict which trades will fail to settle?","answer":"Central securities depositories already offer such predictions. Clearstream's product page says its Settlement Prediction Tool calculates the likelihood that an instruction settles on time and identifies the three primary factors most likely to cause a fail up to four business days in advance. In February 2021 BNY Mellon said its model with Google Cloud aimed to help clients predict about 40 percent of settlement failures in Fed eligible securities with 90 percent accuracy, a goal rather than a result. Neither states a measured accuracy or a measured reduction in fails on the pages cited here."},{"question":"Do we need to build our own model?","answer":"Not to start. Clearstream offers its prediction to clients through its Xact Web Portal, and Euroclear offers EasyFocus+, announced in June 2025, which gives each pending instruction a matching score (the predictive likelihood that it will be matched) and shows the CSDR penalty impact across its CSDs. An internal model adds value for flows those services do not see, such as your own inventory and internal settlements."},{"question":"Where does generative AI help if the prediction is a classic model?","answer":"In the exception work: reading both instructions, drafting counterparty queries and chasing answers. At BNY, more than ten percent of client inquiries about its transactions were resolved or assisted by AI on BNY's Eliza platform, and BNY says this brought eighty percent faster processing of those inquiries. The inquiries are not limited to settlement exceptions."},{"question":"Is this high risk under the EU AI Act?","answer":"No. Settlement fail prediction between professional market participants is not an Annex III use, and the tier depends on the design. With an operator sending every message it is minimal risk. Once the agent sends queries to counterparty or custodian staff itself, Article 50(1) requires telling those recipients they are dealing with an AI system, and the provider of a text generating system must mark its output under Article 50(2) in either case. Beyond that, model risk management, operational resilience under DORA and good records of every approved action are the controls that matter."}],"related":["ledger-and-payment-reconciliation","payment-investigations-and-exceptions"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, written in a discover run for capital markets with four evidence records from Clearstream, Euroclear and BNY, quotes checked against the sources."},{"date":"2026-09-27","note":"Editorial review. Added Article 50(1) to the EU AI Act basis and FAQ, made the high fail rate an explicit editorial assumption, switched the worked example to EUR, used BNY's own quote for the Eliza figure, added Euroclear's product page and archived release, merged the two BNY records under one organization name, and dated the BNY Mellon release (4 February 2021) with the Google Cloud copy."},{"date":"2026-09-27","note":"Second review. Set the EU AI Act tier to context dependent with Article 50(1) and 50(2), softened the prediction FAQ, took the low fail rate and a new problem statistic from ESMA's EEA data, dropped processing time from the tracked KPIs so the adjacent BNY Eliza figure does not form a benchmark, and removed PRA SS1/23."}],"slug":"settlement-fail-prediction-and-exception-management","url":"https://www.blits.ai/ai-use-cases/settlement-fail-prediction-and-exception-management","benchmarks":[],"indicativeValueResult":{"low":250000,"high":3750000},"evidence":["bny-eliza-client-settlement-inquiries","bny-mellon-treasury-settlement-fail-prediction","clearstream-settlement-prediction-tool","euroclear-easyfocus-plus-settlement-analytics"]},{"title":"AI for software vulnerability triage and remediation","shortTitle":"Vulnerability remediation","seoTitle":"AI vulnerability remediation and security autofix","metaDescription":"AI that triages security findings and drafts fixes. Chrome fixed 1,072 security bugs in two milestones, with LLMs drafting candidate fixes for most vulnerabilities.","definition":"AI that takes security findings from scanners, fuzzers and bug reports, filters out duplicates and false positives, reproduces and ranks the real ones, and drafts a code fix with a test for each, which a developer reviews and merges through the normal change process.","aliases":["AI vulnerability autofix","AI security patch generation","automated vulnerability fixing","SAST backlog remediation with AI","AI vulnerability triage"],"industries":["cross-industry","technology","healthcare"],"functions":["security-operations","it-and-engineering"],"patterns":["code-generation","agentic-workflow","classification-and-routing"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Security tools find far more problems than teams can fix. Static analysis (SAST), dependency\nscanners, fuzzers, penetration tests and bug bounty reports all feed the same backlog, and each\nfinding needs someone to confirm it is real, work out how serious it is, find the owner and write\na fix that does not break anything. Developers see security tickets as interruptions to feature\nwork, so the backlog grows and old findings stay open.\n\nAI is making the imbalance sharper on both sides. Models are now used to find vulnerabilities\n(Google's Big Sleep agent found bugs in Chrome's V8 engine in 2025). Separately, the Chrome\nsecurity team reports receiving more bug reports by March 2026 than in all of 2025, without\nnaming a single cause, and says triaging a single report used to take 5 to 30 or more minutes of\nexpert time. At the same\ntime, low quality AI generated reports waste maintainers' time, and developers using coding\nassistants ship more code, and so more findings, than before. Fixing has to scale as fast as\nfinding, without letting unreviewed code into production.","problemStats":[{"statement":"The Chrome security team reports that historically, triaging a single security report took anywhere from 5 to 30 or more minutes and relied primarily on human expertise.","sourceTitle":"Stronger with every update: How we're making Chrome and the web safer in the AI Era","sourceUrl":"https://blog.google/security/chrome-stronger-with-every-update/","year":2026}],"howItWorks":"1. **Collect and deduplicate.** Findings arrive from scanners, fuzzers, CI pipelines and external\n   reports. The system drops spam and duplicates and checks that each one describes a real\n   security issue in scope.\n2. **Validate and rank.** Where possible it reproduces the bug (running the proof of concept or\n   the failing test), checks whether the vulnerable code is reachable, and assigns a severity\n   using written guidelines. Findings that mitigations already neutralise are marked for a human\n   to confirm and close.\n3. **Route to the owner.** The finding goes to the right component and code owner with the stack\n   trace, severity and context attached.\n4. **Draft the fix.** A fixing agent proposes one or more candidate patches; a separate critic\n   step or reviewer model checks them against the style guide, and the fix is rebuilt and rescanned\n   to prove the finding is gone. Test writing agents add a regression test.\n5. **Human review and merge.** The developer reviews the pull request like any other change and\n   merges, edits or rejects it. Rejections and edits feed back into the prompts and examples.","valueDrivers":["risk-reduction","employee-productivity","speed","compliance"],"kpis":["processing-time-reduction","mttr-reduction","automation-rate","false-positive-reduction","productivity-gain","hours-saved","interactions-handled"],"indicativeValue":{"referenceOrg":"A software organization that fixes 2,000 to 5,000 security findings a year","inputs":[{"key":"findingsPerYear","label":"Security findings fixed per year","low":2000,"high":5000,"unit":"findings per year","note":"Editorial assumption for a mid sized engineering organization. Replace with the count from your vulnerability management or code scanning tool."},{"key":"hoursPerFix","label":"Developer hours per manual fix, including review","low":1.5,"high":4,"unit":"hours per finding","note":"GitHub reports a median of 1.5 hours to resolve code scanning alerts manually, measured in its public beta on alerts raised in pull requests on new code; the high end is an editorial allowance for older backlog findings, which take longer.","sourceUrl":"https://github.blog/news-insights/product-news/secure-code-more-than-three-times-faster-with-copilot-autofix/"},{"key":"aiFixShare","label":"Share of findings where an AI drafted fix is accepted","low":0.1,"high":0.2,"unit":"fraction of findings","note":"An editorial range around the only measured acceptance rate, Google's 15% of sanitizer bugs fixed by its LLM pipeline with human review. Chrome's candidate fixes for most vulnerabilities are drafts, not accepted fixes, so they are not used here.","sourceUrl":"https://research.google/pubs/ai-powered-patching-the-future-of-automated-vulnerability-fixes/"},{"key":"timeSavedShare","label":"Share of developer time saved on those findings","low":0.5,"high":0.7,"unit":"fraction of fix time","note":"GitHub reports a median of 28 minutes with Copilot Autofix against 1.5 hours manually, about 69% less, measured in its public beta on new alerts in pull requests; the low end allows for backlog fixes and review of harder fixes.","sourceUrl":"https://github.blog/news-insights/product-news/secure-code-more-than-three-times-faster-with-copilot-autofix/"},{"key":"costPerHour","label":"Fully loaded cost of a developer hour","low":80,"high":120,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"findingsPerYear * hoursPerFix * aiFixShare * timeSavedShare * costPerHour","currency":"USD","period":"per year","resultLabel":"Developer time released on security fixes, valued at cost","caveat":"Counts developer time on accepted fixes only. It leaves out the time saved in triage, the cost of the AI and scanning tools, the value of a smaller exposure window, and the cost of reviewing fixes that are rejected."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Generating a patch is the easy part. The value depends on being able to build, test and rescan each candidate automatically, and on code owners who trust the pipeline enough to review its pull requests promptly.","dataPrerequisites":["Findings with enough detail to act on (rule, location, trace or proof of concept)","A build and test setup that can run a candidate fix automatically","Written severity guidelines and code ownership per component","Past fixes and rejected findings, as examples and as a test set"],"integrations":["Source control and pull requests (for example GitHub, GitLab or Azure DevOps)","Code scanning, dependency scanning and fuzzing tools","CI pipeline to build, test and rescan candidate fixes","Issue tracker or vulnerability management system for ownership and status","Bug bounty or external report intake, where the organization runs one"]},"implementation":{"steps":[{"title":"Measure the backlog and the flow","detail":"Count open findings by severity, age and class, and measure how long triage and fixing take today. Pick one or two high volume classes (for example injection or memory safety bugs in one language) for the first wave."},{"title":"Automate the boring triage first","detail":"Deduplication, reproduction, severity by written rules and routing to the owner deliver value before any AI written code is merged, and they give you the clean data the fixing step needs."},{"title":"Make every candidate fix prove itself","detail":"A candidate fix must build, pass the existing tests, include a regression test and make the scanner or fuzzer stop reporting the finding. Discard candidates that fail, rather than sending them to developers."},{"title":"Review like any other change","detail":"Fixes arrive as ordinary pull requests with an explanation, go through code owner review and the normal release process, and are labelled as AI drafted so acceptance can be measured."},{"title":"Track acceptance per class and tune","detail":"Measure how many fixes are merged unchanged, edited or rejected per vulnerability class, and expand to new classes only where acceptance is good and no regressions were introduced."}],"guardrails":["No AI drafted fix reaches production without human code review and the normal CI checks","Candidate fixes must pass build, tests and a rescan before a developer sees them","Agents that analyse code run in isolated environments without general internet access and with write access limited to the source tree","Severity changes and closures of findings as not exploitable need a named human's confirmation","Every AI drafted change is labelled, so its acceptance and any later regressions can be traced"],"humanInTheLoop":"Developers review and merge every fix, and security engineers confirm severity and any decision to close a finding without a code change. The AI does the reproduction, the drafting and the testing; humans stay accountable for what ships.","kpisToInstrument":["Median time from finding to merged fix, per severity","Share of AI drafted fixes merged unchanged, edited or rejected","Open findings in the backlog by severity and age","Regressions or reopened findings traced to AI drafted fixes","Triage time per incoming report"],"failureModes":[{"title":"A fix that hides the symptom","detail":"The patch silences the scanner (for example by adding a check in the wrong place) but leaves the vulnerability exploitable. Require a regression test that exercises the attack, and have security review high severity fixes."},{"title":"Review fatigue","detail":"A flood of AI pull requests gets merged without real review. Batch fixes by component, limit the daily volume per owner and sample merged fixes for security review."},{"title":"Plausible but false findings","detail":"AI generated reports can look convincing and still be wrong, and each costs expert time to disprove. Require reproduction before a human spends time on a report."},{"title":"Agents with too much reach","detail":"An agent that can run code and reach the network can leak source code or be steered by content in the repository. Sandbox it, restrict its network access and keep its permissions minimal."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Drafting and triaging code fixes for an organization's own software is not an Annex III use, and developers, not the public, interact with the system. The software being fixed remains subject to its own security and resilience rules, whoever wrote the fix."},"regulations":["eu-ai-act","nis2","dora","pci-dss","nist-ai-rmf","iso-42001"],"guidance":[{"title":"NIST SP 800-218, Secure Software Development Framework (SSDF) Version 1.1","issuer":"NIST","region":"north-america","url":"https://csrc.nist.gov/pubs/sp/800/218/final","note":"Recommended practices for mitigating software vulnerabilities, including identifying, analysing and remediating them; the process an AI remediation pipeline has to fit into."},{"title":"NIST SP 800-218A, Secure Software Development Practices for Generative AI and Dual-Use Foundation Models","issuer":"NIST","region":"north-america","url":"https://csrc.nist.gov/pubs/sp/800/218/a/final","note":"An SSDF community profile that adds practices for developing generative AI systems, useful when the remediation pipeline itself is built on models."},{"title":"Secure by Design","issuer":"Cybersecurity and Infrastructure Security Agency (CISA)","region":"north-america","url":"https://www.cisa.gov/securebydesign","note":"CISA's programme asking every technology provider to take ownership of product security at the executive level, with alerts on eliminating specific vulnerability types such as cross site scripting and OS command injection."}],"controls":["Inventory entry for the pipeline with an owner, the vulnerability classes in scope and the repositories it may touch","Code owner review and CI checks on every AI drafted change, with the change labelled as AI drafted","Isolated, network restricted execution environment for agents that read and run code","Metrics on acceptance, regressions and backlog age reviewed by the security lead each month","Documented severity guidelines applied the same way by humans and the pipeline"],"incidents":[{"title":"The I in LLM stands for intelligence (curl project on AI generated vulnerability reports)","url":"https://daniel.haxx.se/blog/2024/01/02/the-i-in-llm-stands-for-intelligence/","note":"The curl maintainer describes AI generated bug bounty reports that looked plausible but were hallucinated, and how each one takes a human's time to disprove; the reason to require reproduction before triage."}]},"blitsAi":{"howToBuild":"On Blits.ai the triage half is an **agentic workflow** triggered by an API token from the CI\npipeline or on a schedule: its **agent loop** calls **custom functions** that read findings\nfrom the scanning tool and the issue tracker as REST calls, or tools exposed through **MCP**\nservers, applies the organization's severity guidelines from a **knowledge base**, and writes\nthe routing and a summary back to the ticket. A **tool execution policy** restricts which tools\nthe agent may call, and **human in the loop confirmation** is required before a finding is\nclosed or its severity changed.\n\nCode generation itself belongs in the developer's own toolchain, where fixes are built, tested\nand reviewed; the workflow on Blits.ai prepares the context, tracks each finding's status and\nkeeps a **full audit trail per run**. The platform is model agnostic, so the security team can\npick the model per task, and **test suites** can replay a labelled set of past findings to check\nthe triage verdicts on every change."},"faq":[{"question":"Can AI fix security vulnerabilities on its own?","answer":"It can draft the fix, but a developer should still review and merge it. Google reports that its Gemini pipeline fixed 15% of sanitizer bugs found in unit tests, with every fix going to human review, and the Chrome team says LLMs now generate candidate fixes for most vulnerabilities while developers evaluate them. Plan for review capacity, not for zero touch."},{"question":"Where should a team start?","answer":"With triage: deduplication, reproduction, severity by written rules and routing. It removes noise before anyone writes code. In a Checkmarx case study, PatientPoint's application security engineer says the Checkmarx triage and remediation tools identified false positives and gave developers the opportunity for human review. Then add AI drafted fixes for one well understood vulnerability class."},{"question":"How fast can a backlog shrink?","answer":"Snyk reports, in a case study, that Labelbox's lead security engineer cleared a backlog of high severity static analysis findings in two to three weeks by pairing an AI coding agent with Snyk's rescanning; he estimated the work at a full calendar year with the old workflow. That is one small company's experience; large codebases with many owners move at the pace of review."}],"related":["developer-coding-assistant","security-alert-triage-and-investigation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the discovery workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with four deployment records (Google Chrome, Google's sanitizer bug pipeline, Labelbox and PatientPoint) checked against their sources; figures not attributable to the AI kept out of the benchmark metrics."}],"slug":"software-vulnerability-remediation","url":"https://www.blits.ai/ai-use-cases/software-vulnerability-remediation","benchmarks":[],"indicativeValueResult":{"low":12000,"high":336000},"evidence":["google-chrome-ai-vulnerability-triage-and-fixing","google-sanitizer-bug-ai-patching","labelbox-snyk-ai-sast-backlog-remediation","patientpoint-checkmarx-ai-triage-and-remediation"]},{"title":"AI for subrogation opportunity detection","shortTitle":"Subrogation detection","seoTitle":"AI subrogation detection for insurance claims","metaDescription":"AI flags claims a third party caused and scores the recovery. Central Insurance uses it; Shift reports over USD 1 million a month recovered at a top 25 US insurer.","definition":"AI that reads open and closed claims to find cases where a third party is wholly or partly liable, estimates liability and the recoverable amount under the applicable negligence and recovery rules, and sends scored recovery opportunities with their reasons to the subrogation team.","aliases":["subrogation AI","recovery opportunity detection","claims recovery analytics","AI subrogation referral"],"industries":["insurance"],"functions":["claims","collections-and-recovery"],"patterns":["classification-and-routing","prediction-and-scoring","document-processing","summarization"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"assist","adoptionStage":"early-adopters","segment":"claims","problem":"When an insurer pays a claim that someone else caused, it can recover the money from that party or\nits insurer. In practice many of those recoveries are never pursued. Handlers focus on settling the\nclaim for the customer, the signs of third party liability sit in free text notes, police reports\nand photos, and the rules on comparative negligence and recovery differ by state or country.\n\nReferrals to the subrogation team therefore depend on individual handlers spotting the opportunity,\noften too late, when evidence is gone or deadlines have passed. Recovery teams in turn spend time on\nreferrals with little chance of success. Missed subrogation is a quiet form of claims leakage: no\ncustomer complains about it, so it rarely surfaces on its own.","problemStats":[],"howItWorks":"1. **Read every claim early.** The AI reads the claim notes, statements, police reports and other\n   documents from the first days of the claim, not only when a handler refers it.\n2. **Identify who else is responsible.** It extracts the parties and facts and assesses whether a\n   third party, product, contractor or other insurer may be liable.\n3. **Apply the rules.** It checks the applicable comparative negligence, recovery and limitation\n   rules for the jurisdiction and line of business.\n4. **Estimate and score.** It estimates liability shares and the recoverable amount and scores the\n   opportunity by expected recovery.\n5. **Refer with reasons.** Scored alerts with the supporting facts and rules go to the subrogation\n   team, which decides whether to pursue, and outcomes feed back into the model.","valueDrivers":["risk-reduction","employee-productivity","speed"],"kpis":["revenue-recovered","handling-time-reduction","productivity-gain"],"indicativeValue":{"referenceOrg":"An auto and property insurer paying USD 500 million in claims a year","inputs":[{"key":"claimsPaid","label":"Claims paid per year","low":500000000,"high":500000000,"unit":"USD per year","note":"The reference insurer."},{"key":"recoverableShare","label":"Share of claims paid that is recoverable from third parties","low":0.03,"high":0.06,"unit":"fraction of claims paid","note":"Editorial assumption; depends heavily on the lines of business and jurisdictions. Replace with your own recovery history."},{"key":"missedShare","label":"Share of recoverable amounts not pursued today","low":0.1,"high":0.25,"unit":"fraction of recoverable amounts","note":"Editorial assumption; estimate it by auditing a sample of closed claims."},{"key":"collectedShare","label":"Share of newly found opportunities actually collected","low":0.4,"high":0.7,"unit":"fraction of opportunities","note":"Editorial assumption; not every liable party pays in full."}],"formula":"claimsPaid * recoverableShare * missedShare * collectedShare","currency":"USD","period":"per year","resultLabel":"Additional recoveries collected","caveat":"Gross recoveries only. It leaves out the cost of pursuing recoveries (staff, legal, arbitration fees), the time value of money, the reduction in customer excess where recoveries are shared with the policyholder, and the cost of the platform."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The model needs claim notes and documents plus machine readable recovery rules per jurisdiction. Integration is mostly read only on the claims system plus a referral into the recovery workflow, so the operational risk is modest; the effort is in the rules content and in feedback from recovery outcomes.","dataPrerequisites":["Claim notes, statements, police reports and photos linked to each claim","Historical subrogation referrals with outcomes and amounts recovered","Comparative negligence, recovery and limitation rules per jurisdiction and line"],"integrations":["Claims management system (read access to claims, notes and payments)","Subrogation or recovery workflow and case management","Document storage for police reports and evidence","Inter company arbitration or recovery platforms where used"]},"implementation":{"steps":[{"title":"Audit a sample of closed claims","detail":"Have experienced recovery staff review a few hundred closed claims to estimate how much was missed and why. It sizes the prize and creates the first labelled data."},{"title":"Start with one line and one jurisdiction set","detail":"Auto physical damage and the personal injury protection and medical payment rules of a few states or countries are typical starting points, because volume is high and the recovery rules for these exposures are written down per state. The top 25 US insurer on this page started with its auto subrogation team and added property later."},{"title":"Run next to handler referrals","detail":"Keep manual referrals and let the AI add its own, so you can measure how many opportunities it finds that people missed and how early."},{"title":"Tune to what the team accepts","detail":"Track which alerts the recovery team accepts and what is collected, and set thresholds to the team's capacity rather than the model's recall."},{"title":"Keep the rules current","detail":"Recovery law and limitation periods change. Give the rules an owner and a review cycle, and version them with the model."}],"guardrails":["The AI refers opportunities; people decide whether to pursue and what to demand","Every alert shows the facts, liability reasoning and the rule applied","Limitation deadlines tracked from the alert so late referrals are visible","The policyholder's claim is never delayed or reduced because of a recovery opportunity"],"humanInTheLoop":"The subrogation team reviews every alert, decides on pursuit and negotiates recoveries. Claims handlers can still refer claims manually, and recovery leads approve changes to rules and thresholds.","kpisToInstrument":["Share of alerts accepted by the recovery team","Amount recovered from AI raised opportunities per month","Days from first notice of loss to subrogation referral","Opportunities found by the AI that handlers had not referred","Recoveries lost to limitation deadlines"],"failureModes":[{"title":"Alerts nobody pursues","detail":"The model raises more opportunities than the team can work, and value is lost anyway. Match thresholds to capacity and prioritise by expected recovery."},{"title":"Wrong law, wrong demand","detail":"Outdated or misapplied negligence rules lead to demands that fail. Version the rules and review them on a schedule."},{"title":"Handlers stop referring","detail":"Teams assume the AI will catch everything. Keep manual referral and measure both sources."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Detecting recovery opportunities against third parties and other insurers is not listed in Annex III: point 5(c) covers only risk assessment and pricing of natural persons in life and health insurance, and the system does not decide on a natural person's access to a service. It is an internal tool that does not converse with the public or publish generated content, so the deployer transparency duties of Article 50 do not apply. Personal data in claim files, including data about the third party, is still subject to GDPR."},"regulations":["eu-ai-act","gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Sets risk based, proportionate expectations (data governance, explainability, human oversight) for insurers' AI systems that are neither prohibited nor high risk under the AI Act, which includes back office claims models like this one."}],"controls":["Versioned recovery rules with an owner and review dates","Log of alerts, decisions and recovery outcomes for model monitoring","Access controls on claim files used by the model"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that runs on a schedule or on claim events through the\nAPI. An **AI agent** reads the claim notes and documents ingested into the **knowledge base**,\nretrieves the applicable recovery rules with **hybrid retrieval**, and returns a **structured\noutput** with parties, liability reasoning, the rule applied and an estimated recovery. **Custom\nfunctions** read the claim from the claims system and create the referral in the recovery workflow.\n\n**Human in the loop approval** keeps the decision to pursue with the recovery team, the **audit\ntrail** records every run, and **test suites** grade the agent against claims where the recovery\noutcome is known. The platform is model agnostic, so the insurer can choose the model per task."},"faq":[{"question":"How much can AI add to subrogation recoveries?","answer":"Published results are vendor figures for unnamed insurers. Shift Technology reports a recurring average recovery of over USD 1 million per month for a top 25 US property and casualty insurer, and an acceptance rate of 60% or more for a small regional insurer. Size it on your own book by auditing closed claims first."},{"question":"Does AI replace handler referrals to subrogation?","answer":"No, it adds to them. Central Insurance still asks its handlers to refer claims to its recovery team and uses the AI to catch the ones they miss, often on the first day of the claim."},{"question":"Which lines of business are the best starting point?","answer":"Auto claims, including personal injury protection and medical payments recoveries, because volumes are high and the state recovery rules can be encoded. The top 25 US insurer on this page started with auto and later extended the system to property; the system also draws on external data such as product recall lists."}],"related":["claims-triage-and-straight-through-processing","claims-fraud-detection","photo-based-damage-assessment","claims-first-notice-of-loss-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the editor after a targeted blocker check."},{"date":"2026-09-25","note":"First version, written for the insurance vertical with evidence from Central Insurance and two anonymized US insurers, verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added SEO title and description, replaced the collections recovery KPIs with revenue recovered, sharpened the EU AI Act basis and the EIOPA note, and aligned the starting line advice with the evidence."},{"date":"2026-09-26","note":"Second fact check against all sources: metaDescription now attributes the USD 1 million figure to Shift, clarified the Article 50 basis, matched the acceptance rate wording to the source and removed an unsupported claim about recall data in property claims; Central Insurance evidence no longer states a line of business the sources do not give."},{"date":"2026-09-27","note":"Removed the accuracy KPI: the small regional insurer's 60%+ acceptance rate measured alert acceptance, not correctness on a checked sample, so it is no longer filed as an accuracy metric and stays in the evidence summary and FAQ only."},{"date":"2026-09-27","note":"Added evidence from Elephant Insurance, a second named organization, grade B on its own press release; and traced the unnamed top 25 insurer's deployment year to a dated source, the case study's first Wayback capture under its earlier URL on 2024-09-20."}],"slug":"subrogation-opportunity-detection","url":"https://www.blits.ai/ai-use-cases/subrogation-opportunity-detection","benchmarks":[],"indicativeValueResult":{"low":600000,"high":5250000},"evidence":["central-insurance-subrogation-detection","elephant-insurance-subrogation-detection"]},{"title":"AI for supervisory exam and information request responses","shortTitle":"Exam response assembly","seoTitle":"AI for regulatory exam and information requests","metaDescription":"AI drafts cited answers to supervisory exam questions and tracks each commitment. The US DHS already uses generative AI to summarise incoming information requests.","definition":"An assistant for the bank's regulatory affairs team that reads a supervisory information request or exam question, retrieves the relevant evidence, policies and prior correspondence, drafts a response for legal and compliance to approve, and tracks every commitment and remediation action through to closure.","aliases":["regulatory exam response assistant","supervisory request response AI","regulator information request management","regulatory commitment tracking"],"industries":["banking","insurance","capital-markets","payments"],"functions":["regulatory-compliance","legal","case-management"],"patterns":["rag-knowledge-assistant","content-generation","document-processing","agentic-workflow"],"channels":["internal-tools","email"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"second-line","problem":"Supervisors send banks a steady flow of information requests, exam questions, thematic review\nquestionnaires and follow up letters, each with a deadline. Answering them means working out\nwhat is really being asked, finding the evidence across policies, board papers, test results and\nearlier correspondence, getting contributions from several teams and making sure the answer is\nconsistent with everything the bank has told the supervisor before.\n\nUnder time pressure, answers get assembled by email and spreadsheet, evidence is sent without a\nrecord of exactly which version went, and commitments made in a response (\"we will complete the\nreview by Q3\") are tracked by memory. Accuracy is an obligation in its own right: the FCA's SUP\n15.6, for example, requires information given to the regulator to be factually accurate or, for\nestimates and judgements, fairly and properly based after appropriate enquiries.","problemStats":[],"howItWorks":"1. **Log and interpret the request.** Each request is logged with its deadline; the assistant\n   splits it into individual questions and states what each one asks for.\n2. **Retrieve evidence.** Retrieval over policies, procedures, committee papers, test results and\n   prior regulatory correspondence returns candidate evidence and earlier answers on the topic.\n3. **Assign contributors.** Questions that need new input go to named owners with the deadline.\n4. **Draft the response.** The assistant drafts answers from the retrieved evidence, cites each\n   document and flags statements that differ from earlier submissions.\n5. **Review and sign.** Subject matter experts, then legal and compliance, edit and approve the\n   final response. Nothing leaves without sign off.\n6. **Track commitments.** Every commitment in the sent response becomes a tracked action with an\n   owner and date, and the full package (request, evidence, response) is archived.","valueDrivers":["compliance","speed","employee-productivity","risk-reduction"],"kpis":["response-time-reduction","processing-time-reduction","time-saved-per-task","accuracy"],"indicativeValue":{"referenceOrg":"A mid sized bank answering 400 supervisory questions a year","inputs":[{"key":"questions","label":"Supervisory questions and information request items per year","low":400,"high":400,"unit":"questions per year","note":"Editorial assumption for the reference bank, not a sourced figure. Replace with your own request log."},{"key":"hoursPerQuestion","label":"Staff hours per question (search, drafting, review)","low":6,"high":12,"unit":"hours per question","note":"Editorial assumption across all contributors. Replace with your own records."},{"key":"timeSaved","label":"Share of search and drafting time saved","low":0.15,"high":0.3,"unit":"fraction of time","note":"Editorial assumption. No public measured benchmark was found; review and sign off time is not reduced."},{"key":"hourlyCost","label":"Fully loaded cost of compliance and specialist staff","low":90,"high":150,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"questions * hoursPerQuestion * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Staff time released from supervisory responses","caveat":"Time only. It leaves out the value of consistent answers and fewer missed commitments, which this estimate does not price, and the cost of building the evidence repository."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The technology is standard retrieval and drafting. The difficulty is confidentiality: supervisory correspondence is often confidential supervisory information, so hosting, access control and model provider terms must be settled first.","dataPrerequisites":["A repository of prior regulatory correspondence and submissions, with versions","Current policies, procedures, committee papers and test results with owners","A request log with deadlines and owners","Rules on confidential supervisory information from the relevant supervisors"],"integrations":["Document management and regulatory correspondence systems","Governance, risk and compliance platform for issues and actions","Email and secure regulator portals (intake only, sending stays manual)","Workflow and task tools for contributors and commitments"]},"implementation":{"steps":[{"title":"Settle confidentiality first","detail":"Classify which correspondence is confidential supervisory information, check what the supervisor allows, and choose hosting and model providers accordingly before loading data."},{"title":"Build the correspondence memory","detail":"Index prior requests, responses and evidence with dates and versions, so the assistant can show what the bank has already said on a topic."},{"title":"Start with interpretation and retrieval","detail":"Use the assistant to split requests into questions and find evidence and prior answers. Measure how often experts accept its evidence before adding drafting."},{"title":"Draft with citations and consistency checks","detail":"Every drafted sentence cites a document; statements that differ from prior submissions are highlighted for the reviewer."},{"title":"Close the loop on commitments","detail":"Extract commitments from sent responses into the action tracker and report overdue items to senior management."}],"guardrails":["Legal and compliance sign every response; the assistant cannot send anything","Drafts cite the evidence they rely on; unsupported statements are flagged","Confidential supervisory information stays in approved hosting with strict access control","Full record of each request, evidence sent, response version and approver","Consistency check against earlier submissions before sign off"],"humanInTheLoop":"Subject matter experts own the content, legal and compliance approve every response, and the head of regulatory affairs owns the relationship and the commitment log. The assistant prepares, drafts and tracks.","kpisToInstrument":["Time from request receipt to approved response","Share of responses sent on or before the deadline","Reviewer edits per drafted answer and evidence acceptance rate","Commitments tracked, closed on time and overdue","Inconsistencies with prior submissions caught before sending"],"failureModes":[{"title":"Confidential information in the wrong place","detail":"Supervisory material is sent to a model provider or tool that the supervisor has not accepted. Settle hosting and terms first and block uploads elsewhere."},{"title":"Confident answers from stale evidence","detail":"The assistant cites a superseded policy or old test result. Index versions and show dates in every citation."},{"title":"Commitments lost after sending","detail":"The response is filed and the promise is forgotten. Extract commitments automatically and review them in governance."},{"title":"Over polished responses","detail":"Fluent drafts hide that the bank does not actually know the answer. Reviewers must confirm facts, not just wording."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Drafting regulatory correspondence for human approval is not an Annex III use. The main risks are confidentiality and accuracy, which are handled by supervisory information rules, data protection law and internal controls."},"regulations":["eu-ai-act","gdpr","dora","iso-42001","nist-ai-rmf"],"guidance":[{"title":"12 CFR Part 261, Rules Regarding Availability of Information","issuer":"Board of Governors of the Federal Reserve System","region":"north-america","url":"https://www.ecfr.gov/current/title-12/chapter-II/subchapter-A/part-261","note":"Example of rules that restrict disclosure of confidential supervisory information, which govern where exam material may be processed."},{"title":"PRIN 2.1, The Principles","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.handbook.fca.org.uk/handbook/PRIN/2/1.html","note":"Principle 11 requires firms to deal with regulators in an open and cooperative way and to disclose anything the regulator would reasonably expect notice of."},{"title":"SUP 15.6, Inaccurate, false or misleading information","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.handbook.fca.org.uk/handbook/SUP/15/6.html","note":"Information given to the FCA must be factually accurate or, for estimates and judgements, fairly and properly based after appropriate enquiries; a firm must notify the FCA if information it gave may have been false, misleading, incomplete or inaccurate."},{"title":"Artificial Intelligence Risk Management Framework, Generative Artificial Intelligence Profile","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence","note":"Guidance on confabulation and information security risks that apply to drafting assistants handling sensitive material."}],"controls":["Classification and access control for confidential supervisory information","Approved hosting and model provider terms documented for supervisory material","Sign off workflow with recorded approvers for every response","Commitment register reviewed by senior management","Archive of each request, evidence package and final response"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** over a **knowledge base** of policies, committee papers,\ntest results and prior regulatory correspondence, with version control and hybrid retrieval so\nexact prior wording is found. Incoming requests can arrive through the **email channel** or be\nuploaded (PDF, Word, Excel or Outlook .msg), and the agent returns **structured output**: the\nquestions, the evidence found with citations and a draft answer.\n\nAn **agentic workflow** handles multi part requests and extracts commitments into the bank's\ntask system through **custom functions**, with **human in the loop approval** before any draft\nis released to reviewers. **Tenant isolation**, role based access, **PII masking** and EU or UAE\ndata residency keep confidential material in approved boundaries, and because the platform is\nmodel agnostic the bank can pick a model and hosting that its supervisors accept."},"faq":[{"question":"Can we put supervisory correspondence into a generative AI tool?","answer":"Only within the rules on confidential supervisory information that apply to you and with hosting, access control and model provider terms your supervisors would accept. Settle this before loading any correspondence."},{"question":"Who uses AI for this today?","answer":"We found no named bank deployment in public sources. The public examples on this page come from US government bodies: the Executive Secretariat of the Department of Homeland Security has used generative AI since December 2024 to summarise incoming correspondence and information requests in its tracking system, with drafting of responses named as a future capability, and two related DHS and FEMA tools are reported as pre deployment."},{"question":"What delivers the most value first?","answer":"In our view, retrieval of prior answers and evidence, and tracking of commitments. They target inconsistent answers and missed promises, the failure modes described above, while drafting speed matters mainly for tight deadlines. No public measured comparison exists yet."}],"related":["regulatory-report-assembly","policy-drafting-and-gap-analysis","continuous-controls-testing","internal-audit-copilot","regulatory-horizon-scanning","complaints-root-cause-analysis"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog. Evidence comes from comparable government deployments because no named bank deployment was found."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: the three federal inventory entries, guidance links and platform capabilities confirmed. Softened an unsourced claim in the problem section and added an SEO title and meta description."},{"date":"2026-09-27","note":"Adversarial review fixes: evidence summaries narrowed to what the DHS and FEMA inventory entries state, DHS inventory pages added as sources, unsourced comparisons in the problem section and FAQ rewritten, and the reference volume labelled as an editorial assumption."},{"date":"2026-09-27","note":"Fact checked against sources; no changes needed. Inventory entries DHS-2342, DHS-2453 and DHS-2709, the DHS inventory pages, the three guidance links and the Blits.ai capabilities rechecked."}],"slug":"supervisory-exam-response-assembly","url":"https://www.blits.ai/ai-use-cases/supervisory-exam-response-assembly","benchmarks":[],"indicativeValueResult":{"low":32400,"high":216000},"evidence":["dhs-congressional-report-task-extraction","dhs-executive-secretariat-inquiry-summarization","fema-spend-plan-data-call-assistant"]},{"title":"AI for supplier invoice processing in accounts payable","shortTitle":"Supplier invoice processing","seoTitle":"AI invoice processing for accounts payable","metaDescription":"AI reads, matches and codes supplier invoices so AP staff only handle exceptions. Rossum reports 60% touchless extraction at Kingfisher and 8x faster work at Veolia.","definition":"AI that captures supplier invoices from any format, extracts header and line data, matches them to purchase orders and goods receipts, proposes tax and cost centre coding, flags duplicates and suspected fraud, and routes them for approval and posting, leaving only exceptions to accounts payable staff.","aliases":["AI accounts payable automation","invoice data capture","intelligent invoice processing","purchase to pay automation"],"industries":["cross-industry","banking","government","retail-and-ecommerce","energy-and-utilities"],"functions":["finance-and-accounting","procurement"],"patterns":["document-processing","agentic-workflow","anomaly-detection","classification-and-routing"],"channels":["email","internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"mainstream","segment":"back-office","problem":"Every organization pays suppliers, and in many the invoice still arrives as a PDF or paper that\nsomeone keys into the ERP. Staff then chase purchase orders and goods receipts, code the cost\ncentre and tax, and route the invoice to an approver who may take days to respond. Banks are no\nexception: their own procurement of technology, property and services runs through the same\naccounts payable process as any large company.\n\nThe costs are well documented. Manual handling is slow and expensive per invoice, late payment\nloses early payment discounts and damages supplier relationships, and duplicate payments and\ninvoice fraud (a changed bank account on a fake invoice) cause direct losses. Template based OCR\nhelped with the largest suppliers but breaks on the long tail of layouts.","problemStats":[{"statement":"Ardent Partners' State of ePayables 2025 report puts the average cost of processing an invoice at USD 9.84 and finds that the teams it ranks as Best in Class (the 20% with the lowest cost and shortest cycle time) process invoices at a 79% lower cost and 79% faster than their peers.","sourceTitle":"State of ePayables (Part Nine): AP Benchmarks and Best-in-Class Performance","sourceUrl":"https://payablesplace.ardentpartners.com/2026/01/state-of-epayables-part-nine-ap-benchmarks-and-best-in-class-performance/","year":2025}],"howItWorks":"1. **Capture.** Invoices arrive by email, supplier portal, electronic invoicing network or scanned post.\n   Structured electronic invoices are read directly; the rest go through AI extraction.\n2. **Extract and validate.** The AI extracts supplier, invoice number, dates, amounts, tax and\n   line items, and validates them against the vendor master (tax ID, bank account, currency).\n3. **Match.** It performs two or three way matching against purchase order and goods receipt,\n   within tolerances, and explains any mismatch in plain language.\n4. **Code and check.** For invoices without a purchase order it proposes the general ledger\n   account, cost centre and tax code from history and contract terms, and checks for duplicates\n   and fraud signals such as a changed bank account.\n5. **Route, approve and post.** Clean invoices post automatically within limits; others go to the\n   right approver with a summary. Approvers can ask questions in chat or email, and every step is\n   logged for audit.","valueDrivers":["cost-to-serve","speed","risk-reduction","employee-productivity"],"kpis":["automation-rate","productivity-gain","handling-time-reduction","cost-reduction","processing-time-reduction","error-reduction"],"indicativeValue":{"referenceOrg":"A bank or company that processes 300,000 supplier invoices a year","inputs":[{"key":"invoices","label":"Supplier invoices per year","low":300000,"high":300000,"unit":"invoices per year","note":"The reference organization. Replace with your own invoice volume."},{"key":"costPerInvoice","label":"Current fully loaded cost per invoice","low":6,"high":10,"unit":"USD per invoice","note":"Ardent Partners puts the average at USD 9.84 per invoice in its State of ePayables 2025 benchmarks; the low end is an editorial assumption for a team with some automation already in place.","sourceUrl":"https://payablesplace.ardentpartners.com/2026/01/state-of-epayables-part-nine-ap-benchmarks-and-best-in-class-performance/"},{"key":"costReduction","label":"Share of processing cost removed","low":0.3,"high":0.6,"unit":"fraction of cost per invoice","note":"Editorial assumption, well below the 79% cost gap between the teams Ardent Partners ranks as Best in Class and their peers."}],"formula":"invoices * costPerInvoice * costReduction","currency":"USD","period":"per year","resultLabel":"Accounts payable processing cost avoided","caveat":"Processing cost only. It leaves out captured early payment discounts, avoided duplicate and fraudulent payments, and the cost of the platform, supplier onboarding and ERP integration."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Extraction and matching are mature, off the shelf capabilities. The effort is in vendor master data quality, purchase order discipline, ERP integration and the approval rules.","dataPrerequisites":["Clean vendor master with tax IDs and verified bank accounts","Purchase orders and goods receipts in the ERP for matched spend","Historical coded invoices to learn coding for non purchase order spend","Approval matrix and limits by entity, cost centre and amount"],"integrations":["ERP (accounts payable, purchasing, general ledger)","Email inboxes, supplier portal and electronic invoicing networks","Vendor master and supplier onboarding tools","Payment and treasury systems","Approval workflow and collaboration tools"]},"implementation":{"steps":[{"title":"Measure the baseline","detail":"Record cost per invoice, touchless rate, cycle time and exception reasons by supplier and entity. Most of the value case depends on where exceptions come from."},{"title":"Fix the master data first","detail":"Duplicate suppliers, missing tax IDs and unverified bank accounts cause more exceptions than extraction errors. Clean them before tuning any model."},{"title":"Start with extraction and matching","detail":"Automate capture and purchase order matching for the largest suppliers, measure field level accuracy, and keep people validating low confidence fields."},{"title":"Add coding and fraud checks","detail":"Introduce proposed coding for non purchase order invoices and duplicate and bank account checks, with thresholds agreed with the controller."},{"title":"Automate posting within limits","detail":"Allow touchless posting only for matched invoices under a value threshold from suppliers with verified details, and widen gradually as quality holds."}],"guardrails":["Bank account changes are verified out of band before any payment, never from the invoice alone","Segregation of duties and approval limits are enforced by the ERP, not by the model","Duplicate checks run on every invoice before posting","Low confidence fields always go to a person for validation"],"humanInTheLoop":"Accounts payable staff validate low confidence extractions and resolve exceptions. Budget holders approve invoices according to the approval matrix, and the controller samples touchless postings every month.","kpisToInstrument":["Touchless rate (invoices posted without manual intervention)","Field level extraction accuracy on a weekly sample","Cost per invoice and cycle time from receipt to approval","Duplicate and fraudulent invoices caught before payment","Early payment discounts captured"],"failureModes":[{"title":"Invoice fraud through changed bank details","detail":"A fake or altered invoice redirects payment. Verify bank account changes through a separate channel and flag them automatically."},{"title":"Wrong coding at scale","detail":"A learned coding pattern posts spend to the wrong cost centre for months. Sample coded invoices and review coding drift at every close."},{"title":"Automation that moves the queue","detail":"Extraction improves but approvals remain slow, so cycle time does not change. Measure end to end, including approval time."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Processing supplier invoices is not an Annex III use case, is not a practice prohibited by Article 5 and does not involve decisions about natural persons, so it is minimal risk and the AI literacy duty of Article 4 applies. Approvers who ask questions in chat use an internal tool they know is AI; if that is not obvious to the people using it, the provider must also inform them that they are interacting with an AI system (Article 50(1))."},"regulations":["eu-ai-act","gdpr","iso-42001","dora"],"guidance":[{"title":"VAT in the Digital Age (ViDA)","issuer":"European Commission","region":"europe","url":"https://taxation-customs.ec.europa.eu/taxation/vat/vat-digital-age-vida_en","note":"The EU VAT package that lets Member States mandate electronic invoicing and, from 1 July 2030, requires digital reporting based on electronic invoices for cross border B2B supplies, which changes the capture step."},{"title":"AS 2201: An Audit of Internal Control Over Financial Reporting That Is Integrated with An Audit of Financial Statements","issuer":"Public Company Accounting Oversight Board","region":"north-america","url":"https://pcaobus.org/oversight/standards/auditing-standards/details/AS2201","note":"Applies to integrated audits of US public companies (issuers), not to every organization. In those audits, the auditor tests the controls over automated invoice processing like any other control over financial reporting."},{"title":"Digital Operational Resilience Act (Regulation (EU) 2022/2554), Article 28","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2022/2554/oj","note":"Relevant only when the deployer is a financial entity such as a bank or insurer. Its invoice capture platform and AI vendors are then ICT third party service providers, managed under the ICT third party risk principles of Article 28."}],"controls":["Extraction and coding models inventoried with an owner and monitored for accuracy","Full trail of extraction, validation, match, approval and posting per invoice","Out of band verification of supplier bank account changes","Monthly sample of touchless postings reviewed by the controller"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow**. Invoices arrive through the **email channel**,\nwhere a dialog flow triggers the workflow, or directly through an API token. An\n**AI agent** with **structured output** returns the invoice fields, and **custom functions**\nlook up the vendor master, purchase orders and goods receipts in the ERP (for example through\nthe SAP or NetSuite connections in the integration catalog) and create the posting proposal.\nContract terms and coding rules sit in the **knowledge base** with hybrid retrieval.\n\n**Human in the loop confirmation** stops the posting action above a configurable threshold for a\nperson to approve or reject. Other routing rules, such as sending every invoice with a changed\nbank account to a person, are built with **custom functions** in the workflow, and the ready\nmade **Microsoft Teams** tool can\nnotify the approver. Every run keeps a **full audit trail**, **test suites** replay a labelled\nset of invoices before each change goes live, and **monitors** run scheduled checks against the\nagent and alert on failure."},"faq":[{"question":"What does it cost to process a supplier invoice?","answer":"Ardent Partners' State of ePayables 2025 report puts the average at USD 9.84 per invoice, and the accounts payable teams it ranks as Best in Class process invoices at a 79% lower cost and 79% faster than their peers. Your own figure depends on the share of paper, purchase order coverage and approval discipline."},{"question":"What share of invoices can go through without a person?","answer":"It varies with supplier mix, master data and purchase order coverage, and few sources report a true end to end touchless rate. In its Kingfisher customer story, Rossum reports that 60% of invoices pass data extraction with no manual intervention before they reach SAP. That covers the extraction step only: exceptions are still handled in SAP, and the story does not say what share of invoices is posted without a person."},{"question":"How do you stop AI from paying fraudulent invoices?","answer":"Never let the invoice change where money goes. Verify supplier bank account changes out of band, run duplicate checks on every invoice, and keep segregation of duties and approval limits in the ERP rather than in the model."}],"related":["intelligent-document-processing","procurement-contract-review","ledger-and-payment-reconciliation","correspondence-triage-and-routing","vendor-due-diligence"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: industries now include energy and utilities, where its evidence comes from; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: the Ardent Partners figures are now attributed to its State of ePayables 2025 report, the ViDA note gives the 2030 date, the AI Act basis cites Articles 4 and 5, the Blits.ai section matches the feature inventory, and seoTitle and metaDescription were added. FDIC stage set to announced, ICE vendors and the Veolia integrator added."},{"date":"2026-09-27","note":"Review fixes: the Kingfisher 60% figure is now attributed to Rossum and limited to the data extraction step before SAP (in the FAQ, the metaDescription and the evidence, where it no longer counts as an end to end automation rate); handling time reduction added as a KPI; AS 2201 full title and issuer scope; DORA scoped to financial entities; Article 50(1) addressed; the Blits.ai section matches the workflow triggers and approval feature in the inventory."},{"date":"2026-09-27","note":"Fact checked against sources: every figure, source, guidance link and Blits.ai capability rechecked; the Kingfisher evidence gains the 80% reduction in manual work stated by its project manager."}],"slug":"supplier-invoice-processing","url":"https://www.blits.ai/ai-use-cases/supplier-invoice-processing","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":1,"median":80,"min":80,"max":80,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"veolia-ssc-invoice-processing","pooled":false},{"id":"kingfisher-ap-invoice-capture","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"kingfisher-ap-invoice-capture","pooled":true}]}],"indicativeValueResult":{"low":540000,"high":1800000},"evidence":["fdic-invoice-and-contract-data-extraction","kingfisher-ap-invoice-capture","us-immigration-and-customs-enforcement-intelligent-document-processing","veolia-ssc-invoice-processing"]},{"title":"AI for support knowledge article generation and maintenance","shortTitle":"Knowledge article generation","seoTitle":"AI knowledge base article generation for support","metaDescription":"AI drafts knowledge base articles from resolved tickets and flags gaps and stale content. NSF (ServiceNow Now Assist) and an IRS pilot use it to draft articles.","definition":"AI that drafts knowledge base articles from resolved tickets, cases and conversations, detects questions the knowledge base does not answer and articles that are outdated or contradict each other, and proposes new or revised articles for a knowledge owner to review and publish.","aliases":["knowledge base article generation","AI knowledge authoring","knowledge gap detection","case to article","knowledge centered service with AI"],"industries":["cross-industry","government","automotive"],"functions":["knowledge-management","customer-service","it-and-engineering"],"patterns":["content-generation","summarization","classification-and-routing"],"channels":["internal-tools","agent-desktop"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","problem":"Every support organization depends on its knowledge base, and almost every knowledge base is\nbehind. Agents solve a new problem, write a few lines in the ticket and move on; the fix never\nbecomes an article, so the next agent and the next customer start from scratch. Articles that do\nexist go stale when products, prices and procedures change, and nobody knows which ones, because\nchecking hundreds of articles against reality is nobody's full time job.\n\nThis matters more now than it used to. Self service portals, chatbots and AI agents all answer\nfrom the same knowledge base, so a gap or an outdated article is repeated at scale. The work AI can\ntake on is the drafting and the detection: turning resolved cases into structured first drafts,\nfinding clusters of questions with no good article, and flagging articles that conflict with\nnewer ones or have not been touched since a change. Publishing stays with a knowledge owner.","problemStats":[],"howItWorks":"1. **Mine resolved work.** When a ticket, case or chat is closed with a new or unusual resolution,\n   the AI reads the case notes, the conversation and the resolution steps.\n2. **Draft in the house format.** It writes a first draft in the organization's article template\n   (problem, environment, cause, resolution steps, related articles), without customer data.\n3. **Check for duplicates and conflicts.** The draft is compared with existing articles; the AI\n   proposes an update to an existing article instead of a new one where they overlap, and flags\n   contradictions.\n4. **Find the gaps.** Questions from search logs, chatbot conversations and tickets that retrieval\n   could not answer well are clustered and ranked by volume, each with a proposed article.\n5. **Flag stale content.** Articles are checked against release notes, policy changes and\n   feedback (thumbs down, reopened tickets), and the ones at risk are queued for their owner.\n6. **Review and publish.** A knowledge owner edits, approves and publishes; the article then feeds\n   agents, self service and AI assistants alike.","valueDrivers":["employee-productivity","cost-to-serve","customer-experience","speed"],"kpis":["time-saved-per-task","productivity-gain","containment-rate","first-contact-resolution"],"indicativeValue":{"referenceOrg":"A support organization that publishes or revises 2,000 knowledge articles a year","inputs":[{"key":"articles","label":"Articles created or substantially revised per year","low":2000,"high":2000,"unit":"articles per year","note":"The reference organization."},{"key":"hoursPerArticle","label":"Author time per article without AI","low":1.5,"high":3,"unit":"hours per article","note":"Editorial assumption covering research, writing and formatting. Replace with your own."},{"key":"timeSavedShare","label":"Share of author time saved by a reviewed AI draft","low":0.3,"high":0.5,"unit":"fraction of author time","note":"Editorial assumption. No public benchmark on this page quantifies it yet; review and testing time is kept with the author."},{"key":"authorCost","label":"Fully loaded cost of a knowledge author or senior agent","low":45,"high":70,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"articles * hoursPerArticle * timeSavedShare * authorCost","currency":"USD","period":"per year","resultLabel":"Authoring time value released","caveat":"Counts only authoring time. It leaves out the usually larger effect of a more complete, current knowledge base on self service containment, first contact resolution and new agent ramp up, and the cost of the platform and of the reviewers' time."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Drafting from a closed case is straightforward. The effort is in a clear article template, clean ownership of content, removing customer data from drafts and getting authors to review instead of rewrite.","dataPrerequisites":["Closed tickets or cases with usable resolution notes","The current knowledge base with owners, review dates and article templates","Search, chatbot and ticket logs that show unanswered or poorly answered questions","Release notes and policy change records to detect stale articles"],"integrations":["IT or customer service management platform (for example ServiceNow, Zendesk, Salesforce)","Knowledge base or content management system with an approval workflow","Search and chatbot analytics","Product release and policy change feeds"]},"implementation":{"steps":[{"title":"Fix the template and the owners first","detail":"Agree one article template per content type and assign every article and category to an owner with a review date. Drafts without an owner never get published."},{"title":"Start with case to article drafting","detail":"Trigger a draft when an agent marks a resolution as reusable. Measure how much the reviewer changes, and tune the prompt until most drafts need edits rather than rewrites."},{"title":"Strip customer data from drafts","detail":"Mask names, account numbers and environment details that identify a customer before drafting, and have the reviewer confirm nothing personal remains."},{"title":"Add gap detection","detail":"Cluster the questions that search and chatbots failed to answer, rank them by volume and propose articles for the top clusters each week."},{"title":"Add stale content checks","detail":"Compare articles with release notes and policy changes, and use negative feedback and reopened tickets as signals. Queue at risk articles for their owner rather than editing them silently."},{"title":"Close the loop with the assistants","detail":"Track whether new articles actually reduce repeat questions and improve answers from self service and AI agents, and retire articles that nobody uses."}],"guardrails":["No article is published without approval by a named knowledge owner","Customer and personal data removed from drafts before review","Drafts cite the cases and sources they were derived from, for the reviewer to check","Updates to existing articles are proposed as changes with a visible difference, never applied silently","Articles on regulated topics (fees, legal rights, safety) go through the existing compliance review"],"humanInTheLoop":"Knowledge owners approve every new article and every change, and remain accountable for the content. Agents flag which resolutions are worth an article; reviewers sample published AI drafts monthly for accuracy against the source cases.","kpisToInstrument":["Share of AI drafts published with minor edits versus rewritten or rejected","Author time per published article, before and after","Time from first occurrence of a new issue to a published article","Unanswered question clusters closed per month","Self service and first contact resolution on topics with new or refreshed articles"],"failureModes":[{"title":"Publishing the fix for one customer","detail":"A draft generalizes a workaround that only applied to one environment. Require the reviewer to confirm scope and prerequisites."},{"title":"Article sprawl","detail":"Every case becomes a new article and search gets worse. Prefer updating existing articles and merge duplicates."},{"title":"Customer data in the knowledge base","detail":"Details from the source case survive into a published article. Mask before drafting and check before approval."},{"title":"Stale content amplified by AI","detail":"Chatbots answer confidently from an outdated article. Tie review dates and stale content checks to the release process."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Drafting internal or public help content that a person reviews and publishes is not a prohibited practice under Article 5 and is not listed in Annex III, so it is minimal risk. The articles are not a direct AI interaction, and the Article 50(4) disclosure for AI generated text published to inform the public does not apply where a person reviews the text and holds editorial responsibility. The Article 50 transparency duties do apply to chatbots that later answer customers from the articles."},"regulations":["eu-ai-act","gdpr","iso-42001"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Paragraph 4 requires disclosure of AI generated text published to inform the public on matters of public interest, unless it has undergone human review or editorial control and a person holds editorial responsibility."}],"controls":["Content ownership and review dates recorded for every article","Approval workflow that records who approved each AI drafted article","Personal data masking on case data used for drafting","Periodic audit of a sample of published AI drafts against their source cases"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that the service platform triggers through an API token\nwhen a case closes: **custom functions**\nread the ticket and resolution from the service platform (ServiceNow and Zendesk are in the\nintegration catalog, Freshdesk has a ready made tool), **PII masking** strips customer data, and an **AI agent** with\n**structured output** writes the draft in your template. The workflow compares the draft with the\nexisting **knowledge base** using hybrid retrieval to propose an update instead of a duplicate, and\n**human in the loop** approval holds every draft until a knowledge owner accepts or rejects it.\n\nFor gap detection, Blits.ai **analytics** already list untrained questions and unexpected answers\nfrom live conversations, and **response feedback** with thumbs and comments shows which answers\nfailed. Approved articles go into the document library with **version control and revert**, so\nevery assistant on the platform answers from the new version, and **test suites** rerun the\nquestions that exposed the gap to confirm it is closed. The platform is model agnostic and can run\nin the EU or UAE region."},"faq":[{"question":"Can AI write knowledge base articles on its own?","answer":"It can produce complete first drafts from resolved cases, but publishing should stay with a knowledge owner. The IRS IT service desk runs a limited production pilot that generates knowledge base articles from incident and case records, and expects it to feed its existing knowledge review and publication processes."},{"question":"Where do the biggest gains come from?","answer":"Less from faster writing than from a knowledge base that keeps up: fixes captured the day they are found, gaps closed by volume, stale articles flagged. That quality then flows into every self service channel and AI agent that answers from it."},{"question":"Is this a separate tool or part of the service platform?","answer":"Often it is a feature switched on in the service platform. The U.S. National Science Foundation uses ServiceNow Now Assist to generate responses, work notes and knowledge base articles. A separate build makes sense when knowledge lives in several systems or feeds several assistants."}],"related":["enterprise-knowledge-search","email-and-ticket-reply-drafting","live-agent-assist","it-service-desk-resolution-agent","first-line-contact-centre-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical with four evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: rechecked all four evidence records, softened the IRS and NSF wording to match the inventory entries, made the EU AI Act basis and Article 50 note precise, corrected the integration wording in howToBuild, and added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: recorded the IRS deployment as a pilot and reworded the FAQ to match, limited the meta description to article drafting, set adoption to emerging, removed the technology industry, named the API token trigger in howToBuild and aligned the CDC summary with its inventory entry."},{"date":"2026-09-27","note":"Fact checked against sources: rechecked the CDC, IRS and NSF inventory entries, the Rivian entry on the Google Cloud list, the Article 50(4) text and every howToBuild capability against the feature inventory; no changes needed to the page."}],"slug":"support-knowledge-article-generation","url":"https://www.blits.ai/ai-use-cases/support-knowledge-article-generation","benchmarks":[],"indicativeValueResult":{"low":40500,"high":210000},"evidence":["cdc-smartfind-knowledge-bot","irs-service-desk-knowledge-article-generation","nsf-servicenow-now-assist-knowledge-articles","rivian-notebooklm-shared-knowledge-base"]},{"title":"AI for synthetic test data generation","shortTitle":"Synthetic test data generation","seoTitle":"Synthetic test data generation with AI","metaDescription":"AI generates realistic test data without real personal data. Patterson Dental cut test data preparation by 75%, and Merkur Versicherung cut time to data to one day.","definition":"AI that generates realistic synthetic datasets, such as customers, transactions, documents and conversations, which keep the structure and statistical properties of production data without containing real personal data, so teams can test software, train and validate models and run demos safely.","aliases":["synthetic data generation","privacy safe test data","synthetic test data management","AI generated test data"],"industries":["cross-industry","banking","healthcare","insurance"],"functions":["it-and-engineering","analytics-and-reporting"],"patterns":["synthetic-data-generation","content-generation"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Every new system, release and model needs data that looks like the real thing: customers with\nplausible histories, transactions with realistic patterns, claims, statements and conversations.\nThe easy answer is to copy production into test environments, which puts personal and confidential\ndata in places with weaker controls, in the hands of contractors and offshore teams, and in the\ntraining sets of models. Some organizations have sealed production off from development\naltogether, as Kin Insurance did, and data protection law makes such copies hard to justify.\n\nThe alternatives each have a cost. Hand written mock data is safe but thin, so tests pass that\nwould fail on real data. Masking production data keeps realism but can still leave people\nidentifiable, and preparing a usable copy by hand is slow: Patterson Dental needed 2.5 hours per\ndataset before it automated the work. Rare but important\ncases, such as fraud patterns or unusual products, are often missing entirely. Synthetic data\ngeneration aims to give teams realistic, referentially consistent data on demand, with a measured\nprivacy risk.","problemStats":[],"howItWorks":"1. **Profile the source.** A generator learns the schema, relationships and statistical\n   distributions of a production dataset inside the secure environment, or works from a schema and\n   business rules when no production data may be used.\n2. **Generate.** It produces new records (customers, accounts, transactions, claims, documents or\n   conversation transcripts) that follow the same patterns and keep keys consistent across tables,\n   with the option to add rare cases such as fraud or edge conditions on purpose.\n3. **Measure privacy, utility and fidelity.** Automated checks confirm that no generated record\n   copies a real one, test for reidentification risk and compare the synthetic statistics with the\n   real ones.\n4. **Approve and publish.** A data owner or privacy officer signs off each new dataset\n   configuration; routine refreshes then run on a schedule or through an API into test, analytics\n   or sandbox environments.\n5. **Use and monitor.** Teams test, train or demo on the synthetic data, and results that matter\n   (a model going live, a performance benchmark) are confirmed on real data under proper controls.","valueDrivers":["speed","risk-reduction","compliance","employee-productivity"],"kpis":["processing-time-reduction","cycle-time-days","users-served","productivity-gain"],"indicativeValue":{"referenceOrg":"A bank with 30 delivery teams that request test data for their environments","inputs":[{"key":"requests","label":"Test data requests per year","low":200,"high":400,"unit":"requests per year","note":"Editorial assumption for about 30 teams refreshing environments every few weeks. Replace with your own ticket volume."},{"key":"hoursPerRequest","label":"Engineering hours to prepare one masked or hand built dataset","low":4,"high":12,"unit":"hours per request","note":"Editorial assumption covering extraction, masking, fixing broken references and checks. Patterson Dental reports 2.5 hours per dataset before automation."},{"key":"timeSaved","label":"Share of preparation time removed","low":0.5,"high":0.75,"unit":"fraction of hours","note":"Capped at the benchmark on this page (Patterson Dental reports a 75% reduction in test data generation time)."},{"key":"hourlyCost","label":"Fully loaded cost of an engineer hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption. Replace with your own blended rate."}],"formula":"requests * hoursPerRequest * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Engineering time released from test data preparation","caveat":"Counts only the preparation effort. It leaves out licence and compute cost, the risk reduction of removing personal data from lower environments, faster releases and fewer defects found late."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Generating a single table is easy. The work is in multi table data with referential integrity, business rules that must hold (a closed account has no new transactions), privacy measurement a privacy officer will accept, and pipelines that keep the synthetic data in step with schema changes.","dataPrerequisites":["Access to the source data inside a controlled environment, or a complete schema with business rules","A data classification that marks personal, special category and confidential fields","Agreed privacy, utility and fidelity thresholds per use (testing, model training, external sharing)","Examples of the rare cases and edge conditions tests must cover"],"integrations":["Source databases and data warehouse for profiling","Target test, sandbox and analytics environments","CI pipelines and environment provisioning, so data refreshes run with deployments","Data catalogue for lineage and approval records"]},"implementation":{"steps":[{"title":"Start where production copies hurt most","detail":"List the environments and teams that still receive production or lightly masked data, and pick one with a clear pain, such as a performance test that needs volume or a supplier that should never see real customers."},{"title":"Decide the privacy standard before generating anything","detail":"Agree with the privacy officer which checks a dataset must pass (no copied records, distance to closest real record, attribute inference tests) and who signs off. Treat the generator itself as processing of personal data, because it learns from it."},{"title":"Model the relationships, not just the columns","detail":"Map keys and business rules across tables and add them as constraints, so tests exercise real journeys. A customer without accounts or a claim without a policy produces false failures."},{"title":"Add the cases production lacks","detail":"Deliberately oversample fraud patterns, edge values, long names, rare products and multiple languages. This is where synthetic data beats a production copy."},{"title":"Automate refresh and measure use","detail":"Run generation from the pipeline on a schedule or per environment, and track requests, time to data and the defects found with synthetic data versus escaped defects."},{"title":"Keep real data for the final proof","detail":"For model training and validation, compare models trained on synthetic, mixed and real data before relying on synthetic data alone, and confirm go live decisions on real data under controls."}],"guardrails":["Generation runs inside the secured data environment; only the approved synthetic output leaves it","Every dataset passes automated privacy checks for copied records and reidentification risk before release","Stricter thresholds, or rule based generation without production data, for datasets shared outside the organization","Labels on every synthetic dataset so it is never mistaken for, or merged with, real data","Model decisions that affect customers are validated on real data, not on synthetic data alone"],"humanInTheLoop":"A data owner and the privacy officer approve each new dataset configuration and each external release, and review the privacy report. Test and data science leads confirm that the data is fit for purpose, and a person decides when a result on synthetic data is strong enough to act on.","kpisToInstrument":["Time from data request to usable test data","Share of test environments that hold no production personal data","Privacy test results per release (copied records, closest record distance)","Fidelity scores against the source statistics per dataset","Defects found in test versus defects that escape to production"],"failureModes":[{"title":"Synthetic data that leaks real people","detail":"Overfitted generators can reproduce real records or rare outliers. Measure copies and closest records on every release and use differential privacy or stricter settings for sensitive data."},{"title":"Realistic columns, broken journeys","detail":"Distributions match but relationships and business rules do not, so tests fail for the wrong reasons or pass for the wrong reasons. Encode constraints and test the data itself."},{"title":"Models trained on synthetic data alone","detail":"Accuracy can drop sharply when real data is removed entirely. Keep a share of real data or validate on real data before deployment."},{"title":"A second shadow copy of production","detail":"Synthetic datasets proliferate without owners or labels. Register them in the catalogue with purpose, lineage and expiry."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"A generator of synthetic tabular test data is not listed in Annex III and does not interact with people, so it is minimal risk with only the AI literacy duty of Article 4. When the system generates synthetic text, images, audio or video, such as documents or conversation transcripts, Article 50(2) requires its provider to mark the output in a machine readable format as artificially generated. When synthetic data is used to train, validate or test a high risk system, such as credit scoring, it falls under that system's data governance duties in Article 10."},"regulations":["gdpr","uk-gdpr","eu-ai-act","hipaa","dora"],"guidance":[{"title":"What PETs are there? Synthetic data","issuer":"UK Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/data-sharing/privacy-enhancing-technologies/what-pets-are-there/synthetic-data/","note":"Guidance on synthetic data as a privacy enhancing technology. It notes that generating it from real data may involve processing personal information, that closer resemblance to real data raises the chance of revealing someone's information, and that biases carry through. The page says it is under review after the Data (Use and Access) Act."},{"title":"Using Synthetic Data in Financial Services","issuer":"Financial Conduct Authority, Synthetic Data Expert Group","region":"europe","url":"https://www.fca.org.uk/publication/corporate/report-using-synthetic-data-in-financial-services.pdf","note":"March 2024 report with use cases on system testing, model validation and data sharing, and practical advice on evaluating privacy, utility and fidelity."}],"controls":["Data protection impact assessment for the generator and its training data","Documented privacy, utility and fidelity thresholds with sign off per dataset","Access control and logging on the environment where the generator sees real data","Catalogue entries with lineage, purpose and expiry for every synthetic dataset","Contract terms that forbid suppliers from attempting reidentification"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai is not a tabular data synthesizer, but it is a practical way to generate synthetic\n**conversations and documents** and to orchestrate a dedicated generator. An **agentic workflow**\nwith **structured output** asks an agent to produce synthetic customers, transcripts, emails or\ncase files that follow a schema and a set of business rules, and **custom functions** call the\nAPI of the synthetic data engine the organization already uses for large tables. A **SQL\nknowledge base** lets an agent read aggregate statistics and schema information to steer\ngeneration, and **human in the loop confirmation** lets a data owner approve or reject each new\ndataset before the workflow publishes it.\n\n**PII masking** at the gateway keeps real personal data out of prompts, and the platform is\nmodel agnostic with **EU and UAE data residency**, so generation can run in the required region.\nSynthetic conversations can be imported as JSON test cases into **test suites** to evaluate chat\nand voice agents before launch, and **monitors** run scheduled checks against the live agents\nafterwards."},"faq":[{"question":"Is synthetic data still personal data under GDPR?","answer":"It can be. The FCA's Synthetic Data Expert Group notes that synthetic data tends to have a lower privacy risk than real data but cannot guarantee privacy, so a risk assessment per implementation is needed. Generation from real records is itself processing of personal data, and outputs should be tested for copied records and reidentification before release."},{"question":"Can synthetic data replace real data for training models?","answer":"Not always. In a proof of concept reported by the FCA's Synthetic Data Expert Group, a model trained on half synthetic and half real data scored 2.5% below the accuracy of the real data benchmark, while one trained on synthetic data only scored 32% below it. For software testing the bar is lower, because privacy and utility matter more than exact fidelity."},{"question":"How is synthetic data different from masked production data?","answer":"Masking keeps real records and replaces identifying fields, so relationships stay intact but people can sometimes still be identified from what remains. Synthetic data creates new records from learned patterns or rules. Many teams combine the two, as Kin Insurance did with subsetting, masking and differential privacy."},{"question":"How much faster does test data get?","answer":"The published cases report large gains in time to data. Patterson Dental reports test data generation falling from 2.5 hours to 35 minutes per dataset, and Merkur Versicherung reports time to data falling from one month to one day with a daily automated pipeline. Both are vendor case studies."}],"related":["developer-coding-assistant","legacy-code-modernization","requirements-to-test-case-generation","model-risk-validation-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from the federal AI use case inventory, bank, regulator and vendor sources and verified against them."},{"date":"2026-09-26","note":"Fact checked against sources. Replaced an unsourced claim about bans on production copies and the \"hours or weeks\" preparation time with sourced examples (Kin Insurance, Patterson Dental), changed the EU AI Act tier to context dependent to cover Article 50(2) marking of synthetic text, pointed the ICO guidance at its synthetic data page, aligned the Blits.ai section with the feature inventory, tightened the FCA accuracy figures and added SEO title and description."}],"slug":"synthetic-test-data-generation","url":"https://www.blits.ai/ai-use-cases/synthetic-test-data-generation","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":1014,"min":28,"max":2000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"boomi-synthetic-data-environment","pooled":true},{"id":"financial-conduct-authority-synthetic-data-sandbox","pooled":true}]},{"kpi":"cycle-time-days","label":"Cycle time","unit":"days","aggregate":false,"higherIsBetter":false,"n":1,"nUpTo":0,"median":1,"min":1,"max":1,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"merkur-versicherung-synthetic-health-data","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":75,"min":75,"max":75,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"patterson-dental-deidentified-test-data","pooled":true}]}],"indicativeValueResult":{"low":24000,"high":360000},"evidence":["boomi-synthetic-data-environment","financial-conduct-authority-synthetic-data-sandbox","irs-synthetic-data-engine","jpmorgan-synthetic-data-research","kin-insurance-masked-test-data","merkur-versicherung-synthetic-health-data","patterson-dental-deidentified-test-data"]},{"title":"AI for tax compliance risk scoring and audit selection","shortTitle":"Tax compliance risk scoring","seoTitle":"AI for tax audit selection and risk scoring","metaDescription":"The IRS, HMRC and the Belastingdienst use models and rules to choose which tax returns to check. How risk scoring works, how to value it and how to keep it fair.","definition":"Models that score tax returns, taxpayers and transactions for the risk of error, underreporting or fraud, so that a tax administration spends its audit and compliance capacity where the risk is highest, with an officer deciding every compliance action and the selection itself monitored for fairness.","aliases":["AI audit selection","tax risk scoring","tax return risk assessment","case selection for tax audits"],"industries":["government"],"functions":["risk-management","case-management","fraud-prevention"],"patterns":["prediction-and-scoring","anomaly-detection"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"assist","adoptionStage":"early-adopters","problem":"Tax administrations can examine only part of the returns they receive, so the question\nof which returns to open decides both how much revenue is protected and who carries the burden of\nan audit. Traditional selection relies on fixed rules, random samples and the judgment of\nexperienced staff. That misses new schemes, sends officers to returns that turn out to be correct\n(so called no change audits) and struggles with complex taxpayers, such as large partnerships,\nwhere the risk is spread over many entities and schedules.\n\nMore data now supports model based selection: third party reporting, electronic invoicing, bank and\ncross border information exchange. It has also shown the risk. A selection model that is accurate\non average can still concentrate audits on particular groups, and the Dutch childcare benefits\nscandal, in which the tax administration's risk classification used nationality, made fairness,\ntransparency and human review preconditions rather than extras.","problemStats":[{"statement":"The IRS projects an annual gross tax gap of USD 696 billion for tax year 2022, of which USD 539 billion comes from tax understated on timely filed returns.","sourceTitle":"IRS: The tax gap","sourceUrl":"https://www.irs.gov/statistics/irs-the-tax-gap","year":2024}],"howItWorks":"1. **Assemble the risk picture.** Returns are joined with third party data the administration\n   already holds: employer and bank reporting, invoices, customs data, prior audit results and\n   information exchanged with other countries.\n2. **Score.** Anomaly detection flags returns that break a taxpayer's own pattern or deviate from\n   peers; supervised models trained on past audit outcomes estimate the likelihood and size of a\n   correction; business rules encode known risks. The approaches can be combined; the\n   Belastingdienst VAT signal model on this page, for example, is purely rules based.\n3. **Explain the signal.** Each score comes with the features and rules that drove it, so an\n   officer can see why a return was flagged and challenge it.\n4. **Select with people.** Risk teams turn scores into case lists, mixed with a random sample that\n   keeps measuring the unflagged population, and an officer decides whether to open a check.\n5. **Learn and monitor.** Audit results feed back into the model; selection rates and outcomes are\n   compared across groups to detect disparate impact before it becomes a scandal.","valueDrivers":["risk-reduction","employee-productivity","compliance","cost-to-serve"],"kpis":["detection-rate-improvement","false-positive-reduction","users-served","interactions-handled","accuracy"],"indicativeValue":{"referenceOrg":"A national tax administration that completes 10,000 desk and field audits a year","inputs":[{"key":"audits","label":"Audits completed per year","low":8000,"high":12000,"unit":"audits per year","note":"Editorial assumption. Replace with your own audit volume."},{"key":"yieldPerAudit","label":"Average additional tax assessed per audit","low":5000,"high":15000,"unit":"EUR per audit","note":"Editorial assumption. Replace with your own average yield, including audits that end with no change."},{"key":"uplift","label":"Relative increase in yield from better selection","low":0.05,"high":0.15,"unit":"fraction of yield","note":"Editorial assumption, deliberately modest. The public evidence on this page reports usage and selection practice, but no audited tax yield uplift."}],"formula":"audits * yieldPerAudit * uplift","currency":"EUR","period":"per year","resultLabel":"Additional tax assessed from the same audit capacity","caveat":"Assessed tax, not collected tax. It leaves out collection losses, appeals, the cost of building and assuring the models, the deterrence effect and the benefit to compliant taxpayers of fewer no change audits."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The data usually exists; the difficulty is governance. Selection models touch taxpayer rights, need a legal basis for each data source, must be explainable to officers and courts, and need fairness monitoring that often lacks the protected characteristic data to do it well.","dataPrerequisites":["Historical audit outcomes, including no change results, linked to returns","Third party information returns and invoices, with a documented legal basis for each","A random audit or sample programme to measure the unflagged population"],"integrations":["Return processing and taxpayer account systems","Case management for compliance checks and audits","Data warehouse with third party and exchange of information data","Analytics workspace for risk teams and model monitoring"]},"implementation":{"steps":[{"title":"Start where the data is richest and the harm is lowest","detail":"Begin with business taxes such as VAT, where invoices and returns give a strong signal and the population is mostly companies, before models touch individuals and families."},{"title":"Keep a random sample running","detail":"Reserve part of the audit capacity for random selection. It is the only way to measure what the model misses and to prove it beats the old approach."},{"title":"Put explanation in front of the officer","detail":"Show the reasons for a flag next to the return. HMRC's VAT tool, for example, shows expected against observed values for each return period, so the officer judges the anomaly."},{"title":"Test for disparate impact before and after go live","detail":"Compare selection rates and hit rates across groups you can measure, including proxies such as income band and region, and document how you will act on a gap."},{"title":"Publish what you run","detail":"Register every selection model with its purpose, data and human oversight. The Dutch Belastingdienst publishes its selection and signal models in the national algorithm register."}],"guardrails":["No protected characteristic, or obvious proxy such as nationality, as a model feature","A score never triggers an assessment or penalty on its own; an officer decides every action","Random sample alongside model selection to measure performance and fairness","Documented legal basis for every data source used in scoring","Model changes reviewed by an independent validation function before use"],"humanInTheLoop":"Risk analysts own the models and the case lists; officers decide whether to open a check and carry out every compliance action. Taxpayers keep the normal review and appeal rights, and a fairness review of selection outcomes is reported to senior management at least yearly.","kpisToInstrument":["Hit rate (share of selected cases with a correction) against random selection","No change rate of audits, before and after","Additional tax assessed per audit hour","Selection and hit rates across measurable groups","Share of flags overridden by officers, with reasons"],"failureModes":[{"title":"Discriminatory selection","detail":"A model trained on past audits learns past bias, or uses a proxy such as nationality. The Dutch childcare benefits scandal and the IRS earned income tax credit disparities show the harm. Remove proxies and monitor outcomes by group."},{"title":"Feedback loops","detail":"The model only learns from returns it selected, so it keeps finding the same risks. Random audits break the loop."},{"title":"Unexplainable flags","detail":"Officers who cannot see why a return was flagged either ignore the model or trust it blindly. Show the drivers of each score."},{"title":"Blacklists that outlive their purpose","detail":"Risk signals stored about individuals and never removed can follow people for years. Set retention and review rules for every signal list."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Risk selection for administrative tax audits is not listed in Annex III, and Recital 59 says systems used by tax and customs authorities in administrative proceedings should not be treated as high risk law enforcement systems. Use in criminal tax investigations (Annex III point 6, law enforcement), or evaluating the eligibility of natural persons for public assistance benefits run through the tax system (Annex III point 5(a)), can make it high risk. When individuals are scored in administrative tax work, the GDPR applies, including its profiling rules (Member States may restrict some rights for taxation matters under Article 23). Article 22 applies when a decision with legal or similarly significant effect is taken solely by the model. Criminal investigations fall outside the GDPR and under the Law Enforcement Directive (EU) 2016/680 instead."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Regulation (EU) 2024/1689 (AI Act), Recital 59","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Systems intended for administrative proceedings by tax and customs authorities should not be classified as high risk law enforcement systems."},{"title":"Algoritmeregister van de Nederlandse overheid","issuer":"Government of the Netherlands","region":"europe","url":"https://algoritmes.overheid.nl/nl","note":"The Dutch national algorithm register, where the Belastingdienst and other agencies publish their selection and risk models with purpose, method and human oversight."},{"title":"Algorithmic Transparency Recording Standard hub","issuer":"UK government","region":"europe","url":"https://www.gov.uk/government/collections/algorithmic-transparency-recording-standard-hub","note":"UK public bodies, including HMRC, publish transparency records for algorithmic tools used in compliance work."}],"controls":["Register entry per model with purpose, data, owner and human oversight","Fairness monitoring of selection and hit rates, reported at least yearly","Explanation of each flag available to the officer and, on request, in disputes","Independent model validation before first use and after material change","Retention limits on risk signals held about individuals"],"incidents":[{"title":"Amnesty International: Xenophobic machines: Discrimination through unregulated use of algorithms in the Dutch childcare benefits scandal","url":"https://www.amnesty.org/en/wp-content/uploads/2021/10/EUR3546862021ENGLISH.pdf","note":"The Dutch tax authorities used a risk classification model in which nationality counted as a risk factor when checking childcare benefit applications, contributing to wrongful fraud accusations against parents."},{"title":"Stanford SIEPR: Measuring and mitigating racial disparities in tax audits","url":"https://siepr.stanford.edu/publications/working-paper/measuring-and-mitigating-racial-disparities-tax-audits","note":"Researchers estimate that, despite race blind selection, Black taxpayers were audited at 2.9 to 4.7 times the rate of non Black taxpayers, driven mainly by audits of earned income tax credit claims."}]},"blitsAi":{"howToBuild":"Blits.ai does not replace a tax administration's scoring models; it builds the work around them.\nAn **agentic workflow** can pick up a flagged return, pull the return and the third party data\nthrough **custom functions** (REST calls and SQL queries), and prepare a case summary with the\ndrivers of the flag for the officer, stopping for **human in the loop approval** before anything\nis sent to the taxpayer. A **knowledge base** with hybrid retrieval over the audit manual and\nguidance lets officers ask how a risk should be handled.\n\nEvery run keeps a full audit trail, **PII masking** runs at the gateway before text reaches a\nmodel, and **test suites** check the case summaries against known cases on every change. The\nplatform is **model agnostic** and offers EU and UAE data residency, which matters for taxpayer\ndata that must stay in the country or region."},"faq":[{"question":"Do tax administrations really use AI to choose audits?","answer":"Yes, in varying forms. The IRS announced in 2023 that machine learning helped select large partnership returns for examination, HMRC gives around 5,500 VAT officers an anomaly detection tool, and the Dutch Belastingdienst publishes its selection models, such as a rules based VAT signal model for large businesses, in the national algorithm register."},{"question":"Is audit selection by AI high risk under the EU AI Act?","answer":"Usually not for administrative tax audits, which Recital 59 keeps out of the law enforcement category. It can become high risk when used in criminal investigations or to decide on public benefits. When individuals are scored, GDPR profiling rules apply, with Article 22 covering any decision the model takes on its own."},{"question":"How do you keep selection fair?","answer":"Exclude protected characteristics and proxies, keep a random sample, and compare selection and hit rates across groups every year. The childcare benefits scandal and research on US earned income tax credit audits show what happens without these checks."}],"related":["benefit-fraud-and-error-detection","inspection-prioritization","tax-questions-and-filing-assistant","ai-model-inventory"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version for the government vertical, with IRS, HMRC and Belastingdienst evidence and known incidents verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed an unsupported 2.5 times detection figure from the value notes, named the Annex III points in the EU AI Act basis, set the IRS evidence to pilot and grade B, corrected the Belastingdienst summary (the model also takes some decisions itself) and added SEO title and description."},{"date":"2026-09-27","note":"Corrected the GDPR statement in the EU AI Act basis and FAQ (Article 22 only for decisions taken solely by the model, Article 23 restrictions for taxation, criminal investigations under the Law Enforcement Directive), linked Recital 59 to EUR-Lex and the Amnesty report PDF, and tightened the description."},{"date":"2026-09-27","note":"Fact checked against sources again (IRS release and tax gap page, HMRC record, algorithm register, EUR-Lex, SIEPR, Amnesty): replaced an unsourced claim that many administrations combine all three scoring methods; everything else confirmed."}],"slug":"tax-compliance-risk-scoring","url":"https://www.blits.ai/ai-use-cases/tax-compliance-risk-scoring","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1500,"min":1500,"max":1500,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"hmrc-vat-return-analysis-tool","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":5500,"min":5500,"max":5500,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"hmrc-vat-return-analysis-tool","pooled":true}]}],"indicativeValueResult":{"low":2000000,"high":27000000},"evidence":["belastingdienst-vat-signal-model-large-businesses","hmrc-vat-return-analysis-tool","irs-large-partnership-audit-selection"]},{"title":"AI for telecom churn prediction and retention offers","shortTitle":"Churn prediction and retention","seoTitle":"AI for telecom churn prediction and retention","metaDescription":"Telecom AI scores churn risk and picks a retention action. Pega reports 20% less churn at Telenet and 15% in Etisalat's SMB unit, with no published control group.","definition":"AI for telecom operators that scores each subscriber's risk of leaving from usage, service, billing and contact signals, explains the likely reason, and chooses the next best retention action, such as fixing a problem, adjusting a plan or making an offer, delivered through the app, messaging, an agent or an advisor within approved offer budgets.","aliases":["churn prediction model","retention next best action","save desk assistant","customer retention AI","proactive retention"],"industries":["telecommunications"],"functions":["marketing","customer-service","sales","analytics-and-reporting"],"patterns":["prediction-and-scoring","recommendation-and-personalization","conversational-agent"],"channels":["agent-desktop","mobile-app","sms","email","voice"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"mainstream","segment":"middle-office","problem":"In mature mobile and broadband markets, growth often depends on keeping customers as much as on\nwinning new ones. Customers leave for a better price, after repeated faults, after a bill shock or\nat the end of a contract, often without contacting the operator first. Traditional retention reacts\nonly when the customer calls to cancel, when the decision is already made, and relies on a save\ndesk that offers the same discount to everyone.\n\nBlanket discounts are expensive and can teach customers that threatening to leave pays. What operators need is\nearlier warning, a reason for the risk, and an action that fits: fixing the fault that annoyed\nthe customer can be worth more than any discount. At the same time, regulators fine operators\nwho make leaving hard (Ofcom fined Virgin Media £28 million in July 2026), so retention has to help customers, not trap them.","problemStats":[],"howItWorks":"1. **Score the risk.** A model scores every customer regularly, and in real time on events\n   such as a failed repair or a price rise, from usage, network quality, billing, contact history\n   and contract end dates.\n2. **Explain the reason.** For each high risk customer it gives the main drivers (repeated\n   faults, a better competitor price, a bill shock) so the action matches the cause.\n3. **Choose the next best action.** A decisioning engine picks the action with the best value for\n   customer and operator within budget: resolve the fault, move to a better fitting plan, a\n   loyalty benefit, or an offer, and sometimes no action at all.\n4. **Deliver it in the right channel.** The action appears in the app, in a message, as a prompt\n   to the advisor on the next call, or in a conversation with an AI agent, subject to consent.\n5. **Learn from the outcome.** Accepted and ignored offers and actual churn feed back into the\n   models, measured against a control group.","valueDrivers":["revenue-growth","customer-experience","cost-to-serve"],"kpis":["churn-reduction","conversion-rate-uplift","revenue-uplift","customer-satisfaction"],"indicativeValue":{"referenceOrg":"An operator with 2 million postpaid mobile and broadband customers","inputs":[{"key":"customers","label":"Postpaid customers","low":2000000,"high":2000000,"unit":"customers","note":"The reference operator."},{"key":"churnRate","label":"Annual churn rate","low":0.12,"high":0.18,"unit":"fraction of customers per year","note":"Editorial assumption, replace with your own annual postpaid churn."},{"key":"churnReduction","label":"Relative reduction in churn among customers the programme reaches","low":0.03,"high":0.08,"unit":"fraction of churn","note":"Conservative against the benchmarks on this page (Pega reports a 20% churn reduction at Telenet and 15% at Etisalat, a figure it first reported for Etisalat's SMB division), because those are vendor figures without a published control group."},{"key":"annualRevenue","label":"Annual revenue per customer","low":250,"high":400,"unit":"USD per customer per year","note":"Editorial assumption, replace with your own average revenue per postpaid account."},{"key":"retainedMarginShare","label":"Share of retained revenue kept after the cost of retention offers","low":0.5,"high":0.7,"unit":"fraction of retained revenue","note":"Editorial assumption covering discounts and benefits given to retained customers."}],"formula":"customers * churnRate * churnReduction * annualRevenue * retainedMarginShare","currency":"USD","period":"per year","resultLabel":"First year revenue retained, net of retention offers","caveat":"Counts only one year of retained revenue net of offer costs. It leaves out the lifetime value of retained customers, savings from fixing root causes, the cost of models and decisioning, and customers who would have stayed anyway, which only a control group can remove."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Churn models are well understood; the hard parts are joining network, billing and contact data per customer in near real time, running a decisioning engine with offer budgets, and delivering actions in every channel with consent and a measurable control group.","dataPrerequisites":["At least two years of customer history with churn outcomes and reasons","Network quality, fault and repair data linked to customers","Billing events, price changes and contract end dates","Contact history and complaint records","Marketing consent and contact preferences"],"integrations":["Data platform joining network, billing, CRM and contact data","Decisioning or next best action engine with offer rules and budgets","Agent desktop and store systems for advisor prompts","App, messaging and email channels for digital actions","Campaign management with consent and contact frequency rules"]},"implementation":{"steps":[{"title":"Define churn and the outcomes you will measure","detail":"Agree on what counts as churn (port out, cancellation, downgrade) and set up a permanent control group before the first model goes live."},{"title":"Build reasons, not only scores","detail":"Train the model to expose its main drivers per customer, so actions can address the cause. A fault driven risk needs a fix, not a discount."},{"title":"Put actions under a decisioning engine with budgets","detail":"Define the available actions, eligibility and cost, and let the engine choose within budget, including the option to do nothing for customers who will stay anyway."},{"title":"Bring it into the conversation","detail":"Show the reason and recommended action to advisors on every call, and give AI agents the same recommendation, so a customer who says they want to leave gets a relevant answer."},{"title":"Keep cancellation easy","detail":"Design retention so a customer who still wants to leave can do so in the same conversation. Treat that as a hard requirement, tested like any other journey."}],"guardrails":["A customer who confirms they want to leave is helped to leave in the same contact","Offers only within approved budgets and eligibility rules","Marketing consent and contact frequency limits enforced before any outbound action","Protected characteristics and proxies excluded from features, with regular fairness checks","Vulnerable customers routed to trained people, not to automated offers"],"humanInTheLoop":"Retention advisors decide on offers above standard limits and handle vulnerable customers. The commercial team owns the action catalogue and budgets, and an analytics team reviews model performance, fairness and the control group results monthly.","kpisToInstrument":["Churn in treated customers versus the control group","Offer acceptance and cost per retained customer","Share of high risk customers whose root cause was fixed","Time to complete a cancellation for customers who still want to leave","Complaints about retention contacts or cancellation"],"failureModes":[{"title":"Paying customers who would have stayed","detail":"Offers go to customers with a high score who were never leaving. Use uplift models and a control group."},{"title":"Making it hard to leave","detail":"Retention becomes obstruction and a regulatory breach. Measure and protect the cancellation path."},{"title":"Discount addiction","detail":"Customers learn that threatening to leave gets a discount. Favour fixing root causes and limit repeat offers."},{"title":"Unfair outcomes","detail":"Better offers go systematically to some groups. Exclude protected attributes and proxies and audit outcomes."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Churn scoring and offer selection for marketing are not listed in Annex III, so a back office design that only scores customers and prompts human advisors is minimal risk, with no specific obligations. When an AI agent delivers the offer to the customer in chat, messaging or voice, the system is limited risk: Article 50 requires telling customers they are dealing with AI. A design that used manipulative techniques or exploited vulnerabilities to keep customers from leaving could fall under the Article 5 prohibitions. GDPR rules on profiling and the right to object to direct marketing (Article 21) apply in full."},"regulations":["eu-ai-act","gdpr","telecom-consumer-rules","eecc"],"guidance":[{"title":"Ofcom fines Virgin Media £28m for repeatedly preventing customers from cancelling contracts","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/phones-and-broadband/switching-provider/ofcom-fines-virgin-media-28m-for-repeatedly-preventing-customers-from-cancelling-contracts","note":"A £28 million fine (July 2026) for retention practices, including a two tier cancellation process and agents rewarded for deterring cancellations, that caused customers unreasonable effort when trying to leave; the line any AI retention design must not cross."},{"title":"Directive on privacy and electronic communications (Directive 2002/58/EC)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2002/58/oj","note":"Article 13 sets the rules for unsolicited electronic marketing, which apply to proactive retention offers by messaging, email or automated calls. Automated calls need prior consent; under the Article 13(2) soft opt in, an operator may email or message its existing customers about its own similar products or services, provided they can object free of charge with every message."}],"controls":["Model inventory entry with owner, features, validation and fairness results","Permanent control group and monthly incrementality reporting","Documented action catalogue with budgets and approval","Consent, frequency and vulnerability checks enforced in the decisioning engine","Audit of cancellation journeys for unreasonable barriers"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the churn score and next best action usually come from the operator's own models or\ndecisioning engine; Blits.ai delivers the conversation. An **AI agent** reads the risk reason and\nrecommended action through a **custom function**, answers questions from a **knowledge base**\nof approved offers and terms, and completes a plan change or offer acceptance through a\n**flow** with explicit confirmation. **Agentic workflows**, run on a schedule or triggered\nthrough the API, can prepare retention actions and hand them off through a custom function to\nthe operator's messaging platform, with **human in the loop approval** for offers above a\nthreshold.\n\nThe agent works on **WhatsApp, SMS, email and voice**, and inside the operator's own app through\nthe **API channel**, and **human handover** passes customers who want to leave, or who seem\nvulnerable, to a trained advisor with the context. **Guardrails** with policies written by the\noperator check answers before they reach the customer, **test suites** evaluate retention\nconversations before release, **analytics** tracks them in production, and the platform is\nmodel agnostic with EU and UAE data residency."},"faq":[{"question":"How much can AI reduce telecom churn?","answer":"The public figures are vendor reported: Pega reports a 20% reduction in churn at Telenet, and a 15% reduction at Etisalat that it first published for Etisalat's small and medium business division. Neither comes with a published baseline or control group, so plan conservatively and measure against your own holdout."},{"question":"Is it legal to use AI to persuade customers to stay?","answer":"Yes, if it helps rather than obstructs. Offers must respect marketing consent and profiling rules, and a customer who wants to leave must be able to. In July 2026 Ofcom fined Virgin Media £28 million for retention practices that made cancelling unreasonably hard."},{"question":"Where should retention actions be delivered?","answer":"Wherever the customer is: as a prompt to the advisor on a call (Virgin Media O2 piloted its Lumi AI advisor prompts with care, telesales and retentions teams in 2025), in the app, or in a conversation with an AI agent that has the same recommendation."}],"related":["plan-upgrade-and-sales-assistant","bill-explanation-and-billing-dispute-agent","personalized-marketing-at-scale","insurance-renewal-and-retention","customer-feedback-analysis"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the telecommunications vertical from Telenet, Etisalat, Vodafone UK and Virgin Media O2 sources checked word for word."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added European Electronic Communications Code to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: dated the Ofcom fine (July 2026), described the Virgin Media O2 Lumi AI deployment as a pilot, added the Article 5 limit to the EU AI Act basis, limited the Blits.ai channels and guardrail wording to documented capabilities, and added seoTitle and metaDescription. Evidence: Etisalat revenue metric now uses Pega's own incremental revenue line and the year is 2020; Telenet and Vodafone UK summaries and years corrected, Vodafone UK moved to production; Amazon Web Services removed as Lumi AI vendor."},{"date":"2026-09-27","note":"Review fixes: EU AI Act tier set to context dependent (minimal for back office scoring, limited when an AI agent delivers the offer); the 15% Etisalat churn figure is now scoped to the SMB division programme where Pega first reported it, in the meta description, FAQ and value note; the meta description no longer claims there were no control groups; the Blits.ai outreach sentence limited to scheduled or API triggered workflows sending by email or a custom function; the ePrivacy note covers the Article 13(2) soft opt in; unsourced wording in the problem section softened. Evidence: Etisalat year, summary and verification note corrected against the January 2021 archive; Telenet and Etisalat summaries say the figures cover the whole decisioning programme."},{"date":"2026-09-27","note":"Review fix: removed the Blits.ai 'send them by email' claim, which the feature inventory does not support (workflows hand off through a custom function). Evidence: Etisalat revenue uplift metric dropped, since the 20% figure reports growth of the already attributed revenue, not an uplift versus a baseline, and the executive quote frames it as incremental value; removed the unsourced 'etisalat by e&' rebrand claim from the summary."}],"slug":"churn-prediction-and-retention-offers","url":"https://www.blits.ai/ai-use-cases/churn-prediction-and-retention-offers","benchmarks":[{"kpi":"churn-reduction","label":"Churn reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":17.5,"min":15,"max":20,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"telenet-next-best-action-decisioning","pooled":true},{"id":"etisalat-next-best-action-retention","pooled":true}]},{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":75,"min":75,"max":75,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"telenet-next-best-action-decisioning","pooled":true}]}],"indicativeValueResult":{"low":900000,"high":8063999.999999999},"evidence":["etisalat-next-best-action-retention","telenet-next-best-action-decisioning","virgin-media-o2-lumi-ai-advisor-assistant","vodafone-uk-always-on-marketing"]},{"title":"AI for telecom fraud detection (SIM swap, IRSF and Wangiri)","shortTitle":"Telecom fraud detection","seoTitle":"AI telecom fraud detection: SIM swap and IRSF","metaDescription":"Operators use AI on network data to stop SIM swap, IRSF and Wangiri fraud and share risk signals with banks, as Telstra and CommBank do with Fraud Indicator.","definition":"AI that protects the operator's own network, revenue and numbers from fraud: it watches call, messaging, roaming and account activity to detect SIM swap and port out takeovers, international revenue share fraud (IRSF) and Wangiri one ring scams, blocks or flags them in real time, and shares risk signals with banks and other businesses that rely on the phone number for security. Scam calls aimed at subscribers are handled by call blocking.","aliases":["telco fraud management","SIM swap fraud detection","IRSF detection","Wangiri fraud prevention","network fraud signals for banks"],"industries":["telecommunications"],"functions":["fraud-prevention","network-operations","security-operations"],"patterns":["anomaly-detection","prediction-and-scoring","classification-and-routing"],"channels":["api","voice","sms"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"customer-protection","problem":"Telecom fraud hits operators and their customers at once. In a SIM swap or port out attack, a\ncriminal takes over a victim's phone number and with it the one time passcodes that protect bank\naccounts and crypto wallets. In international revenue share fraud, criminals push calls to premium\nnumbers they control, often held in another country, through compromised SIM accounts or hacked\ncompany phone systems (PBXs), and take a share of what the victim is billed. In Wangiri scams, a\ncall rings once so the victim calls back to an expensive premium number.\n\nThese patterns move fast and change constantly. Rules written for last month's attack miss this\nmonth's, and blocking too aggressively cuts off genuine calls and customers. Telstra points out\nthat fraudsters who get enough personal information can persuade customers to give up their one\ntime codes and then access bank and superannuation accounts and investment or crypto wallets. The operator holds the\nsignals that reveal these attacks, such as a recent SIM change, an unusual calling pattern or a\nnumber that behaves like a fraud line, but they are only useful if they are scored in real time and\nshared safely.","problemStats":[{"statement":"Telstra cites the Australian Institute of Criminology's estimate that the annual impact of identity crime in Australia exceeds AUD 3.1 billion.","sourceTitle":"What are SIM swaps and porting fraud, and how are we working to stop it?","sourceUrl":"https://www.telstra.com.au/exchange/what-are-sim-swaps-and-porting-fraud--and-how-are-we-working-to-","year":2022},{"statement":"Vodafone reports that more than GBP 485 million was lost to authorised push payment fraud in the UK in 2022.","sourceTitle":"Vodafone Business launches scam signal to defend against impersonation fraud","sourceUrl":"https://www.vodafone.com/news/newsroom/technology/vodafone-business-launches-scam-signal-to-defend-against-impersonation-fraud","year":2024}],"howItWorks":"1. **Stream the signals.** Call detail records, signalling, SIM and port events, roaming records,\n   account changes and customer reports flow into a real time scoring layer.\n2. **Score behaviour, not just lists.** Models learn normal calling and account behaviour and flag\n   deviations: a burst of short international calls to a high cost range, a new SIM followed by\n   logins elsewhere, a number whose call pattern matches known Wangiri campaigns.\n3. **Act by risk.** High confidence fraud traffic is blocked at the network; account takeovers\n   trigger step up checks; uncertain cases go to fraud analysts with the evidence.\n4. **Share risk signals.** Through standard network APIs, such as SIM Swap and Number Verify,\n   banks and online services can ask whether a number was recently swapped or behaves unusually,\n   and use the answer in their own risk decisions.\n5. **Learn from outcomes.** Analyst decisions, customer reports and chargebacks feed back into the\n   models, so detection keeps up as fraudsters change tactics.","valueDrivers":["risk-reduction","customer-experience","cost-to-serve","revenue-growth"],"kpis":["detection-rate-improvement","fraud-loss-reduction","false-positive-reduction"],"indicativeValue":{"referenceOrg":"A mobile operator with 10 million subscribers","inputs":[{"key":"subscribers","label":"Mobile subscribers","low":10000000,"high":10000000,"unit":"subscribers","note":"The reference operator."},{"key":"fraudCostPerSubscriber","label":"Fraud losses and write offs per subscriber per year (IRSF, subscription and SIM related fraud)","low":0.5,"high":2,"unit":"USD per subscriber per year","note":"Editorial assumption. Replace with your own fraud loss reporting."},{"key":"reduction","label":"Share of fraud losses prevented by AI detection","low":0.15,"high":0.3,"unit":"fraction of fraud losses","note":"Editorial assumption, replace with your own. No source on this page measures prevented operator fraud losses; the only measured result is Vodafone's 30% improvement in bank scam detection in a UK pilot, which concerns authorised push payment scams rather than IRSF, Wangiri or subscription fraud."}],"formula":"subscribers * fraudCostPerSubscriber * reduction","currency":"USD","period":"per year","resultLabel":"Operator fraud losses avoided","caveat":"Operator losses only. It leaves out losses avoided by customers and banks, revenue from selling fraud signals through network APIs, analyst time saved, and the cost of false blocks."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Real time scoring on network events at operator scale is demanding, fraud patterns shift quickly, and sharing signals with banks brings privacy, consent and contractual work.","dataPrerequisites":["Call detail records and signalling data in near real time","SIM change, port out and account change events with timestamps","Labelled fraud cases from the fraud team and customer reports","Number ranges and destinations known for high cost or fraud use"],"integrations":["Network switching and signalling platforms for blocking","Fraud management system and case tools","Customer account and SIM management systems","Network API gateway for SIM Swap, Number Verify and similar services","Customer reporting channels such as short codes for spam and scam reports"]},"implementation":{"steps":[{"title":"Map the fraud types and their cost","detail":"Quantify losses per fraud type (IRSF, Wangiri, subscription fraud, SIM swap) and decide which ones justify real time detection first."},{"title":"Get the events in real time","detail":"Attacks on hacked phone systems are often run outside office hours so they last longer, and a daily batch finds them after the bill has grown. Stream call records and SIM events instead."},{"title":"Combine rules and models","detail":"Keep proven rules for known patterns and add models that catch unusual behaviour, with every model alert reviewed by analysts until precision is proven."},{"title":"Tune the blocking threshold on genuine traffic","detail":"Measure how many genuine calls and customers each threshold would block before switching it on, and give customers a quick way to report wrong blocks."},{"title":"Offer signals to partners carefully","detail":"Expose SIM swap and verification signals through standard APIs with contracts, consent and purpose limits, starting with banks."}],"guardrails":["Blocking only above a validated confidence threshold, with a fast route to unblock genuine customers","Signals shared with partners limited to risk indicators, never raw call or location records","Consent and purpose limitation for every partner use of network data","Analyst review of account level actions such as suspending a SIM"],"humanInTheLoop":"Fraud analysts review model alerts that lead to account actions, set blocking thresholds and approve new rules. Customer service can reverse a block after identity checks, and every reversal is fed back to the models.","kpisToInstrument":["Fraud losses per fraud type, normalised for traffic","Detection rate and time to detect for confirmed fraud cases","False positive rate, including genuine calls blocked and customers wrongly flagged","Partner outcomes from shared signals, such as scams stopped by banks"],"failureModes":[{"title":"Blocking genuine customers","detail":"An aggressive threshold cuts off legitimate international callers or new SIM users. Measure impact on genuine traffic before and after."},{"title":"Fraudsters adapt faster than rules","detail":"Static rules catch last month's pattern only. Retrain often and watch for sudden drops in alerts."},{"title":"Signals without context","detail":"A bank treats a recent SIM swap as proof of fraud and locks out a customer who just replaced a phone. Share signals as risk inputs, not verdicts."},{"title":"Privacy overreach","detail":"Partners ask for more network data than they need. Limit sharing to purpose bound risk indicators."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Fraud detection is not listed as high risk in Annex III, and point 5(b) explicitly excludes systems used to detect financial fraud from the creditworthiness category. Blocking fraud traffic is not normally a safety component of critical digital infrastructure (point 2). The tier can change if the same scores are reused for an Annex III purpose: eligibility for essential public assistance benefits and services (point 5(a)), creditworthiness or credit scoring of natural persons (point 5(b)), or risk assessment and pricing for life and health insurance (point 5(c)). A voice or chat agent that takes fraud reports from customers also carries the Article 50(1) duty to tell people they are dealing with an AI system."},"regulations":["gdpr","uk-gdpr","eu-ai-act","telecom-consumer-rules","nist-ai-rmf","eecc","nis2","au-scams-prevention-framework"],"guidance":[{"title":"Cyber Telecom Crime Report 2019","issuer":"Europol European Cybercrime Centre and Trend Micro Research","region":"europe","url":"https://www.europol.europa.eu/cms/sites/default/files/documents/cyber-telecom_crime_report_2019_public.pdf","note":"Threat models for telecom fraud, including international revenue share fraud through hacked PBXs and SIM accounts, and Wangiri callback fraud to premium numbers."},{"title":"CAMARA SIM Swap API","issuer":"CAMARA project (Linux Foundation)","region":"global","url":"https://camaraproject.org/sim-swap/","note":"Open API standard that lets banks and online services check whether a SIM was recently changed, used by operators including Vodafone."},{"title":"APP scams","issuer":"Payment Systems Regulator","region":"europe","url":"https://www.psr.org.uk/our-work/app-scams/","note":"UK reimbursement rules for authorised push payment scams, split between sending and receiving firms; Vodafone cites this reimbursement duty as a reason banks are turning to network based APIs."}],"controls":["Data protection impact assessment for fraud scoring and signal sharing","Documented thresholds and rules with change control","Monitoring of false positives and customer complaints about blocking","Contracts and technical limits on partner use of network risk signals"],"incidents":[]},"blitsAi":{"howToBuild":"Real time scoring and blocking run in the network and the fraud management system. Blits.ai adds\nthe investigation and customer layers: an **AI agent** with a **SQL knowledge base** over alerts\nand cases helps fraud analysts summarise a case and find related numbers, and a **knowledge base**\nholds fraud playbooks. **Agentic workflows** call **custom functions** to suspend a SIM or reverse\na block, always with **human in the loop approval**.\n\nFor customers, a **voice or chat agent** can handle fraud reports and wrongly blocked numbers,\nverify identity through an authentication step in a **flow**, and hand over to the fraud team with\nthe case summary. **PII masking** keeps phone numbers and identity data out of model prompts, and\nthe per tenant user audit log and per task agentic audit trails record who did what."},"faq":[{"question":"How do operators help banks stop SIM swap fraud?","answer":"By sharing a risk signal rather than data. In 2022 Telstra said it would give banks, on request, a rating on a risk scale that shows whether a mobile service used for identity has had a recent SIM swap or port out, so the bank can ask for more information before a transfer goes ahead."},{"question":"Does network data really improve fraud detection?","answer":"Vodafone reports that scam detection improved by 30% after three months of piloting its Scam Signal service with a UK bank; that service targets authorised push payment scams, and Vodafone does not say whether it uses AI. Telstra says the Scam and Fraud Indicator, built by Quantium Telstra with CommBank, uses AI; Fraud Indicator shares intelligence about unusual mobile usage to help detect fraudulently opened accounts, and the gain of more than 25 per cent announced at launch in 2025 was an expectation, not a measured result."},{"question":"What is Wangiri fraud and can it be blocked?","answer":"Wangiri calls ring once from an international number so the victim calls back to a costly premium number. Operators block known patterns in the network; Telstra described improving its Wangiri blocking in 2021, when its platform blocked around 13 million suspected scam calls a month."}],"related":["spam-and-scam-call-blocking","scam-payment-interception","application-and-identity-fraud-detection","order-to-activation-and-esim-onboarding-assistant","mule-network-detection"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched for the telecommunications vertical and verified against operator sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added European Electronic Communications Code, NIS2, Australian Scams Prevention Framework, MAS Shared Responsibility Framework to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: rewrote the IRSF and Wangiri description and the real time step to match the Europol and Trend Micro Cyber Telecom Crime Report, now cited in guidance; removed the unsupported claim that patterns are learned in days; attributed the APP fraud figure to Vodafone; removed the MAS Shared Responsibility Framework, which covers SMS phishing rather than network fraud; clarified in the FAQ that the Fraud Indicator gain was a forecast; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: metaDescription now anchors on the Telstra and CommBank Scam and Fraud Indicator, which Telstra says uses AI, instead of framing the Vodafone Scam Signal result as AI against operator fraud; the Telstra SIM swap rating is described as announced in 2022; the reduction input is marked as an editorial assumption; named the Annex III points; added UK GDPR; aligned account types with the Telstra source."},{"date":"2026-09-27","note":"Second fact check against all cited sources: quotes, dates and prose claims confirmed; metaDescription rewritten as an answer first sentence; EU AI Act basis now notes the Article 50(1) transparency duty for a customer facing fraud reporting agent; the Payment Systems Regulator guidance note now attributes the link to network APIs to Vodafone."}],"slug":"telecom-fraud-detection","url":"https://www.blits.ai/ai-use-cases/telecom-fraud-detection","benchmarks":[{"kpi":"detection-rate-improvement","label":"Detection improvement","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":30,"min":30,"max":30,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"vodafone-scam-signal","pooled":true}]}],"indicativeValueResult":{"low":750000,"high":6000000},"evidence":["telstra-quantium-fraud-indicator","telstra-scam-call-blocking","telstra-sim-swap-and-porting-risk-signal","vodafone-scam-signal"]},{"title":"AI for third party and vendor risk due diligence","shortTitle":"Vendor due diligence","seoTitle":"AI for vendor due diligence and third party risk","metaDescription":"AI reads vendor assurance reports, researches ownership and sanctions and drafts risk assessments. The US DOJ and USDA use AI tools for supplier risk checks.","definition":"AI that reviews a vendor's security questionnaires, SOC and assurance reports, contracts and model documentation against the organization's control requirements, researches the vendor's ownership, sanctions, financial health and adverse media, drafts the risk assessment for a human to approve and keeps the register of material service providers current with ongoing monitoring.","aliases":["third party risk management AI","vendor risk assessment AI","supplier due diligence AI","AI vendor assessment","TPRM automation"],"industries":["cross-industry","banking","insurance","government","payments"],"functions":["procurement","risk-management","regulatory-compliance"],"patterns":["document-processing","rag-knowledge-assistant","agentic-workflow","summarization"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"second-line","problem":"Banks depend on many third parties, and every material one needs due diligence before onboarding\nand monitoring after: security questionnaires, SOC 2 or ISAE reports, business continuity plans,\nfinancial statements, contracts, sanctions and adverse media checks. Analysts read long assurance\nreports to find the handful of exceptions and carve outs that matter, then chase the vendor for\nanswers. Reviews are slow, and once done they go stale until the next periodic review.\n\nAI vendors add new questions: what data trains or reaches the model, how outputs are tested for\nbias and accuracy, who the model and cloud providers are, and whether the bank can audit any of\nit. At the same time, DORA in the EU, APRA CPS 230 in Australia and the US interagency guidance on\nthird party relationships hold the bank accountable for its providers. DORA requires a register of\ninformation on ICT third party arrangements and CPS 230 a register of material service providers\nthat is submitted to APRA. Outsourcing a service never outsources the responsibility.","problemStats":[{"statement":"In the Bank of England and FCA 2024 survey, a third of all AI use cases at UK financial firms were third party implementations, up from 17% in the 2022 survey.","sourceTitle":"Artificial intelligence in UK financial services 2024","sourceUrl":"https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024","year":2024},{"statement":"The same survey found that 46% of firms using or planning to use AI have only a partial understanding of the AI technologies they use, largely because of third party models.","sourceTitle":"Artificial intelligence in UK financial services 2024","sourceUrl":"https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024","year":2024}],"howItWorks":"1. **Scope the review.** The request comes in with the service, data involved and criticality;\n   the assistant proposes the risk tier and the due diligence set required for it.\n2. **Read the documents.** It reads the questionnaire answers, SOC or ISAE reports, bridge\n   letters, policies, contract and, for AI vendors, model documentation, and maps each to the\n   bank's control requirements.\n3. **Flag the gaps.** It lists exceptions in assurance reports, carve outs, missing controls,\n   unanswered questions and contract clauses that fall short of required terms, with citations.\n4. **Research the vendor.** It gathers ownership, sanctions, litigation, financial health and\n   adverse media from approved data sources.\n5. **Draft the assessment.** It drafts the risk assessment and the follow up questions for the\n   vendor; a third party risk analyst reviews, challenges and decides.\n6. **Monitor and update the register.** Ongoing monitoring flags news, certificate expiries and\n   changes, and keeps the register of material or critical providers current.","valueDrivers":["risk-reduction","compliance","employee-productivity","speed"],"kpis":["processing-time-reduction","time-saved-per-task","productivity-gain","hours-saved"],"indicativeValue":{"referenceOrg":"A bank that runs 500 vendor due diligence reviews a year","inputs":[{"key":"reviews","label":"Vendor due diligence reviews per year (new and periodic)","low":500,"high":500,"unit":"reviews per year","note":"The reference bank. Replace with your own third party inventory and review cycle."},{"key":"hoursPerReview","label":"Analyst hours per review","low":8,"high":20,"unit":"hours per review","note":"Editorial assumption across document review, research and write up. Replace with your own records."},{"key":"timeSaved","label":"Share of review time saved","low":0.2,"high":0.4,"unit":"fraction of time","note":"Editorial assumption. No public measured benchmark was found; keep this conservative."},{"key":"hourlyCost","label":"Fully loaded cost of a third party risk analyst","low":60,"high":110,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"reviews * hoursPerReview * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Analyst time released from vendor reviews","caveat":"Time only. It leaves out faster vendor onboarding for the business, the value of continuous monitoring between reviews and the cost of data sources and integration."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Document review and research are well suited to AI. The work is in encoding the bank's control requirements, integrating with the third party risk platform and keeping judgement with analysts.","dataPrerequisites":["The bank's third party control requirements and risk tiering methodology","Vendor documents such as questionnaires, assurance reports, contracts and policies","Licensed data for ownership, sanctions, financial health and adverse media","Past assessments and findings to test the assistant against"],"integrations":["Third party risk management or GRC platform and the vendor register","Procurement and contract management systems","Sanctions, company data and adverse media providers","Regulatory register submission templates, such as the APRA material service provider register"]},"implementation":{"steps":[{"title":"Encode what good looks like","detail":"Turn the bank's control requirements into a checklist per risk tier, including an AI vendor section on data use, model testing, subprocessors and audit rights."},{"title":"Start with assurance report review","detail":"Have the assistant extract scope, exceptions, carve outs and complementary user entity controls from SOC reports, and compare with analyst reviews on past cases."},{"title":"Add research and monitoring","detail":"Connect approved data sources for ownership, sanctions and adverse media, and schedule monitoring for material vendors between reviews."},{"title":"Draft assessments and questions","detail":"Let the assistant draft the assessment and follow up questions, with every finding cited to a document, and have analysts approve before anything reaches the vendor."},{"title":"Keep the register evidenced","detail":"Link each register entry to its latest assessment, contract and monitoring alerts, so the register submitted to the regulator is backed by evidence."}],"guardrails":["A human analyst owns every risk rating and a named executive owns every risk acceptance","Every finding cites the document page or data source it comes from","Vendor documents are processed under confidentiality terms and never used to train models","Research uses approved, licensed data sources only","Monitoring alerts on material vendors are reviewed within a set time"],"humanInTheLoop":"Third party risk analysts review and decide every assessment, business owners and risk committees accept residual risk, and legal approves contract positions. The AI reads, researches, drafts and monitors.","kpisToInstrument":["Cycle time from review request to approved assessment, by risk tier","Analyst hours per review","Findings per review and analyst agreement with AI flagged gaps","Share of material vendors with current assessments and active monitoring","Time from a monitoring alert to analyst review"],"failureModes":[{"title":"Missing the exception in the report","detail":"The assistant summarises a SOC report as clean while an exception affects the bank's service. Test recall on past reports with known exceptions."},{"title":"Questionnaire answers taken at face value","detail":"The vendor's self reported answers are treated as evidence. Weight independent assurance over self attestation."},{"title":"Stale register","detail":"Monitoring runs but alerts are not reviewed, so the register looks current but is not. Track alert review times."},{"title":"AI vendors assessed like any other vendor","detail":"Model, data and subprocessor risks are not asked about. Keep a dedicated AI section in the checklist."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Assessing organizations as vendors is not an Annex III use. If assessments score individual natural persons, such as sole traders, check the design against Annex III and data protection rules. The EU AI Act also shapes what to ask AI vendors, since providers of high risk systems carry specific obligations."},"regulations":["dora","apra-cps-230","eu-ai-act","gdpr","iso-42001","nist-ai-rmf","nis2"],"guidance":[{"title":"Guidelines on third party risk management","issuer":"European Banking Authority","region":"europe","url":"https://www.eba.europa.eu/regulation-and-policy/internal-governance/guidelines-on-outsourcing-arrangements","note":"The EBA's final guidelines on the sound management of third party risk for non ICT services focus on arrangements that support critical or important functions and, once applicable, repeal the 2019 guidelines on outsourcing arrangements. ICT providers, including most AI vendors, fall under DORA."},{"title":"SR 23-4: Interagency Guidance on Third-Party Relationships: Risk Management","issuer":"Federal Reserve, FDIC and OCC","region":"north-america","url":"https://www.federalreserve.gov/supervisionreg/srletters/sr2304.htm","note":"Joint US guidance on sound risk management for all stages in the life cycle of third party relationships, from planning and due diligence to ongoing monitoring and termination. It imposes no new requirements."},{"title":"Operational risk management (CPS 230)","issuer":"Australian Prudential Regulation Authority","region":"asia-pacific","url":"https://www.apra.gov.au/operational-risk-management","note":"Under paragraph 51 of CPS 230, APRA regulated entities must submit their register of material service providers to APRA; APRA's template is the preferred way to do so."}],"controls":["Risk tiering methodology with required due diligence per tier","Documented AI vendor question set covering data, models, testing and audit rights","Evidence link from every register entry to its assessment and contract","Review of AI generated findings and ratings by a named analyst","Periodic quality review of assessments against a sample of manual reviews"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** started from the third party risk platform through\nthe API. The **knowledge base** ingests the vendor's documents (PDF, Word, Excel, email) and the\nbank's control requirements, retrieved with hybrid search, and **custom functions** call\napproved data providers for ownership, sanctions and adverse media. The agent returns\n**structured output** with findings, citations, a draft rating and follow up questions.\n\n**Agentic tasks** with scheduled rechecks keep material vendors under monitoring, **human in the\nloop approval** routes every draft rating to an analyst, and the **tool execution policy**\nlimits which systems the agent may use. Per run audit trails and **test suites** built on past\nassessments show how each conclusion was reached, and the platform is model agnostic with EU and\nUAE data residency."},"faq":[{"question":"Can AI decide whether a vendor is acceptable?","answer":"No. It can read the documents, research the vendor and draft the assessment, but a human analyst owns the rating and a named executive owns the risk acceptance. Regulators are clear that outsourcing does not move responsibility away from the bank."},{"question":"What should we ask AI vendors specifically?","answer":"What data they use and retain, whether customer data trains their models, which model and cloud providers they rely on, how they test outputs for accuracy and bias, how incidents are reported and what audit and access rights the bank gets."},{"question":"Who uses AI for due diligence today?","answer":"Government buyers publish the clearest examples in their AI inventories: the US Department of Justice has used Exiger's AI platform since 2023 to build vendor risk profiles that inform acquisition decisions, USDA uses an AI search tool for supplier responsibility checks and the IRS pilots machine learning to flag contractors at risk of poor performance."}],"related":["procurement-contract-review","continuous-controls-testing","ai-model-inventory","pep-and-adverse-media-screening","business-onboarding-and-ubo-discovery"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against public sources. The catalog's ISACA source is members only and names no organization, so it is not used."},{"date":"2026-09-25","note":"Consolidation pass: added NIS2 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the EBA guidance note (the new third party guidelines, not the outsourcing guidelines, focus on critical or important functions), the APRA register note and the register requirements in the problem; removed unsourced vendor counts and review times; tightened the four evidence summaries to the inventory wording; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Second fact check against the 2024 and 2025 federal AI inventories and all guidance pages: narrowed the 46% survey statistic to firms using or planning to use AI, stopped the meta description implying the DOJ and USDA tools read assurance reports, and separated DOJ risk scores from research outputs and dropped the unsourced contracting officer detail in the USDA summary."}],"slug":"vendor-due-diligence","url":"https://www.blits.ai/ai-use-cases/vendor-due-diligence","benchmarks":[],"indicativeValueResult":{"low":48000,"high":440000},"evidence":["doj-exiger-supply-chain-risk-management","irs-vendor-risk-analytics","usda-procuresight-responsibility-determination","ustda-exiger-partner-due-diligence"]},{"title":"AI for voice of the customer and feedback analysis","shortTitle":"Customer feedback analysis","seoTitle":"AI customer feedback and VoC analysis","metaDescription":"AI tags survey comments, reviews and call transcripts by theme and sentiment. Microsoft reports Majid Al Futtaim cut feedback processing from 7 days to 3 hours.","definition":"AI that reads every piece of free text customer feedback, such as survey verbatims, NPS comments, reviews, social posts, chat and call transcripts, and turns it into themes, sentiment, drivers and suggested actions that a named owner can act on, so the organization hears all of its customers instead of a sample.","aliases":["voice of the customer analytics","VoC analysis","survey verbatim analysis","customer feedback text analytics","NPS comment analysis"],"industries":["cross-industry","retail-and-ecommerce","government","manufacturing"],"functions":["customer-service","marketing","analytics-and-reporting"],"patterns":["classification-and-routing","summarization","speech-analytics"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"copilot","adoptionStage":"mainstream","problem":"Organizations collect far more feedback than they read. Survey platforms, app store reviews,\nsocial media, chat logs and call recordings produce tens of thousands of comments a week (Majid\nAl Futtaim Retail's marketing team processed 60,000 to 70,000 customer responses a week by hand,\naccording to Microsoft), and most of the value sits in the free text: why a customer gave a low score, what broke, what they wanted\ninstead. Analysts read a sample, tag it by hand against a codebook that drifts over time, and\nreport weeks later, by which time the issue has cost more customers.\n\nKeyword based text analytics helped with volume but struggled with sarcasm, mixed sentiment,\nseveral topics in one comment, other languages and new themes nobody had a keyword for. Language\nmodels change the economics: every comment can be classified against the organization's own\ntaxonomy, summarized per theme and linked to operational data, so a product owner sees the\nproblem in days. The discipline that remains is human: someone has to own each theme, decide what\nit means and close the loop with customers.","problemStats":[],"howItWorks":"1. **Collect every source.** Survey exports, reviews, social mentions, chat logs and transcribed\n   calls land in one store with their metadata (channel, product, date, score, segment).\n2. **Clean and protect.** Personal data is masked before analysis; duplicates, spam and empty\n   answers are removed.\n3. **Classify against your taxonomy.** Each comment is tagged with one or more themes from the\n   organization's own codebook, a sentiment per theme and, where present, a suggested action or a\n   statement of customer effort. New clusters that fit no theme are flagged for review.\n4. **Quantify and explain.** Themes are joined to scores and operational data (store, product,\n   journey step), so dashboards show which themes drive detractors and how they trend.\n5. **Summarize for owners.** Each theme owner gets a short summary with representative, anonymized\n   quotes and the change against last period.\n6. **Close the loop.** Individual comments that need a response (a complaint, a safety issue, a\n   vulnerable customer) are routed to the right team; systemic fixes are tracked to completion.","valueDrivers":["customer-experience","employee-productivity","speed","revenue-growth"],"kpis":["cycle-time-days","accuracy","cost-reduction","productivity-gain"],"indicativeValue":{"referenceOrg":"A consumer brand or public service that receives 300,000 free text feedback items a year","inputs":[{"key":"feedbackItems","label":"Free text feedback items per year","low":300000,"high":300000,"unit":"items per year","note":"The reference organization, across surveys, reviews and chat."},{"key":"hoursPerItem","label":"Analyst time to read and code one item by hand","low":0.01,"high":0.02,"unit":"hours per item","note":"Editorial assumption of 36 to 72 seconds per comment. Replace with your own coding time."},{"key":"shareAutomated","label":"Share of manual coding the AI replaces","low":0.6,"high":0.9,"unit":"fraction of items","note":"Conservative against the benchmarks on this page (Google Cloud reports that SBF Group eliminated manual data analysis work and cut annual costs by about 95%), because theme review and quality checks stay with people."},{"key":"analystCost","label":"Fully loaded analyst cost","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"feedbackItems * hoursPerItem * shareAutomated * analystCost","currency":"USD","period":"per year","resultLabel":"Manual feedback coding cost avoided","caveat":"Counts only the analyst time to read and code comments, and assumes the organization would otherwise read all of them (most read a sample). It leaves out the platform cost and the larger value: issues found and fixed weeks earlier, and churn or complaints avoided."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Classifying text is a mature task and the data is usually already exported from survey and review tools. The effort goes into an agreed taxonomy, joining feedback to operational data, masking personal data and building the habit of acting on the output.","dataPrerequisites":["Exports or APIs from survey, review, social listening and contact centre systems","A theme taxonomy (codebook) agreed with the business, with an owner per theme","Metadata that links feedback to product, location, journey step and score","A labelled sample of a few hundred comments to measure classification accuracy"],"integrations":["Survey and experience management platforms","Review sites, app stores and social listening tools","Contact centre recordings and chat logs, with transcription","Data warehouse and BI dashboards","Case or CRM system for comments that need an individual response"]},"implementation":{"steps":[{"title":"Agree the taxonomy before the model","detail":"Start from the themes the business already reports, add the top new clusters found in a sample, and name an owner for each. A model that classifies against themes nobody owns produces dashboards nobody acts on."},{"title":"Build a labelled test set","detail":"Have two people code a few hundred real comments per language, resolve their disagreements, and use the set to measure accuracy per theme on every prompt or model change."},{"title":"Mask personal data at intake","detail":"Remove names, account numbers and contact details before analysis and keep the link to the original record only where a response is needed."},{"title":"Join feedback to operations","detail":"Link each comment to the store, product, order or journey step it concerns, so a theme can be traced to a cause rather than reported as a trend."},{"title":"Route individual cases, report systemic ones","detail":"Send comments that are complaints, safety issues or signs of vulnerability to the teams that respond to individuals, and give theme owners a periodic summary with trend and quotes."},{"title":"Review the codebook every quarter","detail":"Look at the unclassified cluster, retire themes that no longer occur and add new ones with an owner, then rerun the test set."}],"guardrails":["Classification only against an approved taxonomy, with an explicit \"other or new\" bucket reviewed by people","Personal data masked before comments reach a model or a dashboard","Quotes shown to wide audiences are anonymized and checked","Comments that indicate a complaint, a safety risk or a vulnerable customer are routed to a person, not only counted","No inference of emotion from voice or face in call and video feedback without a separate legal assessment"],"humanInTheLoop":"Analysts own the taxonomy, check a weekly sample of classifications against the test set, and validate every theme before it is reported as a finding. Theme owners decide what action to take; the AI suggests, it does not commit changes to products or policies.","kpisToInstrument":["Classification accuracy per theme and language on the labelled test set","Share of feedback analysed (coverage) versus the previous sampling approach","Time from feedback received to theme reported to its owner","Share of themes with an owner and a tracked action","Share of individual cases routed correctly (complaints, safety, vulnerability)"],"failureModes":[{"title":"Dashboards without owners","detail":"The analysis is excellent and nothing changes. Give every theme an owner and track actions to closure."},{"title":"Confident but wrong themes","detail":"The model fits comments into the nearest theme and hides new issues. Keep a \"new or other\" bucket and review it."},{"title":"Sentiment that misses the point","detail":"A positive score on a comment that describes a serious failure. Report themes and drivers, not sentiment alone."},{"title":"Personal data spread into reports","detail":"Verbatim quotes with names or account details reach wide audiences. Mask at intake and review quotes before sharing."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Classifying and summarizing text feedback is minimal risk. The tier changes if the system infers emotions from customers' voices or faces in calls or video: emotion recognition based on biometric data is listed as high risk in Annex III point 1(c) and triggers the Article 50(3) duty to inform the people exposed. Analysing feedback from employees to evaluate individual workers moves it towards Annex III point 4(b), and emotion recognition in the workplace is prohibited by Article 5(1)(f), except for medical or safety reasons."},"regulations":["eu-ai-act","gdpr","iso-42001"],"guidance":[{"title":"Regulation (EU) 2024/1689, the Artificial Intelligence Act","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Official text on EUR-Lex. Annex III point 1(c) lists emotion recognition systems and point 4(b) systems that monitor and evaluate the performance and behaviour of workers; Article 5(1)(f) prohibits emotion recognition in the workplace and in education, relevant when employee feedback or staff calls are analysed; Article 50(3) requires deployers of emotion recognition systems to inform the people exposed to them."}],"controls":["Data protection impact assessment where feedback includes personal data or call recordings","Documented taxonomy with owners and a change log","Accuracy testing on a labelled set per language before each change","Retention limits on raw verbatims and recordings","Access control on dashboards that show individual comments"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this runs as an **agentic workflow** on a schedule: **custom functions** pull new\nfeedback from survey, review and ticketing systems through their REST APIs (the integration\ncatalog includes Zendesk, Salesforce and ServiceNow) and strip names, account numbers and contact\ndetails before the data reaches the agent, and an **AI agent** with **structured output** tags\neach comment with themes from your codebook, sentiment per theme and any suggested action. The results are written to a SQL\ndatabase registered as a **SQL knowledge base**, so an agent can answer plain language questions\nfrom analysts and theme owners over them.\n\nFor feedback that comes in through Blits.ai channels, **PII masking** at the gateway removes\npersonal data from chat input, and the built in **analytics** already show satisfaction and\nsentiment (with engines from Amazon, Google, IBM and Microsoft), thumbs up and down feedback with\ncomments on responses, and conversation level CSAT ratings. **Test suites** with deterministic\nrules for label matching (or LLM based grading for summaries) let you run your labelled comments\nthrough the classifier after every prompt change,\n**human in the loop** approval holds routing of sensitive comments until a person confirms, and\n**monitors** run scheduled checks against the classifier agent and alert when it fails. The\nplatform is model agnostic and can run in the EU or UAE region."},"faq":[{"question":"How accurate is AI at classifying customer feedback?","answer":"It can be accurate enough to replace most manual coding, but only your own labelled comments tell you whether it is. Google Cloud reports that SBF Group raised feedback classification accuracy from 16% to 84% with Google Cloud AI. Measure accuracy per theme and per language, because averages hide weak themes."},{"question":"How much faster is AI feedback analysis?","answer":"Days become minutes or hours. Microsoft reports that Majid Al Futtaim cut feedback processing for Carrefour from seven days to three hours (the same story quotes its Chief Digital Officer as saying three to four minutes), and Google Cloud reports that Mattel cut analysis from a month to a minute. The binding constraint then becomes how fast owners act."},{"question":"Is sentiment analysis of customer calls high risk under the EU AI Act?","answer":"Analysing the words people say or write is not. Inferring emotions from their voice or face is emotion recognition, which Annex III lists as high risk, with a duty under Article 50(3) to inform the people exposed. Inferring the emotions of your own staff, such as contact centre agents, is prohibited by Article 5(1)(f) except for medical or safety reasons. Keep voice analysis to transcripts unless you have done that assessment."},{"question":"How is this different from complaints root cause analysis?","answer":"Complaints root cause analysis works on regulated complaints, where every case must be handled and systemic causes reported. Feedback analysis covers the much larger stream of surveys, reviews and comments, most of which are not complaints, to find what drives satisfaction and where to invest."}],"related":["complaints-root-cause-analysis","public-consultation-response-analysis","call-quality-and-compliance-monitoring","personalized-marketing-at-scale","governed-text-to-sql-analytics"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical with five evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed the Mattel 100x figure as a productivity metric (it is data processing capacity), attributed the SBF Group figures to Google Cloud, added Grupo SBF's own page as a source, removed an unsupported comparison from the SSA record, corrected the Blits.ai build notes on monitors, feedback and SQL knowledge bases, made the Article 5(1)(f) exception explicit, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Editorial review fixes: limited the Blits.ai PII masking claim to feedback arriving through Blits.ai channels (pulled feedback is masked in the custom function), reworded test suites as run by the user after prompt changes, corrected the Majid Al Futtaim record (year 2025 from the dated story, summary no longer claims dashboards with suggested actions) and noted the vendor's minutes figure, cited EUR-Lex for the AI Act instead of an unofficial mirror, and linked the weekly feedback volume in the problem text to its source."}],"slug":"customer-feedback-analysis","url":"https://www.blits.ai/ai-use-cases/customer-feedback-analysis","benchmarks":[{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":84,"min":84,"max":84,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"sbf-group-nps-feedback-analysis","pooled":true}]},{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":95,"min":95,"max":95,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"sbf-group-nps-feedback-analysis","pooled":true}]},{"kpi":"cycle-time-days","label":"Cycle time","unit":"hours","aggregate":false,"higherIsBetter":false,"n":1,"nUpTo":0,"median":3,"min":3,"max":3,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"majid-al-futtaim-customer-feedback-analysis","pooled":true}]},{"kpi":"cycle-time-days","label":"Cycle time","unit":"minutes","aggregate":false,"higherIsBetter":false,"n":1,"nUpTo":0,"median":1,"min":1,"max":1,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"mattel-feedback-classification","pooled":true}]}],"indicativeValueResult":{"low":72000,"high":378000},"evidence":["hud-voice-of-the-customer-analytics","majid-al-futtaim-customer-feedback-analysis","mattel-feedback-classification","sbf-group-nps-feedback-analysis","ssa-customer-insight-survey-text-analytics"]},{"title":"AI generated client portfolio reports and commentary","shortTitle":"Portfolio commentary","seoTitle":"AI portfolio commentary and client reporting","metaDescription":"AI drafts client portfolio commentary from system data for human approval. Morgan Stanley adopted BlackRock's Aladdin Wealth Auto Commentary for talking points.","definition":"AI that drafts each client's periodic portfolio commentary and report narrative (performance, attribution, what drove returns, positioning and outlook) in plain language and in the client's language, where every figure comes from the portfolio system of record and a reviewer approves the text before delivery.","aliases":["automated portfolio commentary","AI client reporting","fund commentary generation","investment commentary drafting"],"industries":["wealth-and-asset-management","banking"],"functions":["analytics-and-reporting","customer-service","operations"],"patterns":["content-generation","summarization","translation"],"channels":["internal-tools","email"],"audience":"back-office","autonomy":"copilot","adoptionStage":"early-adopters","segment":"middle-office","problem":"Clients expect a periodic report that explains what happened to their money and why, not just a\ntable of numbers. Writing that narrative is slow: portfolio managers and specialist writers draft\ncommentary for each strategy, and advisors adapt it for individual clients, often in several\nlanguages. At Quilter, specialist writers interview a portfolio manager and then need a few days\nto turn that into a commentary; Neurons Lab describes a monthly investor report that took an\ninvestment firm 20 days to complete. The result can be generic text that says little about the\nclient's own portfolio, or reports that reach clients well after the period they describe.\n\nThe risk is also real. A commentary is a client communication under conduct rules; a wrong number,\nan unbalanced claim about performance or a forward looking statement without the right disclaimer\nis a compliance issue.","problemStats":[],"howItWorks":"1. **Take numbers from the record.** Performance, attribution, holdings and transactions come from\n   the portfolio accounting and performance systems, not from the model.\n2. **Add the context.** The house view, market commentary and the portfolio manager's notes are\n   retrieved for the period.\n3. **Draft within a template.** The model writes the narrative sections in an approved structure,\n   inserting figures from the data and explaining drivers in plain language, per client or per\n   strategy.\n4. **Localize.** The approved narrative is adapted to the client's language and segment.\n5. **Check and approve.** Automated checks compare every number in the text with the source data\n   and flag banned phrases; a reviewer approves before the report is assembled and delivered, and a\n   log of sources and edits is kept.","valueDrivers":["employee-productivity","speed","customer-experience","compliance"],"kpis":["cycle-time-days","processing-time-reduction","error-reduction","hours-saved"],"indicativeValue":{"referenceOrg":"A wealth manager sending quarterly reports to 20,000 client portfolios","inputs":[{"key":"portfolios","label":"Client portfolios with a personalized commentary","low":20000,"high":20000,"unit":"portfolios","note":"The reference firm."},{"key":"reportsPerYear","label":"Reports per portfolio per year","low":4,"high":4,"unit":"reports per year","note":"Quarterly reporting."},{"key":"minutesSaved","label":"Minutes of drafting and adaptation saved per report","low":10,"high":30,"unit":"minutes per report","note":"Editorial assumption, replace with your own. No deployment on this page publishes a measured saving per report; the only test on this page (Quilter) was a single commentary drafted in about 45 minutes of prompting and editing instead of a few days."},{"key":"hourlyCost","label":"Fully loaded cost per hour of the staff who write and adapt commentary","low":60,"high":120,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"portfolios * reportsPerYear * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Value of staff time released from commentary drafting","caveat":"Where commentary is not personalized per client today, the saving may show up as better reports rather than fewer hours. The figure leaves out the review effort, platform costs and the effect on client retention."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Data is the hard part: reliable performance and attribution data per portfolio, mapped to a template. The generation is well understood when numbers are inserted, not generated, and output goes through review.","dataPrerequisites":["Performance, attribution and holdings data per portfolio and period","House view and market commentary for the period","Approved templates, disclaimers and banned phrases per market","Client language and segment preferences"],"integrations":["Portfolio accounting and performance measurement systems","Research and CIO content","Report assembly and document generation","Client portal or email for delivery"]},"implementation":{"steps":[{"title":"Start at strategy level","detail":"Generate commentary per strategy or model portfolio first, where one reviewed text serves many clients, before personalizing per client."},{"title":"Insert numbers, do not generate them","detail":"Pass figures as structured data and require the model to reference them, then compare every number in the output with the source automatically."},{"title":"Agree the compliance rules up front","detail":"Encode disclaimers, fair and balanced presentation rules and banned phrases, and have compliance approve the templates."},{"title":"Review by exception at scale","detail":"When personalizing per client, review all outputs at first, then move to risk based sampling once error rates are proven low, keeping full review for outliers."},{"title":"Keep the decision log","detail":"Store the data, retrieved context, draft, edits and approver for every report, so any sentence can be traced later."}],"guardrails":["Every figure from the system of record, checked automatically against the source","Approved templates, disclaimers and banned phrase lists per market","Human approval before delivery, with risk based sampling only after proven accuracy","No forecasts or promises beyond the approved house view wording","Log of data, context, draft and approver for every report"],"humanInTheLoop":"Portfolio managers or specialist writers approve strategy commentary; reviewers or advisors approve personalized reports; compliance approves templates and samples output.","kpisToInstrument":["Days from period end to report delivery","Number mismatches caught by automated checks per thousand reports","Reviewer edit rate and rejection reasons","Share of clients receiving personalized commentary","Client feedback or complaints about reports"],"failureModes":[{"title":"A wrong number reaches a client","detail":"The model restates or rounds a figure incorrectly. Insert numbers from data and check every figure before release."},{"title":"Unbalanced performance claims","detail":"Commentary highlights gains and glosses over losses. Encode fair and balanced rules and review for them."},{"title":"Generic text at scale","detail":"Personalized reports all say the same thing. Require references to the client's own holdings and drivers."},{"title":"Review becomes a formality","detail":"Reviewers approve thousands of reports without reading them. Use risk based sampling with clear accountability."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Drafting client reports for human review is not listed in Annex III and is not a practice prohibited by Article 5, so the tier turns on the firm's role under Article 50. A firm that deploys a third party generator (for example a feature of its portfolio platform) for private client reports has no Article 50 duty: the Article 50(4) disclosure duty covers AI generated text published to inform the public on matters of public interest, which private client reports are not, and it lapses anyway after human review under editorial responsibility. For that firm the tier is minimal. A firm that builds the generating system or places it on the market under its own name is a provider under Article 50(2) and must mark the synthetic text in a machine readable format; drafting whole commentaries goes beyond the exemption for an assistive function for standard editing, so for that firm the tier is limited."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","dora","mas-ai-risk-management","iso-42001","mifid-ii"],"guidance":[{"title":"ESMA public statement on the use of AI in the provision of retail investment services","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf","note":"MiFID II duties on accurate, fair and not misleading client information apply to AI drafted communications, and ESMA expects records of AI use."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Providers must mark AI generated text in a machine readable format (paragraph 2); the disclosure duty for text published on matters of public interest does not apply after human review under editorial responsibility (paragraph 4)."},{"title":"PRIN 2A.5 Consumer Duty: retail customer outcome on consumer understanding","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.handbook.fca.org.uk/handbook/PRIN/2A/5.html","note":"The consumer understanding outcome of the Consumer Duty applies to how performance and risks are explained in reports to retail clients."}],"controls":["Automated number reconciliation between text and source data","Compliance approved templates with version control","Decision log per report kept for the retention period","Risk based sampling with documented reviewer accountability","Inventory entry for the generation system with an accountable owner"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai an **agentic workflow** runs after period end: **custom functions** or a **SQL\nknowledge base** read performance and holdings from the firm's systems, a **knowledge base**\nsupplies the house view and manager notes, and an **AI agent** with **structured output** drafts\neach section in the approved template, referencing figures by field. A second function compares\nevery number in the draft with the source data and checks that the required disclaimers are\npresent.\n\nDrafts pause for **human in the loop** approval, and **multi language** support and **machine\ntranslation** adapt approved text per market. **Guardrails** block banned phrases, the workflow's\n**audit trail** and downloadable run data serve as the decision log,\nand **test suites** replay reference portfolios on every template, prompt or model change."},"faq":[{"question":"Who is doing this today?","answer":"BlackRock announced in October 2025 that Morgan Stanley Wealth Management's Portfolio Risk Platform would be the first to implement Auto Commentary, a generative AI feature of Aladdin Wealth, with advisors getting access from that month. The tool drafts concise talking points for advisors from risk analytics, the Chief Investment Office outlook and the client's portfolio, not full client reports. Quilter tested turning a portfolio manager interview into an investment commentary in about 15 minutes of prompting and half an hour of editing instead of a few days, which it called a one off test."},{"question":"How do you prevent wrong numbers?","answer":"Never let the model produce figures. Pass them in from the system of record, reference them in the draft, and reconcile every number automatically before a human reviews the text."},{"question":"Can personalized reports go out without review?","answer":"Start with full review. Move to risk based sampling only when automated checks and error rates justify it, and keep the decision log for every report."}],"related":["investment-research-summarization","portfolio-drift-monitoring-and-rebalancing","wealth-advisor-knowledge-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Portfolio Reporting and Commentary) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: added MiFID II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the EU AI Act basis (Article 50(4) covers text published on matters of public interest; Article 50(2) marking duty for providers), replaced the Consumer Duty guidance link with PRIN 2A.5, clarified the Morgan Stanley and Quilter FAQ answer, added seoTitle and metaDescription; Neurons Lab evidence year set to its 2026 publication."},{"date":"2026-09-27","note":"Review fixes: the EU AI Act tier is now context dependent (minimal for a firm deploying a third party generator, limited for a firm that provides the generating system under Article 50(2)); removed time saved per task from the KPIs so a meeting notes figure cannot appear as a commentary benchmark; rewrote the minutes saved note; the Morgan Stanley answer now follows BlackRock's own release; disclaimer checks moved into the reconciliation function; softened two unsourced claims."},{"date":"2026-09-27","note":"Fact checked against sources again: all quotes, dates and guidance links confirmed; the problem section now cites the Quilter and Neurons Lab baselines instead of an unsourced generalization; the Morgan Stanley evidence summary follows the InvestmentNews wording (bullet point insights, not full scripts or emails)."},{"date":"2026-09-27","note":"Review fixes: scoped the minutesSaved note to 'the only test on this page (Quilter)'; removed organization.region from the Neurons Lab evidence record, since the source names no region; flagged the Quilter evidence record's year for editor sign off, since no date could be found on the source page after a fresh check."},{"date":"2026-09-27","note":"Editor sign off resolved: the Quilter evidence record's source page carries no publish date, so its year now rests on the first Wayback capture of that page (2025-05-17, confirmed via the CDX API), moving it from 2024 to 2025; the page still holds two public records of named organizations (Quilter and Morgan Stanley), so it keeps status review."}],"slug":"portfolio-reporting-and-commentary","url":"https://www.blits.ai/ai-use-cases/portfolio-reporting-and-commentary","benchmarks":[],"indicativeValueResult":{"low":800000,"high":4800000},"evidence":["morgan-stanley-aladdin-auto-commentary","quilter-copilot-meeting-notes"]},{"title":"AI home loan assistant with pre qualification","shortTitle":"Home loan assistant","seoTitle":"AI mortgage assistant for pre qualification","metaDescription":"An AI home loan assistant answers rate and document questions and gives indicative borrowing estimates, handing qualified customers to a mortgage specialist.","definition":"A customer facing assistant that answers home loan questions (rates, loan to value, fees, the documents needed), runs indicative affordability and borrowing estimates from the bank's published rules, and books the customer with a mortgage specialist, grounded in the bank's current, versioned product and policy documents.","aliases":["mortgage chatbot","home loan pre qualification assistant","mortgage enquiry agent","borrowing power assistant"],"industries":["banking","real-estate"],"functions":["lending-and-credit","sales","customer-service"],"patterns":["rag-knowledge-assistant","conversational-agent","voice-agent"],"channels":["web-chat","mobile-app","whatsapp","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"A home loan is a large, long commitment, and the questions start well before an application: how\nmuch can I borrow, what deposit do I need, fixed or variable, what happens when my fixed rate\nends, which documents will you want. Customers research when it suits them, often outside the\nhours when specialists are available, and a slow or vague answer can send them to a broker or a\ncompetitor.\n\nEvery hour a specialist spends on early questions, or on applicants who were unlikely to qualify,\nis an hour not spent on applications. And because rates and credit policy change, a generic\nassistant, or a web page that is out of date, can give answers that are wrong in a way that\nmatters: a borrowing estimate that is too high, or a rate that no longer exists.","problemStats":[],"howItWorks":"1. **Answer from current, approved content.** Questions about rates, fees, loan to value limits,\n   product features and documents are answered by retrieval over versioned product and policy\n   documents, with the source and date shown, and a refusal when the content does not cover it.\n2. **Estimate, clearly labelled.** Borrowing power and repayment estimates come from the bank's\n   published calculator logic, called as a tool, never from the model's own arithmetic, and are\n   shown as indicative, not an offer.\n3. **Pre qualify against published rules.** Basic checks (deposit, income bands, residency,\n   property type) come from a rules service and tell the customer what is likely to be needed,\n   without a credit decision.\n4. **Prepare the handover.** The assistant captures the customer's situation and questions,\n   lists the documents they will need and books a call or meeting with a specialist.\n5. **Serve existing borrowers too.** Fixed rate expiry, switching products, extra repayments and\n   offset questions come from the same content, with account specific actions behind\n   authentication.","valueDrivers":["revenue-growth","customer-experience","employee-productivity","speed"],"kpis":["conversion-rate-uplift","response-time-reduction","interactions-handled","customer-satisfaction","accuracy"],"indicativeValue":{"referenceOrg":"A lender handling 100,000 home loan enquiries a year","inputs":[{"key":"enquiries","label":"Home loan enquiries per year","low":100000,"high":100000,"unit":"enquiries per year","note":"The reference lender."},{"key":"handledShare","label":"Share of enquiries answered without a specialist","low":0.3,"high":0.5,"unit":"fraction of enquiries","note":"Editorial assumption; early questions about rates, fees and documents are the ones the assistant can answer."},{"key":"minutesPerEnquiry","label":"Specialist minutes per early enquiry","low":10,"high":20,"unit":"minutes per enquiry","note":"Editorial assumption, replace with your own time data."},{"key":"costPerMinute","label":"Fully loaded cost of a specialist minute","low":1,"high":1.5,"unit":"USD per minute","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"enquiries * handledShare * minutesPerEnquiry * costPerMinute","currency":"USD","period":"per year","resultLabel":"Specialist time freed from early enquiries","caveat":"Counts specialist time only, which is capacity freed rather than cash saved unless staffing changes. It leaves out extra settled loans from faster, out of hours answers, retention of borrowers at fixed rate expiry and the cost of the AI and content upkeep."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Answering from documents is straightforward. The hard parts are keeping rate and policy content current to the day, reusing the bank's own calculators instead of letting the model compute, and a clean booking handover to specialists.","dataPrerequisites":["Current rate sheets, product terms and credit policy summaries with owners and effective dates","The bank's borrowing power and repayment calculator logic as a callable service","Published eligibility criteria for pre qualification","Specialist calendars and booking rules"],"integrations":["Rates and product content management","Calculator and rules services","Scheduling system for specialists and brokers","CRM for leads and handover summaries","Loan origination for existing application status, behind authentication"]},"implementation":{"steps":[{"title":"Build the content spine first","detail":"Collect every rate sheet, product term and policy summary, assign an owner and an effective date to each, and remove anything unofficial. The assistant is only as good as this set."},{"title":"Wrap the calculators as tools","detail":"Expose the bank's existing borrowing power and repayment calculators to the assistant, so every number matches what a specialist would show."},{"title":"Write the boundaries","detail":"Decide what the assistant may say about eligibility and what is reserved for a specialist, and phrase estimates as indicative every time."},{"title":"Make booking the default next step","detail":"End qualified conversations with a specialist booking and a summary, so the conversation turns into an application rather than a dead end."},{"title":"Test with rate changes","detail":"Include rate change days in the test plan: update the content, rerun the test suite and check that no old rate survives anywhere."}],"guardrails":["Rates, fees and limits only from dated, approved content, with a refusal when not covered","All numbers from the bank's calculators, labelled as indicative estimates","No credit decision or approval language","Personal financial details masked in prompts and logs","Handover to a specialist for complex situations and vulnerability signals"],"humanInTheLoop":"Mortgage specialists own advice, applications and every credit decision. Product owners approve the content set and the calculators, and a sample of conversations is reviewed weekly for accuracy and for any wording that sounds like advice or approval.","kpisToInstrument":["Share of enquiries answered without a specialist, with repeat contacts counted","Specialist bookings and applications started per conversation","Accuracy of answers on a weekly checked sample, including rates quoted","Out of hours share of conversations","Complaints mentioning the assistant"],"failureModes":[{"title":"Yesterday's rate","detail":"A superseded rate stays in the index. Version content with effective dates and retire old versions on the day of change."},{"title":"The model does the maths","detail":"The assistant calculates a repayment itself and gets it wrong. Route every number through the calculator tool."},{"title":"Advice by accident","detail":"The assistant recommends a product for a person's circumstances in a market where that is regulated advice. Keep to general information and hand over for recommendations."},{"title":"Booking dead ends","detail":"The customer is told a specialist will call and nobody does. Book into real calendars and track kept appointments."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Answering questions and giving indicative estimates from published rules is limited risk with an Article 50(1) disclosure that the customer is talking to an AI system. If the assistant evaluates an individual's creditworthiness to decide or filter access to a loan, it falls under Annex III point 5(b) and is high risk."},"regulations":["eu-ai-act","gdpr","eba-loan-origination","uk-consumer-duty","dora","eu-mortgage-credit-directive","us-ecoa-reg-b"],"guidance":[{"title":"Directive 2014/17/EU on credit agreements for consumers relating to residential immovable property","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2014/17/oj","note":"The Mortgage Credit Directive sets rules on advertising, standard information and creditworthiness assessment that the assistant's answers must respect."},{"title":"Responsible lending","issuer":"Australian Securities and Investments Commission","region":"asia-pacific","url":"https://asic.gov.au/regulatory-resources/credit/responsible-lending/","note":"Example of national responsible lending obligations that shape what pre qualification may say."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Annex III of Regulation (EU) 2024/1689 (the AI Act). Point 5(b) covers creditworthiness evaluation of natural persons."}],"controls":["AI disclosure and a statement that estimates are indicative","Content inventory with owners, effective dates and retirement on change","Calculator and rules services under change control and tested","Log of every estimate shown, with the inputs used","Weekly accuracy sampling and complaint monitoring"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** of rate sheets, product terms and\npolicy summaries, retrieved with hybrid search, using **document version control** so superseded\ncontent is replaced on the day of a rate change. **Custom functions** call the bank's borrowing\npower and repayment calculators and the specialist booking system, so every number and every\nappointment comes from the bank's own systems. Structured output can hand a clean summary to CRM.\n\nThe assistant runs on **web chat**, **WhatsApp** and **voice** telephony, inside the bank's own\nmobile app through the **API channel**, and a **digital human** can present it on a website. **Guardrails** block approval\nlanguage and advice phrasing, **PII masking** protects income and debt details, and **human\nhandover** connects the customer to a specialist. **Test suites** with LLM grading and\nknowledge base evidence check answers after every content change, and **monitors** run scheduled\nchecks that the assistant quotes current rates."},"faq":[{"question":"Can a home loan chatbot tell me how much I can borrow?","answer":"It can give an indicative estimate from the bank's own calculator and published criteria, clearly labelled as such. A real borrowing amount needs a full application and a credit assessment by the bank."},{"question":"Who uses AI assistants for mortgages today?","answer":"Examples include Safe Rate in the US, whose AI assistant answers questions about rates and lenders, and Loft in Brazil, whose Gemini assistant lets real estate brokers run about 900 home financing simulations a week on WhatsApp, according to Google Cloud. Figure uses AI chatbots in home equity lending. None of these sources report conversion results."},{"question":"How do you keep answers current when rates change?","answer":"Treat rates and policy as versioned content with owners and effective dates, retire old versions on the day of change and rerun an automated test set that checks the quoted rates."}],"related":["conversational-loan-application-intake","branch-and-appointment-booking-agent","digital-onboarding-assistant","inbound-lead-qualification-agent","proactive-outbound-engagement-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI use case catalog and verified against Safe Rate, Loft, Figure, Lloyds Banking Group, Oper Credits and regulator sources."},{"date":"2026-09-25","note":"Consolidation pass: added EU Mortgage Credit Directive, ECOA and Regulation B to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: Loft now cites the Google Cloud case study (brokers run about 900 simulations a week, not buyers), Safe Rate moved to production on the evidence of its own website, Lloyds summary corrected, unsupported claims removed from the problem, Blits.ai channels aligned with the feature inventory, and SEO title and description added."},{"date":"2026-09-27","note":"Review fixes: meta description no longer states the Loft vendor figure without attribution, Lloyds income verification removed from the evidence (back office work, not a home loan assistant), Annex III guidance now links to the official EUR-Lex text, and the SEO title spelling matches the page."},{"date":"2026-09-27","note":"Review fixes: removed Oper Credits from this page's evidence, since it describes back office document verification rather than a customer facing home loan assistant."}],"slug":"home-loan-assistant-and-prequalification","url":"https://www.blits.ai/ai-use-cases/home-loan-assistant-and-prequalification","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":900,"min":900,"max":900,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"loft-whatsapp-mortgage-simulations","pooled":true}]}],"indicativeValueResult":{"low":300000,"high":1500000},"evidence":["figure-lending-chatbots","loft-whatsapp-mortgage-simulations","safe-rate-ai-mortgage-agent"]},{"title":"AI internal talent marketplace for matching employees to projects, roles and mentors","shortTitle":"Internal talent marketplace","seoTitle":"AI internal talent marketplace and mobility","metaDescription":"AI matches employees to internal gigs, roles and mentors by skills. Mastercard says 90% of its workforce is on its Unlocked talent marketplace.","definition":"An internal platform that uses AI to infer employees' skills and interests and recommend short term projects, open roles, mentors and learning to them, while showing managers which employees fit an opportunity, so that work is staffed from inside before hiring or contracting externally.","aliases":["talent marketplace","internal mobility platform","opportunity marketplace","AI skills matching","internal gig marketplace","career pathing AI"],"industries":["cross-industry","manufacturing","payments","government"],"functions":["human-resources"],"patterns":["recommendation-and-personalization","prediction-and-scoring"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","problem":"In many large organizations employees find internal opportunities through their network and their\nmanager, if at all. Open roles and short projects are advertised unevenly, skills are recorded in\njob titles rather than in anything searchable, and managers who need help quickly hire a contractor\nbecause they cannot see who inside the company has the skill and the time. People who want to grow\nleave to do it elsewhere: Schneider Electric's internal surveys, as reported by its platform vendor\nGloat, found that nearly half of departing employees cited a lack of internal growth opportunities\nas their main reason.\n\nAt the same time, skills needs change faster than job architectures. HR teams want to know which\nskills they have, where the gaps are and how to redeploy people, but a skills inventory built by\nhand soon falls behind.","problemStats":[],"howItWorks":"1. **Build a skills profile.** The employee imports a CV or professional profile; AI extracts and\n   infers skills from it and from HR data, and the employee confirms, adds aspirations and\n   availability.\n2. **Post opportunities.** Managers post projects, gigs, open roles and mentoring offers with the\n   skills needed and the time involved.\n3. **Match both ways.** The platform recommends opportunities, mentors and learning to each\n   employee, and suggests candidates to the manager, ranked by skills fit and stated interest.\n4. **Apply and agree.** Employees apply, managers choose, and the employee's own manager agrees the\n   time commitment; the platform records the assignment.\n5. **Learn from the data.** Completed assignments update the skills profile, and aggregated skills\n   data shows HR where the gaps are for learning and hiring plans, as Mastercard describes.","valueDrivers":["employee-productivity","cost-to-serve","speed","inclusion-and-access"],"kpis":["employee-adoption","cost-savings","users-served"],"indicativeValue":{"referenceOrg":"A company with 20,000 employees","inputs":[{"key":"employees","label":"Employees with access to the marketplace","low":20000,"high":20000,"unit":"employees","note":"The reference company."},{"key":"gigShare","label":"Share of employees who complete an internal project or gig in a year","low":0.01,"high":0.02,"unit":"fraction of employees","note":"Editorial assumption, replace with your own participation data. None of the evidence records gives an annual participation rate."},{"key":"hoursPerGig","label":"Hours of work per project or gig","low":40,"high":60,"unit":"hours per gig","note":"Editorial assumption, replace with your own data. With these values the company logs 0.4 to 1.2 hours of internal gig work per employee a year. The evidence gives no annual rate and does not support one: Schneider Electric's hours are cumulative since its April 2020 launch with no end date, and Gloat's story is inconsistent (360,000 hours in its text, 550,000 in its header), which is 2.3 to 3.5 hours per employee in total for a workforce of about 155,000, so fewer per year."},{"key":"externalShare","label":"Share of those hours that would otherwise have been bought from contractors or new hires","low":0.2,"high":0.4,"unit":"fraction of hours","note":"Editorial assumption; much internal gig work would otherwise not be done at all."},{"key":"contractorRate","label":"Cost of an external contractor hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption."}],"formula":"employees * gigShare * hoursPerGig * externalShare * contractorRate","currency":"USD","period":"per year","resultLabel":"External contractor and hiring cost avoided","caveat":"Counts only avoided external spend on gig work. It leaves out the effect on retention and hiring for open roles, the value of work that would otherwise not be done, the time employees spend away from their main job, and the platform and change management costs."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The matching technology is available off the shelf. The hard parts are a usable skills taxonomy, manager behaviour (posting work and releasing people), works council and data protection agreements, and making the platform fair and explainable enough to be trusted.","dataPrerequisites":["Employee profiles with skills, experience and aspirations, confirmed by the employee","A skills taxonomy or ontology mapped to roles","Open roles, projects and mentoring offers with required skills and time","Policy on eligibility, time allowance and manager approval"],"integrations":["HR information system (employee records, org structure, job architecture)","Applicant tracking system for internal roles","Learning management system","Collaboration tools for notifications, such as Microsoft Teams"]},"implementation":{"steps":[{"title":"Agree the rules before the platform","detail":"Decide who can take part, how much time employees may spend on gigs, how managers approve and how internal applications are treated. Involve works councils early, because the system processes employee data and influences career decisions."},{"title":"Pilot in one function","detail":"Schneider Electric piloted in HR before a global launch. Choose a function with enough project work, seed it with real opportunities and measure participation and manager feedback."},{"title":"Let employees own their profile","detail":"Show employees the skills the system inferred and let them correct them. Profiles people trust produce matches people accept."},{"title":"Test the matching for fairness","detail":"Compare recommendation and selection rates across groups, check that protected characteristics and proxies are not used, and document the results."},{"title":"Connect it to workforce planning","detail":"Use aggregated skills data to plan learning and hiring, and report outcomes such as roles filled internally and contractor spend avoided, not only registrations."}],"guardrails":["Recommendations only; managers and employees make every selection decision","No protected characteristics or obvious proxies in matching features","Employees can see, correct and delete inferred skills","Regular fairness testing of recommendations and outcomes, with results retained","Transparent explanation of why an opportunity or candidate was suggested"],"humanInTheLoop":"Employees decide what to apply for, managers decide whom to select, and the employee's own manager agrees the time. HR owns the taxonomy, reviews fairness reports and handles complaints about recommendations.","kpisToInstrument":["Active users per month as a share of employees, not only registrations","Roles and projects filled internally, and time to fill","Recommendation acceptance rate and selection rate by group","Contractor spend and external hires avoided for filled gigs","Retention of participants versus comparable non participants"],"failureModes":[{"title":"Registrations without activity","detail":"High sign up figures hide low regular use. Report monthly active users and filled opportunities."},{"title":"Managers do not release people","detail":"Employees apply but their managers block the time. Set a time allowance policy and make releasing talent part of manager goals."},{"title":"Biased matching","detail":"Skills inferred from past roles reproduce past inequalities in who got which job. Test outcomes by group and let people correct their profiles."},{"title":"Stale skills data","detail":"Profiles are filled once and never updated. Refresh them from completed assignments and learning, and prompt employees periodically."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 4 lists AI used for the recruitment or selection of natural persons (4(a)) and AI used to make decisions affecting promotion, or to allocate tasks based on individual behaviour, personal traits or characteristics (4(b)). A marketplace that ranks employees for internal roles or allocates projects on the basis of inferred traits is therefore high risk. Recommending learning content or mentors to an employee who chooses freely is usually not. Deployers of the high risk part must inform workers' representatives and the affected employees before use (Article 26)."},"regulations":["eu-ai-act","gdpr","uk-gdpr","nyc-local-law-144","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4 covers employment, workers' management and access to self employment, including selection, promotion and task allocation."},{"title":"Article 26, obligations of deployers of high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/26/","note":"Human oversight, logs, and informing workers' representatives and affected workers before a high risk system is used at the workplace."},{"title":"Automated employment decision tools (Local Law 144)","issuer":"New York City Department of Consumer and Worker Protection","region":"north-america","url":"https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page","note":"Employers may use an automated employment decision tool only after a bias audit, with the results published and notices given to employees or job candidates. Because the notices also cover employees, check whether a tool that ranks employees for internal roles is in scope under the law's definitions."}],"controls":["AI Act classification per feature, with the high risk features managed as such","Data protection impact assessment and works council agreement where required","Fairness testing before launch and at least yearly, with results retained","Named HR owner for the matching logic and for complaints","Supplier due diligence on model documentation, data use and hosting"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai is not a talent marketplace product, but the conversational layer and the matching\nworkflow can be built on it. An **AI agent** in **Microsoft Teams** or the company's intranet\nhelps employees describe their skills and interests in their own words and answers questions\nabout the programme from a **knowledge base** of the mobility policy. **Custom functions** read\nopen projects and roles from the HR system (the integration catalog includes Workday) and write\nback applications.\n\nMatching runs as an **agentic workflow** with **structured output** that explains, for each\nsuggestion, which skills matched and which are missing, and any suggestion that reaches a\nmanager for an internal role waits for **human in the loop** handling, with a full audit trail.\n**Guardrails** stop the agent from asking about or using protected characteristics, **PII masking**\nand the **GDPR toolkit** handle retention and removal requests, and **test suites** can replay\nmatched employee profiles that differ only in a protected attribute to check for inconsistent\ntreatment."},"faq":[{"question":"What results do companies report from AI talent marketplaces?","answer":"Mastercard says 90% of its workforce is on its Unlocked marketplace, with 500,000 project hours delivered. Gloat reports that more than 2,300 Schneider Electric employees began to explore new roles in the first two months, and over USD 15 million in productivity gains and reduced recruitment costs since the 2020 launch. Registration is not the same as regular use, so track active users and filled opportunities."},{"question":"Is an internal talent marketplace high risk under the EU AI Act?","answer":"Parts of it can be. Ranking employees for internal roles or allocating tasks based on inferred traits falls under Annex III point 4. Recommending mentors or courses that the employee chooses freely usually does not. Classify each feature and inform workers' representatives before using a high risk one."},{"question":"Does it replace internal recruiters?","answer":"No. It makes opportunities visible and suggests matches; managers and recruiters still select. The gain is in speed and reach, especially for short projects that would otherwise go to a contractor."}],"related":["recruitment-screening-and-interview-scheduling","employee-onboarding-assistant","hr-and-policy-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched and written with evidence from Mastercard, Schneider Electric, Unilever and the US Federal Bureau of Prisons. Editor review removed registration and undefined adoption rates from the employee adoption KPI, recorded Schneider Electric's 2,300 employees as users served, and lowered the participation and gig hours assumptions so they stay conservative against the cumulative hours in the evidence."}],"slug":"internal-talent-marketplace-matching","url":"https://www.blits.ai/ai-use-cases/internal-talent-marketplace-matching","benchmarks":[{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":15000000,"min":15000000,"max":15000000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"schneider-electric-open-talent-market","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2300,"min":2300,"max":2300,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"schneider-electric-open-talent-market","pooled":true}]}],"indicativeValueResult":{"low":96000,"high":960000},"evidence":["federal-bureau-of-prisons-pathfinder-career-pathways","mastercard-unlocked-talent-marketplace","schneider-electric-open-talent-market","unilever-flex-experiences-talent-marketplace"]},{"title":"AI knowledge assistant for wealth advisors and relationship managers","shortTitle":"Advisor knowledge assistant","seoTitle":"AI knowledge assistant for wealth advisors","metaDescription":"Advisor assistants answer from house research, product and policy documents, with sources. Morgan Stanley says 98% of its Financial Advisor teams adopted one.","definition":"A conversational assistant that answers a wealth advisor's or relationship manager's questions in seconds from the firm's own research, house view, product documentation and policies, with every answer linked to the source document so the advisor can check it before using it with a client.","aliases":["advisor knowledge copilot","relationship manager assistant","RM copilot","wealth advisor chatbot"],"industries":["wealth-and-asset-management","banking"],"functions":["knowledge-management","sales","customer-service"],"patterns":["rag-knowledge-assistant","conversational-agent"],"channels":["internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"mainstream","segment":"front-office","problem":"A wealth advisor is expected to know the firm's view on markets, sectors and asset classes, the\nterms of hundreds of products, tax and structuring notes for several jurisdictions, and the\npolicies that govern what may be offered to whom. That knowledge lives in research libraries,\nproduct term sheets, intranet pages and email, and it changes often. Finding the right paragraph\nwhile a client waits is slow, so advisors lean on memory, ask a colleague or promise to call back.\n\nThe result is inconsistent answers, lost time before and during client conversations, and a real\nconduct risk when an advisor paraphrases an outdated view or a product rule from memory. Junior\nadvisors and relationship managers in new markets are hit hardest, because they do not yet know\nwhere anything lives.","problemStats":[],"howItWorks":"1. **Curate the corpus.** Research notes, the house view, product documents, tax and structuring\n   notes and policies are loaded into a knowledge base with an owner, a publication date and an\n   audience per document.\n2. **Ask in plain language.** The advisor asks \"what is our current view on European banks\" or\n   \"can this structured note be sold to a client in Hong Kong\" in chat or inside the collaboration\n   tool.\n3. **Retrieve and answer with citations.** Hybrid search finds the relevant passages; the model\n   answers only from them and links each statement to its source document and date.\n4. **Respect entitlements.** The assistant only retrieves documents the advisor may see, and it\n   refuses questions it cannot answer from approved content instead of guessing.\n5. **Learn from feedback.** Thumbs down, unanswered questions and stale document hits go to the\n   content owners, who fix the corpus rather than the prompt.","valueDrivers":["employee-productivity","compliance","customer-experience","speed"],"kpis":["employee-adoption","handling-time-reduction","time-saved-per-task","interactions-handled","accuracy"],"indicativeValue":{"referenceOrg":"A wealth manager with 500 client facing advisors","inputs":[{"key":"advisors","label":"Client facing advisors","low":500,"high":500,"unit":"advisors","note":"The reference firm."},{"key":"questionsPerWeek","label":"Knowledge questions per advisor per week","low":5,"high":15,"unit":"questions per advisor per week","note":"Editorial assumption, replace with your own search and help desk volumes."},{"key":"minutesSaved","label":"Minutes saved per question","low":3,"high":10,"unit":"minutes per question","note":"Conservative against the benchmark on this page (J.P. Morgan reports advisers find the right information up to 95% faster); most questions are short lookups."},{"key":"weeks","label":"Working weeks per year","low":46,"high":46,"unit":"weeks per year","note":"Editorial assumption."},{"key":"hourlyCost","label":"Fully loaded advisor cost per hour","low":80,"high":150,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"advisors * questionsPerWeek * minutesSaved / 60 * weeks * hourlyCost","currency":"USD","period":"per year","resultLabel":"Value of advisor time released from searching","caveat":"Time released is only value if advisors spend it with clients. The estimate leaves out the cost of running the assistant and curating content, and the harder to measure benefit of fewer answers given from outdated material."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"The model work is simple; the effort is in the content. Research, product and policy documents need owners, review dates and audience tags, and entitlements must follow the advisor's booking centre and licence.","dataPrerequisites":["Research library and house view with publication dates and authors","Product term sheets, key information documents and product governance rules per market","Tax, structuring and policy notes with an owner and a review date","Entitlement data that maps advisors to booking centres, licences and client segments"],"integrations":["Research and content management systems","Document repositories such as SharePoint","Identity and access management for document level permissions","Collaboration tools such as Microsoft Teams, or the advisor desktop"]},"implementation":{"steps":[{"title":"Start with one corpus and one audience","detail":"Pick the content advisors search most (usually the house view and product documentation) for one booking centre. Measure what they ask in the first weeks before adding more sources."},{"title":"Fix the content before the model","detail":"Remove superseded documents, tag every document with an owner, audience and expiry date, and agree who updates it. An assistant that retrieves a superseded document gives a wrong answer with a valid looking citation."},{"title":"Make citations mandatory","detail":"Every answer shows its source passages and dates, and the assistant refuses when retrieval finds nothing relevant. Advisors must be able to check an answer in one click."},{"title":"Enforce entitlements at retrieval time","detail":"Filter documents by the advisor's permissions before they reach the model, so a restricted research note or a product not approved in that market never appears in an answer."},{"title":"Build a test set from real questions","detail":"Collect a few hundred real advisor questions with approved answers and run them on every content or model change, including questions that must be refused."},{"title":"Roll out with training and feedback loops","detail":"Train advisors that the assistant informs and they decide, publish usage and feedback per desk, and route every thumbs down to the content owner."}],"guardrails":["Answers only from approved, dated content, with a refusal when nothing relevant is retrieved","Citation of the source document and date on every answer","Document level permissions applied before retrieval, per booking centre and licence","No personalized recommendations; the assistant informs, the advisor advises","PII masking so client names and account data are not sent to the model unless required"],"humanInTheLoop":"The advisor decides what, if anything, to tell a client and stays responsible for the advice. Content owners review flagged answers weekly and research or product governance approves any new corpus before it is added.","kpisToInstrument":["Weekly active advisors as a share of licensed advisors","Share of questions answered with a citation versus refused","Answer accuracy on a monthly human reviewed sample","Median time to answer compared with the previous search process","Feedback rate and top unanswered topics"],"failureModes":[{"title":"Stale house view","detail":"The assistant quotes last quarter's view because the old note was never retired. Expire documents automatically and prefer the newest version at retrieval."},{"title":"Answers that look like advice","detail":"Advisors paste a fluent answer into a client email without checking it. Train on the assist posture, keep citations visible and log what is copied."},{"title":"Entitlement leakage","detail":"A restricted or wrong market document appears in an answer. Filter by permission before retrieval, not after generation."},{"title":"Adoption stalls after launch","detail":"Advisors try it, get a poor answer and stop. Seed the corpus with the most searched content and publish improvements."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1) requires that people who interact directly with an AI system are informed of it, unless this is obvious from the context, as it usually is for an internal assistant labelled as AI; Article 50(2) requires providers of systems that generate text to mark the output as AI generated in a machine readable way. Helping advisors find information is not an Annex III use and not a prohibited practice under Article 5. It would become high risk only if the system were used to evaluate the creditworthiness of clients (point 5(b)) or to evaluate or make decisions about advisors (point 4(b)). If the assistant were opened to clients, they would have to be told they are dealing with AI."},"regulations":["eu-ai-act","gdpr","mifid-ii","dora","mas-ai-risk-management","iso-42001"],"guidance":[{"title":"ESMA public statement on the use of AI in the provision of retail investment services","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf","note":"Decisions remain the responsibility of the management body whether taken by people or AI tools, including third party AI used by staff; MiFID II organisational, conduct and record keeping requirements apply."},{"title":"Artificial Intelligence (AI) Model Risk Management (information paper)","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Good practices observed in MAS's mid 2024 thematic review of banks' AI and generative AI model risk management, covering governance and oversight, key risk management systems and processes, and development and deployment."},{"title":"Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT) in the Use of Artificial Intelligence and Data Analytics in Singapore's Financial Sector","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2018/feat","note":"Foundational principles for firms offering financial products and services on the responsible use of AI and data analytics, including internal governance, accountability and transparency."}],"controls":["Entry in the AI inventory with an accountable business owner and a content owner per corpus","Document ownership, audience tags and expiry dates enforced in the knowledge base","Retrieval and answer logs kept for supervision and investigation","Regression test set run on every model or content change","Clear written rule that the assistant supports but does not replace the advisor's judgment"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** grounded in a **knowledge base** that holds the research\nlibrary, house view and product documents, retrieved with **hybrid search** (vector plus BM25).\nDocument ingestion covers PDF, Word, PowerPoint and email files, and the document library keeps\nversions so an outdated note can be reverted or removed. The agent is instructed to answer only\nfrom retrieved passages and to cite them.\n\nAdvisors reach it in **Microsoft Teams** or through the API inside their own desktop.\n**Guardrails** check inputs and outputs, **PII masking** keeps client data out of prompts, and\nrole based access controls who can change the agent. **Test suites** replay real advisor\nquestions with LLM grading against approved answers on every change, and analytics and feedback\nshow adoption and unanswered topics. The platform is model agnostic and can run in the EU or UAE\nregion for data residency."},"faq":[{"question":"How many advisors actually use these assistants?","answer":"Where firms publish figures, adoption is high. Morgan Stanley said in June 2024 that 98% of its Financial Advisor teams had adopted its assistant, and Bank of America reports more than 23 million interactions with ask MERRILL and ask PRIVATE BANK in 2024."},{"question":"How do you stop the assistant from giving wrong answers to clients?","answer":"Ground it only in approved, dated content, show the source with every answer and make it refuse when nothing relevant is found. The advisor, not the assistant, talks to the client and checks the source first."},{"question":"Is this the same as enterprise knowledge search?","answer":"It uses the same retrieval pattern, but the corpus and controls are specific to advice: research and house view, product governance per market, and entitlements by booking centre and licence. Those controls matter because the answers feed conversations with clients."}],"related":["enterprise-knowledge-search","investment-research-summarization","next-best-action-for-advisors","client-meeting-notes-and-crm-update","suitability-assessment-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Advisor Knowledge Copilot) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: EU AI Act tier corrected to limited (Article 50), MiFID II added, FEAT guidance note and FAQ adoption claim made accurate, unsupported pilot claim removed, SEO title and description added; evidence dates and summaries corrected."},{"date":"2026-09-26","note":"Second fact check against sources: MAS guidance title and note aligned with the page, meta description made more concrete; Yes Bank languages removed (not stated) and UBS rollout scope made precise."}],"slug":"wealth-advisor-knowledge-assistant","url":"https://www.blits.ai/ai-use-cases/wealth-advisor-knowledge-assistant","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":23000000,"min":23000000,"max":23000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-ask-merrill-and-ask-private-bank","pooled":true}]}],"indicativeValueResult":{"low":460000,"high":8625000},"evidence":["bank-of-america-ask-merrill-and-ask-private-bank","citi-wealth-askwealth-and-advisor-insights","jpmorgan-coach-ai-advisers","morgan-stanley-ai-assistant-knowledge-search","ubs-red-client-advisor-assistants","yes-bank-rm-assist-chatbot"]},{"title":"AI lease abstraction for commercial real estate","shortTitle":"Lease abstraction","seoTitle":"AI lease abstraction for commercial real estate","metaDescription":"AI reads commercial leases and letters of intent into structured data. Cushman & Wakefield and JLL use it to abstract leases that used to take hours to days.","definition":"AI that reads a commercial lease, amendment or letter of intent and extracts the key terms, such as parties, dates, rent, escalations, options and renewal notices, into structured data, so a property owner, occupier or brokerage does not retype every clause by hand into its lease administration and portfolio systems.","aliases":["AI lease review","lease data extraction","LOI abstraction","lease administration automation"],"industries":["real-estate"],"functions":["operations","legal"],"patterns":["document-processing","agentic-workflow","rag-knowledge-assistant"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"lease-administration","problem":"Every commercial lease, amendment and letter of intent has to be read before its terms can be\nused: the dates that trigger a renewal or a break option, the rent and how it escalates, the\nclauses buried in a rider or an exhibit. Unframe's case study on Cushman & Wakefield reports that\nabstracting a single lease could take anywhere from six hours to three days, depending on the\ndocument's complexity, and that none of the vendors or internal approaches the firm evaluated\ncombined the accuracy and the speed it needed at scale.\n\nLeases are also rarely standardized and can run to 100 pages or more, Unframe's case study notes,\nwhich makes consistent, accurate extraction harder to achieve at scale.","problemStats":[],"howItWorks":"1. **Ingest the document.** The lease, amendment or letter of intent arrives as a PDF or scan of\n   any length, in any of the organization's operating languages.\n2. **Extract the standard fields.** The AI reads the document and produces a first pass abstract\n   against a defined schema: parties, term dates, rent, escalations, options, renewal notices and\n   similar fields.\n3. **Flag what does not fit the schema.** Non standard clauses, unusual riders and anything the\n   model is not confident about are flagged rather than guessed.\n4. **A person validates the exceptions.** Lease administration or legal staff review the flagged\n   clauses and the model's confidence, not every field on every lease.\n5. **The data flows downstream.** Confirmed fields post into the lease administration or portfolio\n   management system, so accounting, reporting and renewal tracking work from the same structured\n   record.\n6. **Answer questions from the corpus.** The same extracted data and source documents let brokers\n   and lease administrators ask questions about a specific lease or compare draft letters of\n   intent to each other.","valueDrivers":["cost-to-serve","employee-productivity","speed"],"kpis":["processing-time-reduction","productivity-gain","hours-saved","accuracy"],"indicativeValue":{"referenceOrg":"A commercial real estate services firm abstracting 10,000 leases and LOIs a year","inputs":[{"key":"documentsPerYear","label":"Leases and LOIs abstracted per year","low":5000,"high":20000,"unit":"documents per year","note":"Editorial assumption, replace with your own volume."},{"key":"hoursSavedPerDocument","label":"Analyst hours saved per document","low":2,"high":5,"unit":"hours saved per document","note":"Conservative against the two deployments on this page: Unframe's case study reports that abstraction used to take six hours to three days per lease before AI, and Cadastral reports JLL's brokerage teams now generate lease and LOI abstracts within seconds."},{"key":"costPerHour","label":"Fully loaded cost of a lease administration or legal analyst hour","low":40,"high":90,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"documentsPerYear * hoursSavedPerDocument * costPerHour","currency":"USD","period":"per year","resultLabel":"Lease abstraction labor cost avoided","caveat":"Gross labor cost avoided only. It leaves out the software cost, the time still needed to review flagged exceptions, and any recovered revenue from clauses the AI surfaces that a manual review would have missed, which neither deployment on this page reports as a company wide figure."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Extracting standard fields such as dates and base rent is well understood; the hard part is the long tail of non standard clauses, exhibits, amendments and, for a global portfolio, multiple languages and currencies, plus wiring confirmed output into the lease administration system instead of leaving it in a spreadsheet.","dataPrerequisites":["A defined extraction schema of the fields the organization actually uses downstream","A library of past leases, amendments and letters of intent in digital form","A golden set of correctly abstracted leases to measure accuracy against"],"integrations":["Document management or electronic signature system where leases are stored","Lease administration or portfolio management system, the destination for extracted data","Optical character recognition for scanned or image based leases"]},"implementation":{"steps":[{"title":"Define the schema before the pilot","detail":"Agree the exact fields lease administration, accounting and portfolio teams need, not every field a model can technically extract, and use that as the acceptance test."},{"title":"Pilot on one lease type","detail":"Start with the most common, most standardized lease template in the portfolio before adding ground leases, sale and leaseback structures or multi tenant riders."},{"title":"Route exceptions to a person by default","detail":"Treat every non standard clause and every low confidence extraction as a review item, not an accepted answer, until the exception rate on that clause type is proven low."},{"title":"Validate against a golden set on every change","detail":"Keep a fixed set of leases with a confirmed correct abstract and rerun it whenever the model, prompt or schema changes, so accuracy regressions are caught before they reach production."},{"title":"Wire the output into the system that uses it","detail":"Post confirmed fields into the lease administration or portfolio system automatically; extraction that stays in a standalone tool does not save the downstream re entry it is meant to remove."}],"guardrails":["Every non standard or low confidence clause is routed to a person before the data is relied on","Extracted values are shown next to the source clause so a reviewer can check them in seconds","Version control tracks which document version, and which amendment, an abstract came from"],"humanInTheLoop":"Lease administration or legal staff confirm every flagged exception before it reaches the lease administration system, and a sample of fields the model marked confident are spot checked on a schedule so silent accuracy drift is caught early.","kpisToInstrument":["Processing time per document, split by lease type and by whether it needed a review","Share of extracted fields confirmed correct on a review sample","Count of clauses, such as escalations or options, recovered that a prior manual process missed"],"failureModes":[{"title":"Non standard clauses misread as standard","detail":"A clause with unusual wording gets mapped to the wrong field or missed entirely because it does not match the schema; a human review pass and a growing library of confirmed exceptions catch this over time."},{"title":"Abstracted data nobody uses","detail":"Extraction is treated as a one off clean up project rather than wired into the lease administration and portfolio systems, so the structured data goes stale as new leases and amendments arrive."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Extraction alone is minimal risk: reading and structuring the terms of a commercial contract does not decide credit, employment, insurance, biometric identification or another use listed in Annex III, so it carries only the Article 4 AI literacy duty. The conversational assistant that lets employees ask questions about a lease adds Article 50(1): people who interact directly with it must be told they are dealing with an AI system, unless that is obvious from the context, as it usually is for an internal tool."},"regulations":["eu-ai-act","gdpr"],"guidance":[],"controls":["A person confirms every non standard or low confidence clause before it is used in financial reporting or a renewal decision","An audit trail records which user confirmed the abstract that downstream numbers rely on"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai, lease abstraction runs as an AI agent with structured output configuration, so the\nmodel returns each document's fields, such as dates, rent and escalations, as machine readable\ndata instead of prose. A knowledge base ingests lease PDFs and images of exhibits and\namendments, and a small, scoped set of custom functions writes the confirmed fields into the\norganization's lease or portfolio management system through its API, once a person has reviewed\nthe fields the model flagged as an exception.\n\nThe same knowledge base, retrieved with hybrid search, powers a conversational assistant that\nanswers questions about a specific lease or compares draft letters of intent, in the employee's\nown language. Test suites regress extraction accuracy on a fixed set of leases before every\nchange, full execution tracing lets a reviewer inspect every agent turn behind an abstract, with\ntoken counts and trace details, and the platform is model agnostic, so a firm can route\nextraction to a different provider without rebuilding the workflow."},"faq":[{"question":"How accurate is AI lease abstraction?","answer":"Neither deployment on this page discloses a checked accuracy percentage. Cadastral reports that JLL's brokerage teams now generate lease and LOI abstracts within seconds rather than manually; treat every extraction as a first pass that a person confirms before it feeds financial reporting or a renewal decision."},{"question":"What time does lease abstraction actually save?","answer":"Unframe's case study reports that abstracting a single lease used to take six hours to three days before deployment; afterward, it reports that Cushman & Wakefield's brokers can access lease insights in real time, without a stated company wide time or cost figure."},{"question":"Can it replace a lease administrator?","answer":"No. Neither case study on this page describes the review step, so plan for one: route non standard clauses and low confidence extractions to a lease administrator or legal reviewer before the data feeds financial reporting or a renewal decision. Treat the AI as a copilot that produces a first pass, not a replacement for that review."},{"question":"Does it also handle letters of intent, not just signed leases?","answer":"Yes in the JLL deployment. Cadastral reports that JLL's brokers use the platform to generate LOI abstracts and to compare draft letters of intent to each other, alongside signed lease abstraction."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched for the real estate and energy scope, with JLL's Cadastral deployment and Cushman & Wakefield's Unframe deployment checked against the primary sources."},{"date":"2026-09-28","note":"Fact checked against sources: attributed the six hours to three days baseline and the any length, multi language claim to Unframe's case study instead of Cushman & Wakefield or Ross Hodges directly, removed the unsourced OCR sentence and the invented paragraph 47 detail, rewrote the FAQ on replacing a lease administrator as editorial advice rather than a claim about the case studies, corrected the tracing and document ingestion description against the platform feature inventory, set the EU AI Act tier to context dependent to cover the conversational assistant's Article 50(1) exposure, corrected the JLL evidence year to 2026 from its earliest Wayback capture and added that as archivedUrl, fixed two blended or misquoted phrases in the evidence verification notes, and switched remaining British spellings to US spelling."}],"slug":"lease-abstraction","url":"https://www.blits.ai/ai-use-cases/lease-abstraction","benchmarks":[],"indicativeValueResult":{"low":400000,"high":9000000},"evidence":["cushman-wakefield-unframe-lease-abstraction","jll-cadastral-lease-abstraction"]},{"title":"AI legal research and drafting assistant for lawyers","shortTitle":"Legal research and drafting","seoTitle":"AI legal research and drafting assistant","metaDescription":"Generative AI that researches law and drafts first versions for lawyers to verify. Ashurst measured 45% time saved on first draft briefings in controlled trials.","definition":"A generative AI assistant for lawyers in firms, legal departments and public bodies that finds and summarises case law, legislation and internal know how, answers legal questions with citations and drafts first versions of memos, briefings, letters and filings, which a lawyer verifies and signs off.","aliases":["legal AI assistant","generative AI for lawyers","AI legal research","legal copilot","AI legal drafting","legal memo drafting AI"],"industries":["professional-services","cross-industry","government"],"functions":["legal","knowledge-management"],"patterns":["rag-knowledge-assistant","content-generation","summarization","document-processing"],"channels":["internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"A large part of a lawyer's time goes into work that is necessary but repetitive: finding the\nrelevant authorities, reading them, extracting what matters from long documents and turning it\ninto a first draft of a memo, a briefing note or a filing. Much of it falls to junior lawyers,\nclients question paying for it by the hour, and in public bodies the same work competes with heavy\ncaseloads and vacancies.\n\nGeneral purpose chatbots look like an answer but are dangerous here. They produce fluent text with\ninvented case citations, and courts have started to take action against lawyers who filed them. The High\nCourt of England and Wales stated in 2025 that freely available generative AI tools trained on a\nlarge language model are not capable of conducting reliable legal research. The useful version is\nan assistant grounded in authoritative legal sources and the organization's own precedents, that\nshows where every statement comes from and leaves the judgment to a lawyer.","problemStats":[],"howItWorks":"1. **Ask in plain language.** The lawyer asks a question or describes the task (\"summarise the\n   limitation rules for this claim in these three jurisdictions\", \"draft a first briefing on this\n   article of association\").\n2. **Retrieve from trusted sources.** The assistant searches licensed legal databases, the\n   organization's precedent bank and the documents of the matter, not the open web by default.\n3. **Answer with citations.** It answers or drafts with a citation for each proposition, linked to\n   the passage it relied on, and says when it found nothing.\n4. **Extract and compare.** For document heavy tasks it pulls defined points from many documents\n   into a table, for example clauses, dates or obligations, with references back to the source.\n5. **Verify and finish.** The lawyer checks every citation and conclusion, edits the draft and\n   records the result in the matter file.","valueDrivers":["employee-productivity","speed","cost-to-serve","risk-reduction"],"kpis":["productivity-gain","time-saved-per-task","users-served","interactions-handled","employee-adoption"],"indicativeValue":{"referenceOrg":"A law firm or legal department with 100 lawyers","inputs":[{"key":"lawyers","label":"Lawyers using the assistant","low":100,"high":100,"unit":"lawyers","note":"The reference organization."},{"key":"hoursPerLawyer","label":"Hours per lawyer per year on research, extraction and first drafts","low":300,"high":500,"unit":"hours per lawyer per year","note":"Editorial assumption, replace with your own time recording data."},{"key":"timeSaved","label":"Share of those hours saved","low":0.15,"high":0.3,"unit":"fraction of hours","note":"Conservative against Ashurst's controlled experiments on this page (about 45% time saved on first draft briefings), because verification of citations takes time and not every task suits the tool."},{"key":"hourlyCost","label":"Internal cost of a lawyer hour","low":100,"high":200,"unit":"USD per hour","note":"Editorial assumption for a blended internal cost, not the billing rate."}],"formula":"lawyers * hoursPerLawyer * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Lawyer time released","caveat":"Values released lawyer time at internal cost. It leaves out licence and legal database costs, the time spent verifying outputs, any change in billable revenue under hourly billing, and the cost of an error that verification misses."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Buying a legal AI product is easy; making it trustworthy is not. The work is in connecting authoritative sources and the organization's own know how, keeping client data confidential, training lawyers to verify, and defining which tasks the tool may be used for.","dataPrerequisites":["Licensed access to the legal databases lawyers already rely on","A curated precedent and know how bank with owners and review dates","Matter documents with access rights that the assistant respects","A written policy on permitted uses, client consent and verification"],"integrations":["Legal research databases","Document management system and precedent bank","Matter management and time recording","Identity and access management for matter level permissions"]},"implementation":{"steps":[{"title":"Start with a structured trial","detail":"Choose a few task types, such as first draft briefings, research summaries and document extraction, and measure time and quality against a control group. Ashurst ran three trials with 411 people and measured time savings in controlled experiments, and A&O ran a beta from November 2022 before its firm wide rollout in February 2023."},{"title":"Ground it in trusted sources","detail":"Connect licensed legal content and the organization's precedents, and turn off open web answers for legal questions. Require a citation for every proposition."},{"title":"Make verification part of the workflow","detail":"Treat every output as a trainee's draft. Build citation checking into the process and make the reviewing lawyer's sign off visible in the matter file."},{"title":"Protect confidentiality","detail":"Use a deployment where client data is not used to train models, respects matter access rights and stays in the required region. Record client consent where engagement terms require it."},{"title":"Train, then widen","detail":"Teach lawyers what the tool is good and bad at, share good prompts per practice area and track adoption and quality before opening it to more teams."}],"guardrails":["Citation for every legal proposition, linked to the source passage","Refusal when no authority is found, instead of a plausible guess","Matter level access control so the assistant never mixes clients","No client data used for model training, with data kept in the required region","Lawyer verification and sign off before anything leaves the organization or reaches a court"],"humanInTheLoop":"A lawyer owns every output: they check each citation against the source, decide the legal position and sign the document. The assistant never files, sends or advises on its own.","kpisToInstrument":["Time to first draft for defined task types, against a control group","Share of citations that fail verification on a sample","Weekly active lawyers as a share of licensed lawyers","Reviewer edits per draft","Client or court complaints linked to AI assisted work"],"failureModes":[{"title":"Invented or misquoted authorities","detail":"The model cites cases that do not exist or do not say what it claims. Courts in the US and the UK have taken action against lawyers for this. Ground answers in authoritative sources and verify every citation."},{"title":"Automation bias","detail":"Fluent drafts are accepted without enough scrutiny, especially under time pressure. Set a fixed verification checklist per task type and sample reviewed work, instead of relying on reviewers to notice when something reads wrong."},{"title":"Confidentiality breach","detail":"Client documents are pasted into a consumer tool. Provide an approved tool and block the rest."},{"title":"Adoption without measurement","detail":"Licences are bought but usage and quality are not tracked. Measure time saved and error rates per task type and cut the tasks where the tool does not help."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Research and drafting support for lawyers in firms and companies is not listed in Annex III, so it is normally minimal risk with AI literacy duties. Annex III point 8(a) makes it high risk when a judicial authority, or someone on its behalf, uses AI to research and interpret facts and the law and to apply the law to a concrete set of facts, or when it is used in a similar way in alternative dispute resolution, so a deployment for courts, tribunals or arbitration needs its own classification."},"regulations":["eu-ai-act","gdpr","uk-gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Risk Outlook report: The use of artificial intelligence in the legal market","issuer":"Solicitors Regulation Authority","region":"europe","url":"https://www.sra.org.uk/sra/research-publications/artificial-intelligence-legal-market/","note":"Explains hallucination and says that however well controlled a system is, firms still need to check its outputs for accuracy."},{"title":"Generative AI, the essentials","issuer":"The Law Society of England and Wales","region":"europe","url":"https://www.lawsociety.org.uk/topics/ai-and-lawtech/generative-ai-the-essentials","note":"Practical guidance for solicitors on risks such as hallucination, confidentiality and professional duties when using generative AI."},{"title":"Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin)","issuer":"High Court of England and Wales (Divisional Court)","region":"europe","url":"https://www.judiciary.uk/wp-content/uploads/2025/06/Ayinde-v-London-Borough-of-Haringey-and-Al-Haroun-v-Qatar-National-Bank.pdf","note":"Sets out lawyers' duties when using AI for research and warns of sanctions, including referral to regulators, for citing fictitious authorities."}],"controls":["Approved tool list and a written acceptable use policy for generative AI","Citation verification step recorded in the matter file","Data protection impact assessment and supplier due diligence on data use and location","Training for lawyers before access, with refreshers when tools change","Sampling of AI assisted work by a senior lawyer"],"incidents":[{"title":"ChatGPT Reportedly Produced False Court Case Law Presented by Legal Counsel in Court","url":"https://incidentdatabase.ai/cite/541/","note":"AI Incident Database entry 541. In Mata v. Avianca, lawyers filed a brief citing cases invented by ChatGPT, and a US federal judge ordered them to show cause why they should not be sanctioned."},{"title":"Ayinde and Al-Haroun, fictitious citations in court documents","url":"https://www.judiciary.uk/wp-content/uploads/2025/06/Ayinde-v-London-Borough-of-Haringey-and-Al-Haroun-v-Qatar-National-Bank.pdf","note":"The Divisional Court dealt with two cases in which documents placed before the court cited authorities that did not exist, and referred lawyers to their regulators."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** that holds the organization's\nprecedents, practice notes and matter documents, retrieved with **hybrid retrieval** (vector and\nBM25) so exact statutory references and case names are found, and answering with the retrieved\npassages. Licensed legal databases and the document management system can be connected as tools\nthrough **custom functions** or **MCP servers**, and the agent is instructed to refuse when\nretrieval finds nothing. It runs inside **Microsoft Teams** or an internal web app through the\n**REST API**.\n\nLonger tasks, such as extracting defined points from a set of documents into a table, run as an\n**agentic workflow** with **structured output** and an audit trail per run. **PII masking** at\nthe gateway, **role based access control** and **EU or UAE data residency** protect client data,\nand the platform is **model agnostic**, so the legal team can compare models on its own test\nquestions. **Test suites** with lawyer written questions and expected authorities catch\nregressions when prompts, content or models change."},"faq":[{"question":"How much time does a legal AI assistant save?","answer":"In Ashurst's controlled experiments, the firm measured approximate time savings of 45% on first draft legal briefings, 59% on sector research reports and 80% on UK corporate filings that required extracting information from articles of association. Those were trial conditions with small groups; real savings depend on how much verification the task needs."},{"question":"Can lawyers rely on ChatGPT for legal research?","answer":"Not on its own. The High Court of England and Wales said in 2025 that freely available generative AI tools are not capable of conducting reliable legal research, and courts in the UK and the US have taken action against lawyers over fictitious citations. Use a tool grounded in authoritative sources and verify every citation."},{"question":"Who uses it at scale?","answer":"Allen & Overy rolled out Harvey to more than 3,500 lawyers in 43 offices in 2023 after a trial in which lawyers asked around 40,000 queries. In the US federal government, the Department of Justice and the SEC use AI features in legal research services, and the SEC is piloting a generative assistant."},{"question":"Is a legal research assistant high risk under the EU AI Act?","answer":"For law firms and legal departments, normally not. It becomes high risk when a judicial authority uses it to research and interpret facts and law and apply the law to a case, or when it is used in a similar way in alternative dispute resolution (Annex III point 8(a))."}],"related":["ediscovery-and-disclosure-document-review","court-and-case-file-summarization","deal-sourcing-and-due-diligence-assistant","procurement-contract-review"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched and written with evidence from A&O Shearman, Ashurst (now Ashurst Perkins Coie), the US Department of Justice and the SEC. Editor pass cited Ashurst's full Vox PopulAI report on ashurstperkinscoie.com instead of the moved press release, dropped the 2.5 hours figure, corrected the blind study summary and narrowed the DOJ record to what the inventory states."}],"slug":"legal-research-and-drafting-assistant","url":"https://www.blits.ai/ai-use-cases/legal-research-and-drafting-assistant","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40000,"min":40000,"max":40000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"allen-and-overy-harvey-legal-assistant","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":45,"min":45,"max":45,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ashurst-generative-ai-legal-drafting-trials","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":3500,"min":3500,"max":3500,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"allen-and-overy-harvey-legal-assistant","pooled":true}]}],"indicativeValueResult":{"low":450000,"high":3000000},"evidence":["allen-and-overy-harvey-legal-assistant","ashurst-generative-ai-legal-drafting-trials","doj-ai-assisted-legal-research","sec-cocounsel-legal-research-pilot"]},{"title":"AI localization of marketing, product and web content","shortTitle":"Content localization","seoTitle":"AI localization for marketing and web content","metaDescription":"AI translates and adapts campaigns, product pages and web content for each market. Google Cloud reports Swarovski localizes campaigns 10x faster with AI.","definition":"AI that translates and adapts an organization's commercial content, such as campaigns, emails, product pages, help content and websites, for each market and language, using the brand's glossary, style guide and past approved translations, and routes the output to human linguists and local marketers for review in proportion to how visible and risky the content is.","aliases":["AI translation for marketing","AI transcreation","machine translation with post editing","website localization AI","multilingual content generation","LLM localization"],"industries":["cross-industry","retail-and-ecommerce","professional-services","manufacturing"],"functions":["marketing"],"patterns":["translation","content-generation"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"A brand that sells in twenty markets needs every campaign, product page, email and help article in\ntwenty languages, adapted to local tone, regulation and culture, and it needs them at the same time\nas the source market. Traditional localization runs through agencies and translation vendors, and\nevery language adds review and delivery time. The risk is that smaller markets get content late,\nget less of it, or get a literal translation that reads badly.\n\nClassic machine translation solved part of this for high volume, low visibility text, but it did\nnot follow brand voice, terminology or local marketing conventions well enough for campaigns.\nLarge language models can take a glossary, a style guide and examples in the prompt, adapt rather\nthan translate, and check their own output, which moves the human effort from translating to\nreviewing.","problemStats":[],"howItWorks":"1. **Prepare the language assets.** Glossaries, do not translate lists, style guides per market\n   and a memory of past approved translations are loaded as reference material.\n2. **Classify the content.** Each item is tiered by visibility and risk: a legal notice or a hero\n   campaign line gets full human review, a long tail product description or help article may get\n   sampling only.\n3. **Translate and adapt.** The AI produces the target version with the terminology and tone\n   rules applied, adapts idioms, units, currencies and cultural references, and flags passages\n   it is unsure about.\n4. **Check quality.** Automated checks catch terminology violations, missing placeholders, length\n   limits and numbers that changed; a quality estimate decides which segments go to a linguist.\n5. **Review and learn.** Linguists and local marketers edit where needed, and their approved\n   versions flow back into the translation memory and glossary for the next job.","valueDrivers":["speed","cost-to-serve","inclusion-and-access","revenue-growth"],"kpis":["processing-time-reduction","productivity-gain","users-served","cost-reduction"],"indicativeValue":{"referenceOrg":"A consumer brand that localizes 3 million words of commercial content a year","inputs":[{"key":"words","label":"Translated words per year, all target languages","low":3000000,"high":3000000,"unit":"words per year","note":"The reference brand, for example 300,000 source words into 10 languages."},{"key":"costPerWord","label":"Current cost per translated word","low":0.1,"high":0.2,"unit":"USD per word","note":"Editorial assumption for professional human translation with review. Replace with your own vendor rates."},{"key":"savingShare","label":"Share of localization cost saved","low":0.15,"high":0.3,"unit":"fraction of cost","note":"Editorial assumption, replace with your own. No evidence record on this page reports a cost figure (Bosch Digital mentions saving time and costs without a number), and human review remains for visible and regulated content, so the range is kept conservative."}],"formula":"words * costPerWord * savingShare","currency":"USD","period":"per year","resultLabel":"Localization cost avoided","caveat":"Direct translation cost only. It leaves out the value of launching in all markets at the same time, the cost of the models and tooling, and the internal review time that remains."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"The models are good enough for most commercial content. The work is building the glossary and style guides, tiering content by risk, and connecting the content management systems so text does not travel by spreadsheet.","dataPrerequisites":["A glossary with approved and forbidden terms per language","Style guides per market, including tone and formality","A translation memory of past approved translations","A content inventory tiered by visibility, legal weight and risk"],"integrations":["Content management system and ecommerce platform","Translation management system or translation memory","Marketing automation and email platform","Digital asset management for images with text"]},"implementation":{"steps":[{"title":"Tier the content","detail":"Sort content into tiers: legal and regulated text, high visibility campaign copy, product and help content, and internal or user generated text. Decide the review level per tier before choosing any tool."},{"title":"Build the language assets","detail":"Consolidate glossaries, style guides and translation memories per market, and have local marketers approve them. The AI's output will only be as consistent as these assets."},{"title":"Benchmark against your own translations","detail":"Take a sample of content your linguists already approved, translate it with the candidate setup, and have reviewers score both blind before deciding where to use AI."},{"title":"Automate the checks, not only the translation","detail":"Add automatic checks for terminology, numbers, placeholders, length limits and brand names, and use quality estimation to send only uncertain segments to a linguist."},{"title":"Feed corrections back","detail":"Store every approved correction in the translation memory and glossary, and track which languages and content types still need heavy editing."}],"guardrails":["Human review for legal, regulated, safety and price related content in every language","Glossary and do not translate lists enforced automatically on every output","Numbers, prices, dates and product specifications checked against the source","No publication of machine output in a tier that requires review until a reviewer signs off"],"humanInTheLoop":"Local marketers and professional linguists own the glossaries and style guides, review content in the higher tiers, and sample the lower tiers. Legal or compliance reviewers approve regulated text in each market, as they would for a human translation.","kpisToInstrument":["Turnaround time from source approval to publication per language","Edit distance or share of segments changed by reviewers, per language and content type","Cost per word or per asset, including review","Terminology and number errors found after publication"],"failureModes":[{"title":"Fluent but wrong","detail":"The translation reads well but changes a number, a claim or a legal meaning. Check numbers and claims automatically and keep humans on regulated text."},{"title":"Brand voice drift","detail":"Each market's output slowly diverges from the brand's tone. Keep style guides current and sample output against them."},{"title":"Scaled thin pages","detail":"Automatically translated pages published in bulk can be treated as low value by search engines. Localize what people in that market need, and review it."},{"title":"Language law breaches","detail":"Some jurisdictions require certain content in the local language with specific quality or precedence. Check local language requirements per market."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Translating and adapting commercial content is not an Annex III use and makes no decisions about people. When an organization uses a third party translation or generation tool, the use is minimal risk for the organization: the Article 50(2) duty to mark generated text in a machine readable way falls on the provider of that system, and beyond AI literacy no specific deployer obligations apply. When an organization builds and operates its own generating system and puts it into service under its own name, it is the provider and must mark the output, unless the exception for systems that only assist standard editing or do not substantially alter the input or its semantics applies. A faithful translation may fall within that exception; transcreation that rewrites the message for a market alters the semantics and is less likely to. Article 50(4) covers deepfakes and text published to inform the public on matters of public interest, not marketing translations. Consumer protection and advertising rules apply to the translated text as to the original."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"Spam policies for Google web search, scaled content abuse","issuer":"Google Search Central","region":"global","url":"https://developers.google.com/search/docs/essentials/spam-policies","note":"Lists scraping feeds, search results or other content to generate many pages, including through automated transformations such as translating, as an example of scaled content abuse when little value is provided to users."},{"title":"Regulation (EU) 2024/1689 (AI Act), Article 50 on transparency obligations","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Sets out who must mark or disclose AI generated content, including the provider duty to mark generated text and its exception for systems that do not substantially alter the input or its semantics."}],"controls":["A tiering policy that sets the review level for each content type and market","Versioned glossaries and style guides with a named owner per market","Records of which content was machine translated, which model and who reviewed it","Personal data removed or masked before customer content is sent for translation"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered on a schedule or from the content system\nthrough an API token: an **AI agent** translates and adapts each item using a\n**knowledge base** that holds the glossary, style guides and past approved translations\n(retrieved with **hybrid retrieval**), and returns the result with **structured output** per\nsegment. **Custom functions** read from and write back to the content management or ecommerce\nsystem through REST.\n\n**Human in the loop** confirmation makes a linguist or local marketer approve or reject the write\nback of higher tier content before it is published,\n**guardrails** check terminology and block unapproved claims, and **PII masking** at the gateway\nremoves personal data before text reaches a model. **Test suites** with LLM based grading score\noutput against approved reference translations per language, including Arabic with regional\nmodels, and the platform is model agnostic, so each language can use the model that scores best."},"faq":[{"question":"Can AI replace translators for marketing content?","answer":"Not for the content that carries the brand or legal weight, which people should still review. Lionbridge, a localization provider, uses AI to flag sensitive content for human review before delivery and reports up to 30% shorter turnaround times. Google Cloud reports that Swarovski's campaign localization became 10 times faster with AI assisted translation and asset adaptation."},{"question":"What is the difference between machine translation and AI localization?","answer":"Classic machine translation converts sentences one by one. Localization with large language models can also apply a glossary and style guide, adapt tone, idioms and units for the market, and flag its own uncertain passages for a reviewer."},{"question":"Does machine translated content hurt search rankings?","answer":"Google's spam policies give scraping content to generate many pages, including through automated translation, as an example of scaled content abuse when the pages provide little value to users. Translation as such is not the target: localized pages that people in the market actually need, reviewed for quality, fall outside that example."}],"related":["public-service-translation","marketing-content-compliance-copilot","product-content-and-catalog-enrichment"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with evidence from Swarovski, Lionbridge and Bosch Digital checked against the sources. Editor pass attributed the Swarovski figure to Google Cloud, limited the human review claim to Lionbridge, and set autonomy to supervised agent."},{"date":"2026-09-27","note":"Second editor pass set the EU AI Act tier to context dependent (own generating system makes the organization a provider), labelled the cost saving range as an editorial assumption with no cost figure in the evidence, dated the Bosch Digital entry to the Google Cloud list of 19 December 2024 with an archived copy, and aligned the Google spam policy wording with its scraping qualifier."},{"date":"2026-09-27","note":"Review fix: confirmed the Bosch Digital dating against the Wayback captures once the archive was reachable again (absent 2024-12-05, present 2025-01-04, dated 19 December 2024), so the year and archived copy stand unchanged."}],"slug":"marketing-and-product-content-localization","url":"https://www.blits.ai/ai-use-cases/marketing-and-product-content-localization","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":750,"min":500,"max":1000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"swarovski-genie-campaign-localization","pooled":true},{"id":"lionbridge-azure-openai-localization","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10,"min":10,"max":10,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"swarovski-genie-campaign-localization","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"lionbridge-azure-openai-localization","pooled":false}]}],"indicativeValueResult":{"low":45000,"high":180000},"evidence":["bosch-gemini-marketing-localization","lionbridge-azure-openai-localization","swarovski-genie-campaign-localization"]},{"title":"AI marketing personalization at scale","shortTitle":"Marketing personalization at scale","seoTitle":"AI for personalized marketing campaigns at scale","metaDescription":"AI picks the next best offer for each customer and writes copy within approved claims. Google Cloud reports Radisson grew AI campaign revenue by over 20%.","definition":"AI that runs marketing campaigns at the level of the individual: it decides for each customer which product, offer, message or content to show next across email, app, web and paid media, and generates the matching copy and creative variants within brand and compliance rules. It is the marketing team's engine across many campaigns and channels, not an agent that converses with the customer.","aliases":["hyper personalization","one to one marketing","AI personalized campaigns","personalized recommendations and offers","AI decisioning for marketing"],"industries":["cross-industry","travel-and-hospitality","media-and-entertainment","retail-and-ecommerce","banking"],"functions":["marketing","sales"],"patterns":["recommendation-and-personalization","prediction-and-scoring","content-generation"],"channels":["email","mobile-app","web-chat","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"Marketing teams know that relevant messages work better than broadcast ones, but personalization\nhas been limited by two bottlenecks. The first is decisioning: choosing, for millions of customers,\nwhich of hundreds of offers and messages is most relevant right now, while respecting frequency\ncaps, consent and eligibility. The second is content: even a good decision engine was only as\npersonal as the handful of creative variants the studio could produce and legal could approve.\n\nMachine learning has handled the first bottleneck for years, in recommendation engines and next best\naction systems. Generative AI now attacks the second: copy, images and video variants per segment,\nlanguage and context, without a separate studio brief for each one. Together they make segment of\none campaigns practical. The risks grow with it: invented claims in generated copy, offers that\nexploit vulnerable customers, profiling without a lawful basis, and brand damage from content nobody\nreviewed. The workable approach is a library of approved building blocks within which AI\npersonalizes.","problemStats":[],"howItWorks":"1. **Unify the signals.** Customer profile, consent, product holdings, behaviour on web and app, and\n   context (location, time, channel) are brought into one decisioning layer.\n2. **Decide the next best action.** Models predict propensity and value for each eligible action,\n   including service messages and doing nothing, and arbitration rules pick one within caps,\n   eligibility and business priorities.\n3. **Assemble the content.** The chosen action is rendered from approved building blocks; generative\n   AI produces copy and creative variants within brand, claims and disclosure rules.\n4. **Check before sending.** Automated checks cover required disclosures, prohibited claims,\n   consent and vulnerability flags; new templates and campaigns get human approval.\n5. **Deliver and learn.** The message goes out in the right channel; responses feed back into the\n   models, and experiments with control groups measure the real uplift.","valueDrivers":["revenue-growth","customer-experience","employee-productivity","speed"],"kpis":["conversion-rate-uplift","revenue-uplift","processing-time-reduction","productivity-gain","interactions-handled"],"indicativeValue":{"referenceOrg":"A consumer business with USD 200 million a year in revenue from marketing driven campaigns","inputs":[{"key":"campaignRevenue","label":"Annual revenue attributed to targeted campaigns","low":200000000,"high":200000000,"unit":"USD per year","note":"The reference organization. Use revenue measured against a holdout, not last click attribution."},{"key":"uplift","label":"Incremental revenue from personalization versus current targeting","low":0.02,"high":0.06,"unit":"fraction of campaign revenue","note":"Editorial assumption, kept well below the benchmarks on this page (Google Cloud reports revenue from Radisson Hotel Group's AI powered campaigns up by more than 20% and Catchtable's reservation conversion up 30%), because those figures are vendor reported and do not state their baseline or whether a control group was used."},{"key":"margin","label":"Contribution margin on incremental revenue","low":0.2,"high":0.4,"unit":"fraction of revenue","note":"Editorial assumption, replace with your own."}],"formula":"campaignRevenue * uplift * margin","currency":"USD","period":"per year","resultLabel":"Incremental contribution from personalization","caveat":"An uplift estimate that only holds if measured with control groups. It leaves out content production savings, the cost of data, platform and people, and any revenue lost to customers who opt out after poorly judged personalization."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Generating variants is easy; everything around it is not. Unified customer data with consent, a decisioning layer with arbitration, a library of approved content blocks and experiment discipline decide whether personalization pays off.","dataPrerequisites":["Customer profile and behavioural data with recorded marketing consent per channel","Product and offer catalogue with eligibility rules","Approved brand guidelines, claims library and required disclosures","Response history and holdout groups to measure uplift"],"integrations":["Customer data platform or data warehouse","Decisioning or next best action engine","Marketing automation, email and push platforms","Content management and digital asset management","Consent and preference management"]},"implementation":{"steps":[{"title":"Fix consent and data first","detail":"Know for each customer and channel whether marketing and profiling are allowed, and make the decisioning layer enforce it. Personalization on data you may not use is a liability."},{"title":"Build an approved content library","detail":"Break campaigns into building blocks (claims, offers, disclosures, images) that legal and brand have approved, and let generative AI personalize wording and combinations within them."},{"title":"Start with one journey and a holdout","detail":"Pick one high volume journey, such as onboarding, cart abandonment or renewal, and measure uplift against a randomized control group before scaling."},{"title":"Add arbitration across campaigns","detail":"Move from campaign by campaign targeting to one decision per customer that weighs all eligible actions, including service messages and no message, with frequency caps."},{"title":"Automate checks, keep approvals for the new","detail":"Run automatic checks for disclosures, prohibited claims and vulnerability flags on every variant, and require human approval for new templates, offers and audiences."},{"title":"Review fairness and customer outcomes","detail":"Check who receives which offers and prices, and whether customers in vulnerable circumstances are targeted with products that may harm them."}],"guardrails":["Personalization only on data with a lawful basis and recorded consent for the channel","Generated copy restricted to approved claims and must include required disclosures","No targeting that exploits age, disability or financial difficulty; vulnerable customers excluded from high risk offers","Frequency caps and an easy opt out in every message","Synthetic images and video labelled or marked where the law requires it"],"humanInTheLoop":"Marketing owns the strategy, audiences and offers; brand and legal approve templates, claims and new campaigns; analysts own the experiments. The AI selects and assembles within those approvals, and people review samples of what customers actually received.","kpisToInstrument":["Incremental conversion and revenue versus a randomized holdout","Opt out, unsubscribe and complaint rates per campaign","Share of generated variants passing automated compliance checks first time","Campaign production time from brief to launch","Distribution of offers across customer groups, including vulnerable customers"],"failureModes":[{"title":"Uplift that is really attribution","detail":"Personalized campaigns take credit for sales that would have happened anyway. Measure against randomized holdouts."},{"title":"Invented claims at scale","detail":"Generated copy promises a benefit or price that does not exist. Restrict generation to an approved claims library and check every variant."},{"title":"Creepy or harmful targeting","detail":"Messages reveal inferences customers did not expect, or push credit or gambling to people in difficulty. Limit sensitive inferences and exclude vulnerable customers."},{"title":"Too many messages","detail":"Each campaign optimizes for itself and customers are flooded. Arbitrate across campaigns with frequency caps."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Most personalization and content generation is minimal risk. Providers of systems that generate synthetic audio, image, video or text content must mark the output as artificially generated, and deployers must disclose deep fakes (Article 50(2) and 50(4)). Personalization that deploys manipulative or deceptive techniques, or exploits vulnerabilities due to age, disability or a specific social or economic situation, in a way that causes or is reasonably likely to cause significant harm, is prohibited under Article 5(1)(a) and (b). Using AI to assess creditworthiness or to price life and health insurance is high risk under Annex III point 5(b) and 5(c) and belongs on its own page. Outside the AI Act, the FCA Consumer Duty applies only to FCA regulated firms (the financial services slice of this use case), and the Telephone Consumer Protection Act applies only to campaigns delivered by call or text message in the US."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","us-tcpa"],"guidance":[{"title":"Article 5, prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Points 1(a) and 1(b) prohibit manipulative techniques and the exploitation of vulnerabilities that cause significant harm."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Marking of synthetic content and disclosure of deep fakes used in campaigns."},{"title":"Article 21 GDPR, right to object","issuer":"European Union","region":"europe","url":"https://gdpr-info.eu/art-21-gdpr/","note":"People may object at any time to processing for direct marketing, including profiling related to it."},{"title":"Direct marketing and privacy and electronic communications","issuer":"UK Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/","note":"The UK regulator's guidance hub on direct marketing under PECR and data protection law, covering consent for email and text marketing and choosing a lawful basis."}],"controls":["Consent and preference checks enforced in the decisioning layer","Approved claims and disclosure library with owners","Automated pre send checks plus human approval of new templates and audiences","Holdout based measurement for every personalized journey","Periodic review of offer distribution for vulnerable customers and fairness"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the decisioning usually stays in the organization's own next best action engine or\ncustomer data platform, which **custom functions** call through REST (the integration catalog\nincludes Salesforce, Snowflake and Databricks). Blits.ai adds the conversational and content side:\nan **agentic workflow** in which an **AI agent** writes copy variants from a **knowledge base** of\napproved claims and brand guidelines, and **guardrails** with admin authored policies check each\nvariant for prohibited claims and missing disclosures before **human in the loop** approval.\n\nWhen a customer responds or asks about an offer, a customer facing agent presents it on **web chat, WhatsApp, SMS or email**,\nor inside the organization's own app through the **REST or WebSocket API channel**, with rich\n**product recommendation cards**; it explains the offer, answers questions and uses **human\nhandover** when needed. **Multi language** bots serve each customer in their language,\n**analytics** and response feedback show how customers react, and **test suites** check the agent's\noutput against your rules before every release. The platform is model agnostic and can run in the EU or\nUAE region."},"faq":[{"question":"What results do companies report from AI personalization?","answer":"Vendor reported results are large. Google Cloud reports that revenue from Radisson Hotel Group's AI powered campaigns rose by more than 20% and ad team productivity by around 50%, that Catchtable's personalized recommendations raised reservation conversion by 30%, and that Virgin Voyages cut campaign creation time by 40%. None of these states its baseline, so treat them as upper bounds and measure against your own holdout."},{"question":"How is this different from an offers and rewards agent in banking?","answer":"The banking offers page covers choosing and explaining offers and rewards from transaction data inside the bank's app and conversations. This page covers the wider marketing job in any industry: deciding and generating personalized campaign content at scale across email, app, web and paid media."},{"question":"Can generative AI write campaign copy without legal review?","answer":"Not freely. The workable model is an approved library of claims, offers and disclosures that legal and brand sign off once, with AI personalizing wording and combinations inside it, automated checks on every variant and human approval for anything new."}],"related":["offers-and-rewards-agent","marketing-content-compliance-copilot","churn-prediction-and-retention-offers","customer-feedback-analysis","conversational-shopping-assistant","proactive-outbound-engagement-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical with five evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added Telephone Consumer Protection Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; attributed the benchmark figures in the value note and FAQ to Google Cloud and removed an unsupported baseline claim; corrected the EU AI Act basis (Article 50 covers text too, Annex III point 5(b) and 5(c)); added UK GDPR; made the ICO guidance note accurate; limited the Blits.ai channels to those in the feature inventory; fixed source titles and removed unsourced languages in the evidence."},{"date":"2026-09-27","note":"Review fixes: sourced Catchtable's country and app channel to its own Google Play listing and Square Enix's country to its company page; removed unsourced channels from the Catchtable and Radisson evidence; moved editorial text out of the Square Enix summary; the meta description now cites Radisson, a closer fit than Catchtable; clarified the scope of the Consumer Duty and TCPA; test suites now run before every release, not on every change."},{"date":"2026-09-27","note":"Second fact check against sources: all quotes, figures, legal references and guidance links confirmed; removed an unsourced API channel from the Virgin Voyages evidence and sourced its country to the company's own website terms."}],"slug":"personalized-marketing-at-scale","url":"https://www.blits.ai/ai-use-cases/personalized-marketing-at-scale","benchmarks":[{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":30,"min":30,"max":30,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"catchtable-personalized-recommendations","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10,"min":10,"max":10,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"swarovski-genie-campaign-localization","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40,"min":40,"max":40,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"virgin-voyages-personalized-campaigns","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"radisson-hotel-group-personalized-advertising","pooled":true}]},{"kpi":"revenue-uplift","label":"Revenue uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":20,"min":20,"max":20,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"radisson-hotel-group-personalized-advertising","pooled":true}]}],"indicativeValueResult":{"low":800000,"high":4800000},"evidence":["amazon-generative-ai-listing-tools","catchtable-personalized-recommendations","commonwealth-bank-customer-engagement-engine","radisson-hotel-group-personalized-advertising","square-enix-personalized-emails","swarovski-genie-campaign-localization","virgin-voyages-personalized-campaigns"]},{"title":"AI medical coding for clinical encounters","shortTitle":"Medical coding automation","seoTitle":"Autonomous medical coding with AI","metaDescription":"AI assigns diagnosis and procedure codes from clinical notes and routes doubtful cases to coders. Your Health reports it codes 95.5% of encounters automatically.","definition":"AI that reads the clinical documentation of an encounter and assigns the diagnosis and procedure codes (such as ICD-10, CPT and HCPCS) needed for billing and reporting, either as suggestions for a certified coder or autonomously for encounters it can code with high confidence, sending the rest to coders with the reasons.","aliases":["autonomous coding","computer assisted coding","AI medical coder","automated charge capture from clinical notes"],"industries":["healthcare"],"functions":["finance-and-accounting","operations"],"patterns":["classification-and-routing","document-processing"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Every patient encounter has to be translated into standard codes before a provider can bill for\nit. Certified coders read the notes and choose among tens of thousands of diagnosis and procedure\ncodes, modifiers and add ons, following payer rules that change every year. Coders are scarce,\nbacklogs delay cash, and inconsistent coding causes claim denials, rework and lost revenue.\n\nCoding also carries compliance risk in both directions. Undercoding loses revenue that the\ndocumentation supports; overcoding, especially of diagnoses that raise risk adjustment payments,\nleads to audits, repayments and fraud allegations. Computer assisted coding has suggested codes for\nyears; newer systems code the simpler encounters on their own, which raises the question of who\nchecks their work.","problemStats":[],"howItWorks":"1. **Receive the signed documentation.** When an encounter is closed, the notes, orders and results\n   are sent to the coding engine.\n2. **Assign codes.** The model proposes diagnosis and procedure codes, modifiers and sequencing, with\n   the text that supports each code and a confidence level.\n3. **Apply rules.** Payer edits, coding guidelines and the organization's own policies are checked,\n   and documentation gaps are flagged for the clinician.\n4. **Route by confidence.** High confidence encounters of approved types are released to billing\n   automatically; the rest go to a coder's worklist with the suggested codes.\n5. **Audit and learn.** Coders and auditors sample automated encounters, and denials and audit\n   findings feed back into rules and models.","valueDrivers":["cost-to-serve","speed","compliance","revenue-growth","employee-productivity"],"kpis":["automation-rate","accuracy","error-reduction","productivity-gain","processing-time-reduction","cost-reduction"],"indicativeValue":{"referenceOrg":"A physician group with 1 million coded encounters a year","inputs":[{"key":"encounters","label":"Encounters coded per year","low":1000000,"high":1000000,"unit":"encounters per year","note":"The reference physician group."},{"key":"automationShare","label":"Share of encounters coded without a coder","low":0.5,"high":0.85,"unit":"fraction of encounters","note":"Conservative against the 95.5% encounter level automation that Your Health reports on this page. The lower range is an editorial caution, not a sourced figure: one self reported result is thin evidence, and it does not say how \"automated\" is defined or measured."},{"key":"minutesPerEncounter","label":"Coder minutes per encounter today","low":2,"high":5,"unit":"minutes per encounter","note":"Editorial assumption for professional fee coding. Replace with your own productivity data."},{"key":"costPerHour","label":"Fully loaded cost per coder hour","low":35,"high":60,"unit":"USD per hour","note":"Editorial assumption covering internal and outsourced coders. Replace with your own."}],"formula":"encounters * automationShare * minutesPerEncounter / 60 * costPerHour","currency":"USD","period":"per year","resultLabel":"Coder effort released","caveat":"Coder effort only. It leaves out the effect on denials, days to bill and cash, the revenue effect of more complete coding (which must be supported by documentation), audit and compliance costs, and the licence cost, often charged per encounter."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Mature vendors exist and integrate with common record systems. The work is in choosing which encounter types may be released without a coder, calibrating to payer mix and local guidelines, and building the audit loop that compliance teams need.","dataPrerequisites":["Historical encounters with documentation and final codes for calibration and testing","Current code sets, payer rules and the organization's coding policies","Denial and audit history by code and specialty"],"integrations":["Electronic health record for signed documentation","Practice management or billing system for charges and claims","Coder worklist and query tools for clinician documentation questions","Denial management and audit systems"]},"implementation":{"steps":[{"title":"Start with suggestions, then release by encounter type","detail":"Run the engine as a suggestion tool first, measure agreement with coders per encounter type, and allow automatic release only where agreement stays high."},{"title":"Set confidence thresholds with compliance","detail":"Agree with compliance which encounter types, specialties and codes may be released automatically and which, such as risk adjustment diagnoses, always need a coder."},{"title":"Close the documentation loop","detail":"Route documentation gaps back to clinicians as queries instead of coding around them, so codes stay supported by the record."},{"title":"Audit automated encounters continuously","detail":"Sample automated encounters every week against a coder's review, track accuracy by code family and feed errors back."},{"title":"Watch denials and payer feedback","detail":"Compare denial rates and reasons before and after, per payer, and investigate any rise in high value codes."}],"guardrails":["Only encounter types with proven accuracy are released without a coder","Every automated code is supported by text in the signed documentation, stored with the claim","Risk adjustment diagnoses and high value procedures reviewed by a certified coder","Documentation gaps sent to the clinician as a query, never filled in by the AI","Weekly audit sample of automated encounters with results reported to compliance"],"humanInTheLoop":"Certified coders handle every encounter below the confidence threshold and all excluded code families, and they audit a sample of automated encounters. Compliance owns the release rules, and clinicians answer documentation queries.","kpisToInstrument":["Share of encounters released without a coder, by specialty","Coding accuracy on a weekly audit sample","Denials due to coding, per payer","Days from encounter to claim","Shift in the distribution of evaluation and management levels and risk adjustment scores"],"failureModes":[{"title":"Upcoding at scale","detail":"The model systematically selects higher levels or adds unsupported diagnoses, which creates overpayment and fraud exposure. Monitor code distributions and audit high value codes."},{"title":"Accuracy measured on the wrong sample","detail":"Vendor and customer accuracy figures may cover only the encounters that were automated, or a sample chosen by the vendor; the Your Health figures on this page do not say how accuracy was measured. Audit a random sample of automated encounters yourself."},{"title":"Stale rules","detail":"Annual code set and payer rule changes are not reflected in time. Assign owners for updates and test before each effective date."},{"title":"Coders lose the skill to audit","detail":"If coders only handle exceptions, the organization may lose the expertise to check the machine. Keep audit and training time in coder roles."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Assigning billing and statistical codes from clinical documentation is not listed in Annex III and does not decide on a person's access to care, so no specific AI Act obligations apply beyond AI literacy. Health data processing falls under GDPR Article 9, and in the United States under HIPAA and the payment integrity rules of public payers. Minimal under the AI Act does not mean low stakes: the Veterans Health Administration classifies its computer assisted coding deployment on this page as high impact in the 2025 US federal AI use case inventory, even though coders select every code."},"regulations":["eu-ai-act","gdpr","hipaa","nist-ai-rmf"],"guidance":[{"title":"General Compliance Program Guidance","issuer":"Office of Inspector General, US Department of Health and Human Services","region":"north-america","url":"https://oig.hhs.gov/compliance/general-compliance-program-guidance/","note":"Voluntary OIG guidance on the federal fraud and abuse laws and the seven elements of a compliance program, the frame against which US providers audit and monitor billing and coding, whether done by people or software."},{"title":"Medicare Advantage Risk Adjustment Data Validation Program","issuer":"Centers for Medicare & Medicaid Services","region":"north-america","url":"https://www.cms.gov/data-research/monitoring-programs/medicare-risk-adjustment-data-validation-program","note":"Explains how CMS audits whether diagnoses that Medicare Advantage organizations submit for risk adjustment are supported by medical records. Providers are not the audited party, but the plans they code for are, so unsupported diagnoses from automated coding flow into this exposure."},{"title":"ICD-10","issuer":"Centers for Medicare & Medicaid Services","region":"north-america","url":"https://www.cms.gov/medicare/coding-billing/icd-10-codes","note":"CMS page for the ICD-10 code sets. It publishes the annual ICD-10-PCS procedure code files and relays the ICD-10-CM diagnosis code updates that CDC develops and announces; an automated coding engine must track both."}],"controls":["Written release rules per encounter type and code family, approved by compliance","Evidence link from every code to the supporting documentation, kept with the claim","Weekly audit of automated encounters and trend monitoring of code distributions","Change control for model, rule and code set updates","Access controls and logging for clinical documentation used by the engine"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai each closed encounter triggers an **agentic workflow** through the **REST API**. An\n**AI agent** with **structured output** proposes codes with the supporting text and a confidence\nlevel, using a **knowledge base** that holds the coding guidelines and the organization's policies,\nretrieved with hybrid search. **Custom functions** read the documentation from the record system,\napply payer rules and write the result to the billing system.\n\nEvery encounter below the agreed threshold, and all excluded code families, is handed to a coder\nfor approval through **human in the loop approval**. The **audit trail** records each run, **test suites** compare the\nagent's codes with coded reference encounters before any change, **monitors** alert on failures,\nand **PII masking** limits the patient data in prompts. The platform is model agnostic."},"faq":[{"question":"What share of encounters can AI code without a coder?","answer":"Published figures are few and define automation differently. Your Health reports that it codes 95.5% of encounters automatically with Fathom across all service lines, with accuracy rising from 96.3% to 98.3%. A Mass General Brigham executive cited \"a 70% reduction in manual labor\" with CodaMetrix, without stating scope or period. Our editorial advice is to treat these self reported figures as upper bounds and measure on your own audit sample."},{"question":"Does automated coding raise compliance risk?","answer":"It can, because errors repeat at scale. The main exposure is unsupported diagnoses or higher service levels that increase payments, which public payers audit. Keep evidence for every code, audit automated encounters and route risk adjustment diagnoses to coders."},{"question":"Is computer assisted coding the same as autonomous coding?","answer":"No. Computer assisted coding suggests codes that a coder confirms, as in the Veterans Health Administration's deployment; autonomous coding releases confident encounters to billing without a coder and sends the rest to a worklist."}],"related":["ambient-clinical-documentation","health-prior-authorization-and-claims-adjudication"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, with evidence from Your Health, Mass General Brigham and the Veterans Health Administration, verified against the sources. Editor pass recorded the Mass General Brigham figure as \"manual labor\" as quoted, added the VHA high impact classification, softened unsourced advice and moved the Your Health figures to Your Health's own newsroom post (grade B)."}],"slug":"medical-coding-automation","url":"https://www.blits.ai/ai-use-cases/medical-coding-automation","benchmarks":[{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":98.3,"min":98.3,"max":98.3,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"your-health-autonomous-medical-coding","pooled":true}]},{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":95.5,"min":95.5,"max":95.5,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"your-health-autonomous-medical-coding","pooled":true}]},{"kpi":"error-reduction","label":"Error reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":59,"min":59,"max":59,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"mass-general-brigham-codametrix-coding-automation","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":70,"min":70,"max":70,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"mass-general-brigham-codametrix-coding-automation","pooled":true}]}],"indicativeValueResult":{"low":583333.3333333334,"high":4250000},"evidence":["mass-general-brigham-codametrix-coding-automation","veterans-health-administration-computer-assisted-coding","your-health-autonomous-medical-coding"]},{"title":"AI meeting notes and CRM update for wealth advisors","shortTitle":"Advisor meeting notes","seoTitle":"AI meeting notes and CRM updates for advisors","metaDescription":"An AI notetaker drafts the file note, follow up and CRM record after each client meeting. Microsoft reports UniSuper advisers save about 30 minutes per interaction.","definition":"An AI notetaker for wealth advisors that turns a client advice meeting, recorded with the client's consent, into the file note, follow up message and CRM record the firm needs to evidence its advice; unlike a general meeting summarizer, its output becomes part of the regulated client record. It drafts a structured note with the client's goals, circumstances, decisions and action items, and writes it into the CRM once the advisor has approved it.","aliases":["AI notetaker for financial advisors","advice file note automation","meeting summary to CRM","client meeting transcription"],"industries":["wealth-and-asset-management","banking"],"functions":["sales","regulatory-compliance","operations"],"patterns":["summarization","speech-analytics","agentic-workflow","content-generation"],"channels":["microsoft-teams","voice","internal-tools","email"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"mainstream","segment":"front-office","problem":"After every client meeting an advisor has to write a file note, record what the client said about\ngoals, risk and circumstances, list what was agreed, send a follow up and update the CRM. With the\nnext meeting waiting, this work is easy to postpone, and a note written days later can miss what\nthe client actually said, which weakens the firm's record when a recommendation is later questioned.\n\nThe work is also expensive: it takes senior, client facing time, or an assistant who sat in the\nmeeting. And when notes are short, the CRM, which should hold the richest view of the client,\nholds little of the conversation, so the next meeting starts from memory.","problemStats":[],"howItWorks":"1. **Consent first.** The advisor tells the client the meeting will be transcribed by an AI\n   notetaker and records the client's consent; without it, nothing is recorded.\n2. **Capture.** The notetaker joins the video call or the phone line and produces a transcript\n   with speaker separation.\n3. **Draft the file note.** A summary is generated in the firm's template: attendees, topics,\n   client goals and circumstances mentioned, decisions, action items with owners, and anything the\n   client asked to be checked.\n4. **Advisor review.** The advisor corrects and approves the note and the draft follow up email.\n   Nothing is filed or sent without that approval.\n5. **Write back.** The approved note, tasks and key facts are written into the CRM and the\n   transcript is retained under the firm's record keeping policy.","valueDrivers":["employee-productivity","compliance","customer-experience"],"kpis":["time-saved-per-task","hours-saved","productivity-gain","employee-adoption","accuracy"],"indicativeValue":{"referenceOrg":"A wealth manager with 500 client facing advisors","inputs":[{"key":"advisors","label":"Client facing advisors","low":500,"high":500,"unit":"advisors","note":"The reference firm."},{"key":"meetingsPerWeek","label":"Client meetings per advisor per week","low":3,"high":6,"unit":"meetings per advisor per week","note":"Editorial assumption, replace with your own CRM activity data."},{"key":"minutesSaved","label":"Minutes of write up saved per meeting","low":20,"high":40,"unit":"minutes per meeting","note":"Brackets the one reported saving on this page (UniSuper, about 30 minutes per interaction) and sits below the 45 minutes per meeting a Quilter Cheviot investment manager assumes in the firm's estimate; the low end allows for review time. Replace with your own pilot data."},{"key":"weeks","label":"Working weeks per year","low":46,"high":46,"unit":"weeks per year","note":"Editorial assumption."},{"key":"hourlyCost","label":"Fully loaded advisor cost per hour","low":80,"high":150,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"advisors * meetingsPerWeek * minutesSaved / 60 * weeks * hourlyCost","currency":"USD","period":"per year","resultLabel":"Value of advisor time released from meeting write up","caveat":"Counts released time only. It leaves out licence and transcription costs, the time to review each note, and the harder to price benefit of a more complete record in complaints and audits."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Transcription and summarization are mature. The work is in consent capture, the note template, retention rules for recordings and transcripts, and a reliable write back into the CRM.","dataPrerequisites":["An agreed file note template per meeting type (review, new advice, service)","Consent wording and a place to record the client's consent","Retention and deletion rules for audio, transcripts and notes","CRM field mapping for notes, tasks and client facts"],"integrations":["Video meeting and telephony platforms","Speech to text with speaker separation","CRM such as Salesforce or Microsoft Dynamics 365","Email for the draft follow up","Records management or archive for retained transcripts"]},"implementation":{"steps":[{"title":"Agree the note standard with compliance","detail":"Define what a good file note contains for each meeting type and which statements must be captured verbatim (for example a client's stated risk appetite or a refusal of advice)."},{"title":"Design consent and retention","detail":"Write the consent script, record consent in the CRM, and decide how long audio and transcripts are kept and where. Some clients will decline; the process must work without AI."},{"title":"Pilot with a small group of advisors","detail":"Measure time to a finished note, edit rate and advisor satisfaction per meeting type, and collect notes that went wrong to improve the template and prompts."},{"title":"Automate the write back","detail":"Once notes are reliable, write the approved note, tasks and structured facts into the CRM through its API rather than copy and paste."},{"title":"Sample and supervise","detail":"Supervisors review a sample of notes against transcripts each month, with extra attention to meetings where advice was given."}],"guardrails":["No recording without recorded client consent, with an easy opt out","Advisor approval before any note is filed or any message is sent","Notes limited to what was said; no inferred emotions, health conditions or vulnerability labels without a human decision","PII masking and access control on transcripts, with retention limits","Every AI generated note labelled as such in the CRM"],"humanInTheLoop":"The advisor reviews, corrects and approves every note and follow up; supervisors sample notes against transcripts. Suspected vulnerability or complaints mentioned in a meeting go to a human process, not to an automated flag alone.","kpisToInstrument":["Median minutes from meeting end to approved note","Share of meetings with a complete note within 24 hours","Advisor edit rate per note section","Consent rate and opt outs","Supervisor sample findings per month"],"failureModes":[{"title":"Rubber stamped notes","detail":"Advisors approve drafts without reading them, so errors enter the record. Track time spent reviewing and sample notes against transcripts."},{"title":"Missing or misattributed statements","detail":"Speaker separation fails on a phone line and a client's words are attributed to the advisor. Test on real audio and flag low confidence passages."},{"title":"Consent gaps","detail":"Recording starts before consent is captured, or consent is not stored. Make consent a hard gate in the workflow."},{"title":"Over retention","detail":"Audio and transcripts are kept indefinitely by default. Apply retention rules from day one."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Transcribing and summarizing meetings for an employee is not a use listed in Annex III, and the advisor reviews every note before it is filed or sent. The tier would change if the tool inferred emotions: emotion recognition is high risk under Annex III point 1(c), and inferring the emotions of employees at work is prohibited under Article 5(1)(f). Both stay out of scope."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","dora","mas-ai-risk-management","iso-42001","mifid-ii"],"guidance":[{"title":"ESMA public statement on the use of AI in the provision of retail investment services","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf","note":"Expects investment firms to keep comprehensive records on their use of AI, and applies MiFID II conduct and record keeping duties to AI supported advice."},{"title":"Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT)","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2018/feat","note":"Transparency and accountability principles for data driven tools used in client relationships."}],"controls":["Consent capture and storage for every recorded meeting","Retention schedule for audio, transcripts and notes aligned with record keeping rules","Labelling of AI drafted notes and an audit trail of advisor edits","Monthly supervisory sampling of notes against transcripts","Inventory entry with an accountable owner and a documented note template"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the capture step uses **speech to text** with **self hosted transcription and speaker\ndiarization**, so audio can be processed on Blits.ai infrastructure, with EU and UAE data residency. An\n**agentic workflow** then drafts the file note in the firm's template with an **AI agent** using\n**structured output**, and pauses for **human in the loop** approval by the advisor before\nanything is filed or sent.\n\nApproved notes and tasks are written to the CRM through **custom functions** (REST calls) or the\nready made **Microsoft Dynamics 365** tool, and the follow up is drafted for **email**.\n**PII masking** at the gateway and **guardrails** control what reaches the model, the workflow's\n**audit trail** records each run and approval, and **test suites** check note quality against a\nset of reference transcripts on every prompt or model change."},"faq":[{"question":"How much time does an AI notetaker save an advisor?","answer":"The one reported saving is about half an hour: Microsoft reports UniSuper advisers save roughly 30 minutes per client interaction. Quilter's estimate of more than 13,000 hours a month assumes 45 minutes saved per meeting, which is a projection rather than a measured result. Bank of America says the capability can save advisors up to four hours per meeting, which it presents as potential rather than a measured result."},{"question":"Do we need client consent to record and transcribe meetings?","answer":"Morgan Stanley and Merrill both run their notetakers with client consent. Whether consent is legally required depends on the jurisdiction and the lawful basis you rely on, so agree it with legal and compliance, and make sure the process still works when a client says no."},{"question":"Can the AI note replace the advisor's own record?","answer":"No. The note is a draft; the advisor approves it and remains responsible for its accuracy. Supervisors should sample notes against transcripts, especially for meetings where advice was given."}],"related":["client-briefing-and-call-report-copilot","meeting-summarization-and-action-items","suitability-assessment-assistant","next-best-action-for-advisors","wealth-advisor-knowledge-assistant","call-quality-and-compliance-monitoring"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Meeting Capture and CRM) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added MiFID II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed an unsourced claim from the problem statement, cited the exact EU AI Act provisions, qualified the Bank of America four hours figure as potential, named Quilter Cheviot, added seoTitle and metaDescription, and added an archived copy for the Morgan Stanley source."},{"date":"2026-09-27","note":"Review fixes: the Quilter 45 minutes is now presented as the assumption behind an estimate and no longer recorded as a metric; removed an unsupported breakdown of the Bank of America four hours; the consent FAQ is now a design recommendation rather than a legal rule; Morgan Stanley Debrief stage set to production; softened unsourced statements in the problem; added UK GDPR."}],"slug":"client-meeting-notes-and-crm-update","url":"https://www.blits.ai/ai-use-cases/client-meeting-notes-and-crm-update","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":15,"min":15,"max":15,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"seb-wealth-advisor-agent-assist","pooled":true}]},{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":30,"min":30,"max":30,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"unisuper-copilot-advice-file-notes","pooled":true}]}],"indicativeValueResult":{"low":1840000,"high":13800000},"evidence":["bank-of-america-merrill-ai-meeting-journey","commerzbank-client-call-documentation","morgan-stanley-debrief-meeting-notes","quilter-copilot-meeting-notes","seb-wealth-advisor-agent-assist","unisuper-copilot-advice-file-notes"]},{"title":"AI meeting summarization and action items","shortTitle":"Meeting summaries and action items","seoTitle":"AI meeting notes, summaries and action items","metaDescription":"AI turns meeting transcripts into summaries and action items for people to check. UK probation staff summarised over 1.6 million meetings with Justice Transcribe.","definition":"AI that summarizes internal and operational meetings, such as team, project, board and case meetings: it transcribes an online or in person meeting with the participants' knowledge and produces a summary, decisions and action items with owners and dates for the organizer to check and share. It is the general purpose tool; client advice meetings and sales calls, which feed a regulated record or a sales pipeline, have their own pages.","aliases":["AI meeting notes","AI notetaker","meeting recap","automated minutes","meeting transcription and summary"],"industries":["cross-industry","government","technology","professional-services"],"functions":["knowledge-management","operations"],"patterns":["summarization","speech-analytics"],"channels":["microsoft-teams","internal-tools","mobile-app"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"mainstream","problem":"What a meeting decides is often lost. Someone takes notes while trying to participate, the notes are partial and late, action items live in people's\nheads, and colleagues who missed the meeting ask for a recap or watch a recording. In frontline\nroles such as probation, social work, casework and field inspection, the meeting is the work, and\nwriting it up afterwards takes time that could go to the next person.\n\nThe input is often already there: meeting platforms can produce a transcript. The value is not the\ntranscript but the structured record:\ndecisions, owners, dates and the points that matter for the case or project. The risks are\nspecific too. Summaries can be confidently wrong about who agreed to what, recording people raises\nconsent and privacy questions, and transcripts of sensitive meetings create records that must be\nprotected and retained correctly.","problemStats":[],"howItWorks":"1. **Tell everyone.** Participants are informed that the meeting is being transcribed, and anyone\n   can ask for it to stop; for sensitive meetings the organizer decides whether AI is used at all.\n2. **Transcribe with speakers.** Speech is transcribed live or from the recording, with speaker\n   attribution, in the language spoken.\n3. **Summarize in a fixed structure.** The AI produces a summary, decisions, action items with\n   owners and due dates, open questions and, for casework, the fields the record system requires.\n4. **Organizer reviews.** The organizer or caseworker corrects and approves the summary before it\n   is shared or saved to a record.\n5. **Push the actions.** Approved action items go to task tools or case systems; the summary is\n   stored with the meeting or case.\n6. **Keep what you must, delete what you can.** Transcripts and recordings follow the\n   organization's retention rules; often only the approved summary is kept.","valueDrivers":["employee-productivity","speed","compliance"],"kpis":["interactions-handled","time-saved-per-task","hours-saved"],"indicativeValue":{"referenceOrg":"An organization with 5,000 employees who use AI meeting summaries","inputs":[{"key":"users","label":"Employees using meeting summaries","low":5000,"high":5000,"unit":"employees","note":"The reference organization."},{"key":"meetingsPerWeek","label":"Summarized meetings per user per week","low":2,"high":4,"unit":"meetings per week","note":"Editorial assumption. Replace with usage data from your meeting platform."},{"key":"hoursSavedPerMeeting","label":"Note taking and write up time saved per meeting","low":0.1,"high":0.2,"unit":"hours per meeting","note":"Editorial assumption of 6 to 12 minutes. For Justice Transcribe, HM Prison and Probation Service advised a broad operational assumption of about 10 minutes per meeting; the Ministry of Justice calls the hours total derived from it an illustrative estimate only.","sourceUrl":"https://www.gov.uk/government/publications/justice-transcribe/justice-transcribe-data-7-october-2025-to-14-september-2026"},{"key":"weeks","label":"Working weeks per year","low":44,"high":46,"unit":"weeks per year","note":"Editorial assumption."},{"key":"hourlyCost","label":"Fully loaded employee cost","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"users * meetingsPerWeek * hoursSavedPerMeeting * weeks * hourlyCost","currency":"USD","period":"per year","resultLabel":"Employee time released from note taking","caveat":"Values time at cost and assumes the time is used productively, which trials often cannot confirm. It leaves out licence and platform cost, the time to review summaries, and the harder to measure value of better records and fewer missed actions."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Meeting platforms offer it out of the box. The work is in policy (when AI may be used, consent, retention), in structured outputs for casework, and in integration with case or task systems.","dataPrerequisites":["A policy on which meetings may be transcribed, how participants are told and how long records are kept","Summary templates per meeting type (project, casework, customer, board)","Access rules for transcripts and summaries, aligned with the meeting's confidentiality"],"integrations":["Meeting platforms (Microsoft Teams, Zoom, Google Meet) or a recording app for in person meetings","Task and project tools for action items","Case management or record systems for frontline meetings","Records management for retention and deletion"]},"implementation":{"steps":[{"title":"Write the policy before the rollout","detail":"Decide which meetings may be transcribed, how people are informed, who may switch it on, and which meetings are off limits (HR cases, legal privilege, some board discussions)."},{"title":"Define summary templates per meeting type","detail":"A project meeting needs decisions and actions; a probation or social work meeting needs the fields the case record requires. Structure beats free prose."},{"title":"Make review part of the flow","detail":"The organizer or caseworker approves before sharing or saving. Make corrections easy and record who approved."},{"title":"Connect actions and records","detail":"Send approved action items to task tools and approved summaries to the case or project record, instead of leaving them in the meeting chat."},{"title":"Train people on what it gets wrong","detail":"Show examples of wrong attributions, missed nuance and invented actions, and how to check for them, especially in meetings with several speakers or languages."},{"title":"Measure use and quality, not only licences","detail":"Track summaries approved, edit rates and user reported time saved, and audit a sample of summaries against recordings for accuracy."}],"guardrails":["Participants are informed before transcription starts and can object","Summaries are drafts until a named person approves them","Action items and decisions are attributed only to what was said, with low confidence items flagged","No sentiment or emotion scoring of participants, and no use of transcripts to evaluate individual employees","Transcripts and recordings follow retention rules and inherit the meeting's access restrictions"],"humanInTheLoop":"The organizer or caseworker reviews, corrects and approves every summary before it is shared or becomes part of a record, and remains accountable for its content. Records management owns retention; the data protection officer approves the policy for sensitive meeting types.","kpisToInstrument":["Meetings summarized and summaries approved per week","Edit rate on summaries and reported errors (wrong owner, invented action)","User reported time saved per meeting, validated by time studies on a sample","Share of action items completed by their due date","Transcripts deleted on schedule"],"failureModes":[{"title":"Confidently wrong attribution","detail":"The summary says a person agreed to something they did not. Require review before sharing and flag low confidence attributions."},{"title":"Transcripts nobody should have","detail":"Sensitive meetings are recorded and transcripts spread through shared channels. Define no AI meetings and inherit access restrictions."},{"title":"Licences without adoption","detail":"Tools are rolled out but few people use them after the first month. Measure active use per team and train on real meetings."},{"title":"Time saved that cannot be found","detail":"Self reported savings do not show in any outcome. Pair usage data with outcome measures such as case throughput or actions completed."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Transcribing and summarizing meetings for the participants is minimal risk. It becomes high risk under Annex III point 4(b) if transcripts are analysed to monitor or evaluate individual workers' performance or behaviour, and inferring participants' emotions from their voices or faces at work is prohibited by Article 5(1)(f). Recording and transcription also need a lawful basis and clear information to participants under GDPR."},"regulations":["eu-ai-act","gdpr","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Article 5, prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Point 1(f) prohibits AI that infers people's emotions in the workplace, except for medical or safety reasons, which rules out mood scoring of employees from their voices or faces in meetings."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) applies if meeting data is used to monitor or evaluate workers."},{"title":"Employment practices and data protection: monitoring workers","issuer":"UK Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/","note":"Relevant to recording and transcribing employees' meetings and to what the organization may do with the records."}],"controls":["Meeting transcription policy with a list of meeting types where AI is not used","Data protection impact assessment, including for meetings with customers or members of the public","Retention and deletion rules for recordings, transcripts and summaries","Access control that follows the meeting's confidentiality","Periodic accuracy audit of summaries against recordings"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the transcription can run on the **self hosted WhisperX service with speaker\ndiarization**, which gives word level timestamps, automatic language detection and \"who spoke\nwhen\" on Blits.ai infrastructure for data sovereignty, or on one of the other **speech to text**\nproviders. An **agentic workflow** passes the transcript to an **AI agent** with **structured\noutput** that fills the summary template for the meeting type, including decisions and action\nitems with owners.\n\n**Human in the loop** confirmation holds the next step until the organizer approves the summary,\nafter which **custom functions** push action items to task tools or write the approved summary to\nthe case system (the integration catalog includes Jira, Asana, ServiceNow and Salesforce).\n**PII masking** and **data retention controls** limit what is kept, and **test suites** with\nLLM based grading, run after each prompt change, compare summaries with reviewed transcripts. The platform\nis model agnostic and can run in the EU or UAE region."},"faq":[{"question":"How widely is AI meeting summarization used?","answer":"Few organizations publish usage figures for meeting summaries alone. The clearest comes from the UK Ministry of Justice: probation staff summarised more than 1.6 million meetings with Justice Transcribe between October 2025 and September 2026, a count of meetings where the tool was used, not a share of all probation meetings. In the UK government's Microsoft 365 Copilot trial, Teams had the highest Copilot adoption of any application, peaking at 71%, but that figure covers every Copilot feature in Teams, not meeting summaries alone."},{"question":"How much time does it save?","answer":"Published figures are mostly assumptions or self reported. The Ministry of Justice applies an assumption of about 10 minutes per meeting and calls the result illustrative, and UK trial participants estimated 26 minutes a day across all Copilot tasks. Measure it yourself on a sample before building a business case on it."},{"question":"Do we need consent to transcribe meetings?","answer":"Under GDPR you need a lawful basis and must inform participants; consent is one possible basis, not the only one, and your data protection officer decides which fits. Tell people before transcription starts, let them object, and do not use transcripts to evaluate employees without a separate assessment."},{"question":"How is this different from advisor meeting notes in wealth management?","answer":"The wealth version records advice to clients, with suitability and CRM obligations. This page covers the general case: internal, project, casework and frontline meetings, where the output is a checked record and a list of actions."}],"related":["client-meeting-notes-and-crm-update","sales-call-coaching-and-crm-update","client-briefing-and-call-report-copilot","enterprise-knowledge-search","live-agent-assist"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical as the general counterpart of the wealth advisor meeting notes page, with five evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed an unsourced claim about the share of time spent in meetings, tied the adoption point to the UK Copilot trial, attributed the 10 minute assumption to HM Prison and Probation Service with its source, narrowed the Article 5 note to emotion inference, corrected the integration examples, reworded the consent answer, corrected two evidence summaries and added an SEO title and meta description."},{"date":"2026-09-27","note":"Second fact check: removed an unsourced claim that meeting summarization is a common first use of generative AI, stated that the 71% Teams figure covers all Copilot features in Teams rather than meeting summaries, rewrote the usage answer and the meta description to match the Ministry of Justice source, and corrected three evidence details (Softcat description, trial participants, Department of Labor contract and authorization)."}],"slug":"meeting-summarization-and-action-items","url":"https://www.blits.ai/ai-use-cases/meeting-summarization-and-action-items","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1600000,"min":1600000,"max":1600000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ministry-of-justice-justice-transcribe","pooled":true}]}],"indicativeValueResult":{"low":1760000,"high":12880000},"evidence":["dol-note-taking-bot","ministry-of-justice-justice-transcribe","softcat-copilot-meeting-summaries","trace3-microsoft-copilot-recruiting-and-meetings","uk-government-m365-copilot-experiment"]},{"title":"AI metadata tagging and indexing for media archives","shortTitle":"Media archive metadata tagging","seoTitle":"AI metadata tagging for media archives","metaDescription":"AI tags video and audio archives with searchable metadata. The ABC analysed a million video records in two weeks; RTVE has contracted an AI archive tagging service.","definition":"AI that watches and listens to a broadcaster's or publisher's video and audio archive and generates rich, structured metadata, such as what is shown, who appears, spoken content, on screen text, logos and objects, so that content makers can find and reuse footage through natural language search instead of relying on the sparse, inconsistent tags an archive accumulated by hand over decades.","aliases":["AI video archive tagging","automated media indexing","AI metadata generation for broadcast archives","content archive AI cataloguing"],"industries":["media-and-entertainment"],"functions":["knowledge-management","operations"],"patterns":["computer-vision","speech-analytics","classification-and-routing"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"copilot","adoptionStage":"early-adopters","segment":"content operations","problem":"Broadcasters and publishers sit on archives that go back decades: newscasts, interviews,\ndocumentaries and raw, unpublished footage, described by whatever labelling convention was in use\nat the time it was catalogued, often no more than a title, a date and a one line description written\nby hand. Content makers who want to reuse this material, for an anniversary package, a breaking\nstory that needs context, or a compilation, have to search on those sparse terms or scrub through\nhours of tape themselves.\n\nModern content operations make this worse before AI makes it better: the volume of new footage\ngrows every year, on demand and streaming platforms want the archive searchable at the level of a\nclip rather than a programme, and content makers want to search for specific shots in natural\nlanguage, such as the ABC's example of a cricketer with zinc on their nose, rather than the exact\nkeywords a manual index was built around. Multimodal models can watch video and listen to audio,\nand describe both in natural language, which makes richer, clip level metadata practical at\narchive scale.","problemStats":[],"howItWorks":"1. **Ingest.** Video and audio files, old and new, are pulled into an AI enhanced digital asset or\n   archive management system, either as a one off backfill of the historical archive or as part of\n   the daily ingest pipeline for new footage.\n2. **Analyse.** A multimodal model segments the file and, for each segment, describes what is shown\n   (people, objects, scenes, on screen text and logos), transcribes and translates what is said,\n   and recognises known people and public figures, going well beyond fixed keyword lists to\n   describe the actual visual and audio content.\n3. **Structure the output.** The generated descriptions are turned into structured, searchable\n   metadata: entities, timecodes, categories and free text descriptions, consistent across decades\n   of inconsistent source material.\n4. **Human review.** A documentalist or archivist validates and corrects the AI generated metadata,\n   particularly for sensitive categories (identifying named individuals, historically significant\n   events), and that correction feeds back into improving the model over time.\n5. **Search and reuse.** Content makers search the archive in natural language rather than exact\n   keywords, and can retrieve a specific clip in seconds instead of scrubbing through the source\n   tape, freeing time for editorial work rather than manual search.","valueDrivers":["employee-productivity","cost-to-serve","revenue-growth"],"kpis":["search-time-reduction","productivity-gain","interactions-handled","revenue-uplift"],"indicativeValue":{"referenceOrg":"A broadcaster with a 50,000 hour video archive","inputs":[{"key":"archiveHours","label":"Archive hours to catalogue","low":50000,"high":50000,"unit":"hours","note":"The reference archive."},{"key":"humanTaggingCostPerHour","label":"Fully loaded cost of manual cataloguing per hour of footage","low":15,"high":40,"unit":"USD per hour of footage","note":"Editorial assumption for archivist or documentalist time. Replace with your own cost."},{"key":"aiCoverageShare","label":"Share of the archive AI tagging can cover before human review","low":0.5,"high":0.9,"unit":"fraction of archive hours","note":"Editorial assumption, replace with your own. Panorama Audiovisual reports RTVE's medium term goal of cataloguing 198,220 hours of material from its regional centres, and a 2025 contracted scope of 20,000 hours (whose printed components, 18,000 plus 5,000 plus 5,000, sum to 28,000); either figure is a small share of that goal. The ABC's one million video records analysed within two weeks says nothing about the share of an archive AI can cover. Neither source supports a specific coverage share."}],"formula":"archiveHours * humanTaggingCostPerHour * aiCoverageShare","currency":"USD","period":"one off","resultLabel":"Archive cataloguing cost avoided","caveat":"Gross cataloguing cost avoided only. It leaves out the cost of running the AI service, the documentalist time still needed to validate and correct output, and the licensing revenue or editorial value the newly discoverable footage may unlock."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The AI analysis itself is largely a managed multimodal model call; the harder work is the archive's own metadata schema and validation workflow, migrating inconsistent decades old records into a structure the AI's output can populate, and integrating with the existing asset management system rather than building a parallel one.","dataPrerequisites":["Digitized video and audio files, or a digitization pipeline for analogue source material","A target metadata schema (entities, categories, timecodes) that both legacy and AI generated records can populate","Existing catalogue records to migrate or reconcile with newly generated metadata"],"integrations":["Digital asset or media asset management (MAM) system","Multimodal AI model or vendor platform for video and audio analysis","Search index that serves natural language queries over the generated metadata","Documentalist or archivist review and correction interface"]},"implementation":{"steps":[{"title":"Pilot on a bounded, representative slice","detail":"Choose a few thousand hours that span the archive's range of formats and eras before committing to the full backfill, and measure both metadata quality and documentalist correction time."},{"title":"Design the metadata schema before the pipeline","detail":"Agree the target categories, entities and timecodes the organization actually needs to search on, so the AI's output is structured to that schema from the start rather than reworked later."},{"title":"Build the human review step in, not on","detail":"Give documentalists an interface to validate and correct AI generated metadata as part of the ingest workflow, especially for named people and sensitive events, rather than treating review as an afterthought."},{"title":"Prioritise by reuse value, not just volume","detail":"Catalogue the material most likely to be reused first (anniversaries, recurring formats, frequently requested topics) so the pilot demonstrates value quickly, then expand to the full backfill."},{"title":"Integrate into daily ingest","detail":"Once quality is proven on the backfill, apply the same pipeline to new footage as it is ingested, so the archive stops growing its backlog even as historical material catches up."}],"guardrails":["Human validation of AI generated metadata before it is published as authoritative, particularly identification of named individuals","A documented policy for what the system may not do, such as making rights or licensing decisions from metadata alone","Access controls on sensitive archive material that mirror the organization's existing editorial and legal restrictions","Version history so a correction to AI generated metadata is auditable and reversible"],"humanInTheLoop":"Documentalists and archivists validate and correct AI generated metadata as part of the ingest workflow, with particular attention to identifying named people, historically sensitive footage and anything that will inform a licensing or rights decision; their corrections are the mechanism that improves the model's output over time, not a one time quality check.","kpisToInstrument":["Search time for a content maker to locate a specific clip, before and after","Share of the archive with AI generated metadata, backfill and new intake separately","Documentalist correction rate on AI generated metadata, by category","Archive material reused in new productions or licensed, before and after"],"failureModes":[{"title":"Confident but wrong identification","detail":"Multimodal models can misidentify people or events with fluent, plausible sounding metadata. Require human validation before any AI identified person or event is treated as authoritative."},{"title":"Metadata schema drift","detail":"AI generated categories that do not map cleanly to the archive's existing schema fragment search rather than improving it. Fix the target schema before the pipeline runs at scale."},{"title":"Backlog blindness","detail":"A backfill project that does not also cover new intake just moves the backlog forward in time. Build the pipeline for daily ingest from the start, even if the backfill runs separately."},{"title":"Sensitive content surfaced without control","detail":"Making decades of raw, unpublished footage newly searchable can surface material that was never meant for wide internal access. Apply the organization's existing access and editorial controls to AI generated search results, not just to the original files."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Cataloguing objects, scenes, logos and spoken content is not listed in Annex III and is typically minimal risk. Annex III point 1(a) covers remote biometric identification: the automated, one to many matching of a person's face or voice, without their active involvement and typically at a distance, against a reference database of identified individuals to establish who they are, in so far as its use is permitted under relevant Union or national law; it excludes one to one biometric verification. A feature that recognises and names a specific person in archive footage by comparing them against such a database meets that definition and is high risk, while grouping similar looking footage without assigning an identity does not. Article 6(3) lets a provider assess a listed system itself as not high risk when it performs only a narrow procedural task, but that derogation is unlikely to cover a system whose purpose is naming an individual, so treat person recognition as high risk by default. Point 1(b) covers biometric categorisation, inferring a sensitive or protected attribute from a person's face or voice. RTVE's 2025 contract specifies speaker gender identification as a feature; whether that counts as a protected attribute under point 1(b) is contested, since Recital 54 ties that category to attributes protected under GDPR Article 9(1), and sex or gender is not listed there. Treat gender inference from voice as potentially high risk and apply the same governance the organization uses for other biometric systems until that question is settled."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"Annex III: High-Risk AI Systems Referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 1(a) covers remote biometric identification, the automated, one to many comparison of a face or voice against a database of identified individuals, relevant where archive tagging recognises and names specific people. Point 1(b) covers biometric categorisation, inferring a sensitive attribute; whether gender inferred from voice, such as RTVE's 2025 contracted speaker gender identification feature, falls under this point is contested and treated here as potentially high risk rather than settled."}],"controls":["Human validation of any AI generated identification of a named person before publication","Inventory entry for the tagging system, its model provider and the categories of personal data it processes","Documented retention and access policy for AI generated metadata, aligned with the archive's existing editorial controls","Bias and accuracy checks on person recognition across demographic groups before wide rollout"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai's own strength is conversational and structured text retrieval, not multimodal video\nanalysis, so the visual and audio tagging step itself calls an external multimodal model or\nvendor platform through a **custom function**. Where Blits.ai fits directly is the audio side and\neverything downstream: its **self hosted transcription and speaker diarization** produces word\nlevel, speaker separated transcripts with automatic language detection, which can feed the same\nmetadata pipeline as the visual tags, and its **knowledge base** with **hybrid retrieval** makes\nthe resulting transcripts and generated metadata searchable in natural language once they are\ningested as documents.\n\nA practical build is an internal **AI agent** for documentalists: it calls the tagging or\ndiarization pipeline through **custom functions** and presents the generated metadata for review\ninside an internal tool, letting a documentalist approve or reject entries in conversation rather\nthan through a separate validation screen. **Guardrails** shape what the agent may do, the\nplatform's **PII masking at the gateway** covers personal data as transcripts pass through it,\nthe **run history and audit trail** record every decision for later review, and the whole\nworkflow can run as an **agentic workflow** with **human in the loop approval** before AI\ngenerated metadata, especially named person identification, is published as authoritative."},"faq":[{"question":"Can AI tag an entire decades old video archive automatically?","answer":"It can generate a first pass of rich metadata at a scale manual cataloguing cannot match. The ABC analysed one million video records within two weeks using Gemini in Vertex AI. RTVE has pursued automatic metadata for its archive since 2020 and, under a 2025 tender, contracted Crosspoint and Amplify to automatically analyse 20,000 hours of content; separately, a NexTReT case study (undated) describes a similar service built into RTVE's ARCA system, with an interface for documentalists to validate the results."},{"question":"Does AI metadata tagging replace archivists and documentalists?","answer":"No, not in the deployments we found. The ABC describes a human in the loop correction process that reviews and refines the AI generated metadata to ensure consistency and meet its quality standards, and NexTReT's RTVE deployment built an interface for documentalists to validate the results. We recommend treating human review of named people and sensitive material as a guardrail on any such system, not an afterthought."},{"question":"Is recognising people in archive footage a high risk use under the EU AI Act?","answer":"It can be. Object, scene and logo detection is typically minimal risk, but recognising a specific named person by matching their face or voice against a database of known individuals is remote biometric identification under Annex III point 1(a), which is high risk. RTVE's 2025 contract specifies speaker gender identification as a feature, not confirmed live by any source we found; whether inferring gender from voice counts as biometric categorisation under point 1(b) is contested, since that category is tied to attributes protected under GDPR and sex is not among them. We recommend treating gender inference from voice as potentially high risk and applying the same governance as other biometric systems."},{"question":"What does a broadcaster get back for cataloguing its archive with AI?","answer":"Faster search for content makers (the ABC's case study describes finding the clip they need in seconds instead of taking an hour to find the right video and then scrubbing through hours of tape), a larger share of the archive that is actually discoverable and reusable, and, for organizations that license footage, a larger catalogue of material that can be found and cleared for reuse in the first place."}],"related":["audio-and-video-transcription-and-captioning","product-content-and-catalog-enrichment"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched and written from vendor case studies naming the ABC and RTVE."},{"date":"2026-09-28","note":"Fact checked against sources: stopped presenting RTVE's contracted 20,000 hour scope as an achieved annual result, since the vendor announcement says the service had not launched and gives no period; rewrote the metaDescription, FAQ 1 and the indicativeValue note accordingly and relabelled the AI coverage share as an editorial assumption; corrected the cricketer example's attribution to ABC content makers searching CoDA, not audiences asking AI answer engines; rewrote the EU AI Act basis, guidance note and FAQ 3 to cite Annex III point 1(a), remote biometric identification, and point 1(b), biometric categorisation, for RTVE's speaker gender identification; replaced \"agent desktop\" with an internal tool and softened the analytics claim to match the platform feature inventory; corrected the ABC time paraphrase in FAQ 4; and set the revenue KPI to revenue uplift."},{"date":"2026-09-28","note":"Second fact check pass: dropped the unsourced \"first technology\" superlative in the problem statement; narrowed FAQ 2 to only what the ABC and NexTReT sources say about human review, moving the emphasis on named people and sensitive material into the implementation guardrails instead; stopped presenting RTVE's speaker gender identification as live and hedged the EU AI Act point 1(b) classification as contested rather than settled, in FAQ 3, risk.euAiAct.basis and the Annex III guidance note; separated the NexTReT deployment and the 2025 Crosspoint and Amplify contract in the RTVE evidence summary, verification note and FAQ 1, since neither source links them; softened the ABC summary to match the source's prospective wording on human correction; changed the ABC metric qualifier to approximately; fixed the indicativeValue period to match the formula; and added the 198,220 and 28,000 hour qualifiers to the indicativeValue note. Reworded blitsAi.howToBuild to match the platform feature inventory (approve or reject, not correct; PII masking at the gateway; audit trail records decisions, not corrections)."},{"date":"2026-09-28","note":"Third fact check pass: the RTVE evidence record no longer combines the NexTReT ARCA deployment with the 2025 Crosspoint and Amplify tender under one stage, year, vendor list and language list, since no source dates or links them; the record now covers the NexTReT deployment only, dated from the case study page's own publish metadata rather than an invented year, with the 2025 tender kept as context in the verification note. Dropped \"earlier\" from FAQ 1, since no source says the NexTReT deployment predates the 2025 tender. Set the ABC metric qualifier back to exact, since the source quote states the one million figure without an approximating word."}],"slug":"media-archive-metadata-tagging","url":"https://www.blits.ai/ai-use-cases/media-archive-metadata-tagging","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1000000,"min":1000000,"max":1000000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"abc-archive-ai-metadata-tagging","pooled":true}]}],"indicativeValueResult":{"low":375000,"high":1800000},"evidence":["abc-archive-ai-metadata-tagging","rtve-archive-metadata-automation"]},{"title":"AI native platform for drug target discovery and molecule design","shortTitle":"AI drug discovery platform","seoTitle":"AI drug discovery platform for biotech","metaDescription":"Generative AI screens drug targets and designs molecules before chemists synthesize them. Insilico reports 12 to 18 months to a candidate, against 2.5 to 4 years.","definition":"An AI native research platform that prioritizes disease targets from biological data, generates and optimizes candidate drug molecules computationally, and predicts their properties before a chemist synthesizes and tests them, so a pharmaceutical or biotech company reaches a validated preclinical candidate with far fewer molecules made and tested than a conventional medicinal chemistry program.","aliases":["generative AI drug discovery","AI drug design platform","computational drug discovery"],"industries":["pharma-and-life-sciences"],"functions":["operations","analytics-and-reporting"],"patterns":["prediction-and-scoring","content-generation"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"drug discovery","problem":"Conventional small molecule drug discovery starts from a target hypothesis and then screens, makes\nand tests thousands of candidate compounds to find a handful worth advancing, before any clinical\ntesting begins. Insilico Medicine puts traditional early stage discovery at two and a half to four\nyears to a preclinical candidate; Recursion Pharmaceuticals puts the industry average at over 2,500\ncompounds synthesized and 42 months per program.\n\nAn AI native platform tries to spend that cost computationally instead: predicting which targets are\nworth pursuing and which molecules are likely to bind, be selective and be safe before a chemist\nmakes anything, so the synthesis and testing budget is spent on a much shorter list of higher\nprobability candidates. It does not remove the need for real chemistry, real assays or real clinical\ntrials; it changes how many molecules a team has to make to find one worth testing in a person.","problemStats":[],"howItWorks":"1. **Model the biology.** Multi omics data, scientific literature and patent intelligence feed models\n   that score and rank potential disease targets against criteria such as novelty, druggability and\n   safety.\n2. **Generate candidate molecules.** Generative chemistry models propose novel molecular structures\n   against the chosen target, rather than screening only a fixed, existing compound library.\n3. **Predict properties computationally.** Models predict binding, selectivity, toxicity and\n   pharmacokinetic properties before any molecule is made, narrowing a large design space to a short\n   list worth synthesizing.\n4. **Synthesize and test the short list.** Chemists make and test only the highest scoring molecules,\n   and the assay results feed back into the models for the next design round.\n5. **Advance to preclinical and clinical development.** The nominated candidate moves into the same\n   regulated preclinical and clinical pathway as any other drug. In this use case, the AI's role ends\n   at candidate nomination (some platforms, such as Insilico's, also offer separate clinical trial\n   prediction tools, which are a different use case); every later stage here is validated by\n   conventional testing.","valueDrivers":["speed","cost-to-serve","employee-productivity"],"kpis":["processing-time-reduction","productivity-gain"],"indicativeValue":{"referenceOrg":"A biotech starting 10 early discovery programs a year","inputs":[{"key":"programs","label":"Discovery programs started per year","low":10,"high":10,"unit":"programs per year","note":"The reference biotech."},{"key":"monthsSaved","label":"Months saved per program versus a conventional discovery timeline","low":12,"high":24,"unit":"months per program","note":"Conservative against the benchmarks on this page (Insilico Medicine reports reaching preclinical candidate nomination in an average of 12 to 18 months against a traditional 2.5 to 4 years; Recursion Pharmaceuticals reports about 17 months against an industry average of 42 months), because a first program on a new platform rarely matches a mature platform's average."},{"key":"costPerMonth","label":"Fully loaded cost of a discovery program team per month","low":150000,"high":300000,"unit":"USD per program per month","note":"Editorial assumption for a mid sized medicinal chemistry and biology team; replace with your own."}],"formula":"programs * monthsSaved * costPerMonth","currency":"USD","period":"per year","resultLabel":"Annual discovery program cost avoided through faster preclinical candidate nomination","caveat":"Time saved in discovery only. It leaves out the platform's own cost, the preclinical and clinical development that follows candidate nomination (unchanged by this use case), and the fact that a faster nomination is not the same as a successful drug: Recursion Pharmaceuticals' own risk disclosure says the risk of failure in pharmaceutical research and development is high and failure can occur at any stage before or after regulatory approval, AI discovered or not."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The AI shortens discovery, not validation. A nominated candidate still needs full preclinical safety and efficacy testing, an investigational new drug filing and clinical trials under existing pharmaceutical regulation. Getting real predictive power out of the models needs a large, curated internal dataset of assay results to train and validate against, not just public data.","dataPrerequisites":["Curated multi omics data (genomics, proteomics, disease models) for target scoring","A large internal dataset of assay results (binding, ADMET, toxicity) to train and validate predictive models against","Structural biology data for the targets in scope, where available"],"integrations":["Electronic lab notebook and assay data management systems","Compound registration and inventory systems","Computational chemistry and structural biology tools"]},"implementation":{"steps":[{"title":"Pick a target class the model can actually score","detail":"Start where public and internal data on the biology and known chemical matter is strongest; a completely novel target with almost no data starves the model of signal."},{"title":"Set a synthesize and test budget per round","detail":"Cap how many AI proposed molecules a chemistry team commits to making per cycle, and measure hit rate against that budget, rather than running an open ended search."},{"title":"Keep a qualified chemist in every loop","detail":"A medicinal chemist reviews every proposed structure for synthesizability and known liabilities before it is made, and can veto a candidate the model scored well."},{"title":"Track the whole funnel, not just discovery speed","detail":"Preclinical candidate nomination is not success. Follow programs into IND filing and Phase 1 to see whether faster discovery produces better drugs, not just more of them, faster."},{"title":"Validate predictions against your own assay data before trusting them","detail":"A platform's published benchmark numbers come from its own historical programs. Run a blinded internal validation before betting a program's timeline on the model."}],"guardrails":["A qualified medicinal chemist reviews and approves every AI proposed molecule before synthesis","Predictive model performance is validated against blinded internal assay data, not only the platform's own published benchmarks"],"humanInTheLoop":"Chemists and biologists review every AI proposed target and molecule; the AI narrows the design space and predicts properties, it does not decide what gets synthesized or what advances. A program only moves to candidate nomination after conventional wet lab confirmation of the predicted properties.","kpisToInstrument":["Molecules synthesized and tested per program, against the target budget","Time from project initiation to preclinical candidate nomination","Attrition rate of AI nominated candidates through IND filing and Phase 1, compared with the company's historical baseline"],"failureModes":[{"title":"Optimizing for a metric that is not a drug","detail":"A model can hit its binding or property targets and still nominate a molecule that fails for reasons the model was not trained to predict, such as manufacturability or an off target effect found only in later testing. Track attrition through clinical development, not only discovery speed."},{"title":"Overfitting to public and vendor benchmark data","detail":"A platform's published cycle time and molecule count figures come from its own historical programs and target classes. Validate on your own target and internal data before assuming the same numbers apply."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Target scoring and molecule generation are not a safety component of an Annex I product and are not one of the Annex III high risk areas (Article 6), so they do not become high risk on that route. Insofar as the platform and its training are themselves scientific research and development, activity that stays there falls outside the Regulation entirely under the Article 2(6) research exclusion. The resulting drug candidate is separately regulated as a medicine, not as an AI system, through the normal pharmaceutical approval pathway; conventional preclinical and clinical testing validates the AI's outputs before anything reaches a patient."},"regulations":["eu-ai-act","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Article 2: Scope","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/2/","note":"Article 2(6) says the Regulation \"does not apply to AI systems or AI models, including their output, specifically developed and put into service for the sole purpose of scientific research and development\"; Article 2(8) extends this to \"any research, testing or development activity regarding AI systems or AI models prior to their being placed on the market or put into service.\" Article 6 itself, by contrast, says nothing about research: it sets the Annex I safety component route and the Annex III route for high risk classification."},{"title":"Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products","issuer":"US Food and Drug Administration","region":"north-america","url":"https://www.fda.gov/media/184830/download","note":"This January 2025 draft guidance covers AI models used to produce information that supports a regulatory decision on a drug, such as evidence in a submission. It explicitly excludes this use case's discovery stage: \"the use of AI for the purposes of drug discovery is not in the scope of this guidance,\" so a platform that stops at candidate nomination sits outside it; the guidance's credibility assessment framework becomes relevant only once AI generated outputs are used to support a later regulatory decision."}],"controls":["Documented model validation against internal assay data before a program relies on a prediction","Named chemist and biologist sign off on every candidate before it advances to synthesis and formal preclinical testing","Data provenance and access controls over the proprietary assay data used to train and validate models"],"incidents":[]},"blitsAi":{"howToBuild":"Drug target scoring and molecule generation need specialized biology and chemistry models that sit\noutside Blits.ai's scope; Blits.ai is not a computational chemistry or structural biology platform.\nWhat Blits.ai can support well is the research operations layer around a discovery program: an\n**AI agent** backed by a **knowledge base** and **hybrid retrieval** over a program's internal\nliterature reviews, target briefs and assay reports, so scientists can ask plain language questions\nacross a program's own documentation instead of searching file by file.\n\n**Custom functions** can connect the agent to an electronic lab notebook or compound registration\nsystem to look up assay results or a compound's registration status, and an **agentic workflow**\ncan route a newly generated candidate through the internal review and sign off steps described in\nthe implementation guardrails above, with **human in the loop approval** at each checkpoint. Keep\nthe target scoring and molecule generation models themselves outside the platform; use Blits.ai for\nthe knowledge access and review workflow around them."},"faq":[{"question":"Has an AI discovered drug reached patients yet?","answer":"Insilico Medicine's rentosertib, a candidate for idiopathic pulmonary fibrosis whose target and molecule were both identified with its AI platform, became the company's first asset to enter a Phase III clinical trial, in July 2026, after a randomized Phase IIa trial published in Nature Medicine showed a positive lung function signal. According to Insilico, rentosertib remains investigational and has not been approved by any regulatory authority."},{"question":"How much faster is AI native drug discovery?","answer":"Insilico Medicine reports reaching preclinical candidate nomination for 22 programs between 2021 and 2024 in an average of 12 to 18 months, against a traditional 2.5 to 4 years, synthesizing only about 60 to 200 molecules per project. Recursion Pharmaceuticals reports advanced candidates have been delivered by synthesizing about 330 compounds per program in about 17 months, against industry averages of over 2,500 compounds and 42 months. Both figures describe the companies' own historical programs, not a guarantee for any specific target."},{"question":"Does this replace medicinal chemists?","answer":"No. The platform proposes structures and predicts properties, but the molecules still have to be made and tested in real assays, and the programs on this page still rely on chemistry and biology teams to validate what the models propose. We recommend a qualified chemist review every AI proposed structure before synthesis (see the implementation guardrails)."},{"question":"Is this the same as AI used in clinical trials or regulatory writing?","answer":"No. This use case covers the discovery stage, from target identification to a nominated preclinical candidate. Matching patients to trials and drafting clinical study reports are separate, later stage use cases with their own evidence."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched from Insilico Medicine's and Recursion Pharmaceuticals' own press releases."},{"date":"2026-09-28","note":"Editorial review: fixed the metaDescription's unsupported 30 month claim, dropped the unsourced superlative and industry wide approval claim from faq[0], replaced the misdescribed FDA guidance reference with the FDA's drug specific AI guidance, and rebased risk.euAiAct.basis and the Article 6 note on the provisions that actually apply (Article 6 safety component and Annex III routes, Article 2(6) research exclusion)."},{"date":"2026-09-28","note":"Second editorial review: rewrote faq[2] as practice advice instead of an unsourced factual claim about deployments, rebased indicativeValue.caveat on Recursion's own risk disclosure instead of an unsourced \"most candidates fail\" claim, replaced the unsourced \"most of that cost\" sentence and the \"companies in this field commonly describe\" generalization in the problem statement with the two named sources' own figures, scoped howItWorks step 5 to this use case (Insilico's separate clinical trial prediction tools are a different use case), and dropped \"about\" before \"42 months\" in indicativeValue since the source states 42 months without a qualifier."}],"slug":"ai-drug-discovery-platform","url":"https://www.blits.ai/ai-use-cases/ai-drug-discovery-platform","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":60,"min":60,"max":60,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"insilico-medicine-pharma-ai-platform","pooled":true},{"id":"recursion-pharmaceuticals-os-platform","pooled":true}]}],"indicativeValueResult":{"low":18000000,"high":72000000},"evidence":["insilico-medicine-pharma-ai-platform","recursion-pharmaceuticals-os-platform"]},{"title":"AI next best action prompts for wealth advisors","shortTitle":"Advisor next best action","seoTitle":"AI next best action for wealth advisors","metaDescription":"AI next best action engines give wealth advisors a ranked list of client prompts. In 2025, UBS said 80% of its US advisors actively used its STAAT Insights engine.","definition":"An AI engine for wealth advisors, not customers, that scans an advisor's whole book and surfaces a short, ranked list of client specific prompts, such as idle cash, a maturing deposit, a concentration to review, a life event or an early sign of attrition, each with the reasoning and data behind it, for the advisor to act on or dismiss.","aliases":["advisor insights engine","relationship manager nudges","book of business opportunity alerts","wealth client prompts"],"industries":["wealth-and-asset-management","banking"],"functions":["sales","marketing","analytics-and-reporting"],"patterns":["recommendation-and-personalization","prediction-and-scoring","content-generation"],"channels":["internal-tools","email"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"front-office","problem":"An advisor responsible for a large book of clients cannot watch every account every day. The signals are\nthere (cash building up after a sale, a deposit about to mature, a portfolio drifting away from its\nmandate, a client who has stopped logging in), but they sit in different systems and surface too\nlate, often when the client has already moved money or called a competitor.\n\nDashboards do not solve it: they show everything and prioritize nothing. What advisors need is a\nshort list each morning of the few clients worth calling, why, and what to say, with the freedom\nto ignore a prompt that does not fit what they know about the client.","problemStats":[],"howItWorks":"1. **Collect signals.** Holdings, cash flows, maturities, product usage, service contacts,\n   portfolio alignment and permitted external data are gathered per client.\n2. **Score and rank.** Predictive models and business rules score opportunities and risks (for\n   example propensity to invest idle cash, risk of attrition), and a ranking step picks the few that\n   matter most for each advisor.\n3. **Explain.** A language model turns each prompt into a short rationale and suggested talking\n   points, citing the data behind it and, where relevant, the house view or an approved product.\n4. **Advisor decides.** The advisor acts, snoozes or dismisses the prompt, and the feedback is\n   used to improve the ranking.\n5. **Controls on the way out.** Any product recommendation that follows still goes through the\n   firm's suitability and product governance checks.","valueDrivers":["revenue-growth","employee-productivity","customer-experience"],"kpis":["conversion-rate-uplift","revenue-uplift","churn-reduction","employee-adoption","hours-saved","time-saved-per-task"],"indicativeValue":{"referenceOrg":"A wealth manager with 50,000 advised clients","inputs":[{"key":"clients","label":"Advised clients","low":50000,"high":50000,"unit":"clients","note":"The reference firm."},{"key":"promptsActed","label":"Prompts acted on per client per year","low":0.2,"high":0.5,"unit":"prompts per client per year","note":"Editorial assumption, replace with your own advisor capacity and pilot data."},{"key":"conversion","label":"Share of acted prompts that lead to new business","low":0.1,"high":0.2,"unit":"fraction of acted prompts","note":"Editorial assumption. No public source on this page states a conversion rate for advisor prompts."},{"key":"revenuePerWin","label":"Annual revenue per converted prompt","low":300,"high":1000,"unit":"USD per conversion per year","note":"Editorial assumption, for example fees on newly invested cash. Replace with your own margins."}],"formula":"clients * promptsActed * conversion * revenuePerWin","currency":"USD","period":"per year","resultLabel":"Additional annual revenue from acted prompts","caveat":"Gross revenue before cannibalization, costs and the business that advisors would have won anyway. Measure it with a control group of advisors or clients before relying on it."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The explanation layer is easy; the signals are not. It needs a clean client data model across banking, investment and CRM systems, models that are validated and monitored, fairness testing, and adoption work so advisors trust the prompts.","dataPrerequisites":["Client holdings, transactions and cash flows across accounts","Product maturities, mandates and model portfolios","CRM activity, service contacts and prior prompt outcomes","Consent and marketing preference data per client"],"integrations":["Portfolio management and core banking systems","CRM and advisor desktop","Data platform for features and model scoring","Suitability and product governance engine for any resulting recommendation"]},"implementation":{"steps":[{"title":"Start with a handful of high value signals","detail":"Pick three to five prompts with clear value and simple logic (idle cash above a threshold, maturing deposits, large inflows) before building propensity models."},{"title":"Put the reason on every prompt","detail":"Each prompt shows the data that triggered it and a suggested talking point. Advisors ignore prompts they cannot explain to a client."},{"title":"Measure against a control group","detail":"Hold out a random group of clients or advisors and compare outcomes, so the revenue story survives scrutiny from finance and risk."},{"title":"Close the feedback loop","detail":"Capture act, snooze and dismiss with a reason, and retrain or retune rankings on that feedback every cycle."},{"title":"Test for fairness and conduct risk","detail":"Check that prompts do not systematically favor higher fee products or neglect client segments, and review any prompt type that pushes a product."}],"guardrails":["Prompts inform the advisor; nothing is sent to a client automatically","Any product recommendation passes the suitability and product governance checks","Contact respects marketing consent and preferences","Fairness and conflict of interest review of prompt types and rankings","Logged rationale for every prompt shown, with the data used","Prompt and action data are not used to rate individual advisors"],"humanInTheLoop":"The advisor decides whether and how to act on each prompt and owns the resulting advice. Business and risk owners approve new prompt types, and model risk validates the scoring models.","kpisToInstrument":["Prompt action rate and dismiss reasons per prompt type","Conversion and revenue against a control group","Client attrition in treated versus control books","Weekly active advisors using the prompts","Complaints or suitability exceptions linked to acted prompts"],"failureModes":[{"title":"Prompt fatigue","detail":"Too many low value prompts and advisors stop looking. Cap the list and retire prompt types with low action rates."},{"title":"Product push dressed as insight","detail":"Rankings optimize for revenue and drift toward high margin products. Add conduct review and suitability checks."},{"title":"Unexplainable scores","detail":"Advisors cannot tell a client why they called. Show the triggering data with every prompt."},{"title":"Credit shown as credit decision","detail":"A prompt suggests a loan based on a score that is really a creditworthiness assessment, which brings high risk obligations. Keep credit decisions in the regulated credit process."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Ranking investment and service prompts for an advisor is not listed in Annex III. It becomes high risk if the system evaluates the creditworthiness of natural persons, for example to decide which clients are offered lending (Annex III point 5(b)), so keep credit decisions out of the prompt engine. It is also high risk if the system itself is used to monitor or evaluate advisors' performance and behaviour, for example by scoring or ranking advisors on how they act on prompts (Annex III point 4(b)), so keep adoption reporting separate from performance management."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","mas-ai-risk-management","us-sr-11-7","iso-42001","mifid-ii"],"guidance":[{"title":"ESMA public statement on the use of AI in the provision of retail investment services","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf","note":"Expects firms to act in the client's best interest when AI shapes recommendations, and to govern algorithmic bias and data quality."},{"title":"Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT)","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2018/feat","note":"Fairness principles apply to models that decide which customers receive which prompts or offers."},{"title":"Consumer Duty","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/firms/consumer-duty","note":"Sets high standards of consumer protection across financial services and requires firms to put their customers' needs first, which applies to advisor prompts that lead to a sale to retail customers."}],"controls":["Model inventory entries and validation for scoring models","Fairness and conflict of interest testing of rankings per segment","Logging of every prompt, its rationale and the advisor's action","Suitability check on any recommendation that results from a prompt","Periodic review of prompt types by business, risk and compliance"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the scoring stays in the firm's own models or data platform; Blits.ai adds the\nexplanation and delivery layer. An **agentic workflow** runs on a schedule, reads scored signals\nand client context through **custom functions** (REST or SQL) or a **SQL knowledge base**, and an\n**AI agent** writes each prompt's rationale and talking points with **structured output**,\ngrounded in the house view held in a **knowledge base**.\n\nPrompts reach advisors by **email** or through the API inside the advisor desktop, and advisors\ncan ask a **Microsoft Teams** assistant about a client. **Agentic tasks** can watch for\nconditions such as a maturity date and fire a prompt when it is reached. Output **guardrails**\nwith policies the firm writes can block prompts that push products outside policy, and the\nworkflow keeps an **audit trail** of every run. Act, snooze and dismiss on each prompt is\ncaptured and used to improve the ranking."},"faq":[{"question":"Does next best action for advisors actually get used?","answer":"Where it is built into the daily workflow, it can be. In a December 2025 Financial Planning interview about UBS's US wealth management unit, UBS's chief data and analytics officer said 80% of advisors were actively using the STAAT Insights engine. In March 2023 Morgan Stanley listed its Next Best Action engine among its recent AI projects, describing it as an internally built engine that delivers timely, customized messages to clients and prospects, guided by the financial advisor."},{"question":"How do you prove the revenue effect?","answer":"With a control group. Compare treated and untreated advisors or clients over the same period, because the clients an engine flags are often the ones advisors would have called anyway."},{"question":"Is this a high risk AI system?","answer":"Not for investment and service prompts. It becomes high risk under the EU AI Act if it assesses the creditworthiness of individuals, or if the system itself is used to monitor or evaluate advisors' performance and behaviour, so credit decisions should stay in the regulated credit process and prompt data out of performance reviews."}],"related":["wealth-advisor-knowledge-assistant","client-meeting-notes-and-crm-update","offers-and-rewards-agent","suitability-assessment-assistant","portfolio-drift-monitoring-and-rebalancing","insurance-broker-and-agent-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Next Best Action) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added MiFID II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; removed an unsourced advisor book size from the problem; tightened the UBS FAQ answer, the Consumer Duty note and the Blits.ai feedback wording; corrected the Morgan Stanley record (year, stage, channels, wording) and added the Financial Planning date to the UBS record."},{"date":"2026-09-27","note":"Review fixes: removed the UBS time saving metrics (hours per month and hours per meeting) from the STAAT Insights record because the sources attribute them to UBS AI in general, not to the engine; dated the UBS adoption figure to 2025; aligned the Consumer Duty note with the FCA page; limited the Blits.ai feedback claim to Teams or chat delivery; clarified the Morgan Stanley engine and release wording."},{"date":"2026-09-27","note":"Fact checked against sources: all five evidence sources and the three guidance documents reread; added Annex III point 4(b) (monitoring advisor performance) to the EU AI Act basis, the FAQ and the guardrails; dated the UBS interview to December 2025 and aligned the Morgan Stanley FAQ wording with the release."},{"date":"2026-09-27","note":"Review fixes: dropped the unsupported Teams push and Teams thumbs feedback claims from the Blits.ai build notes (limited to email, the API and a Teams assistant, plus act/snooze/dismiss capture); tied the Annex III point 4(b) basis and FAQ answer to the system being used to monitor or evaluate advisor performance, not to any reuse of its data."}],"slug":"next-best-action-for-advisors","url":"https://www.blits.ai/ai-use-cases/next-best-action-for-advisors","benchmarks":[{"kpi":"employee-adoption","label":"Employee adoption","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":80,"min":80,"max":80,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ubs-staat-insights-for-advisors","pooled":true}]}],"indicativeValueResult":{"low":300000,"high":5000000},"evidence":["cimb-niaga-proactive-guidance-agents","citi-wealth-askwealth-and-advisor-insights","jpmorgan-coach-ai-advisers","morgan-stanley-next-best-action","ubs-staat-insights-for-advisors"]},{"title":"AI orchestration of corporate account opening and channel setup","shortTitle":"Corporate onboarding operations","seoTitle":"AI agents for corporate account opening","metaDescription":"After due diligence, an AI agent can read mandates and board resolutions, draft users and payment entitlements, and chase documents; staff approve every change.","definition":"An AI agent that runs the operational setup of a corporate client after the due diligence has been approved: it reads mandates, board resolutions and signatory documents, prepares accounts, users, roles and payment entitlements for approval, configures channel access, and chases outstanding items with the client, turning a manual setup that passes between several teams into a tracked, guided flow.","aliases":["corporate account opening automation","client activation agent","mandate and signatory processing","cash management onboarding"],"industries":["banking"],"functions":["onboarding-and-kyc","operations"],"patterns":["agentic-workflow","document-processing","conversational-agent","content-generation"],"channels":["internal-tools","email","web-chat"],"audience":"customer-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"specialized-businesses","problem":"Winning a corporate mandate is only the start. Before the client can pay a supplier, the bank has\nto open accounts in several entities and currencies, capture authorised signatories and their\nlimits from mandates and board resolutions, set up users and roles in the corporate portal,\nconfigure payment entitlements and approval rules, connect host to host or API channels, and\ncollect missing documents. These steps can involve relationship managers, onboarding teams,\noperations and technical implementation, each working in its own systems.\n\nWhere the same data is entered in several systems, or a case waits for one team or for the client,\nthose handoffs are the natural first places to measure elapsed time and rekeying errors. Mistakes\nin signatory or entitlement setup are also a fraud and control risk.\nAn agent can read the documents, prepare the setup and chase what is missing, but entitlements are\na control: segregation of duties and four eyes approval must stay in place.","problemStats":[],"howItWorks":"1. **Open a case.** When onboarding is approved, the agent opens a setup case with a checklist\n   based on the products, entities and countries in the deal.\n2. **Read the documents.** The agent extracts signatories, signing rules, limits and authorised\n   users from mandates, resolutions and forms, and flags inconsistencies.\n3. **Prepare the setup.** It drafts account opening requests, users, roles and payment\n   entitlements in the target systems' formats, restricted to what the approved mandate allows.\n4. **Chase what is missing.** It sends the client specific requests for missing or unclear items,\n   tracks replies and reads returned documents on arrival.\n5. **Approve and apply.** Operations staff review and approve each setup batch under four eyes;\n   only then are changes applied, and the client and banker get status updates throughout.","valueDrivers":["speed","customer-experience","cost-to-serve","risk-reduction"],"kpis":["cycle-time-days","processing-time-reduction","handling-time-reduction","error-reduction"],"indicativeValue":{"referenceOrg":"A corporate bank activating 1,500 new corporate clients a year","inputs":[{"key":"clients","label":"New corporate clients activated per year","low":1500,"high":1500,"unit":"clients per year","note":"The reference bank."},{"key":"opsHours","label":"Operations and implementation hours per client setup","low":15,"high":30,"unit":"hours per client","note":"Editorial assumption, replace with your own time study."},{"key":"reduction","label":"Share of those hours saved","low":0.2,"high":0.4,"unit":"fraction of time","note":"Editorial assumption; no measured public benchmark was found."},{"key":"hourlyCost","label":"Loaded cost of an operations hour","low":50,"high":80,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."}],"formula":"clients * opsHours * reduction * hourlyCost","currency":"USD","period":"per year","resultLabel":"Operations time released, valued at loaded cost","caveat":"Values operations time only. It leaves out earlier revenue from clients who start transacting sooner, fewer setup errors and the related fraud risk, and the cost of integrating with account and entitlement systems."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Many systems (accounts, portal entitlements, payment channels, CRM) and strict controls on who may grant what. The agent must work through approved interfaces and never bypass segregation of duties.","dataPrerequisites":["Product and country checklists for corporate onboarding","Mandate, resolution and signatory templates and their required fields","Entitlement models of the corporate portal and payment channels"],"integrations":["Client lifecycle or onboarding case management","Core banking account opening","Corporate portal user and entitlement administration","Payment channel and host to host setup","CRM and email for client communication"]},"implementation":{"steps":[{"title":"Map the journey and the handoffs","detail":"Chart every step from approval to first transaction, with owner, system and waiting time, and pick the steps with the most rekeying and waiting."},{"title":"Start with document reading and chasing","detail":"Automate extraction from mandates and forms and the chasing of missing items first; these save time without touching entitlements."},{"title":"Prepare, do not grant","detail":"Let the agent prepare setup batches that humans approve under four eyes, and cap them to the approved mandate so it cannot propose rights beyond it."},{"title":"Give everyone the same status","detail":"Publish one case status to the client, the banker and operations, generated from the case, so nobody chases by email."}],"guardrails":["The agent cannot grant access or entitlements beyond the approved mandate","Four eyes approval on every signatory, user and payment entitlement change","Segregation of duties between the person who prepares and the person who approves","Every change is logged with the source document it was based on"],"humanInTheLoop":"Operations staff approve every setup batch under four eyes before it is applied, and handle any inconsistency the agent flags. The relationship manager owns client communication on sensitive items, and control functions review entitlement changes periodically.","kpisToInstrument":["Elapsed days from approval to first transaction","Operations hours per client setup","Setup errors found after go live","Number of document requests per client and time to receive them"],"failureModes":[{"title":"Entitlements beyond the mandate","detail":"An extraction error gives a user higher limits than approved. Cap proposals to the mandate and require four eyes on every entitlement."},{"title":"Chasing that annoys the client","detail":"Automated reminders repeat or ask for documents already sent. Track what was received and let the banker pause chasing."},{"title":"Silent partial setups","detail":"One system is updated and another fails, leaving an inconsistent state. Apply changes as tracked tasks with reconciliation at the end."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Operational setup of accounts and entitlements for corporate clients is not listed in Annex III and makes no decision about a natural person's access to a service or creditworthiness. The agent chases documents directly with client staff, so Article 50(1) applies: the provider must design the system so that they are informed that they are interacting with an AI system, unless that is obvious from the context. A purely internal version without client contact would be minimal risk."},"regulations":["eu-ai-act","gdpr","dora","mas-ai-risk-management","apra-cps-230","eu-amlr"],"guidance":[{"title":"Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Applies to the conversational chasing of documents with client staff."},{"title":"NIST AI Risk Management Framework","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/itl/ai-risk-management-framework","note":"A structure for mapping and managing the risks of an agent that prepares changes in control relevant systems."}],"controls":["Segregation of duties and four eyes approval enforced by the target systems, not by the agent","Audit trail of every prepared and applied change with its source document","Periodic entitlement reviews for new clients","Inventory entry for the agent with its permitted actions"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that reads mandate and signatory documents, fills a\nstructured setup plan with **structured output**, and prepares requests in the target systems\nthrough **custom functions** (REST calls). The **tool execution policy** restricts the agent to\npreparing requests, and **human in the loop approval**, with its threshold set so that every\nsetup batch needs confirmation, puts each batch in front of an operations approver before it is\napplied. **Agentic tasks** with scheduled rechecks can wait for\nmissing documents and send reminders.\n\nClient communication runs through the **email channel** or a **web widget** in the onboarding\nportal, with **human handover** to the banker when the client asks for it. The per run **audit\ntrail** records each document and change, **guardrails** and **PII masking** protect signatory\ndata, and **test suites** can replay sample onboarding files before each release."},"faq":[{"question":"Can AI grant access to a new corporate client's users?","answer":"It should only prepare the setup. Granting users, signatories and payment entitlements stays behind four eyes approval and segregation of duties, and the agent must not propose rights beyond the approved mandate."},{"question":"Where should a bank look first for time savings in corporate onboarding?","answer":"We found no public benchmark that splits onboarding time by step, so map your own journey first. Handoffs between teams, rekeying the same data into several systems and waiting for client documents are the usual candidates to measure; reading mandates automatically and chasing missing items address those steps without touching the approval controls."},{"question":"Are there published results?","answer":"We found published results for the wider corporate onboarding journey that this operational setup step sits inside, but not for the step measured on its own. Standard Chartered says machine learning document processing cut average client onboarding time in Corporate and Institutional Banking from 41 to 8 days. Citi says AI enhancements to CitiDirect Commercial Banking, including automated KYC renewals and a digitized onboarding process, have significantly reduced onboarding turnaround times, without giving a figure. Neither source measures the mandate reading, signatory or entitlement setup step by itself, so pilot with document reading and chasing, and measure days from approval to first transaction."}],"related":["business-onboarding-and-ubo-discovery","corporate-client-servicing-assistant","developer-api-integration-assistant","account-servicing-execution","intelligent-document-processing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, kept as draft because no public deployment evidence was found; the catalog cited a McKinsey article that timed out and gives no deployment."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added EU Anti Money Laundering Regulation to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: EU AI Act tier corrected from minimal to limited (Article 50(1) applies because the agent contacts client staff directly); removed the unsourced multi week and days claims from the definition, problem and FAQ; guidance URLs and regulation ids confirmed; added seoTitle and metaDescription. Stays draft: no public deployment evidence."},{"date":"2026-09-27","note":"Review fixes: rewrote the unsourced claims about email, spreadsheets and repeated client requests in the problem and FAQ as design considerations; Article 50 guidance now cites EUR-Lex and names the provider duty; blitsAi wording aligned with the feature inventory (approval threshold, test suites before release); vendor neutral wording in step 2. Stays draft: no public deployment evidence, so the emerging stage is not yet supported by a source."},{"date":"2026-09-27","note":"Review fixes: audience changed from back office to customer facing to match the direct client contact the Article 50 basis relies on; autonomy changed from supervised agent to copilot to match the mandatory four eyes approval on every setup batch; removed the automation rate KPI, since nothing here runs end to end without human touch. Stays draft: still no public deployment or pilot evidence for this slug, so the emerging adoptionStage and the missing evidence records remain open."},{"date":"2026-09-27","note":"Evidence found and added: Standard Chartered's own 2018 press release on its Instabase deployment (average CIB client onboarding time cut from 41 to 8 days) and Citi's 2025 press release on AI enhancements to CitiDirect Commercial Banking (automated KYC renewals, a digitized onboarding process, and a Digital Servicing Hub for document submission). Both cover the wider onboarding journey rather than the mandate, signatory and entitlement setup step in isolation, so the FAQ on published results was rewritten and status moves to review; adoptionStage stays emerging since that narrower step still has no dedicated public evidence."}],"slug":"corporate-account-onboarding-orchestration","url":"https://www.blits.ai/ai-use-cases/corporate-account-onboarding-orchestration","benchmarks":[{"kpi":"cycle-time-days","label":"Cycle time","unit":"days","aggregate":false,"higherIsBetter":false,"n":1,"nUpTo":0,"median":8,"min":8,"max":8,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"standard-chartered-instabase-client-onboarding","pooled":true}]}],"indicativeValueResult":{"low":225000,"high":1440000},"evidence":["citi-citidirect-commercial-onboarding","standard-chartered-instabase-client-onboarding"]},{"title":"AI portfolio drift monitoring and rebalancing proposals","shortTitle":"Drift and rebalancing","seoTitle":"AI portfolio drift monitoring and rebalancing","metaDescription":"AI flags portfolios that drift outside their bands and drafts rebalancing proposals for approval. SimCorp's Wealth Lens scores portfolios from 0.00 to 1.00.","definition":"Continuous monitoring of every client portfolio against its mandate or model, which detects drift beyond agreed bands and prepares a tax aware, low turnover rebalancing proposal with its rationale for an advisor or portfolio manager to approve before any trade is placed.","aliases":["portfolio drift alerts","automated rebalancing proposals","mandate monitoring","model portfolio drift detection"],"industries":["wealth-and-asset-management","banking"],"functions":["operations","risk-management","analytics-and-reporting"],"patterns":["anomaly-detection","agentic-workflow","prediction-and-scoring","content-generation"],"channels":["internal-tools","email"],"audience":"back-office","autonomy":"copilot","adoptionStage":"emerging","segment":"middle-office","problem":"Portfolios drift as markets move, cash comes in and clients make their own trades. After a strong\nrun in equities, a balanced mandate can hold more equity than its model allows, and a single stock\ncan grow into a concentration the client never agreed to. Where drift is checked on a calendar,\nquarterly or annually, portfolios can move outside their bands between reviews.\n\nA calendar check also spends time on portfolios that did not need attention. A sound proposal has to\naccount for taxes, costs, restrictions and client preferences, which takes time when it is done by\nhand for each account, and supervisors need a record of why a trade was or was not made.","problemStats":[],"howItWorks":"1. **Monitor continuously.** Holdings are compared daily with each portfolio's mandate or model,\n   restrictions and tolerance bands; breaches and near breaches are scored by size and urgency.\n2. **Prioritize.** Portfolios are ranked for attention, for example by a rebalancing score, so the\n   team sees the few that matter first.\n3. **Propose.** An optimizer builds a rebalancing proposal that respects taxes, costs, restrictions\n   and minimum trade sizes; the model writes a short rationale explaining the drift and the trades.\n4. **Approve.** The advisor or portfolio manager reviews, adjusts and approves the proposal; for\n   advisory accounts the client's consent is obtained before execution.\n5. **Execute and record.** Approved trades go to order management, and the trigger, proposal,\n   approval and executed trades are logged against the mandate.","valueDrivers":["risk-reduction","employee-productivity","compliance","customer-experience"],"kpis":["time-saved-per-task","hours-saved","processing-time-reduction","cycle-time-days","error-reduction"],"indicativeValue":{"referenceOrg":"A wealth manager with 20,000 managed or advised portfolios","inputs":[{"key":"portfolios","label":"Portfolios monitored against a mandate or model","low":20000,"high":20000,"unit":"portfolios","note":"The reference firm."},{"key":"reviewsPerYear","label":"Manual drift reviews per portfolio per year today","low":4,"high":4,"unit":"reviews per year","note":"Quarterly calendar review, an editorial assumption."},{"key":"minutesSaved","label":"Minutes saved per review","low":10,"high":20,"unit":"minutes per review","note":"Editorial assumption, replace with your own. No public source on this page states a time saving."},{"key":"hourlyCost","label":"Fully loaded cost per hour of portfolio managers and support staff","low":60,"high":120,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"portfolios * reviewsPerYear * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Value of staff time released from manual drift reviews","caveat":"Productivity only. It leaves out the effect on client outcomes (risk kept within mandate, tax savings), which depends on markets and is not measured by any source on this page, and the cost of the monitoring and optimization tools."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Monitoring against models is mature and often rules based already. Adding proposals needs clean tax lot data, restrictions and preferences per account, an optimizer, and an approval path into order management.","dataPrerequisites":["Daily holdings and cash per portfolio, with tax lots where relevant","Mandates, model portfolios, tolerance bands and restrictions per account","Client preferences such as exclusions and tax sensitivity","Cost and minimum trade size parameters"],"integrations":["Portfolio management and accounting systems","Optimization or rebalancing engine","Order management system for approved trades","CRM for client communication and consent on advisory accounts"]},"implementation":{"steps":[{"title":"Agree the bands and triggers","detail":"Define tolerance bands per mandate type and what counts as a breach, a near breach and an exception, with investment committee sign off."},{"title":"Monitor before you propose","detail":"Run daily drift detection and prioritization first; measure how many portfolios breach and how quickly they are handled."},{"title":"Add proposals with a rationale","detail":"Connect an optimizer for trades and use the language model only to explain the drift and the proposed trades in plain language for the approver and, where needed, the client."},{"title":"Keep execution behind approval","detail":"Route every proposal to a named approver with limits; for advisory accounts capture client consent. Automatic execution, if ever allowed, stays within narrow discretionary limits."},{"title":"Log for supervision","detail":"Store the trigger, proposal, approval, overrides and executed trades per portfolio so reviewers can see that each portfolio stayed within its mandate."}],"guardrails":["No trade without approval by an authorized person, and client consent on advisory accounts","Trades generated by the optimizer, not by the language model","Restrictions, exclusions and tax preferences enforced as hard constraints","Limits on turnover and trade size per proposal","Full log of trigger, proposal, approval and execution"],"humanInTheLoop":"Advisors or portfolio managers approve every proposal and can override it with a recorded reason; the investment committee owns models and bands; supervisors review exceptions and overrides.","kpisToInstrument":["Share of portfolios outside tolerance bands and median days to resolution","Proposals approved, modified or rejected, with reasons","Turnover and realized tax impact of approved rebalances","Time spent per review before and after","Mandate breaches found in supervisory review"],"failureModes":[{"title":"Alert floods","detail":"Tight bands trigger constant alerts and staff start ignoring them. Tune bands and prioritize by size and urgency."},{"title":"Tax blind proposals","detail":"Rebalancing realizes gains the client did not need to. Use tax lot data and constraints in the optimizer."},{"title":"Rationale that does not match the trades","detail":"The narrative explains a different trade than the optimizer produced. Generate the text from the proposal data and check it."},{"title":"Silent execution creep","detail":"Approval becomes a click through. Monitor approval times and override rates and sample proposals."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Monitoring portfolios and proposing trades for human approval is not listed in Annex III and is not a prohibited practice under Article 5, so the tier turns on the firm's role under Article 50. A firm that builds or brands the rationale writer in house is a provider under Article 50(2) and must mark the generated text in a machine readable format: drafting a rationale for the drift and the proposed trades goes beyond the exemption for an assistive function for standard editing, so for that firm the tier is limited. Article 50(1) also applies once the rationale reaches the client, as this page's own implementation step allows. A firm that only deploys a third party feature for internal approver use has no Article 50 duty, and for that firm the tier is minimal. Investment conduct rules such as MiFID II suitability and best execution still apply to the resulting trades."},"regulations":["eu-ai-act","gdpr","dora","mas-ai-risk-management","apra-cps-230","iso-42001","mifid-ii"],"guidance":[{"title":"ESMA public statement on the use of AI in the provision of retail investment services","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf","note":"Covers AI used to manage and rebalance client portfolios and expects heightened diligence on suitability in portfolio management."},{"title":"Artificial Intelligence (AI) Model Risk Management (information paper)","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Good practices for AI and generative AI model risk management observed in a 2024 thematic review of banks, covering governance and oversight, key risk management systems and processes, and the development and deployment of AI."}],"controls":["Tolerance bands and models approved and version controlled by the investment committee","Monitoring and optimization tools inventoried with owners and validation","Approval limits per role and client consent capture for advisory accounts","Audit trail from trigger to executed trade per portfolio","Periodic review of overrides and exceptions"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai monitoring runs as **agentic tasks**: condition triggered checks (\"when a portfolio\ndrifts beyond its band\") with scheduled rechecks, calling the firm's portfolio and optimization\nsystems through **custom functions** (REST or SQL) or a **SQL knowledge base**. An **agentic\nworkflow** then collects the optimizer's proposal and an **AI agent** writes the rationale with\n**structured output** tied to the proposed trades.\n\nEvery proposal pauses for **human in the loop** approval before anything reaches order management\n(the configurable confirmation threshold is set so that no proposal skips it), and the **tool\nexecution policy** keeps order placement outside the agent's allowed tools unless explicitly\npermitted. The workflow's **audit trail** records each trigger, proposal and decision, notices go\nto approvers by **email** or **Microsoft Teams**, and **monitors** run scheduled health checks on\nthe agents and alert the team when they fail."},"faq":[{"question":"Is this already automated today?","answer":"Rules based rebalancing is already automated in digital advice: Vanguard Digital Advisor, for example, rebalances when a portfolio drifts more than 5% from its recommended allocation. The AI steps in the evidence here sit on top of that: ranking large books for attention (SimCorp's Wealth Lens tool scores each portfolio from 0.00 to 1.00 for rebalancing) and drafting plain language talking points, as BlackRock announced its Aladdin Wealth Auto Commentary would do for Morgan Stanley advisors to help them identify issues such as portfolio overweights."},{"question":"Should the AI place trades on its own?","answer":"In advisory books, no: trades need advisor approval and client consent. Even in discretionary mandates, keep execution behind an authorized approver or narrow limits and log every decision."},{"question":"What about claims of extra returns or tax savings?","answer":"Measure them on your own books against a control before relying on them; no source on this page publishes an independent figure for extra returns or tax savings. What you can measure directly is the share of portfolios outside their bands, the time to resolve a breach and the completeness of the record from trigger to trade."}],"related":["portfolio-reporting-and-commentary","suitability-assessment-assistant","next-best-action-for-advisors","goal-based-financial-planning-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Drift Monitoring and Rebalancing) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: added MiFID II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; softened the unsourced claim about calendar reviews; added the Article 5 and Article 50 basis; corrected the MAS paper note; tied the FAQ answers to the SimCorp and Morgan Stanley evidence; described monitors as the inventory does; SimCorp evidence summary no longer says Wealth Vision was built on Azure."},{"date":"2026-09-27","note":"Review fixes: corrected the Article 50 basis (the trigger is direct interaction with clients, not the absence of human review); rewrote the unsourced problem claims about spreadsheets, thin records and quarterly drift as general statements; the FAQ no longer calls rules based rebalancing long established or claims what is newer; adoption stage set to emerging, since the only AI evidence is an announced vendor capability, a commentary tool and an internal record; every proposal now routes to approval in the Blits.ai build; metaDescription attributes the score to Wealth Lens."},{"date":"2026-09-27","note":"Fact checked against sources: the SimCorp post, the BlackRock release, the InvestmentNews article, the Vanguard page, the ESMA statement and the MAS paper were re read and support the page as written; no content changes needed. SimCorp evidence recheck date updated."},{"date":"2026-09-27","note":"Review fixes: rewrote the Article 50 basis so a firm that builds or brands the rationale writer is a provider (tier now context dependent); set autonomy to copilot, since every proposal needs advisor or portfolio manager approval; reworded the FAQ to say BlackRock announced Auto Commentary rather than describing it as live; the Morgan Stanley evidence record's stage is now announced, matching the launch announcement it is sourced from."}],"slug":"portfolio-drift-monitoring-and-rebalancing","url":"https://www.blits.ai/ai-use-cases/portfolio-drift-monitoring-and-rebalancing","benchmarks":[],"indicativeValueResult":{"low":800000,"high":3200000},"evidence":["morgan-stanley-aladdin-auto-commentary","simcorp-wealth-lens-rebalancing","vanguard-digital-advisor"]},{"title":"AI predictive maintenance for freight rail rolling stock","shortTitle":"Rail rolling stock predictive maintenance","seoTitle":"AI predictive maintenance for rail wheels","metaDescription":"AI inspects freight rail wheels for cracks before failure. BNSF analyses over 35 million wayside readings daily and Norfolk Southern built its own AI wheel scanner.","definition":"Machine vision and machine learning that inspect freight railcar wheels, bearings and other running gear as trains pass wayside sensors and camera portals at track speed, learn what a healthy wheel or a healthy reading looks like, and flag the ones that need attention before a crack, an overheating bearing or a worn wheel causes a service failure or a derailment.","aliases":["AI wheel defect detection","wayside detector analytics","railcar condition monitoring","machine vision train inspection","rail predictive maintenance"],"industries":["logistics-and-transportation"],"functions":["operations","field-service"],"patterns":["computer-vision","anomaly-detection","prediction-and-scoring"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"mechanical-and-safety","problem":"A wheel, bearing or brake component on freight rolling stock can start to fail long before\nanyone sees it. Norfolk Southern says wheel defects are among the most serious\nmechanical defects in the industry. Railroads monitor the condition of railcars with sensors\nalong the track: BNSF says it has \"long used advanced technologies to monitor the condition of\nour railcars and tracks,\" and the Association of American Railroads describes trackside sensors\nthat \"capture large amounts of data as trains pass by, including wheel profiles, wheel impact\nloads, bearing acoustics, and component temperatures.\" But Norfolk Southern says the defects its\nWheel Integrity System looks for are \"subtle defects difficult for the human eye to identify\nconsistently.\"\n\nThe Association of American Railroads describes Norfolk Southern's Wheel Integrity System as\ncatching wheel defects early, helping prevent failures and derailments.","problemStats":[{"statement":"The Association of American Railroads says that by analysing large volumes of real time and historical data, AI enabled systems help detect equipment and infrastructure issues early and support predictive maintenance across the US freight rail network.","sourceTitle":"How Freight Rail Uses AI | Safety & Efficiency","sourceUrl":"https://www.aar.org/issue/how-freight-railroads-use-ai-for-safety-efficiency/","year":2026}],"howItWorks":"1. **Sense the wheel or the train.** Thermal sensors, acoustic sensors and high resolution\n   cameras mounted at fixed wayside portals capture temperature, sound and images of every\n   wheel, bearing and railcar that passes, without slowing the train.\n2. **Build a picture of healthy equipment.** Models trained on large volumes of past readings\n   and images learn what a normal wheel profile, a normal bearing temperature curve or a normal\n   surface looks like for that equipment type.\n3. **Flag the deviations.** The system scores each passing wheel or railcar and flags cracks,\n   overheating, unusual wear or other deviations that need a closer look, ranked by urgency.\n4. **Route to the right team.** A flagged defect goes to the monitoring desk or the mechanical\n   team nearest the railcar's next stop, with the image or reading attached, so the car can be\n   pulled from service or repaired before its next long run.\n5. **Confirm and feed back.** A mechanical inspector confirms the defect on the physical car;\n   confirmed and missed cases both feed back into the model so thresholds improve over time.","valueDrivers":["risk-reduction","cost-to-serve","speed"],"kpis":["mttr-reduction","cost-savings","false-positive-reduction"],"indicativeValue":{"referenceOrg":"A mid sized freight railroad running 1 million loaded railcar trips a year","inputs":[{"key":"carTrips","label":"Loaded railcar trips per year","low":1000000,"high":1000000,"unit":"railcar trips per year","note":"The reference railroad."},{"key":"inServiceFailureRate","label":"Share of trips with a wheel, bearing or brake related in service failure or setout","low":0.0005,"high":0.002,"unit":"fraction of trips","note":"Editorial assumption, replace with your own mechanical setout and derailment data."},{"key":"earlyCatchShare","label":"Share of those failures caught early instead of happening in service","low":0.1,"high":0.3,"unit":"fraction of at risk trips","note":"Editorial assumption. BNSF and Norfolk Southern both report scaled, production wheel inspection systems, but neither discloses a share of failures prevented, so this range is not benchmarked to their evidence; replace it with your own catch rate."},{"key":"costPerInServiceFailure","label":"Cost of an in service wheel, bearing or brake failure (setout, delay, repair, claims)","low":15000,"high":60000,"unit":"USD per event","note":"Editorial assumption; replace with your own cost per mechanical setout or incident, which varies hugely by whether it causes a derailment."}],"formula":"carTrips * inServiceFailureRate * earlyCatchShare * costPerInServiceFailure","currency":"USD","period":"per year","resultLabel":"In service failure and derailment cost avoided","caveat":"Gross avoided cost only, built entirely on editorial assumptions because neither named deployment on this page discloses a catch rate or a percentage outcome. It leaves out the cost of the sensors, cameras and models themselves, false positive handling, and any effect on dwell time or car cycle time."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The sensing hardware (thermal, acoustic, high resolution camera portals) is a capital project in its own right; the AI work is training reliable models per equipment type, keeping false positive rates low enough that mechanical teams trust the flags, and routing a flagged railcar to a stop where it can actually be inspected or repaired.","dataPrerequisites":["Historical wayside detector readings and images linked to confirmed defects and failures","Railcar and wheel maintenance history, by equipment type","A network map of where inspection and repair capacity exists, to route flags usefully"],"integrations":["Wayside detector and camera portal network","Mechanical and maintenance management system for work orders","Car scheduling or yard system, to know where a flagged car will next be reachable","Monitoring desk alerting and dashboards"]},"implementation":{"steps":[{"title":"Start with one defect type and one corridor","detail":"Pick a high value defect type (wheel cracks, bearing overheating) with existing sensor coverage on a busy corridor, so there is enough data and enough traffic to prove the model quickly."},{"title":"Get the sensor data trusted before trusting the model","detail":"Many first flags turn out to be a dirty lens or a miscalibrated sensor, not a real defect. Build in sensor health checks before tuning detection thresholds."},{"title":"Route flags to where action is possible","detail":"A flagged car is only useful information if it reaches a point on the network with the inspection or repair capacity to act on it before the car's next long run."},{"title":"Keep a person on every removal decision","detail":"A model flag advises; a qualified inspector confirms the physical defect and decides whether the car is pulled from service, in line with the railroad's mechanical rules."},{"title":"Track false positives as closely as catches","detail":"A high false positive rate burns the trust of mechanical teams faster than a missed defect does. Report both the confirmed catch rate and the false positive rate every month."},{"title":"Expand by defect type, then by corridor","detail":"Reuse the sensing and model pattern for the next defect type or corridor once the first one has a track record of confirmed catches and a manageable false positive rate."}],"guardrails":["A qualified mechanical inspector confirms every flagged defect before a car is pulled or repaired","Flags feed maintenance planning; they do not directly command a train to stop or slow","Sensor and camera health checks, so faulty hardware is not read as a defective railcar","Model changes tested against a held out set of confirmed past defects before release"],"humanInTheLoop":"Monitoring desk staff and mechanical inspectors review every flagged wheel or railcar and decide whether it is pulled from service or scheduled for repair. Reliability engineers review confirmed catches and any missed failures on a regular cycle to tune detection thresholds.","kpisToInstrument":["Confirmed defect catches per period, and the estimated cost of the in service failure avoided","False positive rate on flagged wheels and railcars","Time from a flag to a confirmed inspection","In service wheel, bearing and brake related setouts and derailments, before and after coverage"],"failureModes":[{"title":"Alert fatigue at the monitoring desk","detail":"Too many low confidence flags and staff start clearing them without a real look. Report the confirmation rate and tune thresholds until flags are worth opening."},{"title":"A flag nobody can act on","detail":"A defect is flagged on a railcar that has already left the network segment with inspection capacity. Route flags to where the car will next be reachable, not just to a dashboard."},{"title":"Sensor drift read as a fleet problem","detail":"A miscalibrated camera or thermal sensor produces a wave of false flags that looks like a real fleet issue. Monitor sensor health separately from defect detection."},{"title":"Treating an advisory flag as a control action","detail":"A flag is wired directly into an automatic stop or slow order without a human check. Keep the model advisory to mechanical teams unless the safety case and conformity assessment for an automated control action have been done separately."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"A system that flags a wheel or railcar for a qualified inspector to confirm is advisory and usually minimal risk. Under Article 6(1) it is high risk when both conditions hold: the same detection logic is built into a safety component of rolling stock or track equipment (or is itself such a product) covered by Directive (EU) 2016/797 on the interoperability of the rail system, which sits in Annex I Section B, for example if a flag were wired to trigger an automatic stop or speed restriction without a human check, and that directive requires a third party conformity assessment of the product. Under Article 2(2), as amended by Regulation (EU) 2026/1744, a high risk system of that kind is not subject to the full AI Act: only Article 6(1), Article 60a and Articles 102 to 112 apply directly, and Articles 57, 58 and 59 apply only so far as the high risk requirements have been integrated into the interoperability directive. The substantive high risk requirements reach the system through that directive instead, which Article 106 of the AI Act amends to require rail delegated and implementing acts to take those requirements into account."},"regulations":["eu-ai-act","nis2","iso-42001"],"guidance":[{"title":"Article 2, scope","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/2/","note":"Article 2(2), as amended by Regulation (EU) 2026/1744, says that for high risk AI systems under Article 6(1) related to products covered by the Annex I Section B laws, such as rail interoperability, only Article 6(1), Article 60a and Articles 102 to 112 apply directly; Articles 57, 58 and 59 apply only so far as the high risk requirements have been integrated into that Union harmonisation legislation."},{"title":"Annex I, list of Union harmonisation legislation","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/1/","note":"Section B, point 17, lists Directive (EU) 2016/797 on the interoperability of the rail system as the product law behind Article 6(1) for rail equipment."},{"title":"Article 6, classification rules for high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/6/","note":"Sets the two conditions in paragraph 1: (a) the AI system is a safety component of a product, or is itself a product, covered by the Union harmonisation legislation listed in Annex I, such as rail interoperability; and (b) that product or AI system is required to undergo a third party conformity assessment under that legislation with a view to being placed on the market or put into service."},{"title":"Article 106, amendment to Directive (EU) 2016/797","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/106/","note":"Adds a paragraph to Article 5 of the rail interoperability directive requiring that, when the Commission adopts delegated and implementing acts on rail safety components, it takes into account the AI Act's Chapter III, Section 2 requirements for high risk AI systems. This is how the AI Act's substantive requirements reach rail equipment instead of applying directly."},{"title":"Digital Omnibus on AI, Regulation (EU) 2026/1744","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/ai-act-explorer/digital-omnibus/","note":"The regulation of 8 July 2026 that amended Article 2(2) to limit the AI Act's direct application to products covered by Annex I Section B, including rail interoperability, to Article 6(1), Article 60a and Articles 102 to 112, with Articles 57 to 59 applying only so far as integrated into the sectoral law."}],"controls":["Model and sensor inventory with an owner per detection model and the equipment it covers","Documented separation between advisory detection models and certified train control and protection systems","Written record of every confirmed catch and every known missed defect","Cybersecurity controls on data flows from wayside sensors and camera portals, in line with NIS2 where it applies"],"incidents":[]},"blitsAi":{"howToBuild":"The wayside sensor, camera and detection models are specialist rail engineering systems that\nstay outside Blits.ai. Blits.ai adds the layer where mechanical and monitoring desk staff act\non the flags. An **AI agent** with a **SQL knowledge base** over a database of flagged\ndefects, railcar history and repair records lets staff ask which railcars are flagged, where\nthey are, and what the history on that wheel or car looks like, in plain language.\n\n**Agentic workflows** with **human in the loop approval** turn a confirmed high urgency flag\ninto a work order through a **custom function** that calls the mechanical or maintenance\nmanagement system's API, once an inspector approves it; lower urgency flags route to a queue\nfor the next scheduled inspection instead of paging someone immediately. Staff reach the agent\nthrough the **REST or WebSocket API** wired into internal tools, or through **Microsoft\nTeams**. A scheduled **agentic task** queries the SQL knowledge base of flags and\nconfirmations and alerts the monitoring desk when volumes cross a threshold, and **monitors**\nrun scheduled health checks on the agent itself so a broken integration is caught early. Every\napproval leaves a full **audit trail** in the agentic workflow's **run history**, and staff\ncan ask the agent for confirmed catch trends straight from the same SQL knowledge base instead\nof a separate chart. The platform is model agnostic, with EU and UAE data residency."},"faq":[{"question":"What results have freight railroads reported from AI wheel and rolling stock inspection?","answer":"BNSF says its AI algorithms sift through more than 35 million wayside detector readings a day and its machine vision processes over 2 million wheel images daily, across a network that monitors more than 1.5 million wheels in motion. Norfolk Southern built its own, standalone Wheel Integrity System with its in house data science and AI team; it follows the railroad's earlier, separate Digital Train Inspection portals, which had already identified and removed from service over 50 wheels with issues since January 2025. The new system pinpointed a vendor casting flaw, and Norfolk Southern says that finding, coupled with its root cause investigation, triggered an industry recall. Neither railroad has published a percentage reduction in failures or derailments from these systems."},{"question":"Does this replace wayside detectors and mechanical inspectors?","answer":"No. It adds AI analysis on top of wayside sensors and camera portals that many railroads already operate, and a qualified mechanical inspector still confirms every flagged defect before a railcar is pulled from service or repaired."},{"question":"Is this the same as predictive maintenance for industrial equipment?","answer":"The pattern, anomaly detection on sensor data, is the same, but freight rail's own wayside detector and camera portal network, its safety regime and its car routing constraints (a flagged car has to be reachable at a point with repair capacity) make it different enough in practice to plan separately from an industrial or energy plant's predictive maintenance programme."},{"question":"What should stay with a qualified inspector?","answer":"The final call on whether a flagged defect is real and whether a railcar is pulled from service or repaired. Neither BNSF nor Norfolk Southern discloses how its AI flags are wired into day to day operations, so treat the recommendation, not a documented fact about either deployment: keep the AI flags advisory to mechanical and monitoring desk staff, and do not wire them into a certified train control system, unless a separate safety case and conformity assessment for an automated control action have been completed."}],"related":["industrial-asset-predictive-maintenance"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by the editor after the Opus targeted check's remaining text fixes."},{"date":"2026-09-28","note":"Second editor pass: added Article 6, Article 106 and the Digital Omnibus on AI (Regulation (EU) 2026/1744) as guidance, since the risk basis named them without a citation; removed the unsourced \"rolling stock runs for decades\" claim and the \"especially on a train moving at track speed\" and hazardous cargo clauses from the problem text, since no source supports them; rewrote the wayside detector sentence to rest only on the BNSF quote and AAR's trackside sensor quote, without the \"long used\" claim AAR does not make about sensors."},{"date":"2026-09-28","note":"Editor pass after adversarial review: corrected the EU AI Act Article 2(2) basis and guidance note to the amended text (Article 6(1), Article 60a and Articles 102 to 112, with the Articles 57 to 59 qualifier, per Regulation (EU) 2026/1744) and added the third party conformity assessment condition to Article 6(1); removed unsourced claims about conventional hot bearing detectors, decades long detector use and human sampling limits from the problem and FAQ text, and rephrased the \"not wired into a certified train control system\" claim as a design recommendation; corrected the blitsAi paragraph to use agentic workflow run history and audit trail instead of a false SQL knowledge base dashboard widget claim, and named the REST or WebSocket API instead of \"internal tools\" as a Blits.ai channel; attributed the Wheel Integrity System recall to the finding plus Norfolk Southern's root cause investigation; softened \"large\" to \"mid sized\" for the reference railroad; tightened the metaDescription's BNSF figure to \"over 35 million\"; linked the industrial and energy asset predictive maintenance page in related."},{"date":"2026-09-28","note":"First version. Covers AI wheel and rolling stock inspection for freight rail, sourced from BNSF and Norfolk Southern, both fetched and quote checked with usecases:source; the Association of American Railroads fact sheet is used only as a problem statistic, not as deployment evidence."}],"slug":"freight-rail-rolling-stock-predictive-maintenance","url":"https://www.blits.ai/ai-use-cases/freight-rail-rolling-stock-predictive-maintenance","benchmarks":[],"indicativeValueResult":{"low":750000,"high":36000000},"evidence":["bnsf-ai-wheel-and-wayside-inspection","norfolk-southern-wheel-integrity-system"]},{"title":"AI predictive maintenance for industrial and energy assets","shortTitle":"Industrial predictive maintenance","seoTitle":"AI predictive maintenance for industrial assets","metaDescription":"C3 AI said in 2022 that Shell monitors over 10,000 pieces of equipment with it; AVEVA says one Duke Energy catch avoided $34 million in cost.","definition":"Machine learning that learns the normal behaviour of industrial and energy equipment from sensor and process data, flags early signs of degradation weeks or months before a failure, and turns them into prioritised maintenance work, so plants and utilities plan repairs instead of reacting to breakdowns.","aliases":["predictive maintenance for industrial equipment","equipment failure prediction","condition based maintenance with AI","asset performance management","remote monitoring and diagnostics centre"],"industries":["energy-and-utilities","manufacturing"],"functions":["operations","field-service"],"patterns":["anomaly-detection","prediction-and-scoring"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"asset-management","problem":"Plants and power stations run on pumps, compressors, valves, turbines, motors and conveyors that\nwear out. Many are maintained on a fixed calendar or run until they fail. Calendar maintenance\nreplaces parts that still had life in them; run to failure means an unplanned stop, emergency\nparts at premium prices, lost production and, in energy, safety and environmental risk.\n\nThe signals of an approaching failure are usually there: a bearing runs slightly warmer, vibration\ncreeps up, a valve takes longer to close. But a large site has thousands of sensors, and the few\nexperienced engineers who can read those patterns cannot watch all of them. Some companies also\nexpect to lose that knowledge: Georgia-Pacific said many of its site experts were \"retiring soon\"\n(AWS). Predictive maintenance lets models watch every asset continuously and send\nthe experts only the cases that need their judgment.","problemStats":[],"howItWorks":"1. **Collect the signals.** Temperature, vibration, pressure, flow, current and process data\n   stream from the control systems and historians into one data platform.\n2. **Learn normal behaviour.** For each asset, a model learns how its readings relate to each\n   other under normal operation, or is trained on labelled past failures where they exist.\n3. **Detect and predict.** The model flags deviations early and, where history allows, estimates\n   the likely failure mode and the time left, weeks or months ahead.\n4. **Triage centrally.** Analysts in a central monitoring and diagnostics centre review early\n   warnings, set aside false alerts and confirm real issues with the site. At the time of the AVEVA\n   story, Duke Energy ran such a centre, with five analysts, for over 87% of its generating fleet\n   (AVEVA). At most Shell assets, a remote engineer vets each alert before it goes to the asset\n   engineers (Shell).\n5. **Plan the work.** Confirmed issues become prioritised work orders with the parts and the\n   window for the repair, planned into the next outage instead of forcing an unplanned one.\n6. **Feed back.** What the technicians find on the equipment is recorded against the alert, so the\n   models and thresholds improve.","valueDrivers":["risk-reduction","cost-to-serve","speed","employee-productivity"],"kpis":["cost-savings","cost-reduction","mttr-reduction","detection-rate-improvement","false-positive-reduction"],"indicativeValue":{"referenceOrg":"A process plant or power station with 150 critical rotating and process assets","inputs":[{"key":"downtimeHours","label":"Unplanned downtime hours per year on critical assets","low":100,"high":150,"unit":"hours per year","note":"Editorial assumption, replace with your own downtime records."},{"key":"costPerHour","label":"Cost of one hour of unplanned downtime","low":20000,"high":60000,"unit":"USD per hour","note":"Editorial assumption covering lost production, emergency repair and restart. Replace with your own figure."},{"key":"avoidedShare","label":"Share of unplanned downtime avoided or converted into planned work","low":0.1,"high":0.3,"unit":"fraction of downtime hours","note":"Editorial assumption. The evidence on this page reports one early catch at Duke Energy that AVEVA says avoided more than 34 million US dollars in cost for that single event (AVEVA), but no fleet wide downtime reduction, so the range stays conservative."}],"formula":"downtimeHours * costPerHour * avoidedShare","currency":"USD","period":"per year","resultLabel":"Unplanned downtime cost avoided","caveat":"Counts avoided downtime only. It leaves out sensors, data platform and analyst costs, the cost of the planned repairs that replace breakdowns, savings from fewer unnecessary calendar maintenance tasks, and the safety and environmental value of avoided failures."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The models are well understood; the hard parts are clean sensor data from many different control systems, too few recorded failures to learn from, and changing the maintenance process so that alerts actually become planned work.","dataPrerequisites":["Historian or control system data per asset at a useful sampling rate","An asset register with criticality, failure modes and maintenance history","Past failures and work orders linked to the sensor data where possible","Engineering knowledge of normal operating ranges per asset type"],"integrations":["Plant historian and control systems (SCADA, DCS)","Data platform for streaming and model training","Enterprise asset management or computerised maintenance management system for work orders","Alerting to the monitoring centre and site teams"]},"implementation":{"steps":[{"title":"Start with critical assets and known failure modes","detail":"Rank assets by the cost and risk of failure and pick one asset class with sensor coverage and a few documented failures, such as large pumps or compressors."},{"title":"Get the data flowing and trusted","detail":"Connect the historian, fix tag names and units, and agree with the site which readings are reliable. Many first alerts turn out to be sensor faults, not machine faults."},{"title":"Stand up a central monitoring team","detail":"Put a small group of experienced engineers between the models and the sites to triage alerts, so the sites only see confirmed issues."},{"title":"Wire alerts into maintenance planning","detail":"Create work orders in the maintenance system from confirmed alerts, with a priority and a repair window, and record what the technician found."},{"title":"Scale by asset class, then by site","detail":"Reuse models and templates across identical equipment, track avoided failures with a written case per catch, and widen coverage one asset class at a time. Shell set itself a target of 10,000 monitored pieces of critical equipment for 2021 and reported reaching it."}],"guardrails":["Alerts advise; protection systems and trips stay in the certified control and safety systems","A human engineer confirms every alert before a work order or shutdown is raised","Model changes tested against past data before release, with version history","Sensor health checks so that faulty instruments are not read as failing machines"],"humanInTheLoop":"Monitoring and diagnostics engineers triage every alert and decide whether it becomes work. Site maintenance planners choose when to repair, and reliability engineers review missed failures and false alarms each month to tune the models.","kpisToInstrument":["Unplanned downtime hours on monitored assets, before and after","Documented early catches and their estimated avoided cost","Share of alerts confirmed as real issues","Lead time between first alert and failure or repair","Failures on monitored assets that the models missed"],"failureModes":[{"title":"Alert fatigue","detail":"Too many alerts with too little context and sites stop reacting. Triage centrally and report the confirmation rate."},{"title":"No link to the maintenance process","detail":"Alerts land in a dashboard nobody plans from. Create work orders from confirmed alerts in the maintenance system."},{"title":"Too few failures to learn from","detail":"Critical assets rarely fail, so supervised models lack examples. Use anomaly detection on normal behaviour and engineering rules alongside."},{"title":"Savings nobody believes","detail":"Claimed savings that cannot be traced to a specific catch lose credibility. Write up each early catch with what would have happened."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"A system that advises engineers on the condition of equipment is usually minimal risk. Annex III point 2 lists AI systems intended as safety components in the management and operation of critical digital infrastructure, road traffic and the supply of water, gas, heating or electricity; if predictive maintenance acts on protection or control in a utility network, it can become high risk. Article 6(1) can also apply when the AI is a safety component of machinery or another product covered by Annex I legislation and that product must undergo a third party conformity assessment."},"regulations":["eu-ai-act","nis2","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 2 covers safety components in the management and operation of critical infrastructure, including the supply of water, gas, heating and electricity."},{"title":"Article 6, classification rules for high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/6/","note":"Explains when an AI system that is a safety component of a product under Annex I legislation, such as machinery, is high risk."},{"title":"AI Risk Management Framework","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/itl/ai-risk-management-framework","note":"Voluntary framework to map, measure and manage the risks of AI systems, useful for documenting model limits and monitoring."}],"controls":["Asset and model inventory with an owner per model and the assets it covers","Documented separation between advisory models and certified protection systems","Written record of every confirmed catch and every missed failure","Cybersecurity controls on data flows from operational technology, in line with NIS2 where it applies"],"incidents":[]},"blitsAi":{"howToBuild":"The anomaly and failure models run on the manufacturer's or utility's asset analytics platform.\nBlits.ai adds the layer where people act on them. An **AI agent** with a **SQL knowledge base**\nover a PostgreSQL or SQLite copy of alerts, asset data and work orders lets engineers ask which\nassets are trending toward failure and why, and a **knowledge base** with hybrid retrieval over manuals and past failure\nreports explains the likely cause and repair.\n\n**Agentic tasks** run condition triggered checks (\"when a confirmed alert is older than two days\nwithout a work order, draft one\") and **human in the loop confirmation**, with the threshold set\nso that it covers work orders and escalations, lets an engineer approve or reject each one before\n**custom functions** create it through the maintenance system's API. Engineers use it in **Microsoft Teams**, **monitors** check its answers on a\nschedule, and the platform is model agnostic with EU and UAE data residency."},"faq":[{"question":"What results do companies report from predictive maintenance?","answer":"AVEVA reports that a single early catch by Duke Energy's monitoring and diagnostics centre in 2016 avoided more than 34 million US dollars in cost had the problem gone undetected; that is one event, not a yearly saving. AWS reports that Georgia-Pacific can predict failure of selected assets 60 to 90 days ahead, without an outcome figure. Shell writes that at one Dutch refinery its models flagged 65 control valves in need of repair that traditional methods would have missed. Published fleet wide downtime figures are rare, so measure your own avoided failures case by case."},{"question":"How many assets can one programme cover?","answer":"In a March 2022 press release, C3 AI said Shell's predictive maintenance programme on its platform monitors more than 10,000 pieces of equipment; Shell, quoted in the release, called monitoring 10,000 pieces of critical equipment a target set for 2021 and achieved. Coverage grows asset class by asset class, reusing models across identical equipment."},{"question":"Is predictive maintenance high risk under the EU AI Act?","answer":"Usually not while it advises engineers. It can become high risk if it acts as a safety component in the supply of water, gas, heating or electricity (Annex III point 2), or a safety component of machinery covered by Annex I legislation that is subject to third party conformity assessment (Article 6(1))."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the energy and manufacturing scope with Shell, Duke Energy and Georgia-Pacific evidence checked against the sources. The researcher's own revision left out the Georgia-Pacific paper tear metric (process optimization, not maintenance), attributed the Shell figure to C3 AI, and called the Duke Energy figure an estimated avoided cost for one event. Review pass the same day separated the Duke Energy reference from the generic triage practice, attributed the retiring experts point to Georgia-Pacific, and added the third party conformity assessment condition to the EU AI Act answer. Editor pass added Shell's own TechXplorer Digest article to the Shell evidence (now grade B) and used it for the triage step and the results answer."},{"date":"2026-09-27","note":"Second editor pass restored Shell's scope qualifier (a remote engineer vets alerts first \"for most assets\"), dated the Duke Energy centre figures to the AVEVA story, dropped the word \"estimated\" that AVEVA never uses, and softened the unsourced claim that most assets run on calendar or run to failure maintenance."}],"slug":"industrial-asset-predictive-maintenance","url":"https://www.blits.ai/ai-use-cases/industrial-asset-predictive-maintenance","benchmarks":[{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":34000000,"min":34000000,"max":34000000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"duke-energy-monitoring-and-diagnostics-center","pooled":true}]}],"indicativeValueResult":{"low":200000,"high":2700000},"evidence":["duke-energy-monitoring-and-diagnostics-center","georgia-pacific-predictive-asset-analytics","shell-c3-ai-predictive-maintenance"]},{"title":"AI prioritization of radiology and imaging worklists","shortTitle":"Radiology worklist triage","seoTitle":"AI radiology worklist triage software","metaDescription":"AI flags urgent CT and MRI findings and reorders the radiologist worklist. A regional stroke center cut transfer time 44% with a program that included Viz.ai.","definition":"An AI system that analyzes a medical image immediately after a scan, flags time sensitive findings such as a brain bleed, a stroke causing large vessel occlusion or a pulmonary embolism, and reorders the radiologist's worklist and notifies the care team so the most urgent cases are read and acted on first, while a radiologist confirms every finding before it changes a patient's treatment.","aliases":["AI radiology triage","imaging worklist prioritization","critical findings AI","AI stroke detection"],"industries":["healthcare"],"functions":["operations"],"patterns":["computer-vision","classification-and-routing","anomaly-detection"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"mainstream","segment":"emergency and inpatient imaging","problem":"Within the same priority class, a radiology worklist is normally read in roughly the order scans\narrive. A scan that shows a brain bleed or a blood clot blocking a major vessel can sit behind\nseveral routine studies at the same priority level before a radiologist opens it, and for time\nsensitive conditions every extra minute has a cost: in acute ischemic stroke, treatment delay is\ndirectly linked to worse outcomes. The problem compounds at\nregional or community hospitals, where a patient needing specialist treatment must first be\nidentified, then transferred to a comprehensive center, a handoff that traditionally depends on a\nradiologist's read, a phone call to a specialist, and a manual transfer process.","problemStats":[],"howItWorks":"1. **Scan and analyze.** As soon as a CT or MRI is acquired, an AI model, cleared for that specific\n   use, analyzes it for the patterns it is trained to detect, such as intracranial hemorrhage, large\n   vessel occlusion or pulmonary embolism.\n2. **Flag and notify.** A positive finding pushes the case to the top of the radiologist's worklist\n   and sends a mobile alert to the on call specialist and care team, often within seconds of the scan\n   completing.\n3. **Confirm and act.** The radiologist reviews the flagged images and confirms or rules out the\n   finding; the specialist team begins the treatment pathway, such as a thrombectomy, transfer or\n   surgery, based on the confirmed read, not the AI flag alone.\n4. **Coordinate transfer.** At a regional hospital without full stroke or trauma capability, the same\n   alert can trigger a standardized transfer protocol and direct communication with a comprehensive\n   center.\n5. **Audit and monitor.** Every flagged and missed case feeds back into ongoing monitoring of\n   sensitivity, specificity and turnaround time by pathology, site and shift.","valueDrivers":["speed","risk-reduction","employee-productivity"],"kpis":["processing-time-reduction","response-time-reduction","detection-rate-improvement"],"indicativeValue":{"referenceOrg":"A 400 bed hospital reading 40,000 CT and MRI studies a year for time sensitive pathologies","inputs":[{"key":"studies","label":"Time sensitive CT and MRI studies read per year","low":40000,"high":40000,"unit":"studies per year","note":"Editorial assumption for a 400 bed hospital's time sensitive imaging volume; replace with your own case mix."},{"key":"positiveShare","label":"Share of studies with a time sensitive positive finding","low":0.02,"high":0.05,"unit":"fraction of studies","note":"Editorial assumption across intracranial hemorrhage, large vessel occlusion and pulmonary embolism screening; replace with your own case mix."},{"key":"minutesSaved","label":"Minutes of care team activation time saved per positive case","low":10,"high":30,"unit":"minutes per case","note":"Below the figure in Viz.ai's release (describing the Adventist Health + Rideout deployment, reporting care team notification time falling from 45 minutes to 7 minutes, a 38 minute reduction, for large vessel occlusion stroke at one hospital), because this input averages across intracranial hemorrhage, large vessel occlusion and pulmonary embolism, and the 38 minute figure covers only large vessel occlusion. This figure is not a recorded metric or a computed benchmark; the evidence record does not report it as a metric because it measures a narrower step, care team notification, than the end to end transfer time the taxonomy's Cycle time reduction KPI defines, and no KPI in this taxonomy covers that step on its own."},{"key":"valuePerMinute","label":"Value of a minute of faster time sensitive treatment","low":50,"high":120,"unit":"USD per minute","note":"Editorial assumption combining avoided length of stay, disability and readmission cost for time sensitive conditions; replace with your own health economic estimate."}],"formula":"studies * positiveShare * minutesSaved * valuePerMinute","currency":"USD","period":"per year","resultLabel":"Annual value of faster time sensitive treatment","caveat":"A rough proxy for the value of speed only. It leaves out the cost of the software and its integration, the value of pathologies not modeled here, and the fact that faster notification does not guarantee a faster or better clinical outcome for every patient."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Analyzing the image is a small part of the work. The device must be cleared or CE marked for the exact indication, scanner types and patient population in use, integrated with PACS and the hospital's paging and care team activation system, and validated on local data, all before it changes a single worklist.","dataPrerequisites":["A regulatory clearance (FDA clearance or CE mark) that covers the pathology, patient population and scanner protocols in use","PACS and imaging archive integration for the modalities in scope","A defined care team activation and paging workflow to wire the alert into"],"integrations":["Picture archiving and communication system (PACS)","Radiology information system and worklist","Clinician paging and care team activation system","Electronic health record, for the confirmed finding and downstream care pathway"]},"implementation":{"steps":[{"title":"Start with one time critical pathology with a clear clinical owner","detail":"Pick a pathology such as large vessel occlusion stroke or intracranial hemorrhage where a stroke or trauma lead can own the rollout, rather than deploying every available module at once."},{"title":"Confirm the clearance matches your population before go live","detail":"Check that the FDA clearance or CE mark covers your scanner models, contrast protocols and patient population; performance on a mismatched population can be unreliable and go unnoticed."},{"title":"Wire the alert into the paging system specialists already use","detail":"Route the notification through the existing on call and care team activation workflow, not a new inbox nobody checks at 3am."},{"title":"Set a turnaround time target per pathology and measure it before and after","detail":"Track time from scan completion to notification and to radiologist confirmation, not only sensitivity and specificity."},{"title":"Add pathologies and sites one at a time, each with its own validation","detail":"Treat every new pathology or site as a new rollout with its own local validation, not an automatic extension of the first one."}],"guardrails":["Every flagged finding is confirmed by a radiologist before it changes a treatment plan; the AI reorders the queue, it does not diagnose","The system only runs within its cleared indications, scanner types and patient population","A defined fallback (acuity based or FIFO ordering) applies when the AI is unavailable or a study fails triage"],"humanInTheLoop":"Radiologists confirm every AI flagged finding before it drives a clinical decision, and review a sample of unflagged cases to catch missed findings. A clinical safety lead owns the pathology's performance against its cleared claims and approves any expansion to a new site or population.","kpisToInstrument":["Time from scan completion to critical finding notification, by pathology and shift","Radiologist report turnaround time for flagged versus unflagged cases","False positive and false negative rate against a sampled radiologist read"],"failureModes":[{"title":"Alert fatigue from false positives","detail":"Too many low value alerts and clinicians start deprioritizing all of them, including true positives. Track and act on the false positive rate per pathology and site, not only overall sensitivity."},{"title":"Silent underperformance outside the cleared population","detail":"The model performs unreliably on a scanner protocol or patient group it was not validated on, and nobody notices because the model gives no signal that it is out of its depth. Validate on local data before go live and monitor for performance drift by site."}]},"risk":{"euAiAct":{"tier":"high","basis":"Article 6(1) and Annex I: software that analyzes a medical image to detect or prioritize a disease finding is itself, or is a safety component of, a device in scope of the EU Medical Device Regulation, and typically needs a notified body conformity assessment as software as a medical device (the FDA's AI Enabled Medical Device List shows US market authorization for devices in this category, listing authorized stroke triage devices from Viz.ai and Aidoc's BriefCase triage devices), which makes it high risk under the EU AI Act regardless of Annex III. The radiologist's own diagnostic read stays a human decision; the AI narrows and reorders the queue. Annex I high risk classification under Article 6(1) applies from 2 August 2028 (Article 113(c)); until then, Article 4 (AI literacy obligations) and Article 5 (prohibited practices), which bind the hospital as a deployer, already apply."},"regulations":["eu-ai-act","gdpr","hipaa","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Regulation (EU) 2017/745 on medical devices","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2017/745/oj","note":"Does not mention AI by name. Software that provides information used to take decisions with diagnostic or therapeutic purposes, such as a finding used to prioritize or route a patient, is classified under Annex VIII Rule 11, usually as class IIa or higher, which requires a notified body conformity assessment before CE marking."},{"title":"Article 6: Classification Rules for High-Risk AI Systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/6/","note":"A safety component of, or a product that is itself, a CE marked medical device under EU harmonisation legislation requiring third party conformity assessment is high risk under the EU AI Act. Per Article 113(c), this Annex I route applies from 2 August 2028, later than the 2 December 2027 date for the Annex III use cases."},{"title":"List of Artificial Intelligence-Enabled Medical Devices","issuer":"US Food and Drug Administration","region":"north-america","url":"https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices","note":"FDA's list of AI enabled medical devices authorized for marketing in the United States. The list names each device and its company, for example \"Viz.ai, Inc.\", along with a decision date and submission number; the cleared or granted indications are in the linked 510(k) or De Novo records. It lists authorized stroke triage devices from Viz.ai and Aidoc's BriefCase triage devices."}],"controls":["Maintain the regulatory clearance and intended use statement for each detection module, and never enable a pathology it is not cleared for","Radiologist confirms every flagged finding before it changes a treatment plan","Track sensitivity, specificity and turnaround time by pathology, site and shift against the cleared performance claims","Defined fallback ordering when the AI is unavailable or a study fails triage, so the worklist never silently reverts to an unmanaged queue"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai does not read medical images; the FDA cleared or CE marked detection model does that, and\nstays outside Blits.ai's scope. What Blits.ai adds is the coordination layer around the alert: an\n**agentic workflow**, triggered through its own API token (or the REST API channel), receives the\nfinding from the imaging system. An **AI agent** turns it into a plain language page to the right\non call specialist, with the patient's context pulled from a **SQL knowledge base**, and a\n**custom function** makes the outbound call to the hospital's paging or care team activation\nsystem. For a regional hospital without full stroke capability, the same workflow, with **human in\nthe loop approval** before any transfer request goes out, encodes the standardized transfer\nprotocol to a comprehensive center.\n\n**Monitors** run scheduled health checks against the agent that writes the specialist facing\npage, with email or webhook alerts when it fails, and **analytics** on the workflow's run\nhistory track time from finding to notification. **Guardrails** apply LLM based content\nchecks to keep the specialist facing page in plain, unambiguous language; they reduce the risk of\na garbled or misleading message but cannot themselves guarantee that only a confirmed finding is\never relayed. That guarantee comes from the **human in the loop approval** step, not from\nguardrails, which is why no transfer request goes out before a radiologist has acted."},"faq":[{"question":"What does AI radiology triage actually do?","answer":"It analyzes an image right after the scan, flags time sensitive findings it is cleared to detect, and reorders the radiologist's worklist and alerts the care team so urgent cases are read first. It does not replace the radiologist's diagnostic read."},{"question":"Does the AI diagnose the patient?","answer":"No. The AI flags a likely finding and reprioritizes the queue; a radiologist confirms or rules out the finding before any treatment decision is made."},{"question":"Is AI radiology triage regulated as a medical device?","answer":"Generally yes. In the United States these tools typically hold FDA clearance as software as a medical device, and in the EU they are usually CE marked medical devices under Annex VIII Rule 11 of the Medical Device Regulation. Because that classification requires a notified body conformity assessment, they are high risk under the EU AI Act's Annex I route, regardless of whether the specific use appears in Annex III, though that Annex I classification only takes effect on 2 August 2028."},{"question":"How much time does it actually save?","answer":"It depends heavily on the pathology, the baseline workflow and the hospital. According to Viz.ai, a study led by Adventist Health + Rideout's stroke program manager and presented at the 2026 International Stroke Conference found that average door in door out transfer time for large vessel occlusion stroke patients fell by 44%, from 202 to 113 minutes, after a quality improvement program that included the Viz.ai platform, a partnership with a comprehensive stroke center and standardized transfer protocols."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched from Sheba Medical Center's own site and Viz.ai's press release naming Adventist Health + Rideout."},{"date":"2026-09-28","note":"Editorial review: reworded metaDescription and the Monitors claim in blitsAi.howToBuild to match the platform feature inventory, fixed the EU AI Act basis grammar and its Article 4/5 and FDA list wording, softened the FAQ CE marking claim, renamed howItWorks step 5 to Audit and monitor, and clarified two indicativeValue notes."},{"date":"2026-09-28","note":"Review fix round: removed the Adventist Health + Rideout evidence record's second metric (an 84% figure for time from CTA completion to detection, filed under the end to end Cycle time reduction KPI, which double counted the 44% end to end figure) and reworded its verification note and this page's indicativeValue.minutesSaved note to match; corrected the Sheba evidence record's year note to point to the earliest Wayback Machine capture of the source page instead of a wrong date range for neighbouring articles, and pointed archivedUrl at that capture; corrected the FDA guidance note, which wrongly said the list shows device names only, and changed \"cleared\" to \"authorized\" for the Viz.ai and Aidoc stroke triage devices here and in euAiAct.basis; and reattributed the 44% transfer time figure in the FAQ to Viz.ai's account of the study, rather than presenting it as Adventist Health + Rideout's own report."}],"slug":"radiology-worklist-triage","url":"https://www.blits.ai/ai-use-cases/radiology-worklist-triage","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":44,"min":44,"max":44,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"adventist-health-rideout-viz-ai-stroke-transfer","pooled":true}]}],"indicativeValueResult":{"low":400000,"high":7200000},"evidence":["adventist-health-rideout-viz-ai-stroke-transfer","sheba-medical-center-aidoc-triage"]},{"title":"AI product content and catalog enrichment for online retail","shortTitle":"Product content and catalog enrichment","seoTitle":"AI product descriptions and catalog enrichment","metaDescription":"AI drafts product titles, descriptions and attributes at catalog scale. Walmart created or improved over 850 million catalog data points using several LLMs.","definition":"AI that writes and repairs product content at catalog scale: it drafts titles, descriptions and image alt text, and extracts missing attributes such as color, size and material from supplier text and product images, then checks its own output before the content is published to the store and to search engines. A human owns the rules, the quality thresholds and the exceptions.","aliases":["AI product description generator","product attribute extraction","catalog enrichment","AI product listing generation","product information management AI","ecommerce SEO content generation"],"industries":["retail-and-ecommerce","cross-industry"],"functions":["marketing","operations"],"patterns":["content-generation","computer-vision","classification-and-routing"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"An online catalog is only as findable as its data. Shoppers filter by size, color and material,\nsearch engines index titles, descriptions and alt text, and marketplaces such as Amazon use\nattributes like color to index products in their own search. Yet most product records arrive thin or inconsistent: a supplier spreadsheet with a\ncryptic name, a few bullet points and a photo. Retailers with hundreds of thousands or millions\nof items cannot write and check every record by hand, so attributes stay empty, descriptions stay\ncopied from the manufacturer, and items that are in stock are never found.\n\nWalmart describes the stakes plainly: the quality of catalog data affects nearly everything it\ndoes, from helping customers find products to sorting inventory and delivering orders. Small\nsellers feel the same problem from the other side. Writing a good listing takes time they do not\nhave, which is why the large marketplaces now offer to draft it for them from a photo, a few words\nor an existing web page.","problemStats":[],"howItWorks":"1. **Collect the inputs.** Supplier feeds, existing descriptions, product images, the category\n   and the attribute specification for that category (allowed values, units, required fields).\n2. **Extract attributes.** A model reads the text and the images and proposes a value for each\n   attribute in the specification, with the source it used.\n3. **Write the content.** A model drafts the title, description, bullet points and image alt\n   text in the house style, using only the extracted attributes and approved claims, and adds\n   the keywords shoppers use in search.\n4. **Check before publishing.** A second model or rule set checks each value and each draft:\n   does the attribute match the image, is the unit valid, does the description claim anything the\n   data does not support? Values above a set accuracy threshold are published; the rest go to a\n   human reviewer, whose decisions become new training and test data.\n5. **Measure and repeat.** Search conversion, returns for \"not as described\" and reviewer\n   corrections show which categories and attributes need better prompts or a human.","valueDrivers":["revenue-growth","employee-productivity","customer-experience","speed"],"kpis":["interactions-handled","users-served","productivity-gain","quality-score-uplift","accuracy"],"indicativeValue":{"referenceOrg":"An online retailer with 200,000 active product listings","inputs":[{"key":"listings","label":"Active product listings","low":200000,"high":200000,"unit":"listings","note":"The reference retailer."},{"key":"shareTouched","label":"Share of listings written or repaired per year","low":0.3,"high":0.6,"unit":"fraction of listings","note":"Editorial assumption covering new items and repairs of thin or inconsistent records. Replace with your own assortment change rate."},{"key":"minutesSaved","label":"Content work saved per listing","low":10,"high":25,"unit":"minutes per listing","note":"Editorial assumption. A seller quoted by Amazon in May 2025 (in the Amazon evidence record for this page) said listings used to take an hour and that the AI content is now generated in under 15 minutes; that 15 minutes is generation time, not total listing time. This range is more conservative because a retailer's content team already works faster than a small seller and still reviews each item."},{"key":"hourlyCost","label":"Fully loaded cost of a content specialist","low":30,"high":50,"unit":"USD per hour","note":"Editorial assumption, replace with your own cost or agency rate."}],"formula":"listings * shareTouched * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Content production cost avoided","caveat":"Counts only the writing and data entry time saved. It leaves out the running cost of the models, the human review that stays in place, and the revenue effect of better findability, which Etsy reports as 3% more conversions for its sellers (alongside 5% more visits from search engines) from improved alt text, but which depends on the catalog."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Drafting text is easy; getting reliable attributes at catalog scale is the work. It needs an attribute specification per category, a labelled benchmark set, a quality check that decides what is published automatically, and a connection to the product information system.","dataPrerequisites":["An attribute specification per category, with allowed values and units","A human validated sample of products per category to benchmark accuracy","Brand style guide and a list of claims that may and may not be made","Product images and supplier data linked to each item"],"integrations":["Product information management (PIM) or catalog system","Digital asset management for product images","Ecommerce platform or marketplace listing API","Search and analytics for conversion and search performance"]},"implementation":{"steps":[{"title":"Write the specification first","detail":"For each category, list the attributes that matter for search and filters, their allowed values and units, and which ones are safety or compliance relevant. The model is only as good as this list."},{"title":"Build a benchmark before a prompt","detail":"Have specialists label a few hundred items per large category, keep them out of any tuning data, and measure precision and recall per attribute for every model or prompt change."},{"title":"Separate writing from checking","detail":"Use one step to extract and write and another to verify, as Walmart describes. Publish automatically only the attributes whose measured accuracy clears your threshold, and send the rest through a further check or to reviewers."},{"title":"Ground descriptions in the data","detail":"Generate descriptions from the verified attributes and approved claims only, never from the model's general knowledge, so a description cannot promise a feature the item lacks."},{"title":"Put people where the risk is","detail":"Keep human review for safety relevant attributes (allergens, age ratings, electrical ratings), regulated categories and low confidence items, and sample the automatic output every week."},{"title":"Measure on the storefront","detail":"Track search conversion, zero result searches, filter usage and returns for \"not as described\" per category, so enrichment is judged by shoppers, not by word count."}],"guardrails":["Descriptions generated only from verified attributes and an approved claims list","Automatic publication only above a measured accuracy threshold per attribute","Human review for safety, legal and regulated product attributes","Filters that block model refusals, placeholder text and competitor brand names from publication","Versioned prompts and a benchmark run before every change"],"humanInTheLoop":"Content and category specialists own the attribute specifications, label the benchmark sets, review low confidence and safety relevant values, and sample automatically published content. On a marketplace the seller submits each AI draft: Amazon encourages sellers to review drafts before they submit them, and eBay's listing flow asks the seller to review and approve the suggestions.","kpisToInstrument":["Attribute precision and recall per category against the human labelled benchmark","Share of AI output published without edits, and the edit rate by reviewers","Attribute fill rate for the fields shoppers filter on","Search conversion and zero result searches before and after, per category","Returns with the reason \"not as described\""],"failureModes":[{"title":"Confident but wrong attributes","detail":"A model fills a color, size or material that the image contradicts, and the item is returned. Measure accuracy per attribute and only automate what clears the threshold."},{"title":"Invented features","detail":"A fluent description promises a feature the product does not have, which becomes a consumer protection problem. Generate only from verified data and approved claims."},{"title":"Unreviewed output goes live","detail":"In January 2024 Amazon hosted listings whose titles were model refusal messages. Block refusals and placeholder text automatically and keep a human submit step."},{"title":"Thin pages at scale","detail":"Thousands of near identical generated pages can be treated as spam by search engines. Write for shoppers with real product facts, not for keyword variants."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Writing product content and extracting catalog attributes is not an Annex III use and makes no decisions about people. When a retailer uses a third party generator, the use is minimal risk for the retailer: the Article 50(2) duty to mark generated text in a machine readable way falls on the provider of that system. When a retailer builds and operates its own generating system and puts it into service under its own name, it is the provider and must mark the output, unless the exception for systems that only assist standard editing or do not substantially alter the input applies. Article 50(4) covers text published to inform the public on matters of public interest, not product listings. Consumer protection law applies to what the listing says in every case."},"regulations":["eu-ai-act","eu-accessibility-act"],"guidance":[{"title":"Google Search's guidance about AI generated content","issuer":"Google Search Central","region":"global","url":"https://developers.google.com/search/blog/2023/02/google-search-and-ai-content","note":"Google rewards helpful content however it is produced, and treats automation used mainly to manipulate rankings as spam."},{"title":"Spam policies for Google web search, scaled content abuse","issuer":"Google Search Central","region":"global","url":"https://developers.google.com/search/docs/essentials/spam-policies","note":"Lists pages generated at scale with little value for users, including automated transformations such as synonymizing and translating, as scaled content abuse."},{"title":"Final rule banning fake reviews and testimonials","issuer":"Federal Trade Commission","region":"north-america","url":"https://www.ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commission-announces-final-rule-banning-fake-reviews-testimonials","note":"Prohibits fake reviews and testimonials, including AI generated ones; product content generation must not extend to reviews."},{"title":"Unfair Commercial Practices Directive","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/dir/2005/29/oj","note":"Misleading product information is an unfair commercial practice, whoever or whatever wrote it."}],"controls":["A named owner per category for the attribute specification and the approved claims list","Accuracy benchmarks per attribute with thresholds for automatic publication","An audit trail that records which model, prompt version and reviewer produced each published value","Alt text that describes the image for people using screen readers, not only for search engines","A takedown route for content reported as wrong by customers or sellers"],"incidents":[{"title":"AI refusal messages published as Amazon product titles","url":"https://oecd.ai/en/incidents/2024-01-12-c37b","note":"Amazon hosted listings whose titles were model refusal messages (\"I'm sorry but I cannot fulfill this request\"); Amazon removed them and said it was improving its review systems. The date comes from the incident ID (2024-01-12). It shows what happens when generated content is published without review."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** that runs on a schedule or through an API token: an\n**AI agent** with **structured output configuration** reads the product record and the category's\nattribute specification from a **knowledge base** (supplier sheets and style guides ingested from\nXLSX, CSV, PDF and DOCX), proposes attribute values and drafts the title, description and alt\ntext. **Custom functions** read from and write back to the product information system or\necommerce platform through REST, and **SQL knowledge bases** can hold the catalog tables the\nagent queries.\n\nA second agent step checks each draft against the specification, and **human in the loop**\nconfirmation asks a content specialist to approve or reject the write back to the catalog before\nit happens. **Guardrails** block unsupported claims and refusal text, **test suites** with LLM\nbased grading run the labelled benchmark on every prompt version, and the run history keeps a\nfull audit trail per run.\n**Machine translation** and multi language support can extend the same content to other\nmarkets, and the platform is model agnostic, so each step can use the model that performs best\non the benchmark."},"faq":[{"question":"Can AI write product descriptions without a human checking them?","answer":"For extracted attributes with measured accuracy, yes, and that is how Walmart describes its catalog work: a second model checks the first, and values for attributes above an accuracy threshold go straight into the catalog. Generated descriptions are a different matter: on Amazon and eBay the seller submits each draft, and Amazon reported in 2025 that sellers accept AI content with little or no edits about 90% of the time."},{"question":"Does AI generated product content hurt SEO?","answer":"Not by itself. Google says it rewards helpful content however it is produced, but treats pages generated at scale mainly to manipulate rankings as spam. Etsy reports that better alt text generated with Gemini increased visits from search engines by 5% and conversions by 3% for its sellers."},{"question":"What is harder, the text or the attributes?","answer":"The attributes. Fluent descriptions are easy to generate; correct size, color, material and safety data across millions of items need a specification, a labelled benchmark and a quality check per attribute."}],"related":["conversational-shopping-assistant","marketing-content-compliance-copilot","personalized-marketing-at-scale"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the editor after a targeted blocker check."},{"date":"2026-09-27","note":"First version, with evidence from Walmart, Amazon, eBay and Etsy checked against the sources. Editor pass corrected the marketplace review wording, the EU AI Act tier (context dependent) and the FAQ answers. Second editor pass dated the Amazon figures to the May 2025 update of its post, dropped Walmart's counterfactual 100 times headcount estimate as a metric, added a Wayback capture for the Etsy case study and aligned the Blits.ai section with the feature inventory. Third editor pass removed the conversion rate uplift KPI from the kpis list, since no evidence file for this page files that KPI and the Etsy record could not support a conversion rate metric."}],"slug":"product-content-and-catalog-enrichment","url":"https://www.blits.ai/ai-use-cases/product-content-and-catalog-enrichment","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":475000000,"min":100000000,"max":850000000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"walmart-generative-ai-product-catalog","pooled":true},{"id":"ebay-magical-listing-generative-ai","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":5450000,"min":900000,"max":10000000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"ebay-magical-listing-generative-ai","pooled":true},{"id":"amazon-generative-ai-listing-tools","pooled":true}]},{"kpi":"quality-score-uplift","label":"Quality score uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40,"min":40,"max":40,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"amazon-generative-ai-listing-tools","pooled":true}]}],"indicativeValueResult":{"low":300000,"high":2500000},"evidence":["amazon-generative-ai-listing-tools","ebay-magical-listing-generative-ai","etsy-gemini-listing-enrichment","walmart-generative-ai-product-catalog"]},{"title":"AI quality and compliance monitoring of every customer interaction","shortTitle":"Call quality and compliance","seoTitle":"AI call quality and compliance monitoring","metaDescription":"AI scores every call and chat against your QA rubric and required disclosures. US bank group Central Bank went from 24 to 167,000 calls checked a quarter.","definition":"Automated quality assurance that transcribes and scores every customer interaction, voice and chat, against the organization's own rubric, checking required disclosures and script adherence, flagging conduct and mis selling risk, and surfacing coaching opportunities, instead of the small sample a human QA team can review.","aliases":["automated QA","AI quality management","conversation intelligence","speech analytics for compliance","interaction analytics"],"industries":["cross-industry","banking","insurance","energy-and-utilities","telecommunications","retail-and-ecommerce"],"functions":["customer-service","regulatory-compliance","operations"],"patterns":["speech-analytics","classification-and-routing","summarization"],"channels":["voice","web-chat","agent-desktop"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Traditional contact centre QA reviews a tiny sample: a few calls per agent per month, scored by\nhand. That sample is too small to find systematic problems, too late to coach while the call is\nremembered, and inconsistent between reviewers. For regulated firms it is also weak evidence:\nwhen a supervisor asks whether required disclosures were given, or whether vulnerable customers\nwere treated fairly, a small sample is not much of an answer.\n\nSpeech analytics and language models make it possible to evaluate every interaction against the\nsame rubric within a day. The value is coverage and consistency: finding the missed disclosure,\nthe mis sold product or the recurring complaint driver, and coaching on patterns rather than\nanecdotes. The risk is treating a model score as a verdict on a person, which is both unfair and,\nin the EU, a high risk use of AI.","problemStats":[],"howItWorks":"1. **Capture and transcribe.** Calls are transcribed after the fact (or in near real time) with\n   speaker separation; chat and email are ingested as text. Card data is redacted.\n2. **Score against the rubric.** Each interaction is checked against configurable criteria:\n   required disclosures, identity checks, script steps, prohibited statements, complaint and\n   vulnerability indicators, and service behaviours.\n3. **Flag for review.** Interactions with likely breaches, complaints or vulnerability signals go\n   to a QA or compliance reviewer with the relevant excerpt, not a bare score.\n4. **Calibrate against humans.** Reviewers regularly score the same interactions as the model;\n   disagreements tune the rubric and prompts.\n5. **Coach on patterns.** Team leaders see recurring gaps per team and topic and coach from real\n   examples; agents can see and dispute their own results.\n6. **Report oversight evidence.** Compliance gets coverage statistics, breach rates and trends for\n   conduct reporting and root cause analysis.","valueDrivers":["compliance","risk-reduction","customer-experience","employee-productivity"],"kpis":["quality-score-uplift","interactions-handled","handling-time-reduction","accuracy","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A contact centre with 500 agents and a QA team of 10 to 20 analysts","inputs":[{"key":"qaAnalysts","label":"QA analysts","low":10,"high":20,"unit":"analysts","note":"Editorial assumption of one analyst per 25 to 50 agents. Replace with your own team size."},{"key":"analystCost","label":"Fully loaded cost per QA analyst","low":50000,"high":70000,"unit":"USD per analyst per year","note":"Editorial assumption. Replace with your own cost."},{"key":"timeFreed","label":"Share of analyst time moved from listening and scoring to coaching and root cause work","low":0.3,"high":0.5,"unit":"fraction of analyst time","note":"Editorial assumption. Automated scoring replaces most manual listening, but calibration and review of flagged interactions remain."}],"formula":"qaAnalysts * analystCost * timeFreed","currency":"USD","period":"per year","resultLabel":"QA analyst capacity redirected","caveat":"Covers QA effort only. It leaves out the main value, which is finding compliance breaches, mis selling and complaint drivers that a sample misses, and the platform and transcription costs."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Transcription and scoring are mature. The effort is in writing a rubric specific enough for a machine to apply, calibrating it against human reviewers, integrating recordings from the telephony platform, and agreeing with HR and employee representatives how results may be used.","dataPrerequisites":["The QA rubric and regulatory disclosure requirements per product and journey","A calibration set of interactions scored by experienced reviewers","Access to recordings and chat logs with metadata (agent, team, product, outcome)"],"integrations":["Call recording and telephony platform","Chat and messaging platforms","QA, coaching and workforce management tools","Complaints and conduct risk reporting"]},"implementation":{"steps":[{"title":"Rewrite the rubric for machines","detail":"Turn vague criteria (\"showed empathy\") into observable checks (\"acknowledged the problem before offering a solution\"), and list every mandatory disclosure per journey."},{"title":"Calibrate before you publish scores","detail":"Have experienced reviewers score a few hundred interactions and compare with the model. Publish only criteria where agreement is at least as good as between two humans. British Gas tested until its automated scores reached at least 80% agreement with human reviewers before going live."},{"title":"Start with compliance checks","detail":"Mandatory disclosures and prohibited statements are the clearest criteria and the strongest oversight evidence. Add softer service behaviours later."},{"title":"Agree the rules of use","detail":"Decide with HR, legal and employee representatives how results feed coaching and whether they may affect evaluation, pay or discipline, and document the assessment."},{"title":"Give agents visibility and a dispute route","detail":"Let agents see their scored interactions and challenge them. Disputes are also a calibration signal."},{"title":"Close the loop to root causes","detail":"Feed recurring breaches and complaint drivers to product, process and training owners, not only to individual coaching."}],"guardrails":["Scores are inputs for human review, never automatic sanctions","Rubric criteria published only after calibration against human reviewers","No inference of agents' emotions","Card data and special category data redacted before scoring and storage","Agents can see and dispute their results"],"humanInTheLoop":"QA and compliance reviewers confirm every flagged breach before it is recorded or acted on. Team leaders decide on coaching, and any consequence for an individual follows the normal HR process with human judgment. Reviewers recalibrate the rubric at least quarterly.","kpisToInstrument":["Share of interactions evaluated automatically","Agreement between model and human reviewers per criterion","Confirmed breach rate for mandatory disclosures, per journey","Disputed scores and their outcome","Complaint and repeat contact trends after coaching"],"failureModes":[{"title":"Scores treated as facts","detail":"An uncalibrated score drives performance ratings. Calibrate per criterion and keep humans in every consequential decision."},{"title":"Rubric too vague for a model","detail":"Criteria such as \"professional tone\" give noisy scores. Rewrite into observable behaviours."},{"title":"Surveillance backlash","detail":"Agents experience total monitoring without transparency, and trust and retention fall. Be open about what is measured and let agents dispute."},{"title":"Finding problems nobody fixes","detail":"Breaches are counted but root causes in products or processes remain. Route themes to owners with deadlines."}]},"risk":{"euAiAct":{"tier":"high","basis":"Scoring individual agents' interactions to monitor and evaluate their performance and behaviour falls under Annex III point 4(b), employment and worker management. The Article 6(3) exception does not apply where the system profiles natural persons. Inferring agents' emotions is prohibited under Article 5(1)(f), except for medical or safety reasons. Inferring customers' emotions from their voice is emotion recognition on biometric data: high risk under Annex III point 1(c), and Article 50(3) requires informing the people exposed to it. Analytics that only aggregate interaction themes without evaluating individuals can fall outside the high risk category."},"regulations":["eu-ai-act","gdpr","uk-gdpr","uk-consumer-duty","pci-dss","iso-42001"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) covers AI systems intended to monitor and evaluate the performance and behaviour of workers."},{"title":"Article 5, prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Point 1(f) prohibits inferring emotions of natural persons in the workplace, except for medical or safety reasons."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Point 3 requires deployers of an emotion recognition system to inform the people exposed to it, which matters when voice analytics infers customer sentiment from speech."},{"title":"Employment practices and data protection: monitoring workers","issuer":"UK Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/","note":"Guidance on transparency, necessity and impact assessments when monitoring workers, including through automated tools. The ICO marks it as under review after the Data (Use and Access) Act."}],"controls":["Data protection impact assessment and, in the EU, the high risk obligations for the deployer","Documented calibration results per criterion and per model version","Written rules on how scores may and may not be used for individual decisions","Agent access to their results and a dispute process","Retention limits and access controls on recordings and transcripts"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the scoring runs as an **agentic workflow**: interactions from the platform's own\n**voice** and chat channels, or recordings fetched through **custom functions**, are transcribed\nwith streaming speech recognition or the **self hosted transcription** option with speaker\ndiarization, then an **AI agent** with **structured output** scores each one against the\nrubric held in a **knowledge base**, citing the excerpt behind each finding.\n\nFindings above a threshold go to a reviewer through **human in the loop approval** before\nanything is recorded, and every run keeps an audit trail. Chat traffic on the platform's own\nchannels passes the gateway, where **PII masking** and card number tokenization apply before it\nreaches the chat backend. Recordings fetched into the workflow do not pass the gateway, so add a\nredaction step in the workflow before scoring. **Test suites** with LLM grading hold the\nhuman calibration set and are rerun after every rubric or model change, so agreement with\nreviewers is tracked over time. The platform is model agnostic and can run in the EU or UAE region."},"faq":[{"question":"Can AI really review every call?","answer":"Yes, coverage is the main change. Observe.AI reports that Central Bank, a US community bank group, evaluated 167,000 calls in the third quarter of 2024, up from 24 per quarter before automated QA, and that DoorDash reached nearly 100% automated quality coverage across 19,000 agents."},{"question":"Is automated agent scoring high risk under the EU AI Act?","answer":"When it evaluates individual workers, yes: Annex III point 4(b) covers monitoring and evaluating the performance and behaviour of workers. Inferring agents' emotions from biometric data such as voice is prohibited under Article 5(1)(f), except for medical or safety reasons. Aggregate analytics on themes, without scoring individuals, carry less risk."},{"question":"How do we know the scores are right?","answer":"Calibrate. Have experienced reviewers score the same interactions as the model, publish only criteria where agreement is good enough, and repeat after every rubric or model change. British Gas, for example, went live only after its automated scores reached at least 80% agreement with human reviewers, and lets agents and team leaders challenge scores."}],"related":["live-agent-assist","conversation-roleplay-training","complaints-root-cause-analysis","sales-call-coaching-and-crm-update","customer-feedback-analysis"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog as an industry neutral page, with five evidence records verified against their sources. The catalog cited a vendor blog without a named deployment; it was replaced by named case studies."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed an unsourced 2% sample figure, added the British Gas 80% calibration threshold, added customer emotion recognition (Annex III point 1(c), Article 50(3)) to the risk basis, corrected the VitalityHealth deployment year and a source title, narrowed DoorDash channels to voice, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: limited the Blits.ai PII masking claim to chat traffic through the gateway, removed the DoorDash coverage figure from the automation rate benchmark (the taxonomy has no QA coverage KPI), replaced automation rate with quality score uplift in the KPIs, traced the Oportun (2024) and British Gas (2026) evidence years, added UK GDPR and the Article 5(1)(f) exceptions, and noted that the ICO guidance is under review."}],"slug":"call-quality-and-compliance-monitoring","url":"https://www.blits.ai/ai-use-cases/call-quality-and-compliance-monitoring","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":167000,"min":167000,"max":167000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"central-bank-automated-call-quality","pooled":true}]},{"kpi":"quality-score-uplift","label":"Quality score uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10,"min":10,"max":10,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"british-gas-quality-and-regulatory-assurance","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"central-bank-automated-call-quality","pooled":false}]}],"indicativeValueResult":{"low":150000,"high":700000},"evidence":["british-gas-quality-and-regulatory-assurance","central-bank-automated-call-quality","doordash-automated-quality-coverage","oportun-ai-quality-management","vitalityhealth-automated-quality-assurance"]},{"title":"AI quality inspection on the production line","shortTitle":"Production quality inspection","seoTitle":"AI quality inspection for manufacturing lines","metaDescription":"AI checks each unit by camera, sound or sensor. In 2023 Audi reported about 1.5 million welds analysed per shift; NVIDIA cites a 67% lower defect rate at Pegatron.","definition":"AI that inspects every unit on a production line, from camera images, sound or machine process data, to find defects, missing parts and wrong variants in real time, and routes the few anomalies it flags to a quality inspector instead of relying on manual sampling at the end of the line.","aliases":["AI visual inspection","AI defect detection","machine vision quality control","automated optical inspection with deep learning","acoustic quality inspection"],"industries":["manufacturing","automotive"],"functions":["operations"],"patterns":["computer-vision","anomaly-detection","synthetic-data-generation","agentic-workflow"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"production","problem":"Classic quality control samples. An inspector checks a share of the units, technicians test a\nrandom selection of welds with ultrasound, a specialist walks round the finished product at the\nend of the line. Defects that fall between samples travel on to the next station, to the customer\nor into a warranty claim, further away from the station that caused them.\n\nRule based machine vision suits simple, stable parts, but every new variant, lighting change or\ndefect type means writing and tuning new rules. The shift is to learning models that check every unit, from images, sound or the process data a\nmachine already produces, and send people only the cases that look wrong.","problemStats":[],"howItWorks":"1. **Capture every unit.** Cameras, microphones or the machine controller record each part or\n   vehicle at the station, for example images of an assembly step, driving noise from the seats or\n   the process readings a machine logs for each joint.\n2. **Score it against what good looks like.** A model trained on labelled examples, or on normal\n   production when defects are rare, classifies the unit or scores how far it deviates from normal.\n   Synthetic defect images can fill the gap when real defects are too rare to train on.\n3. **Check it against the order.** The expected variant, parts list and assembly steps come from\n   the production system, so the model knows what this specific unit should look like.\n4. **Alert the right person at the right station.** An anomaly goes straight to the worker or\n   inspector on a smart device, with the image or clip, while the unit can still be fixed in line.\n5. **Close the loop.** Confirmed and rejected findings are logged, used to retrain the model and\n   analysed for root causes such as a drifting machine setting, so the process improves rather\n   than just the inspection.","valueDrivers":["risk-reduction","cost-to-serve","employee-productivity","speed"],"kpis":["error-reduction","detection-rate-improvement","false-positive-reduction","cost-reduction","accuracy","interactions-handled"],"indicativeValue":{"referenceOrg":"An assembly plant building 200,000 units a year","inputs":[{"key":"units","label":"Units produced per year","low":200000,"high":200000,"unit":"units per year","note":"The reference plant."},{"key":"defectRate","label":"Share of units with a defect found late (final inspection, customer or warranty)","low":0.02,"high":0.04,"unit":"fraction of units","note":"Editorial assumption, replace with your own late defect and warranty rate."},{"key":"costPerDefect","label":"Cost of a defect found late","low":150,"high":400,"unit":"USD per defect","note":"Editorial assumption covering rework, scrap and warranty handling. Replace with your own cost of poor quality."},{"key":"reduction","label":"Share of late defects avoided by inspecting every unit in line","low":0.2,"high":0.5,"unit":"fraction of late defects","note":"Conservative against the benchmark on this page (NVIDIA reports a 67% decrease in defect rates on Pegatron assembly lines using its visual AI agent), because that figure is one vendor reported case in electronics assembly. The benchmark is not like for like: it measures fewer defects produced, while this input measures fewer defects escaping to late stages."}],"formula":"units * defectRate * costPerDefect * reduction","currency":"USD","period":"per year","resultLabel":"Cost of late defects avoided","caveat":"Counts avoided rework, scrap and warranty cost only. It leaves out cameras, sensors, edge computing and integration, the labelling effort, any reduction in inspection staff hours and the value of fewer recalls and a better brand reputation."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The model is rarely the hard part. The work is in stable image or signal capture on a moving line, enough labelled defects, a link to the order so the model knows the expected variant, and a validation approach that quality auditors and certification bodies accept.","dataPrerequisites":["Images, sound or process signals captured consistently per unit and station","Labelled examples of good units and of each defect class, or long runs of normal production","The expected variant, parts list and work steps per unit from the production system","Confirmed outcomes of flagged cases to retrain and to measure false alarms"],"integrations":["Manufacturing execution system for orders, variants and unit identity","Cameras, microphones or machine controllers at the stations","Edge computing near the line for low latency scoring","Quality management system for defect records and corrective actions","Worker devices or line displays for alerts"]},"implementation":{"steps":[{"title":"Pick one station with a costly, visible defect","detail":"Choose a defect that escapes today, is expensive later and can be seen, heard or measured at one station, for example a missing fastener or a weak weld. Record its current escape rate."},{"title":"Fix the capture before the model","detail":"Stabilise lighting, camera angle, microphone placement or signal logging first. A better model cannot make up for inconsistent capture."},{"title":"Build the defect library","detail":"Collect and label real defects with the quality team. When real defects are rare, train on normal production for anomaly detection or add synthetic defect images, and keep a real holdout set to test on."},{"title":"Run in shadow mode next to the current inspection","detail":"Let the model score every unit while the existing sampling continues, and compare both on catches, misses and false alarms before anyone relies on it."},{"title":"Agree the validation with quality and auditors","detail":"Document how a result is produced, the acceptance thresholds and the retraining rules, so the process survives audits and certification. Audi worked with DGQ and Fraunhofer for this."},{"title":"Go live with a human on every alert, then widen","detail":"Route anomalies to the station, measure confirmation rates per defect class, and only then reduce manual sampling or add stations, plants and suppliers."}],"guardrails":["A documented acceptance threshold per defect class, with a human decision on every flagged unit","A fallback to the previous inspection method when the model or capture is unavailable","Retraining only through change control, tested on a fixed holdout set before release","Drift monitoring on input images or signals after changes to materials, lighting or machines","Recording limited to the product and process, with worker privacy rules for any video of people"],"humanInTheLoop":"Quality inspectors confirm or reject every flagged unit and own the decision to release, rework or scrap. Quality engineers approve each new defect class, threshold and model version, and a sample of passed units is still inspected manually to measure what the model misses.","kpisToInstrument":["Escape rate of each defect class to later stations, the customer and warranty, before and after","False alarm rate per defect class and station","Share of flagged units confirmed as real defects","Inspection coverage (share of units inspected) and time to detection","Cost of poor quality per unit produced"],"failureModes":[{"title":"False alarms that train people to ignore alerts","detail":"A model that flags too much gets overridden by habit. Tune thresholds per defect class and report confirmation rates to the line."},{"title":"Silent drift after a process change","detail":"A new supplier, paint batch or light fitting changes the input and accuracy drops without anyone noticing. Monitor input drift and recheck on a holdout set after every change."},{"title":"A model trained on too few real defects","detail":"Rare defects are exactly the ones that matter. Use anomaly detection or synthetic data, but always test on real defects."},{"title":"Inspection without root cause","detail":"The model catches more defects but nobody fixes the cause. Feed findings into corrective actions and machine settings."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Inspecting products is not an Annex III use, so a system that only judges parts, welds or assemblies is usually minimal risk. Two designs change that. Under Article 6(1) it is high risk when both conditions hold: it is a safety component of a product (or itself a product) covered by the Union harmonisation legislation in Annex I, and that law requires a third party conformity assessment of the product. For a production line the relevant product laws are the Machinery Regulation (EU) 2023/1230 and, for cars, the vehicle type approval regulations. Since the Digital Omnibus on AI, Regulation (EU) 2026/1744, moved the Machinery Regulation into Annex I Section B, where the vehicle type approval regulations already sat. Article 6(1) still classifies such a safety component as high risk, but under Article 2(2) only Article 6(1), Article 60a and Articles 102 to 112 of the AI Act apply directly. The requirements reach the system through the sectoral law instead: delegated acts amending Annex III of the Machinery Regulation, and type approval for vehicles. The AI Act rules for Article 6(1) high risk systems apply from 2 August 2028. An inspection system on the assembly line is usually not a safety component of the product it inspects. If it monitors and evaluates the performance and behaviour of individual workers, for example by scoring who made an assembly error, it falls under Annex III point 4(b) and is high risk. Keep the output about the unit, not the person."},"regulations":["eu-ai-act","gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Article 2, scope","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/2/","note":"Article 2(2), as amended by Regulation (EU) 2026/1744, says that for high risk AI systems related to products under the Annex I Section B laws, such as machinery and vehicle type approval, only Article 6(1), Article 60a and Articles 102 to 112 apply."},{"title":"Annex I, list of Union harmonisation legislation","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/1/","note":"The product laws behind Article 6(1). Since Regulation (EU) 2026/1744, Section B lists the Machinery Regulation (EU) 2023/1230 as point 21, next to motor vehicle type approval (Regulations 2018/858 and 2019/2144), and the Machinery Directive 2006/42/EC is deleted from Section A."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) lists AI used to monitor and evaluate the performance and behaviour of workers, which matters when inspection video also shows people."},{"title":"Regulation (EU) 2023/1230 on machinery","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2023/1230/oj","note":"The EU Machinery Regulation replaces Directive 2006/42/EC and applies from 14 January 2027. Since Regulation (EU) 2026/1744 it is listed in Annex I Section B of the AI Act. It is the product law that can make an AI safety component of a machine high risk under Article 6(1), and the AI requirements for such machines are to be added to its Annex III by delegated acts that apply by 2 August 2028."},{"title":"IATF 16949 automotive quality management","issuer":"International Automotive Task Force","region":"global","url":"https://www.iatfglobaloversight.org/","note":"The body behind the IATF 16949 automotive quality management standard. At certified plants, plan for an AI based inspection process to be reviewed as part of the quality management system in certification audits."},{"title":"Digital Omnibus on AI, Regulation (EU) 2026/1744","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/ai-act-explorer/digital-omnibus/","note":"Entered into force on 27 July 2026. It moves machinery from Annex I Section A to Section B, limits the AI Act's direct application to machinery to the provisions in Article 2(2), and sets 2 August 2028 as the application date for Article 6(1) high risk AI systems."}],"controls":["Validation record per model version with holdout results per defect class","Change control for thresholds, training data and model releases","Traceability from every flagged or passed unit to the model version and input data","Data protection impact assessment where cameras capture workers","Periodic manual audit of passed units to estimate the miss rate"],"incidents":[]},"blitsAi":{"howToBuild":"The inspection model itself runs on the line, on the camera or edge platform of the manufacturer's\nchoice. Blits.ai adds the layer around it that people work with. An **AI agent** connected to a\n**SQL knowledge base** of inspection results lets quality engineers ask in plain language which\nstations, variants or shifts show rising defect rates, and a **knowledge base** with hybrid\nretrieval over work instructions and defect catalogues answers how to rework a finding.\n\n**Agentic workflows** watch the results and, when a defect pattern crosses a threshold, draft a\nquality case or corrective action and wait for **human in the loop approval** before anything is\nraised in the quality system through **custom functions**. Line staff reach the agent in\n**Microsoft Teams** or through the REST API channel on their devices, **monitors** check the\nagent's answers on a schedule, and the platform is model agnostic, with EU and UAE data residency."},"faq":[{"question":"How much can AI inspection reduce defects?","answer":"Published results vary widely by defect and line. NVIDIA reports that Pegatron saw a 67% decrease in defect rates on assembly lines using its visual AI agent, a vendor figure for electronics assembly. Plan for a smaller effect at first and measure escapes per defect class against your current sampling."},{"question":"Does AI inspection need cameras?","answer":"Not always. At Plant Dingolfing, BMW's Acoustic Analytics listens to driving noises through microphones on the seats as a final check before handover. Choose the signal that shows the defect most reliably at the lowest cost."},{"question":"Is AI quality inspection high risk under the EU AI Act?","answer":"Usually not when it only judges the product. It is classified as high risk if it is a safety component of a product under Annex I legislation, such as machinery or vehicles, and that law requires a third party conformity assessment. Since the Digital Omnibus on AI, machinery is in Annex I Section B, like vehicle type approval already was, so the requirements for those systems come through the Machinery Regulation (delegated acts that apply by 2 August 2028) and vehicle type approval rather than the AI Act directly. It is also high risk if it scores the performance of individual workers, which Annex III point 4(b) covers."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Editor pass: updated the EU AI Act basis, the Article 2, Annex I and Machinery Regulation notes and the FAQ for the Digital Omnibus on AI (Regulation (EU) 2026/1744), which moved machinery to Annex I Section B and set 2 August 2028 for Article 6(1) systems; added the omnibus as guidance, fixed the Annex III 4(b) wording, softened the IATF note and the unsourced problem claims."},{"date":"2026-09-27","note":"Editor pass: corrected the Annex I Section B sentence (Article 6(1) does apply; Article 2(2) routes the requirements through vehicle type approval) and cited Article 2 and Annex I, added the Machinery Regulation application date, qualified the IATF note, dated the Audi figure, flagged the defect benchmark as not like for like, and added the synthetic data pattern."},{"date":"2026-09-27","note":"First version, researched for the energy, manufacturing and automotive scope with BMW Group, Audi and Pegatron evidence checked against the sources. Editor pass removed claims the Audi release does not support (input data, year, rollout status), unsourced problem claims, and completed the Article 6(1) conditions."}],"slug":"production-line-quality-inspection","url":"https://www.blits.ai/ai-use-cases/production-line-quality-inspection","benchmarks":[{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":7,"min":7,"max":7,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"pegatron-visual-ai-assembly-inspection","pooled":true}]},{"kpi":"error-reduction","label":"Error reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":67,"min":67,"max":67,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"pegatron-visual-ai-assembly-inspection","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1500000,"min":1500000,"max":1500000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"audi-spot-weld-quality-analytics","pooled":true}]}],"indicativeValueResult":{"low":120000,"high":1600000},"evidence":["audi-spot-weld-quality-analytics","bmw-group-aiqx-quality-inspection","pegatron-visual-ai-assembly-inspection"]},{"title":"AI quote and estimate generation from customer requirements","shortTitle":"Quote and estimate generation","seoTitle":"AI quote generation and sales estimating","metaDescription":"Draft quotes from drawings, item lists and roof photos. dida says Enpal cut solar quote work from 120 to 15 minutes; Microsoft says ODP bids take hours, not days.","definition":"AI that turns what a customer sends, such as a product list, a drawing, a roof photo or a request for quotation, into a draft quote: it reads the input, matches items to the catalog, calculates quantities and applies the organization's price rules, and hands the draft to a seller or estimator who checks and sends it.","aliases":["AI quoting","AI estimating","automated quote generation","RFQ to quote automation","AI configure price quote","SKU matching for quotes"],"industries":["cross-industry","manufacturing","energy-and-utilities","retail-and-ecommerce"],"functions":["sales","product-and-pricing"],"patterns":["document-processing","agentic-workflow","computer-vision"],"channels":["internal-tools","microsoft-teams","email"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"In many business to business and project sales, the quote is the bottleneck. A distributor\nreceives a customer's list of thousands of items in the customer's own descriptions and has to\nmatch each one to its own catalog. A solar installer has to measure a roof from a satellite image\nbefore it can say how many panels fit. A steel or building products supplier has to read\narchitectural drawings to know what to price. The work is skilled, slow and repetitive, so\ncustomers can wait days for an answer.\n\nManual quoting also produces errors: a miscounted roof, a wrong substitute product, a price rule\napplied inconsistently. dida, which built the roof assessment step of Enpal's quotes, describes Enpal's old process as\ntaking a salesperson 120 minutes per quote and as error prone, leading to inaccurate projections of\ncost and energy production.","problemStats":[],"howItWorks":"1. **Read the request.** The AI reads what the customer sent: an email, a spreadsheet of items,\n   a PDF drawing or an image of a roof or site, and extracts items, dimensions and quantities.\n2. **Match to the catalog.** Each item is matched to the organization's own products or\n   configurable options, with a confidence score and alternatives where no exact match exists.\n3. **Calculate.** Quantities, dimensions and technical constraints are calculated with\n   deterministic rules or models (for example the usable roof area and the number of panels),\n   not left to a language model's arithmetic.\n4. **Price.** List prices, customer contracts, discounts and margin rules are applied from the\n   pricing system, with anything outside the seller's authority flagged for approval.\n5. **Review and send.** The seller or estimator checks the draft, adjusts it, and sends the\n   quote from the CRM or quoting system; accepted and rejected quotes feed back into matching.","valueDrivers":["speed","revenue-growth","employee-productivity","risk-reduction"],"kpis":["handling-time-reduction","processing-time-reduction","users-served","accuracy","conversion-rate-uplift"],"indicativeValue":{"referenceOrg":"A distributor or installer that issues 10,000 custom quotes a year","inputs":[{"key":"quotes","label":"Custom quotes per year","low":10000,"high":10000,"unit":"quotes per year","note":"The reference organization."},{"key":"minutesPerQuote","label":"Staff time per quote today","low":60,"high":120,"unit":"minutes per quote","note":"Editorial assumption; dida reports 120 minutes per solar quote at Enpal before automation. Replace with a time study of your own quotes."},{"key":"reduction","label":"Share of quoting time saved","low":0.4,"high":0.8,"unit":"fraction of time","note":"Conservative against the benchmark on this page (dida reports an 87.5% reduction at Enpal), because most catalogs are less uniform than solar roofs."},{"key":"hourlyCost","label":"Fully loaded cost of a seller or estimator","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption, replace with your own cost."}],"formula":"quotes * minutesPerQuote / 60 * reduction * hourlyCost","currency":"USD","period":"per year","resultLabel":"Quoting time released","caveat":"Time value only. It leaves out the build and integration cost, and the revenue effect of faster quotes, which Microsoft reports as 20% more sales opportunities a quarter at ODP but which depends on the market."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Reading the input is usually the easier part; the difficulty is a clean product catalog with attributes good enough to match against, and a pricing system the AI can call instead of guessing prices.","dataPrerequisites":["A product catalog with structured attributes and approved substitutes","Price lists, customer contract prices and discount rules in a system of record","A set of past requests with the quotes that were sent, to measure matching accuracy","Technical rules for configuration and quantity calculation"],"integrations":["CRM or configure, price, quote (CPQ) system","ERP or pricing engine for prices, stock and margins","Document and email intake for requests","Mapping or imaging services where the quote depends on a site"]},"implementation":{"steps":[{"title":"Pick one quote type with volume","detail":"Start with a quote type that is frequent and fairly standard, such as a list of catalog items or a single product family, and measure the minutes and errors per quote today."},{"title":"Keep prices out of the model","detail":"Let the AI match and count, and let the pricing system price. Every price in the draft should come from a call to the system of record, never from generated text."},{"title":"Measure matching accuracy","detail":"Use past requests and the quotes actually sent as a test set, and report precision per product family before sellers rely on it."},{"title":"Show confidence and alternatives","detail":"Mark low confidence matches and unusual quantities so the seller checks those lines first, and keep the seller able to adjust anything."},{"title":"Learn from sent quotes","detail":"Feed the seller's corrections and the quote outcome back into matching and into the test set, and review the lines sellers change most often."}],"guardrails":["Prices, discounts and availability only from the pricing and ERP systems, never generated","Discounts or margins outside a seller's authority routed for approval","Human review of every quote before it is sent to a customer","Technical calculations done by deterministic rules or validated models, with the inputs shown"],"humanInTheLoop":"The seller or estimator reviews and sends every quote and owns the price. Pricing managers own the rules and approve exceptions, and product specialists maintain the catalog attributes and approved substitutes the matching relies on.","kpisToInstrument":["Minutes of staff time per quote and elapsed time from request to quote","Line level matching accuracy and the share of lines sellers change","Quote to order conversion, before and after","Quotes with pricing or quantity errors found after sending"],"failureModes":[{"title":"Plausible but wrong matches","detail":"The AI picks a similar product that does not meet the customer's specification. Show confidence per line and test matching on past quotes."},{"title":"Invented prices","detail":"A language model fills in a price or discount it was never given. Keep pricing in the system of record and block generated prices."},{"title":"Garbage catalog in, garbage quote out","detail":"Matching fails because product attributes are incomplete. Fix the catalog data first; product content enrichment helps here."},{"title":"Speed without margin control","detail":"Faster quotes with inconsistent discounts erode margin. Enforce discount authority in the workflow, not in the prompt."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Drafting business quotes for a seller to review is not an Annex III use and does not interact with the customer as an AI system. It would need a fresh assessment if the system set individual consumer prices or terms in areas such as credit or insurance, where Annex III point 5 can apply."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"Algorithms: how they can reduce competition and harm consumers","issuer":"UK Competition and Markets Authority","region":"europe","url":"https://www.gov.uk/government/publications/algorithms-how-they-can-reduce-competition-and-harm-consumers","note":"Describes how pricing algorithms and personalized pricing can harm competition and consumers, relevant when quote tools also set prices."}],"controls":["Pricing rules and discount authority enforced in the quoting system, with an audit trail","A record of the AI draft and the seller's changes for every quote sent","Periodic review of matching accuracy and of quote errors reported by customers","Data protection review when quotes use images or data about a customer's home or site"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow**. A dialog flow on the **email channel** or\n**Microsoft Teams** receives the request and its attachment and triggers the workflow, which can\nalso start on a schedule or from an API token. An **AI agent** extracts the items with\n**structured output** and matches them to the catalog held in a **SQL knowledge base** or\nsearched through **hybrid retrieval**. **Custom functions** call the pricing engine, ERP or CRM\nthrough REST (SAP, Salesforce and NetSuite are in the integration catalog, and Microsoft\nDynamics 365 is a ready made business system tool) so that every price comes from the system of\nrecord, and the **file generation** tool returns the draft quote.\n\n**Human in the loop** confirmation above a configurable threshold holds large or unusual quotes\nfor the seller or a pricing manager, and the **tool execution policy** limits what the agent may\ncall. **Test suites** replay past requests against the quotes actually sent, the run history\nkeeps an audit trail per quote, and the platform is model agnostic."},"faq":[{"question":"Can AI produce a customer quote without a salesperson?","answer":"For simple, standard requests it can draft the whole quote. At Enpal the tool is used by Enpal staff and at ODP by sales representatives; DeAcero's public description does not say how its proposals are reviewed. The safer design keeps a seller reviewing every quote and takes prices from the pricing system, never from a language model."},{"question":"How much faster does AI make quoting?","answer":"It depends on the input. dida reports that Enpal's solar quote went from 120 minutes to 15, a reduction of 87.5%, and Microsoft reports that ODP Business Solutions now returns pricing bids in hours instead of one to two days."},{"question":"Is this the same as CPQ software?","answer":"It sits in front of it. Configure, price, quote systems hold the rules and prices; the AI reads unstructured customer requests and fills the quote, so sellers spend less time on data entry."}],"related":["business-connectivity-quoting-and-service-assistant","product-content-and-catalog-enrichment"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Unpublished by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review). Enpal summary limited to the roof assessment step dida describes, unsourced languages removed from the Enpal and ODP records, ODP bid time stated as hours instead of days."},{"date":"2026-09-27","note":"First version, with evidence from Enpal, The ODP Corporation and DeAcero checked against the sources. Claims attributed to dida and Microsoft as the claimants after review."}],"slug":"sales-quote-and-estimate-generation","url":"https://www.blits.ai/ai-use-cases/sales-quote-and-estimate-generation","benchmarks":[{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":87.5,"min":87.5,"max":87.5,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"enpal-solar-quote-automation","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":150,"min":150,"max":150,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"enpal-solar-quote-automation","pooled":true}]}],"indicativeValueResult":{"low":160000,"high":1120000},"evidence":["deacero-blueprint-cost-proposal-agents","enpal-solar-quote-automation","odp-corporation-sales-quote-assistant"]},{"title":"AI recommendation and personalization engine for streaming and media","shortTitle":"Content recommendation and personalization","seoTitle":"AI recommendation engine for streaming","metaDescription":"AI decides what each viewer sees next. Netflix researchers estimate a popularity ranking cuts engagement 12%; Spotify's Discover Weekly hit 100 billion streams.","definition":"A recommendation system that decides, for each individual viewer or listener, what to show next on a home page, in search or in a personalized playlist, learned from that person's own viewing or listening history, ratings and context, and continuously updated as new content is added and behavior changes. It ranks the catalog's own content; it is not the marketing engine that decides which offers or campaigns to send, which is a separate use case in this library.","aliases":["AI content recommendation","streaming personalization engine","recommendation algorithm","personalized playlist and homepage","next best content"],"industries":["media-and-entertainment"],"functions":["marketing","analytics-and-reporting"],"patterns":["recommendation-and-personalization","prediction-and-scoring"],"channels":["mobile-app","api"],"audience":"back-office","autonomy":"autonomous","adoptionStage":"mainstream","segment":"content discovery","problem":"A streaming service or publisher with a large catalog has a discovery problem, not a supply\nproblem: most of what would delight a given viewer or listener is not what they would have found by\nbrowsing. Left to manual curation or simple popularity ranking, the same hit titles and tracks\ndominate every home page, niche and new content struggles to be found, and people who cannot find\nsomething they like churn.\n\nRecommendation systems at large platforms have evolved considerably, and the work keeps shifting.\nEarly systems used collaborative filtering, matching people with similar histories. Netflix\nillustrates a newer direction: rather than a variety of specialized models each covering one need\n(for example \"Continue Watching\" or \"Today's Top Picks for You\"), a single foundation model learns\nfrom a person's comprehensive interaction history, tokenized the way text is tokenized for a large\nlanguage model, and shares that learning with other models through embeddings or fine tuning.","problemStats":[],"howItWorks":"1. **Collect signals.** Every play, pause, skip, rating, search and scroll is logged as an event,\n   alongside metadata about the content itself (genre, cast, tempo, mood, release date).\n2. **Build a shared representation.** A model learns embeddings for people and for content from\n   this interaction history at scale, so that people and titles with similar patterns end up close\n   together in the model's internal representation, and new, unwatched titles can still be placed\n   using their metadata (a cold start problem).\n3. **Rank for each surface.** The shared model, or models fine tuned from it, rank candidates for a\n   specific surface: the home page, a personalized playlist, a search result, an autoplay queue.\n4. **Serve within a latency budget.** Ranking has to return in milliseconds, so systems trade off\n   how much history they can consider against how fast they can score it, often using sparse\n   attention or similar techniques to fit long histories into a short serving budget.\n5. **Measure causally, not just by clicks.** Because recommendations are also what people see, raw\n   engagement with recommended content overstates the system's effect. Mature teams run\n   experiments, including replacing the system with a simpler baseline for a slice of users, to\n   isolate how much of the engagement the recommender actually causes.","valueDrivers":["customer-experience","revenue-growth"],"kpis":["users-served","revenue-uplift","churn-reduction"],"indicativeValue":{"referenceOrg":"A streaming service with 5 million active subscribers","inputs":[{"key":"subscribers","label":"Active subscribers","low":5000000,"high":5000000,"unit":"subscribers","note":"The reference service."},{"key":"annualArpuUsd","label":"Average annual revenue per subscriber","low":60,"high":150,"unit":"USD per subscriber per year","note":"Editorial assumption for a mid tier subscription service. Replace with your own ARPU."},{"key":"retentionEffect","label":"Share of subscribers retained per year because of personalized recommendations","low":0.01,"high":0.03,"unit":"fraction of subscribers per year","note":"Netflix's own finding is that replacing its recommender with a popularity based ranking would cut member engagement by 12% (arXiv:2511.07280, not attached here as a source since the paper does not measure retention). Editorial assumption, replace with your own: this range assumes only a small, unverified fraction of that engagement effect converts into an avoided cancellation."}],"formula":"subscribers * retentionEffect * annualArpuUsd","currency":"USD","period":"per year","resultLabel":"Annual subscription revenue retained through personalization","caveat":"Gross retention value only, built on an unverified assumption about how much of Netflix's engagement effect converts into retention. It leaves out the cost of running the recommendation system, any effect on new subscriber acquisition, and the fact that churn has many causes besides content discovery, so treat this as a rough illustration, not a forecast."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"A production recommendation system is a standing data and machine learning platform, not a single project: it needs a real time event pipeline, a feature and embedding store, offline and online evaluation, and a serving layer that meets a strict latency budget across every surface it feeds.","dataPrerequisites":["Complete interaction event logs (plays, skips, ratings, searches) at the individual level","Content metadata (genre, cast, language, release date and, ideally, richer descriptors)","A way to measure outcomes beyond clicks, such as a holdout or interleaving experiment"],"integrations":["Event streaming or logging pipeline from every client (app, web, TV)","Content catalog and metadata management system","Low latency online serving infrastructure for the ranking model","Experimentation platform for A/B and holdout testing"]},"implementation":{"steps":[{"title":"Start from a strong baseline, not a blank page","detail":"Ship a well tuned popularity or collaborative filtering baseline first, and measure every later model against it with a real experiment, not just an offline metric."},{"title":"Instrument the full interaction history","detail":"Capture every meaningful event, not just completions, and decide early how to tokenize or aggregate them (for example, summing watch duration per title) so the signal survives compression into a manageable sequence length."},{"title":"Solve cold start explicitly","detail":"New content and new users have no history. Use content metadata to place new titles near similar existing ones, and use onboarding preferences or early signals to place new users, rather than defaulting everyone to the same popular list."},{"title":"Separate ranking from presentation","detail":"Keep the model's job (rank candidates) separate from product decisions (how many rows, how much diversity, whether to explain a pick), so the product team can adjust presentation without retraining the model."},{"title":"Run holdout and interleaving experiments","detail":"Periodically hold out a small population on an older or simpler algorithm, or interleave results from two algorithms in the same session, so you can measure the model's true incremental value instead of trusting raw engagement with recommended content."},{"title":"Watch diversity, not only accuracy","detail":"A model optimized purely for predicted engagement will over serve the same popular titles. Track catalog coverage and the share of recommendations going to mid and long tail content alongside accuracy metrics."}],"guardrails":["Content and safety filters applied before anything is recommended to a person, independent of the ranking model (age appropriate content, platform policy compliance)","A minimum diversity or exploration budget so the system keeps surfacing content outside a person's established pattern, rather than narrowing to a filter bubble","Human review of what the model associates with sensitive categories (for example, content aimed at children) before those associations reach production","Rate limits and monitoring on any interactive or agentic layer built on top of the ranking model"],"humanInTheLoop":"Editorial and content teams own the guardrails (what may never be recommended, and to whom), review model behavior on sensitive content categories, and set the exploration and diversity targets the ranking has to respect; data scientists own experiment design and causal measurement so engagement gains are not mistaken for value the system did not create.","kpisToInstrument":["Incremental engagement from a holdout or interleaving experiment, not raw engagement with recommended content","Catalog coverage and the share of engagement going to content that is not already popular","Subscriber retention or return rate for people who do and do not engage with recommendations","Time to first meaningful recommendation for a new user or new title (cold start latency)"],"failureModes":[{"title":"Engagement that is not incremental","detail":"Recommended content also gets promoted placement, so raw clicks overstate the model's effect. Measure against a randomized or interleaved baseline, or model the counterfactual explicitly and validate it with a randomized experiment, rather than against a no recommendation control that never ships."},{"title":"Filter bubbles and catalog concentration","detail":"A model that only optimizes predicted engagement converges on already popular titles. Instrument and target catalog coverage explicitly, not just top line engagement."},{"title":"Silent bias in what gets amplified","detail":"Embeddings learned from historical behavior can encode and reinforce existing skew (for example, under exposing content from smaller creators). Audit exposure by creator or content category, not only by predicted relevance."},{"title":"Cold start dead zones","detail":"New titles and new users get poor recommendations until enough interaction data accumulates, which can suppress exactly the content a catalog most needs to surface. Use metadata based placement and monitor exposure for new content specifically."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Recommendation and personalization systems are not listed in Annex III, so most deployments are minimal risk under the EU AI Act. They become a compliance question elsewhere: manipulative or deceptive techniques that materially distort a person's behavior in a way that causes significant harm, or that exploit vulnerabilities linked to age, disability or a specific social or economic situation, are a prohibited practice under Article 5(1)(a) and (b), which is relevant to recommendation systems that target children. A decision based solely on automated processing, including profiling, that produces legal or similarly significant effects on a person falls under GDPR Article 22, though routine content ranking rarely meets that bar on its own."},"regulations":["eu-ai-act","gdpr","uk-gdpr"],"guidance":[{"title":"EU AI Act Explorer: Article 5, prohibited AI practices","issuer":"Future of Life Institute","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Third party plain language explainer of the AI Act, checked against its Article 5 text. Points 1(a) and 1(b) prohibit manipulative or deceptive techniques and the exploitation of vulnerabilities of a person or group \"due to their age, disability or a specific social or economic situation\", in each case only where they materially distort behavior in a way that causes or is reasonably likely to cause significant harm."},{"title":"Guidelines 05/2020 on consent under Regulation 2016/679","issuer":"European Data Protection Board","region":"europe","url":"https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-052020-consent-under-regulation-2016679_en","note":"Relevant where personalization relies on tracking or profiling that needs a lawful basis."}],"controls":["Inventory entry for the recommendation system with an accountable owner and a documented list of what it may never recommend, and to whom","Regular experiment based measurement of the system's true incremental effect, not only engagement dashboards","Bias and exposure audits by content category and, where relevant, by protected characteristic of the audience segment","Age appropriate design review for any surface reachable by minors"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai is a conversational and agentic AI platform, not a ranking or embeddings platform, so\nthe core recommendation model here is built and served outside it, typically as an existing\ncatalog or personalization service the organization already runs or licenses. Where Blits.ai adds\nvalue is the layer people actually talk to: an **AI agent** with **custom functions** that call\nthe organization's own recommendation API can turn a ranked list into a conversation, explaining\nwhy a title or track was suggested, taking feedback (\"more like this\", \"not interested\"), and\nhandling requests the ranking model cannot, such as \"something short for tonight\" or \"what should\nI watch with my kids\", grounded in a **knowledge base** of the catalog's own metadata through\nhybrid retrieval.\n\nThis kind of recommendation concierge can run as an **agentic workflow** that calls the\nrecommendation API as a tool, respects a **tool execution policy** that limits which catalog\nactions it may take, and is delivered through **web chat, the mobile app's REST or WebSocket API\nchannel, or a digital human** for a more visual browsing experience. **Guardrails** keep the\nconversation inside age appropriate content policy, **analytics** and **response feedback** show\nwhich explanations and conversational picks users rate well, and **test suites** catch\nregressions when the underlying catalog or ranking API changes."},"faq":[{"question":"How do streaming services measure whether their recommendations actually work?","answer":"Raw engagement with recommended content overstates the effect, since recommended items also get more visibility. Netflix researchers, with one academic coauthor, isolated the causal effect with a structural model of viewing choices, validated by a randomized experiment that allocated members into eight treatment arms with different recommendation salience: their modeled counterfactual shows that replacing the current recommender with a simpler popularity based ranking would cut engagement by 12%, with most of the effect coming from effective targeting rather than exposure alone."},{"question":"Is a content recommendation engine high risk under the EU AI Act?","answer":"Usually not; recommendation and personalization are not listed in Annex III, so most deployments are minimal risk. The exception is manipulative or deceptive design that materially distorts behavior and causes significant harm, or that exploits a vulnerability linked to age, disability or a specific social or economic situation, which is a prohibited practice under Article 5, and matters most for surfaces reachable by children."},{"question":"How does a new title or a new user get good recommendations before there is any history?","answer":"This is the cold start problem. Modern systems place new content using its metadata (genre, cast, description) rather than waiting for interaction data, and place new users using onboarding preferences or early signals, blending in more behavioral data as it accumulates."},{"question":"Does personalization mean everyone sees a narrower catalog?","answer":"It can, if the system is optimized purely for predicted engagement, which tends to concentrate recommendations on already popular titles. Teams that also track catalog coverage and the share of engagement going to content that is not already popular, and build in a deliberate exploration budget, reduce this failure mode."}],"related":["personalized-marketing-at-scale","churn-prediction-and-retention-offers"],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-28","note":"First version, researched and written from Netflix's and Spotify's own published sources."},{"date":"2026-09-28","note":"Editorial pass on review findings: corrected the Netflix problem and evidence summaries to match the source blog (dropped \"hundreds\", \"replaced\", search and in app messaging), moved the Netflix stage from scaled to production and dropped the unsupported 300 million users served metric, dropped the two Spotify metrics that did not map to interactions handled, fixed the Spotify summary's personalization claim, corrected the description of Netflix's causal method in the implementation steps, failure modes and FAQ, limited the Blits.ai paragraph to features in the inventory, removed the EU Accessibility Act and inclusion and access value driver, added the \"social or economic situation\" limb to Article 5(1)(b) references, fixed the guidance issuer, tightened the metaDescription and indicativeValue caveat, and narrowed related."},{"date":"2026-09-28","note":"Second editorial pass on an adversarial review: rewrote the Spotify summary so it no longer claims Discover Weekly itself is built from listening history (only the 2025 genre controls are), moved the Netflix stage from production to pilot since no cited source confirms the foundation model serves members and the record's own note flagged this, dropped \"interactions handled\" from the KPIs (streams and artist discoveries are not conversations, cases or documents), rewrote the metaDescription to attribute the 12% figure to Netflix researchers' estimate instead of stating it as fact, tightened the GDPR Article 22 basis to the \"decision based solely on automated processing\" test, replaced \"entire interaction history\" with the blog's own \"comprehensive interaction history\", replaced \"a slice of members\" with the paper's literal \"eight treatment arms\" language in the FAQ, and removed the arXiv sourceUrl from the retention indicative input since that paper does not measure retention."}],"slug":"content-recommendation-and-personalization","url":"https://www.blits.ai/ai-use-cases/content-recommendation-and-personalization","benchmarks":[],"indicativeValueResult":{"low":3000000,"high":22500000},"evidence":["netflix-foundation-model-recommendation","spotify-discover-weekly-personalization"]},{"title":"AI recommendations for loan restructuring and hardship arrangements","shortTitle":"Restructuring recommendations","seoTitle":"AI for loan restructuring and hardship plans","metaDescription":"AI tests term extensions, payment holidays and rate relief against policy and affordability, then recommends the best fit for a hardship specialist to approve.","definition":"An assistant that assembles a stressed borrower's position, tests restructuring options such as a term extension, rate relief, payment holiday or due date change against policy and affordability, and recommends the best fit with a written rationale for a person to approve.","aliases":["forbearance recommendation","hardship arrangement assistant","loan modification recommendation"],"industries":["banking"],"functions":["collections-and-recovery","lending-and-credit","risk-management"],"patterns":["agentic-workflow","rag-knowledge-assistant","document-processing","recommendation-and-personalization"],"channels":["agent-desktop","internal-tools","mobile-app"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","segment":"lending","problem":"When a borrower gets into difficulty the right intervention early is cheaper for everyone than\nenforcement later. But finding it is slow. A hardship or workout specialist has to pull together\nbalances, arrears history, income and expense evidence, collateral and the borrower's own\nexplanation, then work through policy to see which options are allowed and affordable. Queues grow\nexactly when times are hard, decisions vary between specialists, and the reasons are not always\nwritten down.\n\nRegulators expect lenders to treat borrowers in financial difficulty fairly, to choose sustainable\nsolutions over short term fixes that fail, and to document why. Inconsistent or undocumented\nconcessions are a conduct risk and a credit risk at the same time.","problemStats":[],"howItWorks":"1. **Assemble the position.** The assistant gathers balances, arrears, payment history, other\n   exposures, collateral and any income or hardship evidence the customer has provided.\n2. **Read the evidence.** Document AI extracts figures from payslips, bank statements and letters,\n   and flags gaps.\n3. **Test the options.** For each option the policy allows (due date change, payment holiday, term\n   extension, temporary rate relief, capitalisation), it calculates the new payment, the effect on\n   arrears and whether it fits the stated budget.\n4. **Recommend with reasons.** It ranks the options, cites the policy clause behind each, and\n   writes a short rationale and the risks.\n5. **Decide and record.** A specialist accepts, changes or rejects the recommendation; the decision,\n   the reasons and the evidence are stored with the case.\n6. **Follow up.** Review dates are scheduled and the arrangement is monitored for early signs that\n   it is not working.","valueDrivers":["risk-reduction","customer-experience","employee-productivity","compliance"],"kpis":["handling-time-reduction","processing-time-reduction","recovery-rate-uplift"],"indicativeValue":{"referenceOrg":"A retail bank handling 10,000 hardship and restructuring requests a year","inputs":[{"key":"requests","label":"Hardship and restructuring requests per year","low":10000,"high":10000,"unit":"requests per year","note":"The reference bank."},{"key":"hoursSaved","label":"Specialist hours saved per request on assembling the case and testing options","low":0.5,"high":1.5,"unit":"hours per request","note":"Editorial assumption. Replace with your own time study."},{"key":"hourlyCost","label":"Fully loaded cost of a hardship specialist hour","low":45,"high":70,"unit":"USD per hour","note":"Editorial assumption."}],"formula":"requests * hoursSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Specialist time released","caveat":"Counts specialist time only. It leaves out the credit effect of earlier and more sustainable arrangements, fewer broken plans and lower complaint volumes, which can be larger but need a controlled measurement."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The option calculations are deterministic once policy is written down. The effort is in encoding policy, reading hardship evidence reliably and fitting the assistant into the specialist's case workflow.","dataPrerequisites":["Restructuring and hardship policy with eligibility rules per product","Account, arrears and payment history per borrower","Income and expense evidence, or a structured budget from the customer","Outcomes of past arrangements to learn which options last"],"integrations":["Loan servicing and collections systems","Document intake for hardship evidence","Case management for hardship and workout teams","Customer channels for evidence requests and outcome letters"]},"implementation":{"steps":[{"title":"Encode the policy","detail":"Turn the restructuring policy into explicit rules per product: which options, for how long, with which limits and approvals. The assistant can only recommend what the rules allow."},{"title":"Automate the case pack","detail":"Start by assembling the borrower's position and evidence automatically. Specialists gain time even before any recommendation is shown."},{"title":"Add option testing and ranking","detail":"Calculate each allowed option's payment and effect, rank them on affordability and sustainability, and show the calculation behind every number."},{"title":"Measure agreement and outcomes","detail":"Track how often specialists accept the recommendation and how arrangements perform after six and twelve months, by option and segment."},{"title":"Extend to proactive outreach","detail":"Once recommendations are trusted, combine them with early warning signals to offer support before customers fall behind."}],"guardrails":["Recommendations only; every restructure is approved by a person with authority","Options limited to what policy allows, with the policy clause cited","A written rationale stored with every decision","Consistency checks that flag similar cases receiving different outcomes","Vulnerability flags shown prominently and never used to reduce support"],"humanInTheLoop":"Hardship and workout specialists decide every case and can override any recommendation, with a reason. Credit risk approves the rules and reviews arrangement performance; conduct risk reviews consistency and outcomes for vulnerable customers.","kpisToInstrument":["Time from hardship request to decision","Share of recommendations accepted without change","Arrangements still performing after six and twelve months, by option","Complaints about hardship decisions","Outcome differences between comparable customers"],"failureModes":[{"title":"Short term fixes that fail","detail":"The assistant optimises for the lowest payment now and the arrangement breaks later. Rank on sustainability and track long term outcomes."},{"title":"Rubber stamping","detail":"Specialists accept recommendations without reading them. Show the reasoning, sample decisions for review and measure override quality."},{"title":"Evidence misread","detail":"Wrong income or expense figures lead to an unaffordable plan. Show the source document next to each extracted figure."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Recommending restructuring terms for individuals involves assessing their ability to pay, which can amount to evaluating the creditworthiness of natural persons under Annex III point 5(b). Human approval alone does not remove that: the Article 6(3) exception covers only systems that do not materially influence the decision, such as a narrow procedural or preparatory task, and never applies when the system profiles natural persons. A tool that only assembles the case file can fall under the exception; restructuring for companies is outside point 5(b)."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","eba-loan-origination","us-sr-11-7","us-ecoa-reg-b"],"guidance":[{"title":"Guidance to banks on non-performing loans","issuer":"European Central Bank","region":"europe","url":"https://www.bankingsupervision.europa.eu/ecb/pub/pdf/guidance_on_npl.en.pdf","note":"Sets expectations for viable forbearance solutions, borrower affordability assessments and clearly defined, consistent decision making procedures."},{"title":"FG21/1: guidance for firms on the fair treatment of vulnerable customers","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publication/finalised-guidance/fg21-1.pdf","note":"Fair treatment of customers in vulnerable circumstances, relevant to every hardship decision."},{"title":"REP 782 Hardship, hard to get help: Findings and actions to support customers in financial hardship","issuer":"Australian Securities and Investments Commission","region":"asia-pacific","url":"https://asic.gov.au/regulatory-resources/find-a-document/reports/rep-782-hardship-hard-to-get-help-findings-and-actions-to-support-customers-in-financial-hardship/","note":"Review of how 10 large lenders handled home loan customers in financial hardship (May 2024), with good and poor practices; ASIC says the insights are relevant to hardship involving all types of credit."}],"controls":["Policy rules under change control, approved by credit risk","Decision log with recommendation, final decision, override reason and evidence","Periodic consistency review across comparable cases","Outcome monitoring for vulnerable customers"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** started from the hardship case. **Custom functions**\npull balances, arrears and payment history from the servicing system, and hardship evidence the\ncustomer sends as a **file upload** (payslips, statements) is attached to the case. The restructuring\npolicy sits in the **knowledge base**, so the **agent** cites the clause behind each option, while\ndeterministic **custom functions** (JavaScript sandbox) calculate the payments. **Structured\noutput** returns the ranked options and the rationale in a fixed format.\n\nThe recommendation waits for **human in the loop approval** by a specialist, and the **audit\ntrail** stores the run, the recommendation and the decision. Customer facing parts, such as\ncollecting evidence or confirming the arrangement, can run through an **AI agent** in the bank's\napp (through the **API channel**), on WhatsApp or by voice with **human handover** to the specialist. **PII masking** protects the evidence,\nand **test suites** check recommendations against reference cases on every policy change."},"faq":[{"question":"Can AI decide a loan restructure?","answer":"It should recommend, not decide. A person with authority approves every restructure, with the assistant's calculation and rationale in front of them, because hardship cases need judgment and fair treatment."},{"question":"Is anyone using AI to offer hardship support proactively?","answer":"In 2022 Commonwealth Bank said it used its Customer Engagement Engine and a weather data model to reach customers hit by natural disasters with same day support, such as deferring a loan or an emergency overdraft. That is proactive hardship outreach, not a restructuring recommender. A US auto lender's collections agent, as described by its vendor, spots borrowers' pay patterns on the call and triggers a change of due date in the lender's system, and routes hardship cases to its dealerships."},{"question":"What makes a good restructuring recommendation?","answer":"One that the borrower can sustain, that policy allows, and whose reasoning is written down. Track arrangements for six to twelve months to learn which options actually last."}],"related":["collections-and-hardship-agent","credit-early-warning-monitoring","adverse-action-explanations","financial-wellbeing-coach","outbound-notice-drafting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added ECOA and Regulation B to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription, sharpened the EU AI Act basis (Article 6(3)), replaced the ASIC guidance with REP 782 on hardship, corrected the auto lender FAQ answer and the ECB guidance note, aligned the Blits.ai build with the feature inventory."},{"date":"2026-09-27","note":"Removed the interactions handled KPI, so the bank wide decision volume of Commonwealth Bank no longer shows as a restructuring benchmark; dated and narrowed the Commonwealth Bank example to proactive hardship outreach in 2022."}],"slug":"loan-restructuring-recommendations","url":"https://www.blits.ai/ai-use-cases/loan-restructuring-recommendations","benchmarks":[],"indicativeValueResult":{"low":225000,"high":1050000},"evidence":["commonwealth-bank-customer-engagement-engine"]},{"title":"AI regulatory horizon scanning and obligation mapping","shortTitle":"Regulatory horizon scanning","seoTitle":"AI regulatory horizon scanning for compliance","metaDescription":"AI reads regulator publications, flags relevant changes and maps new obligations to controls. Corlytics reports a 25% efficiency gain at a European tier 1 bank.","definition":"An AI system that continuously reads publications from the regulators and standard setters an organization answers to, classifies each item by relevance and urgency, breaks new rules into individual obligations and maps them to the internal policies and controls that meet them, so compliance owners see what changed and where the gaps are.","aliases":["regulatory change management","regulatory intelligence","obligation mapping","regulatory monitoring"],"industries":["cross-industry","banking","insurance","payments","wealth-and-asset-management","pharma-and-life-sciences","government"],"functions":["regulatory-compliance","legal","risk-management"],"patterns":["classification-and-routing","document-processing","rag-knowledge-assistant","summarization","agentic-workflow"],"channels":["internal-tools","email","microsoft-teams"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","segment":"compliance","problem":"A bank operating in a few countries answers to many regulators and standard setters, each\npublishing consultations, rules, guidance, speeches and enforcement actions. Compliance\nteams read feeds and newsletters by hand, decide what is relevant, and then work out which internal\npolicies and controls a new rule touches. The work is repetitive, depends on who is reading, and\nleaves little audit trail of why an item was judged irrelevant.\n\nThe cost of missing something is high: a late implementation, a finding in an examination, or a\nboard that cannot show how it stays current on regulatory change. Answering an examiner or auditor who\nasks how a rule is met takes a traceable line from each obligation to the policy and control that\nmeets it, and that line is hard to keep up to date by hand.","problemStats":[],"howItWorks":"1. **Collect.** The system monitors regulator websites, official journals, standard setters and\n   enforcement publications for every jurisdiction in scope.\n2. **Classify.** Each item is tagged by jurisdiction, topic, document type, business line and\n   urgency, and irrelevant items are filtered with a recorded reason.\n3. **Summarise.** Relevant items get a short summary, key dates and what is new compared with the\n   previous version or consultation.\n4. **Extract obligations.** Final rules are broken into individual obligations in a consistent\n   structure (who must do what, by when).\n5. **Map to the library.** Each obligation is matched to existing policies and controls through\n   retrieval over the internal obligation and control library, and unmatched or partly matched\n   obligations are flagged as gaps.\n6. **Route and record.** Items go to the owner of the affected area, who confirms materiality and\n   accepts, changes or rejects each mapping; the decision and reasoning are kept.","valueDrivers":["compliance","employee-productivity","risk-reduction"],"kpis":["productivity-gain","hours-saved","time-saved-per-task","interactions-handled"],"indicativeValue":{"referenceOrg":"A bank monitoring regulatory change across 10 jurisdictions","inputs":[{"key":"monitoringFte","label":"Full time staff spent on monitoring and first assessment of regulatory change","low":6,"high":12,"unit":"full time equivalents","note":"Editorial assumption. Replace with your own team size."},{"key":"efficiency","label":"Share of that effort saved","low":0.15,"high":0.2,"unit":"fraction of effort","note":"The high end is derived from the only published efficiency figure on this page, a 25% efficiency gain reported by a vendor, in the benefits section of a case study, for the compliance monitoring and compliance risk insight teams of an anonymous European tier 1 bank. A 25% efficiency (output per unit of effort) increase corresponds to about 20% of effort released, so it does not match the source figure directly. Treat it as a ceiling, not a typical result. The low end is an editorial assumption for teams that keep more manual review."},{"key":"costPerFte","label":"Fully loaded cost per compliance analyst","low":120000,"high":180000,"unit":"USD per year","note":"Editorial assumption."}],"formula":"monitoringFte * efficiency * costPerFte","currency":"USD","period":"per year","resultLabel":"Compliance monitoring effort released","caveat":"Counts analyst effort only. It leaves out licence and content costs, the cost of building and maintaining the obligation library, and the harder to price value of fewer missed changes and a cleaner audit trail for supervisors."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Monitoring and classification are mature, and commercial regulatory content feeds exist. The hard part is a clean internal library of policies and controls to map against, and owners who review the mappings.","dataPrerequisites":["A list of regulators, jurisdictions and topics in scope","A policy and control library with owners, ideally already linked to obligations","A taxonomy of business lines, products and risk types","Historical regulatory change decisions to test classification against"],"integrations":["Regulatory content feeds or website monitoring","Governance, risk and compliance (GRC) platform","Policy management system","Collaboration tools for routing and sign off"]},"implementation":{"steps":[{"title":"Define scope and relevance","detail":"List jurisdictions, regulators and topics, and write down what makes an item relevant for each business line. That definition is what the classifier is tested against."},{"title":"Clean the obligation and control library","detail":"Mapping only works against a library that is current, owned and consistently written. Fix the library before automating the mapping."},{"title":"Run in parallel with the manual process","detail":"For a quarter, let the system classify and map alongside the team and compare: what it missed, what it flagged that people missed, and where mappings differ."},{"title":"Route to owners with the evidence","detail":"Send each relevant item to the accountable owner with the summary, the proposed mappings and the source, and require a recorded decision."},{"title":"Report to management and the board","detail":"Use the recorded decisions to show open changes, gaps and implementation status per regulator and business line."}],"guardrails":["A named owner confirms materiality and accepts or overrides every mapping","Every filtered out item keeps its reason, so exclusions can be audited","Summaries always link to the official source text","Obligations are extracted from final texts, not from secondary commentary"],"humanInTheLoop":"Compliance owners confirm relevance and materiality, approve obligation mappings and decide on gaps. Legal interprets ambiguous rules. The system proposes; people decide, and their reasoning is retained for supervisors.","kpisToInstrument":["Share of relevant items found within a set number of days of publication","Items missed by the system but found by people, and the reverse","Owner agreement rate with proposed mappings","Analyst hours per week on monitoring and first assessment","Open gaps and their age"],"failureModes":[{"title":"Silent misses","detail":"A source changes its website or feed and nothing arrives. Monitor each source's volume and alert when it drops to zero."},{"title":"Confident but wrong mappings","detail":"The system maps an obligation to a control that sounds similar but does not meet it. Require owner sign off and sample accepted mappings."},{"title":"Summaries treated as the rule","detail":"Staff act on a summary that missed a qualification. Always link to and quote the source text."}]},"risk":{"euAiAct":{"tier":"limited","basis":"An internal tool that monitors and classifies regulatory publications for staff makes no decisions about natural persons, so it is not listed in Annex III and is not a prohibited practice under Article 5. Staff know they are using an AI tool and its summaries are not published to the public, so the Article 50 duties to inform users and to disclose published generated text add little for the deploying organization. Article 50(2) still requires the provider of a system that generates text to mark its output, in a machine readable format, as AI generated: usually the vendor, but an organization that builds its own summariser can itself be that provider, which is what puts this use case at the limited tier rather than minimal. Beyond this and AI literacy (Article 4), no specific obligations apply. General model risk and third party rules still apply."},"regulations":["dora","iso-42001","nist-ai-rmf","mas-ai-risk-management","apra-cps-230"],"guidance":[{"title":"MAS Guidelines for Artificial Intelligence Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Proposed in a consultation paper of 13 November 2025 that closed on 31 January 2026; no final Guidelines were listed on the consultation page when checked on 27 September 2026. The proposed Guidelines will apply to all financial institutions, cover different AI applications and technologies, including generative AI and AI agents, and expect controls proportionate to the assessed risk materiality of each AI use."},{"title":"AI Risk Management Framework","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/itl/ai-risk-management-framework","note":"Voluntary framework to govern, map, measure and manage the risks of AI systems, including generative AI."}],"controls":["Documented source list with monitoring of each source's availability","Decision log of relevance, materiality and mapping decisions with owners and dates","Periodic sample review of excluded items and accepted mappings","Inventory entry for the tool with its intended use and limitations"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the scanning runs as scheduled **agentic workflows**. The **knowledge base** crawls and\nrecrawls regulator web pages, ingests PDFs and email alerts through the **incoming email** source,\nand holds the internal policy and control library for **hybrid retrieval**. An **agent** with\n**structured output** classifies each new item, summarises it, extracts obligations and proposes\nmappings to the library with citations to the source text. **Custom functions**, ready made\ntools such as SharePoint, or systems from the integration catalog such as ServiceNow write the\nresults to the GRC or policy system.\n\nOwners receive items by email or in **Microsoft Teams**. Set the **human in the loop** threshold\nof the workflow so that every mapping waits for an owner to approve or reject it before it is\nrecorded. The **run history and audit trail** keep every classification and\ndecision, **monitors** run scheduled checks on the agent and alert by email or webhook when it\nfails, and **test suites** with LLM based\ngrading check classification against past decisions. Compliance staff can ask questions about the\nlibrary through an **agent** that answers from the **knowledge base** with retrieval augmented\ngeneration. The platform is model agnostic and can run in the EU\nor UAE region."},"faq":[{"question":"Can AI replace a regulatory change team?","answer":"No. It removes the reading and first sorting, and proposes obligation mappings, but relevance, materiality and interpretation stay with compliance owners and legal. The value is coverage, speed and an audit trail."},{"question":"What results have organizations reported?","answer":"Published figures are scarce and come from vendors. Corlytics reports a 25% efficiency gain for the compliance monitoring and compliance risk insight teams of an anonymous European tier 1 bank, stated in the benefits section of its case study rather than as a measured outcome. Other deployments on this page, such as the UK FCA Intelligent Handbook, where a machine learning framework developed with Corlytics auto tags Handbook content with review and a full audit trail, and the US Administration for Children and Families review of documents against new directives, describe their benefits in words only."},{"question":"Is regulatory horizon scanning high risk under the EU AI Act?","answer":"No, it is not a high risk Annex III use. It is limited risk: an organization that builds its own summariser can be the provider of a text generating system under Article 50(2), which requires marking its output as AI generated. Treat it as a model with an owner, documented limits and human sign off on every mapping."}],"related":["policy-drafting-and-gap-analysis","continuous-controls-testing","regulatory-report-assembly","supervisory-exam-response-assembly","marketing-content-compliance-copilot","ai-model-inventory"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI catalog as an industry neutral page and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: industries now include government, where its evidence comes from; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; softened unsourced claims in the problem; corrected the FCA and ACF wording in the FAQ; MAS guidance marked as a November 2025 consultation; EU AI Act basis now cites Articles 4, 5 and 50; Blits.ai build text aligned with the feature inventory; Corlytics product page added as the AI source for two case studies."},{"date":"2026-09-27","note":"Adversarial review: removed the Arbuthnot Latham record, whose regulatory monitoring text on the case study was copied from another client; removed the FCA 18,000 provisions metric, which is rulebook scope, not AI throughput; FAQ now states that the 25% figure is a vendor benefit claim and describes the FCA tagging as the source does; MAS note matches the press release and records the consultation status; EU AI Act basis now covers Article 50(2); human in the loop approval described as a configured threshold."},{"date":"2026-09-27","note":"Fact checked against sources: all quotes, dates, the MAS consultation status and the regulation ids confirmed; the 25% figure now names both teams the vendor cites; the tier 1 bank and insurer summaries follow the case study wording more closely; the Blits.ai build text now names only inventory capabilities."}],"slug":"regulatory-horizon-scanning","url":"https://www.blits.ai/ai-use-cases/regulatory-horizon-scanning","benchmarks":[],"indicativeValueResult":{"low":107999.99999999999,"high":432000.00000000006},"evidence":["financial-conduct-authority-intelligent-handbook","hhs-acf-directive-alignment-document-review"]},{"title":"AI reply drafting for customer email and support tickets","shortTitle":"Email and ticket reply drafting","seoTitle":"AI reply drafting for support emails and tickets","metaDescription":"AI drafts replies to customer emails and tickets for agents to check and send. Google Cloud reports that Turing cut its HR ticket processing time by a third.","definition":"A copilot for asynchronous service work that drafts the reply to an incoming customer email, message or ticket once it has reached an agent: it summarizes the request, pulls the relevant customer data and approved knowledge, and drafts a reply in the organization's tone and the customer's language for the agent to check, edit and send. Live calls and chats, and the sorting of the inbox itself, are separate use cases.","aliases":["AI email response assistant","ticket reply suggestions","suggested replies for service agents","AI drafted customer replies","written channel agent assist"],"industries":["cross-industry","government","banking","telecommunications","technology"],"functions":["customer-service","operations"],"patterns":["content-generation","summarization","rag-knowledge-assistant","classification-and-routing"],"channels":["email","agent-desktop","social-messaging","whatsapp"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"mainstream","problem":"Written service channels are slow in a way phone is not. An email or ticket waits in a queue, an\nagent opens it, reads the history, looks up the account, searches the knowledge base, writes a\nreply and often has to ask a colleague. Much of that effort is repeated for questions that have\nbeen answered many times before, and the reply quality depends on who happens to pick the ticket\nup. Backlogs build after product changes and incidents, and response time targets slip.\n\nFully automated replies are tempting but risky for anything beyond simple, well understood\nrequests: a confident wrong answer in writing is a record the customer can forward. A common\nmiddle ground is a drafted reply that the agent owns. The AI does the reading, the lookup and the\nfirst draft; the agent brings judgment, corrects and sends. Real time assist during calls and\nchats is covered on its own page; this page is about the asynchronous queue.","problemStats":[],"howItWorks":"1. **Read and summarize.** When a message or ticket arrives, the AI summarizes the request and\n   the thread, detects the intent, language and urgency, and labels it for routing.\n2. **Gather context.** It retrieves the customer's relevant data (orders, account status, open\n   cases) through approved system calls and the relevant passages from the approved knowledge base.\n3. **Draft the reply.** It writes a response in the house tone and the customer's language that\n   addresses every point raised, cites the knowledge it used and leaves placeholders where it lacks\n   information.\n4. **Agent reviews and sends.** The agent edits, completes and sends. Drafts for regulated\n   topics, complaints or vulnerable customers are marked for extra care.\n5. **Automate only the safe tail.** For a narrow set of simple, low risk intents, the organization\n   may send replies automatically, with disclosure and sampling.\n6. **Learn from edits.** The difference between draft and sent reply is logged to improve prompts\n   and to find missing knowledge.","valueDrivers":["employee-productivity","cost-to-serve","customer-experience","speed"],"kpis":["handling-time-reduction","processing-time-reduction","response-time-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A service team that answers 400,000 emails and tickets a year","inputs":[{"key":"tickets","label":"Emails and tickets answered per year","low":400000,"high":400000,"unit":"tickets per year","note":"The reference organization."},{"key":"minutesPerTicket","label":"Agent handling time per ticket today","low":6,"high":10,"unit":"minutes per ticket","note":"Editorial assumption for written service. Replace with your own handling time."},{"key":"reduction","label":"Reduction in handling time with drafted replies","low":0.2,"high":0.33,"unit":"fraction of handling time","note":"Editorial assumption, replace with your own. No source on this page measures agent handling time per email or ticket: Google Cloud reports Turing's HR ticket processing time down by 33%, without defining what that processing time covers, and Microsoft reports HYPE's agents resolving WhatsApp chat conversations in half the time. The range is set at or below both."},{"key":"costPerMinute","label":"Fully loaded agent cost per minute","low":0.6,"high":0.9,"unit":"USD per minute","note":"Editorial assumption, replace with your own."}],"formula":"tickets * minutesPerTicket * reduction * costPerMinute","currency":"USD","period":"per year","resultLabel":"Agent handling cost released","caveat":"Counts only agent time. It leaves out faster responses and more consistent quality, the cost of the platform, and any effect of automating simple replies end to end. Savings usually show as backlog cleared and capacity redeployed rather than immediately lower cost."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Drafting from a knowledge base is quick to prototype. The work is in reading customer data through proper integrations, keeping knowledge current, handling many languages and designing which topics may never be answered without extra review.","dataPrerequisites":["Approved, current knowledge articles and reply templates with owners","A contact reason taxonomy with volumes for email and ticket channels","Historical tickets with the replies that resolved them, for testing","Customer and order data reachable through APIs"],"integrations":["Ticketing or case management platform (for example Zendesk, Salesforce, Freshdesk, ServiceNow)","Email and messaging channels","CRM, order and account systems for context","Knowledge base","Quality assurance tooling for sampling sent replies"]},"implementation":{"steps":[{"title":"Pick intents by volume and risk","detail":"From the contact reason report, choose high volume topics where the answer is in approved knowledge. Mark complaints, legal and regulated topics for extra review from day one."},{"title":"Ground every draft","detail":"Draft only from approved knowledge and data from system calls, show the agent the sources, and leave explicit gaps instead of guessing when information is missing."},{"title":"Put the draft where agents work","detail":"Embed summaries and drafts in the existing ticket view. A separate tool that needs copy and paste loses most of the gain."},{"title":"Measure edits, not only speed","detail":"Track how much agents change each draft and why. Heavy edits point at missing knowledge or a bad prompt; no edits on complex topics may point at over reliance."},{"title":"Test in every language you serve","detail":"Build a test set of real tickets per language and intent, and run it on every prompt, model or knowledge change."},{"title":"Automate the safe tail last","detail":"Only after months of low edit rates on an intent consider sending automatically, with AI disclosure, sampling and an easy route to a person."}],"guardrails":["The agent sends; no reply leaves without human action unless the intent is on an approved automation list","Drafts use only approved knowledge and data from authorized system calls, with sources shown","Complaints, legal threats and signs of vulnerability are flagged and routed, not just answered","Personal and payment data masked in prompts and logs","Regulated statements (fees, rights, deadlines) come from approved templates, not free generation"],"humanInTheLoop":"Agents review and own every reply they send. Team leads sample sent replies weekly for accuracy and tone, knowledge owners fix the gaps that heavy edits reveal, and any move to automatic sending for an intent needs sign off from the service owner and compliance.","kpisToInstrument":["Handling time per ticket by intent, before and after, on the same case mix","Draft acceptance rate and edit distance per intent","Reopen and repeat contact rate within seven days","Quality assurance score of sent replies","Time to first response and backlog age"],"failureModes":[{"title":"Rubber stamping","detail":"Agents send drafts unread under time pressure. Sample sent replies and watch for near zero edit rates on complex topics."},{"title":"Fluent but wrong","detail":"A plausible reply built on outdated knowledge. Show sources, refuse when retrieval finds nothing and keep knowledge owned and current."},{"title":"Missed complaint","detail":"A complaint is answered as a routine question and never logged. Detect complaint language and route it to the complaints process."},{"title":"Data from the wrong customer","detail":"Context pulled for a similar name or a shared email address. Match on verified identifiers only and show the agent what was used."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A drafting copilot whose output an agent reviews and sends falls under the transparency tier at most. When replies are sent without human review, customers interact with the AI system directly, and Article 50(1) requires that they are informed unless this is obvious from the context. Article 50(2) separately requires the provider of a system that generates text to mark its output in a machine readable format as artificially generated, whether or not a person reviews the draft. It becomes high risk only if it is used for a purpose listed in Annex III, such as evaluating eligibility for public benefits or creditworthiness (point 5), or evaluating the performance of the agents who use it (point 4)."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty"],"guidance":[{"title":"Regulation (EU) 2024/1689 (AI Act), Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj","note":"Official text on EUR-Lex. Article 50(1) is relevant when replies are sent automatically, so the customer interacts with the AI system directly; Article 50(2) covers machine readable marking of generated text by the provider."}],"controls":["Documented list of intents eligible for drafting and, separately, for automatic sending","Logging of draft, edits and sent version for quality review and disputes","Weekly quality sampling with feedback into knowledge and prompts","Personal data masking and retention limits on drafts and logs","Complaint detection linked to the complaints handling process","For UK retail financial services firms, drafted replies checked against the consumer understanding outcome of the FCA Consumer Duty (PRIN 2A.5) before they are sent"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the **email channel** (inbound and outbound) or a **REST API** call from the ticketing\nplatform sends each message to an **AI agent** that summarizes it, classifies the intent with\n**structured output** and drafts a reply from the **knowledge base** with hybrid retrieval.\n**Custom functions** read order, account or case data from the systems of record (the integration\ncatalog includes Zendesk, Salesforce and ServiceNow, and Freshdesk is a ready made tool), and\n**PII masking** at the gateway keeps personal and card data out of prompts.\n\nWhen the ticketing platform calls the agent through the API, the draft returns to the ticket for\nthe agent to review and send. The email channel path suits the intents you approve for\nautomation: an **agentic workflow** sends the reply, with **human in the loop** approval above a\nthreshold you set. **Guardrails** check every draft for tone and policy, **multi language**\nsupport answers in the customer's language, and **test suites** replay real tickets before each\nchange goes live. The platform is\nmodel agnostic and can run in the EU or UAE region."},"faq":[{"question":"How much time does AI reply drafting save?","answer":"The figures on this page come from vendors and measure different things. Google Cloud reports that Turing cut its HR ticket processing time by 33% with a model that drafts replies, and Microsoft reports that HYPE's human agents resolve WhatsApp conversations in half the time with Copilot case and conversation summaries and email assistance. It is unclear whether either figure measures agent handling time per email or ticket, so measure your own baseline before relying on either."},{"question":"Should AI replies be sent automatically?","answer":"Only for a narrow set of simple, low risk intents after a period of drafts that agents barely change, and with AI disclosure. Everything else is better drafted by the AI and sent by an agent. TSA's AskTSA service, for example, uses AI to summarize inquiries and recommend replies to the human agents who answer them."},{"question":"How is this different from live agent assist?","answer":"Live agent assist works in real time during a call or chat. Reply drafting works on the asynchronous queue of emails, tickets and messages, where the agent has time to review and the draft is the main output."}],"related":["correspondence-triage-and-routing","live-agent-assist","support-knowledge-article-generation","complaints-handling-agent","civil-servant-drafting-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical with five evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: attributed benchmarks to their vendor claimants, softened unsupported wording in the problem and FAQ, made the EU AI Act basis cite Article 50(1) and Annex III points 4 and 5, corrected the integration list in the Blits.ai section, added an AskTSA channel source and set the SEO title and description."},{"date":"2026-09-27","note":"Second fact check: the description names Google Cloud as the source of the Turing figure; the FAQ and the handling time assumption now say what each vendor figure measures; the EU AI Act basis adds Article 50(2) and cites EUR-Lex; the Consumer Duty is scoped to UK retail financial firms; the Blits.ai section separates the review path from the automation path."}],"slug":"email-and-ticket-reply-drafting","url":"https://www.blits.ai/ai-use-cases/email-and-ticket-reply-drafting","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":41.5,"min":33,"max":50,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"hype-customer-service-email-and-chat-assist","pooled":true},{"id":"turing-hr-ticket-reply-drafting","pooled":true}]}],"indicativeValueResult":{"low":288000,"high":1188000},"evidence":["cdc-smartfind-knowledge-bot","first-national-bank-copilot-for-sales","hype-customer-service-email-and-chat-assist","nomad-esim-support-ticket-replies","tsa-asktsa-response-assist","turing-hr-ticket-reply-drafting"]},{"title":"AI roleplay training for customer conversations","shortTitle":"Conversation roleplay training","seoTitle":"AI roleplay training for sales and service teams","metaDescription":"AI roleplay lets staff rehearse hard calls with a simulated customer and get scored feedback. Bank of America staff completed over 1 million simulations in 2024.","definition":"A training simulator in which generative AI plays a realistic customer, by voice or text, so service, sales and crisis staff can rehearse difficult conversations as often as they need before they handle live ones, and receive structured feedback against the organization's own standards.","aliases":["AI roleplay","conversation simulator","AI sales roleplay","simulated customer training","AI practice calls"],"industries":["cross-industry","banking","insurance","telecommunications","government","healthcare"],"functions":["human-resources","customer-service","sales"],"patterns":["conversational-agent","voice-agent","content-generation"],"channels":["internal-tools","voice","web-chat"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","problem":"New contact centre, sales and crisis line staff learn the hardest conversations on real\ncustomers: the angry caller, the fraud victim, the customer in financial hardship, the person in\ncrisis. Classroom roleplay with colleagues or trainers is limited by trainer time, feels\nartificial and rarely covers the full range of situations, so new hires often reach the floor with\nlittle practice. The risk is long ramp up times, inconsistent handling of disclosures and\nvulnerability, and avoidable harm to the first customers each new hire serves.\n\nRegulated firms have an extra reason to care. Conduct rules expect staff to recognise\nvulnerability, give required disclosures and treat customers fairly; in the UK, the FCA's guidance\non vulnerable customers asks firms to ensure frontline staff have the skills and capability to\nrecognise and respond to vulnerability. Trainer led roleplay leaves little evidence of what was\npractised and how well.","problemStats":[],"howItWorks":"1. **Build scenarios from real work.** Training and quality teams write scenarios from real,\n   anonymized contact reasons: a disputed charge, a lost card abroad, a hardship request, a\n   complaint, a sales conversation with required disclosures. Each scenario has a persona, a goal,\n   facts the trainee must find out and behaviours to test.\n2. **The AI plays the customer.** A model plays the persona by voice or text, reacts to what the\n   trainee says, becomes calmer or more upset depending on how the conversation goes, and raises\n   the objections or cues the scenario calls for.\n3. **Score against the rubric.** After the conversation the AI scores the transcript against the\n   organization's rubric (verification steps, required disclosures, empathy, accuracy of\n   information, next steps) and quotes the moments behind each score.\n4. **Give targeted feedback and repeat.** The trainee gets specific feedback and can retry the\n   same scenario or a harder variant immediately.\n5. **Report to trainers, not to discipline.** Trainers see progress per skill and per cohort and\n   spend their time coaching where the simulator shows gaps.","valueDrivers":["employee-productivity","customer-experience","compliance","speed"],"kpis":["time-to-proficiency-reduction","interactions-handled","conversion-rate-uplift","quality-score-uplift","users-served"],"indicativeValue":{"referenceOrg":"A contact centre that hires 200 new agents a year","inputs":[{"key":"hires","label":"New agents trained per year","low":200,"high":200,"unit":"agents per year","note":"The reference organization. Replace with your own hiring volume."},{"key":"rampWeeks","label":"Weeks from start to full proficiency today","low":6,"high":10,"unit":"weeks","note":"Editorial assumption. Replace with your own ramp time."},{"key":"rampReduction","label":"Share of ramp time removed by simulated practice","low":0.1,"high":0.3,"unit":"fraction of ramp time","note":"Conservative against the evidence on this page (GoHealth's vendor reports onboarding cut from nine weeks to four, a 55% saving), because that figure is a single vendor reported case."},{"key":"productivityGap","label":"Productivity shortfall of a new agent during ramp up","low":0.3,"high":0.5,"unit":"fraction of a fully proficient agent","note":"Editorial assumption."},{"key":"weeklyCost","label":"Fully loaded weekly cost of an agent","low":900,"high":1500,"unit":"USD per week","note":"Editorial assumption. Replace with your own cost."}],"formula":"hires * rampWeeks * rampReduction * productivityGap * weeklyCost","currency":"USD","period":"per year","resultLabel":"Value of productive time gained by faster ramp up","caveat":"Ramp up value only. It leaves out trainer time saved, lower early attrition, fewer complaints and conduct breaches from new hires, and the licence and scenario authoring costs of the simulator."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"No integration with customer systems is needed. The effort is in writing good scenarios and rubrics with the quality and compliance teams, calibrating the scoring against human assessors and making voice latency low enough to feel like a real call.","dataPrerequisites":["The contact reason report, to choose scenarios by volume and risk","The quality assurance rubric and required disclosures per conversation type","Anonymized example transcripts or call recordings for realistic personas","Vulnerability and complaint handling policies the scenarios must test"],"integrations":["Learning management system for assignments and completion records","Single sign on for trainees and trainers","Voice or telephony softphone for realistic practice calls"]},"implementation":{"steps":[{"title":"Pick scenarios by risk and volume","detail":"Start with five to ten scenarios that new hires find hardest and that carry conduct risk, such as a hardship request, a scam victim or a complaint. Add routine ones later."},{"title":"Write rubrics with quality and compliance","detail":"Use the same rubric the quality team uses on live calls, so practice and assessment measure the same things. Mark which items are mandatory, such as identity checks and disclosures."},{"title":"Calibrate scoring against humans","detail":"Have experienced assessors score a sample of simulated conversations and compare with the AI scores. Fix rubric items where they disagree before trainees see scores."},{"title":"Keep personas realistic, not cruel","detail":"Let the persona escalate and de escalate in response to the trainee, but keep abuse within what staff actually meet, and give trainees a way to stop a session."},{"title":"Blend into the programme","detail":"Integrate practice into training weeks rather than bolting it on, as GoHealth did by folding practice calls into training, and let trainers use the reports to target coaching."},{"title":"Measure on the floor","detail":"Compare ramp time, quality scores and complaint rates of trained cohorts with earlier cohorts on the same contact mix."}],"guardrails":["Scores are used for practice and coaching, not for promotion, pay or termination decisions without human review","No inference of trainees' emotions from voice or face, which the EU AI Act prohibits in the workplace outside medical or safety reasons","Scenarios and model answers use only approved policies, products and disclosure wording","No real customer data in personas; examples are anonymized before they become scenarios","Trainees can see their transcripts and scores and contest a score with a trainer"],"humanInTheLoop":"Trainers and quality leads own the scenarios and rubrics, review the AI's scoring on a sample every cohort, and make every certification or sign off decision. The AI gives practice and feedback; a human decides whether someone is ready for live customers.","kpisToInstrument":["Time to proficiency per cohort, before and after","Practice sessions per trainee and scenario coverage","Agreement between AI scores and human assessor scores on a sample","Live quality scores and complaint rates in the first three months on the floor","Trainee rating of realism and usefulness"],"failureModes":[{"title":"Scoring that trainees do not trust","detail":"Scores that disagree with what trainers say undermine the tool. Calibrate against human assessors and show the transcript evidence behind each score."},{"title":"Unrealistic customers","detail":"Personas that are too easy or cartoonishly hostile teach the wrong lessons. Build personas from real contact reasons and review them with experienced agents."},{"title":"Training on outdated policy","detail":"The simulated conversation rewards a disclosure or process that has changed. Tie scenarios to the policy owner and review them when policy changes."},{"title":"Practice data used as surveillance","detail":"Using practice scores in performance management kills honest practice and can make the system high risk. Keep practice and performance evaluation separate by design."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Used only for practice and feedback, the simulator is limited risk. Article 50 requires that people know they are interacting with AI unless that is obvious from the context, as it usually is in a training session, and the provider must mark synthetic voice or text output as AI generated in a machine readable format. It becomes high risk under Annex III point 4(b) if its scores are used to evaluate the performance of workers or to decide on their promotion or termination, and can fall under point 3(b) when a vocational training institution uses it to evaluate learning outcomes. Inferring trainees' emotions from voice or face in the workplace is prohibited under Article 5(1)(f), except for medical or safety reasons."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","iso-42001"],"guidance":[{"title":"Annex III: High risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) covers AI used to monitor and evaluate the performance and behaviour of persons in work related relationships, which is where training scores can end up. Point 3(b) covers AI that evaluates learning outcomes in educational and vocational training institutions."},{"title":"Article 5: Prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Point (f) prohibits AI that infers the emotions of a natural person in the workplace or in education institutions, except for medical or safety reasons, which rules out scoring trainees' emotions from their voice or face."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Requires that people are told they are interacting with AI unless it is obvious from the context, and that providers mark synthetic audio and text output as artificially generated."},{"title":"FG21/1: guidance for firms on the fair treatment of vulnerable customers","issuer":"Financial Conduct Authority","region":"europe","url":"https://www.fca.org.uk/publication/finalised-guidance/fg21-1.pdf","note":"Asks UK financial services firms to ensure frontline staff have the skills and capability to recognise and respond to customers in vulnerable circumstances, which scenario practice can support."}],"controls":["Written purpose limitation that keeps practice scores out of performance management","Periodic calibration of AI scores against human assessors, with results recorded","Scenario and rubric change control owned by training and compliance","Trainee notice of how transcripts and scores are stored, who sees them and for how long","Inventory entry for the simulator with an accountable owner"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai each scenario is an **AI agent** with a persona prompt, a goal and the facts it may\nreveal, with **prompt versioning** so training teams can refine scenarios safely. Trainees\npractise by **voice** with streaming speech recognition and synthesis, including **emotion\naware TTS** so the simulated customer sounds calmer or more upset as the conversation develops,\nor in **web chat** and **Microsoft Teams**; a **digital human** can add a face for in person\ntraining rooms. A **knowledge base** holds the approved policies and disclosure wording the\nrubric checks against.\n\nAfter each session a second agent scores the transcript with **structured output** against the\nrubric, quoting the evidence per item. **Test suites** with **LLM based grading** keep scenarios\nand scoring consistent across changes, **conversation logs** give trainers the transcripts,\nand **role based access control** limits who sees individual results. The platform is model\nagnostic, so the persona and the grader can run on different models."},"faq":[{"question":"Does AI roleplay actually shorten ramp up time?","answer":"Public evidence is early and mostly vendor reported. GoHealth's vendor reports onboarding cut from nine weeks to four after folding AI practice into training, and Bank of America reports more than one million simulations completed by employees in 2024. Measure ramp time and live quality on your own cohorts before and after."},{"question":"Is AI roleplay training high risk under the EU AI Act?","answer":"Not when it is used only for practice and feedback; then the Article 50 transparency rules apply. It becomes high risk if its scores are used to evaluate employees' performance or decide on promotion or termination (Annex III point 4(b)), and inferring trainees' emotions in the workplace is prohibited (Article 5(1)(f))."},{"question":"Can it be used for sensitive conversations such as crisis calls?","answer":"Yes, with care. The US Department of Veterans Affairs trains new Veterans Crisis Line responders on AI simulations with eight Veteran personas, and each simulated call produces a scoring summary of strengths and areas of growth. Scenarios for sensitive topics need expert review and a way for trainees to stop a session."}],"related":["call-quality-and-compliance-monitoring","live-agent-assist","sales-call-coaching-and-crm-update","first-line-contact-centre-agent","employee-onboarding-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog, made industry neutral and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: GoHealth pilot figure moved to conversion uplift, unconfirmed voice channel removed from GoHealth and VA records, page KPIs aligned with the evidence, Article 50 and Annex III point 3(b) added to the risk basis, FCA FG21/1 cited for the staff skills statement, SEO title and description added."}],"slug":"conversation-roleplay-training","url":"https://www.blits.ai/ai-use-cases/conversation-roleplay-training","benchmarks":[{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":21,"min":21,"max":21,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"gohealth-ai-roleplay-sales-training","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":1000000,"min":1000000,"max":1000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bank-of-america-academy-conversation-simulators","pooled":true}]},{"kpi":"time-to-proficiency-reduction","label":"Time to proficiency reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":55,"min":55,"max":55,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"gohealth-ai-roleplay-sales-training","pooled":true}]}],"indicativeValueResult":{"low":32400,"high":450000},"evidence":["bank-of-america-academy-conversation-simulators","gohealth-ai-roleplay-sales-training","veterans-crisis-line-reflexai-training-simulations"]},{"title":"AI sales call coaching and CRM update","shortTitle":"Sales call coaching and CRM update","seoTitle":"AI sales call coaching and automatic CRM updates","metaDescription":"AI analyses sales calls to coach sellers and proposes CRM updates. Hughes cut call audit costs by 90%; Sandvik sellers save three minutes per Outlook lookup.","definition":"AI for sales teams that analyses sales calls and meetings against the team's own sales method to coach sellers and their managers, and writes the call summary, next steps and opportunity updates into the CRM for the seller to confirm. Its purpose is winning deals and building selling skill, not the regulated advice record or general meeting notes.","aliases":["conversation intelligence for sales","revenue intelligence","AI sales coaching","automatic CRM logging","sales call analysis"],"industries":["cross-industry","telecommunications","manufacturing","insurance"],"functions":["sales"],"patterns":["speech-analytics","summarization","content-generation"],"channels":["voice","microsoft-teams","email","internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Sellers spend most of their week on work that is not selling: writing up calls, updating\nopportunities, logging contacts, searching past emails before a meeting. The CRM suffers first.\nNotes are short, late or missing, stages and next steps are out of date, and forecasts are built on\nwhat sellers remembered to type. Managers coach from the few calls they join and from pipeline\nreports that do not show what was actually said.\n\nAI can take over much of the administration and make coaching evidence based. Call and meeting transcripts\nbecome summaries, next steps and CRM updates the seller confirms in one step. Across many calls,\nthe same analysis shows where deals stall, which questions top performers ask, and where a seller\nneeds help. The sensitive part is the second one: once calls are analysed to judge individual\nsellers, it is worker monitoring, with legal limits and a trust cost if handled badly.","problemStats":[{"statement":"Salesforce's State of Sales survey of 7,775 sales professionals found that reps spend 28% of their week actually selling, with most of their time taken by tasks such as deal management and data entry.","sourceTitle":"New Research Reveals Sales Reps Need a Productivity Overhaul, Spend Less than 30% Of Their Time Actually Selling","sourceUrl":"https://www.salesforce.com/news/stories/sales-research-2023/","year":2023},{"statement":"The same Salesforce survey found that only 26% of sales professionals receive one to one coaching at least weekly.","sourceTitle":"New Research Reveals Sales Reps Need a Productivity Overhaul, Spend Less than 30% Of Their Time Actually Selling","sourceUrl":"https://www.salesforce.com/news/stories/sales-research-2023/","year":2023}],"howItWorks":"1. **Capture with consent.** Calls and online meetings are recorded and transcribed only where the\n   participants have been informed, and customers can decline.\n2. **Summarize and extract.** The AI writes a summary, the customer's stated needs and objections,\n   agreed next steps, and any changes to contacts, stage, amount or close date.\n3. **Propose the CRM update.** The proposed changes appear next to the opportunity for the seller to\n   confirm or correct, instead of being typed from memory.\n4. **Draft the follow up.** A recap email to the customer with the agreed next steps is drafted for\n   the seller to edit and send.\n5. **Coach against the method.** Across calls, the AI marks moments linked to the team's sales\n   method (discovery questions, next step agreed, pricing discussed) and surfaces examples, for the\n   seller's own review and for coaching conversations with their manager.\n6. **Improve the playbook.** Aggregated, anonymized patterns show which objections are rising and\n   which approaches work, feeding training and enablement content.","valueDrivers":["employee-productivity","revenue-growth","speed"],"kpis":["time-saved-per-task","cycle-time-days","cost-reduction","conversion-rate-uplift","users-served"],"indicativeValue":{"referenceOrg":"A sales organization with 300 quota carrying sellers","inputs":[{"key":"sellers","label":"Sellers using the tool","low":300,"high":300,"unit":"sellers","note":"The reference organization."},{"key":"adminHoursPerWeek","label":"Hours per seller per week on call notes, CRM updates and follow up emails","low":3,"high":5,"unit":"hours per seller per week","note":"Editorial assumption. Replace with a time study of your own sellers."},{"key":"shareSaved","label":"Share of that administration the AI takes over","low":0.3,"high":0.5,"unit":"fraction of admin hours","note":"Editorial assumption. For scale, Sandvik Coromant reports three minutes saved per transaction several times a day per account manager."},{"key":"weeks","label":"Working weeks per year","low":44,"high":46,"unit":"weeks per year","note":"Editorial assumption."},{"key":"hourlyCost","label":"Fully loaded seller cost","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption, replace with your own."}],"formula":"sellers * adminHoursPerWeek * shareSaved * weeks * hourlyCost","currency":"USD","period":"per year","resultLabel":"Seller time released from administration","caveat":"Values seller time at cost. It leaves out the effect on revenue of more selling time and better coaching, the value of a more accurate CRM for forecasting, and the cost of the platform. Time released only becomes value if it goes into customer work."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Summaries and CRM suggestions are available in many CRM and meeting tools. The effort is in consent and recording rules per country, clean CRM field definitions, and a coaching approach that sellers and works councils accept.","dataPrerequisites":["Call and meeting recordings or transcripts, captured with notice and consent","A CRM with defined opportunity stages, fields and next step conventions","The team's sales method or playbook, written down","Agreement with sellers (and employee representatives where required) on how analysis is used"],"integrations":["CRM (for example Salesforce, Microsoft Dynamics 365, HubSpot)","Telephony and meeting platforms (Microsoft Teams, Zoom, dialers)","Email and calendar","Sales enablement and learning content"]},"implementation":{"steps":[{"title":"Settle consent and purpose first","detail":"Decide which calls are recorded, how customers are told and can opt out, and in writing what the analysis will and will not be used for. Involve employee representatives where required."},{"title":"Start with summaries and CRM updates","detail":"Deliver the part sellers feel immediately: a summary, next steps and proposed CRM changes they confirm in one step. Measure time saved and CRM completeness."},{"title":"Define the method you coach against","detail":"Turn the sales method into observable moments (for example budget discussed, decision maker identified, next step agreed) and test that the AI detects them reliably on real calls."},{"title":"Give sellers their own insight first","detail":"Let sellers review their own calls and scores before managers see them, and use insight in coaching conversations rather than as a league table."},{"title":"Audit the scoring","detail":"Check detection accuracy by language, accent and call type, and make sure no metric depends on tone of voice or inferred emotion."},{"title":"Feed enablement","detail":"Use aggregated patterns (rising objections, winning questions) to update training and playbooks, with anonymized examples."}],"guardrails":["Recording and analysis only with notice to all participants and an opt out for customers","CRM changes proposed to the seller, never written without confirmation","No inference of emotions from voice or face; analysis based on what was said","Coaching insight is not used alone for pay, promotion or dismissal decisions","Transcripts masked for payment and sensitive personal data, with defined retention"],"humanInTheLoop":"Sellers confirm every CRM update and follow up message. Managers use the insight in coaching conversations and own any judgment about performance, based on more than the AI's analysis. Sales operations reviews extraction accuracy monthly.","kpisToInstrument":["Seller time on administration per week, before and after","Share of opportunities with a next step and updated fields after each meeting","Acceptance rate of proposed CRM updates","Accuracy of detected sales method moments on a reviewed sample","Win rate and cycle length for coached versus not yet coached sellers"],"failureModes":[{"title":"Surveillance, not coaching","detail":"Sellers experience scores as monitoring and game or avoid the tool. Agree the purpose up front, show sellers their data first and coach rather than rank."},{"title":"Confident but wrong CRM data","detail":"The AI records a close date or amount that was never agreed. Propose changes for confirmation, never write them silently."},{"title":"Recording without a lawful basis","detail":"Calls are recorded in a country or channel where notice or consent was not given. Map the rules per country and enforce them in the tool."},{"title":"Biased scoring","detail":"Detection works worse for some accents or languages and penalizes those sellers. Test accuracy per group before scores are shown."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Summaries, CRM suggestions and follow up drafts that the seller reviews are not an Annex III use and are minimal risk. Using call analysis to monitor and evaluate the performance and behaviour of individual sellers, or to allocate leads to sellers based on their behaviour or personal traits, is high risk under Annex III point 4(b). Inferring sellers' emotions from their voice is prohibited in the workplace by Article 5(1)(f). Emotion recognition applied to customers' voices is high risk under Annex III point 1(c), and Article 50(3) requires deployers to inform the people exposed to it."},"regulations":["eu-ai-act","gdpr","uk-gdpr","iso-42001"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) covers monitoring and evaluating workers' performance and behaviour; point 1(c) covers emotion recognition."},{"title":"Article 5, prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Point 1(f) prohibits emotion recognition in the workplace."},{"title":"Employment practices and data protection: monitoring workers","issuer":"UK Information Commissioner's Office","region":"europe","url":"https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/","note":"Expectations for transparency, proportionality and data protection impact assessments when monitoring workers, including a section on monitoring telephone calls."}],"controls":["Data protection impact assessment for call recording and analysis, per country","Written purpose limitation for coaching data, agreed with employee representatives where required","Customer notice and opt out at the start of recorded calls and meetings","Retention limits and access control on recordings and transcripts","Accuracy testing of summaries and detected moments by language"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the post call part is an **agentic workflow**: the **self hosted transcription with\nspeaker diarization** (who spoke when, processed on Blits.ai infrastructure) or transcripts from\nthe meeting platform feed an **AI agent** with **structured output** that returns the summary,\nnext steps and proposed CRM field changes. **Custom functions** read the opportunity and write the\nconfirmed changes back (ready made tools cover Microsoft Dynamics 365 and HubSpot, and the\nintegration catalog includes Salesforce), with **human in the loop** approval so the seller\nconfirms every update.\n\nCoaching against the sales method runs as a second agent that marks moments in the transcript\nagainst your written playbook, grounded in a **knowledge base** of enablement content. **PII\nmasking** removes payment and personal data before text reaches a model, **test suites** with\nLLM based grading check extraction accuracy on reviewed calls, and **role based access control**\nlimits who in the platform can see transcripts and results. The platform is model agnostic and can run in the EU or\nUAE region."},"faq":[{"question":"How much time does AI save sellers on admin and CRM updates?","answer":"Microsoft reports that adding an email summary as a CRM note takes 10 seconds instead of three minutes or longer with Copilot for Sales. Sandvik Coromant says its account managers save three minutes per transaction multiple times a day, but that saving comes from seeing a customer's full situation in the Outlook side panel, not from updating the CRM. Microsoft also reports that Lumen cut the time sellers spend summarizing past sales interactions and researching an account from up to four hours to 15 minutes."},{"question":"Is AI analysis of sales calls high risk under the EU AI Act?","answer":"Summaries and CRM updates are not. Using the analysis to monitor and evaluate individual sellers is high risk under Annex III point 4(b), and inferring sellers' emotions from their voice is prohibited in the workplace. Design coaching around what was said, and let managers own judgments about people."},{"question":"Can call analysis replace manual call audits?","answer":"For coverage, largely. Microsoft reports that Hughes cut the cost of a sales call audit by 90%, from USD 26 to USD 2 per call hour, by replacing manual listening with automated transcription and analysis. Keep people reviewing the calls the analysis flags."}],"related":["meeting-summarization-and-action-items","client-briefing-and-call-report-copilot","client-meeting-notes-and-crm-update","conversation-roleplay-training","call-quality-and-compliance-monitoring","inbound-lead-qualification-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written for the cross industry employee and insight vertical with four evidence records verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added two cited Salesforce problem statistics, corrected the Lumen and Sandvik Coromant wording, tightened the EU AI Act basis (Article 50(3)), added UK GDPR, corrected the integration wording in the Blits.ai section, and added SEO title and meta description."},{"date":"2026-09-27","note":"Reframed the CRM update FAQ so Sandvik Coromant's three minutes per transaction is attributed to the Outlook side panel rather than CRM updates, and corrected the meta description, which implied the same saving came from call analysis and CRM updates."}],"slug":"sales-call-coaching-and-crm-update","url":"https://www.blits.ai/ai-use-cases/sales-call-coaching-and-crm-update","benchmarks":[{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":1,"median":3,"min":3,"max":3,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"lumen-copilot-sales-account-research","pooled":false},{"id":"sandvik-coromant-copilot-for-sales","pooled":true}]},{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"hughes-sales-call-auditing","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":300,"min":300,"max":300,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"zurich-copilot-for-sales-crm-updates","pooled":true}]}],"indicativeValueResult":{"low":712800,"high":3450000},"evidence":["hughes-sales-call-auditing","lumen-copilot-sales-account-research","sandvik-coromant-copilot-for-sales","zurich-copilot-for-sales-crm-updates"]},{"title":"AI scam intervention for instant payments","shortTitle":"Scam payment interception","seoTitle":"AI scam intervention for instant payments","metaDescription":"AI scam intervention questions customers before a risky payment and holds it for a specialist. Revolut and Commonwealth Bank report lower scam losses.","definition":"AI that talks to the customer when they are about to authorise an instant payment that looks like a scam: it combines the payee check and the risk score, asks targeted questions about the payment in plain language, explains the specific scam pattern, and holds, delays or escalates the payment to a human specialist when the risk stays high. Unlike fraud scoring, which stops payments the customer did not make, it protects customers from payments they are being manipulated into making.","aliases":["APP scam intervention","authorised push payment scam warning","dynamic scam warnings","scam payment friction"],"industries":["banking","payments"],"functions":["fraud-prevention","customer-service"],"patterns":["conversational-agent","prediction-and-scoring","agentic-workflow","voice-agent"],"channels":["mobile-app","web-chat","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"In an authorised push payment scam the customer sends the money themselves, usually after a\nconvincing story: a romance, an investment, a fake invoice, a caller posing as the bank. The\npayment passes every authentication check because the real customer makes it, and on instant\npayment rails such as Faster Payments in the UK or the New Payments Platform in Australia it\nleaves the account in seconds.\n\nA generic \"are you sure?\" warning is easy to click through for a customer who is being guided by\na scammer, which is why Revolut and Starling both describe their tools as breaking the scammer's\n\"spell\". At the same time regulators are moving the cost onto banks. In the UK, payment firms must\nreimburse most APP scam victims on Faster Payments and CHAPS, with the cost split 50:50 between\nthe sending and receiving firm. Singapore's Shared Responsibility Framework requires banks and\ntelcos to pay phishing scam victims when they breach set duties, and Australia's Scams Prevention\nFramework sets obligations to prevent, detect, disrupt and respond to scams, next to the banks'\nown Scam-Safe Accord. That makes the quality of the intervention, and the record of it, a\nfinancial and a regulatory question.","problemStats":[{"statement":"UK Finance data cited by Starling Bank show that Britons lost GBP 576.4 million to authorised push payment fraud in 2025, an increase of 19% on the previous year.","sourceTitle":"New AI feature detects romance scammers, investment heists and deepfake phishing attempts","sourceUrl":"https://www.starlingbank.com/news/new-ai-feature-detects-romance-scammers/","year":2026}],"howItWorks":"1. **Score the payment in real time.** The fraud and scam models, the Confirmation of Payee or\n   name check result, mule account signals on the payee and the customer's own behaviour produce\n   a risk level before the payment is sent.\n2. **Choose the intervention by risk.** Low risk payments go straight through. Medium risk gets a\n   warning specific to the payment's purpose, not a generic one. High risk opens a short\n   conversation in the app.\n3. **Ask, listen and explain.** The agent asks why the customer is paying, how they met the\n   payee and who suggested the payment, looks for signs of coaching or urgency, and explains the\n   matching scam pattern in plain words. Starling's in app assistant does this for transfers a\n   customer describes, and Revolut runs a similar flow for card payments its model has declined.\n4. **Hold and escalate.** When the risk stays high the payment is held and the customer is offered\n   a call with a scam specialist. On the call, an assistant can transcribe and flag indicators for\n   the banker, as Westpac reported piloting in 2025.\n5. **Decide within policy.** The agent can release low and medium risk payments after the\n   conversation; releasing a held high risk payment, or declining it, is a human decision with a\n   documented reason.\n6. **Record everything.** Every warning shown, every answer given and every override is stored,\n   because reimbursement and shared responsibility regimes ask what the bank did and when.\n7. **Learn.** Confirmed scams and false alarms flow back into the models and into the questions\n   the agent asks.","valueDrivers":["risk-reduction","customer-experience","compliance"],"kpis":["fraud-loss-reduction","detection-rate-improvement","false-positive-reduction","customer-satisfaction","interactions-handled"],"indicativeValue":{"referenceOrg":"A retail bank with 1 million digitally active customers","inputs":[{"key":"customers","label":"Digitally active customers","low":1000000,"high":1000000,"unit":"customers","note":"The reference bank."},{"key":"scamLossPerCustomer","label":"Scam losses per customer per year","low":1,"high":4,"unit":"USD per customer per year","note":"Editorial assumption, replace with your own reported scam losses divided by active customers."},{"key":"lossReduction","label":"Reduction in scam losses from better intervention","low":0.1,"high":0.3,"unit":"fraction of scam losses","note":"Conservative against the evidence on this page. Revolut reports a 30% fall in card scam losses for investment scams; Commonwealth Bank's 76% fall since the peak covers its whole program, not the intervention alone."},{"key":"bankBorneShare","label":"Share of scam losses the bank bears","low":0.5,"high":0.9,"unit":"fraction of scam losses","note":"Editorial assumption; depends on the reimbursement regime in your market and your own policy."}],"formula":"customers * scamLossPerCustomer * lossReduction * bankBorneShare","currency":"USD","period":"per year","resultLabel":"Scam losses borne by the bank that are avoided","caveat":"Counts avoided losses the bank would carry. It leaves out the losses customers avoid, the cost of added friction on genuine payments, specialist call time, the cost of the models and the reputational effect."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The intervention sits in the payment path, so it must answer within the payment's latency budget, work with the fraud engine and payee checks, and hold a payment without breaking payment scheme rules. The conversation design and the evidence trail matter as much as the model.","dataPrerequisites":["Labelled scam cases by typology (romance, investment, purchase, impersonation, invoice)","Payee check results and mule account intelligence","Payment purpose and customer behaviour signals in real time","Approved scam education content per typology"],"integrations":["Payment initiation in the app and online banking, with the ability to hold or delay","Fraud and scam scoring engine","Confirmation of Payee or equivalent name check service","Contact centre platform for specialist calls with context","Case management for held payments and reimbursement claims"]},"implementation":{"steps":[{"title":"Start from your scam typologies","detail":"Take last year's confirmed scams, group them by typology and value, and write for each the signals, the questions that expose it and the words that land with a customer."},{"title":"Tier the interventions","detail":"Agree risk bands with the fraud team and a different response for each: no friction, a tailored warning, a conversation, a hold with a call. Measure how many genuine payments each band touches."},{"title":"Design the conversation with victims","detail":"Test the questions with people who were scammed; Starling designed its romance scam feature with advice from a romance scam survivor. Scripts written by fraud analysts alone tend to sound like accusations."},{"title":"Wire the hold and the specialist route","detail":"Make sure a held payment has an owner, a service level and a callback, and that the specialist sees the conversation so the customer does not repeat the story."},{"title":"Build the evidence trail","detail":"Store each warning, answer and override with timestamps in a form your reimbursement and complaints teams can retrieve per payment."},{"title":"Run it as a champion and challenger test","detail":"Compare scam losses, cancelled payments and complaints between the new intervention and the current warnings on a random split before full rollout."}],"guardrails":["Releasing a held high risk payment or declining it requires a human with a recorded reason","The agent never asks for passcodes, card details or remote access, and says so","Warnings and questions come from approved content per typology","Vulnerability signals route to a specialist rather than to more automated questions","Latency budget and a safe default when the scoring service is unavailable"],"humanInTheLoop":"Scam specialists handle every held high risk payment and decide on release, delay or decline, with the conversation in front of them. The fraud team reviews new detection rules before they go live, as Commonwealth Bank does with its detection agent, and samples released payments that later turned out to be scams.","kpisToInstrument":["Scam losses per million payments, by typology, against a control group","Share of high risk payments cancelled after the intervention","Share of genuine payments that received friction, and their abandonment rate","Time to specialist contact for held payments","Reimbursement claims where the record shows no effective warning"],"failureModes":[{"title":"Warning fatigue","detail":"Too many warnings on genuine payments train customers to click through. Keep friction for the risk bands that justify it and measure how often genuine customers see it."},{"title":"The scammer coaches the answers","detail":"Victims are often told what to say. Ask questions that are hard to script, watch for coaching signals and escalate to a human rather than accepting a clean answer."},{"title":"Held payments with no owner","detail":"A hold without a fast specialist call angers genuine customers and pushes them to other banks. Staff the queue before switching the hold on."},{"title":"No usable evidence trail","detail":"The bank did intervene but cannot show it per payment. Store the exact warning and the answers, not just a flag."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Annex III point 5(b) expressly excludes AI systems used to detect financial fraud from the high risk creditworthiness category, so the scoring is not high risk. The conversational part must disclose that it is AI under Article 50(1). If a voice component infers the customer's emotions from their voice, it becomes an emotion recognition system under Annex III point 1(c), which is high risk and needs the Article 50(3) notice, so keep coaching detection to what is said rather than to biometric signals."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","dora","mas-ai-risk-management","apra-cps-230","eu-psd2","uk-psr-app-reimbursement","mas-shared-responsibility-framework","au-scams-prevention-framework"],"guidance":[{"title":"APP scams","issuer":"Payment Systems Regulator","region":"europe","url":"https://www.psr.org.uk/our-work/app-scams/","note":"UK reimbursement requirement for APP scam victims paying by Faster Payments or CHAPS, with costs split 50:50 between sending and receiving firms and most victims reimbursed within five business days."},{"title":"Guidelines on Shared Responsibility Framework","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/guidelines/guidelines-on-shared-responsibility-framework","note":"Assigns duties to financial institutions and telcos to mitigate phishing scams and requires payouts to victims where those duties are breached; in force since 16 December 2024."},{"title":"Keeping Australia Scam Safe","issuer":"Australian Banking Association","region":"asia-pacific","url":"https://www.ausbanking.org.au/priorities/scam-safe-accord/","note":"The Australian banks' Scam-Safe Accord, including Confirmation of Payee and commitments to more warnings, payment delays and security questions on risky payments."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) excludes AI systems used for detecting financial fraud from the creditworthiness high risk category."}],"controls":["AI disclosure in the intervention conversation","Documented risk bands and intervention per band, approved by the fraud risk owner","Per payment record of warnings, answers, overrides and human decisions","Model monitoring for detection, false positives and drift, with human approval of new rules","Regular review of outcomes for vulnerable customers"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the intervention is an **AI agent** called from the payment journey through the\n**REST API or WebSocket API channel**, or shown in the embedded **chat widget**, with the payment\ncontext and the risk band passed in. A **flow** fixes the regulated steps per band (warning, questions, hold offer)\nand the agent handles the open conversation. **Custom functions** call the bank's scoring\nservice, the payee check and the payment hold API; a **knowledge base** holds the approved scam\neducation content per typology.\n\nHeld payments go through an **agentic workflow with human in the loop approval**, so a scam\nspecialist approves or rejects release with the conversation attached, and every step lands in\nthe run's audit trail. **Human handover** passes the customer and the conversation history to a\nspecialist in chat or on a **voice** call, so they do not repeat the story. **Guardrails** and **PII masking** keep account data\nout of prompts, **test suites** replay scam scripts (including coached answers) on every change,\nand **analytics** track cancellations and handovers per typology. The platform is model agnostic\nand runs in EU and UAE regions for data residency."},"faq":[{"question":"Does AI actually reduce scam losses?","answer":"The published results point that way, with caveats. Revolut reports a 30% fall in losses from card scams involving investment opportunities after launching its AI detection and intervention flow, and Commonwealth Bank reports a 76% fall in customer scam losses since the peak, across its whole scam program. Measure it yourself with a control group."},{"question":"Should the AI be allowed to block a payment?","answer":"It can pause a payment and start a conversation within agreed risk bands. Releasing or declining a held high risk payment should stay with a trained specialist who has the full conversation in front of them, and the decision should be recorded."},{"question":"Is scam detection high risk under the EU AI Act?","answer":"Not as such. Annex III point 5(b) excludes AI used to detect financial fraud from the creditworthiness category. The customer conversation still needs an AI disclosure, a voice component that recognises emotions would be high risk under Annex III point 1(c), and GDPR applies to the personal data used."}],"related":["real-time-fraud-scoring","mule-network-detection","fraud-alert-confirmation","spam-and-scam-call-blocking","telecom-fraud-detection","card-dispute-and-chargeback-intake"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI use case catalog and verified against Commonwealth Bank, Starling, Westpac, Mastercard, Revolut and regulator sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added PSD2, UK APP scam reimbursement rules, MAS Shared Responsibility Framework, Australian Scams Prevention Framework to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; replaced unsourced claims in the problem (payment rail list, warning effectiveness, Scam-Safe Accord described as duties) with sourced wording on the UK, Singapore and Australian rules; corrected the Starling, Revolut and Westpac references; added the emotion recognition caveat to the EU AI Act basis and FAQ; replaced the webview reference in the Blits.ai block with inventory capabilities; moved the Revolut metric to Revolut's own release via its archived copy (grade B); regraded Starling to C because its only metric is a vendor claim; added source dates."}],"slug":"scam-payment-interception","url":"https://www.blits.ai/ai-use-cases/scam-payment-interception","benchmarks":[{"kpi":"detection-rate-improvement","label":"Detection improvement","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":165,"min":30,"max":300,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"starling-bank-scam-intelligence","pooled":true},{"id":"vodafone-scam-signal","pooled":true}]},{"kpi":"fraud-loss-reduction","label":"Fraud loss reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":53,"min":30,"max":76,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"commonwealth-bank-scam-and-fraud-interventions","pooled":true},{"id":"revolut-card-scam-detection","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":40000,"min":40000,"max":40000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"commonwealth-bank-scam-and-fraud-interventions","pooled":true}]}],"indicativeValueResult":{"low":50000,"high":1080000},"evidence":["commonwealth-bank-scam-and-fraud-interventions","mastercard-consumer-fraud-risk","revolut-card-scam-detection","starling-bank-scam-intelligence","vodafone-scam-signal","westpac-scam-call-assistant"]},{"title":"AI scoring of essays and written answers in assessments","shortTitle":"Essay and written answer scoring","seoTitle":"AI essay scoring for tests and exams","metaDescription":"Engines score essays and written answers while humans read a sample and every unclear case, as in Texas. Massachusetts had to rescore about 1,400 essays in 2025.","definition":"AI that scores students' essays and short written answers against a rubric, trained on responses scored by human raters, with human raters rescoring a sample of responses and every response the engine is unsure about. In hybrid programmes such as Texas, a human score is the score of record whenever a human scores a response.","aliases":["automated essay scoring","AI grading of constructed responses","AI marking assistant","machine scoring of written answers"],"industries":["education","government"],"functions":["operations"],"patterns":["classification-and-routing","prediction-and-scoring"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"mainstream","problem":"Written answers show what students can do in ways multiple choice cannot, but every one has to be\nread against a rubric by a trained rater. When Texas redesigned its STAAR tests to include more\nwriting, the number of constructed responses to score each year grew six to sevenfold, and the\nagency estimated that scoring them all by hand would cost 15 to 20 million US dollars more per\nyear. Human scoring is not perfectly consistent either: in a TEA study of extended essays, two\ntrained raters gave exactly the same conventions score on only 67 to 72 percent of responses.\nIn classrooms, the UK Department for Education describes feedback and marking as a burden on\nteachers.\n\nAutomated essay scoring is not new: ETS has used its engine alongside human raters on the GRE\nsince at least 2012, and TEA said in 2024 that at least 21 states use automated scoring for their\nstate assessments. What has changed is scale and scope. State assessment programmes now let an\nengine give the first score for most written answers, and the UK Department for Education funds\nAI tools meant to reduce the burden of feedback and marking on teachers. The stakes are high: a wrong score can\naffect a student's record, a school's rating and public trust, as the Massachusetts rescoring of\nabout 1,400 essays in 2025 showed.","problemStats":[{"statement":"The Texas Education Agency says the STAAR redesign brought 6 to 7 times more constructed responses to grade each year, and that maintaining full human scoring would have cost 15 to 20 million US dollars more per year.","sourceTitle":"Hybrid Scoring Key Questions (Texas Education Agency, March 2024)","sourceUrl":"https://web.archive.org/web/20240512091507/https://tea.texas.gov/student-assessment/testing/hybrid-scoring-key-questions.pdf","year":2024},{"statement":"In a Texas Education Agency study of spring 2023 STAAR extended constructed responses, two human raters gave exactly the same conventions score on 67 to 72 percent of responses, depending on the item.","sourceTitle":"Hybrid Scoring Key Questions (Texas Education Agency, March 2024)","sourceUrl":"https://web.archive.org/web/20240512091507/https://tea.texas.gov/student-assessment/testing/hybrid-scoring-key-questions.pdf","year":2024}],"howItWorks":"1. **Set the standard with humans.** Educators score a set of field test responses against the\n   rubric and agree anchor responses for every score point.\n2. **Train and qualify the engine.** The engine is trained on human scored responses for each\n   question and must agree with human raters at the same rate human raters agree with one\n   another, with a similar score distribution, before it is used live.\n3. **Score and flag.** The engine gives each response a first score and a confidence value, and\n   flags responses that are blank, too short, off topic, copied, in another language or unlike\n   anything in its training data.\n4. **Route to humans.** Flagged and low confidence responses, and a fixed share of all responses,\n   go to trained human raters. Where a human scores a response, the human score is the score of\n   record.\n5. **Monitor and correct.** Agreement between engine and humans is monitored daily, preliminary\n   results go to schools with a window to report discrepancies, and rescoring is available.","valueDrivers":["cost-to-serve","speed","employee-productivity","compliance"],"kpis":["cost-reduction","processing-time-reduction","accuracy","automation-rate","hours-saved"],"indicativeValue":{"referenceOrg":"A state assessment programme scoring 1 million written responses a year","inputs":[{"key":"responses","label":"Written responses scored per year","low":1000000,"high":1000000,"unit":"responses per year","note":"The reference programme."},{"key":"engineOnlyShare","label":"Share of responses scored by the engine without a human read","low":0.6,"high":0.75,"unit":"fraction of responses","note":"The Texas Education Agency routes at least 25 percent of responses, plus flagged and low confidence ones, to human raters, so at most 75 percent are engine only. The low value is an editorial assumption."},{"key":"humanScoringCost","label":"Cost of human scoring per response","low":0.6,"high":1.7,"unit":"USD per response","note":"Derived from the Texas Education Agency deck cited above: 15 to 20 million US dollars a year avoided, spread over the roughly 75 percent of 15.8 million annual responses that the engine now scores alone and that were previously scored by two humans, is about 1.3 to 1.7 USD per response (about 0.6 to 0.8 USD per single read). The low value assumes one human read per response. Replace with your own contract rates."}],"formula":"responses * engineOnlyShare * humanScoringCost","currency":"USD","period":"per year","resultLabel":"Human scoring cost avoided","caveat":"Gross avoided rater cost only. It leaves out engine licensing, validation studies per question, monitoring, rescoring and appeals, and the cost of errors, which can be large in reputation and student outcomes even when rare."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The engine is the smaller part. The work is psychometric: per question training and validation, agreement thresholds, fairness checks across student groups, confidence routing, daily monitoring and a rescoring process that schools and families trust.","dataPrerequisites":["Rubrics and anchor responses approved by educators for every question","Human scored responses for each question from field tests, enough to train and validate","Student group information to test for differences between engine and human scores"],"integrations":["Test delivery platform that captures responses","Human scoring platform for routed responses and second reads","Results reporting to schools with a discrepancy and rescore process"]},"implementation":{"steps":[{"title":"Decide where machine scoring is appropriate","detail":"Use it for tasks scored for writing quality or a clear rubric, not for tasks where the correctness of claims or creative reasoning is what counts. Keep human scoring for any language or test version the engine has not been validated on; Texas, for example, scores its Spanish STAAR responses entirely by hand."},{"title":"Validate per question","detail":"Require the engine to agree with human raters at the same rate human raters agree with one another, with a similar score distribution, on responses the engine has not seen, for every question."},{"title":"Route by confidence and condition","detail":"Send low confidence, borderline and unusual responses to humans, plus a fixed random share of all responses as an ongoing check. Make the human score the score of record."},{"title":"Check fairness","detail":"Compare engine and human scores for student groups (for example by language background and disability) and investigate systematic differences before release."},{"title":"Build in a discrepancy window","detail":"Release preliminary results to schools with time to report issues and a clear rescoring route, and publish how the process works."}],"guardrails":["Human rescoring of a fixed share of responses and of all low confidence or flagged responses","Human score as the score of record whenever a human scores a response","Per question validation against human agreement before live use","Daily monitoring of engine and human agreement during the scoring window","A discrepancy period and rescoring process for schools and families"],"humanInTheLoop":"Educators set the rubric and anchor responses; trained human raters score every flagged, low confidence and sampled response; scoring directors monitor agreement daily; and schools can challenge preliminary scores before results are final.","kpisToInstrument":["Exact and adjacent agreement between engine and human scores per question","Share of responses routed to humans, by reason","Differences between engine and human scores by student group","Rescore requests and changed scores after release","Cost and time per scored response"],"failureModes":[{"title":"Systematic scoring errors on a set of responses","detail":"A technical issue makes the engine score a set of essays incorrectly, as with about 1,400 Massachusetts essays in 2025, which DESE attributed to a temporary technical issue in the process. Sample human reads across the score range and give schools a discrepancy window."},{"title":"Responses unlike the training data","detail":"Writing styles, structures or vocabulary that are rare in the training responses can be scored less reliably. Route responses the engine flags as unusual to humans and check agreement for different student groups."},{"title":"Gaming the engine","detail":"Answers written to exploit surface features an engine may reward, such as length or rubric vocabulary, can score higher than their content deserves. Flag unusual responses, keep a random share of human reads and review what drives scores."},{"title":"Scoring what the engine can see","detail":"Rubrics drift toward surface features such as length and punctuation. Keep educators in charge of the rubric and review what drives scores."},{"title":"Loss of trust","detail":"Families and teachers distrust machine scores when the process is opaque. Publish how scoring works and how to request a rescore."}]},"risk":{"euAiAct":{"tier":"high","basis":"Annex III point 3(b): AI systems intended to be used to evaluate learning outcomes in educational and vocational training institutions at all levels are high risk. Scoring that determines access to an institution or the level of education a student will receive is also covered by points 3(a) and 3(c). Schools and exam bodies that use such a system have the deployer obligations of Article 26."},"regulations":["eu-ai-act","gdpr","uk-gdpr","nist-ai-rmf","iso-42001"],"guidance":[{"title":"Annex III, high risk AI systems (point 3, education and vocational training)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Lists AI systems intended to evaluate learning outcomes as high risk, which brings the requirements of Chapter III, Section 2 (Articles 8 to 15) on risk management, data governance, logging and human oversight, and the deployer obligations of Article 26."},{"title":"Scoring Process for STAAR Constructed Responses","issuer":"Texas Education Agency","region":"north-america","url":"https://tea.texas.gov/data-reports/staar/scoring-process-staar-constructed-response-1.pdf","note":"A published description of a hybrid scoring process, including engine qualification against human agreement, confidence and condition code routing and the human score as the score of record."},{"title":"Generative artificial intelligence (AI) in education","issuer":"UK Department for Education","region":"europe","url":"https://www.gov.uk/government/publications/generative-artificial-intelligence-in-education/generative-artificial-intelligence-ai-in-education","note":"The department's position on generative AI tools in schools and colleges, including its funding for AI tools that aim to reduce the burden of feedback and marking on teachers; relevant when teachers use AI to mark classroom work."}],"controls":["Published description of the scoring process and how to request a rescore","Per question validation records and agreement thresholds","Fairness analysis across student groups before each release","Logging of engine scores, confidence values and routing decisions","Contract terms requiring the scoring vendor to report and correct errors"],"incidents":[{"title":"AI grading issue affects hundreds of MCAS essays in Massachusetts","url":"https://www.nbcboston.com/investigations/ai-grading-massachusetts-mcas/3807392/","note":"In 2025 the state's testing contractor Cognia found that roughly 1,400 MCAS essays (of about 750,000 statewide) had not received the correct scores under AI scoring, which DESE attributed to a temporary technical issue in the process. District leaders, including in Lowell, raised the problem after preliminary results were released, and the essays were rescored."}]},"blitsAi":{"howToBuild":"On Blits.ai a scoring assistant is an **agentic workflow** that reads each response and applies\nthe rubric through an **AI agent** with **structured output** (score per trait, rationale and a\nconfidence value), using the approved anchor responses in a **knowledge base** as reference.\nResponses below a confidence threshold, flagged by **guardrails** (blank or off topic) or by\n**language detection** (another language) pause for **human in the loop** confirmation, where a\ntrained rater approves or rejects the engine score. A human rescore is recorded in the\nprogramme's own scoring system, which a **custom function** can write to.\n\n**Test suites** with deterministic and LLM grading run the rubric prompt against a set of human\nscored responses and report a pass rate and per case verdicts before use; agreement statistics\nsuch as exact and adjacent agreement are calculated by the programme. **Prompt versioning** keeps\na history of prompt versions, and the **run history** and **audit trail** per run support\nrescoring and appeals. The platform is model agnostic and offers EU and UAE data residency.\nValidation against human agreement and fairness checks remain the programme's responsibility."},"faq":[{"question":"Is AI already used to score state tests?","answer":"Yes. Massachusetts uses AI to score MCAS essays, trained on human scored examples of each score point, with humans giving 10 percent of AI scored essays a second read. Texas scores STAAR written responses with an automated scoring engine first and routes at least 25 percent, plus low confidence and unusual ones, to human raters; TEA says its engine is not AI in the sense of a system that teaches itself, but is programmed on about 3,000 human scored responses per question."},{"question":"How accurate is AI essay scoring?","answer":"For suitable tasks, engines can agree with human raters as closely as two humans agree with each other. Texas requires this before an engine is used, and ETS cites a 2010 study in which its engine's agreement with a human rater on the TOEFL Independent and GRE Issue tasks was higher than between two human raters. Errors still happen: in 2025 about 1,400 Massachusetts essays did not receive the correct scores and were rescored."},{"question":"Is AI essay scoring high risk under the EU AI Act?","answer":"Yes. Annex III point 3(b) lists AI systems intended to evaluate learning outcomes as high risk, which brings the requirements of Chapter III, Section 2 (Articles 8 to 15) on risk management, data governance, logging and human oversight. Schools and exam bodies that use such a system also have the deployer obligations of Article 26."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the education scope with evidence from the Texas Education Agency, ETS and the Massachusetts Department of Elementary and Secondary Education, checked against the sources. Editor pass the same day sourced the problem statement from TEA, corrected the evidence years (ETS 2012, Texas 2023) and tightened the incident notes. A second editor pass rechecked every claim against the sources, stated the TEA rater agreement figure exactly and softened the Ofqual note. A third editor pass stated that TEA does not call its engine AI, described the Massachusetts incident as DESE did, limited the Blits.ai section to listed features, derived the human scoring cost from TEA figures, set adoption to mainstream (TEA reports at least 21 states), removed the Ofqual incident and corrected the EU AI Act articles."}],"slug":"automated-scoring-of-written-responses","url":"https://www.blits.ai/ai-use-cases/automated-scoring-of-written-responses","benchmarks":[],"indicativeValueResult":{"low":360000,"high":1275000},"evidence":["ets-gre-automated-essay-scoring","massachusetts-dese-mcas-ai-essay-scoring","texas-education-agency-staar-automated-scoring"]},{"title":"AI screening of trade finance transactions for trade based money laundering","shortTitle":"Trade crime screening","seoTitle":"AI screening for trade based money laundering","metaDescription":"AI screens trade finance deals for sanctioned vessels, dual use goods and mispriced invoices. See United Bank Limited's automation and a proof of concept with HSBC.","definition":"AI that screens every trade finance transaction for financial crime risk: it checks parties, vessels and ports against sanctions and watchlists, tests goods descriptions against dual use and controlled goods lists, compares unit prices with benchmarks for over or under invoicing, and reads trade documents and messages for laundering red flags, then prepares a case narrative for a human investigator.","aliases":["TBML screening","trade based money laundering detection","trade compliance screening","dual use goods screening"],"industries":["banking"],"functions":["financial-crime-compliance","operations"],"patterns":["document-processing","anomaly-detection","classification-and-routing","summarization"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"emerging","segment":"specialized-businesses","problem":"Trade is a well known channel for moving criminal money and evading sanctions: goods can be over\nor under invoiced, misdeclared, traded through shell companies, or routed through restricted\njurisdictions, sanctioned ports and vessels. The evidence is spread across letters of credit, invoices, bills of\nlading, SWIFT messages and vessel tracking data, much of it unstructured.\n\nMuch of this checking has been manual. United Bank Limited in Pakistan, for example, relied on\nmanual processes for vessel screening, tracking and dual use goods identification before it\nselected an automated platform in 2020. Done by hand, the work means looking up vessels, reading\ngoods descriptions against control lists and searching for counterparties one by one, which is\nslow and open to errors and delays. Regulators expect a risk based, documented process, and\ncompliance gaps can lead to penalties.","problemStats":[{"statement":"LexisNexis Risk Solutions states, citing an outside source, that trade based money laundering represents up to 80% of capital flight from developing nations.","sourceTitle":"Targeting Trade-Based Money Laundering in APAC","sourceUrl":"https://risk.lexisnexis.com/global/en/insights-resources/article/trade-based-money-laundering-with-data-driven-response","year":2025}],"howItWorks":"1. **Extract.** Document AI and message parsing pull parties, goods descriptions, quantities,\n   prices, ports, vessels and dates from the letter of credit, trade documents and SWIFT messages.\n2. **Screen names and routes.** Parties, banks, vessels and ports are screened against sanctions\n   and watchlists, and vessel history and tracking are checked for suspicious port calls or\n   transponder gaps.\n3. **Check the goods.** Goods descriptions are classified against dual use and controlled goods\n   lists, including vague or unusual descriptions that need a closer look.\n4. **Test the economics.** Unit prices and quantities are compared with benchmarks and the\n   client's usual trade pattern to spot over or under invoicing and unusual routes.\n5. **Prioritise and explain.** Alerts are scored, duplicates and known false positives are\n   suppressed, and a case narrative with the evidence is drafted for the investigator, who decides\n   and, where needed, files a suspicious activity report.","valueDrivers":["risk-reduction","compliance","speed","employee-productivity"],"kpis":["false-positive-reduction","detection-rate-improvement","processing-time-reduction","handling-time-reduction","automation-rate"],"indicativeValue":{"referenceOrg":"A trade bank handling 40,000 trade finance transactions a year","inputs":[{"key":"transactions","label":"Trade finance transactions per year","low":40000,"high":40000,"unit":"transactions per year","note":"The reference bank."},{"key":"alertRate","label":"Share of transactions that need a manual compliance review","low":0.2,"high":0.4,"unit":"fraction of transactions","note":"Editorial assumption, replace with your own alert volumes."},{"key":"minutesPerAlert","label":"Minutes of review per alerted transaction","low":20,"high":40,"unit":"minutes per alert","note":"Editorial assumption, replace with your own time study."},{"key":"reduction","label":"Share of review time saved by extraction, suppression and drafted narratives","low":0.2,"high":0.4,"unit":"fraction of time","note":"Editorial assumption, replace with your own pilot results."},{"key":"hourlyCost","label":"Loaded cost of a compliance analyst hour","low":50,"high":80,"unit":"USD per hour","note":"Editorial assumption, replace with your own loaded cost."}],"formula":"transactions * alertRate * minutesPerAlert / 60 * reduction * hourlyCost","currency":"USD","period":"per year","resultLabel":"Compliance review time released, valued at loaded cost","caveat":"Values review time only. Possible further value, such as fewer missed red flags and lower regulatory penalty risk, is not quantified here and is not yet shown by public evidence, nor are faster turnaround for clients or the cost of data feeds."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Needs good trade document extraction, vessel and commodity data feeds, price benchmarks, and integration with the trade processing and case management systems, all under model risk and financial crime governance.","dataPrerequisites":["Sanctions, watchlist, vessel and port data kept current","Dual use and controlled goods lists for the relevant jurisdictions","Price benchmarks per commodity and a history of client trade patterns","Labelled past alerts and investigation outcomes to tune thresholds"],"integrations":["Trade finance processing system","SWIFT messaging","Sanctions screening engine","Vessel tracking and maritime data provider","Financial crime case management"]},"implementation":{"steps":[{"title":"Map the red flags you must cover","detail":"Start from the typologies your regulators and industry bodies list (pricing anomalies, dual use goods, vessel and route risks, unusual parties) and map each to data and a check."},{"title":"Automate extraction and list checks first","detail":"Begin with the checks that are still manual: data extraction from documents and vessel, port, party and goods screening. United Bank Limited, for example, replaced manual vessel screening, tracking and dual use goods identification with automated checks."},{"title":"Add anomaly detection with explanations","detail":"Add price and pattern anomaly models only with clear reasons per alert, so investigators can act on them and examiners can follow them."},{"title":"Draft, do not decide","detail":"Let the system draft the case narrative with evidence links; the investigator decides and signs any suspicious activity report."},{"title":"Monitor as model metrics","detail":"Track list freshness, false positive rates and missed cases found later as model performance metrics with an owner."}],"guardrails":["Humans decide on every suspicious activity report and on declining or exiting a transaction","Screening lists and vessel data freshness monitored, with alerts on stale feeds","Full data lineage and reasoning retained for every alert and decision","Suppression rules for false positives are documented, approved and reviewed periodically"],"humanInTheLoop":"Trade compliance analysts review every alert above threshold and every sanctions or dual use match. Investigators decide on escalation and suspicious activity reports. The financial crime function approves suppression rules and thresholds, and model risk validates the models.","kpisToInstrument":["False positive rate and alerts per thousand transactions","Time from presentation to compliance clearance","True positives and escalations, including those found later by other means","Freshness of screening lists and vessel data","Analyst hours per alert"],"failureModes":[{"title":"Over suppression","detail":"Tuning to cut false positives also hides real risk. Test suppression rules against past true positives and sample suppressed alerts."},{"title":"Vague goods descriptions slip through","detail":"Generic descriptions (\"machinery parts\") evade keyword checks. Flag vague descriptions for review rather than passing them."},{"title":"Stale data","detail":"A sanctions list or vessel feed stops updating unnoticed. Monitor feed timestamps and block clearance when data is stale."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Financial crime screening of trade transactions is not listed in Annex III. It still processes personal data of individual parties, so GDPR applies, and supervisors expect it to be governed like any financial crime model."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","mas-ai-risk-management","us-sr-11-7","dora","eu-amlr","us-bsa","mas-notice-626"],"guidance":[{"title":"MAS Guidelines for Artificial Intelligence (AI) Risk Management","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Consultation paper of 13 November 2025 (consultation closed 31 January 2026; no final Guidelines were found on the MAS consultation page when checked in September 2026) proposing supervisory expectations for AI oversight, inventories, life cycle controls, human oversight and monitoring, relevant for models that prioritise or suppress financial crime alerts."},{"title":"ICC trade finance rules and financial crime risk controls","issuer":"International Chamber of Commerce","region":"global","url":"https://iccwbo.org/business-solutions/trade-finance/","note":"ICC's trade finance hub links the documentary credit rules (UCP 600, eUCP) with its financial crime risk control guides on dual use goods, price checking and vessel checking, and with the Wolfsberg Group, ICC and BAFT Trade Finance Principles."}],"controls":["Model inventory, validation and ongoing performance monitoring","Documented red flag coverage mapped to typologies","Audit trail of data, rules and model versions behind each alert","Periodic independent testing of screening effectiveness"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the screening engine and list data stay where they are; Blits.ai adds the extraction,\norchestration and narrative layer. An **agentic workflow**, triggered through the API when a\npresentation arrives, takes the document content from the trade system, extracts the fields with\n**structured output**, and calls the\nscreening, vessel data and price benchmark services through **custom functions** or **MCP**. A\n**knowledge base** holds the bank's red flag typologies and dual use guidance for retrieval.\n\nThe agent drafts the case narrative with links to the evidence, and **human in the loop\napproval** is required before any escalation is recorded. The per run **audit trail** keeps every\ntool call and result, **guardrails** and **PII masking** protect party data in prompts, and\n**test suites** replay past cases, including known true positives, on every change."},"faq":[{"question":"Can AI replace trade compliance analysts?","answer":"No. It removes manual look ups and drafts the case, but humans decide on escalation and on every suspicious activity report. The gains reported so far are faster turnaround and fewer false positives, claimed by a vendor for automated screening without AI; any gain in detection should be measured in your own pilot."},{"question":"What does a real deployment look like?","answer":"The clearest production example is automation, not AI: United Bank Limited in Pakistan automated vessel screening and tracking and dual use goods identification in one platform to meet the Pakistan Single Window directive, and its vendor's case study, which does not mention AI, reports faster turnaround and fewer false positives without figures. On the AI side, Microsoft, ANZ, HSBC and Lloyds showed a proof of concept at Sibos 2025 in which an AI agent parses letters of credit, and Microsoft says such agents can help flag references to sanctioned entities and ambiguous dual use descriptions. We found no bank that has published results for AI trade crime screening in production."},{"question":"Is this high risk under the EU AI Act?","answer":"No, financial crime screening is not in Annex III. It still needs strong governance because supervisors treat these models as part of the bank's financial crime controls."}],"related":["trade-document-examination","sanctions-screening-adjudication","aml-alert-triage","suspicious-activity-report-drafting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version. The catalog's claim that banks name these checks as top automation priorities was not found on the cited LexisNexis page and was dropped."},{"date":"2026-09-25","note":"Consolidation pass: added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: problem text tied to the United Bank Limited case study and the LexisNexis article, ICC guidance note corrected to what the page offers (financial crime risk control guides), MAS note marked as a November 2025 consultation, Blits.ai build text aligned with the feature inventory, FAQ adds the Microsoft proof of concept; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: adoption stage set to emerging (the only production record, United Bank Limited, is automated screening and its case study does not mention AI; the AI record is a proof of concept); metaDescription and FAQ no longer present United Bank Limited as an AI deployment; problem text dates the manual work to before the 2020 platform selection and drops the unsourced generalisation; FAQ gain claim limited to the vendor reported turnaround and false positive results; MAS guidance checked, still a consultation; removed a draft page from related."},{"date":"2026-09-27","note":"Fact checked against sources again (UBL case study and PDF, LexisNexis article dated 10 July 2025, Microsoft blog, ICC hub, MAS consultation still without final Guidelines): implementation step no longer attributes document extraction or \"fastest gains\" to United Bank Limited, FAQ softens the unprovable claim that no bank has published production results, metaDescription now names United Bank Limited and the proof of concept with HSBC."}],"slug":"trade-finance-crime-screening","url":"https://www.blits.ai/ai-use-cases/trade-finance-crime-screening","benchmarks":[],"indicativeValueResult":{"low":26666.666666666668,"high":341333.3333333334},"evidence":["anz-hsbc-lloyds-trade-finance-agent-proof-of-concept","standard-bank-automated-trade-document-checking","united-bank-limited-trade-compliance-screening"]},{"title":"AI shopping assistant for product discovery and recommendations","shortTitle":"Conversational shopping assistant","seoTitle":"AI shopping assistants for product discovery","metaDescription":"AI shopping assistants answer product questions and recommend catalog items. Amazon reports over 350 million users in a year; Zalando runs one in 25 markets.","definition":"A conversational assistant on a retailer's site or app that answers product questions, compares items and recommends products from the retailer's own catalog for a need, project or occasion described in the shopper's own words, grounded in product data, reviews and stock, and hands the shopper to a basket, a store or a human expert.","aliases":["AI shopping assistant","conversational commerce assistant","product recommendation chatbot","virtual shopping advisor","agentic shopping assistant"],"industries":["cross-industry","retail-and-ecommerce"],"functions":["sales","marketing","customer-service"],"patterns":["conversational-agent","recommendation-and-personalization","rag-knowledge-assistant","agentic-workflow"],"channels":["mobile-app","web-chat","whatsapp","voice"],"audience":"customer-facing","autonomy":"autonomous","adoptionStage":"early-adopters","problem":"Search boxes and filters work when the shopper knows the product name. They fail when the shopper\nknows the problem: \"what do I need to fix a leaky faucet\", \"what should I wear to a wedding in\nBarcelona in November\", \"will these bindings fit these boots\". In a store an experienced associate\nanswers those questions and sells the right basket; online the shopper reads reviews in ten tabs,\nguesses, or leaves.\n\nRetailers with large or technical assortments (home improvement, sporting goods, fashion,\nelectronics) feel this most, and their best experts are scarce and seasonal. Earlier product\nrecommendation engines ranked items from behaviour but could not hold a conversation or explain a\ntrade off. Generative models grounded in the catalog, reviews and how to content can, and the\nlarger retailers have now put them in front of hundreds of millions of shoppers.","problemStats":[],"howItWorks":"1. **Understand the need.** The assistant takes a free text or spoken question about a product, a\n   project or an occasion, and asks clarifying questions (skill level, budget, size, location) the\n   way a good associate would.\n2. **Retrieve from the retailer's own data.** It searches the product catalog, specifications,\n   customer reviews, questions and answers and the retailer's how to content, and checks price and\n   local stock.\n3. **Recommend and explain.** It proposes a short list or a complete basket, compares options and\n   says why each item fits, including compatibility between items.\n4. **Personalise with consent.** For logged in customers it can use purchase history and\n   preferences, and it adapts to the page the shopper is on.\n5. **Hand over to the purchase.** It adds items to the basket, reserves in store or books a\n   service, and routes complex or high value questions to a human expert.\n6. **Stay inside the rules.** Prices, promotions and availability come from live systems, never\n   from the model, and sponsored products are labelled.","valueDrivers":["revenue-growth","customer-experience","cost-to-serve","inclusion-and-access"],"kpis":["conversion-rate-uplift","revenue-uplift","users-served","customer-satisfaction","customer-satisfaction-uplift"],"indicativeValue":{"referenceOrg":"An online retailer with EUR 200 million in annual online sales","inputs":[{"key":"onlineSales","label":"Annual online sales","low":200000000,"high":200000000,"unit":"EUR per year","note":"The reference retailer."},{"key":"engagedShare","label":"Share of online sales from sessions where the shopper uses the assistant","low":0.02,"high":0.06,"unit":"fraction of online sales","note":"Editorial assumption for the first years; usage grows slowly because most shoppers still search and browse. Replace with your own adoption data."},{"key":"incrementalShare","label":"Share of those sales that is truly incremental","low":0.05,"high":0.15,"unit":"fraction of engaged sales","note":"Editorial assumption, deliberately far below the conversion multiples on this page (Sierra, the vendor, reports triple the conversion for Sun & Ski Sports shoppers who engage), because engaged shoppers are self selected. Measure with a holdout group."},{"key":"grossMargin","label":"Gross margin on incremental sales","low":0.3,"high":0.4,"unit":"fraction of sales","note":"Editorial assumption for a general merchandise retailer. Replace with your own margin."}],"formula":"onlineSales * engagedShare * incrementalShare * grossMargin","currency":"EUR","period":"per year","resultLabel":"Incremental gross margin from assisted sessions","caveat":"A rough estimate of incremental margin only. It leaves out the cost of the assistant and its model usage, service contacts avoided, the effect on returns (better advice can lower them), and it assumes a holdout test confirms the uplift."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The conversation is the easy part. The work is in clean, rich product data, live price and stock, compatibility rules, and evaluation of recommendation quality at the scale of a full catalog.","dataPrerequisites":["A product catalog with complete attributes and specifications, not only marketing copy","Customer reviews and questions and answers, where the retailer owns and may use them","Live price, promotion and stock by channel and store","How to and project content, with an owner and review date","Consent records for using purchase history in recommendations"],"integrations":["Product information management and search or recommendation engine","Pricing, promotions and inventory systems","Basket, checkout and store reservation APIs","Customer profile and consent management","Human expert chat or video channel for handover"]},"implementation":{"steps":[{"title":"Pick categories where advice matters","detail":"Start where shoppers ask compatibility or project questions and conversion is low (tools, sporting goods, fashion occasions), not in commodity categories where search already works."},{"title":"Fix the product data first","detail":"Missing attributes produce confident but wrong recommendations. Measure attribute completeness per category and fill the gaps before launch."},{"title":"Keep price, stock and promotions live","detail":"Call the pricing and inventory systems at answer time. Never let the model state a price or a discount from its own memory or from stale retrieved text."},{"title":"Build an evaluation set per category","detail":"Write real shopper questions with the right answers, including compatibility traps and questions the assistant should refuse (medical, safety critical), and run them on every change."},{"title":"Launch with a holdout group","detail":"Measure conversion, basket size and returns against shoppers who do not get the assistant, so the business case rests on incremental effect rather than on self selected users."},{"title":"Connect the assistant to the humans","detail":"Offer a human expert for high value or complex projects, and give store associates the same assistant so online and in store advice agree."}],"guardrails":["Prices, promotions and availability only from live systems, never generated","Recommendations only from the retailer's current catalog, with sponsored items labelled","Refusal and a safe pointer for safety critical, medical or legal questions","Personal data used for personalisation only with consent, and never inferred sensitive traits","Protection against prompt injection that tries to obtain unauthorised discounts or commitments"],"humanInTheLoop":"Merchandising and category experts own the content and review a sample of conversations per category every week. Human experts take over complex or high value projects on request. Any new capability that commits the retailer (adding to a basket, reserving stock, booking a service) is signed off before launch.","kpisToInstrument":["Conversion and average order value against a holdout group","Share of recommended items in stock and correctly priced at answer time","Return rate of products bought after an assisted session","Satisfaction and thumbs down rate per category","Share of sessions handed to a human expert and why"],"failureModes":[{"title":"Confident recommendations from thin data","detail":"When attributes are missing the model fills the gap with plausible text. Measure data completeness and make the assistant say when it does not know."},{"title":"Commitments the retailer did not make","detail":"Shoppers try to talk the assistant into prices, discounts or promises. Keep all commercial terms in systems of record and test for manipulation."},{"title":"Conversion claims built on self selection","detail":"Engaged shoppers were already more likely to buy. Without a holdout the business case is overstated."},{"title":"Advice beyond the catalog's safety limits","detail":"Recipes, chemicals, electrical work and health products need refusals and safety content, not creative answers."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A shopping assistant interacts directly with people, so under Article 50(1) shoppers must be informed that they are dealing with an AI system unless that is obvious. It is not listed in Annex III, so it is not high risk. Manipulative or deceptive techniques that materially distort a shopper's behaviour and cause significant harm are prohibited under Article 5(1)(a), which matters for how persuasion and urgency are designed."},"regulations":["eu-ai-act","gdpr","eu-accessibility-act"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Article 5, prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Prohibits AI that uses manipulative or deceptive techniques to materially distort behaviour in a way that causes significant harm; relevant to persuasive recommendation design."},{"title":"Unfair commercial practices directive","issuer":"European Commission","region":"europe","url":"https://commission.europa.eu/law/law-topic/consumer-protection-law/unfair-commercial-practices-law/unfair-commercial-practices-directive_en","note":"Misleading claims, hidden advertising and aggressive practices are unfair whether a person or an assistant makes them, so sponsored items and claims need the same care."},{"title":"Digital Services Act, Regulation (EU) 2022/2065, Article 27 on recommender system transparency","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2022/2065/oj/eng","note":"Online platforms that use recommender systems must set out the main parameters in their terms and conditions (Article 27), which applies to marketplaces that add a conversational recommender."}],"controls":["AI disclosure and labelling of sponsored recommendations","Price and stock answers traceable to the live system call","Evaluation set per category run on every model or content change","Consent check before any use of purchase history","Monitoring of conversations for manipulation attempts and unsafe advice"],"incidents":[{"title":"Incident 622: Chevrolet dealer chatbot agrees to sell Tahoe for $1","url":"https://incidentdatabase.ai/cite/622/","note":"A dealer's sales chatbot was talked into \"agreeing\" to an absurd price and recommending a competitor's car, showing why commercial terms must stay outside the model."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** that holds product guides and how\nto content, retrieved with hybrid search, plus **SQL knowledge bases** or **custom functions** that\nquery the product catalog, live price and stock (REST calls or SQL queries). The agent asks\nclarifying questions, recommends and compares, and the chat widget shows the result as **product\nrecommendation, retail and order cards** or a carousel; basket and reservation actions are custom\nfunctions with their own limits.\n\nThe same assistant runs on **web chat, WhatsApp, voice and a mobile app** (through the REST or\nWebSocket API channel), and can appear as a **digital human**, a photorealistic avatar streamed to\nthe browser. **Guardrails**\nblock prompt injection and apply the retailer's own policies on unsafe advice, **PII masking**\nprotects personal data, and **human handover** routes complex projects to an expert. **Test\nsuites** run category question sets on every change, **analytics** show satisfaction, response\nfeedback and top intents, and the platform is model agnostic, with **EU and UAE data residency**\nwhere needed."},"faq":[{"question":"Do AI shopping assistants increase sales?","answer":"The published figures show strong associations. Sierra, the vendor, reports that Sun & Ski Sports shoppers who engage with its agent convert at triple the rate of those who do not, and Amazon says US customers who use Alexa for Shopping spend over 40% more per order. Those shoppers are self selected, so measure the effect with a holdout group before building a business case on it."},{"question":"How many shoppers actually use them?","answer":"At the largest retailers, many. Amazon reports that over 350 million customers used its AI shopping assistant in the twelve months to its second quarter 2026 results. Zalando reported in March 2025 that over 2 million customers had used its fashion assistant, which has been live in all 25 of its markets since October 2024."},{"question":"Where should the assistant get prices and stock?","answer":"Only from live systems at the moment of the answer. Prices, promotions and availability that come from the model or from stale text lead to wrong promises, and a dealer chatbot that \"agreed\" to sell a car for one dollar shows how easily a model can be talked into commitments."}],"related":["order-status-and-returns-agent","personalized-marketing-at-scale","inbound-lead-qualification-agent","travel-and-hotel-booking-concierge","first-line-contact-centre-agent","agentic-payment-initiation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, cross industry customer facing vertical; evidence from Amazon, Zalando, Walmart, Lowe's and Sun & Ski Sports."},{"date":"2026-09-25","note":"Consolidation pass: added European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: dated the Amazon and Zalando sources and the Zalando user period, added Amazon's order value comparison, made the EU AI Act basis precise (Article 50(1), Article 5(1)(a)), aligned the Blits.ai section with the feature inventory, added SEO title and description."},{"date":"2026-09-27","note":"Review fixes: attributed the Sun & Ski Sports conversion figure to the vendor Sierra, corrected the Zalando user figure to cumulative since launch (beta from 2023), cited DSA Article 27 on EUR-Lex, tightened the Amazon figure and the digital human wording."}],"slug":"conversational-shopping-assistant","url":"https://www.blits.ai/ai-use-cases/conversational-shopping-assistant","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":176000000,"min":2000000,"max":350000000,"byClaimant":{"organization":2,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"amazon-rufus-and-alexa-for-shopping","pooled":true},{"id":"zalando-ai-fashion-assistant","pooled":true}]},{"kpi":"conversion-rate-uplift","label":"Conversion uplift","unit":"multiplier","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":3,"min":3,"max":3,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"sun-and-ski-sports-sunny-ai-agent","pooled":true}]},{"kpi":"customer-satisfaction","label":"Customer satisfaction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"sun-and-ski-sports-sunny-ai-agent","pooled":true}]},{"kpi":"customer-satisfaction-uplift","label":"Satisfaction uplift","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"sun-and-ski-sports-sunny-ai-agent","pooled":true}]}],"indicativeValueResult":{"low":60000,"high":720000},"evidence":["amazon-rufus-and-alexa-for-shopping","lowes-mylow-virtual-advisor","sun-and-ski-sports-sunny-ai-agent","walmart-sparky-shopping-assistant","zalando-ai-fashion-assistant"]},{"title":"AI spam and scam call blocking for mobile and landline subscribers","shortTitle":"Spam and scam call blocking","seoTitle":"AI scam and spam call blocking for telcos","metaDescription":"Operators use AI to block and label scam calls before customers answer. O2 labels about 70 million calls a month; Bell has blocked or labelled over 540 million.","definition":"AI in the operator's network that protects subscribers from unwanted calls: it analyses incoming calls in real time, blocks known fraudulent calls, and labels suspected scam, spam and spoofed calls on the customer's screen before they answer, so subscribers can decide whether to pick up. Fraud against the operator itself, such as SIM swap or revenue share fraud, is a separate use case.","aliases":["scam call protection","spam call labelling","robocall blocking","spoofed call detection","caller risk warning"],"industries":["telecommunications"],"functions":["fraud-prevention","customer-service"],"patterns":["anomaly-detection","classification-and-routing","prediction-and-scoring"],"channels":["voice","mobile-app","api"],"audience":"customer-facing","autonomy":"autonomous","adoptionStage":"mainstream","segment":"customer-protection","problem":"Phone scams are one of the main ways criminals reach victims. Callers pose as a bank, a tax\nauthority, an online retailer or the operator itself, often with a spoofed number that looks local\nor familiar, and push people to hand over details or move money. In the UK, Virgin Media O2 and\nHiya found fake Amazon, HMRC and banking calls at the top of the list of nuisance calls in early\n2026, and in Australia Telstra cites the ACCC's\nfinding that phone scams accounted for the highest overall financial losses among all contact\nmethods in Australia in 2024. The side effect is that people stop answering unknown numbers: in\nTelstra's research, 42% of Australians with a mobile device say they are less likely to answer\ncalls because of scam fears, which also hurts the legitimate organizations that need to reach them.\n\nScammers constantly adapt their numbers and tactics, so static block lists fall behind. Operators\nalso have to avoid blocking genuine calls, which is why some malicious calls still slip through.\nThe network sees signals no single phone can see: how many calls a number makes and whether a call\nclaiming a local number actually arrives from abroad. Using those signals at scale, in real time,\nis where machine learning helps; BT says its vendor Hiya uses machine learning to improve scam\ndetection the more malicious calls it encounters.","problemStats":[{"statement":"Telstra cites the ACCC's Targeting Scams report, under which phone scams accounted for the highest overall financial losses among all contact methods in Australia in 2024, with AUD 107.2 million reported lost across 2,179 reports.","sourceTitle":"Suspicious phone calls: what Telstra is doing to raise the alarm","sourceUrl":"https://www.telstra.com.au/exchange/suspicious-phone-calls--what-telstra-is-doing-to-raise-the-alarm","year":2025},{"statement":"Telstra reports research with YouGov under which 42% of Australians who own a mobile device are less likely to answer calls out of concern about being scammed.","sourceTitle":"Suspicious phone calls: what Telstra is doing to raise the alarm","sourceUrl":"https://www.telstra.com.au/exchange/suspicious-phone-calls--what-telstra-is-doing-to-raise-the-alarm","year":2025},{"statement":"O2 cites Hiya's State of the Call report, under which 16% of UK consumers fell victim to phone scams in the previous year, losing an average of GBP 798 each.","sourceTitle":"O2 launches free AI-powered scam call detection service to help combat fraud and nuisance calls","sourceUrl":"https://news.virginmediao2.co.uk/o2-launches-free-ai-powered-scam-call-detection-service-to-help-combat-fraud-and-nuisance-calls/","year":2024}],"howItWorks":"1. **Analyse every unknown call.** When a call arrives from an unknown number, the model scores it\n   in real time on the behaviour of the calling number (for example a high volume of calls from a\n   single number), where the call really comes from, customer reports and other data points.\n2. **Detect spoofing.** Models and network checks flag calls whose presented number does not match\n   where the call really comes from, such as an overseas call showing a local mobile number.\n3. **Block or label.** Known fraudulent calls are blocked or diverted to voicemail; suspected scam or\n   spam calls are delivered with a warning label on the handset or landline display; verified\n   businesses can show their name.\n4. **Learn from reports.** Customer reports (for example to the 7726 short code in the UK) and\n   investigations are used to block the numbers behind them and to refine the blocking services,\n   so new scam trends are identified and blocked faster.\n5. **Trace and shut down.** Operators work with other carriers and regulators to trace scam calls\n   back to their origin and stop the parties bringing them into the network.","valueDrivers":["risk-reduction","customer-experience","inclusion-and-access"],"kpis":["interactions-handled","users-served","fraud-loss-reduction","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A mobile operator with 5 million subscribers","inputs":[{"key":"subscribers","label":"Subscribers protected","low":5000000,"high":5000000,"unit":"subscribers","note":"The reference operator."},{"key":"victimRate","label":"Share of subscribers who lose money to a phone scam in a year","low":0.002,"high":0.005,"unit":"fraction of subscribers","note":"Editorial assumption, deliberately far below the survey figure cited on this page (Hiya reports 16% of UK consumers fell victim to phone scams), because survey victimisation includes small and unreported losses."},{"key":"averageLoss","label":"Average loss per victim","low":500,"high":1000,"unit":"USD per victim","note":"Editorial assumption; Hiya's survey cited on this page reports an average loss of GBP 798 per UK victim."},{"key":"preventedShare","label":"Share of those losses prevented by blocking and warnings","low":0.1,"high":0.3,"unit":"fraction of losses","note":"Editorial assumption. O2 reports that calls labelled suspected scam are answered 42% less often; not every unanswered scam call is a prevented loss."}],"formula":"subscribers * victimRate * averageLoss * preventedShare","currency":"USD","period":"per year","resultLabel":"Customer scam losses prevented","caveat":"Customer losses only, and a rough order of magnitude. It leaves out the operator's savings on scam related complaints and contacts, the value of customers trusting calls again, and the cost of genuine calls that are wrongly labelled or blocked."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Specialist vendors can provide the scoring and labelling (BT and O2 use Hiya), while Telstra and Bell describe their own network capabilities. Either way, the work is integration with the voice network, handset support, handling of wrongly labelled businesses, and regulatory alignment on blocking.","dataPrerequisites":["Real time call signalling and call detail records","Caller authentication data where available (for example STIR/SHAKEN in North America)","Customer scam reports and complaints","Registry of verified business numbers"],"integrations":["Voice core and signalling platforms for blocking and diversion","Handset or network based caller display for labels","Customer reporting channels such as the 7726 short code in the UK","Industry traceback and intelligence sharing with other carriers and regulators"]},"implementation":{"steps":[{"title":"Start with the network signals you control","detail":"Block numbers that should never originate calls and calls that present a domestic number but arrive from abroad, before adding model based scoring."},{"title":"Label before you block","detail":"For uncertain calls, a warning label lets the customer decide and gives you feedback. Block only above a high confidence threshold."},{"title":"Handle false positives fast","detail":"Give businesses a way to register numbers and dispute labels, and give customers a way to report missed scams and wrongly flagged calls."},{"title":"Cover every customer group","detail":"Extend protection to landlines and older handsets, where vulnerable customers are concentrated, not only to the newest smartphones."},{"title":"Measure harm, not only volume","detail":"Track answer rates on labelled calls and scam reports per thousand customers, not just the number of calls blocked."}],"guardrails":["Blocking only above a validated confidence threshold; everything else is labelled, not blocked","Emergency and priority numbers never blocked","A dispute process for businesses whose calls are wrongly labelled","Clear customer information about what is analysed and how to switch labelling off where permitted"],"humanInTheLoop":"Fraud analysts set blocking thresholds and review new campaign patterns, a team handles disputes from businesses, and customer reports are treated as training signals. Blocking rules follow the national regulator's requirements.","kpisToInstrument":["Calls blocked and labelled per month, per category","Answer rate and call duration for labelled versus unlabelled calls","Scam reports per thousand customers","Disputes from businesses and share upheld","Share of customers covered, including landlines and older devices"],"failureModes":[{"title":"Genuine calls marked as scam","detail":"Hospitals, schools or delivery firms get labelled and people stop answering them. Run a fast dispute process and verified caller programmes."},{"title":"Volume as the only measure","detail":"Billions of calls blocked says little about harm prevented. Track answer rates and scam reports."},{"title":"Protection only for new phones","detail":"Handset based labels miss older devices and landlines. Combine network blocking with display features for all lines."},{"title":"Scammers move to other channels","detail":"Scammers shift between calls, texts and messaging apps as each channel gets harder to use. Coordinate call and message protection."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Scoring, blocking and labelling calls is not listed in Annex III, is not a prohibited practice under Article 5, and the system does not interact with people or generate content, so Article 50 does not apply. A conversational scambaiting agent such as O2's Daisy talks to callers with a synthetic voice, which raises separate Article 50 transparency questions and should be assessed on its own."},"regulations":["telecom-consumer-rules","gdpr","uk-gdpr","eu-ai-act","eecc","au-scams-prevention-framework"],"guidance":[{"title":"Scam calls and messages","issuer":"Ofcom","region":"europe","url":"https://www.ofcom.org.uk/phones-and-broadband/scam-calls-and-messages","note":"UK regulator hub on scam calls and messages, including its statement on tackling scam calls from abroad, which covers its calling line identification guidance on how providers should process calls from abroad that present a UK mobile number."}],"controls":["Documented blocking and labelling policy aligned with national rules","Monitoring of false positives and business disputes","Privacy notice and legal basis for analysing call metadata","Regular review of thresholds against new scam campaigns"],"incidents":[]},"blitsAi":{"howToBuild":"Call scoring and labelling run in the voice network with specialist vendors. Blits.ai covers the\nconversations around it: a **voice or chat agent** on **web chat, WhatsApp, SMS and the phone**\nhelps customers report a scam, check whether a call was genuine, and get advice on what to do\nnext, from a **knowledge base** of approved scam guidance with hybrid retrieval, and hands\ndistressed or vulnerable callers to a human with **human handover**.\n\nA second **AI agent** can help business customers dispute a wrong label, collecting evidence in\na **flow** and routing it to the fraud team through **custom functions**. **Guardrails** stop the\nagent from giving advice outside approved content, **PII masking** protects numbers and personal\ndata, and **analytics** show report volumes and themes."},"faq":[{"question":"How many calls do operators block or label?","answer":"Bell reports analysing more than 4.4 billion calls and blocking or labelling over 540 million since launching Suspicious Call Detection in 2025. Virgin Media O2 labels around 70 million suspected scam and spam calls a month. Telstra reports blocking more than 11 million scam calls a month on average and showing Scam Protect warnings on about 12 million calls a month."},{"question":"Do warning labels actually change behaviour?","answer":"Virgin Media O2 reports that calls labelled suspected scam are answered 42% less often and last 89% less time than unflagged calls."},{"question":"Can AI also fight back against scammers?","answer":"O2 built Daisy, an AI voice persona that answers scam calls and keeps fraudsters talking, in some cases for 40 minutes, to waste their time and expose their tactics. It is an awareness and disruption tool rather than a replacement for network blocking."}],"related":["telecom-fraud-detection","scam-payment-interception","mule-network-detection"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched for the telecommunications vertical and verified against operator sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added European Electronic Communications Code, Australian Scams Prevention Framework, MAS Shared Responsibility Framework to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: rewrote unsourced details in the problem and flow (call duration, answer rates, catching campaigns within hours, delivery firms); added Telstra's 42% answer hesitancy statistic and two 2025 Telstra metrics; corrected the Bell and Telstra evidence summaries; noted that Telstra and Bell describe their own capabilities; removed the MAS Shared Responsibility Framework (its telco duties cover phishing SMS, not calls) and added UK GDPR; sharpened the EU AI Act basis and the Ofcom guidance note; added the SEO title and meta description."}],"slug":"spam-and-scam-call-blocking","url":"https://www.blits.ai/ai-use-cases/spam-and-scam-call-blocking","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":3,"nUpTo":0,"median":70000000,"min":2430000,"max":540000000,"byClaimant":{"organization":3,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bell-suspicious-call-detection","pooled":true},{"id":"virgin-media-o2-call-defence","pooled":true},{"id":"bt-enhanced-call-protect","pooled":true}]},{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2500000,"min":2500000,"max":2500000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"bt-enhanced-call-protect","pooled":true}]}],"indicativeValueResult":{"low":500000,"high":7500000},"evidence":["bell-suspicious-call-detection","bt-enhanced-call-protect","telstra-scam-call-blocking","virgin-media-o2-call-defence","virgin-media-o2-daisy-ai-scambaiter"]},{"title":"AI spend classification and spend analytics for procurement","shortTitle":"Spend classification","seoTitle":"AI spend classification for procurement","metaDescription":"AI assigns purchase lines to spend categories so buyers see what they buy. GSA uses it for category management and the VHA for executive spend analysis.","definition":"AI that reads purchase orders, invoices, card transactions and contracts and assigns each line of spend to a category in the organization's taxonomy, and to the right supplier, so that procurement can see what is bought, from whom and where to consolidate or negotiate.","aliases":["spend analytics AI","AI spend categorization","spend cube automation","UNSPSC classification with AI","procurement category classification"],"industries":["cross-industry","government","manufacturing","healthcare"],"functions":["procurement","finance-and-accounting","analytics-and-reporting"],"patterns":["classification-and-routing","document-processing","summarization"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Procurement can only manage what it can see, and raw spend data does not show it on its own.\nPurchases can arrive from several ERP systems, purchasing cards and expense tools, with free\ntext descriptions (\"gloves blk L 100\"), inconsistent supplier names and general ledger codes that\ndescribe the budget, not the thing bought. Building a reliable view of spend by category means\ncleaning and classifying large volumes of lines. Where that is done by hand in a periodic\nexercise, the view can be out of date by the time it is finished.\n\nThe consequences are practical. Category managers cannot tell how much the organization spends on\na category across units, so they cannot consolidate demand or negotiate on volume, contract\ncompliance and maverick buying go unmeasured, and savings claims are hard to prove. The US federal\ngovernment manages its buying through a government wide category management taxonomy. In the\n2025 inventory, USDA says that to plan for the upcoming fire season all of the previous year's\nincident related purchases are categorized by hand, a process it describes as time consuming; it\nhas piloted a machine learning classifier for this since December 2024.","problemStats":[],"howItWorks":"1. **Gather and clean.** Purchase order lines, invoice lines, card transactions and contract\n   records are extracted from each source system, and supplier names are normalized and matched\n   to one supplier record.\n2. **Classify each line.** A model trained on lines already labeled by buyers, or a language\n   model given the taxonomy with definitions and examples, assigns each line to a category and a\n   subcategory and returns a confidence score.\n3. **Read the contracts.** For contract spend, the AI reads the contract document and summarizes\n   what goods or services it covers, which gives category managers a view of what each contract\n   buys. The IRS uses generative AI to summarize its contracts in support of category management.\n4. **Review the uncertain.** Low confidence lines and high value lines go to a category analyst,\n   whose corrections are fed back as training examples.\n5. **Analyze.** The classified spend feeds dashboards by category, supplier, unit and period,\n   showing consolidation opportunities, contract coverage and trends.","valueDrivers":["employee-productivity","cost-to-serve","speed","compliance"],"kpis":["automation-rate","accuracy","hours-saved","cost-savings","processing-time-reduction"],"indicativeValue":{"referenceOrg":"An organization with 2 million purchase lines a year across several ERP and card systems","inputs":[{"key":"lines","label":"Purchase, invoice and card lines per year","low":2000000,"high":2000000,"unit":"lines per year","note":"The reference organization. Replace with your own volumes."},{"key":"manualShare","label":"Share of lines that need manual classification today","low":0.2,"high":0.4,"unit":"fraction of lines","note":"Editorial assumption, replace with the share your rules or suppliers do not classify."},{"key":"automatedShare","label":"Share of that manual work the AI takes over","low":0.5,"high":0.8,"unit":"fraction of manual lines","note":"Editorial assumption; analysts still review low confidence and high value lines."},{"key":"minutesPerLine","label":"Analyst minutes per manually classified line","low":0.5,"high":1,"unit":"minutes per line","note":"Editorial assumption for a trained analyst working in batches."},{"key":"hourlyCost","label":"Fully loaded analyst hour","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption."}],"formula":"lines * manualShare * automatedShare * minutesPerLine / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Classification labor avoided","caveat":"Counts only the labor of classifying spend. It leaves out the larger but harder to attribute value of better category decisions (consolidation, negotiation, contract compliance), the cost of the platform and model, and the one time work of building the taxonomy and training data."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The model is the easy part. The work is in extracting and joining data from several source systems, agreeing a taxonomy with clear definitions, cleaning supplier names and building a labeled sample that reflects the organization's own language.","dataPrerequisites":["Purchase order, invoice and card line data with free text descriptions","A spend taxonomy with definitions and examples per category","A supplier master, or at least a supplier normalization table","A labeled sample of lines classified by experienced buyers"],"integrations":["ERP and purchasing systems","Purchasing card and expense platforms","Contract repository","Analytics or business intelligence tools for the spend dashboards"]},"implementation":{"steps":[{"title":"Fix the taxonomy first","detail":"Choose the taxonomy (your own, UNSPSC or a public sector category structure) and write a definition and examples for each category. A model cannot be more consistent than the definitions it is given."},{"title":"Build a labeled benchmark","detail":"Have buyers classify a random sample of a few thousand lines, stratified by source and value, and keep it aside to measure accuracy on every change."},{"title":"Classify with confidence thresholds","detail":"Let the model classify all lines, accept high confidence lines automatically and route low confidence and high value lines to analysts. Tune the threshold on the benchmark."},{"title":"Close the loop","detail":"Feed analyst corrections back into training data or prompt examples, and rerun the benchmark before each release."},{"title":"Put the result to work","detail":"Connect classified spend to category plans and supplier negotiations, and report savings by category, so the classification earns its keep."}],"guardrails":["Every line keeps its original description and source, so a classification can be traced and corrected","Confidence score on every line, with low confidence and high value lines reviewed by a person","Accuracy measured on a fixed benchmark before each change to the model or taxonomy","Taxonomy changes versioned, with reclassification of history when definitions change"],"humanInTheLoop":"Category analysts own the taxonomy, review low confidence and high value lines and sample accepted lines each period. Classifications inform decisions but do not trigger purchases or payments.","kpisToInstrument":["Share of spend value and of lines classified automatically","Accuracy on the labeled benchmark, by category and source","Analyst hours spent on classification per period","Time from period end to an updated spend view","Savings identified and realized per category"],"failureModes":[{"title":"High accuracy by line, low accuracy by value","detail":"The model classifies many small lines well and a few large ones badly. Measure accuracy weighted by spend value and review large lines by hand."},{"title":"Taxonomy drift","detail":"Categories are renamed or split but old spend is not reclassified, so trends break. Version the taxonomy and reclassify history."},{"title":"Garbage descriptions","detail":"Lines with empty or generic descriptions cannot be classified reliably by any model. Use the supplier, contract and ledger code as extra signals and fix the capture at the source."},{"title":"Dashboards nobody uses","detail":"Spend is classified but category plans do not change. Tie the output to specific sourcing decisions from the start."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Classifying the organization's own purchase lines into categories is not listed in Annex III and is used internally by procurement staff, so no specific obligations apply beyond AI literacy. The data can still contain personal data, for example in purchasing card and expense lines, which brings GDPR duties. Using the classified card and expense lines to monitor or evaluate individual employees would move the system towards Annex III point 4 (employment and worker management) and a high risk assessment."},"regulations":["eu-ai-act","gdpr","iso-42001","nist-ai-rmf"],"guidance":[{"title":"M-19-13: Category Management: Making Smarter Use of Common Contract Solutions and Practices","issuer":"Office of Management and Budget","region":"north-america","url":"https://www.whitehouse.gov/wp-content/uploads/2019/03/M-19-13.pdf","note":"The OMB memorandum that implements category management government wide, defines the role of the Category Management Leadership Council and asks agencies to use spending data and data analytics tools to make data driven buying decisions. It is not specific to AI."},{"title":"Category management","issuer":"U.S. General Services Administration","region":"north-america","url":"https://www.gsa.gov/buy-through-us/category-management","note":"Describes category management as identifying categories of spend and using data to consolidate contracts, manage suppliers and demand, and reduce total cost of ownership."}],"controls":["Named owner for the taxonomy and for classification quality","Benchmark results and model versions kept with each release","Masking or exclusion of personal data in expense and card lines before classification","Periodic sample audit of automatically accepted lines"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this runs as an **agentic workflow** on a schedule or through the **REST API**:\n**custom functions** read new purchase, invoice and card lines from the ERP, or from a **SQL\nknowledge base** that holds the extract, and an **agent** with the taxonomy definitions in its\n**knowledge base** returns a category, a confidence and a short reason per line as **structured\noutput**. The workflow routes lines below a confidence score or above a value limit to an\nanalyst, and **human in the loop** confirmation can be required before results are written back\nto the ERP. Every run keeps a full **audit trail**.\n\n**Test suites** built from the buyers' labeled sample act as the accuracy benchmark on each\nchange of model, prompt or taxonomy, and the platform is **model agnostic**, so a cheaper model\ncan classify routine lines and a stronger one the hard cases. **PII masking** applies to traffic\nthrough the gateway; personal data that custom functions pull from expense and card lines should\nalso be masked in the extract before it reaches a model. Category managers can then query the\nclassified results in plain language through an **agent** connected to the SQL knowledge base."},"faq":[{"question":"Who uses AI for spend classification?","answer":"In the US federal government, GSA classifies transactions into the government wide category management taxonomy, the Veterans Health Administration uses generative AI to categorize purchase order lines for an executive spend dashboard, and the IRS has generative AI summarize what each contract buys. None of the inventory entries reports accuracy or savings figures, so measure your own."},{"question":"Should we use a language model or a trained classifier?","answer":"Both work. A classifier trained on your own labeled lines is cheap and fast at volume; a language model given the taxonomy and examples needs less training data and copes better with new categories. One option is to use the classifier for routine lines and a language model for uncertain ones; whichever you choose, compare both on the same benchmark."},{"question":"How accurate does it need to be?","answer":"Accurate enough by value for the decisions it supports. Measure accuracy weighted by spend, review high value lines by hand, and report the share of spend classified with high confidence rather than a single accuracy number."}],"related":["supplier-invoice-processing","procurement-contract-review","vendor-due-diligence","ledger-and-payment-reconciliation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the discovery workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched and written with evidence from GSA, the Veterans Health Administration, the IRS and USDA as recorded in the 2025 US federal AI use case inventory. Edited after review: unsupported practice claims rewritten as advice, OMB M-19-13 cited as the category management guidance, American spelling throughout. Second review: unsourced claims about how common the problem and manual practice are removed, the USDA example tied to its 2024 pilot, FAQ and Blits.ai build notes narrowed to what the sources and platform support. Third pass: the USDA manual practice restated as the inventory states it, and the opening claim about spend data quality softened."}],"slug":"procurement-spend-classification","url":"https://www.blits.ai/ai-use-cases/procurement-spend-classification","benchmarks":[],"indicativeValueResult":{"low":66666.66666666667,"high":746666.6666666666},"evidence":["gsa-acquisition-analytics-spend-classification","irs-procurement-contract-spend-summaries","usda-incident-procurement-classification","va-executive-spend-analysis"]},{"title":"AI summaries of investment research and the house view","shortTitle":"Research summaries","seoTitle":"AI investment research summaries for advisors","metaDescription":"AI assistants turn research and the house view into cited briefings. Deutsche Bank analysts report saving up to two hours per research report with DB Lumina.","definition":"An AI assistant that condenses long research reports, overnight market moves and the house view into short, sourced briefings for advisors and analysts, answers \"what is our view on X\" on demand, and adapts approved research for different client segments and languages, with every figure traced to the original research.","aliases":["research summarization assistant","house view briefing generator","research question answering for sales and advisors","market commentary summarizer"],"industries":["wealth-and-asset-management","capital-markets","banking"],"functions":["analytics-and-reporting","sales","knowledge-management"],"patterns":["summarization","rag-knowledge-assistant","content-generation","translation"],"channels":["internal-tools","email","microsoft-teams"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"front-office","problem":"Morgan Stanley alone publishes more than 70,000 proprietary research reports a year. No advisor or\nsalesperson can read that, so client conversations lean on the few notes someone happened to see,\nand the firm's own view reaches clients unevenly. Analysts, in turn, spend much of their time on the\nmechanical parts of writing: sifting through financial statements, regulatory filings and industry\nreports, and summarizing earnings releases and investor transcripts.\n\nRewriting research for segments and languages multiplies the work, and every rewrite is a chance\nfor a number to drift from the approved report. The job is to make the research usable at the\nmoment of need without changing what it says.","problemStats":[],"howItWorks":"1. **Ingest approved research.** Published reports, the house view, earnings summaries and market\n   notes are indexed with their publication date, author and distribution rules.\n2. **Answer and summarize on demand.** An advisor or salesperson asks a question or requests a\n   briefing; the assistant retrieves the relevant passages and writes a short answer with links to\n   each source report.\n3. **Produce standard briefings.** A morning note, a sector summary or a \"what changed\" digest is\n   generated from a template, with figures and price targets copied from the source rather than\n   generated.\n4. **Adapt for audience and language.** Approved summaries are rewritten for a client segment or\n   translated, and the adapted version is checked against the source before use.\n5. **Review before anything branded goes out.** Research or compliance signs off on any client\n   facing summary; internal answers carry citations so the reader can verify them.","valueDrivers":["employee-productivity","speed","customer-experience","compliance"],"kpis":["time-saved-per-task","users-served","response-time-reduction","employee-adoption","accuracy"],"indicativeValue":{"referenceOrg":"A research and advisory team of 100 analysts and strategists","inputs":[{"key":"analysts","label":"Analysts and strategists producing research","low":100,"high":100,"unit":"people","note":"The reference team."},{"key":"documentsPerYear","label":"Notes and reports per person per year","low":50,"high":100,"unit":"documents per person per year","note":"Editorial assumption, replace with your own publication volumes."},{"key":"hoursSaved","label":"Hours saved per document","low":0.5,"high":0.75,"unit":"hours per document","note":"In line with the Deutsche Bank figure of 30 to 45 minutes saved on earnings note templates. The reported ceiling of up to two hours applies to full research reports and roadshow updates and is not assumed here."},{"key":"hourlyCost","label":"Fully loaded analyst cost per hour","low":100,"high":200,"unit":"USD per hour","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"analysts * documentsPerYear * hoursSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Value of analyst time released from summarizing and drafting","caveat":"Covers production time only. It leaves out the value on the distribution side (faster answers to client questions), the cost of running the tool and the review effort for client facing output."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Retrieval and summarization are straightforward. The effort is in distribution rights (which research may be shown to whom), keeping numbers exact, and a review workflow for anything that reaches clients under the firm's name.","dataPrerequisites":["Research archive with publication dates, authors, ratings and distribution permissions","Current house view and CIO publications","Market data feeds where briefings reference prices or moves","Style guides and disclaimers per client segment and market"],"integrations":["Research publishing platform and archive","Market data provider","Document management and translation workflow","Advisor or sales desktop, email and collaboration tools"]},"implementation":{"steps":[{"title":"Start internal, with citations","detail":"Launch as an internal question answering tool over published research, where every answer links to the source report. Internal use builds trust and shows which questions matter."},{"title":"Copy numbers, never generate them","detail":"Extract figures, ratings and price targets from the source with structured extraction and insert them into the text, then check the final output against the source automatically."},{"title":"Respect distribution rules","detail":"Filter research by audience, market and embargo before retrieval so restricted or institutional only content never appears in an answer for the wrong reader."},{"title":"Add templated briefings","detail":"Once answers are reliable, generate recurring digests (morning note, weekly house view changes) from templates owned by the research team."},{"title":"Put client facing output through review","detail":"Any summary sent to clients goes through the same approval as other research or marketing communications, with the AI draft and the source retained."}],"guardrails":["Answers only from published, approved research with citations and dates","Numbers, ratings and price targets copied from source and verified, never generated","Distribution and embargo rules applied before retrieval","Human sign off before any branded or client facing summary is sent","Refusal to give personalized recommendations; the assistant summarizes the firm's view"],"humanInTheLoop":"Analysts approve summaries of their own work before external use; research management or compliance signs off on client facing templates; readers of internal answers verify through the cited source.","kpisToInstrument":["Share of answers with a valid citation and a matching figure check","Time from report publication to advisor ready summary","Weekly active users by desk","Error rate on a monthly sample checked by analysts","Client facing summaries rejected at review, with reasons"],"failureModes":[{"title":"Drifting numbers","detail":"A price target or percentage in the summary differs from the report. Copy numbers from structured extraction and check them automatically."},{"title":"Stale view presented as current","detail":"An older note outranks the latest update. Weight recency, show dates and retire superseded views."},{"title":"Distribution breach","detail":"Institutional or embargoed research reaches a retail audience. Filter by entitlement before retrieval."},{"title":"Summaries that read as advice","detail":"A generic summary is sent to a client as if it were personal advice. Keep client facing use behind review and templates."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Summarizing research for staff is not an Annex III use and is not a practice prohibited by Article 5, so the tier turns on the firm's role under Article 50. It is minimal for a purchased internal tool with no client or public facing exposure. Article 50 transparency applies when the firm builds the generating system itself, which brings the Article 50(2) duty to mark synthetic text in a machine readable format; when the assistant is offered to clients as a chatbot, which brings the Article 50(1) duty to tell them they are interacting with AI; or when AI generated text is published to inform the public on matters of public interest, which brings the Article 50(4) disclosure duty unless the text has gone through human review or editorial control and a person holds editorial responsibility for it."},"regulations":["eu-ai-act","gdpr","dora","mas-ai-risk-management","iso-42001","mifid-ii","eu-mar"],"guidance":[{"title":"ESMA public statement on the use of AI in the provision of retail investment services","issuer":"European Securities and Markets Authority","region":"europe","url":"https://www.esma.europa.eu/sites/default/files/2024-05/ESMA35-335435667-5924__Public_Statement_on_AI_and_investment_services.pdf","note":"Names hallucination and overreliance as risks and expects accuracy controls and records when AI supports investment services."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Providers must mark AI generated text in a machine readable format (paragraph 2); the disclosure duty for text published on matters of public interest does not apply after human review under editorial responsibility (paragraph 4)."},{"title":"Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT)","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2018/feat","note":"Principles for the responsible use of AI and data analytics by Singapore financial firms, including accountability for AI driven outputs and transparency to customers about AI use."}],"controls":["Source traceability for every summary, retained with the output","Automated figure check against the source before release","Review and approval workflow for client facing summaries","Entitlement and embargo filtering tested on every change","Inventory entry with owners in research and distribution"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the research archive sits in a **knowledge base** with **hybrid retrieval**, loaded\nfrom PDF, Word and PowerPoint reports or crawled from the research portal, with version control in\nthe central document library. An **AI agent** answers questions from the retrieved research, using\n**structured output** to return the answer, its figures and the source report in separate fields,\nso a **custom function** can check each figure against the source before release. An **agentic\nworkflow** triggered on a schedule produces the recurring briefings.\n\nSending a client facing draft is an agentic action that waits for **human in the loop** approval.\n**Machine translation** adapts approved summaries per market, and the translated version goes\nthrough the same figure check and review.\n**Guardrails** block personalized recommendations, **test suites** check answers and figures on\nevery change, and the platform is **model agnostic**, so the firm can switch models per agent and\nuse regional model routing to keep data in the EU or UAE region."},"faq":[{"question":"Will the AI invent numbers or price targets?","answer":"It can, if you let it write numbers freely. The safe design copies every figure from the source report, checks the output against it, and keeps citations visible. Deutsche Bank's DB Lumina, for example, grounds answers in internal research with inline citations and source viewers."},{"question":"How much time does it save?","answer":"Deutsche Bank analysts report saving 30 to 45 minutes on earnings note templates and up to two hours on research reports and roadshow updates. On the distribution side, Morgan Stanley's global director of research told CNBC that a salesperson needs one tenth of the time to answer the average client inquiry with AskResearchGPT."},{"question":"Can summaries go straight to clients?","answer":"Only through the same review as other research and marketing communications. Internal use with citations is the low risk starting point; client facing output needs templates and sign off."}],"related":["wealth-advisor-knowledge-assistant","portfolio-reporting-and-commentary","next-best-action-for-advisors","marketing-content-compliance-copilot","client-briefing-and-call-report-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog (Investment Research Summaries) with evidence verified against public sources."},{"date":"2026-09-25","note":"Consolidation pass: added MiFID II, EU Market Abuse Regulation to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added the SEO title and description, sharpened the problem wording to match the Deutsche Bank source, restated the EU AI Act basis by Article 50 paragraph, corrected the MAS FEAT note, limited the Blits.ai build notes to listed capabilities, and updated both owned evidence records (DB Lumina stage and source date, AskResearchGPT channels and archive link)."},{"date":"2026-09-27","note":"Review fixes: corrected the Article 50 note and added the Article 50(2) marking duty for firms that build their own system, lowered the high hours saved assumption to 0.75 to match the 30 to 45 minute earnings note figure (high scenario now USD 1.5 million), restated the problem paragraph to what the Morgan Stanley and Deutsche Bank sources say, and dropped multi language authoring from the build notes."},{"date":"2026-09-27","note":"Review fix: changed the EU AI Act tier from minimal to context dependent, since the design's basis already ties the outcome to the firm's role under Article 50."}],"slug":"investment-research-summarization","url":"https://www.blits.ai/ai-use-cases/investment-research-summarization","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":5000,"min":5000,"max":5000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"deutsche-bank-db-lumina-research","pooled":true}]},{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":0,"nUpTo":1,"median":null,"min":null,"max":null,"byClaimant":{"organization":0,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"deutsche-bank-db-lumina-research","pooled":false}]}],"indicativeValueResult":{"low":250000,"high":1500000},"evidence":["citi-wealth-askwealth-and-advisor-insights","deutsche-bank-db-lumina-research","morgan-stanley-askresearchgpt","ubs-red-client-advisor-assistants"]},{"title":"AI summarization of medical evidence for life and health underwriting","shortTitle":"Life underwriting medical summaries","seoTitle":"AI medical record summaries for life underwriting","metaDescription":"AI turns medical records into cited summaries for life underwriters. Manulife uses generative AI to digitize and summarize underwriting documents in Singapore.","definition":"AI that reads the medical evidence behind a life or health insurance application (attending physician statements, electronic health records, lab results and disclosures), turns it into a structured, cited summary of conditions, treatments and dates, and maps it to the insurer's underwriting manual so an underwriter can decide faster and more consistently.","aliases":["attending physician statement summarization","APS summarization","medical evidence review for underwriting"],"industries":["insurance"],"functions":["underwriting"],"patterns":["document-processing","summarization","rag-knowledge-assistant"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"underwriting","problem":"For fully underwritten life and health cover, much of the work is reading medical evidence. An\nattending physician statement or a set of health records can be long, mixing handwritten notes,\nscans and lab printouts in no particular order. Underwriters or nurse reviewers read it to find\nthe handful of facts that matter (diagnoses, dates, medications, test values, smoking status) and\nthen look them up in the insurer's underwriting manual.\n\nThat reading is slow, expensive and can differ between reviewers, and applicants wait while it\nhappens. Electronic health records can make evidence available sooner, but someone still has to\nread it.","problemStats":[],"howItWorks":"1. **Collect and digitize.** Medical records arrive from providers, labs and record retrieval\n   vendors; the system converts scans and handwriting to text and splits the file into encounters.\n2. **Extract clinical facts.** A model extracts diagnoses, procedures, medications, vitals and lab\n   values with dates, and codes them to a standard vocabulary, each linked to its source page.\n3. **Build the timeline.** Facts are ordered into a timeline and checked against the applicant's\n   own disclosures to highlight differences.\n4. **Map to the manual.** Retrieval over the underwriting manual suggests the relevant impairment\n   guidance and any further evidence needed, without setting the final rating.\n5. **Underwriter decides.** The underwriter reviews the summary, opens the source pages where it\n   matters and makes the decision; simple, clean cases can go to straight through rules that the\n   insurer already governs.","valueDrivers":["speed","employee-productivity","customer-experience","compliance"],"kpis":["processing-time-reduction","time-saved-per-task","automation-rate","cycle-time-days","accuracy"],"indicativeValue":{"referenceOrg":"A life insurer that fully underwrites 20,000 applications a year with medical records","inputs":[{"key":"applications","label":"Applications with medical records reviewed per year","low":20000,"high":20000,"unit":"applications per year","note":"The reference insurer."},{"key":"reviewHours","label":"Underwriter or nurse reading time per file","low":1,"high":2,"unit":"hours per file","note":"Editorial assumption. Replace with a time study of your own medical evidence review."},{"key":"timeSavedShare","label":"Share of reading time the summary removes","low":0.3,"high":0.5,"unit":"fraction of reading time","note":"Editorial assumption. No insurer on this page publishes a reading time figure, so replace it with the result of your own parallel run."},{"key":"costPerHour","label":"Fully loaded cost of an underwriting hour","low":50,"high":80,"unit":"USD per hour","note":"Editorial assumption. Replace with your own fully loaded cost."}],"formula":"applications * reviewHours * timeSavedShare * costPerHour","currency":"USD","period":"per year","resultLabel":"Underwriting review time released","caveat":"Time released only. It leaves out the usually larger effect of faster decisions on placement rates (fewer applicants dropping out), record retrieval costs, platform costs and the effort of validating the model for a high risk use."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Clinical extraction from poor scans is hard, errors carry real consequences for applicants, and in the EU the use is high risk under the AI Act. Special category health data raises the bar on data protection, residency and access control. Expect formal model validation and a long parallel run.","dataPrerequisites":["Historical medical files with the underwriting decisions made on them, for testing","The underwriting manual (the insurer's own or a reinsurer's) in a form the insurer may use for retrieval","A clinical vocabulary and mapping for impairments used in the manual","Consent and authorization records for every medical record used"],"integrations":["New business and underwriting workbench","Medical record retrieval vendors and electronic health record sources","Document management with page level retention","Rules engine for straight through decisions"]},"implementation":{"steps":[{"title":"Start with summarization, not decisions","detail":"The first release produces a cited summary and timeline for the underwriter. Straight through decisions stay with existing, validated rules until the summary is proven."},{"title":"Build a gold standard set","detail":"Have senior underwriters and medical officers annotate a few hundred real files so recall of critical facts (for example a cancer history or abnormal lab value) can be measured, not guessed."},{"title":"Measure what is missed, not just what is right","detail":"A missed impairment is far worse than an extra one. Track recall on critical conditions separately and set release thresholds with the chief underwriter and medical officer."},{"title":"Run in parallel","detail":"For a period, underwriters review files the usual way and compare with the summary. Only reduce reading once differences are understood and documented."},{"title":"Govern it as a high risk system where applicable","detail":"In the EU, meet the AI Act requirements for high risk systems (risk management, data governance, logging and human oversight) and, as the deploying insurer, carry out the fundamental rights impact assessment that Article 27 requires. Elsewhere follow the insurer's AI governance framework and local rules, such as Colorado's Regulation 10-1-1 where external consumer data or predictive models are involved."}],"guardrails":["Every extracted fact links to the page and passage it came from","The summary never sets the rating or declines an applicant on its own","Health data processed only in approved regions, with access limited to underwriting roles","Explicit consent or legal basis recorded for every record processed","No inference of protected characteristics or genetic information beyond what law permits"],"humanInTheLoop":"Underwriters make every decision and review source pages for any material fact. Medical officers own the clinical vocabulary and the gold standard set, and the chief underwriter signs off release thresholds and reviews a monthly sample of summaries against full reads.","kpisToInstrument":["Median days from application to decision, before and after","Reading time per file from workbench logs","Recall of critical impairments on the audited sample","Share of summaries corrected by underwriters, by error type","Placement rate (offers accepted) for summarized versus non summarized cases"],"failureModes":[{"title":"Missed conditions in poor scans","detail":"Handwritten notes and faxed pages drop out of extraction and the summary looks complete. Flag low quality pages and require a human to read them."},{"title":"Automation bias in the underwriter","detail":"A clean summary is trusted and the source is never opened. Sample decisions and show the pages behind each material fact."},{"title":"Unfair outcomes through proxies","detail":"Summaries emphasise facts that correlate with protected characteristics. Test outcomes by group and keep the underwriting manual, not the model, in charge of rating."},{"title":"Consent and residency gaps","detail":"Records are sent to a model outside the approved region or without a valid basis. Route health data through controlled infrastructure only."}]},"risk":{"euAiAct":{"tier":"high","basis":"Annex III point 5(c): AI intended for risk assessment and pricing in relation to natural persons in life and health insurance. Article 6(3) exempts some purely preparatory tasks, but never a system that profiles natural persons. Extracting an applicant's health conditions and mapping them to the underwriting manual evaluates their health, which is profiling, so treat the system as high risk. Under the timeline as amended, the obligations for Annex III high risk systems apply from 2 December 2027, and Article 27 requires deployers of point 5(c) systems to assess the impact on fundamental rights before first use."},"regulations":["eu-ai-act","gdpr","hipaa","nist-ai-rmf","iso-42001","dora","solvency-ii"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(c) lists life and health insurance risk assessment and pricing of natural persons as high risk."},{"title":"Article 27, fundamental rights impact assessment for high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/27/","note":"Deployers of high risk systems referred to in Annex III points 5(b) and (c), which includes life and health insurance risk assessment and pricing, must assess the impact on fundamental rights before deploying the system."},{"title":"Implementation timeline of the EU AI Act","issuer":"Future of Life Institute (artificialintelligenceact.eu)","region":"europe","url":"https://artificialintelligenceact.eu/implementation-timeline/","note":"Lists 2 December 2027 as the date from which Chapter III, Sections 1 to 3, apply to high risk systems classified under Article 6(2) and Annex III."},{"title":"Opinion on Artificial Intelligence governance and risk management","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/publications/opinion-artificial-intelligence-governance-and-risk-management_en","note":"Supervisory expectations for AI in insurance, including data governance, fairness, record keeping and human oversight."},{"title":"SB21-169, Protecting Consumers from Unfair Discrimination in Insurance Practices","issuer":"Colorado Division of Insurance","region":"north-america","url":"https://doi.colorado.gov/for-consumers/sb21-169-protecting-consumers-from-unfair-discrimination-in-insurance-practices","note":"Amended Regulation 10-1-1, effective 15 October 2025, sets governance and risk management requirements for life insurers' and health benefit plan insurers' use of external consumer data and information sources, algorithms and predictive models. The Division has also held stakeholder meetings on life insurance underwriting."}],"controls":["Registration as a high risk system in the EU and a conformity assessment before use","Fundamental rights impact assessment under Article 27 of the AI Act before the insurer first uses the system, including when it is bought from a vendor","Documented data governance for training and test medical data, including consent","Logging of every summary, its sources and the underwriter's decision","Periodic fairness testing of decisions made with the summary","Data protection impact assessment for special category health data","In the US, a valid HIPAA authorization from the applicant for every record requested from a provider: life insurers are generally not HIPAA covered entities themselves, while health insurers acting as health plans are and must meet the Privacy and Security Rules directly"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** started from the new business system through the API.\nThe medical file is loaded through **document ingestion** (PDF, Office documents and images), an **agent with structured output**\nextracts clinical facts with page references into a fixed schema, and a **knowledge base** holding\nthe underwriting manual is searched with hybrid retrieval to suggest the relevant guidance. The\nresult is returned to the workbench for the underwriter; nothing is decided by the workflow.\n\nBecause this is health data and, in the EU, a high risk use, it runs in the **EU or UAE data\nresidency regions** with **PII masking** configured for identifiers that the model does not need,\n**tenant isolation** and **role based access**. Every run keeps a full **audit trail**,\nand **test suites** replay an annotated gold standard set on every prompt or model change so recall\non critical conditions is measured before release. The platform is model agnostic, so the insurer\ncan choose a model it has validated and switch without rebuilding."},"faq":[{"question":"Can AI decide life insurance applications from medical records?","answer":"For some applications, AI does decide. Manulife says its partnership with Munich Re Life US on alitheia, an AI driven risk assessment platform, raised instant underwriting decision eligibility from US$3 million to US$5 million. Summarization tools, such as Manulife's generative AI in Singapore, support an underwriter instead, and Prudential launched MedScreen+ to provide a faster, simpler and more transparent process for its underwriters. In the EU, AI used for risk assessment and pricing of individual life or health cover is high risk under the AI Act, so keep an underwriter accountable for decisions the summary informs."},{"question":"How much faster does underwriting get?","answer":"No insurer on this page publishes a figure for medical record summarization. Manulife says its generative AI in Singapore reduces processing time for policy applications but gives no number, and Prudential discloses no outcome for MedScreen+. Measure it yourself in a parallel run, and separate reading time from time spent waiting for records."},{"question":"What is the biggest risk?","answer":"Missing a material condition. Measure recall on critical impairments against a gold standard, keep source pages one click away and sample decisions, because a fluent summary is easy to trust too much."}],"related":["underwriting-risk-assessment-copilot","health-prior-authorization-and-claims-adjudication","intelligent-document-processing","insurance-pricing-and-actuarial-copilot"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from insurer annual reports and results filings verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added Solvency II to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed unsourced figures from the problem text, made the EU AI Act basis address Article 6(3), updated the Colorado Regulation 10-1-1 note, corrected the first FAQ answer and added an SEO title and meta description."},{"date":"2026-09-27","note":"Removed the Sun Life response time figure, which covers unspecified AI tools and straight through decisions rather than summarization, from the meta description, FAQ, value assumption and evidence; rewrote the first FAQ to acknowledge instant decisions; added the Article 27 fundamental rights impact assessment, the Annex III application date and a HIPAA note."}],"slug":"life-underwriting-medical-record-summarization","url":"https://www.blits.ai/ai-use-cases/life-underwriting-medical-record-summarization","benchmarks":[],"indicativeValueResult":{"low":300000,"high":1600000},"evidence":["manulife-generative-ai-life-underwriting","prudential-plc-medscreen-ai-underwriting"]},{"title":"AI support for emergency call triage (112 and 911)","shortTitle":"Emergency call triage support","seoTitle":"AI for 911 and 112 emergency call triage","metaDescription":"AI helps 911 and 112 call takers with transcripts, translation and alerts. In a Copenhagen trial, alerts did not significantly raise cardiac arrest recognition.","definition":"AI that supports emergency call takers and dispatchers during 112 and 911 calls, with live transcription, translation, summaries, location cues and alerts for critical conditions such as cardiac arrest, while the call taker keeps every triage and dispatch decision.","aliases":["911 call taking AI","112 emergency call AI","dispatcher assist","AI for public safety answering points"],"industries":["government","healthcare"],"functions":["citizen-services","operations"],"patterns":["speech-analytics","translation","summarization","prediction-and-scoring"],"channels":["voice","agent-desktop"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","problem":"Emergency communications centres work under constant volume: Baltimore answers about 1.4 million\n911 calls a year, and at Galt Police Department a single dispatcher is sometimes on duty alone. Call\ntakers must understand panicked, noisy or non native callers, get an address, follow a protocol\nand type everything at once, often while also handling radio. Critical conditions are missed:\nCopenhagen EMS researchers note that dispatchers fail to identify roughly a quarter of out of\nhospital cardiac arrests. Callers who do not speak the local language wait for a telephone\ninterpreter. Quality assurance often covers only a sample of calls (roughly 30% in Baltimore\nbefore automation), with feedback arriving long after the call.\n\nThe stakes are high. A wrong alert, a mistranslation or a missed cue can cost a life, which is\nwhy the EU AI Act lists emergency call classification and dispatch as high risk.","problemStats":[{"statement":"Copenhagen Emergency Medical Services researchers report that emergency medical dispatchers fail to identify approximately 25% of out of hospital cardiac arrests.","sourceTitle":"Effect of Machine Learning on Dispatcher Recognition of Out-of-Hospital Cardiac Arrest During Calls to Emergency Medical Services: A Randomized Clinical Trial","sourceUrl":"https://pmc.ncbi.nlm.nih.gov/articles/PMC7788469/","year":2021}],"howItWorks":"1. **Transcribe live.** Speech recognition produces a running transcript of the call on the call\n   taker's screen, with key details (address, weapons, symptoms) highlighted.\n2. **Translate.** For callers in another language, the system detects the language and shows a\n   translated transcript, or voices the call taker's questions in the caller's language.\n3. **Flag critical conditions.** A model listens for patterns of time critical conditions, such as\n   cardiac arrest, and alerts the call taker to consider the matching protocol.\n4. **Summarise and document.** At the end of the call, the system drafts a summary for the incident\n   record, which the call taker edits.\n5. **Review quality.** Automated QA checks every call against protocol and flags calls for\n   supervisor review and coaching.\n6. **Keep humans in charge.** Call takers decide the triage category, protocol and dispatch; the\n   AI never dispatches or downgrades a call.","valueDrivers":["risk-reduction","speed","inclusion-and-access","employee-productivity"],"kpis":["accuracy","detection-rate-improvement","handling-time-reduction","time-saved-per-task"],"indicativeValue":{"referenceOrg":"An emergency communications centre handling 1 million calls a year","inputs":[{"key":"calls","label":"Emergency calls per year","low":1000000,"high":1000000,"unit":"calls per year","note":"The reference centre; Baltimore answers about 1.4 million a year."},{"key":"minutesSaved","label":"Call taker minutes saved per call on documentation","low":0.5,"high":1.5,"unit":"minutes per call","note":"Editorial assumption for summaries replacing manual narrative entry. No public benchmark states this yet."},{"key":"hourlyCost","label":"Fully loaded call taker cost per hour","low":40,"high":60,"unit":"USD per hour","note":"Editorial assumption. Replace with your own cost."}],"formula":"calls * minutesSaved / 60 * hourlyCost","currency":"USD","period":"per year","resultLabel":"Call taker time released, valued at cost","caveat":"Values documentation time only. It leaves out the value of faster recognition of critical conditions, interpreter costs avoided, full QA coverage and the cost of the system; released time in understaffed centres usually goes to answering calls faster, not to savings."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Integration with call handling and CAD systems, real time latency, noisy audio and many languages, plus the need for clinical and operational validation of every alert, make this a demanding public sector deployment.","dataPrerequisites":["Recorded calls with outcomes for testing (for example confirmed cardiac arrest)","Current protocols (medical, fire, police) and local address data","Language mix of callers, to prioritise translation"],"integrations":["Call handling system and audio stream","Computer aided dispatch (CAD) for incident records","Location services and maps","QA and training systems"]},"implementation":{"steps":[{"title":"Start with transcription and summaries","detail":"Live transcripts and draft summaries support every call and change no decision; staff in Baltimore and Galt describe both as practical help with addresses and documentation."},{"title":"Add translation with a fallback","detail":"Offer machine translation for the most common languages, with a clear way to bring in a human interpreter when the call is complex or the translation is doubtful."},{"title":"Treat alerts as clinical interventions","detail":"Before switching on alerts for conditions such as cardiac arrest, run a controlled trial on your own calls. Copenhagen's randomized trial showed a model that beat dispatchers on sensitivity did not significantly improve dispatcher recognition."},{"title":"Design the alert for the call taker","detail":"A low positive predictive value floods call takers with false alarms. Tune thresholds with dispatchers and measure whether alerts are acted on."},{"title":"Use automated QA for coaching","detail":"Review every call against protocol and feed findings into training. Baltimore moved from a third party reviewing roughly 30% of calls to automated QA on all of them."}],"guardrails":["The AI never dispatches, downgrades or closes a call; call takers decide","Alerts are advisory, logged and evaluated against confirmed outcomes","Human interpreter always available as a fallback to machine translation","Transcripts and summaries edited and approved by the call taker before they enter the record","Fail safe design, so an outage of the AI never blocks call handling"],"humanInTheLoop":"Call takers and dispatchers make every triage and dispatch decision. Medical directors approve any clinical alert and its thresholds; supervisors review AI assisted QA findings before they reach staff files; every alert and translation is auditable against the call recording.","kpisToInstrument":["Recognition rate of target conditions with and without alerts, on confirmed outcomes","Alert positive predictive value and the share of alerts acted on","Time to address confirmation and to dispatch","Transcription and translation accuracy on a sampled set of calls","QA coverage and protocol compliance scores"],"failureModes":[{"title":"Better model, same outcome","detail":"In Copenhagen the model had higher sensitivity for cardiac arrest than dispatchers (85.0% against 77.5%), but alerting dispatchers did not significantly improve their recognition. Measure the human outcome, not the model."},{"title":"Alert fatigue","detail":"Alerts with low positive predictive value (17.8% in the Copenhagen trial, against 55.8% for dispatchers) are easy to discount. Tune thresholds and track actions on alerts."},{"title":"Mistranslation under pressure","detail":"A wrong word in a translated address or symptom can send help to the wrong place. Show the original and translation, confirm critical details and keep interpreters available."},{"title":"Dependence during outages","detail":"Staff who rely on AI transcripts can lose practice in manual call taking. Keep manual procedures practised and the system fail safe."}]},"risk":{"euAiAct":{"tier":"high","basis":"Annex III point 5(d): AI systems intended to evaluate and classify emergency calls or to dispatch or set priority for emergency first response services (police, fire, medical aid) are high risk. Pure transcription that performs a narrow procedural or preparatory task may fall outside it under the Article 6(3) exceptions, but alerts that influence triage are in scope. An AI agent that speaks with callers directly, for example on a non emergency line, must also tell them they are interacting with AI (Article 50)."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(d) lists emergency call classification and dispatch as high risk."},{"title":"Article 27, fundamental rights impact assessment for high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/27/","note":"Deployers that are bodies governed by public law, or private entities providing public services, must assess the impact on fundamental rights before first using a high risk system such as one covered by point 5(d)."},{"title":"NIST AI Risk Management Framework","issuer":"NIST","region":"north-america","url":"https://www.nist.gov/itl/ai-risk-management-framework","note":"A framework for mapping, measuring and managing risk that US public safety agencies can apply to call taking AI."}],"controls":["Controlled evaluation or trial on local calls before any alert goes live","Conformity assessment, risk management and logging as required for high risk AI in the EU","Continuous monitoring of alert performance against confirmed outcomes","Documented fallback procedures and regular drills without the AI","Clear records of which AI outputs the call taker saw on each call"],"incidents":[]},"blitsAi":{"howToBuild":"Blits.ai is not a computer aided dispatch system and should not sit in the 911 or 112 decision\npath. Where it fits is around it: its **self hosted transcription and speaker diarization**\n(WhisperX on GPU, processed on Blits.ai infrastructure) can transcribe recorded calls for\nquality review and training, and an **agent** grounded in the centre's protocols through the\n**knowledge base** can draft protocol checks that supervisors review.\n\nOn the **non emergency line** next to the emergency number, a **voice agent** on telephony with\nreal time streaming speech recognition can answer routine calls and **redirect the call** to the\ndispatcher when emergency language is detected, with **test suites** to check that\nbehaviour before go live. **Language detection and translation**, **PII masking**\nand **EU and UAE data residency** support multilingual and sovereign deployments; anything that\nclassifies emergency calls should be treated as high risk and validated with the medical director."},"faq":[{"question":"Does AI improve emergency call triage?","answer":"The best public evidence is mixed. In Copenhagen's randomized trial on 112 calls, a model flagged cardiac arrest with higher sensitivity than dispatchers (85.0% against 77.5%), but dispatchers who received alerts did not recognize significantly more cases. For transcription, translation, summaries and automated QA, the evidence on this page comes from vendor case studies from US centres such as Baltimore, which report practical benefits but no controlled comparison."},{"question":"Is emergency call AI high risk under the EU AI Act?","answer":"Yes, when it evaluates or classifies emergency calls or sets dispatch priority (Annex III point 5(d)). That brings risk management, logging, human oversight and conformity requirements, and public bodies must also carry out a fundamental rights impact assessment (Article 27)."},{"question":"Where do centres start?","answer":"With live transcripts, summaries, translation and automated QA, as Baltimore did, and with AI on the non emergency line, as Galt Police Department did to keep routine calls away from dispatchers."}],"related":["non-emergency-service-request-routing","public-service-translation","live-agent-assist","citizen-information-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from a randomized trial abstract and emergency centre case studies, with every quote verified against the source."},{"date":"2026-09-25","note":"Consolidation pass: added European Electronic Communications Code, UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription; tied the problem statement to Baltimore and Galt figures; softened unsupported claims (hardest deployment, alerts ignored, outage dependence, clearer benefits); stated that Copenhagen alerts did not significantly improve recognition; removed HIPAA, which applies only where the centre is part of a covered entity; added the Article 6(3) and Article 50 notes and corrected the Article 27 scope; limited the Blits.ai section to inventory capabilities; set the Copenhagen trial year to its 2018 start; the problem statistic now links the open access full text on PubMed Central, because the journal page blocks automated access."},{"date":"2026-09-27","note":"Review fixes: removed \"California\" (no cited source names the state); removed the containment rate KPI, because the AI does not handle emergency calls and the only containment figure comes from a non emergency line; removed the UK Algorithmic Transparency Recording Standard (mandatory only for UK central government, not for police, ambulance or fire services) and the European Electronic Communications Code (telecom operator rules, no obligation on triage AI); set adoption to early adopters, since Baltimore, Galt and Delaware County run it in production; Baltimore's unmeasured 98% accuracy impression is no longer a metric; Copenhagen figures now follow the full text (169,049 of 226,130 calls processed, 5,847 flagged, 5,242 randomized)."}],"slug":"emergency-call-triage-support","url":"https://www.blits.ai/ai-use-cases/emergency-call-triage-support","benchmarks":[],"indicativeValueResult":{"low":333333.3333333334,"high":1500000},"evidence":["baltimore-911-assistive-call-taking","copenhagen-ems-cardiac-arrest-recognition-trial","galt-police-department-non-emergency-call-triage"]},{"title":"AI support for property valuation and appraisal","shortTitle":"Property valuation support","seoTitle":"AI property valuation and appraisal support","metaDescription":"Fannie Mae estimates appraisal alternatives saved US mortgage borrowers over $2.5 billion since 2020; C3 AI reports a 40% accuracy gain at Riverside County.","definition":"AI, most often an automated valuation model, that estimates a property's market value from comparable sales, property characteristics and location data, and either offers to replace a full appraisal within set limits or gives a professional valuer a first pass estimate, the closest comparable sales and a reliability score, so the valuer's time goes to the properties that need a person's judgment.","aliases":["automated valuation model","AVM","AI property appraisal","mass appraisal AI"],"industries":["real-estate","banking","government"],"functions":["lending-and-credit","case-management","operations"],"patterns":["prediction-and-scoring","classification-and-routing"],"channels":["internal-tools","api"],"audience":"employee-facing","autonomy":"supervised-agent","adoptionStage":"mainstream","segment":"lending","problem":"Every mortgage, remortgage and property tax reassessment needs a value, and a full, on site\nappraisal takes a professional's time to inspect the property, research comparable sales and write\na report for each individual property. The Royal Institution of Chartered Surveyors, a global\nprofessional body for the sector, notes that residential valuation data is often publicly\navailable while commercial property data is often less widely available, which makes automated\nmodels less reliable outside standard, homogeneous housing.\n\nAutomated valuation models are not new, and not all of them use machine learning; RICS points out\nthat some still apply fixed, rule based formulas. What has changed is how many of them now learn\nfrom large, continuously updated datasets of sales, tax records and property characteristics, and\nhow far organizations are willing to let a model's output stand in for a person's inspection.","problemStats":[],"howItWorks":"1. **Collect the comparables.** The model pulls comparable sales, tax assessment and land registry\n   records, property characteristics and location data covering the area.\n2. **Predict and compare.** A regression or machine learning model estimates the property's value\n   and identifies the closest comparable properties that have sold.\n3. **Score the reliability.** The system scores how confident the estimate is, typically from how\n   closely it tracks the comparable sales it used.\n4. **Route by confidence.** High confidence, low risk cases are auto accepted, for example as an\n   appraisal waiver or a direct enrollment at the sale price; low confidence or high value cases\n   are queued for a person.\n5. **A professional reviews the queue.** Valuers or appraisers spend their time on the batches the\n   model flagged as uncertain or close to a decision boundary, not on every case.","valueDrivers":["cost-to-serve","speed","employee-productivity","customer-experience"],"kpis":["cost-reduction","customer-savings","accuracy","automation-rate"],"indicativeValue":{"referenceOrg":"A residential mortgage lender originating 50,000 loans a year","inputs":[{"key":"loans","label":"Loans originated per year","low":20000,"high":100000,"unit":"loans per year","note":"Editorial assumption, replace with your own origination volume."},{"key":"avmEligibleShare","label":"Share of loans eligible for an automated valuation instead of a full appraisal","low":0.2,"high":0.5,"unit":"fraction of loans","note":"Editorial assumption, replace with your own eligibility policy and loan mix."},{"key":"appraisalFeeAvoided","label":"Appraisal fee avoided per automated valuation","low":400,"high":700,"unit":"USD per loan","note":"Editorial assumption for a typical US conventional appraisal fee, replace with your own."}],"formula":"loans * avmEligibleShare * appraisalFeeAvoided","currency":"USD","period":"per year","resultLabel":"Appraisal fee cost avoided","caveat":"Only the appraisal fee the borrower avoids paying, for the lender's loans that qualify for an automated valuation instead of a full appraisal. It leaves out the cost of building and validating the model, the appraisals still needed for higher risk or higher value loans, and any difference in collateral risk between an automated and a fully manual valuation, which none of the deployments on this page report as a single figure."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Building or buying a model a regulator, lender or model risk team will accept needs deep, clean sales and property data and a documented method for testing accuracy and fairness across property types and areas; the Valuation Office Agency aligned its in house testing to International Association of Assessing Officers AVM standards and had the International Association of Assessing Officers review its model's development process before relying on it.","dataPrerequisites":["Comparable sales, tax assessment and property characteristic data covering the geography","A documented method to test accuracy and fairness across property types, values and areas","A defined confidence threshold that decides which properties get an automated value and which go to a person"],"integrations":["Automated underwriting system or tax assessment system that consumes the estimate","Geospatial, land registry or multiple listing service data feeds","A case queue for the properties the model routes to manual valuation"]},"implementation":{"steps":[{"title":"Decide where an automated value can stand in for a person","detail":"Set the loan to value, price band or property type limits within which an automated value is acceptable, matching what a regulator or an internal model risk function expects."},{"title":"Build the confidence score, not just the estimate","detail":"Score every estimate's reliability from its distance to comparable sales, so low confidence cases route to a person automatically rather than being accepted at face value."},{"title":"Test for accuracy and fairness before launch","detail":"Run a ratio study across property types, price bands and geographies, not just an overall error rate, against a recognized mass appraisal standard."},{"title":"Keep a professional in the loop for the exceptions","detail":"Route low confidence and high value properties to a valuer or appraiser, and feed that person's corrections back into ongoing monitoring of the model."},{"title":"Publish the method","detail":"A public register entry or a documented internal policy explaining what the model does and why builds the trust an automated valuation program needs from regulators and customers."}],"guardrails":["A confidence score below a set threshold always routes to a human valuer, never an automated value alone","Independent testing against a recognized mass appraisal standard before launch and after every material model change","Loan to value, price band or property type limits on when an automated value replaces a full appraisal"],"humanInTheLoop":"A qualified valuer or appraiser reviews every property the model flags as low confidence or above a value threshold, and a central analytics or model risk team monitors overall accuracy and fairness against completed sales and manual valuations.","kpisToInstrument":["Automated valuations issued versus properties routed to a person, split by price band and area","Accuracy and dispersion of automated valuations against confirmed sale prices or completed manual valuations","Appeals or challenges to automated valuations as a share of all automated valuations issued"],"failureModes":[{"title":"Confident but wrong in a fast moving market","detail":"A model trained on past sales lags a market that is moving quickly and prices systematically low or high; monitor the gap between automated valuations and completed sales every month, not only at launch."},{"title":"Uneven accuracy across areas and property types","detail":"Data rich urban areas get a more accurate value than rural or unusual properties, so the model quietly serves some customers worse than others; a ratio study by property type and area, not an overall figure alone, is what catches this."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"An automated valuation model values the collateral, not the person, so it is not itself listed in Annex III; the EU Mortgage Credit Directive treats property valuation (Article 19) and the creditworthiness assessment of the borrower (Article 18) as separate steps, and Article 18(3) says the creditworthiness assessment must not be based predominantly on the value of the property exceeding the amount of credit, or on an assumption that the property's value will increase. The valuation becomes relevant to Annex III point 5(b), creditworthiness assessment of natural persons, only where its output is built into a separate system that evaluates the borrower's creditworthiness, and whether that happens depends on how the lender designs the credit decision, not on the valuation model itself."},"regulations":["eu-ai-act","gdpr","eu-mortgage-credit-directive","us-ecoa-reg-b","uk-gdpr","uk-atrs"],"guidance":[{"title":"Responsible use of AI case study: valuation","issuer":"Royal Institution of Chartered Surveyors","region":"europe","url":"https://www.rics.org/profession-standards/rics-standards-and-guidance/conduct-competence/responsible-use-of-ai/ruai-case-studies-02","note":"Explains that automated valuation models should support, not replace, a qualified valuer's judgment for high risk valuations such as those informing lending, legal disputes or investment decisions."},{"title":"Quality Control Standards for Automated Valuation Models","issuer":"CFPB, OCC, Federal Reserve, FDIC, NCUA and FHFA","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/final-rules/quality-control-standards-for-automated-valuation-models/","note":"The 2024 US interagency final rule requiring mortgage originators and secondary market issuers to adopt policies and controls so that automated valuation models used to value a consumer's principal dwelling maintain a high level of confidence in the estimates, protect data integrity, avoid conflicts of interest, are tested by random sample review and comply with nondiscrimination law."}],"controls":["A qualified valuer signs off on every high value, high risk or low confidence automated valuation before it is relied on","An independently reviewed accuracy and fairness study by property type and area, refreshed on a set cycle"],"incidents":[]},"blitsAi":{"howToBuild":"The valuation model itself, the regression, comparable matching and confidence scoring, runs on\nthe organization's own data science stack or a specialist valuation vendor; Blits.ai is not where\nyou build a mass appraisal model. What fits on Blits.ai is the layer that puts the estimate to\nwork: an SQL knowledge base over the valuation, sales and property data lets an agent answer a\nvaluer's question about why a property scored the way it did and which comparable sales it used,\nin plain language rather than a raw output table.\n\nAgentic workflows with human in the loop confirmation route every low confidence or above\nthreshold estimate to a valuer with approve and reject controls and the comparable evidence\nattached, instead of a queue nobody checks, and every override a valuer makes is logged for the\naccuracy study. Analytics track the\nsplit between automated and manually reviewed valuations over time, and the platform's model\nagnostic routing and EU and UAE data residency options fit a regulated valuation function without\nmoving data outside a required region."},"faq":[{"question":"Can AI fully replace a professional valuer?","answer":"It depends on the deployment. Riverside County routes properties outside its configured AVM variance thresholds to a person, and the Valuation Office Agency routes the batches its model scores as least reliable or closest to a Council Tax band boundary to a valuer; RICS is explicit that automated valuation models should support, not replace, the valuation process for lending, legal or investment decisions. Fannie Mae's Value Acceptance goes further and replaces the appraisal entirely for eligible loans, within set loan to value and loan type limits, rather than routing anything to a human valuer."},{"question":"How much do lenders and governments save with automated valuation?","answer":"Fannie Mae estimates that appraisal alternatives such as Value Acceptance saved US mortgage borrowers more than $2.5 billion since early 2020, though that figure also covers Value Acceptance + Property Data, which uses a third party data collector rather than a fully automated valuation. The UK Valuation Office Agency's own early estimate for its Wales Council Tax revaluation model is a reduction in the cost of a revaluation by one third compared with a fully manual valuation, an estimate for a revaluation planned for 2028 that has not happened yet."},{"question":"How accurate does an automated valuation need to be?","answer":"There is no single number. C3 AI reports a 40% improvement in model accuracy at Riverside County, over the county's previous linear regression models, demonstrating the ability to directly enroll up to 97% of all property sales. Independent testing against a recognized mass appraisal standard, not a vendor's own figure alone, is what a buyer should ask for."},{"question":"Is an automated valuation high risk under the EU AI Act?","answer":"An automated valuation model is not itself listed in Annex III; it values the collateral, not the borrower. It becomes relevant to Annex III point 5(b), creditworthiness assessment of natural persons, only where its output is built into a separate system that assesses the borrower, which depends on how the lender designs that system, not on the valuation model alone."}],"related":[],"datePublished":"2026-09-28","dateModified":"2026-09-28","lastVerified":"2026-09-28","changelog":[{"date":"2026-09-28","note":"Published after review by the editor after the Opus targeted check's remaining text fixes."},{"date":"2026-09-28","note":"First version, researched for the real estate and energy scope, with the UK Valuation Office Agency's own transparency record, Riverside County's C3 AI case study and Fannie Mae's own reporting checked against the primary sources."},{"date":"2026-09-28","note":"Editorial pass: corrected the Riverside County accuracy metric, year and stage; removed the unsupported five year commitment claim; dropped the VOA cost reduction forecast as a metric and fixed its year; rewrote the Fannie Mae summary to match its source; reattributed the 40% and 97% figures to C3 AI in the FAQ and meta description; narrowed the EU AI Act basis and dropped the FCRA reference in favor of the 2024 US interagency AVM rule; removed unsourced superlatives from the problem statement; and fixed the indicativeValue caveat, complexityNote wording and spelling consistency."},{"date":"2026-09-28","note":"Adversarial review pass: reattributed the $2.5 billion figure in the meta description to appraisal alternatives rather than Value Acceptance alone; dropped the Riverside County automation rate metric, which the C3 AI page frames as a demonstrated capability of an initial deployment rather than an achieved production result, and clarified in the evidence note that the 80% accuracy figure is a model selection and validation result; reworded the Fannie Mae loan to value ratio change as an announcement rather than an accomplished fact; corrected the RICS quote and its description as a global professional body; corrected the VOA summary and note to say the model covers the vast majority of Wales's 1.5 million properties rather than around 1.5 million; dropped \"about\" before \"one third\" to match the VOA's exact wording; split the FAQ's routing answer so the decision boundary language is attributed only to the VOA and Riverside's own AVM variance threshold language is used for Riverside; refined the EU Mortgage Credit Directive Article 18(3) paraphrase; added the missing \"high level of confidence in the estimates\" factor to the 2024 US interagency AVM rule guidance note; added UK GDPR and the UK Algorithmic Transparency Recording Standard to risk.regulations; toned down blitsAi.howToBuild's human in the loop language to match the platform feature inventory; and changed segment from a one off value to the shared \"lending\" grouping."}],"slug":"property-valuation-support","url":"https://www.blits.ai/ai-use-cases/property-valuation-support","benchmarks":[{"kpi":"customer-savings","label":"Customer savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2500000000,"min":2500000000,"max":2500000000,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"fannie-mae-value-acceptance-appraisal-alternatives","pooled":true}]}],"indicativeValueResult":{"low":1600000,"high":35000000},"evidence":["fannie-mae-value-acceptance-appraisal-alternatives","riverside-county-c3-ai-property-appraisal","uk-voa-wales-automated-valuation-model"]},{"title":"AI system and model inventory with shadow AI discovery","shortTitle":"AI model inventory","seoTitle":"AI model inventory and shadow AI discovery","metaDescription":"An AI inventory records every AI system with its owner, risk tier and approval. See how the DOJ, the Federal Reserve and Unilever build and run theirs.","definition":"A governed register of every AI system and model an organization builds, buys or uses, with its owner, purpose, data, risk tier and approval status, kept current by AI that discovers unregistered use, reads the documentation and assembles the evidence a board, auditor or supervisor asks for.","aliases":["AI inventory","AI register","AI use case inventory","model inventory","shadow AI discovery"],"industries":["cross-industry","banking","insurance","government","manufacturing"],"functions":["risk-management","regulatory-compliance","it-and-engineering"],"patterns":["agentic-workflow","document-processing","rag-knowledge-assistant","classification-and-routing"],"channels":["internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"AI governance frameworks start from the same question: which AI systems do you run? The NIST AI\nRisk Management Framework asks for mechanisms to inventory AI systems, the Monetary Authority of\nSingapore's proposed AI risk management guidelines expect financial institutions to keep accurate\nand up to date AI inventories, US federal agencies (with limited exceptions) must inventory their\nAI use cases every year,\nand the EU AI Act's deployer and registration duties presume an organization knows which high risk\nsystems it uses. Answering the question per system, with an owner, a purpose, a risk tier and\nevidence of approval, is harder than it looks.\n\nA register kept as a spreadsheet that project teams fill in once, at approval, misses the AI that\narrives inside software someone bought, the assistant a team switched on in a SaaS tool, and the\nprompts employees paste into public chatbots. Use also grows fast: the US Department of Justice\nreports that its 2025 inventory holds 315 entries, 30.7% more than the year before. Gaps\nhave consequences. In May 2026 CB Financial Services reported a material cybersecurity incident to\nthe SEC after non public customer information at its subsidiary Community Bank was handled with an\nunauthorized AI based application, the kind of unregistered use that discovery aims to surface early.","problemStats":[{"statement":"In its 2025 workshops with 13 banks, ECB Banking Supervision observed that all banks preparing for the AI Act had built up AI systems inventories and a set process to put AI models into production, while data governance adapted to AI was emerging only in a small number of cases.","sourceTitle":"AI workshops with banks 2025, annex","sourceUrl":"https://www.bankingsupervision.europa.eu/ecb/pub/pdf/annex/ssm.nl251120_1_annex.en.pdf","year":2025}],"howItWorks":"1. **Define the record.** One schema for every AI system: owner, business purpose, users,\n   vendor or in house, model and version, data categories (including personal data), autonomy\n   level, risk tier under internal policy and the EU AI Act, approval status, review date and\n   links to documentation.\n2. **Discover what is actually running.** An agent reconciles the register against evidence\n   sources: procurement and contract records, SaaS and API usage logs, cloud and model platform\n   accounts, code repositories and network egress to AI services. Anything that looks like AI and\n   has no entry becomes a candidate record for an owner to confirm or retire.\n3. **Draft the entry from the documents.** Document AI reads model cards, vendor documentation,\n   data protection impact assessments and approval minutes and pre fills the record, citing the\n   page each field came from. The owner confirms or corrects every field.\n4. **Classify and route.** The draft risk tier and the triggers for deeper review (personal data,\n   decisions about people, customer facing use, material service provider) are proposed by rules\n   plus a model, and a governance officer decides.\n5. **Keep it current.** Scheduled checks flag records past their review date, models whose\n   version changed, vendors whose terms changed and systems whose usage jumped.\n6. **Answer and assemble.** Staff and auditors ask questions in plain language (\"which customer\n   facing systems use personal data and a third party model?\") and get answers with links to the\n   records, and the agent assembles the evidence pack for a named system on request.","valueDrivers":["compliance","risk-reduction","employee-productivity","speed"],"kpis":["hours-saved","time-saved-per-task","productivity-gain","accuracy"],"indicativeValue":{"referenceOrg":"A bank or insurer with 150 AI systems and models in scope of its AI policy","inputs":[{"key":"systems","label":"AI systems and models in the register","low":100,"high":200,"unit":"systems","note":"Editorial assumption for a mid sized regulated firm, counting vendor AI and generative AI tools. Replace with your own count."},{"key":"hoursPerSystem","label":"Manual effort per system per year to discover, document, attest and review","low":8,"high":20,"unit":"hours per system per year","note":"Editorial assumption covering the owner, the second line reviewer and the inventory administrator. Replace with your own time study."},{"key":"automationShare","label":"Share of that effort the discovery and drafting removes","low":0.25,"high":0.45,"unit":"fraction of effort","note":"Editorial assumption; no public benchmark exists yet. Owners still confirm every field."},{"key":"requests","label":"Evidence requests per year from supervisors, auditors and the board","low":6,"high":15,"unit":"requests per year","note":"Editorial assumption. Replace with your own count of inventory related requests."},{"key":"hoursPerRequest","label":"Hours to assemble one evidence pack by hand","low":20,"high":60,"unit":"hours per request","note":"Editorial assumption."},{"key":"requestReduction","label":"Share of evidence assembly time saved","low":0.3,"high":0.5,"unit":"fraction of time","note":"Editorial assumption; assembly still needs a human review before anything leaves the firm."},{"key":"hourlyCost","label":"Fully loaded cost of a risk or technology specialist hour","low":60,"high":110,"unit":"USD per hour","note":"Editorial assumption. Replace with your own rate."}],"formula":"(systems * hoursPerSystem * automationShare + requests * hoursPerRequest * requestReduction) * hourlyCost","currency":"USD","period":"per year","resultLabel":"Specialist time released from inventory upkeep and evidence assembly","caveat":"Time released only. It leaves out the value that matters most and is hardest to price: the incidents, fines and failed audits avoided because shadow AI is found early, and the faster approval of new AI use cases once the register is trusted. It also leaves out the cost of the discovery integrations."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Keeping a register is easy; keeping it true is the work. Discovery needs read access to procurement, SaaS, cloud and network data, which crosses several owners, and the risk classification needs a written policy before any AI can apply it.","dataPrerequisites":["A written AI policy with a definition of what counts as AI and a risk tiering method","The current register, however incomplete, and the list of approved AI tools and vendors","Procurement and contract records that show which suppliers provide AI features","Usage logs from SaaS administration, cloud accounts, API gateways and network egress","Model cards, vendor documentation, impact assessments and approval records"],"integrations":["Governance, risk and compliance platform or the existing model inventory","Procurement and contract management system","SaaS management, identity provider and cloud accounts for usage signals","Code repositories and model registries","Document stores holding model documentation and approvals","Ticketing or workflow tool for owner attestations"]},"implementation":{"steps":[{"title":"Write the definition and the schema first","detail":"Agree what counts as an AI system (include vendor features and general purpose assistants), the mandatory fields and the risk tiers. Map the tiers to the EU AI Act categories and to your sector rules so one record answers every framework."},{"title":"Seed from what you already know","detail":"Load the existing register, the model risk inventory, the approved tool list and the procurement records. Deduplicate before adding anything new, as the US Department of Justice did when it combined similar, widely adopted AI use cases into single department wide entries."},{"title":"Add discovery sources one at a time","detail":"Start with the highest yield signals (SaaS admin consoles, API keys to model providers, network egress to AI domains), and route each unregistered hit to a named owner as a candidate record, not as an accusation."},{"title":"Let AI draft, owners confirm","detail":"Pre fill records from documents with a citation per field, then ask the owner to confirm. Measure how often owners correct the draft and fix the extraction where corrections cluster."},{"title":"Put the register in the approval path","detail":"No production release, contract signature or tool enablement without a record. This is what keeps the register current after the first clean up."},{"title":"Rehearse the evidence request","detail":"Pick one high risk system and ask for the full evidence pack as a supervisor would, time it, and fix the gaps before the real request arrives."}],"guardrails":["The AI proposes records, tiers and links; a named human owner confirms every record and a governance officer approves every risk tier","Every field drafted from a document cites the source document and page","Discovery reads metadata and usage signals, not the content of employee prompts, unless policy and law allow it","The register itself is access controlled and logged, because it maps the organization's most sensitive systems","The inventory agent is itself an entry in the register, with its own owner and review date"],"humanInTheLoop":"Owners attest to their records, the second line approves risk tiers and exceptions, and the AI governance committee decides on retiring or blocking unregistered systems. The AI never changes an approval status or blocks a tool on its own.","kpisToInstrument":["Coverage, the share of discovered AI systems that have a confirmed record","Number of unregistered systems found per month and time from discovery to confirmed record or retirement","Share of records past their review date","Owner correction rate on AI drafted fields","Hours to assemble an evidence pack for one system"],"failureModes":[{"title":"The register is complete on paper only","detail":"Teams fill it in once at approval and never again. Tie the record to release, contract renewal and tool enablement, and flag records whose model version or usage changed."},{"title":"Discovery as surveillance","detail":"Reading employee prompts to find shadow AI creates a privacy and trust problem. Use metadata and usage signals first, involve the works council or employee representatives where required, and offer an approved alternative for every tool you block."},{"title":"Confident but wrong classification","detail":"A model tier that looks authoritative gets copied into reports without review. Keep the tier as a proposal until a named person approves it and record who did."},{"title":"Definition too narrow","detail":"Only in house machine learning models get registered, while vendor features and generative AI assistants carry most of the new risk. Include anything that infers outputs from input, as the EU AI Act definition does."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Minimal for a system level register of systems and owners with no monitoring of individual employees; it is not listed in Annex III and is the instrument deployers use to meet obligations such as the Article 26 duties for high risk systems and the Article 49 registration of Annex III systems in the EU database. Limited where the plain language assistant that staff and auditors query is not obviously an AI system to its users: under Article 50(1) its provider must then design it so people are told they are dealing with AI. Possibly high risk under Annex III point 4(b) on worker management if the discovery process monitors or evaluates the behavior of individual employees rather than staying at the level of systems and owners."},"regulations":["eu-ai-act","iso-42001","nist-ai-rmf","mas-ai-risk-management","dora","gdpr","apra-cps-230","us-sr-11-7","pra-ss1-23"],"guidance":[{"title":"Article 49: Registration","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/49/","note":"Providers register high risk systems listed in Annex III in the EU database before placing them on the market or putting them into service (critical infrastructure systems under point 2 are registered nationally), and deployers that are public authorities register their use; an internal register is the practical source for that data."},{"title":"Article 26: Obligations of deployers of high risk AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/26/","note":"Deployers must monitor the operation of high risk systems, keep logs and assign human oversight, which presumes they know which systems they deploy."},{"title":"MAS Guidelines for Artificial Intelligence (AI) Risk Management, consultation paper","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/news/media-releases/2025/mas-guidelines-for-artificial-intelligence-risk-management","note":"Financial institutions are expected to identify AI usage across the firm, maintain accurate and up to date AI inventories and assess risk materiality."},{"title":"AI RMF Core, GOVERN 1.6","issuer":"NIST","region":"north-america","url":"https://airc.nist.gov/airmf-resources/airmf/5-sec-core/","note":"Mechanisms are in place to inventory AI systems, resourced according to organizational risk priorities."},{"title":"M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust","issuer":"US Office of Management and Budget","region":"north-america","url":"https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf","note":"Federal agencies must inventory AI use cases at least annually, submit them to OMB and publish a public version; a useful template for the fields a register needs."},{"title":"AI workshops with banks 2025, annex","issuer":"ECB Banking Supervision","region":"europe","url":"https://www.bankingsupervision.europa.eu/ecb/pub/pdf/annex/ssm.nl251120_1_annex.en.pdf","note":"Supervisory observations from 13 bank workshops, including AI systems inventories, AI Act self assessments and, as an emerging practice, automated tools that monitor the inventory and workflow."}],"controls":["A single register with an accountable owner per record and a documented definition of AI","Record required before production release, contract signature or tool enablement","Periodic attestation by owners and validation of the register by the second line","Audit trail of every change to a record, tier or approval status","Approved alternatives for common generative AI tasks, so blocking shadow AI does not stop the work"],"incidents":[{"title":"CB Financial Services 8-K, Item 1.05: customer data handled in an unauthorized AI application","url":"https://www.sec.gov/Archives/edgar/data/1605301/000160530126000021/cbfv-20260507.htm","note":"CB Financial Services disclosed in May 2026 that its subsidiary Community Bank had found non public customer information, including names, social security numbers and dates of birth, handled with an unauthorized artificial intelligence based application, and judged the incident material."}]},"blitsAi":{"howToBuild":"On Blits.ai the inventory assistant is an **AI agent** that answers questions over the register,\nwhich lives in a **SQL knowledge base** so every answer comes from the records themselves, and a\n**knowledge base** with **hybrid retrieval** over model cards, vendor documents and approvals. An\n**agentic workflow**, triggered **on a schedule**, calls **custom functions** (REST calls to the\nprocurement system, SaaS admin consoles and the governance platform) to compare usage signals\nwith the register and draft candidate records, with **human in the loop confirmation** before\nanything is written back. Connectors from the **integration catalog** (for example ServiceNow,\nSAP, Okta, Jira, SharePoint) and **MCP** servers extend the discovery sources.\n\nOwners receive attestation requests and answer in **Microsoft Teams** or email. **PII masking**\nkeeps personal data out of prompts, **run history with a full audit trail** records every draft\nand every approval, and **test suites** check the agent's answers against the register before\neach change. The platform is model agnostic, so the inventory agent itself can run on the model\nthe organization has approved, and **EU and UAE data residency** keeps the register in region."},"faq":[{"question":"What should an AI inventory record for each system?","answer":"At minimum the owner, purpose, users, vendor or in house status, model and version, data used (including personal data), risk tier, approval status and review date. The US federal inventory is a useful public template: the Federal Reserve Board records stage, purpose, vendor, data, personal data involvement and high impact designation for each use case."},{"question":"How do you find shadow AI without monitoring employees' prompts?","answer":"Start with metadata: SaaS administration consoles, API keys to model providers, procurement records and network egress to AI services. Route each hit to a named owner to confirm or retire, and offer an approved alternative, because blocking alone pushes use further out of sight."},{"question":"Is an AI inventory required by the EU AI Act?","answer":"The Act does not set a general inventory duty, but its obligations presume one: deployers of high risk systems must monitor them and assign human oversight, and providers (and deployers that are public authorities) must register Annex III high risk systems in the EU database. Sector supervisors go further: the Monetary Authority of Singapore's proposed guidelines expect financial institutions to maintain accurate, up to date AI inventories."},{"question":"Can AI maintain the inventory on its own?","answer":"No, and it should not. AI can discover candidates, pre fill records from documents and flag stale entries, but an accountable owner confirms each record and a governance officer approves the risk tier. The inventory agent is itself a system in the register."}],"related":["model-risk-validation-copilot","internal-audit-copilot","continuous-controls-testing","vendor-due-diligence","regulatory-horizon-scanning"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog, made industry neutral and verified against public inventories and supervisory sources."},{"date":"2026-09-25","note":"Consolidation pass: industries now include manufacturing, where its evidence comes from; added SR 11-7, PRA SS1/23 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: rewrote the problem to cite only sourced claims and name CB Financial Services correctly, tightened the ECB statistic, Article 49 note, incident note and EU AI Act FAQ, added the Annex III point 4(b) caveat for employee monitoring, added seoTitle and metaDescription, and updated the federal inventory evidence to the 2025 consolidation."},{"date":"2026-09-27","note":"Removed the Blits.ai agent governance layer record, which described an agent orchestration layer rather than an AI inventory and claimed capabilities not in the platform inventory; the DOJ growth figure no longer reads as a stale register, DOJ consolidation now says use cases, added the Article 50(1) transparency note and the federal exceptions; MAS guidelines rechecked and still a closed consultation."}],"slug":"ai-model-inventory","url":"https://www.blits.ai/ai-use-cases/ai-model-inventory","benchmarks":[],"indicativeValueResult":{"low":14160,"high":247500},"evidence":["federal-reserve-board-ai-use-case-inventory","omb-federal-ai-use-case-inventory","unilever-ai-inventory-and-assurance","us-department-of-justice-ai-use-case-inventory"]},{"title":"AI that turns requirements into user stories, acceptance criteria and test cases","shortTitle":"Requirements to test cases","seoTitle":"AI test case generation from requirements","metaDescription":"AI drafts user stories and test cases from requirements for QA review. In a Tricentis case study, LTIMindtree cut simple test case design from 30 to 10 minutes.","definition":"AI that reads product requirements, specifications or recorded sessions, checks them for gaps, ambiguity and contradictions, and drafts structured user stories, acceptance criteria and test cases (for example in Gherkin) that QA engineers review, with each item traced back to the requirement it covers.","aliases":["AI test case generation from requirements","requirements engineering assistant","Gherkin and BDD scenario generation","acceptance criteria generator"],"industries":["technology","government","banking","cross-industry"],"functions":["it-and-engineering"],"patterns":["content-generation","document-processing"],"channels":["internal-tools"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"emerging","problem":"Before anyone writes code or a test, someone has to read the requirements. In banking, insurance,\nautomotive and government projects these arrive as long documents, regulatory change notices,\nspreadsheets or workshop recordings, often with hundreds or thousands of individual requirements.\nAnalysts break them into user stories, testers write test cases by hand, and the link between a\nrequirement and the tests that prove it lives in a spreadsheet that goes stale.\n\nThe cost shows up late. Ambiguous or contradictory requirements are found during testing or in\nproduction, coverage gaps are invisible until an audit asks which test proves a control, and\nskilled testers spend their time typing steps rather than thinking about risk. Generative AI is\ngood at reading and restructuring text, which makes this front end of the delivery cycle a natural\nplace to apply it, as long as people stay accountable for what is tested.","problemStats":[],"howItWorks":"1. **Ingest the source.** The assistant reads requirement documents, backlog items, change\n   requests, regulations or transcripts of recorded sessions and walkthroughs.\n2. **Check the requirements.** It flags ambiguity, missing acceptance conditions, duplicates and\n   contradictions, and classifies each requirement (functional or not, safety or security relevant,\n   in or out of scope) for an analyst to confirm.\n3. **Draft stories and criteria.** For accepted requirements it drafts user stories and acceptance\n   criteria in the team's template.\n4. **Draft test cases.** It proposes positive, negative and boundary test cases, in Gherkin or the\n   team's test format, each tagged with the requirement it covers.\n5. **Review and publish.** A QA engineer edits and approves the drafts, which are then pushed to the\n   backlog and test management tool, where a coverage view shows which requirements have no\n   approved test yet.","valueDrivers":["employee-productivity","speed","risk-reduction"],"kpis":["processing-time-reduction","accuracy","productivity-gain","time-saved-per-task"],"indicativeValue":{"referenceOrg":"A bank's delivery organization with 50 QA engineers and analysts","inputs":[{"key":"people","label":"QA engineers and analysts","low":50,"high":50,"unit":"people","note":"The reference organization."},{"key":"designShare","label":"Share of their time spent analysing requirements and writing test cases","low":0.2,"high":0.35,"unit":"fraction of working time","note":"Editorial assumption. Replace with your own time split."},{"key":"timeSaved","label":"Share of that time saved after review","low":0.25,"high":0.5,"unit":"fraction of design time","note":"Conservative against the benchmarks on this page (a Tricentis case study reports 67% to 83% less time per test case in an LTIMindtree pilot of 10 to 12 test cases), because review and correction take time and not all work is test drafting."},{"key":"loadedCost","label":"Fully loaded cost per person","low":70000,"high":120000,"unit":"USD per person per year","note":"Editorial assumption. Replace with your own blended cost, including contractors."}],"formula":"people * designShare * timeSaved * loadedCost","currency":"USD","period":"per year","resultLabel":"QA and analysis capacity released","caveat":"Capacity released, not cash saved, unless headcount or contractor spend actually changes. It leaves out the cost of the tool, integration work, and the value of defects caught earlier."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Drafting a test case from one clear requirement is easy. The work is in messy source documents, domain terms the model does not know, a house style for stories and tests, and a reliable link into the backlog and test management tools so traceability survives change.","dataPrerequisites":["Requirement sources in machine readable form (documents, backlog items, transcripts)","Templates and examples of good user stories, acceptance criteria and test cases","A glossary of domain terms, systems and products","Existing test cases linked to requirements, as examples and to find gaps"],"integrations":["Backlog and requirements tool","Test management tool","Document stores that hold specifications and change requests","Test automation framework, when drafts become automated scripts"]},"implementation":{"steps":[{"title":"Pick a well documented product area","detail":"Start with a system whose requirements are written down and whose testers are willing to compare drafts with their own work, not with the messiest legacy area."},{"title":"Teach it your formats","detail":"Give the assistant your story and test templates, Gherkin conventions, glossary and a set of approved examples, and version the prompts like code."},{"title":"Validate requirements before generating tests","detail":"Run the ambiguity, duplicate and contradiction checks first and send findings back to the business analyst. Tests generated from a bad requirement only automate the misunderstanding."},{"title":"Make traceability part of the output","detail":"Require every story and test case to carry the id of its source requirement, and reject drafts that do not. Build the coverage view from those links."},{"title":"Measure against a baseline","detail":"Time test design with and without the assistant on comparable requirements, and track how much of each draft reviewers change, not just how fast drafts appear."},{"title":"Connect to the tools last","detail":"Push approved drafts into the backlog and test management tools only after the quality of drafts is stable, and keep a human approval step on every push."}],"guardrails":["No generated story or test case enters the backlog or test library without a named reviewer's approval","Every generated item carries the id of the requirement it covers","Confidential specifications stay within approved models and regions, with no training on the organization's data","Findings about requirement quality go back to the requirement owner rather than being silently fixed in the tests","Regulated controls keep tests designed by an accountable person, with AI drafts as input only"],"humanInTheLoop":"Business analysts own the requirements and decide on every flagged ambiguity or contradiction. QA engineers review, edit and approve every story and test case, decide what is in scope and add the risk based and exploratory tests the AI does not think of. A test lead signs off coverage for each release.","kpisToInstrument":["Time to design tests per requirement, before and after","Share of each draft changed by the reviewer","Requirements with at least one approved test (coverage)","Requirement defects found before development versus found in testing or production","Defects that escape to production in areas designed with the assistant"],"failureModes":[{"title":"Plausible tests that test nothing","detail":"Drafts restate the requirement without meaningful checks or data. Review samples against a checklist and track reviewer edit rates."},{"title":"Coverage theatre","detail":"Many generated tests inflate coverage numbers while risky paths stay untested. Measure coverage by requirement and risk, not by test count."},{"title":"Invented requirements","detail":"The model fills gaps with assumptions that look like requirements. Flag assumptions separately and send them to the requirement owner."},{"title":"Traceability that breaks on change","detail":"Requirements change and generated tests are not updated. Regenerate or flag linked tests whenever a requirement changes."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"An internal assistant that drafts requirements artifacts and test cases for engineers is not listed in Annex III and does not interact with the public, so no specific obligations apply beyond AI literacy (Article 4). The system under test may itself fall under the Act."},"regulations":["eu-ai-act","dora","iso-42001","nist-ai-rmf"],"guidance":[{"title":"Article 4, AI literacy","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/4/","note":"Providers and deployers must take measures on the AI literacy of staff who use AI systems (the amended wording shown on this page asks them to support it rather than ensure a sufficient level). Here that means training analysts and testers on the limits of generated drafts."},{"title":"SP 800-218A, Secure Software Development Practices for Generative AI and Dual-Use Foundation Models","issuer":"NIST","region":"north-america","url":"https://csrc.nist.gov/pubs/sp/800/218/a/final","note":"NIST's SSDF community profile (July 2024) that adds secure development practices for producers of AI models, producers of AI systems that use them, and acquirers of those systems. A reference for the controls around an AI tool that feeds the delivery pipeline."}],"controls":["Approved tool with contractual data terms and no training on the organization's specifications","Mandatory human approval recorded against each generated test case","Traceability matrix from requirement to approved test, kept current on change","Periodic sample review of generated tests by the test lead","Change control on prompts and templates"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** built around an **AI agent** with **structured\noutput**: the workflow reads requirement documents from the **knowledge base** (PDF, Word,\nPowerPoint, spreadsheets and Markdown are ingested and retrieved with hybrid search), runs a\nvalidation step that lists ambiguities and contradictions, and then drafts user stories,\nacceptance criteria and Gherkin test cases as structured records that each carry the source\nrequirement id. **Human in the loop confirmation**, set so that every push needs approval, lets a\nQA engineer approve or reject each batch before **custom functions** or connected **MCP** tools\nwrite it to the backlog or test management system.\n\n**Prompt versioning** keeps the story and test templates under change control, **execution\ntracing** shows how each draft was produced, **PII masking** keeps personal data out of prompts,\nand **guardrails** check inputs and outputs. The platform is model agnostic, with EU and UAE data\nresidency. For teams that build conversational agents, generated test cases can be imported as JSON into Blits.ai\n**test suites**, which run them against agents with deterministic or LLM based grading."},"faq":[{"question":"How much time does AI save on writing test cases?","answer":"The published results come from small pilots. A Tricentis case study of an LTIMindtree pilot covering 10 to 12 test cases reports low complexity test cases falling from 30 minutes to 10 and high complexity ones from 2 hours to 20 to 30 minutes. Review time and edits by testers should be counted before scaling these numbers."},{"question":"Can it check the requirements themselves, not just write tests?","answer":"Yes, and that is often where the value starts. Continental Automotive built a proof of concept that uses generative AI to scan requirement documents of up to 30,000 requirements, categorize them and compare them with its feature catalogue, and NASA's verification and validation programme is building tools that assess requirement quality and traceability for analysts to review."},{"question":"How is this different from an AI coding assistant?","answer":"A coding assistant works on code in the developer's editor and drafts unit tests for that code. This use case works upstream on requirements and produces stories, acceptance criteria and functional test cases for QA, traced to the requirement, before or alongside development."},{"question":"Should generated test cases go straight into the test library?","answer":"No. Treat them as drafts: a named QA engineer reviews and approves each one, and every test keeps a link to its requirement so coverage and change impact stay visible."}],"related":["developer-coding-assistant","legacy-code-modernization","synthetic-test-data-generation","continuous-controls-testing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Unpublished by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from the federal AI use case inventory and vendor customer stories and verified against them."},{"date":"2026-09-26","note":"Fact checked against sources. Pilot figures now attributed to the Tricentis case study rather than to LTIMindtree, government added as an industry, Article 4 and NIST SP 800-218A notes corrected, the VA record no longer describes VA GPT beyond its source, and an SEO title and meta description added."},{"date":"2026-09-26","note":"Fact checked again against all sources. LTIMindtree record now describes the company and its SAP GUI pilot as the case study does, BrowserStack record notes that its year is an estimate, and the Blits.ai build notes describe human in the loop confirmation as the platform offers it."}],"slug":"requirements-to-test-case-generation","url":"https://www.blits.ai/ai-use-cases/requirements-to-test-case-generation","benchmarks":[{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":1,"median":67,"min":67,"max":67,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"browserstack-ai-test-case-generation","pooled":false},{"id":"ltimindtree-agentic-test-case-creation","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":90,"min":90,"max":90,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"browserstack-ai-test-case-generation","pooled":true}]}],"indicativeValueResult":{"low":175000,"high":1050000},"evidence":["browserstack-ai-test-case-generation","continental-ai-requirements-engineering","ltimindtree-agentic-test-case-creation","nasa-ivv-requirements-and-test-analysis","va-provar-test-case-generation"]},{"title":"AI transcription, subtitles and captions for audio and video","shortTitle":"Transcription and captioning","seoTitle":"AI transcription and captions for audio and video","metaDescription":"AI turns audio and video into timed transcripts, captions and subtitles for an editor to check. Warner Bros. Discovery cut the time to caption a file by 80%.","definition":"AI that transcribes recorded audio and video, such as podcasts, broadcasts, lessons, interviews and hearings, in several languages, separates the speakers and produces timed transcripts, subtitles and captions for a human editor to check, delivered as files for publishing or the archive.","aliases":["AI captioning","automatic subtitling","AI subtitle generation","automated closed captions","AI media transcription"],"industries":["media-and-entertainment","education","cross-industry"],"functions":["operations","knowledge-management"],"patterns":["speech-analytics","translation","content-generation"],"channels":["api","internal-tools"],"audience":"back-office","autonomy":"copilot","adoptionStage":"mainstream","problem":"Every organization that publishes audio or video has the same backlog: captions for people who\nare deaf or hard of hearing, subtitles for viewers who speak another language, and searchable\ntranscripts for the archive. Done by hand, captioning one hour of content takes many hours of\ntranscription, timing and translation. Ateme describes up to 15 hours of manual work per hour of\nvideo per language, and external providers that were slow and hard to scale. So organizations\ncaption the flagship content and leave the rest, or subtitle into one language only.\n\nThe gap is growing as content volumes grow and accessibility rules tighten. SVT, Sweden's public\nbroadcaster, says it simply could not caption its local news without automation, because it\npublishes for 21 regional stations several times a day. Speech recognition is now accurate enough\nthat the human role shifts from typing to editing: the machine produces a timed draft with\nspeakers marked, and an editor fixes names, terms and meaning before publication.","problemStats":[{"statement":"The World Health Organization estimates that over 5% of the world's population, or 430 million people, require rehabilitation for disabling hearing loss.","sourceTitle":"Deafness and hearing loss","sourceUrl":"https://www.who.int/news-room/fact-sheets/detail/deafness-and-hearing-loss","year":2026}],"howItWorks":"1. **Ingest the file.** A new recording arrives from the media asset system, learning platform,\n   podcast host or court recording system, and a job starts automatically.\n2. **Transcribe with timings.** Speech recognition produces text with word level timestamps and\n   detects the spoken language, using a custom vocabulary of names and terms. Pacers Sports &\n   Entertainment reduced its transcription error rate by 87% by tuning the model to its own\n   broadcasts.\n3. **Separate speakers.** Diarization marks who spoke when, so transcripts read as a dialogue and\n   captions can show speaker changes.\n4. **Translate and segment.** The text is translated into the target languages and cut into\n   subtitle lines that respect reading speed, line length and shot changes.\n5. **Edit and approve.** An editor reviews the draft in a subtitle editor, corrects names, numbers\n   and meaning, and approves each language before release.\n6. **Deliver files.** The system exports standard formats such as SRT and WebVTT captions and a\n   plain transcript, and returns them to the publishing platform and the archive.\n\nLive captioning, as on arena screens or live broadcasts, uses the same speech models but removes\nthe editor, so accuracy tuning and filters matter even more.","valueDrivers":["cost-to-serve","speed","inclusion-and-access","compliance","employee-productivity"],"kpis":["cost-reduction","processing-time-reduction","error-reduction","accuracy","hours-saved"],"indicativeValue":{"referenceOrg":"A regional broadcaster or publisher that captions 5,000 hours of content a year","inputs":[{"key":"hours","label":"Hours of content captioned per year","low":5000,"high":5000,"unit":"hours of content per year","note":"The reference organization."},{"key":"costPerHour","label":"Current cost of captioning one hour of content","low":150,"high":600,"unit":"USD per hour of content","note":"Editorial assumption covering in house or outsourced transcription, timing and review in one language. Replace with your own rates."},{"key":"costReduction","label":"Share of captioning cost removed","low":0.3,"high":0.5,"unit":"fraction of cost","note":"Conservative against the benchmark on this page (Google Cloud reports a 50% reduction in overall costs for Warner Bros. Discovery's AI captioning tool), because editing time remains."}],"formula":"hours * costPerHour * costReduction","currency":"USD","period":"per year","resultLabel":"Captioning cost avoided","caveat":"Covers the existing captioning volume only. It leaves out the cost of running the speech models and the editing tool, the value of content that is captioned or subtitled for the first time, extra languages, and the audience and compliance benefits."},"macroEstimates":[],"feasibility":{"complexity":"low","complexityNote":"Speech recognition, diarization and machine translation are mature and available from many providers. The work is in the workflow around them: getting files in and out of the media or learning platform, a glossary of names and terms, an editing step that editors actually like, and output formats that the players accept.","dataPrerequisites":["Access to the source audio or video files, in a quality good enough for speech recognition","A glossary of names, places, brands and technical terms per programme, course or court","House style for captions and subtitles (line length, reading speed, speaker labels, sound descriptions)","A sample of manually captioned content to measure accuracy before and after"],"integrations":["Media asset management system, learning platform, podcast host or recording system","Subtitle or transcript editor for human review","Publishing platform or video player that accepts SRT or WebVTT files","Archive or search index for transcripts"]},"implementation":{"steps":[{"title":"Measure your baseline first","detail":"Take a sample of content across programme types, speakers and languages, caption it the current way and record the time and cost. Without this baseline, no one can say whether the AI draft saves time once editing is counted."},{"title":"Compare speech engines on your own audio","detail":"Run the same sample through several speech recognition engines and compare word error rate, especially on names, accents, dialects and overlapping speech. Results on vendor benchmarks rarely match results on your material."},{"title":"Build the glossary and tune","detail":"Feed names and domain terms to the engine as custom vocabulary or a correction step. Pacers Sports & Entertainment tuned its model with its own broadcasts and name lists and pushed the error rate far below its original target."},{"title":"Put the editor at the centre","detail":"Give editors a tool that shows the draft with timings, speaker labels and low confidence words highlighted, and measure editing time per hour of content. The editor, not the model, signs off what is published."},{"title":"Add languages one at a time","detail":"Start with same language captions, then add subtitle languages where audience data shows demand, each with a native speaker review of a sample before going live."},{"title":"Automate delivery","detail":"Trigger jobs when files arrive and return approved SRT, WebVTT and transcript files to the publishing platform and archive automatically, so nothing depends on manual uploads."}],"guardrails":["A human editor approves every caption and subtitle file before publication","Custom vocabulary and a correction step for names, places and terms","Low confidence words and segments flagged for the editor, not silently published","Profanity and sensitive word checks on output, especially for content aimed at children","Recordings and transcripts of identifiable people processed and stored under the organization's retention rules"],"humanInTheLoop":"Editors review and correct every file before it is published, with extra attention to names, numbers, quotes and anything legally sensitive. For translated subtitles, a native speaker checks a sample per language and programme type. For live captioning, where no editor can intervene, a producer monitors the output and can switch captions off.","kpisToInstrument":["Word error rate per programme type, language and speaker profile, on a monthly sample","Editing time per hour of content, compared with the manual baseline","Cost per hour of captioned content, including model and editing costs","Share of published content with captions and with subtitles, per language","Errors reported by viewers or learners after publication"],"failureModes":[{"title":"Offensive or embarrassing misrecognition","detail":"The engine hears an innocent word as an offensive one, and it reaches the screen. Prevent with sensitive word checks and editor review, above all for children's content."},{"title":"Names and terms mangled","detail":"People, places and technical terms are transcribed wrongly, which undermines trust and can be defamatory. Maintain a glossary per programme and check names first in review."},{"title":"Review that becomes a rubber stamp","detail":"Editors under time pressure approve drafts without reading them. Track editing time and sample published files for errors."},{"title":"Captions that do not fit the screen","detail":"Correct text that is badly timed or too long to read. Enforce reading speed and line length rules in the segmentation step."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Transcription and captioning are not listed in Annex III and are not a prohibited practice under Article 5, so the tier depends on how captions are published. Article 50(4) requires deployers to disclose AI generated or manipulated text published to inform the public on matters of public interest, such as news captions, unless it has undergone human review or editorial control and someone holds editorial responsibility, so the editor step keeps most deployments outside this duty. The provider duty to mark output in Article 50(2) does not apply where the system does not substantially alter the input or its semantics, which fits same language transcription better than translated subtitles. Unreviewed news captions or subtitles should therefore be disclosed as automatic."},"regulations":["eu-ai-act","gdpr","eu-accessibility-act"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Paragraph 4 sets the disclosure duty for AI generated text published to inform the public on matters of public interest and exempts content under human review or editorial control; paragraph 2 exempts systems that do not substantially alter the input or its semantics."},{"title":"Guidelines 02/2021 on virtual voice assistants","issuer":"European Data Protection Board","region":"europe","url":"https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-022021-virtual-voice-assistants_en","note":"Final version of July 2021 on processing voice data in voice assistants; its analysis of voice recordings as personal data, and when they become biometric data, also applies to recordings and transcripts of identifiable speakers."}],"controls":["Documented editorial sign off per published file","Glossary and custom vocabulary with an owner per programme or course","Retention and access rules for recordings and transcripts of identifiable people","Monthly accuracy sample with word error rate reported per language","Labelling of automatic captions where no human review takes place"],"incidents":[{"title":"YouTube's Captions Insert Explicit Language in Kids' Videos","url":"https://www.wired.com/story/youtubes-captions-insert-explicit-language-kids-videos/","note":"Researchers found that automatic captions on videos from top children's channels contained inappropriate words the speakers never said, a risk for any unreviewed captioning."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **agentic workflow** triggered through an API token or on a schedule\nwhen a new file arrives. The first step uses the **self hosted transcription and speaker\ndiarization** option, with word level timestamps and automatic language detection, processed on\nBlits.ai infrastructure, or one of the other **speech to text** providers the platform supports; the\n**speech to text quality check** compares providers on word error rate with your own samples. An\nagent step corrects names and terms against a glossary in the **knowledge base**, and **machine\ntranslation** or the chosen model produces the subtitle languages.\n\nA **custom function** in the JavaScript sandbox turns the timed segments into SRT or WebVTT, and\nthe **file generation** tool returns the files. **Human in the loop confirmation** holds each\nfile until an editor approves it, and the run history keeps a full audit trail with downloadable\nrun data. **PII masking** at the gateway can redact personal data before text reaches an external model, the platform is **model\nagnostic**, and **EU and UAE data residency** keeps recordings of identifiable people in region."},"faq":[{"question":"How much does AI captioning save?","answer":"Published results are large but come from vendors. Google Cloud reports that Warner Bros. Discovery's AI captioning tool delivered a 50% reduction in overall costs and cut the time to caption a file by 80%. Your saving depends on how much editing your content needs, so measure editing time per hour on a sample first."},{"question":"Can AI captions be published without a human check?","answer":"For live events there is often no human in the loop before the caption appears, and organizations such as Pacers Sports & Entertainment rely on tuned models and moderation filters. For recorded content, an editor should check names, numbers and sensitive words, and under the EU AI Act human editorial review also removes the duty to label published text as AI generated."},{"question":"How is this different from AI meeting notes?","answer":"Meeting tools produce summaries and action items for participants. This use case produces a complete, timed transcript and caption files for publication or the archive, where every word and its timing matter and an editor signs off."},{"question":"Which accessibility rules apply?","answer":"In the EU, the European Accessibility Act requires services that give access to audiovisual media, such as players and apps, to transmit subtitles for the deaf and hard of hearing with adequate quality and in sync with sound and video. Automation is how broadcasters such as SVT caption content they could not caption by hand."}],"related":["meeting-summarization-and-action-items","public-service-translation","court-and-case-file-summarization","call-quality-and-compliance-monitoring"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with five public evidence records verified against their sources."},{"date":"2026-09-26","note":"Fact checked against sources. Sharpened the EU AI Act basis (Article 50(2) and 50(4)), the accessibility FAQ (EAA versus the Audiovisual Media Services Directive) and the Warner Bros. Discovery wording to match the source; fixed evidence details for Pacers, Comeen and Ateme; added seoTitle and metaDescription."}],"slug":"audio-and-video-transcription-and-captioning","url":"https://www.blits.ai/ai-use-cases/audio-and-video-transcription-and-captioning","benchmarks":[{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":50,"min":50,"max":50,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"warner-bros-discovery-ai-captioning-tool","pooled":true}]},{"kpi":"processing-time-reduction","label":"Cycle time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":80,"min":80,"max":80,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"warner-bros-discovery-ai-captioning-tool","pooled":true}]},{"kpi":"error-reduction","label":"Error reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":87,"min":87,"max":87,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"pacers-sports-and-entertainment-live-arena-captions","pooled":true}]}],"indicativeValueResult":{"low":225000,"high":1500000},"evidence":["ateme-multilingual-subtitle-generation","comeen-multilingual-video-subtitles","pacers-sports-and-entertainment-live-arena-captions","sveriges-television-automated-closed-captions","warner-bros-discovery-ai-captioning-tool"]},{"title":"AI translation and interpretation for multilingual public services","shortTitle":"Public service translation","seoTitle":"AI translation and interpretation for government","metaDescription":"Governments use AI to translate documents and conversations. EU eTranslation translated 891 million pages in 2025; the State Department pilots it at visa windows.","definition":"AI that translates government content, documents and conversations between officials and the public, in writing and in real time speech, so people can use public services in their own language, with human translators and interpreters reviewing what carries legal or safety weight.","aliases":["government machine translation","AI interpretation for public services","language access AI","multilingual citizen service"],"industries":["government"],"functions":["citizen-services","customer-service","operations"],"patterns":["translation","conversational-agent","speech-analytics","document-processing"],"channels":["web-chat","voice","kiosk","internal-tools","api"],"audience":"customer-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Every public service has residents who do not speak the official language well, and they tend to\nbe the people who most need services: new arrivals, disaster survivors, people in crisis.\nProfessional translation of documents is slow and costly, so agencies translate a few key pages\nand summarise the rest; interpreters are limited and take time to connect, especially for rarer\nlanguages; staff fall back on ad hoc translation, including free consumer apps, outside any\ngovernance.\n\nMachine translation and speech models now make many languages workable in seconds, but a\nmistranslated address in an emergency call, a symptom at a clinic or a sentence in an asylum\ninterview can change an outcome. The design question is where AI translation is enough and where\na human must check it.","problemStats":[],"howItWorks":"1. **Translate published content.** Web pages, letters and guidance are machine translated with\n   domain glossaries and style settings, then reviewed by a human for high impact texts.\n2. **Translate what the public sends.** Documents residents submit in other languages are\n   translated in full, with the original kept alongside for the case file.\n3. **Converse across languages.** Chat and voice assistants detect the resident's language and\n   answer in it, grounded in content in the official language.\n4. **Interpret live.** At counters, interview windows and on the phone, speech is transcribed,\n   translated and voiced or shown on screen for both sides, with a transcript kept when enabled.\n5. **Escalate to people.** Where the law or the stakes require it, a certified interpreter or\n   translator takes over, and staff can call one in at any moment.","valueDrivers":["inclusion-and-access","speed","cost-to-serve","customer-experience"],"kpis":["accuracy","processing-time-reduction","cost-reduction","interactions-handled","users-served"],"indicativeValue":{"referenceOrg":"An agency that commissions translation of 20,000 incoming documents a year","inputs":[{"key":"documents","label":"Incoming documents translated per year","low":20000,"high":20000,"unit":"documents per year","note":"The reference agency. Editorial assumption, replace with your own volume."},{"key":"humanCost","label":"Cost of human translation per document","low":30,"high":50,"unit":"USD per document","note":"FEMA puts its current cost at approximately USD 40 per document; the range brackets that figure.","sourceUrl":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv"},{"key":"machineShare","label":"Share of documents where machine translation plus light review is enough","low":0.5,"high":0.8,"unit":"fraction of documents","note":"Editorial assumption; the rest still need full human translation."},{"key":"reviewCost","label":"Cost of light human review of a machine translation","low":5,"high":10,"unit":"USD per document","note":"Editorial assumption. Replace with your own review cost."}],"formula":"documents * machineShare * (humanCost - reviewCost)","currency":"USD","period":"per year","resultLabel":"Document translation cost avoided","caveat":"Covers incoming documents only. It leaves out faster case decisions, interpreter costs on calls and at counters, the value of wider language access and the cost of the translation service."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Machine translation itself is widely available, in some cases free to public administrations; the work is in glossaries, quality review for high impact texts, keeping originals with translations, and integrating speech translation into counters, phones and interview rooms without breaking legal rights to an interpreter.","dataPrerequisites":["Language demand by service and channel, to pick languages","Domain glossaries and approved translations of key terms","Rules on which document and conversation types need certified human translation"],"integrations":["Content management system and website (translation by API)","Case management and document stores, to keep originals and translations together","Telephony, counters and interview rooms for speech translation","Contact centre and chat channels for multilingual assistants"]},"implementation":{"steps":[{"title":"Classify content and conversations by stakes","detail":"Decide which texts and conversations can use machine translation alone, which need human review, and which require a certified interpreter by law."},{"title":"Use a governed service, not phones","detail":"Replace ad hoc free apps with an approved service with data protection terms, such as the European Commission's eTranslation for eligible administrations."},{"title":"Build glossaries","detail":"Load approved translations of programme names and legal terms so the same concept is translated the same way everywhere."},{"title":"Keep the original next to the translation","detail":"Store both in the case file so a reviewer can check a disputed phrase. FEMA plans to keep the original and the translation together in survivors' files as substantiating documents."},{"title":"Pilot live interpretation with staff","detail":"Start with short, structured interactions (such as visa windows or Spanish 911 calls), keep transcripts and measure repeat questions and escalations to interpreters."}],"guardrails":["Certified human interpretation where the law requires it, and always on request","Original text or audio retained alongside every translation used in a decision","Glossaries for programme names and legal terms, maintained by the service","Data protection terms that keep public data out of model training","Visible notice to the public that a translation is machine generated"],"humanInTheLoop":"Human translators review high impact published texts and any translation used in a decision; staff can call an interpreter at any point in a conversation. Language leads sample machine translations each month by language and correct the glossary.","kpisToInstrument":["Quality scores by language on a monthly human reviewed sample","Time from document receipt to translated file","Share of conversations escalated to a human interpreter, by language","Translation cost per document and per call","Complaints and corrections linked to translation"],"failureModes":[{"title":"Errors that change a case","detail":"The Guardian reported machine translation errors that affected US asylum applications. Use certified interpreters for interviews that decide status, and keep transcripts."},{"title":"Rare languages quietly worse","detail":"Translators quoted by The Guardian in 2023 said AI tools are particularly unreliable for less documented languages, and that major tools did not offer some languages at all. Measure quality by language and route rare languages to people."},{"title":"Shadow translation","detail":"Staff paste case data into free consumer apps. Provide an approved tool and block the rest."},{"title":"Inconsistent terminology","detail":"The same benefit gets different names on different pages. Maintain glossaries, as the IRS does with its Publication 850 glossary of English and Spanish tax terms."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Assistants that talk with residents must tell people they are interacting with AI (Article 50(1)), and AI generated text published to inform the public on matters of public interest must be disclosed unless it has had human review under editorial responsibility (Article 50(4)). Internal translation that neither talks with people nor is published carries no specific obligation. Translation can also sit inside an Annex III process, such as examining asylum, visa or residence permit applications (point 7(c)) or evaluating emergency calls and dispatching emergency services (point 5(d)). Whether the translation component is itself high risk depends on its intended purpose (Article 6(3) exempts systems that only perform a narrow procedural task); either way it should be governed with that high risk process."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001","uk-gdpr","uk-atrs"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Sets when people must be told they are interacting with AI, when providers must mark generated content, and when deployers must disclose AI generated text published to inform the public. Applies from 2 August 2026."},{"title":"Limited English Proficiency (notice on the suspension of lep.gov)","issuer":"US Department of Justice, Civil Rights Division","region":"north-america","url":"https://www.justice.gov/crt/limited-english-proficiency","note":"The Department of Justice has temporarily suspended lep.gov to implement Executive Order 14224, pending an internal review; its language access materials will be replaced when new guidance is issued. Check the current federal position before relying on older guidance."},{"title":"AI translation and language tools","issuer":"European Commission","region":"europe","url":"https://commission.europa.eu/resources-partners/etranslation_en","note":"Describes eTranslation and related tools available free to eligible public administrations."}],"controls":["Language access policy stating where machine translation is allowed","Register entry for translation tools used in decisions","Data processing agreement that excludes training on public data","Monthly quality sampling by language","Staff training on when to call a human interpreter"],"incidents":[{"title":"Lost in AI translation: growing reliance on language apps jeopardizes some asylum applications","url":"https://www.theguardian.com/us-news/2023/sep/07/asylum-seekers-ai-translation-apps","note":"Reporting on US asylum cases harmed by AI translation, such as a city name translated literally in an application; volunteers describe applications denied after mistranslations."}]},"blitsAi":{"howToBuild":"On Blits.ai bots are **multi language**: each bot has a language list, flows hold localized\ncontent per language, the platform detects the resident's language and can switch mid conversation, and\n**machine translation** runs through Amazon, Google, IBM or Microsoft. An **AI agent** grounded\nin a **knowledge base** in the official language can answer in the resident's language, with\nstrong **Arabic** support (normalization, Arabic voices and regional Arabic models).\n\nOn the phone, **speech to text across nine providers** and **text to speech across thirteen**\nlet a **voice agent** converse in many languages, and **self hosted transcription with language\ndetection** keeps audio on Blits.ai infrastructure. **PII masking**, **guardrails**, **human\nhandover** to a staff member or interpreter, and **EU and UAE data residency** complete the\ngoverned setup; **test suites** check answers per language."},"faq":[{"question":"Is machine translation good enough for public services?","answer":"For information and routine conversations it can be, with glossaries and regular quality checks; the European Commission's eTranslation, free to eligible public administrations, translated 891 million pages in 2025. For decisions, treat it as a draft and keep the original: FEMA plans to translate survivors' documents in full and store the original and the translation together as substantiating documents in the survivor's file, and the State Department's citizen services pilot is assistive only, not a replacement for certified interpreters where they are required."},{"question":"Can AI interpret live at a counter or on the phone?","answer":"It is being piloted and used. The State Department is piloting live interpretation at the visa interview window, and Baltimore 911 operators can dial in an automated Spanish voice translator instead of a third party interpreter."},{"question":"Which languages can assistants cover?","answer":"Many. Montgomery County's Monty 2.0 answers in 140 languages and Madrid's visitor assistant in more than 95. Quality varies by language, so measure it for the languages your residents speak."}],"related":["citizen-information-assistant","immigration-and-visa-application-assistant","emergency-call-triage-support","benefits-eligibility-and-application-assistant","non-emergency-service-request-routing"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, researched from the US federal AI inventory, European Commission pages and public safety and city case studies, with every source checked."},{"date":"2026-09-25","note":"Consolidation pass: added UK GDPR, UK Algorithmic Transparency Recording Standard to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Article 50 basis and guidance note, replaced the suspended LEP.gov guidance, softened unsupported claims on language coverage, FEMA's status and translation quality, fixed dates, languages and wording in the IRS, FEMA, Madrid, Delaware County and eTranslation records, and added an SEO title and meta description."},{"date":"2026-09-26","note":"Second fact check against sources: added the European Commission's own 2025 eTranslation volume (891 million pages) as a metric, with the 2017 launch year, and cited it in the FAQ and meta description; removed an unsupported interpreter claim from the Delaware County record; dated the Guardian language coverage point to 2023; added the Article 6(3) nuance to the EU AI Act basis; retitled the LEP.gov guidance and toned down the Blits.ai language wording."}],"slug":"public-service-translation","url":"https://www.blits.ai/ai-use-cases/public-service-translation","benchmarks":[{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":2,"nUpTo":0,"median":445510000,"min":20000,"max":891000000,"byClaimant":{"organization":1,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"european-commission-etranslation","pooled":true},{"id":"montgomery-county-monty-chatbot","pooled":true}]}],"indicativeValueResult":{"low":250000,"high":640000},"evidence":["baltimore-911-assistive-call-taking","delaware-county-911-spanish-voice-translation","european-commission-etranslation","fema-individual-assistance-document-translation","irs-machine-translation","madrid-destino-visitmadridgpt","montgomery-county-monty-chatbot","us-department-of-state-consular-ai-interpretation"]},{"title":"AI travel and hotel booking concierge","shortTitle":"Travel and hotel booking concierge","seoTitle":"AI assistants for travel and hotel booking","metaDescription":"An AI travel concierge searches live inventory, answers booking questions and handles changes. Airbnb resolves nearly 45% of its assistant's issues without a human.","definition":"A customer facing AI assistant that turns an open travel question into a concrete trip by searching live inventory for flights, hotels, rentals, cruises and activities, comparing options and answering questions about the property and the booking, then completes or hands off the booking and supports the traveller with changes and questions before and during the stay.","aliases":["AI trip planner","AI travel assistant","hotel booking chatbot","travel concierge chatbot","virtual travel agent"],"industries":["travel-and-hospitality"],"functions":["sales","customer-service"],"patterns":["conversational-agent","recommendation-and-personalization","rag-knowledge-assistant","agentic-workflow"],"channels":["mobile-app","web-chat","whatsapp","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","problem":"Planning a trip means comparing many options. Traditional online travel search, as Priceline\ndescribes it, requires navigating filters, tabs and separate browser windows. Search boxes work\nwhen the traveller already knows the destination and dates; they fail at the questions people\nactually have (\"where is warm in February with a direct flight\", \"which of these hotels is quiet\nand walkable\", \"can we bring the dog\"). On the service side, the same pre and post booking\nquestions come back again and again: when Booking.com widened access to its Booking Assistant\nservice chatbot in 2017, it listed payment, transportation, arrival and departure times, date\nchanges, cancellations, parking, extra beds, pet policies and WiFi among the most frequently asked\ntopics.\n\nA concierge that only chats about destinations adds little. The value comes when the assistant\nis grounded in live prices and availability, in the operator's own property content and\npolicies, and in the traveller's booking, so it can recommend, answer precisely, book or change,\nand hand the conversation to a human agent or the property when needed. Several companies on\nthis page, among them Priceline, Trip.com and Holland America Line, run one assistant for both\nplanning and service.","problemStats":[],"howItWorks":"1. **Understand the trip.** The assistant asks for or infers the essentials (who travels, dates\n   or flexibility, budget, what matters) in the traveller's own words and language.\n2. **Search live inventory.** It queries the booking engine or supplier APIs for flights, hotels,\n   rentals, cruises and activities with current prices and availability, never prices from\n   memory.\n3. **Compare and explain.** It shortlists options and explains the tradeoffs from property\n   content, reviews and policies, with links to each listing.\n4. **Book or hand off.** It builds the basket and passes the traveller to checkout, or books\n   within set limits, and hands group, complex or high value requests to a human travel agent.\n5. **Support the trip.** After booking, the same assistant answers questions about the\n   reservation, makes changes and cancellations within the policy, passes requests to the\n   property and hands complaints and exceptions to a person with the context.","valueDrivers":["revenue-growth","customer-experience","cost-to-serve","inclusion-and-access"],"kpis":["containment-rate","automation-rate","cost-reduction","time-saved-per-task","conversion-rate-uplift","customer-satisfaction","users-served"],"indicativeValue":{"referenceOrg":"An online travel company or hotel group with 2 million bookings a year","inputs":[{"key":"bookings","label":"Bookings per year","low":2000000,"high":2000000,"unit":"bookings per year","note":"The reference organization."},{"key":"contactsPerBooking","label":"Assisted service contacts per booking","low":0.2,"high":0.4,"unit":"contacts per booking","note":"Editorial assumption for pre and post booking questions and changes. Replace with your own contact rate."},{"key":"containment","label":"Share of service contacts the assistant resolves","low":0.3,"high":0.45,"unit":"fraction of contacts","note":"The high end matches the benchmark on this page (Airbnb reports nearly 45% of issues that begin with its AI assistant resolved without a human agent in Q2 2026); the low end allows for a first year."},{"key":"costPerContact","label":"Cost of a human handled contact","low":4,"high":8,"unit":"USD per contact","note":"Editorial assumption for a blended chat and phone contact. Replace with your own fully loaded cost."}],"formula":"bookings * contactsPerBooking * containment * costPerContact","currency":"USD","period":"per year","resultLabel":"Human handled service contact cost avoided","caveat":"Service cost only. It leaves out the revenue effect of better conversion (the most important and least published benefit), the cost of running the AI and the integrations, and any change in cancellations or complaints."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"A planning chatbot on public content is quick to build. A concierge that quotes real prices, books and changes reservations needs live access to inventory, pricing and the booking system, authentication for post booking changes, and strict rules so it never states a price or policy it did not retrieve.","dataPrerequisites":["Live availability and pricing through the booking engine or supplier APIs","Structured property, cabin or room content (amenities, accessibility, parking, pets) with an owner","Booking, change and cancellation policies per rate and product","Contact reasons and search logs to choose the first intents"],"integrations":["Booking engine, central reservation system or GDS and supplier APIs","Customer account and loyalty system for authentication and personalization","Payment service for checkout, deposits and change fees","Property management system or messaging to pass requests to the hotel or host","Contact centre platform for handover with the conversation context"]},"implementation":{"steps":[{"title":"Start where the booking already exists","detail":"Post booking questions (parking, check in times, what is included, change a date) are frequent, well defined and measurable. Launch there first, prove containment and satisfaction, then move up the funnel into planning and search."},{"title":"Ground every price and fact","detail":"Let the assistant quote prices, availability and policies only from live tool calls and approved content, and show where the answer came from. A fluent but invented price or policy costs more than no answer."},{"title":"Keep checkout deterministic","detail":"Let the model build the basket, but run payment, terms acceptance and confirmation in a fixed flow, with the price and cancellation conditions shown exactly as the booking engine returns them."},{"title":"Roll out in waves","detail":"Follow the Holland America Line pattern: internal agents first, then employees, then a small share of website visitors, widening only when resolution and satisfaction hold."},{"title":"Measure revenue, not just deflection","detail":"Run the assistant against a control group and measure conversion, basket value and cancellations, not only contained conversations. Priceline reports higher conversion and fewer support contacts for Penny users, Trip.com reports growth in orders assisted by TripGenie, and Airbnb reports service effects only."},{"title":"Test before travellers do","detail":"Keep a regression set of planning and service conversations per market and language, including requests for prices the rate does not allow and attempts to change someone else's booking, and run it on every change."}],"guardrails":["Prices, availability and policies only from live tool results and approved content, never from the model's memory","Authentication before showing or changing a booking; changes only through an allow list of actions with limits","Checkout, payment and terms acceptance in a deterministic flow with card data tokenized","Handover for complaints, accessibility needs, groups and high value or complex itineraries","Recommendations free of undisclosed paid placement, with sponsored results labelled"],"humanInTheLoop":"Human travel agents handle complex, group and high value trips, complaints and exceptions to policy. A content owner approves property and policy content, and a team reviews a weekly sample of conversations for wrong prices, wrong policies and unfair recommendations.","kpisToInstrument":["Conversion and basket value for assistant users versus a control group","Containment on post booking contacts, counting repeat contacts within seven days as not contained","Share of answers with a price or policy that did not match the booking engine, on a sampled review","Handover rate and reasons","Customer satisfaction on assistant conversations versus human handled ones"],"failureModes":[{"title":"Invented prices and policies","detail":"The assistant states a price, fee or refund rule from memory, and the company is held responsible for it, as in the Air Canada tribunal case. Retrieve every fact and refuse when retrieval finds nothing."},{"title":"A planner nobody books from","detail":"Engagement grows but conversion does not, because the assistant is not connected to live inventory and checkout. Measure bookings, not conversations."},{"title":"Steering that breaks consumer law","detail":"Recommendations favour higher commission options without disclosure, or hide fees until checkout. Label sponsored results and show the full price early."},{"title":"Handover without context","detail":"The traveller has to repeat the trip details to a human agent. Pass the summary, the booking and the options already shown."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A customer facing assistant must tell people they are interacting with AI unless that is obvious from the context (Article 50(1), applicable from 2 August 2026). Recommending and booking travel is not listed in Annex III, so it is not high risk; consumer protection law on price transparency and fair commercial practices still applies to what it says."},"regulations":["eu-ai-act","gdpr","pci-dss","eu-accessibility-act"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Travellers must be informed that they are interacting with an AI system unless this is obvious from the context."},{"title":"Package travel, package holidays and linked travel arrangements in the EU (Your Europe)","issuer":"European Union","region":"europe","url":"https://europa.eu/youreurope/citizens/travel/holidays/package-travel/index_en.htm","note":"When an assistant combines flights, hotels and other services, the combination can become a package or linked travel arrangement with information duties and traveller rights."},{"title":"Unfair commercial practices directive","issuer":"European Commission","region":"europe","url":"https://commission.europa.eu/law/law-topic/consumer-protection-law/unfair-commercial-practices-law/unfair-commercial-practices-directive_en","note":"Rules against misleading information and practices, relevant to how an assistant presents prices, fees, rankings and sponsored results."}],"controls":["AI disclosure at the start of every conversation","Price and policy statements logged with the tool result they came from","Versioned property and policy content with an owner and review date","Labelling of sponsored or commission based recommendations","Change control and regression tests for every new market, language or action"],"incidents":[{"title":"Incident 639: Air Canada Chatbot Reportedly Provides Inaccurate Bereavement Fare Information, Leading to Customer Overpayment","url":"https://incidentdatabase.ai/cite/639/","note":"A Canadian small claims tribunal held Air Canada responsible in 2024 for its website chatbot's wrong statement about bereavement fare eligibility and ordered it to pay damages, rejecting the argument that the chatbot was a separate legal entity. A company can be held responsible for what its assistant says about fares and refund rules."}]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with **custom functions** that call the booking engine,\navailability and pricing APIs and the reservation system, and a **knowledge base** with\nproperty, room and policy content retrieved with hybrid search. **SQL knowledge bases** can\nhold structured property attributes for precise filtering. Checkout and booking changes run as\n**flows** with deterministic steps, and the **payment** service sends payment links with card\ntokenization; the agent recommends and explains, and **rich cards** for flights and hotels show\nthe options in the chat.\n\nThe same agent serves **web chat, WhatsApp and voice**, reaches a mobile app through the REST or\nWebSocket API channel, and can run as a **digital human** streamed to a browser, for example on\na lobby screen. It detects the traveller's language and answers in it. **Guardrails** and **PII\nmasking** check every turn, **human handover** passes the trip and the conversation to a travel\nagent, and **test suites** replay planning and service conversations per market before every\nchange. **Analytics** show interactions, satisfaction and sentiment per channel, conversation\nlogs show why travellers were handed over, and the platform is model agnostic."},"faq":[{"question":"Do AI travel assistants actually increase bookings?","answer":"Some operators report it, few publish numbers. Trip.com says TripGenie assisted order volume grew about 400% year on year, and Priceline says Penny users show higher conversion in early testing without giving a figure. Measure conversion against a control group before claiming revenue."},{"question":"How much service volume can the assistant take?","answer":"Airbnb reports that nearly 45% of issues that begin with its AI assistant were resolved without a human agent in Q2 2026, and that support cost per booking fell about 16% year on year, partly because of the assistant. Priceline estimates Penny users saved nearly ten minutes per trip compared with calling support."},{"question":"Should the assistant book on its own?","answer":"It can build the basket and handle simple changes within limits, but payment, terms and confirmation should run in a fixed flow with the exact price and conditions from the booking engine. Complex, group and high value trips belong with a human agent."}],"related":["flight-disruption-and-rebooking-agent","conversational-shopping-assistant","personalized-marketing-at-scale","outbound-reminder-and-confirmation-agent","first-line-contact-centre-agent"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with five travel and hospitality deployments verified against their sources."},{"date":"2026-09-25","note":"Consolidation pass: added European Accessibility Act to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Booking.com 2017 chatbot description (an expansion of the pilot, not a first launch), made the revenue measurement step match what Priceline, Trip.com and Airbnb actually report, limited the Blits.ai build notes to capabilities in the feature inventory, set the Holland America Line deployment year to 2024 from the story's first archive date, and named Trip.com as the Singapore based brand that issued the TripGenie release."}],"slug":"travel-and-hotel-booking-concierge","url":"https://www.blits.ai/ai-use-cases/travel-and-hotel-booking-concierge","benchmarks":[{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":30,"min":30,"max":30,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"booking-com-ai-trip-support-and-voice","pooled":true}]},{"kpi":"containment-rate","label":"Containment rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":45,"min":45,"max":45,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"airbnb-ai-customer-support-assistant","pooled":true}]},{"kpi":"cost-reduction","label":"Cost reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":16,"min":16,"max":16,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"airbnb-ai-customer-support-assistant","pooled":true}]},{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":10,"min":10,"max":10,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"priceline-penny-agentic-travel-assistant","pooled":true}]}],"indicativeValueResult":{"low":480000,"high":2880000},"evidence":["airbnb-ai-customer-support-assistant","booking-com-ai-trip-support-and-voice","holland-america-line-anna-digital-concierge","priceline-penny-agentic-travel-assistant","trip-com-tripgenie-ai-travel-assistant"]},{"title":"AI tutor that coaches students through problems","shortTitle":"AI tutor for students","seoTitle":"AI tutor for students: results and risks","metaDescription":"AI tutors guide students with hints, not answers. Harvard's CS50 has run one since 2023; a Khanmigo trial in 18 Tennessee schools found small gains and rare use.","definition":"An AI tutor that works with a student on course material in a conversation, asking questions and giving hints instead of handing over answers, grounded in the course content and set up by the school or teacher, with limits on use and a clear route to a human teacher.","aliases":["AI tutoring assistant","Socratic AI tutor","virtual tutor for students","AI homework helper"],"industries":["education"],"functions":["customer-service"],"patterns":["conversational-agent","rag-knowledge-assistant"],"channels":["web-chat","mobile-app"],"audience":"customer-facing","autonomy":"assist","adoptionStage":"emerging","problem":"A 2026 working paper on AI tutoring sums up the research on human tutoring: the best one to one\nprogrammes produce gains of a third of a standard deviation or more, but high dosage tutoring often\ncosts several thousand dollars per student per year. In large courses the queue for help is the\nbottleneck: Harvard's CS50 recalls times when office hours became unmanageable and the average\nwait could be as long as an hour.\n\nGeneral purpose chatbots answer every question fully and fluently, which is exactly what a\nlearner does not need: a finished answer skips the struggle that produces learning, and it makes\ncheating trivial. The job is different from a help desk. A tutor has to hold back, ask the next\nquestion, spot the misconception and push the student to do the work, within the course's rules\non academic honesty.\n\nThe evidence so far is mixed. A randomized trial of Khan Academy's Khanmigo in 18 Tennessee\nmiddle schools found small gains that resembled those from Khan Academy practice without AI, and\nthat almost every student tried the tutor but rarely engaged it in substantive mathematical\ndialogue; the authors suggest low engagement as one explanation. A World Bank pilot in Nigeria, run with\nteacher support, reported large gains in six weeks, and larger gains for students who\nattended more sessions. The Khanmigo authors conclude that realizing the promise of AI tutoring\nwill require getting students to use it, not just giving them access.","problemStats":[{"statement":"A 2026 EdWorkingPaper by Philip Oreopoulos and Nina Low, citing Nickow et al. (2024), puts the average effect of tutoring programmes at roughly 0.3 standard deviations, and notes that high dosage programmes often cost several thousand dollars per student per year.","sourceTitle":"One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment","sourceUrl":"https://edworkingpapers.com/sites/default/files/ai26-1551.pdf","year":2026}],"howItWorks":"1. **Set the scope.** The school or teacher defines the course, the material the tutor may use\n   and the rules: no full solutions to graded work, adherence to the academic honesty policy,\n   escalation topics such as wellbeing concerns.\n2. **Ground the tutor in the course.** Lecture notes, readings and worked examples are indexed so\n   the tutor explains in the course's own terms and cites where a concept is taught.\n3. **Coach, do not solve.** The tutor asks what the student has tried, gives the smallest useful\n   hint, checks understanding with a question and only then moves on. A second check reviews each\n   reply before the student sees it and can reject or retry replies that give away an answer.\n4. **Limit and pace use.** A cap on questions per period (CS50 uses a \"heart\" system) stops\n   students from replacing thinking with hundreds of prompts, and prompts inside the exercise flow nudge students who are\n   stuck but not asking.\n5. **Keep the teacher in the loop.** Teachers see aggregated topics and misconceptions, can read\n   conversations under the school's policy, and take over for anything personal or sensitive.","valueDrivers":["inclusion-and-access","customer-experience","employee-productivity"],"kpis":["users-served","interactions-handled","customer-satisfaction"],"indicativeValue":{"referenceOrg":"A school district with 10,000 students in grades 6 to 12","inputs":[{"key":"students","label":"Students with access to the tutor","low":10000,"high":10000,"unit":"students","note":"The reference district."},{"key":"activeShare","label":"Share of students who use the tutor substantively","low":0.15,"high":0.35,"unit":"fraction of students","note":"Conservative on purpose. In the Khanmigo trial on this page, 96 percent of students tried the tutor but the median student messaged it on only a third of practice days. Editorial assumption, replace with your own usage data."},{"key":"hoursPerStudent","label":"Tutoring hours per active student per year","low":5,"high":15,"unit":"hours per student per year","note":"Editorial assumption, replace with your own usage data."},{"key":"equivalence","label":"Value of an AI tutoring hour relative to a human tutoring hour","low":0.1,"high":0.4,"unit":"fraction of a human tutoring hour","note":"Editorial assumption. The Khanmigo trial on this page found gains similar to Khan Academy practice without AI, so the trial evidence does not support treating an AI tutoring hour as equal to a human one."},{"key":"tutorCostPerHour","label":"Cost of an hour of human tutoring","low":20,"high":40,"unit":"USD per hour","note":"Editorial assumption for group or online tutoring, replace with your local rate."}],"formula":"students * activeShare * hoursPerStudent * equivalence * tutorCostPerHour","currency":"USD","period":"per year","resultLabel":"Equivalent value of tutoring time delivered","caveat":"A proxy, not a learning outcome. It prices tutoring time, discounted heavily because AI tutoring is not equivalent to a human tutor, and leaves out licence and integration costs, teacher time to supervise and the risk that students use the tutor to avoid work. The larger randomized trial on this page (Khanmigo, 18 Tennessee middle schools) found no clear gain from the AI tutor beyond what the same practice platform delivers without it, so the value may be close to zero where students rarely engage; the positive Nigerian pilot was short, ran with teacher support and was reported by the World Bank team ahead of formal publication. Measure learning gains against a comparison group before claiming more."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"A chatbot is easy; a tutor that reliably holds back answers, stays inside the course and is used well by students is hard. Most of the work is pedagogy, safeguarding and embedding the tutor in the exercise flow, not integration.","dataPrerequisites":["Course materials, worked examples and rubrics the school has the right to use","The course's academic honesty policy written as rules the tutor can follow","A set of real student questions and misconceptions to test against"],"integrations":["Learning management system or course platform (single sign on, course roster)","Exercise or practice platform, so the tutor can appear where students get stuck","Safeguarding and wellbeing referral process"]},"implementation":{"steps":[{"title":"Start with one course and one teacher team","detail":"Pick a course with high demand for help and a teacher team willing to shape the tutor's behaviour. Write down what the tutor must never do (give a full solution to graded work) and what it should always do (ask what the student tried)."},{"title":"Ground it in the course","detail":"Index the course's own notes, readings and examples, and make the tutor cite where a concept is taught. Refuse or redirect questions outside the course."},{"title":"Add an answer check","detail":"Review every reply before the student sees it with a second pass that rejects replies that contain full solutions or break the honesty policy. CS50 reports that instructions alone were not enough."},{"title":"Put the tutor where students get stuck","detail":"Embed it in the exercise flow and prompt it after a wrong answer, rather than as a separate chat tab students must choose to open. The Khanmigo trial shows that optional access alone produces little use."},{"title":"Pace use and involve teachers","detail":"Cap questions per student per period, show teachers the common misconceptions each week and agree how teachers follow up with students who over rely on the tutor or show signs of distress."},{"title":"Evaluate learning, not usage","detail":"Compare learning outcomes with a comparison group over at least a term, and report engagement honestly, including off topic use and attempts to extract answers."}],"guardrails":["A reply check that blocks full solutions to graded work and anything against the honesty policy","Answers grounded in approved course material, with a refusal outside the course","A cap on questions per student per period","Age appropriate content filters and escalation of wellbeing or safeguarding signals to staff","No emotion recognition of students, which the EU AI Act prohibits in education"],"humanInTheLoop":"Teachers own the course scope, the rules and the follow up. They review aggregated topics and a sample of conversations each week, handle any safeguarding signal, and decide grades; the tutor never grades or places a student.","kpisToInstrument":["Share of students who use the tutor substantively each week, not just once","Share of stuck moments (wrong answers) in which the student asks the tutor for help","Replies blocked or retried by the answer check","Learning gains against a comparison group over a term","Student and teacher satisfaction"],"failureModes":[{"title":"Access without engagement","detail":"Students try the tutor once and stop, or ask it off topic questions. Embed it in the work and prompt it at the moment of error."},{"title":"Answer vending","detail":"Students talk the tutor into giving the solution, which removes the learning. Check replies before they are shown and log extraction attempts."},{"title":"Over reliance","detail":"A few students ask hundreds of questions instead of thinking. Cap questions per period and let teachers follow up."},{"title":"Confidently wrong explanations","detail":"The tutor explains a concept incorrectly. Ground it in course material, test it on known misconceptions and let students flag errors to teachers."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"A tutor that only converses with students falls under the transparency duty of Article 50. It becomes high risk under Annex III point 3(b) when it evaluates learning outcomes, including when those outcomes are used to steer a student's learning process, and under point 3(c) when it assesses the level of education a student should receive. Inferring students' emotions is prohibited in education institutions under Article 5(1)(f)."},"regulations":["eu-ai-act","gdpr","uk-gdpr","nist-ai-rmf","iso-42001"],"guidance":[{"title":"Guidance for generative AI in education and research","issuer":"UNESCO","region":"global","url":"https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research","note":"UNESCO's first global guidance on generative AI in education; it proposes protecting learners' data privacy and setting an age limit for independent conversations with generative AI platforms."},{"title":"Generative artificial intelligence (AI) in education","issuer":"UK Department for Education","region":"europe","url":"https://www.gov.uk/government/publications/generative-artificial-intelligence-in-education","note":"The department's position on generative AI tools in schools and colleges, to be read together with its product safety expectations for generative AI."},{"title":"Regulation (EU) 2024/1689 (AI Act), Annex III point 3, education and vocational training","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng","note":"Lists as high risk AI that evaluates learning outcomes, including when those outcomes steer the learning process, and AI that assesses the appropriate level of education a person will receive (point 3(b) and 3(c))."},{"title":"Article 5, prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Prohibits AI systems that infer the emotions of a natural person in education institutions, except for medical or safety reasons."}],"controls":["AI disclosure to students and parents, with the school's rules for use","A data protection impact assessment covering minors and conversation logs","Inventory entry with an accountable owner per course","Regression tests for answer giving and off topic behaviour on every prompt or model change","Teacher review of aggregated topics and a sample of conversations"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a tutoring persona and prompt versioning, grounded in a\n**knowledge base** that holds only the course's own material, retrieved with hybrid search.\n**Output guardrails** with an admin authored policy check every reply before the student sees\nit, so replies that hand over a full solution or break the honesty policy are blocked, and the\ndeterministic content scanner and profanity detection keep the conversation age appropriate.\n\nThe tutor runs in the **web chat** widget or inside the school's own learning platform through\nthe **REST or WebSocket API** channel, with voice input and spoken replies where that helps\nyounger or struggling readers, and multi language support for students learning in a second\nlanguage. **Human handover** routes wellbeing signals to staff, **test suites** with LLM grading,\nrun after each change, check that the tutor still refuses to give answers, and **analytics** and\nconversation logs let teachers follow usage and review a sample of conversations. The platform is model agnostic and supports EU data residency."},"faq":[{"question":"Do AI tutors improve learning?","answer":"Sometimes, and it depends on how they are used. A World Bank pilot in Edo, Nigeria, run with teacher support, reported gains of about 0.3 standard deviations in six weeks. A two year randomized trial of Khanmigo in 18 Tennessee middle schools found small gains similar to Khan Academy practice without AI; the authors suggest low engagement as one explanation, as the median student messaged the tutor on only a third of the days they practiced."},{"question":"How do you stop an AI tutor from giving students the answers?","answer":"Instructions alone are not enough. Harvard's CS50 combines prompting with code that tries to evaluate each reply before the student sees it and sometimes rejects or retries it, plus a limit on questions per period. Test the tutor against real attempts to extract answers on every change."},{"question":"Is an AI tutor high risk under the EU AI Act?","answer":"A tutor that only converses is subject to the transparency duty. It is high risk under Annex III point 3(b) if it evaluates learning outcomes, including when those outcomes steer a student's learning, or under point 3(c) if it assesses the level of education a student should receive. Inferring students' emotions in education is prohibited outright."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the education scope with evidence from Harvard CS50, a Khanmigo randomized trial and a World Bank pilot in Nigeria, all checked against the sources. Editor pass the same day kept the Khanmigo paper's hedges, removed unsourced and overstated claims, sourced the tutoring effect and cost, corrected the Annex III note (now linked to EUR-Lex) and removed a rate limiting claim not in the feature inventory; a second pass aligned the Khanmigo engagement wording with the paper and limited the analytics claim to what the feature inventory lists. A third pass corrected the value caveat (both studies are randomized), recorded the Khanmigo trial as a pilot under the district's name and set adoption to emerging (one production deployment, two pilots), restored the World Bank hedges and teacher support wording, and replaced ill fitting taxonomy terms."}],"slug":"ai-tutor-for-students","url":"https://www.blits.ai/ai-use-cases/ai-tutor-for-students","benchmarks":[],"indicativeValueResult":{"low":15000,"high":840000},"evidence":["hamilton-county-schools-khanmigo-trial","harvard-cs50-duck-ai-tutor","world-bank-edo-nigeria-ai-tutor-pilot"]},{"title":"AI vegetation management for power lines","shortTitle":"Power line vegetation management","seoTitle":"AI vegetation management for utility power lines","metaDescription":"Satellite imagery and AI show where trees threaten power lines, so utilities trim by risk, not by calendar. National Grid in Massachusetts and Entergy use it.","definition":"AI that analyses satellite, aerial or lidar imagery of the land along power lines to estimate where and how fast vegetation will grow into the lines or fall onto them, and turns that into a risk based trimming and hazard tree removal plan, replacing fixed trimming cycles and manual patrols.","aliases":["satellite vegetation management","utility vegetation management with AI","risk based tree trimming","hazard tree detection","right of way vegetation monitoring"],"industries":["energy-and-utilities"],"functions":["network-operations","field-service"],"patterns":["computer-vision","prediction-and-scoring"],"channels":["internal-tools","mobile-app"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","segment":"grid","problem":"Vegetation growing into or falling onto lines interrupts supply. The NERC transmission vegetation\nstandard notes that major outages and operational problems have resulted from overgrown vegetation\ninterfering with transmission lines, and AiDASH calls vegetation one of the grid's biggest threats.\nAiDASH says vegetation programs have for decades largely followed a fixed formula: trim a set share\nof the system each year and repeat the cycle. Its case studies describe Entergy on a standardized\nfive year cycle, maintaining about 20% of its system each year, and National Grid's Massachusetts\nnetwork on a typical five year cycle, where deciding whether a circuit needed pruning took manual\nfield reviews.\n\nAiDASH notes that growth rates vary by region, species, weather and circuit, so a uniform cycle\nmeans unnecessary work in some areas and elevated risk in others. According to AiDASH,\nNational Grid had deferred work for four years running rather than fund it, and in the year after\npruning its average circuit saw only an 8% reduction in customers interrupted and no significant\nimprovement in tree events or customer minutes interrupted. The same case study describes rising\ncosts for routine maintenance and strict regulations to prevent outages and manage fire risk.","problemStats":[{"statement":"NERC's transmission vegetation standard FAC-003-5 states that major outages and operational problems have resulted from interference between overgrown vegetation and transmission lines.","sourceTitle":"FAC-003-5 Transmission Vegetation Management","sourceUrl":"https://www.nerc.com/pa/Stand/Reliability%20Standards/FAC-003-5.pdf","year":2021}],"howItWorks":"1. **Image the whole network.** Satellite imagery, supplemented by aerial or lidar data where\n   needed, covers every span, repeated as often as the budget allows.\n2. **Measure the vegetation.** Computer vision models identify trees, their height, their distance\n   to the conductors and signs of poor health, span by span.\n3. **Predict growth and risk.** Models combine species, growth rates, weather and outage history to\n   estimate when each span will become a risk.\n4. **Plan the work.** Circuits are scheduled for trimming when their risk warrants it, hazard trees\n   are prioritised for removal, and budgets are allocated where they avoid the most outages.\n5. **Dispatch and audit.** Work goes to contractors with maps, and new imagery checks that the\n   work was done and done well.\n6. **Measure reliability.** Tree related outages, customers interrupted and minutes interrupted are\n   tracked per circuit, and the results tune the risk model.","valueDrivers":["cost-to-serve","risk-reduction","customer-experience","employee-productivity"],"kpis":["cost-savings","cost-reduction"],"indicativeValue":{"referenceOrg":"An electric distribution utility with 20,000 line miles","inputs":[{"key":"vegBudget","label":"Annual vegetation management spend","low":20000000,"high":40000000,"unit":"USD per year","note":"Editorial assumption for a network of this size, replace with your own budget."},{"key":"savingShare","label":"Share of spend saved by trimming on risk instead of on a fixed cycle","low":0.03,"high":0.1,"unit":"fraction of vegetation spend","note":"Editorial assumption, deliberately below the 20% expense reduction AiDASH claims in its own marketing on the Entergy page, which is a vendor average and not an evidence record. For comparison, AiDASH reports USD 2M in efficiencies in National Grid's first few years and USD 1M in avoided cost on about 13,500 line miles, without saying whether the two overlap."}],"formula":"vegBudget * savingShare","currency":"USD","period":"per year","resultLabel":"Vegetation management spend avoided or redeployed","caveat":"Counts spend only. It leaves out imagery and software costs, the reliability value of fewer tree events and customer minutes interrupted (which AiDASH reports for National Grid's worked circuits) or of beating reliability targets (which AiDASH reports for Entergy), avoided storm restoration cost and reduced wildfire risk."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Vendors deliver the imagery and models as a service. The work for the utility is a clean network model in GIS, outage history per circuit, and changing contracts and planning from fixed cycles to risk based work.","dataPrerequisites":["A GIS model of the network with spans, circuits and voltage","Outage history per circuit with cause codes","Trimming history and contractor work records","Local knowledge of species, growth and regulatory clearance rules"],"integrations":["Geographic information system of the network","Outage management system for cause coded outage history","Work management system and contractor portals","Mobile apps for crews and auditors"]},"implementation":{"steps":[{"title":"Prove it on real territory first","detail":"Run the imagery and risk model on a large, representative area and compare its risk ranking with outage history and a field check before changing the plan. National Grid, according to AiDASH, ran its 2020 proof of concept on its entire Massachusetts footprint rather than a portion of it, and the first model run produced its FY2021 work plan."},{"title":"Agree the clearance rules and risk appetite","detail":"Encode regulatory clearance requirements and decide how much risk the utility accepts per circuit type, so the model plans to rules, not just to growth."},{"title":"Move the plan from cycles to risk","detail":"Schedule circuits by predicted risk, keep a floor of mandatory inspections, and give planners the final say on the plan."},{"title":"Change the contracts","detail":"Contractor agreements built on miles trimmed per cycle need to change to work orders by span and risk, with imagery based audits."},{"title":"Measure reliability per circuit","detail":"Track tree related events, customers interrupted and minutes interrupted on treated circuits against the previous cycle and against untreated comparable circuits."}],"guardrails":["Regulatory clearance and inspection obligations always override the model's schedule","Planners approve the annual plan and any deferral of work on a high risk circuit","Field verification of high risk findings before removal of trees on private land","Imagery of private property used only for network maintenance purposes"],"humanInTheLoop":"Vegetation planners and arborists review the model's risk ranking, approve the work plan and decide on hazard tree removals. Field crews confirm conditions on site and report back, and reliability engineers review outcomes per circuit each year.","kpisToInstrument":["Tree related outage events per 100 line miles, before and after","Customers interrupted and customer minutes interrupted from tree causes (SAIFI and SAIDI contribution)","Vegetation spend per line mile","Share of high risk spans treated before the storm or fire season","Audit pass rate of contractor work checked against new imagery"],"failureModes":[{"title":"Model trusted over the rulebook","detail":"A span with low predicted risk still has a legal clearance obligation. Keep regulatory rules as hard constraints."},{"title":"Stale imagery","detail":"Imagery that is too old misses storm damage and fast growth. Agree refresh frequency by region and season."},{"title":"No change in contracts","detail":"Contractors paid per mile on a cycle keep trimming on the cycle. Align contracts with risk based work orders."},{"title":"Benefits that cannot be shown","detail":"Reliability varies with weather, so one good year proves little. Compare treated and comparable untreated circuits over several years."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Annex III point 2 makes AI systems high risk when they are intended as safety components in the management and operation of the supply of electricity. Recital 55 defines such components as systems used to directly protect the physical integrity of critical infrastructure or the health and safety of persons and property. A system that only feeds a multi year trimming plan, which vegetation planners review and approve before crews act, informs maintenance rather than directly protecting the network, and is then usually minimal risk. The assessment changes when the design acts directly on protection, for example when vegetation risk scores automatically trigger fire risk protection settings or switch lines off without a person deciding; such a system should be assessed as a possible safety component. Standard GDPR duties apply where imagery shows private property or people."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001"],"guidance":[{"title":"FAC-003-5 Transmission Vegetation Management","issuer":"North American Electric Reliability Corporation (NERC)","region":"north-america","url":"https://www.nerc.com/pa/Stand/Reliability%20Standards/FAC-003-5.pdf","note":"The mandatory reliability standard for vegetation clearances on North American transmission lines, which any AI based plan must still meet."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 2 lists safety components in the management and operation of critical infrastructure, including electricity supply."},{"title":"Recital 55, safety components of critical infrastructure","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/recital/55/","note":"Defines safety components as systems that directly protect the physical integrity of critical infrastructure or the health and safety of persons and property, the test that decides the tier here."}],"controls":["Documented mapping of regulatory clearance rules into the planning model","Annual review of model performance against tree related outages per circuit","Records of approvals for deferred work on high risk spans","Data protection rules for imagery of private land"],"incidents":[]},"blitsAi":{"howToBuild":"The imagery analysis and risk model come from the utility's vegetation platform. Blits.ai adds\nthe conversations around the work. An **AI agent** with a **SQL knowledge base** over risk scores,\nwork plans and outage history lets planners ask which circuits are due and why, and a\n**knowledge base** with the utility's clearance rules and vegetation policy answers questions\nconsistently.\n\nFor customers and landowners, an agent on **web chat, WhatsApp, SMS and voice** explains planned\ntrimming on their property, answers policy questions from the approved knowledge base and books\nor changes access appointments through **custom functions**, with **human handover** to a\nvegetation specialist for disputes and tree removal requests. **Agentic workflows** can prepare\nadvance notices for approval, **guardrails** and **PII masking** protect customer data, and the\nplatform is model agnostic with EU and UAE data residency."},"faq":[{"question":"What results do utilities report from AI vegetation management?","answer":"AiDASH reports three sets of figures for National Grid. Its latest account of the Massachusetts program gives average improvements of 22% in tree events, 29% in customers impacted and 43% in customer minutes interrupted on circuits worked in FY2022 to 2025, measured 12 months after each circuit is worked; a case study on the Massachusetts service area reports declines of 30%, 38% and 55% on the same measures in the year after pruning; and its page on a NextGrid Alliance Summit 2025 talk by National Grid's vegetation strategy manager lists decreases of 26.4%, 30.2% and 46.5%, without naming a service area. The pages do not say which period the improvements are compared against, so treat them as indicative. For Entergy, AiDASH reports that it beat its vegetation reliability (Veg SAIFI) targets by over 30% in 2020 and 2021 on flat budgets, and quotes an Entergy vice president saying vegetation impacts to customers improved by more than 20% within the first year."},{"question":"Does satellite imagery replace field patrols and lidar?","answer":"Not fully. AiDASH's National Grid case study describes satellite imagery as a quick way to see vegetation conditions over an entire service area and a foundation for building a plan, and its later account notes that National Grid now also has lidar data and drone data from three in house drone pilots alongside the satellite outputs. Our advice: keep lidar, drones or field checks for precise clearance measurement and for verifying high risk findings before work on private land."},{"question":"Is AI vegetation management regulated as high risk AI?","answer":"Usually not under the EU AI Act when it only informs a maintenance plan that planners approve, because it does not directly protect the network in the sense of Recital 55. A design in which the risk scores directly trigger protective actions, such as fire risk settings or switching lines off, needs a proper assessment as a possible safety component. Existing clearance rules, such as NERC FAC-003 for North American transmission lines, still apply to the plan it produces."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the discovery workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the energy and utilities scope with National Grid and Entergy evidence checked against the sources."},{"date":"2026-09-27","note":"Editor review fixes. Removed an unsourced forecast horizon and unsourced problem claims, scoped National Grid to Massachusetts, reworked the EU AI Act tier around Recital 55, noted both sets of National Grid figures and recorded its cost results."},{"date":"2026-09-27","note":"Second review fixes. Sourced or attributed the fixed cycle problem claims, corrected National Grid's earlier practice and proof of concept, added AiDASH's FY2022 to 2025 National Grid figures and a second Entergy case study, attributed the satellite imagery claim to AiDASH, limited the value caveat to what the sources report and lowered the high saving share."},{"date":"2026-09-27","note":"Third review pass. Rechecked every problem, evidence and FAQ claim against the sources with the source tool and recorded the NERC outage statement in problemStats."}],"slug":"power-line-vegetation-management","url":"https://www.blits.ai/ai-use-cases/power-line-vegetation-management","benchmarks":[{"kpi":"cost-savings","label":"Cost savings","unit":"currency","currency":"USD","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":2000000,"min":2000000,"max":2000000,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"national-grid-satellite-vegetation-management","pooled":true}]}],"indicativeValueResult":{"low":600000,"high":4000000},"evidence":["entergy-satellite-vegetation-management","national-grid-satellite-vegetation-management"]},{"title":"AI visitor and tour guide for cities, museums and events","shortTitle":"Visitor and tour guide","seoTitle":"AI tour guides for museums, cities and events","metaDescription":"An AI guide tells visitors the story behind each artwork or place in their own language. See how National Gallery Singapore and Art Basel use one.","definition":"A location and context aware AI guide, often spoken, that tells visitors of cities, museums, heritage sites and events the stories behind what is around them and answers their questions in their own language, grounded in the organization's curated content and in the visitor's position or the object they scan.","aliases":["AI tour guide","AI museum guide","AI docent","AI audio guide","virtual city guide"],"industries":["travel-and-hospitality","media-and-entertainment","government"],"functions":["customer-service","marketing","knowledge-management"],"patterns":["rag-knowledge-assistant","conversational-agent","voice-agent","computer-vision","translation"],"channels":["mobile-app","web-chat","whatsapp","voice","kiosk","digital-human"],"audience":"customer-facing","autonomy":"autonomous","adoptionStage":"early-adopters","problem":"Museums, heritage sites, cities and events hold far more knowledge than a visitor ever sees. Wall\nlabels are short, recorded audio guides cover a fixed route in a handful of languages, and human\nguides are limited by schedules, group sizes and the languages they speak. Visitors who are\ncurious but not experts, who speak another language, or who cannot read small print or see the\nobject well get the thinnest experience. National Gallery Singapore puts it plainly: people\nwalking into a museum often feel lost or intimidated.\n\nUpdating guide content is slow as well. Every new exhibition, route or audience (children,\nspecialists, first time visitors) means rewriting and rerecording the same stories. Tourism\nboards face the same problem at city scale, with visitor questions arriving in dozens of languages\nand outside office hours. A generative guide changes the unit of work: curators maintain one body\nof approved content, and the guide tells it in the visitor's language, at the visitor's level,\nabout the thing in front of them.","problemStats":[{"statement":"UN Tourism estimates that 1.52 billion international tourists were recorded worldwide in 2025, almost 60 million more than in 2024.","sourceTitle":"UN Tourism World Tourism Barometer","sourceUrl":"https://www.unwto.org/un-tourism-world-tourism-barometer-data","year":2026}],"howItWorks":"1. **Know where the visitor is.** The guide receives context with each question: the room or\n   stop, a GPS position, a QR code or object number, or a photo of the artwork that is matched\n   against the collection (Art Basel's Lens returns artist and gallery details in about two\n   seconds).\n2. **Retrieve approved content.** It looks up the object or place in the collection database or\n   points of interest list and retrieves the curated texts, research and practical information\n   (opening hours, accessibility, routes) that belong to it.\n3. **Tell the story in the visitor's terms.** The model turns that content into a short spoken or\n   written answer in the visitor's language and at their level, and can connect it to interests the\n   visitor mentions, as National Gallery Singapore's G(ai)le does with pop culture references.\n4. **Answer follow up questions.** Visitors ask in their own words, by voice or text; the guide\n   stays within the approved content and says so when it does not know.\n5. **Suggest what next.** It recommends the next stop, event or exhibit based on the visitor's\n   position, time and interests.\n6. **Feed insight back.** Anonymized questions show curators and marketing teams what visitors\n   actually want to know, and where the content has gaps.","valueDrivers":["customer-experience","inclusion-and-access","revenue-growth","employee-productivity"],"kpis":["users-served","interactions-handled","customer-satisfaction","accuracy"],"indicativeValue":{"referenceOrg":"A city museum with 1 million visitors a year and a guided route of about 150 stops","inputs":[{"key":"stops","label":"Stops or objects with guide content","low":100,"high":200,"unit":"stops","note":"Editorial assumption for a medium sized museum or city route. Replace with your own."},{"key":"languages","label":"Additional languages offered","low":4,"high":8,"unit":"languages","note":"Editorial assumption. For comparison, National Gallery Singapore's guide supports four languages in total. Replace with your own visitor language mix."},{"key":"costPerStopLanguage","label":"Cost to write, translate and record one stop in one language","low":100,"high":250,"unit":"USD per stop per language","note":"Editorial assumption for professional translation and voice recording. Replace with your own agency rates."},{"key":"refreshShare","label":"Share of stops rewritten or added each year","low":0.25,"high":0.5,"unit":"fraction of stops per year","note":"Editorial assumption covering new exhibitions, rotations and audience versions."}],"formula":"stops * languages * costPerStopLanguage * refreshShare","currency":"USD","period":"per year","resultLabel":"Multilingual guide content production cost avoided","caveat":"Counts only the avoided cost of writing, translating and recording guide content in extra languages. It leaves out the cost of running the AI and curating the source content, any revenue from longer visits, return visits or bookings, and the accessibility benefit, which is the main reason many institutions build a guide."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The conversation is the easy part. The work is in the content: a clean collection or points of interest database with identifiers, approved texts per object, rights to use images and research, and a reliable way to know where the visitor is (QR codes, beacons, GPS or image matching). Spoken delivery in a noisy hall and accurate recognition of local names and accents need testing on site.","dataPrerequisites":["A collection management or points of interest database with stable identifiers per object or place","Approved, curated texts per object or place, with an owner and a review date","Practical visitor information (opening hours, routes, accessibility, events) from one source","A pronunciation list of names of artists, places and local terms for speech recognition and synthesis","For image recognition, reference photos of each object and the rights to use them"],"integrations":["Collection management system or content management system","The organization's visitor app, website or kiosk software","Positioning (GPS, beacons, QR codes) or an image matching service","Ticketing and events calendar for recommendations and practical questions","Analytics for anonymized question and usage reporting"]},"implementation":{"steps":[{"title":"Start from the content, not the model","detail":"Pick one gallery, route or district and bring its content into shape: an identifier per object or place, the approved texts, practical information and a short brief on tone. The quality of the guide will never exceed the quality of this content."},{"title":"Decide how the guide knows where the visitor is","detail":"QR codes or object numbers are the most reliable and cheapest. Image matching needs no codes on the wall, but it needs reference photos and tests in real lighting and angles; GPS works outdoors but not between rooms. Many deployments use two methods with a fallback."},{"title":"Write the voice of the guide","detail":"Agree with curators how the guide speaks, what it may interpret and what it must leave open. National Gallery Singapore spent a long time tuning prompts so its docent informs without imposing a single view of the art."},{"title":"Test with real visitors and real languages","detail":"Build a test set of questions per stop in every supported language, including names that are hard to pronounce and questions the content cannot answer, and run it on every content or model change. Then run a pilot on the floor and listen to what visitors ask."},{"title":"Design for access from day one","detail":"Offer audio only and large text modes, captions for spoken replies, and a way to use the guide without looking at the screen. These features serve visually impaired visitors and everyone who wants to look at the object rather than a phone."},{"title":"Close the loop with curators","detail":"Review anonymized questions every month: frequent questions without a good answer become new content, and questions that show confusion feed back into labels and routes."}],"guardrails":["Answers only from the approved collection and visitor content, with a clear \"I do not know\" when the content is silent","Facts such as dates, attributions and prices are taken from the source record, never generated","Clear disclosure that the guide is an AI, and a way to reach staff for practical or safety questions","No identification of people in camera images; image matching limited to objects in the collection","Location and conversation data kept only as long as needed, anonymized for analytics"],"humanInTheLoop":"Curators and educators own the content and the tone of the guide, approve new stories before they go live, and review a sample of conversations each month for errors of fact and tone. Front of house staff handle anything practical the guide cannot, such as lost items, accessibility assistance and safety.","kpisToInstrument":["Share of visitors who use the guide, by language and entry point","Questions per session and share answered from content versus declined","Factual accuracy on a monthly sample reviewed by curators","Visitor satisfaction with the guide compared with the recorded audio guide or human tour","Speech recognition errors on names and local terms, per language"],"failureModes":[{"title":"Confident but wrong stories","detail":"The guide invents a date, an attribution or an anecdote that sounds right. Prevent with retrieval from approved records, facts taken from structured fields and curator review of samples."},{"title":"Wrong object, wrong story","detail":"Image matching or positioning picks the neighbouring object and the guide tells the wrong story. Set a confidence threshold and ask the visitor to confirm or scan the code when unsure."},{"title":"A screen between visitor and art","detail":"Visitors stare at their phones instead of the object. Design audio first and short answers, as National Gallery Singapore did with its \"eyes up\" mode."},{"title":"Stale practical information","detail":"Opening hours, closed rooms or event times in the guide differ from reality. Pull practical information from one live source instead of copying it into the content."}]},"risk":{"euAiAct":{"tier":"limited","basis":"A visitor facing assistant must make clear that people are interacting with AI (Article 50), and synthetic speech should be identifiable as AI generated. It is not high risk. It would change if the camera feature were used to identify or categorise visitors by biometric data, which a guide does not need: remote biometric identification, biometric categorisation and emotion recognition are high risk under Annex III point 1, and biometric categorisation that infers sensitive traits is prohibited under Article 5."},"regulations":["eu-ai-act","gdpr"],"guidance":[{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system, and providers of systems that generate synthetic audio must mark it in a machine readable format as artificially generated."},{"title":"Guidelines 02/2021 on virtual voice assistants","issuer":"European Data Protection Board","region":"europe","url":"https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-022021-virtual-voice-assistants_en","note":"How GDPR applies to voice assistants, including transparency, purpose limitation and retention of voice recordings."}],"controls":["AI disclosure at the start of every session and on spoken replies","Content ownership per object or place, with review dates and a change log","Test set per stop and language, run before every content or model change","Data protection impact assessment covering location data, voice and camera images","Retention limits and anonymization for conversation logs used in analytics"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the guide is an **AI agent** grounded in a **knowledge base** built from the\ncurated texts, documents and crawled pages of the organization, retrieved with hybrid search,\nplus a **SQL knowledge base** with the collection or points of interest records (identifier,\ntitle, location, opening hours). The organization's own app sends the visitor's stop, position\nor scanned object number with each message through the **REST or WebSocket API channel**, and\n**custom functions** call the collection system, ticketing or events calendar for live details.\n\nThe same agent runs in the embeddable **web chat**, **WhatsApp** and a **digital human** on a\nkiosk, and speaks through **text to speech** from many providers with custom voices, with **TTS\ncaching** so fixed stop narration plays instantly. **Multi language** support detects the\nvisitor's language and switches mid conversation. **Guardrails** keep answers inside the\napproved content, **analytics** show what visitors ask, **test suites** replay\nquestions per stop before each change, and the platform is **model agnostic**, with **EU and UAE\ndata residency**."},"faq":[{"question":"Does an AI guide replace human guides and docents?","answer":"In the deployments on this page, no. National Gallery Singapore describes its AI docent as a new teammate alongside the tours team, and uses it to draft tour versions that staff refine. The Gallery notes that in person tours are not always available in the language a visitor prefers, which is where the guide helps."},{"question":"How does the guide know which object or place the visitor means?","answer":"Through context sent with each question: a QR code or object number, GPS outdoors, or image recognition. Art Basel matches a photo of an artwork against its index and returns the artist and gallery in about two seconds. Codes are the most reliable; image matching needs reference photos and a confidence threshold."},{"question":"How do you stop the guide from making up facts about the collection?","answer":"Ground every answer in approved records, take hard facts such as dates and attributions from structured fields, make the guide say when the content is silent, and have curators review a sample of conversations every month."},{"question":"Is an AI tour guide high risk under the EU AI Act?","answer":"Not as described here. It carries the Article 50 transparency duties: visitors must know it is an AI, and synthetic speech must be identifiable. Location, voice and camera data still fall under the GDPR, so a data protection impact assessment is advisable. Using the camera to identify or categorise visitors would change the assessment."}],"related":["travel-and-hotel-booking-concierge","citizen-information-assistant","public-service-translation","retail-store-and-kiosk-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-26","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, with three public evidence records verified against their sources."},{"date":"2026-09-26","note":"Fact checked against sources. Removed the National Gallery Singapore speech recognition figure (a component claim, not a guide outcome), tightened the EU AI Act basis and FAQ answers, added source dates, SEO title and meta description."}],"slug":"ai-visitor-and-tour-guide","url":"https://www.blits.ai/ai-use-cases/ai-visitor-and-tour-guide","benchmarks":[],"indicativeValueResult":{"low":10000,"high":200000},"evidence":["art-basel-companion-app-and-lens","bloomberg-connects-gemini-audio-guides","madrid-destino-visitmadridgpt","national-gallery-singapore-gaile-ai-docent"]},{"title":"Conversational AI for insurance quote and buy","shortTitle":"Conversational quote and buy","seoTitle":"AI agents for insurance quote and buy","metaDescription":"AI agents that quote, explain cover and bind insurance policies in a chat. Lemonade says its bot AI Maya and its APIs sell 98% of its policies.","definition":"A customer facing AI agent that sells insurance directly in a conversation: it asks the rating questions in plain language, explains cover options, returns a price from the insurer's rating engine, handles objections and takes payment to bind the policy, with a licensed human available for advice and anything outside its limits.","aliases":["insurance sales chatbot","conversational insurance onboarding","digital insurance sales agent"],"industries":["insurance"],"functions":["sales","customer-service"],"patterns":["conversational-agent","agentic-workflow","recommendation-and-personalization","voice-agent"],"channels":["web-chat","mobile-app","whatsapp","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"distribution","problem":"Buying insurance online often means long forms with questions customers do not understand (\"what\nis your property's construction type?\"), and each confusing question can make a customer leave\nbefore the price. A form can show a price, but it cannot answer a question about what is and is not\ncovered. Intermediaries still dominate some lines: Lemonade's 2025 annual report notes that\nhomeowners insurance in the United States is sold primarily via agents.\n\nA conversation can ask fewer, better questions, explain terms in plain language and answer\nquestions about cover at the moment of purchase. The hard part is doing that while staying inside\ndistribution rules: demands and needs testing, product information disclosure, suitability for\ninvestment based products and honest explanations of exclusions.","problemStats":[],"howItWorks":"1. **Understand the need.** The agent asks what the customer wants to protect and captures the\n   demands and needs the law requires before recommending a product.\n2. **Collect rating data conversationally.** It asks only the rating questions the product needs,\n   prefills what it can from approved data sources and explains why each question matters.\n3. **Price through the rating engine.** The agent calls the insurer's own rating and underwriting\n   rules; it never calculates or negotiates price itself.\n4. **Explain and compare cover.** Retrieval over the product documents answers questions on\n   limits, excesses and exclusions, and the required product information is shown before purchase.\n5. **Bind and pay.** The customer confirms the details, accepts the documents and pays through a\n   secure payment link; the policy is issued and documents are sent.\n6. **Hand over when needed.** Advice requests, referrals from underwriting rules, vulnerable\n   customers and complex needs go to a licensed human with the conversation so far.","valueDrivers":["revenue-growth","customer-experience","cost-to-serve","inclusion-and-access"],"kpis":["conversion-rate-uplift","automation-rate","revenue-uplift","customer-satisfaction","users-served"],"indicativeValue":{"referenceOrg":"A direct personal lines insurer with 200,000 online quote journeys a year","inputs":[{"key":"quoteJourneys","label":"Online quote journeys started per year","low":200000,"high":200000,"unit":"journeys per year","note":"The reference insurer."},{"key":"baseConversion","label":"Baseline conversion from quote start to purchase","low":0.08,"high":0.12,"unit":"fraction of journeys","note":"Editorial assumption for a direct channel. Replace with your own funnel data."},{"key":"relativeUplift","label":"Relative conversion uplift from the conversational journey","low":0.05,"high":0.15,"unit":"fraction of baseline conversion","note":"Editorial assumption; no insurer on this page publishes a controlled uplift. Measure it with an A/B test."},{"key":"averagePremium","label":"Average annual premium per new policy","low":400,"high":600,"unit":"USD per policy","note":"Editorial assumption. Replace with your own average premium."}],"formula":"quoteJourneys * baseConversion * relativeUplift * averagePremium","currency":"USD","period":"per year","resultLabel":"Additional gross written premium","caveat":"Premium, not profit. It leaves out loss ratio effects of the new business, acquisition costs saved or added, the cost of running the agent and regulatory work, and the risk that a poorly designed flow lowers conversion."},"macroEstimates":[{"statement":"Evident reports that sales and distribution matched underwriting and pricing with nine new AI use cases among the 30 insurers it tracked in the second quarter of 2026, and notes that Aviva, Liberty Mutual Insurance and Allianz now generate live quotes directly inside ChatGPT.","sourceTitle":"Evident: Insurance Use Case Trends Q2 2026","sourceUrl":"https://evidentinsights.com/insights/insurance-use-case-trends-q2-2026","year":2026}],"feasibility":{"complexity":"high","complexityNote":"The conversation is the easy part. Binding real policies needs the rating engine, underwriting rules, document generation, payments and policy issuance behind APIs, plus distribution compliance (demands and needs, product information, record keeping) built into the flow.","dataPrerequisites":["Rating and underwriting rules exposed through an API","Product documents (terms, product information documents) per product version","A mapping of every rating question to plain language explanations and allowed answers","Approved prefill data sources and their use conditions"],"integrations":["Rating engine and underwriting rules","Policy administration for issuance","Payment service provider","Document generation and delivery","CRM and handover to licensed sales staff"]},"implementation":{"steps":[{"title":"Start with a simple product","detail":"Renters, travel, pet or simple home cover with few rating factors and low advice needs are the right first products. Leave life and investment based products for later."},{"title":"Put compliance in the flow, not the prompt","detail":"Demands and needs capture, required disclosures and document acceptance should be deterministic steps that cannot be skipped, with a record of each."},{"title":"Let the rating engine own the price","detail":"The agent passes answers to the rating API and presents the result; it must not estimate, discount or negotiate prices."},{"title":"Test for misselling","detail":"Build test conversations where customers ask leading questions (\"so I'm covered for flood?\") and check that exclusions are explained correctly."},{"title":"Run an A/B test","detail":"Compare the conversational journey with the existing form on conversion, cancellations in the cooling off period and complaints before switching traffic."}],"guardrails":["Price only from the rating engine; no free text price statements","Required disclosures and demands and needs steps enforced by the flow","Coverage answers only from the current product documents, with refusal when unsure","Clear AI disclosure and a route to a licensed human","Payment by secure link or tokenized card, never card numbers in the chat transcript"],"humanInTheLoop":"Licensed sales staff take advice requests, underwriting referrals and vulnerable customers. Compliance reviews a sample of completed sales every month, and product owners approve every change to questions, explanations and flows.","kpisToInstrument":["Conversion from quote start to purchase versus the form journey, by product","Cancellations within the cooling off period","Complaints and misselling indicators per 1,000 sales","Handover rate to licensed staff and reasons","Satisfaction at purchase"],"failureModes":[{"title":"The agent talks about price","detail":"A model rounds, estimates or promises a discount. Keep all price statements tied to the rating engine output."},{"title":"Exclusions glossed over","detail":"The agent reassures instead of explaining, which surfaces later as a declined claim. Test leading questions and cite the wording."},{"title":"Unsuitable sales","detail":"The agent recommends a product without capturing needs. Enforce the demands and needs step."},{"title":"Invisible fairness issues","detail":"Conversational data (language, typing style) leaks into underwriting or pricing. Keep the conversation layer separate from rating inputs."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"The conversational layer carries the Article 50 transparency duty. If the system assesses risk or sets prices for life or health insurance of natural persons, that part is high risk under Annex III point 5(c); pricing for property and casualty products is not listed."},"regulations":["eu-ai-act","gdpr","uk-consumer-duty","pci-dss","dora","eu-idd"],"guidance":[{"title":"Insurance Distribution Directive (IDD)","issuer":"European Insurance and Occupational Pensions Authority","region":"europe","url":"https://www.eiopa.europa.eu/browse/regulation-and-policy/insurance-distribution-directive-idd_en","note":"Sets the rules for how insurance is distributed in the EU, including demands and needs, product information and conduct, which also apply to AI sales journeys."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"People must be informed that they are interacting with an AI system, unless that is obvious from the context."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(c) makes AI for life and health insurance risk assessment and pricing of natural persons high risk."}],"controls":["Record of demands and needs, disclosures shown and documents accepted for every sale","Rating engine version logged with each quote","Monthly compliance sample of AI completed sales","Complaints and cancellation monitoring by journey","PCI DSS scope kept outside the conversational platform through tokenization"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the sales journey is a **flow** for the regulated steps (demands and needs,\ndisclosures, confirmation) with an **AI agent** block for the open conversation, and a **knowledge\nbase** of product documents for coverage questions. **Custom functions** call the rating engine and\npolicy system over REST, so the price always comes from the insurer's own rules, and **payment\nlinks** through Stripe, Mollie or Adyen take payment with card data tokenized at the payment\nprovider.\n\nThe same journey runs on **web chat, WhatsApp and voice**, and inside the insurer's own app through\nthe API channel, with rich cards for options and product recommendations and **multi language** support. Output **guardrails** with admin authored\npolicies catch price or cover statements the agent should not make, **human handover** routes\nadvice requests to licensed staff, and\n**test suites** replay misselling scenarios on every change. **Analytics** and flow statistics show\nhow the journey performs, and **human in the loop approval** can be required for agent actions\nabove a set threshold."},"faq":[{"question":"Do insurers really sell policies through chatbots?","answer":"Some do. Lemonade's 2025 annual report says its bot AI Maya and its APIs sell 98% of its policies, with Maya collecting information, personalizing coverage, quoting and taking payment in a chat; the filing does not split that share between Maya and the APIs. Evident also notes that Aviva, Liberty Mutual Insurance and Allianz now generate live quotes directly inside ChatGPT."},{"question":"Can the AI give advice?","answer":"Only within the distribution rules that apply. The cautious design keeps the agent to information and non advised sales on simple products, captures demands and needs in the flow, and hands advice requests to licensed staff."},{"question":"Is a sales chatbot high risk under the EU AI Act?","answer":"The chat itself needs AI disclosure under Article 50. Any component that assesses risk or sets prices for individual life or health insurance is high risk under Annex III; for home, motor or travel pricing the Act does not list it as high risk."}],"related":["insurance-policy-servicing-agent","insurance-broker-and-agent-assistant","insurance-renewal-and-retention","insurance-pricing-and-actuarial-copilot","conversational-shopping-assistant"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from insurer filings, analyst research and vendor case studies verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: added Insurance Distribution Directive to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: tied the agent distribution point to Lemonade's 10-K, made the Waterdrop FAQ answer concrete, aligned the Blits.ai section with the feature inventory, refined the Article 50 note and added an SEO title and description."},{"date":"2026-09-27","note":"Review fixes: removed the Waterdrop AI Medical Insurance Expert figure from the FAQ (the source does not say it is a customer facing sales chatbot), described Lemonade's 98% as a share of sales rather than an automation rate, narrowed the agent distribution point and aligned the Evident and Blits.ai wording with the sources."}],"slug":"conversational-insurance-quote-and-buy","url":"https://www.blits.ai/ai-use-cases/conversational-insurance-quote-and-buy","benchmarks":[],"indicativeValueResult":{"low":320000,"high":2160000},"evidence":["lemonade-ai-maya-and-cx-ai"]},{"title":"Conversational AI for loan application intake","shortTitle":"Loan application intake","seoTitle":"AI assistant for loan application intake","metaDescription":"Chat and voice assistants can take loan applications and hand origination a complete file, while the credit decision stays with the bank's governed credit process.","definition":"A conversational assistant on web, app, messaging or voice that explains loan products, captures the application through dialogue in the customer's language, checks documents and basic eligibility rules, and hands a complete, structured application to origination, without making the credit decision.","aliases":["loan application chatbot","conversational loan origination","AI loan intake assistant","digital loan application assistant"],"industries":["banking"],"functions":["lending-and-credit","sales","customer-service"],"patterns":["conversational-agent","document-processing","rag-knowledge-assistant","voice-agent"],"channels":["web-chat","mobile-app","whatsapp","voice"],"audience":"customer-facing","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"front-office","problem":"Loan application forms can lose customers. They ask for terms the applicant does not understand,\nin a language that may not be their first, and they rarely explain which product fits. Some\napplicants give up, and many completed applications arrive with missing documents or wrong\nfigures. According to Google Cloud, Oper Credits reports that most loan applications in Belgium\nare returned for missing or incorrect information, and each return is another round between the\ncustomer and the credit team. Where branches are few, a phone or messaging channel can be the most\npractical way to apply.\n\nThe bank also has to be careful. Intake collects personal and financial data, touches product\nsuitability and responsible lending duties, and sits right next to the credit decision. An\nassistant that drifts from collecting information into telling people they will or will not be\napproved creates regulatory and fairness risk.","problemStats":[{"statement":"Oper Credits, a mortgage digitisation company serving about 20 banks, says only 30 to 40% of loan applications in Belgium are complete and compliant on first submission, as reported by Google Cloud.","sourceTitle":"1,302 real-world gen AI use cases from the world's leading organizations","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","year":2025}],"howItWorks":"1. **Understand the need.** The assistant asks what the customer wants to finance and explains\n   the relevant products from approved, current product content, in the customer's language.\n2. **Check the basics first.** Published eligibility rules (age, residency, minimum income,\n   product limits) are checked by a rules service before the customer invests time, and the\n   assistant explains any rule that is not met.\n3. **Capture the application in conversation.** It asks one question at a time, fills the\n   application fields, and reads uploaded payslips, statements and identity documents with\n   document AI, asking again when something is unclear.\n4. **Validate completeness.** Before submission it checks that every required field and\n   document is present and consistent, so the application arrives complete.\n5. **Hand over a structured file.** The complete application goes into the loan origination\n   system with a summary; the credit decision is made there, by the bank's governed credit\n   process and people.\n6. **Keep the customer informed.** The assistant tells the customer what happens next and when,\n   and hands over to a lending specialist on request or when the case is complex.","valueDrivers":["revenue-growth","customer-experience","speed","inclusion-and-access","cost-to-serve"],"kpis":["conversion-rate-uplift","processing-time-reduction","automation-rate","customer-satisfaction","handling-time-reduction"],"indicativeValue":{"referenceOrg":"A lender receiving 50,000 personal loan applications a year","inputs":[{"key":"applications","label":"Applications received per year","low":50000,"high":50000,"unit":"applications per year","note":"The reference lender."},{"key":"incompleteShare","label":"Share of applications that need follow up for missing or wrong information","low":0.3,"high":0.5,"unit":"fraction of applications","note":"Editorial assumption. According to Google Cloud, Oper Credits, a mortgage digitisation company, says only 30 to 40% of loan applications in Belgium are complete on first submission, which implies 60 to 70% need follow up; this range stays below that to be conservative for other markets and products."},{"key":"followUpAvoided","label":"Share of that follow up the assistant prevents","low":0.3,"high":0.6,"unit":"fraction of follow up","note":"Editorial assumption, replace with your own pilot result."},{"key":"minutesPerFollowUp","label":"Staff minutes per follow up","low":20,"high":40,"unit":"minutes per application","note":"Editorial assumption covering calls, emails and re checking documents."},{"key":"costPerMinute","label":"Fully loaded cost of a lending staff minute","low":0.6,"high":1,"unit":"USD per minute","note":"Editorial assumption, replace with your own fully loaded cost."}],"formula":"applications * incompleteShare * followUpAvoided * minutesPerFollowUp * costPerMinute","currency":"USD","period":"per year","resultLabel":"Application follow up cost avoided","caveat":"Counts avoided follow up work only. It leaves out additional completed applications (the larger prize, but hard to predict), faster time to decision, the cost of the AI and the integration with origination."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Capturing an application in conversation is well understood. The work is in the eligibility rules service, reliable document reading, writing clean records into loan origination and keeping the assistant on the right side of the line between information and a credit decision.","dataPrerequisites":["Approved, versioned product content and eligibility rules","The application data model of the origination system","Sample documents per type for testing extraction","Past applications with the reasons they were returned or declined"],"integrations":["Loan origination system","Document storage and document AI","Identity verification and authentication","CRM for leads and follow up","Contact centre or lending specialists for handover"]},"implementation":{"steps":[{"title":"Pick one product and one channel","detail":"Start with a simple product (a personal loan or a card) on the channel your applicants use most, often the app or WhatsApp, and measure completion and quality against the web form."},{"title":"Separate rules from conversation","detail":"Put eligibility checks and required fields in a rules service the assistant calls. The model explains; it does not decide whether someone qualifies."},{"title":"Map every field to a question and a check","detail":"For each application field write the plain language question, the validation and what the assistant says when the answer does not fit."},{"title":"Make documents conversational","detail":"Let the customer upload a photo, read it, show what was extracted and ask them to confirm, instead of retyping."},{"title":"Define the handover and the wording","detail":"Agree the phrases the assistant may use about outcomes (\"your application is complete and will be reviewed\") and ban the ones it may not (\"you are approved\")."},{"title":"Monitor fairness from day one","detail":"Track drop off and completion by language, channel and customer segment, so the assistant does not become a filter that disadvantages some groups."}],"guardrails":["The assistant never states or implies a credit decision","Eligibility checks come only from the versioned rules service","Product explanations come only from approved, current content","Personal and financial data masked in prompts and logs, and kept in region","Handover to a specialist on request, for complex cases and on vulnerability signals"],"humanInTheLoop":"Credit officers or the bank's governed credit models decide every application. Lending specialists take over complex or vulnerable cases, and a sample of conversations is reviewed each week for accuracy, suitability and fair treatment.","kpisToInstrument":["Application completion rate versus the web form, by segment and language","Share of applications complete at first submission","Time from first message to complete application","Handover rate and reasons","Complaints mentioning the assistant"],"failureModes":[{"title":"Implied approvals","detail":"The assistant says something that sounds like a decision. Restrict outcome wording and test for it."},{"title":"Stale rules or rates","detail":"The assistant quotes last month's criteria. Version the rules and content with owners and review dates."},{"title":"Silent filtering","detail":"Customers are discouraged before they apply, unevenly across groups; in the US, Regulation B bars discouraging applicants on a prohibited basis. Measure drop off by segment and review the eligibility wording."},{"title":"Bad data, faster","detail":"Extraction errors enter origination unnoticed. Show extracted values to the customer for confirmation and sample them."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"Explaining products and capturing an application is limited risk with an Article 50 disclosure. If the assistant evaluates creditworthiness or filters applicants on its own assessment, it falls under Annex III point 5(b) and becomes high risk, so keep the decision in the governed credit process."},"regulations":["eu-ai-act","gdpr","eba-loan-origination","uk-consumer-duty","dora","us-ecoa-reg-b"],"guidance":[{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) lists AI systems used to evaluate the creditworthiness of natural persons or establish their credit score as high risk, except systems used to detect financial fraud; intake must stay outside that scope or meet the high risk duties."},{"title":"Regulation B, § 1002.4 General rules","issuer":"Consumer Financial Protection Bureau","region":"north-america","url":"https://www.consumerfinance.gov/rules-policy/regulations/1002/4/","note":"Section 1002.4(b) bars statements to applicants or prospective applicants that would discourage them on a prohibited basis, and the official commentary names interview scripts that do so; an intake assistant's eligibility wording falls under the same rule."},{"title":"Responsible lending","issuer":"Australian Securities and Investments Commission","region":"asia-pacific","url":"https://asic.gov.au/regulatory-resources/credit/responsible-lending/","note":"Example of national responsible lending obligations that an intake assistant's questions and wording must support."}],"controls":["AI disclosure at the start and a clear route to a person","Versioned eligibility rules and product content with change control","Log of every question, answer and document captured","Fairness monitoring of drop off and completion by segment","Data protection impact assessment for the personal and financial data collected"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the application runs as a **flow** that holds the regulated structure: **ask\nquestion** blocks with the **sensitive data flag**, **multiple entity check** (slot filling),\n**validation**, **receive attachment** for documents and an **authentication** step. An **AI\nagent** inside the flow explains products from a **knowledge base** of approved content with\nhybrid retrieval, and **custom functions** call the eligibility rules service and write the\ncompleted application into the origination system.\n\nThe same flow serves **web chat**, **WhatsApp**, **voice** and, through the **API channel**, the\nbank's mobile app, in every bot language, with **language detection** and per language content. **PII masking** keeps personal\ndata out of prompts, **guardrails** block outcome wording the bank has banned, and **human\nhandover** passes complex cases to lending specialists. **Test suites** check the flow and the\nagent's wording on every change, and **flow statistics** show where applicants stop. EU and UAE\nregions support data residency."},"faq":[{"question":"Can a chatbot approve a loan?","answer":"It should not. The assistant collects and checks information; the credit decision stays with the bank's governed credit process. An assistant that assessed creditworthiness itself would be a high risk system under the EU AI Act."},{"question":"Are banks using AI assistants for loan applications today?","answer":"Yes, though outcome data is scarce. Absa's Agentforce based assistant Abby is reported to help customers apply for loans, and Figure uses Gemini powered chatbots in its home equity lending. Rocket Mortgage's assistant, built with Sierra, goes further than intake: it also pulls credit and takes clients to preapproval, and the vendor says it handles more than 400,000 chat conversations a month. Oper Credits, which serves about 20 banks, uses Vertex AI to automate document checks on mortgage applications."},{"question":"What should we measure first?","answer":"Completion rate against your web form and the share of applications that arrive complete, both split by language and customer segment. They show value and fairness at the same time."}],"related":["home-loan-assistant-and-prequalification","digital-onboarding-assistant","alternative-data-credit-scoring","application-and-identity-fraud-detection","inbound-lead-qualification-agent","adverse-action-explanations"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI use case catalog and verified against Absa, Figure, Oper Credits, Lloyds Banking Group and regulator sources."},{"date":"2026-09-25","note":"Consolidation pass: added ECOA and Regulation B to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added seoTitle and metaDescription, a cited Oper Credits problem statistic and the Regulation B discouragement rule; softened unsourced claims in the problem; corrected the Absa pricing guideline attribution; limited the mobile app claim in How to build to the API channel; added Rocket Mortgage to the FAQ."},{"date":"2026-09-27","note":"Review fixes: removed the unsupported organization claim from the metaDescription; the FAQ now says Rocket Mortgage's assistant goes past intake to credit pulls and preapproval; clarified the Oper Credits assumption note; How to build now names flow statistics."}],"slug":"conversational-loan-application-intake","url":"https://www.blits.ai/ai-use-cases/conversational-loan-application-intake","benchmarks":[],"indicativeValueResult":{"low":54000,"high":600000},"evidence":["absa-abby-agentic-assistant","figure-lending-chatbots","lloyds-mortgage-income-verification","oper-credits-mortgage-document-verification","rocket-mortgage-digital-assistant"]},{"title":"Dynamic AML customer risk rating with machine learning","shortTitle":"Dynamic customer risk rating","seoTitle":"AML customer risk rating with machine learning","metaDescription":"How banks use explainable machine learning to rate and update AML customer risk, with a playbook, a value model and EU AI Act status.","definition":"Explainable machine learning that produces the money laundering risk rating itself: it computes and continuously updates each customer's rating from due diligence data, products, geography, behaviour and screening results, and shows which factors drive the rating and when enhanced due diligence is warranted.","aliases":["customer risk rating","AML customer risk scoring","dynamic risk assessment","behavioural customer risk rating"],"industries":["banking","payments","wealth-and-asset-management"],"functions":["financial-crime-compliance","risk-management"],"patterns":["prediction-and-scoring","anomaly-detection"],"channels":["internal-tools","api"],"audience":"back-office","autonomy":"supervised-agent","adoptionStage":"early-adopters","segment":"middle-office","problem":"Every bank must rate the money laundering risk of each customer and apply more scrutiny to the\nhigher risk ones. In many banks the rating is a static scorecard filled in at onboarding and\nrefreshed at fixed intervals; Fenergo describes periodic KYC checks as typically carried out\nannually or every two years. Between reviews the customer's behaviour can change completely\nwithout the rating moving.\n\nStatic scorecards can also drift out of line with reality: customers can end up high risk\nbecause of a single attribute, which adds work for enhanced due diligence teams, while risky behaviour in\na low rated customer can go unnoticed until an alert or a law enforcement request. Supervisors\nexpect a documented, risk based approach, and industry principles such as the Wolfsberg Group's\nask that machine learning results can be explained from the data that went in.","problemStats":[{"statement":"A Fenergo study found that more than half of financial institutions spend between 61 and 150 days on client KYC reviews, at an average cost of $2,200 per review.","sourceTitle":"Ongoing Customer Due Diligence with Perpetual KYC","sourceUrl":"https://resources.fenergo.com/blogs/perpetual-kyc-pkyc","year":2026}],"howItWorks":"1. **Combine the data.** Due diligence attributes, products and channels, geographies, screening\n   results, transaction behaviour and alert history are brought together per customer.\n2. **Score with explanations.** An interpretable model, or a model with feature attribution,\n   produces a risk score and the factors that raise or lower it, alongside the bank's\n   regulatory minimum rules (for example PEPs are always high risk).\n3. **Update on events.** The score is recalculated when behaviour or data changes, not only on the\n   review date, and a material move creates a task.\n4. **Route the work.** Customers moving into higher risk bands are queued for enhanced due\n   diligence with the drivers listed; moves down are reviewed before they reduce scrutiny.\n5. **Govern the model.** Rating distribution, stability and outcomes (alerts, reports, exits per\n   band) are monitored, and the model is validated like any other risk model.","valueDrivers":["compliance","risk-reduction","employee-productivity"],"kpis":["detection-rate-improvement","productivity-gain","processing-time-reduction"],"indicativeValue":{"referenceOrg":"A bank with 500,000 customers and 15,000 enhanced due diligence reviews a year","inputs":[{"key":"eddReviews","label":"Enhanced due diligence reviews per year","low":15000,"high":15000,"unit":"reviews per year","note":"The reference bank."},{"key":"hoursPerReview","label":"Hours per enhanced due diligence review","low":2,"high":5,"unit":"hours per review","note":"Editorial assumption. Replace with your own."},{"key":"misratedShare","label":"Share of reviews avoided because customers were rated high only by static rules","low":0.1,"high":0.25,"unit":"fraction of reviews","note":"Editorial assumption. Replace with the results of your own rerating exercise."},{"key":"costPerHour","label":"Fully loaded analyst cost per hour","low":40,"high":70,"unit":"USD per hour","note":"Editorial assumption. Replace with your own."}],"formula":"eddReviews * hoursPerReview * misratedShare * costPerHour","currency":"USD","period":"per year","resultLabel":"Enhanced due diligence effort redirected","caveat":"Assumes the released effort is redirected to genuinely risky customers rather than cut. It leaves out the value of catching risk between reviews and the cost of model validation."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The rating drives regulatory obligations for every customer, so it must be explainable, stable, validated and consistent with the bank's documented risk assessment. Data joining across KYC, transactions and screening is usually the biggest piece of work.","dataPrerequisites":["Customer due diligence attributes with data quality measures","Transaction behaviour aggregated per customer","Screening, alert and suspicious activity report history","The bank's enterprise wide money laundering risk assessment and rating methodology"],"integrations":["KYC and customer lifecycle management system","Transaction monitoring and case management","Screening engines","Core banking and product systems","Periodic review and enhanced due diligence workflow"]},"implementation":{"steps":[{"title":"Keep the regulatory floor explicit","detail":"Write down the ratings that rules dictate (for example PEPs and high risk jurisdictions) and keep them as hard constraints above the model."},{"title":"Choose explainability over marginal accuracy","detail":"Prefer a model whose drivers an analyst can read out to an examiner. A slightly better score that nobody can explain is not usable."},{"title":"Back test against outcomes","detail":"Check that higher bands contain more alerts, reports and exits than lower ones, on history, and compare with the current scorecard."},{"title":"Rerate in parallel","detail":"Rerate the whole book in parallel and review the largest moves with compliance before switching, including customers that would move down."},{"title":"Turn on event driven updates","detail":"Recalculate on material events and route moves into higher bands to enhanced due diligence with the drivers attached."}],"guardrails":["Regulatory minimum ratings enforced as hard rules above the model","Every rating shows its contributing factors in plain language","Moves to a lower band that reduce scrutiny are reviewed by a human","Bias testing so that nationality or other protected characteristics do not drive ratings beyond what the risk assessment justifies","Model inventory entry with validation and stability monitoring"],"humanInTheLoop":"The model rates and updates; analysts confirm moves into and out of high risk, and compliance owns the methodology, approves the model and reviews distribution and outcome reports. Decisions to restrict or exit a customer remain human decisions.","kpisToInstrument":["Distribution of customers across bands and month on month stability","Alerts, reports and exits per band (the rating should separate them)","Enhanced due diligence volume and time per review","Share of rating moves overturned by analysts","Time from a material event to the updated rating"],"failureModes":[{"title":"The black box rating","detail":"A rating the bank cannot explain is hard to defend to a supervisor and gives analysts nothing to investigate, however accurate it is. Use interpretable models or reliable attribution."},{"title":"Rating churn","detail":"Customers flip between bands with every transaction, creating work without insight. Smooth scores and use materiality thresholds."},{"title":"Proxy discrimination","detail":"The model leans on nationality or postcode beyond what the risk assessment supports. Test and constrain features."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"An AML customer risk rating is not listed in Annex III. Article 5(1)(d) prohibits AI risk assessments that predict whether a natural person will commit or will likely commit a criminal offence based solely on profiling of that person or on assessing their personality traits and characteristics; it exempts only AI that supports the human assessment of a person's involvement in a criminal activity, which is already based on objective and verifiable facts directly linked to a criminal activity. An AML customer risk rating built from due diligence attributes, transaction behaviour and screening results is itself an automated evaluation of a person's situation and behaviour, which is profiling under GDPR Article 4(4), and due diligence facts such as occupation, geography and products are not facts directly linked to a criminal activity, so the rating does not sit squarely inside the exemption. What keeps it a defensible AML due diligence tool rather than an offence prediction is that it does not itself accuse a person of an offence: it sets a level of scrutiny, a human analyst reviews material moves, and regulatory minimum rules sit above the model as hard constraints. A rating driven mainly by nationality or other personal attributes weakens that position further, which is why the proxy discrimination guardrail matters. If the same score is used to evaluate the creditworthiness of natural persons or to establish their credit score, that use falls under Annex III point 5(b) and is high risk, so keep the AML rating and credit decisions separate."},"regulations":["eu-ai-act","gdpr","fatf-recommendations","us-sr-11-7","mas-ai-risk-management","cbuae-ai-guidance","nist-ai-rmf","iso-42001","eu-amlr","us-bsa","mas-notice-626"],"guidance":[{"title":"Supporting Artificial Intelligence Adoption in AML/CFT","issuer":"Hong Kong Monetary Authority","region":"asia-pacific","url":"https://brdr.hkma.gov.hk/eng/doc-ldg/docId/getPdf/20251118-3-EN/20251118-3-EN.pdf","note":"Describes banks moving from rules to holistic, data driven AML approaches and the supervisor's support programme."},{"title":"Principles for Using Artificial Intelligence and Machine Learning in Financial Crime Compliance","issuer":"Wolfsberg Group","region":"global","url":"https://wolfsberg-group.org/resources/202/93","note":"Industry principles on accountability, openness and transparency that apply directly to a risk rating model."},{"title":"Notice 626 Prevention of Money Laundering and Countering the Financing of Terrorism, Banks","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/notices/notice-626","note":"Example of national rules on risk based customer due diligence and enhanced measures for higher risk customers."}],"controls":["Documented rating methodology linked to the enterprise wide risk assessment","Factor level explanation stored with every rating","Independent validation, stability monitoring and annual review of the model","Bias testing on protected characteristics","Human review of material moves in both directions"],"incidents":[{"title":"De Nederlandsche Bank fines bunq for insufficient customer due diligence","url":"https://www.dnb.nl/en/general-news/enforcement-measures-2025/fine-for-bunq-b-v-for-insufficient-customer-due-diligence/","note":"On 6 May 2025 the Dutch central bank fined bunq EUR 2.6 million because it did not sufficiently follow up signals and transaction monitoring alerts in four customer files it had itself identified as high risk (period January 2021 to May 2022); bunq has appealed. The notice does not blame bunq's data driven approach, but it shows that a rating is only as good as the human follow up it triggers."}]},"blitsAi":{"howToBuild":"The rating model belongs in the bank's analytics platform. Blits.ai handles the work it\ntriggers: when a rating moves into a higher band, an **agentic workflow**, started through the\nAPI, collects the drivers and the customer file through **custom functions** and **SQL knowledge\nbases**, drafts the enhanced due diligence request and the questions to ask, and routes it to an\nanalyst through **human in the loop approval**.\n\nWhere the bank needs information from the customer, a **conversational agent** in the bank's\nown app through the **REST API channel**, by **email** or on **WhatsApp** can collect source of funds or updated details and documents, with\n**PII masking**, **guardrails** and a **human handover** for anything sensitive. Runs keep a\nfull audit trail, **test suites** cover the drafting, and the platform is model agnostic with EU\nand UAE data residency."},"faq":[{"question":"What is a dynamic customer risk rating?","answer":"A money laundering risk rating that updates when the customer's data or behaviour changes, instead of only at a scheduled review, and shows which factors drive it."},{"question":"Does a machine learning risk rating have to be explainable?","answer":"In practice yes. The Wolfsberg Group's principles for AI and machine learning in financial crime compliance ask that results \"can be adequately explained or proven given the data inputs\", and analysts need the drivers to run enhanced due diligence. Choose interpretable models or reliable attribution over a small gain in accuracy."},{"question":"Is the risk rating high risk under the EU AI Act?","answer":"Not as an AML tool: AML risk rating is not listed in Annex III. It becomes high risk if the same score is used to evaluate the creditworthiness of natural persons or to set their credit score (Annex III point 5(b)), so keep the uses separate and documented. Article 5(1)(d) bans predicting that a person will commit an offence from profiling or personality traits alone; it is designed as a due diligence tool, not an offence prediction, so do not let it drive a rating from nationality or other personal traits alone, keep the regulatory minimum rules and analyst review in place, and treat the rating itself as a level of scrutiny rather than an accusation."},{"question":"How is this different from a machine learning transaction monitoring score?","answer":"A monitoring score such as HSBC's Dynamic Risk Assessment, which Google Cloud describes as a customer risk score offered as an alternative to rules based transaction alerting, decides which customers investigators look at for suspicious activity (see AML alert triage). The rating on this page decides the level of due diligence each customer gets. The two share data and methods, so HSBC's deployment is related evidence here rather than a direct example."}],"related":["perpetual-kyc","aml-alert-triage","pep-and-adverse-media-screening","business-onboarding-and-ubo-discovery","source-of-wealth-diligence"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added EU Anti Money Laundering Regulation, Bank Secrecy Act, MAS Notice 626 to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: attributed the explainability expectation to the Wolfsberg principles; tightened the EU AI Act basis to creditworthiness under Annex III point 5(b); updated the bunq incident with the four files and bunq's appeal; added the CBb ruling to the bunq evidence; named the REST API channel in the Blits.ai section; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: dated the Fenergo statistic to the cited page (2026) and attributed the review frequency claim to Fenergo's own wording; attributed the explainability expectation in the FAQ to the Wolfsberg principles and softened the black box failure mode; unlinked HSBC and UOB from this use case and dropped the false positive reduction KPI, because neither deployment reports a customer risk rating result; rewrote the meta description so it no longer presents bunq or HSBC as deployments; corrected the EU AI Act basis to quote Article 5(1)(d) in full and to explain the rating on that basis, and updated the matching FAQ answer."}],"slug":"dynamic-customer-risk-rating","url":"https://www.blits.ai/ai-use-cases/dynamic-customer-risk-rating","benchmarks":[],"indicativeValueResult":{"low":120000,"high":1312500},"evidence":["bunq-machine-learning-customer-due-diligence"]},{"title":"Generative AI copilot for internal audit","shortTitle":"Internal audit copilot","seoTitle":"Generative AI copilot for internal audit teams","metaDescription":"An internal audit copilot drafts plans, evidence summaries and findings. Microsoft cites 55% less reporting time at Bradesco and 30% less report writing at BCI.","definition":"A copilot for internal auditors that drafts planning memos and document request lists from prior audits, summarises large evidence sets, builds risk and control matrices from policies and process documents, and drafts findings and reports, with every statement traceable to its evidence and a qualified auditor accountable for every conclusion.","aliases":["AI for internal audit","audit copilot","generative AI audit assistant","AI audit workpapers"],"industries":["cross-industry","banking","insurance","government","capital-markets","wealth-and-asset-management"],"functions":["risk-management","regulatory-compliance","finance-and-accounting"],"patterns":["rag-knowledge-assistant","summarization","content-generation","document-processing","anomaly-detection"],"channels":["internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"copilot","adoptionStage":"early-adopters","problem":"Internal audit functions are asked to cover a widening risk universe (cyber, third parties, AI,\nconduct, operational resilience), and every new area competes for the same auditor hours. A large share\nof each engagement is reading and writing: going through prior workpapers, policies and process\ndocuments to plan the scope, summarising hundreds of pages of evidence, documenting walkthroughs,\nand drafting findings and reports that go through several rounds of review.\n\nGenerative AI fits that reading and writing work, and audit leaders are adopting it quickly. The\nquestion for a third line function is how to use it without weakening what makes audit valuable:\nindependence, evidence that stands up to challenge and a qualified auditor who owns every\nconclusion. A fluent sentence in a workpaper that no evidence supports is worse than no sentence.","problemStats":[{"statement":"Gartner reported in March 2024 that 41% of chief audit executives were using or planning to use generative AI that year; in its survey of 112 chief audit executives in mid 2023, 12% were already using it.","sourceTitle":"Gartner Survey Shows 41% of Internal Audit Teams Use or Plan to Use Generative AI this Year","sourceUrl":"https://web.archive.org/web/20240523072602/https://www.gartner.com/en/newsroom/press-releases/2024-03-11-gartner-survey-shows-41-percent-of-internal-audit-teams-use-or-plan-to-use-generative-ai-this-year","year":2024},{"statement":"Gartner reported in August 2026 that 93% of audit leaders report some level of AI use but only 38% have an AI strategy; in its webinar poll of 743 audit professionals, 60% used AI to draft audit issues, ratings or reports.","sourceTitle":"Gartner Survey Finds Audit Teams' AI Use is Common, But Most Teams are Lacking Strategic Adoption and Application","sourceUrl":"https://web.archive.org/web/20260820091915/https://www.gartner.com/en/newsroom/press-releases/2026-08-10-gartner-survey-finds-audit-teams-ai-use-is-common-but-most","year":2026}],"howItWorks":"1. **Plan from what is known.** The copilot retrieves prior workpapers, previous findings and their\n   status, the risk assessment and relevant policy changes, and drafts the planning memo, scope and\n   document request list for the audit lead to edit.\n2. **Build the control picture.** From policies, procedures and process documents it drafts a risk\n   and control matrix and walkthrough narratives, citing the document behind each control.\n3. **Digest the evidence.** It summarises evidence files, meeting notes and management responses,\n   and answers the auditor's questions about them with page level citations.\n4. **Point at anomalies.** Analytics on full populations (payments, access logs, journal entries)\n   surface outliers and exceptions for the auditor to test; the AI explains the pattern, it does\n   not conclude on it.\n5. **Draft findings and reports.** It drafts findings in the house structure (condition, criteria,\n   cause, effect, recommendation) from the auditor's notes and tested evidence, and checks draft\n   reports for consistency and tone.\n6. **Keep the trail.** Every AI draft, the prompt context and the auditor's edits are stored with\n   the workpaper, so reviewers can see what the AI wrote and what the auditor concluded.","valueDrivers":["employee-productivity","speed","compliance","risk-reduction"],"kpis":["productivity-gain","processing-time-reduction","handling-time-reduction","hours-saved","time-saved-per-task"],"indicativeValue":{"referenceOrg":"A bank internal audit function with 60 auditors","inputs":[{"key":"auditors","label":"Auditors using the copilot","low":60,"high":60,"unit":"auditors","note":"The reference organization."},{"key":"hoursPerYear","label":"Productive hours per auditor per year","low":1500,"high":1600,"unit":"hours per auditor per year","note":"Editorial assumption after leave, training and administration."},{"key":"draftingShare","label":"Share of auditor time spent on planning documents, evidence summaries, workpaper write ups and reports","low":0.25,"high":0.4,"unit":"fraction of time","note":"Editorial assumption. Replace with your own time recording."},{"key":"timeSaved","label":"Share of that drafting and summarising time saved","low":0.15,"high":0.3,"unit":"fraction of time","note":"Conservative against the evidence on this page (Microsoft reports 30% less time writing internal audit reports at BCI and 55% less time in reporting at Bradesco), because those are vendor reported best cases."},{"key":"hourlyCost","label":"Fully loaded auditor hour","low":70,"high":120,"unit":"USD per hour","note":"Editorial assumption. Replace with your own rate."}],"formula":"auditors * hoursPerYear * draftingShare * timeSaved * hourlyCost","currency":"USD","period":"per year","resultLabel":"Auditor capacity released for testing and additional coverage","caveat":"Capacity only, which most functions reinvest in coverage rather than headcount. It leaves out the value of broader risk coverage and full population testing, and the cost of reviewing AI drafts and maintaining the workpaper corpus."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Drafting and summarising on a secure platform is quick to start. The harder parts are giving the copilot permission aware access to workpapers and evidence without breaching confidentiality between engagements, and agreeing the methodology changes for documenting AI assistance.","dataPrerequisites":["Prior workpapers, findings and reports in the audit management system","The audit methodology, templates and finding structure","Policies, procedures and process documents for the audited areas","Population data (transactions, logs, journals) for analytics where testing needs it"],"integrations":["Audit management system (workpapers, issues, action tracking)","Document management and policy repositories","Data platform for population analytics","Microsoft Teams or the audit team's collaboration tool"]},"implementation":{"steps":[{"title":"Update the methodology first","detail":"Decide where AI may assist (planning, summarising, drafting) and where it may not (forming opinions, rating findings), and how AI assistance is recorded in workpapers. Brief the audit committee and the external auditor."},{"title":"Start with low risk drafting","detail":"Planning memos, document request lists and summaries of prior audits give quick wins with little risk, because the auditor edits every output before it is used."},{"title":"Ground everything in the audit corpus","detail":"Connect prior workpapers, the methodology and policies with retrieval and require a citation for every statement, so reviewers can check it quickly."},{"title":"Respect engagement boundaries","detail":"Apply the audit management system's permissions, so the copilot only retrieves from engagements the auditor may see, especially for investigations and sensitive reviews."},{"title":"Add analytics and findings drafting","detail":"Move to anomaly detection on full populations and drafting of findings once reviewers trust the citations, and measure review time and rework per report."},{"title":"Audit the copilot","detail":"Treat the copilot as an AI system in the inventory, with testing of its outputs on past audits and periodic review by someone independent of the team that built it."}],"guardrails":["A qualified auditor reviews, edits and signs every workpaper, finding and report; the AI never forms an audit opinion or rates a finding","Every AI generated statement cites the evidence document and page it relies on","Retrieval respects engagement and document permissions, including restrictions on investigations","AI assistance is recorded in the workpaper, with the draft and the auditor's changes retained","Anomalies surfaced by analytics are leads for testing, never conclusions","Audit data stays within the approved tenancy and region and is not used to train external models"],"humanInTheLoop":"Auditors own every conclusion and the chief audit executive owns the methodology. Reviewers check AI assisted workpapers with the same rigour as any other, using the citations, and the audit committee is informed how AI is used in the function.","kpisToInstrument":["Hours per engagement phase (planning, fieldwork documentation, reporting), before and after","Time from fieldwork end to final report","Share of AI drafted statements changed or removed by the auditor or reviewer","Review notes raised on AI assisted workpapers compared with others","Audit plan coverage (auditable entities covered per year)"],"failureModes":[{"title":"Unsupported statements in workpapers","detail":"A fluent summary includes a claim that no evidence supports and survives review. Require citations and train reviewers to check them."},{"title":"Confidentiality leaks across engagements","detail":"The copilot retrieves material from an investigation or another engagement the auditor may not see. Enforce document level permissions in retrieval."},{"title":"Independence eroded by shared tooling","detail":"Audit relies on the same AI tools and prompts as the first line it audits. Keep audit's own configuration and test it independently."},{"title":"Anchoring on the AI draft","detail":"Auditors accept the drafted risk and control matrix rather than challenging it. Have auditors edit, not approve, and track the change rate."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"An internal drafting and analysis assistant for auditors that makes no decisions about natural persons. It would need reassessment if used to evaluate individual employees' behaviour or performance, which falls under Annex III point 4(b)."},"regulations":["eu-ai-act","gdpr","iso-42001","nist-ai-rmf","dora"],"guidance":[{"title":"2024 Global Internal Audit Standards","issuer":"The Institute of Internal Auditors","region":"global","url":"https://www.theiia.org/en/standards/2024-standards/global-internal-audit-standards/","note":"A mandatory component of the IPPF, and The IIA expects every internal audit function to conform. Its principles (among them objectivity, due professional care and confidentiality) and the Domain V requirements for planning engagements and developing findings apply to AI assisted work as to any other."}],"controls":["Methodology section on AI use approved by the chief audit executive and shared with the audit committee","Workpaper field recording AI assistance, with the draft retained","Periodic independent testing of the copilot's outputs on past engagements","Access control aligned with engagement permissions in the audit management system","Inventory entry for the copilot with an accountable owner"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** with a **knowledge base** holding prior workpapers, the audit\nmethodology, templates and policies, retrieved with **hybrid search**. Evidence files in PDF, Word, Excel and Outlook email formats are ingested\nper engagement, and **SQL knowledge bases** let the agent query extracted population data for\nanomaly analysis. **Structured output** produces findings in the house\nformat (condition, criteria, cause, effect, recommendation).\n\nAuditors use the agent in **Microsoft Teams** or a web chat. Fine grained **role based access\ncontrol**, with custom roles per tenant and per bot roles, lets investigations and sensitive\nreviews run on their own bot, agent and knowledge base, and **tenant isolation** keeps audit data apart from other tenants. **PII masking** keeps personal\ndata out of prompts, **conversation logs** and **execution tracing** retain what the AI produced\nfor the workpaper, and **test suites** check the agent against past engagements before every\nchange. The platform is model agnostic, so the audit function can choose its own model\nindependently of the first line."},"faq":[{"question":"How much time can generative AI save in internal audit?","answer":"Reported figures are early and vendor published, without a stated method. Microsoft reports that BCI reduced time spent writing internal audit reports by 30%, that Bradesco achieved 55% less time in reporting with its AILA audit assistant, and that XP Inc. increased audit team efficiency by 30%. Measure hours per engagement phase on your own audits before and after."},{"question":"Does using AI compromise auditor independence?","answer":"Not if the auditor owns every conclusion, the AI's contribution is recorded in the workpaper, and audit's tooling is configured and tested independently of the first line it audits. The Global Internal Audit Standards, including objectivity and due professional care, apply to AI assisted work as to any other."},{"question":"What should the AI never do in an audit?","answer":"Form an audit opinion, rate a finding, or treat an anomaly as a conclusion. It drafts, summarises and points at exceptions; qualified auditors test, judge and sign."}],"related":["continuous-controls-testing","ai-model-inventory","policy-drafting-and-gap-analysis","supervisory-exam-response-assembly","governed-text-to-sql-analytics"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog, made industry neutral and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: industries now include wealth and asset management, where its evidence comes from; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: made the Gartner survey bases precise, corrected the IIA standards note and the Blits.ai capability wording, and added an SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: removed a grant audit follow up record as out of scope, aligned the Blits.ai access control wording with the feature inventory, softened the IIA conformance note, tightened the Bradesco, BCI and XP Inc. records to what the Microsoft source states, and removed an unsupported source citation claim from the Blits.ai build description."}],"slug":"internal-audit-copilot","url":"https://www.blits.ai/ai-use-cases/internal-audit-copilot","benchmarks":[{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":42.5,"min":30,"max":55,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"banco-bradesco-aila-audit-assistant","pooled":true},{"id":"bci-copilot-internal-audit-reports","pooled":true}]},{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":47.5,"min":30,"max":65,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"banco-bradesco-aila-audit-assistant","pooled":true},{"id":"xp-inc-copilot-audit-team","pooled":true}]}],"indicativeValueResult":{"low":236250,"high":1382400},"evidence":["banco-bradesco-aila-audit-assistant","bci-copilot-internal-audit-reports","xp-inc-copilot-audit-team"]},{"title":"Generative AI voice assistant in the car","shortTitle":"In car voice assistant","seoTitle":"In car AI voice assistant for car makers","metaDescription":"LLM voice assistants let drivers talk to the car in plain language. Volkswagen added ChatGPT to IDA and Mercedes-Benz offered its update to over three million cars.","definition":"A voice assistant built into the vehicle that uses a large language model to hold a natural conversation with the driver and passengers, controls comfort, navigation and media functions, answers questions about the car and the world, and keeps the vehicle's own commands and data under the car maker's control.","aliases":["in car voice assistant","connected car AI assistant","generative AI car assistant","LLM voice assistant for vehicles","automotive virtual assistant"],"industries":["automotive"],"functions":["customer-service","product-and-pricing"],"patterns":["voice-agent","conversational-agent","rag-knowledge-assistant","agentic-workflow"],"channels":["voice"],"audience":"customer-facing","autonomy":"autonomous","adoptionStage":"early-adopters","segment":"connected-car","problem":"Modern cars have many functions behind touchscreens and menus, and drivers should keep\ntheir eyes on the road. First generation voice control only understood fixed commands in a fixed\norder: say it slightly differently and it failed, which leaves drivers with menus, the manual or\na question to the dealer.\n\nLarge language models change what the assistant can understand, but a car is not a phone. The\nassistant has to work with patchy connectivity, answer in a split second, never distract, keep\nvehicle and location data private, and separate harmless requests (a warmer seat, a restaurant\nnearby) from anything that touches driving. Car makers also want to own the experience and the\nbrand voice rather than hand the cabin to a third party assistant.","problemStats":[],"howItWorks":"1. **Wake and listen.** The driver says the wake word or presses the steering wheel button; speech\n   recognition runs on board for commands and in the cloud for open questions.\n2. **Decide who answers.** Vehicle commands (climate, seats, media, navigation) stay in the car\n   maker's own system. Only questions it cannot answer go to a language model, anonymized and\n   without vehicle data, as Volkswagen describes for IDA.\n3. **Ground the answer.** Knowledge questions use a web search or a maps platform, and questions\n   about the car use the owner's manual and vehicle status, so answers are current and specific.\n4. **Keep the conversation.** The assistant remembers the dialogue for a limited time, so follow up\n   questions work without repeating the context.\n5. **Answer in the brand voice and act.** The reply is spoken in the car maker's voice and, where\n   allowed, the assistant sets the function or starts navigation, then confirms what it did.\n6. **Learn safely.** Voice data is stored anonymized, answers are screened for harmful content,\n   and usage patterns can suggest routines the driver can accept or ignore.","valueDrivers":["customer-experience","revenue-growth","inclusion-and-access","cost-to-serve"],"kpis":["users-served","interactions-handled","customer-satisfaction","contact-deflection","nps-change"],"indicativeValue":{"referenceOrg":"A car maker with 1 million connected vehicles on the road","inputs":[{"key":"vehicles","label":"Connected vehicles with the assistant","low":1000000,"high":1000000,"unit":"vehicles","note":"The reference car maker."},{"key":"featureContacts","label":"Contacts to brand customer service or the dealer about how a vehicle function works","low":0.1,"high":0.3,"unit":"contacts per vehicle per year","note":"Editorial assumption, replace with your own contact reason data."},{"key":"deflection","label":"Share of those contacts the assistant makes unnecessary","low":0.1,"high":0.3,"unit":"fraction of feature contacts","note":"Editorial assumption. None of the car makers on this page discloses usage or deflection figures."},{"key":"costPerContact","label":"Cost of a handled contact","low":5,"high":10,"unit":"USD per contact","note":"Editorial assumption for a blended phone and chat contact, replace with your own cost."}],"formula":"vehicles * featureContacts * deflection * costPerContact","currency":"USD","period":"per year","resultLabel":"Customer service cost avoided on feature questions","caveat":"Covers service cost only. It leaves out the main reasons car makers build this (product differentiation, connected services revenue and brand loyalty), the cost of cloud models and speech processing per vehicle, and the integration and validation work in the vehicle."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"Hybrid on board and cloud processing, low latency speech in a noisy cabin, many languages and accents, strict separation from driving functions, over the air updates and privacy by design across millions of vehicles make this a platform program rather than a feature.","dataPrerequisites":["A catalog of vehicle commands and the functions the assistant may set, per model and market","Owner's manuals and feature descriptions in every supported language","Vehicle status signals the assistant may read, with a privacy classification per signal","Anonymized samples of real utterances to test recognition and answers"],"integrations":["Head unit and vehicle operating system for commands and status","Connected car cloud and over the air update pipeline","Speech recognition and synthesis with a brand voice","Language model and web search or maps platform for open questions","Companion app and customer account for consent and settings"]},"implementation":{"steps":[{"title":"Draw the line between commands and conversation","detail":"List which requests the vehicle system keeps, which go to the language model and which are refused, and never let the model reach functions that affect driving."},{"title":"Design privacy in from the start","detail":"Decide what leaves the car, how it is anonymized, how long dialogue memory lasts, how the driver switches the online assistant off and how consent is recorded in the account."},{"title":"Ground answers about the car","detail":"Load owner's manuals and feature descriptions per model and market into retrieval, so the assistant explains the driver's own car rather than a generic one."},{"title":"Test in the cabin, not the lab","detail":"Test with road noise, accents, passengers talking and weak connectivity, and measure recognition, latency and task success per language before launch."},{"title":"Launch per language and market, then update over the air","detail":"Start with a few languages in new vehicles, extend to cars on the road by software update and add functions such as routines only after they pass the same tests."}],"guardrails":["No access for the language model to driving, safety or security functions","Only anonymized text leaves the vehicle, without vehicle identifiers or location unless the driver asked for a place","Content screening of answers for harmful, illegal or distracting content","A clear way to switch the online assistant off, and short retention of dialogue history","Spoken answers kept short to limit driver distraction"],"humanInTheLoop":"There is no human in the conversation, so the human control sits in design and operations: product and safety teams approve every new function the assistant may set, reviewers sample anonymized dialogues for wrong or unsafe answers, and customer service handles complaints and feedback about the assistant.","kpisToInstrument":["Share of active vehicles that use the assistant each month","Task success rate per request type and language","Latency from end of speech to start of the answer","Share of answers flagged as wrong or unsafe in reviews","Customer service contacts about vehicle functions per 1,000 vehicles"],"failureModes":[{"title":"Confident answers about the wrong car","detail":"The model explains a feature the driver's model or trim does not have. Ground answers in the manual for that vehicle and say when a feature is not available."},{"title":"Latency that drivers will not accept","detail":"A cloud round trip that takes several seconds makes people give up. Keep commands on board and stream the answer."},{"title":"Privacy surprises","detail":"Drivers discover that voice or location data left the car without them knowing. Make data flows, retention and the off switch visible."},{"title":"Distraction by design","detail":"Long, chatty answers pull attention from the road. Limit answer length and avoid anything that invites the driver to look at a screen."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1): people must be informed that they are interacting with an AI system unless that is obvious from the context. Article 50(2): synthetic audio output must be marked as artificially generated. A cabin assistant for comfort, media, navigation and knowledge questions is not an Annex III use. It would move towards the high risk regime if it became a safety component of the vehicle: vehicle type approval legislation is listed in Annex I Section B, and under Article 2(2) the high risk requirements reach those products only through the amendments the AI Act makes to that legislation. Keep driving and safety functions out of its reach."},"regulations":["eu-ai-act","gdpr","nist-ai-rmf","iso-42001"],"guidance":[{"title":"Guidelines 01/2020 on processing personal data in the context of connected vehicles and mobility related applications","issuer":"European Data Protection Board","region":"europe","url":"https://www.edpb.europa.eu/our-work-tools/our-documents/guidelines/guidelines-012020-processing-personal-data-context_en","note":"Sets out how GDPR applies to data processed in and sent from connected vehicles, including consent, minimization and local processing."},{"title":"Visual manual NHTSA driver distraction guidelines for in vehicle electronic devices","issuer":"US National Highway Traffic Safety Administration","region":"north-america","url":"https://www.federalregister.gov/documents/2013/04/26/2013-09883/visual-manual-nhtsa-driver-distraction-guidelines-for-in-vehicle-electronic-devices","note":"US guidelines to limit the distraction caused by in vehicle devices, relevant to how much the assistant shows on screen and asks of the driver by hand. Auditory vocal interaction itself is outside their scope."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Requires disclosure that people are interacting with AI and marking of synthetic audio."}],"controls":["Documented allow list of vehicle functions the assistant may control, approved by product safety","Data protection impact assessment covering voice, location and vehicle data","Anonymization and retention rules enforced in the cloud pipeline","Regression test suites per language and model before every over the air release","Monitoring of harmful or wrong answers with a fast rollback path"],"incidents":[]},"blitsAi":{"howToBuild":"Commands that run in the head unit stay in the car maker's vehicle software. Blits.ai fits the\ncloud side: an **AI agent** reached through the REST or WebSocket API channel from the vehicle\ncloud or the companion app, with a **knowledge base** of owner's manuals and feature descriptions\nretrieved with hybrid search, and **custom functions** that call the car maker's connected car APIs\nfor allowed actions such as preconditioning or finding a charging point.\n\nSpeech recognition and synthesis stay on the car maker's side, in the head unit or its own\nspeech stack, which exchanges text with the agent. **Multi language** support covers several\nmarkets from one agent, **guardrails** screen input and output, and **PII masking** removes personal data before text reaches a model. **Test suites**\nrun multi turn conversations per language on every change, and the platform is model agnostic,\nwith EU and UAE data residency, so the car maker can choose or switch the model per market."},"faq":[{"question":"Which car makers use a generative AI voice assistant?","answer":"Volkswagen added ChatGPT to its IDA voice assistant in 2024 through Cerence, for new ID. models, the Golf, Tiguan and Passat. Mercedes-Benz made a ChatGPT and Bing based knowledge feature in the MBUX Voice Assistant available to over three million vehicles in December 2024, and BMW introduced the German language version of its Alexa+ based Intelligent Personal Assistant in the iX3, from mid April 2026 production."},{"question":"How do car makers keep voice data private?","answer":"Volkswagen forwards only questions its own system cannot answer, anonymously, gives ChatGPT no vehicle data and deletes questions and answers immediately. Mercedes-Benz stores voice data anonymized in its own cloud. The EDPB guidelines on connected vehicles set the GDPR baseline."},{"question":"Is an in car AI assistant high risk under the EU AI Act?","answer":"Usually not: it carries the Article 50 transparency duties. It would move towards the high risk regime if it became a safety component of the vehicle, so keep it away from driving functions."}],"related":[],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by the fact check workflow (independent skeptic review)."},{"date":"2026-09-27","note":"First version, researched for the automotive scope with Volkswagen, Mercedes-Benz and BMW evidence checked against the sources. Editor pass corrected the Article 50 paragraphs, the BMW rollout status and languages, and the Blits.ai speech description."}],"slug":"in-car-ai-voice-assistant","url":"https://www.blits.ai/ai-use-cases/in-car-ai-voice-assistant","benchmarks":[],"indicativeValueResult":{"low":50000,"high":900000},"evidence":["bmw-intelligent-personal-assistant-alexa-plus","mercedes-benz-mbux-generative-ai-voice-assistant","volkswagen-ida-voice-assistant-chatgpt"]},{"title":"Governed text to SQL analytics assistant","shortTitle":"Governed SQL analytics","seoTitle":"Text to SQL assistant for governed data analytics","metaDescription":"Governed text to SQL lets staff query data in plain language under their own permissions. See how LinkedIn and Uber built it, with accuracy data and controls.","definition":"An assistant that turns a business user's plain language question into a query against governed data, runs it under that user's own data permissions and returns the table or chart together with the SQL and the tables used, so routine ad hoc questions no longer queue for the data team.","aliases":["text to SQL","natural language to SQL","conversational BI","chat with your data","self service analytics copilot"],"industries":["cross-industry","banking","insurance","retail-and-ecommerce","technology","pharma-and-life-sciences"],"functions":["analytics-and-reporting","it-and-engineering"],"patterns":["conversational-agent","code-generation","rag-knowledge-assistant"],"channels":["internal-tools","microsoft-teams"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"early-adopters","problem":"Most data questions in a business are small and urgent: how many applications came in from this\nchannel last week, what is the arrears rate by region, which branches missed target. Dashboards\nanswer the questions someone anticipated; everything else becomes a ticket for an analyst who\nknows which of thousands of tables holds the answer and how to join them. The queue slows\ndecisions and consumes analysts on work that is repetitive rather than analytical.\n\nGeneric text to SQL is not the answer on its own. On a real warehouse it picks the wrong table,\nmisreads a column or applies the wrong business definition, and returns a confident, wrong number.\nIn a bank or insurer the second risk is access: a query tool must never let a user see rows or\ncolumns their role does not permit. The job is therefore governed text to SQL: a curated semantic\nlayer, the user's own permissions, and the SQL always visible.","problemStats":[{"statement":"Snowflake, citing a Forrester report, says anecdotal evidence shows a best case rate of 70% for generating accurate, executable code on simple single table queries and around 20% at worst for queries with multiple tables or complex joins.","sourceTitle":"Cortex Analyst: Paving the Way to Self-Service Analytics with AI","sourceUrl":"https://www.snowflake.com/en/blog/cortex-analyst-ai-self-service-analytics/","year":2024}],"howItWorks":"1. **Understand the question.** The assistant restates the question in business terms and asks\n   for what is missing (\"which period?\", \"gross or net?\") instead of guessing.\n2. **Find the right data.** It retrieves from a curated semantic layer: certified tables, metric\n   definitions, join paths and example queries for the business domain. Uber narrows the search\n   with domain \"workspaces\"; LinkedIn had domain experts certify and describe key tables, and\n   draws example queries from notebooks that users have certified.\n3. **Write the query.** A model generates SQL against those definitions only, and the query is\n   validated (syntax, allowed tables, row limits) before it runs.\n4. **Run it as the user.** The query executes with the user's own credentials, so row and column\n   level security in the data platform decides what comes back.\n5. **Show the work.** The answer comes with the SQL, the tables and the definitions used, plus a\n   short explanation, so the user or an analyst can check it.\n6. **Learn from corrections.** Queries that analysts correct or certify become new examples in\n   the semantic layer.","valueDrivers":["speed","employee-productivity","cost-to-serve"],"kpis":["accuracy","users-served","time-saved-per-task","employee-adoption","productivity-gain"],"indicativeValue":{"referenceOrg":"A bank with 1,500 regular data consumers and a central data and BI team","inputs":[{"key":"requests","label":"Ad hoc data requests to the data team per year","low":6000,"high":12000,"unit":"requests per year","note":"Editorial assumption, about four to eight requests per data consumer per year. Replace with your own ticket volume."},{"key":"selfServeShare","label":"Share of requests the assistant answers without an analyst","low":0.15,"high":0.35,"unit":"fraction of requests","note":"Editorial assumption, deliberately conservative because accuracy falls on complex multi table questions (see the Forrester figure cited on this page)."},{"key":"analystHours","label":"Analyst hours per ad hoc request","low":1.5,"high":3,"unit":"hours per request","note":"Editorial assumption including clarification, query writing and checking."},{"key":"hourlyCost","label":"Fully loaded analyst hour","low":60,"high":100,"unit":"USD per hour","note":"Editorial assumption. Replace with your own rate."}],"formula":"requests * selfServeShare * analystHours * hourlyCost","currency":"USD","period":"per year","resultLabel":"Analyst time released from routine ad hoc requests","caveat":"Analyst time only. It leaves out the value of faster decisions for the business users, the cost of building and maintaining the semantic layer, and the cost of any wrong answers that are not caught, which is why the self serve share is kept low."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"The model is the easy part. The work is the semantic layer (certified tables, metric definitions and example queries), enforcing each user's data permissions, and an evaluation set of real questions with known answers.","dataPrerequisites":["A curated set of certified tables or a semantic model with business definitions","Row and column level security defined in the data platform per role","A library of example questions with correct, reviewed SQL","Data classification, so sensitive columns can be excluded or masked"],"integrations":["Data warehouse or lakehouse (SQL endpoint)","Semantic layer or data catalog","Identity provider for passing the user's identity to the data platform","BI tool or chat front end such as Microsoft Teams"]},"implementation":{"steps":[{"title":"Pick one domain with a clean model","detail":"Start where definitions are settled and tables are few, for example sales pipeline or contact centre volumes. Write down the metric definitions before any model sees a question."},{"title":"Build the evaluation set first","detail":"Collect 100 to 200 real questions from the request queue with correct SQL and results. Run every change of model, prompt or semantic layer against it and publish the pass rate."},{"title":"Enforce permissions in the data platform, not the prompt","detail":"Run each query with the user's own identity so row and column security applies. Never rely on instructions to the model to hide data."},{"title":"Always show the SQL and the definitions","detail":"Users and analysts must be able to see how a number was produced. Label answers built on uncertified tables clearly."},{"title":"Close the loop with analysts","detail":"Route questions the assistant cannot answer confidently to the data team, and turn their corrected queries into new certified examples."},{"title":"Widen by domain, not by table count","detail":"Add the next domain only when the first meets its accuracy target, and track accuracy per domain separately."}],"guardrails":["Queries run with the user's own credentials; row and column level security is enforced by the data platform","Read only access, an allow list of schemas and a row limit on every query","The generated SQL, tables and definitions are shown with every answer","The assistant asks for clarification or declines when confidence is low instead of guessing","Every question, query and result is logged for audit and evaluation","Sensitive columns (personal data, account numbers) excluded or masked unless the role needs them"],"humanInTheLoop":"Analysts own the semantic layer and certify example queries. Any number that goes into a board pack, regulatory report or customer communication is checked by an analyst, not taken straight from the assistant. Users can flag a wrong answer, which goes to the data team for review.","kpisToInstrument":["Execution accuracy on the evaluation set, per domain","Share of questions answered without analyst involvement, and share later flagged as wrong","Weekly active users and repeat usage","Time from question to answer compared with the request queue","Denied or blocked queries by reason (permission, schema, row limit)"],"failureModes":[{"title":"Confident but wrong numbers","detail":"The query runs and returns a plausible figure from the wrong table or definition. Show the SQL, restrict to certified tables, and measure accuracy on real questions."},{"title":"Permission leaks through a service account","detail":"The assistant queries with a powerful technical account and bypasses row level security. Pass the user's identity to the data platform."},{"title":"Definitions drift","detail":"The business changes a metric definition but the semantic layer does not. Give each definition an owner and a review date."},{"title":"Adoption stalls on trust","detail":"One visible wrong answer and users return to the queue. Start with a narrow domain, label uncertainty and publish the accuracy number."}]},"risk":{"euAiAct":{"tier":"limited","basis":"Article 50(1) requires providers to design AI systems that interact directly with people so that those people are informed they are dealing with AI, unless this is obvious from the context, as it usually is for an internal assistant. An analytics assistant that makes no decisions about people is not a prohibited practice under Article 5 and is not listed in Annex III. It would be high risk only if it were intended for an Annex III purpose, such as assessing the creditworthiness of natural persons (point 5(b))."},"regulations":["gdpr","eu-ai-act","dora","iso-42001","nist-ai-rmf"],"guidance":[{"title":"General Data Protection Regulation, Article 25 data protection by design and by default","issuer":"European Union","region":"europe","url":"https://eur-lex.europa.eu/eli/reg/2016/679/oj","note":"Access to personal data must be limited to what each purpose needs, which argues for running every generated query under the user's own permissions and masking personal data by default."},{"title":"Article 50, transparency obligations for providers and deployers of certain AI systems","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/50/","note":"Providers must design AI systems that interact directly with people so that those people are informed they are interacting with AI, unless this is obvious from the context. Applies from 2 August 2026."}],"controls":["Data access enforced by the data platform per user, with no shared privileged service account","Query and answer logs retained and reviewable by the data owner","Evaluation set pass rate recorded for every release of model, prompt or semantic layer","Owner and review date for every certified table and metric definition","Inventory entry for the assistant with an accountable owner"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai this is an **AI agent** connected to a **SQL knowledge base**: a curated reporting\ntable, loaded into the platform's own storage, that holds only data every user of that\nassistant may see. The agent queries it with structured filters, sorting, column selection and\na row limit, not raw SQL. A **knowledge base** holds the metric definitions, table descriptions\nand certified example queries, retrieved with **hybrid search** so the agent applies the\napproved filters and definitions to the right table. For questions that need joins or\naggregation across tables, a **custom function** calls a governed endpoint of the data\nplatform (for example a text to SQL or semantic layer API) with a scoped, read only credential;\nthe platform does not pass each user's own identity through, so that endpoint must also expose\nonly data every user of the assistant may see, and data that differs by role belongs in a\nseparate assistant per audience. **Structured output** returns the tables used and the result\ntogether, plus the generated SQL for questions answered through the custom function.\n\nBusiness users ask questions in **Microsoft Teams** or a web chat embedded in the intranet, and\n**role based access control** in the tenant console decides who may build, change and publish\nthe assistant. **Guardrails** decline out of\nscope requests, **PII masking** removes personal data that users type into their questions; the\nloaded table itself must exclude personal data columns, because query results are not masked.\n**Test suites** run the\nevaluation set of real questions on every change, and **analytics** with thumbs up and down\nfeedback show which answers users reject. The platform is model agnostic, so the data team can\npick the model that scores best on its own evaluation set."},"faq":[{"question":"How accurate is text to SQL on real company data?","answer":"It depends heavily on the question and the data model. Snowflake, citing anecdotal evidence in a Forrester report, gives a best case of 70% on simple single table queries and around 20% at worst on complex joins. LinkedIn reports that about 95% of surveyed users rated SQL Bot's query accuracy \"Passes\" or above, a user rating rather than a measured accuracy rate. A curated semantic layer and a narrow scope make the difference."},{"question":"How do you stop users seeing data they are not entitled to?","answer":"Run every query with the user's own identity so the data platform's row and column level security applies, give the assistant read only access to an allow list of schemas, and mask sensitive columns. Instructions in a prompt are not an access control."},{"question":"Does this replace the BI team?","answer":"No. It takes routine, repetitive questions off the queue. Analysts still own the definitions, certify example queries and handle complex or novel analysis, and anything that goes into a regulatory report or board pack is checked by an analyst."}],"related":["enterprise-knowledge-search","customer-feedback-analysis","regulatory-report-assembly","internal-audit-copilot","treasury-cash-flow-forecasting"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog, made industry neutral and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: industries now include pharma and life sciences, where its evidence comes from; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: corrected the Forrester figures as cited by Snowflake, sharpened the LinkedIn certification detail and the EU AI Act basis (Article 5, Annex III point 5(b), Article 50(1)), aligned the Blits.ai build notes with the platform's capabilities, added source dates, SEO title and meta description."},{"date":"2026-09-27","note":"Review fixes: the LinkedIn user survey no longer counts toward the accuracy benchmark; EU AI Act tier set to limited (Article 50(1)) in line with enterprise knowledge search; the Blits.ai build notes no longer imply end user role based access or per user identity passthrough; removed an internal data consolidation record that is not this use case; corrected the LinkedIn certification detail and the Forrester wording."}],"slug":"governed-text-to-sql-analytics","url":"https://www.blits.ai/ai-use-cases/governed-text-to-sql-analytics","benchmarks":[{"kpi":"users-served","label":"Users served","unit":"count","aggregate":false,"higherIsBetter":true,"n":1,"nUpTo":0,"median":300,"min":300,"max":300,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"uber-querygpt-natural-language-to-sql","pooled":true}]}],"indicativeValueResult":{"low":81000,"high":1260000},"evidence":["bayer-cortex-analyst-self-service-analytics","linkedin-sql-bot-text-to-sql","uber-querygpt-natural-language-to-sql"]},{"title":"Real time AI assist for contact centre agents","shortTitle":"Live agent assist","seoTitle":"AI agent assist for live contact centre calls","metaDescription":"AI agent assist transcribes live calls, surfaces approved knowledge and drafts wrap up notes. Definity says call summaries cut 3.5 minutes from each call.","definition":"A real time copilot for human contact centre agents during a live call or chat: it transcribes the conversation as it happens, surfaces the relevant knowledge and next step, drafts responses, and writes the after call summary and CRM notes, while the agent stays in control of what is said and done.","aliases":["agent assist","contact centre copilot","real time agent guidance","AI call summarization"],"industries":["cross-industry","banking","insurance","telecommunications","healthcare","retail-and-ecommerce","technology"],"functions":["customer-service","operations"],"patterns":["speech-analytics","rag-knowledge-assistant","summarization","content-generation"],"channels":["agent-desktop","voice","web-chat"],"audience":"employee-facing","autonomy":"assist","adoptionStage":"mainstream","problem":"Even with good self service, the contacts that reach a human are the hard ones: complex products,\nexceptions, complaints, distressed customers. Agents juggle several systems while the customer\nwaits, search the knowledge base mid call, and then spend minutes writing notes after every\ncontact (three to five minutes per call at Definity before it automated summaries). New agents\ntake months to become proficient, and attrition means there are always new agents.\n\nAgent assist attacks the time around the conversation (searching, typing, summarizing) and the\nknowledge gap of less experienced staff, without handing the customer to a machine. That makes it\none of the lower risk ways to bring generative AI into regulated customer service, provided the\nsuggestions are grounded in approved content and the recording is handled correctly.","problemStats":[{"statement":"Industry estimates cited by Brynjolfsson, Li and Raymond suggest that 60% of contact centre agents leave each year, costing firms $10,000 to $20,000 per agent.","sourceTitle":"Generative AI at Work (NBER Working Paper 31161)","sourceUrl":"https://www.nber.org/system/files/working_papers/w31161/w31161.pdf","year":2023},{"statement":"In the support operation studied by Brynjolfsson, Li and Raymond, agents without AI assistance needed more than six months of tenure to perform as well as assisted agents with two months.","sourceTitle":"Generative AI at Work (NBER Working Paper 31161)","sourceUrl":"https://www.nber.org/system/files/working_papers/w31161/w31161.pdf","year":2023}],"howItWorks":"1. **Transcribe live.** Streaming speech recognition turns both sides of the call into text in\n   real time; on chat the text is already there.\n2. **Understand the moment.** The system detects the customer's intent and key details (product,\n   account type, the problem) as the conversation develops.\n3. **Surface knowledge and next steps.** Relevant procedure snippets, eligibility rules and next\n   best actions appear in the agent desktop, retrieved from approved content.\n4. **Draft, do not send.** On chat and email it drafts replies the agent edits; on voice it\n   suggests wording for disclosures and explanations.\n5. **Wrap up automatically.** After the contact it writes a structured summary, fills CRM fields\n   and service request forms, and the agent confirms them.\n6. **Pause on sensitive data.** Card numbers and authentication answers are not transcribed or\n   stored, using pause and resume or redaction.","valueDrivers":["employee-productivity","cost-to-serve","customer-experience","compliance"],"kpis":["handling-time-reduction","productivity-gain","time-saved-per-task","first-contact-resolution","customer-satisfaction","accuracy"],"indicativeValue":{"referenceOrg":"A contact centre with 500 agents","inputs":[{"key":"agents","label":"Agents using the assistant","low":500,"high":500,"unit":"agents","note":"The reference organization."},{"key":"agentCost","label":"Fully loaded cost per agent","low":40000,"high":60000,"unit":"USD per agent per year","note":"Editorial assumption. Replace with your own cost, including outsourced seats."},{"key":"onContactShare","label":"Share of paid time spent on contacts and wrap up","low":0.7,"high":0.8,"unit":"fraction of paid time","note":"Editorial assumption for occupancy. Replace with your workforce management data."},{"key":"timeReduction","label":"Reduction in handling time, including wrap up","low":0.05,"high":0.15,"unit":"fraction of handling time","note":"Conservative against the evidence on this page (Definity says automated summaries saved 3.5 minutes per call; Google Cloud reports a 20% cut in call handle time for Definity's whole program, which also automated caller authentication; the NBER field study measured 14% more issues resolved per hour)."}],"formula":"agents * agentCost * onContactShare * timeReduction","currency":"USD","period":"per year","resultLabel":"Agent capacity released","caveat":"Released capacity, realized only if staffing or service levels are adjusted. It leaves out platform and transcription costs, and quality effects such as fewer errors, better compliance and faster onboarding of new agents."},"macroEstimates":[],"feasibility":{"complexity":"medium","complexityNote":"Summaries after the call are straightforward. Real time guidance needs low latency streaming transcription, integration with the telephony platform and agent desktop, and a knowledge base that is clean enough to surface the right snippet in seconds.","dataPrerequisites":["Approved knowledge articles and procedures with owners","Call recordings or transcripts to tune intents and test summaries","The CRM fields and disposition codes the summary must fill"],"integrations":["Telephony or contact centre platform with access to the audio stream","Agent desktop and CRM","Knowledge base","Card data redaction or pause and resume for payment calls"]},"implementation":{"steps":[{"title":"Start with after call summaries","detail":"Automated wrap up is the fastest, safest win: the agent reviews and confirms every summary, and time saved is easy to measure."},{"title":"Clean the knowledge before surfacing it","detail":"Real time suggestions are only as good as the articles behind them. Fix duplicates and outdated procedures for the top call reasons first."},{"title":"Add real time guidance for a few intents","detail":"Pick high volume intents with clear procedures and add knowledge surfacing and disclosure prompts. Watch whether agents use or ignore suggestions."},{"title":"Handle sensitive data by design","detail":"Make sure card numbers, authentication answers and special category data are redacted or never captured, and that transcripts are stored in region with defined retention."},{"title":"Measure with a control group","detail":"Roll out by team and compare handling time, resolution and satisfaction with teams that do not yet have it, on the same contact mix."},{"title":"Agree how the data will not be used","detail":"Tell agents what is recorded and agree that assist data is not used for individual performance scoring unless that is assessed and disclosed separately."}],"guardrails":["Suggestions only from approved knowledge, with the source visible to the agent","The agent confirms every summary and every drafted reply before it is saved or sent","Card data and authentication answers redacted or excluded from transcription","No inference of agents' emotions","Transcripts stored in region with a defined retention period"],"humanInTheLoop":"The agent decides what is said and done on every contact and confirms summaries before they are saved. Team leaders review a sample of summaries and suggestions for accuracy, and knowledge owners fix content behind wrong suggestions.","kpisToInstrument":["Average handling time and wrap up time, against a control group","Summary accuracy on a weekly sample","Suggestion acceptance rate per intent","First contact resolution and customer satisfaction","Time to proficiency for new agents"],"failureModes":[{"title":"Summaries nobody checks","detail":"Agents approve summaries without reading them and errors enter the CRM. Sample and score them, and make edits easy."},{"title":"Suggestion noise","detail":"Too many prompts distract agents mid call. Limit suggestions to what is relevant and measure acceptance."},{"title":"Card data in transcripts","detail":"A payment call is transcribed and stored with the card number. Build redaction or pause and resume in from the start."},{"title":"Assist becomes surveillance","detail":"Transcripts are reused to score individuals without disclosure or assessment. That changes the risk class and erodes trust."}]},"risk":{"euAiAct":{"tier":"context-dependent","basis":"As a pure assist tool for agents it is minimal risk; the customer does not interact with the AI. It becomes high risk under Annex III point 4(b) if its data is used to monitor and evaluate individual agents' performance, and inferring agents' emotions at work is prohibited under Article 5(1)(f)."},"regulations":["eu-ai-act","gdpr","pci-dss","uk-consumer-duty","dora","mas-ai-risk-management"],"guidance":[{"title":"Article 5, prohibited AI practices","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/article/5/","note":"Point 1(f) prohibits AI systems that infer emotions of natural persons in the workplace, except for medical or safety reasons, which rules out emotion scoring of agents."},{"title":"Annex III, high risk AI systems referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 4(b) covers systems that monitor and evaluate the performance and behaviour of workers."},{"title":"Artificial Intelligence (AI) Model Risk Management, information paper","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/publications/monographs-or-information-paper/2024/artificial-intelligence-model-risk-management","note":"Example of supervisory good practice for testing and monitoring generative AI applications at banks."}],"controls":["Data protection impact assessment covering call transcription and retention","PCI DSS scoping of the transcription and storage path","Written limits on the use of assist data for individual performance management","Inventory entry with an accountable owner in customer operations","Regular accuracy sampling of summaries and suggestions"],"incidents":[]},"blitsAi":{"howToBuild":"On Blits.ai the building blocks are **voice** telephony with real time streaming speech\nrecognition across several providers (with **self hosted transcription** and speaker\ndiarization available where recordings must stay on Blits.ai infrastructure), an **AI agent** grounded in a **knowledge base** of approved\nprocedures with hybrid retrieval, and **summarization** through an agent with **structured\noutput** that maps to CRM fields. The assistant is exposed to the agent desktop through the\n**REST or WebSocket API**, and **custom functions** write the confirmed summary to the CRM.\n\n**PII masking** at the gateway keeps sensitive data away from the model and logs, and card\nnumbers typed in chat are detected and tokenized there; for payment calls on voice, plan\nredaction or pause and resume as a design step. **Test suites** grade summaries and suggestions against\nreviewed examples on every change, and the platform is model agnostic with EU or UAE data\nresidency."},"faq":[{"question":"How much handling time does agent assist save?","answer":"Results vary with what is measured. Definity says automated call summaries cut three and a half minutes from each call, which Google Cloud's customer story puts at 33% of average handle time; Google Cloud's customer list reports a 20% cut for Definity's whole program, which also automated caller authentication, and a 15% efficiency gain at SEB. The NBER field study measured 14% more issues resolved per hour, and DBS expects up to 20% for its CSO Assistant once fully deployed, which is a forecast, not yet a result."},{"question":"Who benefits most?","answer":"Less experienced agents. The NBER field study of 5,179 support agents found 14% more issues resolved per hour on average, with a 34% improvement for novice and low skilled workers and little effect on the most experienced."},{"question":"Is agent assist high risk under the EU AI Act?","answer":"Not as an assist tool. It becomes high risk if the same data is used to evaluate individual agents (Annex III point 4(b)), and inferring agents' emotions at work is prohibited. Keep those uses separate and assessed."}],"related":["email-and-ticket-reply-drafting","first-line-contact-centre-agent","call-quality-and-compliance-monitoring","enterprise-knowledge-search","conversation-roleplay-training","meeting-summarization-and-action-items"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-25","note":"First version, written from the banking catalog as an industry neutral page, with six evidence records verified against their sources. The catalog presented DBS's 20% handling time cut as achieved; the DBS release states it as an expectation."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; industries now include technology, where its evidence comes from; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: added Google Cloud's Definity customer story with a 3.5 minute per call saving quoted from Definity, two problem statistics from the NBER working paper, and the time saved per task KPI; the FAQ range now matches the evidence (14% to 20%); removed an unsupported superlative from the NBER record; added seoTitle and metaDescription."},{"date":"2026-09-27","note":"Review fixes: the meta description no longer attributes Google Cloud's 15% figure to SEB; the handling time FAQ drops the 14% to 20% range, adds the 33% of average handle time from Definity's customer story and says the 20% covers Definity's whole program including automated caller authentication (also in the indicative value note); tightened the Blits.ai transcription and card data wording to match the feature inventory."},{"date":"2026-09-27","note":"Review fixes: SEB evidence year set from the earliest Wayback snapshot that shows the entry (2025); Definity evidence year requalified as the check date, since the entry could not be dated against a reachable Wayback snapshot; the Blits.ai agent copilot evidence now notes that its production stage rests on the internal portfolio entry alone and still needs confirmation from Blits.ai."}],"slug":"live-agent-assist","url":"https://www.blits.ai/ai-use-cases/live-agent-assist","benchmarks":[{"kpi":"productivity-gain","label":"Productivity gain","unit":"percent","aggregate":true,"higherIsBetter":true,"n":2,"nUpTo":0,"median":15,"min":15,"max":15,"byClaimant":{"organization":0,"vendor":2,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"definity-contact-centre-agent-assist","pooled":true},{"id":"seb-wealth-advisor-agent-assist","pooled":true}]},{"kpi":"accuracy","label":"Accuracy","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":100,"min":100,"max":100,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"dbs-cso-assistant","pooled":true}]},{"kpi":"handling-time-reduction","label":"Handling time reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":20,"min":20,"max":20,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"definity-contact-centre-agent-assist","pooled":true}]},{"kpi":"time-saved-per-task","label":"Time saved per task","unit":"minutes","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":3.5,"min":3.5,"max":3.5,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"definity-contact-centre-agent-assist","pooled":true}]}],"indicativeValueResult":{"low":700000,"high":3600000},"evidence":["dbs-cso-assistant","definity-contact-centre-agent-assist","oportun-ai-quality-management","seb-wealth-advisor-agent-assist","signal-iduna-co-si-knowledge-assistant"]},{"title":"Real time fraud scoring for card and instant payments","shortTitle":"Real time fraud scoring","seoTitle":"AI fraud scoring for card and instant payments","metaDescription":"Machine learning approves, challenges or blocks each card and instant payment in milliseconds. At NatWest, Featurespace reports 57% more value of fraud detected.","definition":"Machine learning that decides in milliseconds, without any conversation, how likely each card authorization and account to account payment is to be fraudulent, combining behavioural, device and network signals, so the bank can approve, challenge or block a payment before the money leaves. Working the resulting alerts and talking to the customer about them are separate use cases.","aliases":["transaction fraud scoring","real time payment fraud detection","card authorization fraud model","instant payment fraud screening"],"industries":["banking","payments"],"functions":["fraud-prevention"],"patterns":["prediction-and-scoring","anomaly-detection"],"channels":["api"],"audience":"back-office","autonomy":"autonomous","adoptionStage":"mainstream","segment":"middle-office","problem":"Fraud has moved to the fastest rails. Authorised push payment scams and account takeover often end in\ninstant account to account payments that settle in seconds and are hard to recall, and card not\npresent fraud has to be caught while the card authorization is still open. The decision to stop a\npayment is made inside that window, with no time for a human: Mastercard, for example, says\nDecision Intelligence Pro returns its improved score in less than 50 milliseconds.\n\nRule based engines struggle on both sides of that decision. Rules written for last quarter's\nattack miss the new one, and the rules that do fire decline many genuine customers, who then call\nthe contact centre or abandon the purchase. Scams are the hardest case: the customer is\nauthenticating the payment themselves, so strong authentication does not help, and only a change\nin their behaviour or the payee's profile gives the attack away.","problemStats":[{"statement":"The US Federal Trade Commission reports that consumers reported losing more than USD 12.5 billion to fraud in 2024, a 25% increase over the prior year.","sourceTitle":"New FTC Data Show a Big Jump in Reported Losses to Fraud to $12.5 Billion in 2024","sourceUrl":"https://www.ftc.gov/news-events/news/press-releases/2025/03/new-ftc-data-show-big-jump-reported-losses-fraud-125-billion-2024","year":2025},{"statement":"In a Feedzai survey of 562 fraud and financial crime professionals, 90% of financial institutions said they use AI to expedite fraud investigations and detect new tactics in real time.","sourceTitle":"AI Fraud Trends 2025: Banks Fight Back","sourceUrl":"https://www.feedzai.com/pressrelease/ai-fraud-trends-2025/","year":2025}],"howItWorks":"1. **Enrich the event.** Each authorization or payment request is joined in real time with the\n   customer's profile, recent behaviour, device and session data, merchant or payee history and\n   any confirmation of payee result.\n2. **Score it.** One or more models (gradient boosted trees, behavioural sequence models, graph\n   features that link accounts, devices and payees) return a risk score and the top reasons, within\n   the latency budget of the rail.\n3. **Decide with a strategy layer.** Score bands and business rules turn the score into an action:\n   approve, approve and monitor, step up authentication, hold for review, warn the customer about a\n   likely scam, or decline.\n4. **Learn from outcomes.** Confirmed fraud, chargebacks, scam reports and customer confirmations\n   flow back as labels, and challenger models are trained and compared before promotion.\n5. **Hand over the grey zone.** Holds and scam warnings create cases for the fraud team and, where\n   the customer is involved, a short interaction in the app or by phone.","valueDrivers":["risk-reduction","customer-experience","cost-to-serve"],"kpis":["fraud-loss-reduction","detection-rate-improvement","false-positive-reduction","automation-rate","interactions-handled"],"indicativeValue":{"referenceOrg":"A retail bank with 1 million active card and payment customers","inputs":[{"key":"fraudLosses","label":"Annual gross fraud losses on cards and payments","low":5000000,"high":15000000,"unit":"USD per year","note":"Editorial assumption for a bank of this size. Replace with your own gross fraud loss figure."},{"key":"lossReduction","label":"Share of fraud losses avoided by better scoring","low":0.1,"high":0.2,"unit":"fraction of losses","note":"Conservative against the achieved results on this page (Stripe reports an over 30% reduction in fraud on eligible transactions for early users of its new Radar interventions; Commonwealth Bank reports fraud losses down by over 20% year on year, with its detection technology playing a role), because not all of a bank's losses sit in the segment where scoring improves."},{"key":"manualReviews","label":"Transactions sent to manual review per year","low":50000,"high":150000,"unit":"reviews per year","note":"Editorial assumption. Replace with your own review queue volume."},{"key":"reviewReduction","label":"Share of manual reviews avoided","low":0.1,"high":0.25,"unit":"fraction of reviews","note":"Capped at the figure Visa reports for active Decision Manager users (manual reviews reduced by 25% or more)."},{"key":"costPerReview","label":"Cost of one manual review","low":4,"high":10,"unit":"USD per review","note":"Editorial assumption for a few minutes of fully loaded analyst time per review. Replace with your own."}],"formula":"fraudLosses * lossReduction + manualReviews * reviewReduction * costPerReview","currency":"USD","period":"per year","resultLabel":"Fraud losses and review cost avoided","caveat":"Leaves out the revenue recovered from fewer false declines, the effect on scam reimbursement liabilities, and the cost of the platform, data engineering and model validation."},"macroEstimates":[],"feasibility":{"complexity":"high","complexityNote":"The model is the easy part. The work is streaming data at authorization latency, a clean label pipeline from chargebacks and scam reports, a strategy layer that the fraud team can tune, and model risk validation for a model that decides on customers' payments without a human.","dataPrerequisites":["Labelled history of fraud, chargebacks and scam reports linked to the original transactions","Real time customer, device and session data available within the latency budget","Payee and merchant history, including confirmation of payee results where available","Customer contact outcomes for held or challenged payments"],"integrations":["Card authorization host or issuer processor","Instant payment and account to account payment hub","Digital banking channels for device and session signals and in app scam warnings","Case management for held payments and confirmed fraud","Network or consortium scores from card schemes and industry data sharing schemes"]},"implementation":{"steps":[{"title":"Map the decision points and latency budgets","detail":"List every place a payment can be stopped (authorization, payment initiation, payee creation, login) and the time available at each. This decides which features and models are feasible."},{"title":"Build the label pipeline first","detail":"Link chargebacks, customer fraud claims and scam reports back to the transactions, with dates, so you can train on what really happened and measure detection honestly."},{"title":"Run champion and challenger in shadow mode","detail":"Score live traffic with the new model without acting on it, and compare detection and false positive rates against the current engine on the same transactions for several weeks."},{"title":"Design the strategy layer with the fraud team","detail":"Agree score bands and actions per segment and rail, including when to warn a customer about a likely scam instead of declining, and document who can change thresholds."},{"title":"Validate and inventory the model","detail":"Put the model through independent validation, record it in the model inventory with an owner, and set monitoring for drift, data quality and fairness across customer groups."},{"title":"Close the loop with customer contact","detail":"Make sure held or declined payments can be released quickly through the app or the contact centre, and feed those outcomes back as labels."}],"guardrails":["Every automated decline or hold returns reason codes that staff can explain to the customer","Threshold changes go through change control with a documented owner and a rollback plan","A fallback rule set takes over automatically if the model or its data feeds fail","Regular fairness testing so that false declines do not concentrate on particular customer groups","Card data handled only inside the PCI DSS scope, with tokenized identifiers elsewhere"],"humanInTheLoop":"The model decides autonomously within the latency window, so human control sits around it: the fraud strategy team owns thresholds and actions, analysts work the held payments and scam warnings, and model risk validates every material change before it goes live.","kpisToInstrument":["Fraud detection rate and value detection rate on confirmed fraud, by rail and segment","False positive ratio (genuine transactions declined or held per fraud caught)","Gross fraud and scam losses normalised for volume","Share of held payments released by the customer, and time to release","Model latency at the 99th percentile and fallback activations"],"failureModes":[{"title":"Label leakage and optimistic backtests","detail":"Models trained on labels that were only known after the fact look excellent offline and disappoint live. Build features only from data available at decision time and trust shadow mode results over backtests."},{"title":"Fraud moves to the next rail","detail":"Tightening card controls pushes attackers to instant payments or account takeover. Score all rails and watch the loss mix, not one channel."},{"title":"False declines hidden in the dashboard","detail":"A model tuned only for detection quietly declines good customers. Track the false positive ratio and complaints as closely as losses."},{"title":"Silent model drift","detail":"Behaviour changes (a new wallet, a holiday season) degrade the model without an alert. Monitor feature distributions and score stability daily."}]},"risk":{"euAiAct":{"tier":"minimal","basis":"Annex III point 5(b) lists creditworthiness assessment and credit scoring of natural persons as high risk but explicitly excludes AI systems used for the purpose of detecting financial fraud, and payment fraud scoring is not otherwise listed in Annex III or prohibited by Article 5. Behavioural biometrics used only to confirm that customers are who they claim to be fall under the biometric verification exclusion in Annex III point 1(a). The model does not interact with people, so Article 50 does not apply. GDPR Article 22 can still apply to solely automated declines with significant effects on customers."},"regulations":["eu-ai-act","gdpr","dora","pci-dss","eu-psd2","uk-consumer-duty","uk-psr-app-reimbursement","pra-ss1-23","us-sr-11-7","nist-ai-rmf","mas-ai-risk-management","mas-shared-responsibility-framework","apra-cps-230"],"guidance":[{"title":"Annex III: High-Risk AI Systems Referred to in Article 6(2)","issuer":"European Union","region":"europe","url":"https://artificialintelligenceact.eu/annex/3/","note":"Point 5(b) excludes AI systems used for the purpose of detecting financial fraud from the high risk creditworthiness category."},{"title":"APP scams","issuer":"Payment Systems Regulator","region":"europe","url":"https://www.psr.org.uk/our-work/app-scams/","note":"The UK reimbursement requirement for authorised push payment scams over Faster Payments and CHAPS has sending and receiving firms split the cost of reimbursing victims 50:50, which puts the cost of missed scams on both sides of the payment."},{"title":"Guidelines on Shared Responsibility Framework","issuer":"Monetary Authority of Singapore","region":"asia-pacific","url":"https://www.mas.gov.sg/regulation/guidelines/guidelines-on-shared-responsibility-framework","note":"Implemented from 16 December 2024, it assigns financial institutions and telcos duties to mitigate phishing scams and requires payouts to victims where those duties are breached, which raises the value of real time detection."}],"controls":["Model inventory entry with owner, validation report and monitoring plan","Reason codes stored with every automated decline or hold","Documented threshold governance with change control and rollback","Fairness and customer outcome monitoring on false declines","Tested fallback to a rule set when the model or data feeds are unavailable"],"incidents":[]},"blitsAi":{"howToBuild":"The scoring model itself runs in the bank's payment stack, next to the authorization host,\nbecause it has to answer within milliseconds. Blits.ai covers the parts around it that involve\npeople. When a customer gets in touch about a held payment, in the bank's app through the **REST\nor WebSocket API channel**, on **WhatsApp**, by **SMS** or on the phone, an **agent** confirms\nwhether they made the payment, gives the scam warnings the fraud team wrote in the **knowledge\nbase**, and calls **custom functions** to release the payment or keep the block. The bank's\nsystems can also trigger an **agentic workflow** through the API to prepare the case, with **human\nin the loop confirmation** so an analyst approves a release above a set amount.\n\nOn the phone, **voice** telephony uses real time streaming speech recognition and synthesis, and\nanything unusual goes to a fraud specialist through **human handover**, with the conversation\nhistory passed along. **Guardrails** and **PII masking** keep card and account data out of model\nprompts, **test suites** replay scam scenarios on every change, and the platform is **model\nagnostic**, with EU and UAE data residency options."},"faq":[{"question":"How much does machine learning improve fraud detection?","answer":"Published results come mostly from vendors and networks. Featurespace reports, citing NatWest data from 2025, 57% more value of fraud detected and 75% fewer false positives on scams, and Stripe says early users of its new Radar interventions saw fraud on eligible transactions fall by over 30%. Pre launch modelling figures, such as those Mastercard published for Decision Intelligence Pro, are not measured results, so measure your own gain in shadow mode on your own traffic."},{"question":"Is a fraud scoring model high risk under the EU AI Act?","answer":"Not by default. Annex III point 5(b) explicitly excludes AI used to detect financial fraud from the high risk creditworthiness category. GDPR rules on automated decisions and your model risk framework still apply, so keep reason codes and a route for customers to challenge a decline."},{"question":"Can real time scoring stop authorised push payment scams?","answer":"Partly. The customer authorises the payment, so the signal is in behaviour and in the payee: a new payee, an unusual amount, a remote access session or a mule account on the receiving side. Banks that report results pair scoring with an intervention: Revolut declines card payments its model judges likely to be scams and sends the customer through an in app intervention flow, and reports a 30% reduction in fraud losses from card scams where money was sent for investment opportunities. Commonwealth Bank sends more than 40,000 proactive warning alerts a day in its app."}],"related":["scam-payment-interception","fraud-alert-triage","fraud-alert-confirmation","mule-network-detection","application-and-identity-fraud-detection","agentic-payment-initiation"],"datePublished":"2026-09-27","dateModified":"2026-09-27","lastVerified":"2026-09-27","changelog":[{"date":"2026-09-27","note":"Published after review by an automated review workflow (independent skeptic review)."},{"date":"2026-09-27","note":"Blocker fix: corrected the Revolut FAQ sentence to match its source (decline, not an intervention on the payment itself; the 30% figure covers card scams for investment opportunities, not all investment scams); removed the Pay dot UK and Visa pilot's detection rate metric, a retrospective backtest, and set outcomeDisclosed to false; removed the real time fraud scoring link from the US Treasury evidence, since its figure is post payment recovery, not real time scoring."},{"date":"2026-09-25","note":"First version, written from the Banking AI Use Case Catalog and verified against the sources."},{"date":"2026-09-25","note":"Consolidation pass: sharpened the definition so the difference from neighbouring use cases is explicit; added PSD2, UK APP scam reimbursement rules to the regulations; revised the related use cases."},{"date":"2026-09-26","note":"Fact checked against sources: removed Mastercard's pre launch modelling figures as metrics and from the FAQ and value inputs; removed a Treasury figure not attributed to machine learning; tied the latency claim to its source; corrected the EU AI Act basis (biometric verification, Article 50), the PSR and MAS guidance notes and the Blits.ai build notes; added PRA SS1/23 and the MAS Shared Responsibility Framework; added SEO title and description."}],"slug":"real-time-fraud-scoring","url":"https://www.blits.ai/ai-use-cases/real-time-fraud-scoring","benchmarks":[{"kpi":"fraud-loss-reduction","label":"Fraud loss reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":3,"nUpTo":0,"median":30,"min":30,"max":76,"byClaimant":{"organization":3,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"commonwealth-bank-scam-and-fraud-interventions","pooled":true},{"id":"revolut-card-scam-detection","pooled":true},{"id":"stripe-radar-payments-foundation-model","pooled":true}]},{"kpi":"interactions-handled","label":"Interactions handled","unit":"count","aggregate":false,"higherIsBetter":true,"n":3,"nUpTo":0,"median":180000000,"min":40000,"max":3200000000,"byClaimant":{"organization":2,"vendor":1,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"visa-decision-manager","pooled":true},{"id":"biocatch-trust-australia-mule-intelligence","pooled":true},{"id":"commonwealth-bank-scam-and-fraud-interventions","pooled":true}]},{"kpi":"automation-rate","label":"Automation rate","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":98.7,"min":98.7,"max":98.7,"byClaimant":{"organization":1,"vendor":0,"regulator":0,"independent":0},"vendorOnly":false,"evidence":[{"id":"visa-decision-manager","pooled":true}]},{"kpi":"detection-rate-improvement","label":"Detection improvement","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":135,"min":135,"max":135,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"natwest-featurespace-fraud-and-scam-detection","pooled":true}]},{"kpi":"false-positive-reduction","label":"False positive reduction","unit":"percent","aggregate":true,"higherIsBetter":true,"n":1,"nUpTo":0,"median":75,"min":75,"max":75,"byClaimant":{"organization":0,"vendor":1,"regulator":0,"independent":0},"vendorOnly":true,"evidence":[{"id":"natwest-featurespace-fraud-and-scam-detection","pooled":true}]}],"indicativeValueResult":{"low":520000,"high":3375000},"evidence":["biocatch-trust-australia-mule-intelligence","commonwealth-bank-scam-and-fraud-interventions","mastercard-consumer-fraud-risk","mastercard-decision-intelligence-pro","natwest-featurespace-fraud-and-scam-detection","pay-uk-visa-account-to-account-fraud-pilot","revolut-card-scam-detection","stripe-radar-payments-foundation-model","visa-decision-manager"]}],"evidence":[{"title":"7-Eleven Vietnam: internal IT support chatbot for employees","useCases":["it-service-desk-resolution-agent"],"organization":{"name":"7-Eleven Vietnam","anonymized":false,"country":"VN","region":"asia-pacific","industry":"retail-and-ecommerce"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"7-Eleven Vietnam, with 140 stores, built an internal IT support chatbot on Vertex AI Agent Builder and Gemini so employees can resolve technical issues on their own. Google Cloud reports that it halved the time spent fixing IT issues and lightened the IT team's workload.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"productivity-gain","value":50,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"The chatbot reduced time spent fixing IT issues by 50%, lightening the workload for the IT team.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2024-04-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"7-eleven-vietnam-it-support-chatbot"},{"title":"8x8: AI SDR agent that engages and qualifies inbound website buyers","useCases":["inbound-lead-qualification-agent"],"organization":{"name":"8x8","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Qualified","role":"platform"}],"summary":"8x8 had rising inbound traffic but falling meeting volume: chats went unanswered, more than half of website sessions fell outside business hours and sales development reps spent time sorting job seekers and support requests from buyers. It deployed an AI SDR agent that engages visitors around the clock, qualifies them by company size, country and intent, routes meeting and pricing requests and hands ready conversations to sales, with automated email follow up for buyers who did not book. The vendor reports gains across the funnel in the first nine months.","stage":"production","year":2025,"channels":["web-chat","email"],"languages":["en"],"metrics":[{"kpi":"conversion-rate-uplift","value":19,"unit":"percent","qualifier":"exact","period":"first nine months live","baseline":"MQL to SQL conversion before the agent","claimant":"vendor","quote":"+19% MQL → SQL conversion","sourceUrl":"https://www.qualified.com/customers/8x8"},{"kpi":"conversion-rate-uplift","value":24,"unit":"percent","qualifier":"exact","period":"first nine months live","baseline":"MQL to closed won deals before the agent","claimant":"vendor","quote":"+24% MQL → closed-won deals","sourceUrl":"https://www.qualified.com/customers/8x8"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.qualified.com/customers/8x8","title":"8x8 Drives 24% More Closed-Won Inbound Revenue with Piper the AI SDR Agent","publisher":"Qualified"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"8x8-ai-sdr-inbound-qualification"},{"title":"A-LIGN: automated account research for competitive displacement outreach","useCases":["outbound-sales-prospecting-agent"],"organization":{"name":"A-LIGN","anonymized":false,"country":"US","region":"north-america","industry":"professional-services"},"vendors":[{"name":"Clay","role":"platform"}],"summary":"A-LIGN, a security and compliance firm, previously paid a contractor to research 2,000 target accounts by hand over six months, which produced yes or no answers that reps found of little use. It replaced this with AI research workflows in Clay that find which of 15 compliance services each account uses, why it needs them and which competitor provides them, and push the result into Salesforce, so reps open conversations with a specific reason instead of a generic pitch. The workflow went live between March and May 2025 and the vendor reports lower research costs and displacement pipeline tracked through a rep incentive program. The vendor states the cost saving inconsistently: an 83% reduction in the headline and results, but a Clay contract that cost USD 10,000 less than the USD 60,000 manual contract in the body.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.clay.com/customers/a-lign","title":"How A-LIGN cut research costs by 83% and generated millions in pipeline","publisher":"Clay"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"a-lign-outbound-account-research"},{"title":"Australian Broadcasting Corporation: AI tagging of the CoDA video archive","useCases":["media-archive-metadata-tagging"],"organization":{"name":"Australian Broadcasting Corporation","anonymized":false,"country":"AU","region":"asia-pacific","industry":"media-and-entertainment"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"The ABC modernized its 90 year old archive with Gemini in Vertex AI, adding rich, AI generated metadata to CoDA (Content Digital Archives), its open access platform for journalists, producers and editors. Gemini catalogs video segments, flags incidental footage, and powers semantic search so content makers can describe what they are looking for in natural language instead of scanning vague, human written tags. The editability of the AI generated metadata is presented as a key advantage, with human oversight reviewing and refining output for consistency and to meet the ABC's quality standards, and a human in the loop correction process feeding back into the models over time. The ABC has since integrated the tagging into its daily content pipeline so new uploads are described automatically.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":1000000,"unit":"count","qualifier":"exact","period":"within two weeks","claimant":"vendor","quote":"Analysed one million video records within two weeks with the Gemini API in Vertex AI","sourceUrl":"https://cloud.google.com/customers/abc"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/abc","title":"Unlocking Australian history: How the ABC used Gemini in Vertex AI to tag a million video archives in weeks","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"abc-archive-ai-metadata-tagging"},{"title":"Absa: Abby, an agentic assistant that guides customers to products and applications","useCases":["conversational-loan-application-intake"],"organization":{"name":"Absa Bank","anonymized":false,"country":"ZA","region":"africa","industry":"banking"},"vendors":[{"name":"Salesforce","role":"platform"}],"summary":"Absa runs Abby, a customer facing assistant built on Salesforce Agentforce, in production. At a Salesforce event in June 2025 Absa's Relationship Banking technology chief described it asking business customers about their needs, such as working capital, and recommending matching products. The report says that, according to Absa's website, the agent can help customers apply for loans, open investment accounts and make international payments. Absa staff set guidelines on what the agent may not do, and Salesforce says it urges its partners to route topics such as pricing to employees. Salesforce could not share the impact on Absa customers, and no outcome figures were disclosed.","stage":"production","year":2025,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://htxt.co.za/2025/06/south-african-banking-giant-reaches-global-first-with-ai-agent/","title":"South African banking giant reaches global first with AI","publisher":"Hypertext","date":"2025-06-05"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"absa-abby-agentic-assistant"},{"title":"Abu Dhabi Government: TAMM AI assistant for government services","useCases":["citizen-information-assistant","non-emergency-service-request-routing"],"organization":{"name":"Abu Dhabi Government (TAMM)","anonymized":false,"country":"AE","region":"middle-east","industry":"government"},"vendors":[{"name":"Microsoft (Azure OpenAI Service)","role":"model-provider"},{"name":"G42 (Compass 2.0)","role":"platform"}],"summary":"TAMM is Abu Dhabi's single platform for about 950 government services from many entities, from car registration and visa renewals to traffic fines. The platform, including its AI assistant, is powered by Azure OpenAI Service and G42 Compass 2.0 (which also gives access to the Arabic JAIS model). The assistant answers questions about processes, shows the status of a user's requests and speaks several languages. A photo reporting feature lets residents photograph a problem such as a pothole or a broken traffic light; the assistant helps fill in the report and updates the reporter on the repair.","stage":"scaled","year":2025,"channels":["mobile-app","web-chat"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://news.microsoft.com/source/emea/features/tamm-app-abu-dhabi-government-services/","title":"Consider it done: How TAMM is transforming government services in Abu Dhabi with AI","publisher":"Microsoft Source EMEA","date":"2025-02-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"abu-dhabi-tamm-ai-assistant"},{"title":"Accelerant: AI agents for data ingestion and AI tools for underwriting and actuarial analysis","useCases":["underwriting-risk-assessment-copilot","insurance-pricing-and-actuarial-copilot"],"organization":{"name":"Accelerant Holdings","anonymized":false,"region":"global","industry":"insurance"},"vendors":[{"name":"Accelerant Holdings","role":"in-house"}],"summary":"Accelerant runs a risk exchange that connects specialty MGAs (its Members) with risk capital. Its 2025 annual report says incoming data, from Member bordereaux to third party sources, is validated, transformed and governed using AI agents, that internally developed AI tools and models assist Members' underwriting, and that its risk evaluation tools help Members identify, classify, validate, research and price underwriting opportunities. Members also get AI supported claims insights, actuarial analysis and portfolio management to manage rate adequacy. Engineers, data scientists, product managers and designers made up 34% of its workforce at the end of 2025. No outcome figures are disclosed.","stage":"production","year":2025,"channels":["internal-tools","api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1997350/000199735026000003/arx-20251231.htm","title":"Accelerant Holdings Form 10-K for 2025","publisher":"Accelerant Holdings via SEC EDGAR","date":"2026-03-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"accelerant-ai-underwriting-and-actuarial-tools"},{"title":"Accenture: randomized controlled trial of GitHub Copilot","useCases":["developer-coding-assistant"],"organization":{"name":"Accenture","anonymized":false,"country":"IE","region":"global","industry":"professional-services"},"vendors":[{"name":"GitHub","role":"platform"}],"summary":"GitHub and Accenture ran a randomized controlled trial in which Accenture developers were randomly given GitHub Copilot or not, and measured DevOps telemetry such as pull requests and build success. The Copilot group opened more pull requests and saw many more successful builds, and surveyed developers reported less mental effort on repetitive tasks. GitHub also ran a company wide adoption analysis of installation and suggestion acceptance at Accenture.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":8.69,"unit":"percent","qualifier":"exact","baseline":"Pull requests per developer in the control group","claimant":"vendor","quote":"Ultimately, an increase in pull requests represents an increase in value delivered, and Accenture developers saw an 8.69% increase in pull requests.","sourceUrl":"https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-in-the-enterprise-with-accenture/"}],"outcomeDisclosed":true,"sources":[{"url":"https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-in-the-enterprise-with-accenture/","title":"Research: Quantifying GitHub Copilot's impact in the enterprise with Accenture","publisher":"GitHub","date":"2024-05-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"accenture-github-copilot-trial"},{"title":"Acentra Health: MedScribe drafting of Medicare appeal determination letters","useCases":["health-prior-authorization-and-claims-adjudication","outbound-notice-drafting"],"organization":{"name":"Acentra Health","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Acentra Health, which reviews Medicare appeals, built MedScribe on Azure OpenAI Service to turn a physician's clinical rationale into a plain language, empathetic appeal determination letter for the beneficiary and the provider, a task specially trained nurses used to do by hand in an appeals process where a decision can be due within 24 hours. It was tested with 10 nurses who rated every draft, then rolled out to all nurses who write these letters. Microsoft reports that time per letter fell by about 50%, from six to three minutes, saving 11,000 nursing hours and nearly USD 800,000 since deployment, and that nurses gave the generated letters a 99% approval rating.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":50,"unit":"percent","qualifier":"approximately","period":"Nurse time per appeal determination letter","claimant":"vendor","quote":"With MedScribe, Acentra Health reduced the time that its specially trained nursing staff spent on each appeal determination letter by approximately 50%.","sourceUrl":"https://www.microsoft.com/en/customers/story/19280-acentra-health-azure"},{"kpi":"hours-saved","value":11000,"unit":"hours","qualifier":"exact","period":"Nursing hours saved by mid 2024","claimant":"vendor","quote":"By mid-2024, the company had saved 11,000 nursing hours and begun preparations to expand MedScribe to other areas to drive organizational efficiency.","sourceUrl":"https://www.microsoft.com/en/customers/story/19280-acentra-health-azure"},{"kpi":"cost-savings","value":800000,"unit":"currency","currency":"USD","qualifier":"approximately","period":"Saved since deployment","claimant":"vendor","quote":"This adds up to 11,000 nursing hours and nearly $800,000 that have been saved since deploying MedScribe.","sourceUrl":"https://www.microsoft.com/en/customers/story/19280-acentra-health-azure"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/19280-acentra-health-azure","title":"Acentra Health boosts employee productivity with generative AI and Azure OpenAI Service, saving 11,000 nursing hours and nearly $800,000","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"acentra-health-medscribe-appeal-letters"},{"title":"Adelphi University: Adele, an AI assistant for student questions and text reminders","useCases":["student-enrollment-and-services-assistant"],"organization":{"name":"Adelphi University","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[{"name":"Mainstay","role":"platform"}],"summary":"Adelphi University launched Adele, a conversational AI assistant on the Mainstay platform, on its website in January 2022 and extended it to two way text messaging for about 6,000 current students in March 2023. Adele sends reminders about academic and financial deadlines, answers questions from a shared knowledge base, enabled generative AI in July 2025 and requests a human from the right office by email when needed. A cross office task force coordinates campaigns, and the university adopted a policy for text messaging in June 2024. The vendor reports 81,167 messages handled by the bot in the past year. Assuming approximately one minute per message, the vendor estimates this at 1,353 staff capacity hours; that is a modelled figure, not a measured saving, so it is not recorded as a metric.","stage":"production","year":2022,"channels":["web-chat","sms"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":81167,"unit":"count","qualifier":"exact","period":"in the past year, per the case study","claimant":"vendor","quote":"81,167 messages handled by the bot","sourceUrl":"https://mainstay.com/case-study/adelphi-student-support-capacity-case-study/"}],"outcomeDisclosed":true,"sources":[{"url":"https://mainstay.com/case-study/adelphi-student-support-capacity-case-study/","title":"When IT Thinks Beyond IT: How Adelphi Gained Capacity by Reducing Barriers for Students","publisher":"Mainstay"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"adelphi-university-adele-student-assistant"},{"title":"Admiral Seguros: touchless motor claims with AI damage estimates","useCases":["photo-based-damage-assessment","claims-triage-and-straight-through-processing"],"organization":{"name":"Admiral Seguros","anonymized":false,"country":"ES","region":"europe","industry":"insurance"},"vendors":[{"name":"Tractable","role":"platform"}],"summary":"Admiral Seguros, the Spanish business of Admiral Group, sends motor claimants a link to a web app in which they photograph the damage, and Tractable's AI produces the repair estimate within minutes. Tractable reports that Admiral Seguros processed 12,000 touchless claims this way in 2021, that 90% of claim estimates were processed without human appraisers and that 98% of claims were completed in less than 15 minutes.","stage":"production","year":2021,"channels":[],"languages":["es"],"metrics":[{"kpi":"interactions-handled","value":12000,"unit":"count","qualifier":"exact","period":"Touchless claims in 2021","claimant":"vendor","quote":"In 2021, Admiral Seguros processed 12,000 touchless claims using Tractable AI.","sourceUrl":"https://tractable.ai/case-studies/admiral-seguros/"}],"outcomeDisclosed":true,"sources":[{"url":"https://tractable.ai/case-studies/admiral-seguros/","title":"Admiral Seguros: Tractable delivers outstanding customer services through touchless claims","publisher":"Tractable","date":"2023-04-21"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"admiral-seguros-ai-vehicle-damage-estimates"},{"title":"AdvanceCare: AI extraction and automation of health insurance claims","useCases":["health-prior-authorization-and-claims-adjudication"],"organization":{"name":"AdvanceCare","anonymized":false,"country":"PT","region":"europe","industry":"insurance"},"vendors":[{"name":"Sprout.ai","role":"platform"}],"summary":"AdvanceCare, a health insurance group in Portugal that is part of the Generali Group and acts as a third party administrator for 1.7 million members, has used Sprout.ai's platform since January 2023 to extract, structure and validate data from hospital, dental, pharmacy and other healthcare invoices, map claim descriptions to codes with confidence levels, and automate settlement of routine claims. Sprout.ai reports that some routine claims are now settled in 60 seconds, that over a million claims went through the platform in the past year, that automation levels rose by more than 10%, and that a pilot on a sample of invoices before the partnership was 98% accurate.","stage":"scaled","year":2023,"channels":["api"],"languages":["pt"],"metrics":[{"kpi":"interactions-handled","value":1000000,"unit":"count","qualifier":"at-least","period":"Claims processed in the past year","claimant":"vendor","quote":"The milestone comes after the health insurer TPA (Third-Party Administrator) processed over a million claims in the past year using Sprout.ai’s patented AI platform.","sourceUrl":"https://sprout.ai/resource/advancecare-slashes-claims-processing-times-to-60-seconds-using-sprout-ais-technology/"},{"kpi":"accuracy","value":98,"unit":"percent","qualifier":"exact","period":"Pilot on a sample of healthcare invoices across categories, before the partnership","claimant":"vendor","quote":"The results were 98% accurate across all categories.","sourceUrl":"https://sprout.ai/resource/case-study-advancecare-and-sprout-ai/"}],"outcomeDisclosed":true,"sources":[{"url":"https://sprout.ai/resource/advancecare-slashes-claims-processing-times-to-60-seconds-using-sprout-ais-technology/","title":"AdvanceCare slashes claims processing times to 60 seconds using Sprout.ai’s technology","publisher":"Sprout.ai","date":"2025-07-16"},{"url":"https://sprout.ai/resource/case-study-advancecare-and-sprout-ai/","title":"AdvanceCare & Sprout.ai collaborate on AI to achieve 60-second claim turnaround","publisher":"Sprout.ai"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"advancecare-health-claims-automation"},{"title":"Adventist Health + Rideout: AI stroke transfer time reduction","useCases":["radiology-worklist-triage"],"organization":{"name":"Adventist Health + Rideout","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Viz.ai","role":"platform"}],"summary":"Adventist Health + Rideout, a regional primary stroke center in a hub and spoke network, used the Viz.ai platform's real time imaging analysis and automated care coordination as part of a quality improvement initiative to speed up the transfer of large vessel occlusion stroke patients to a comprehensive stroke center. The program combined the AI platform with a partnership with a comprehensive stroke center and standardized transfer protocols. Data presented at the American Heart Association's 2026 International Stroke Conference, led by the hospital's stroke program manager Caezar G. Jara, showed the changes cut average door in door out transfer time to 113 minutes, below the Joint Commission's 120 minute national benchmark.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":44,"unit":"percent","qualifier":"approximately","period":"quality improvement initiative combining Viz.ai platform deployment, a partnership with a comprehensive stroke center, and standardized transfer protocols, presented at ISC 2026","baseline":"202 minutes average door in door out (DIDO) time before the quality improvement initiative","claimant":"vendor","quote":"Viz.ai, the leader in AI-powered disease detection and intelligent care coordination, today announced the presentation of new clinical data at the American Heart Association's International Stroke Conference (ISC) 2026 demonstrating a 44% reduction in door-in-door-out (DIDO) time — the time required to evaluate, coordinate, and transfer a patient to a comprehensive stroke center — for patients with large vessel occlusion (LVO) stroke in regional care settings.","sourceUrl":"https://www.viz.ai/news/viz-ai-study-demonstrates-44-reduction-in-interfacility-stroke-transfer-times"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.viz.ai/news/viz-ai-study-demonstrates-44-reduction-in-interfacility-stroke-transfer-times","title":"Viz.ai Study Demonstrates 44% Reduction in Interfacility Stroke Transfer Times","publisher":"Viz.ai","date":"2026-03-05"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"adventist-health-rideout-viz-ai-stroke-transfer"},{"title":"AIG: AIG Underwriter Assistance for submission ingestion, prioritization and augmentation","useCases":["commercial-underwriting-submission-triage","underwriting-risk-assessment-copilot"],"organization":{"name":"American International Group","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Anthropic","role":"model-provider"},{"name":"Palantir","role":"platform"},{"name":"Amazon Web Services","role":"platform"}],"summary":"At its March 2025 investor day, AIG presented AIG Underwriter Assistance, a generative AI solution in production in Financial Lines that extracts data from broker and agent submissions, augments it with AIG and approved third party data, summarizes each submission and ranks submissions by appetite and propensity to bind, after which the underwriter analyzes the output and quotes. AIG describes the prior submission to quote process as taking about three to four weeks, with underwriters unable to review every submission, and says the assistant prepares submissions for review within one day. AIG frames the build around a human in the loop principle and was extending the same components to claims.","stage":"production","year":2025,"channels":["internal-tools","api"],"languages":["en"],"metrics":[{"kpi":"cycle-time-days","value":1,"unit":"days","qualifier":"up-to","period":"submission to underwriter ready file, in production lines","baseline":"about three to four weeks in the previous submission to quote process, as shown by AIG","claimant":"organization","quote":"AIG Underwriter Assistance Synthesizes and Prepares Submissions for Underwriter Review Within One Day","sourceUrl":"https://www.sec.gov/Archives/edgar/data/5272/000000527225000017/aig_investorxdayx2025.htm"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/5272/000000527225000017/aig_investorxdayx2025.htm","title":"AIG Investor Day 2025 presentation (Form 8-K, Exhibit 99.1)","publisher":"American International Group via SEC EDGAR","date":"2025-03-31"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"aig-underwriter-assistance"},{"title":"Air India: AI.g generative AI virtual assistant","useCases":["flight-disruption-and-rebooking-agent","first-line-contact-centre-agent"],"organization":{"name":"Air India","anonymized":false,"country":"IN","region":"asia-pacific","industry":"travel-and-hospitality"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Air India launched AI.g in May 2023, a virtual assistant on Azure OpenAI that is integrated with the reservation system and answers questions across 1,300 topic areas including bookings, flight status, baggage, check in, frequent flyer awards and lounge access, and escalates automatically to contact centre staff when it detects the need. The airline says it has kept contact centre call volume flat while its passenger count doubled.","stage":"scaled","year":2023,"channels":["web-chat","mobile-app"],"languages":[],"metrics":[{"kpi":"automation-rate","value":97,"unit":"percent","qualifier":"exact","period":"cumulative, of nearly 4 million queries","claimant":"vendor","quote":"To date, AI.g has successfully handled nearly 4 million customer queries, 97% of them with full automation.","sourceUrl":"https://customers.microsoft.com/en-us/story/1836108400811529412-airindia-azure-ai-search-travel-and-transportation-en-india"},{"kpi":"interactions-handled","value":10000,"unit":"count","qualifier":"approximately","period":"per day","claimant":"organization","quote":"That's because AI.g is handling about 10,000 a day.","sourceUrl":"https://customers.microsoft.com/en-us/story/1836108400811529412-airindia-azure-ai-search-travel-and-transportation-en-india"}],"outcomeDisclosed":true,"sources":[{"url":"https://customers.microsoft.com/en-us/story/1836108400811529412-airindia-azure-ai-search-travel-and-transportation-en-india","title":"Air India elevates customer support while saving money with Azure AI, data, and apps","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"air-india-aig-virtual-assistant"},{"title":"Airbnb: AI assistant for guest and host support","useCases":["travel-and-hotel-booking-concierge","first-line-contact-centre-agent"],"organization":{"name":"Airbnb","anonymized":false,"country":"US","region":"global","industry":"travel-and-hospitality"},"vendors":[],"summary":"Airbnb runs an AI assistant as the first line of customer support for guests and hosts. It was expanded to all US users in 2025 and then rolled out to more countries and languages, reaching more than 50 languages by mid 2026. Airbnb reports the share of issues resolved without a human agent in each quarterly letter and links part of the fall in support cost per booking to the assistant; it plans an AI voice assistant.","stage":"scaled","year":2025,"channels":["mobile-app","web-chat"],"languages":[],"metrics":[{"kpi":"containment-rate","value":45,"unit":"percent","qualifier":"approximately","period":"Q2 2026, issues that begin with the AI assistant","claimant":"organization","quote":"Nearly 45 percent of issues that begin with our AI assistant are now resolved without a human agent, up from Q1, while delivering much faster resolution times.","sourceUrl":"https://news.airbnb.com/airbnb-q2-2026-financial-results/"},{"kpi":"cost-reduction","value":16,"unit":"percent","qualifier":"approximately","period":"Q2 2026 year on year, customer support cost per booking","baseline":"Q2 2025","claimant":"organization","quote":"In Q2, our customer support related cost per booking declined approximately 16 percent year-over-year, driven in part by improvements to our AI assistant.","sourceUrl":"https://news.airbnb.com/airbnb-q2-2026-financial-results/"}],"outcomeDisclosed":true,"sources":[{"url":"https://news.airbnb.com/airbnb-q2-2026-financial-results/","title":"Airbnb Q2 2026 financial results","publisher":"Airbnb","date":"2026-08-06"},{"url":"https://news.airbnb.com/airbnb-q1-2026-financial-results/","title":"Airbnb Q1 2026 financial results","publisher":"Airbnb","date":"2026-05-07"},{"url":"https://news.airbnb.com/airbnb-q2-2025-financial-results/","title":"Airbnb Q2 2025 financial results","publisher":"Airbnb","date":"2025-08-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"airbnb-ai-customer-support-assistant"},{"title":"Airbnb: LLM driven migration of about 3,500 test files finished in 6 weeks instead of an estimated 1.5 years","useCases":["legacy-code-modernization"],"organization":{"name":"Airbnb","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Airbnb","role":"in-house"}],"summary":"Airbnb migrated around 3,500 React test files from Enzyme to React Testing Library using a combination of frontier language models and automation. Airbnb had originally estimated the migration would take about 1.5 years of engineering time to do by hand, and instead completed it in 6 weeks.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://medium.com/airbnb-engineering/accelerating-large-scale-test-migration-with-llms-9565c208023b","title":"Accelerating large scale test migration with LLMs","publisher":"Airbnb Engineering","date":"2025-01-01","archivedUrl":"https://web.archive.org/web/2026/https://medium.com/airbnb-engineering/accelerating-large-scale-test-migration-with-llms-9565c208023b"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"airbnb-test-migration-llm"},{"title":"Airwallex: generative AI website screening in business onboarding","useCases":["merchant-underwriting-and-risk-monitoring"],"organization":{"name":"Airwallex","anonymized":false,"region":"global","industry":"payments"},"vendors":[{"name":"Airwallex","role":"in-house"}],"summary":"Airwallex, a global payments and financial platform for businesses, announced in December 2023 early results from a generative AI tool it uses in its know your customer and onboarding process. The tool scans new customers' websites to check what they really sell and whether it fits Airwallex's acceptable use policies. Compared with its earlier rules based and NLP scanner, Airwallex says the model can better distinguish, for example, a retailer selling a military style jacket from a merchant selling prohibited military goods. Early results from Airwallex's internal analysis show 50 percent fewer false positives on average and 20 percent more customers passing through onboarding without human intervention.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"false-positive-reduction","value":50,"unit":"percent","qualifier":"exact","period":"early results, website screening in onboarding, against the earlier rules based scanner; Airwallex's footnote defines the reduction as fewer customers whose websites need a manual check for red flags","baseline":"legacy rules based and NLP website scanner","claimant":"organization","quote":"The new tool reduces ‘false positives’ by 50 percent on average, while boosting the number of customers that pass through the onboarding process without human intervention by 20 percent [1].","sourceUrl":"https://www.airwallex.com/newsroom/airwallex-improves-customer-onboarding-with-generative-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.airwallex.com/newsroom/airwallex-improves-customer-onboarding-with-generative-ai","title":"Airwallex improves customer onboarding with generative AI","publisher":"Airwallex","date":"2023-12-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"airwallex-generative-ai-website-screening"},{"title":"AJ Bell: machine learning customer screening with ComplyAdvantage","useCases":["sanctions-screening-adjudication"],"organization":{"name":"AJ Bell","anonymized":false,"country":"GB","region":"europe","industry":"wealth-and-asset-management"},"vendors":[{"name":"ComplyAdvantage","role":"platform"}],"summary":"UK investment platform AJ Bell uses ComplyAdvantage's AI powered Customer Screening and Ongoing Monitoring, whose matching combines fuzzy logic with machine learning that learns aliases and global naming conventions. AJ Bell's Head of Financial Crime and MLRO says that optimising the system's settings cut alert volume by 82 percent, which lets analysts clear alerts faster and focus on the highest risk areas.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"alert-volume-reduction","value":82,"unit":"percent","qualifier":"exact","period":"customer screening alert volume","claimant":"organization","quote":"Through optimizing the levers within the system, we’ve been able to reduce our alert volume by 82 percent.","sourceUrl":"https://complyadvantage.com/customer-stories/aj-bell-reduces-alert-volumes-by-82-percent/"}],"outcomeDisclosed":true,"sources":[{"url":"https://complyadvantage.com/customer-stories/aj-bell-reduces-alert-volumes-by-82-percent/","title":"AJ Bell Reduces Alert Volumes by 82% with AI-Powered Customer Screening","publisher":"ComplyAdvantage","date":"2025-08-05","archivedUrl":"https://web.archive.org/web/20251204194052/https://complyadvantage.com/customer-stories/aj-bell-reduces-alert-volumes-by-82-percent/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"aj-bell-complyadvantage-customer-screening"},{"title":"Akamai: Cast AI cuts Kubernetes cloud costs 40 to 70%","useCases":["cloud-cost-optimization-agent"],"organization":{"name":"Akamai Technologies","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Cast AI","role":"platform"}],"summary":"Akamai runs a large, complex Kubernetes estate on Microsoft Azure (AKS) behind services with strict SLAs. It uses Cast AI's automated bin packing, selection of the most cost efficient compute instances, Spot instance lifecycle automation and Kubernetes cost analytics to optimize the cost of its core production infrastructure. Akamai reports savings of 40 to 70% depending on the workload, its engineering team reports reclaiming time previously spent manually tuning capacity, and the interviewee describes the value of being able to \"turn on and forget\" the platform.","stage":"production","year":2024,"channels":["api"],"languages":[],"metrics":[{"kpi":"cost-reduction","value":40,"unit":"percent","qualifier":"at-least","period":"ongoing","baseline":"Compute cost before Cast AI, varies by workload (40 to 70% range reported)","claimant":"organization","quote":"The core savings we got are just brilliant, falling between 40-70%, depending on the workload.","sourceUrl":"https://cast.ai/case-studies/akamai/"}],"outcomeDisclosed":true,"sources":[{"url":"https://cast.ai/case-studies/akamai/","title":"Akamai Kubernetes Optimization Case Study","publisher":"Cast AI","archivedUrl":"https://web.archive.org/web/20250219143453/https://cast.ai/case-studies/akamai/"},{"url":"https://cast.ai/","title":"Kubernetes Optimization Platform for Performance - Cast AI","publisher":"Cast AI","archivedUrl":"https://web.archive.org/web/20260822010142/https://cast.ai/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"akamai-kubernetes-cost-optimization"},{"title":"Albert Heijn: daily AI demand forecasts for more than 15 million store and item combinations","useCases":["retail-demand-forecasting-and-replenishment"],"organization":{"name":"Albert Heijn","anonymized":false,"country":"NL","region":"europe","industry":"retail-and-ecommerce"},"vendors":[],"summary":"Albert Heijn, the leading supermarket chain in the Netherlands with more than 1,200 stores, forecasts daily sales for more than 15 million store and item combinations more than five weeks ahead, which its VP of Product Operations describes as almost one billion predictions a day, with the aim of bringing just enough stock to each store and cutting food waste. The retailer says it hopes forecasting will cut its total food waste by more than 10%, a target rather than a result. A related Dynamic Markdown initiative raises discounts on electronic shelf labels during the day for products close to their date, and the story says it now saves 250,000 kilos of wasted food per year.","stage":"scaled","year":2024,"channels":["api"],"languages":["nl"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1739352737304784739-albertheijn-azure-open-ai-service-retailers-en-netherlands","title":"How Albert Heijn is using Azure OpenAI to reduce food waste and help its customers eat healthy","publisher":"Microsoft","archivedUrl":"https://web.archive.org/web/20241224222112/https://www.microsoft.com/en/customers/story/1739352737304784739-albertheijn-azure-open-ai-service-retailers-en-netherlands"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"albert-heijn-ai-demand-forecasting"},{"title":"Albo: Albot AI chatbot for onboarding and support of first time bank users","useCases":["digital-onboarding-assistant"],"organization":{"name":"Albo","anonymized":false,"country":"MX","region":"latin-america","industry":"banking"},"vendors":[{"name":"Google","role":"model-provider"}],"summary":"Albo, a Mexican neobank, uses Gemini models in Albot, a chatbot that handles customer onboarding and offers support and financial guidance around the clock to millions of first time banking users. Google Cloud lists the deployment and says it supports financial inclusion and regulatory compliance; no outcome figures are disclosed.","stage":"production","year":2025,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"albo-albot-ai-chatbot"},{"title":"Allen & Overy (now A&O Shearman): Harvey generative AI assistant for lawyers across 43 offices","useCases":["legal-research-and-drafting-assistant"],"organization":{"name":"A&O Shearman","anonymized":false,"country":"GB","region":"global","industry":"professional-services"},"vendors":[{"name":"Harvey","role":"platform"},{"name":"OpenAI","role":"model-provider"}],"summary":"Allen & Overy, which merged with Shearman & Sterling in 2024 to form A&O Shearman, trialled Harvey, a generative AI platform built on OpenAI models and adapted for legal work, from November 2022 and in February 2023 integrated it into its global practice for more than 3,500 lawyers in 43 offices. The release describes Harvey as supporting work such as contract analysis, due diligence, litigation and regulatory compliance, and a quoted A&O executive says it can work in multiple languages; it does not say which of these tasks A&O lawyers use it for. The firm states that the output needs careful review by an A&O lawyer. The published figures describe usage in the trial, not time saved.","stage":"scaled","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":3500,"unit":"count","qualifier":"approximately","period":"beta trial, November 2022 to February 2023","claimant":"organization","quote":"At the end of the trial, around 3500 of A&O’s lawyers had asked Harvey around 40,000 queries for their day-to-day client work.","sourceUrl":"https://www.aoshearman.com/en/news/ao-announces-exclusive-launch-partnership-with-harvey"},{"kpi":"interactions-handled","value":40000,"unit":"count","qualifier":"approximately","period":"beta trial, November 2022 to February 2023","claimant":"organization","quote":"At the end of the trial, around 3500 of A&O’s lawyers had asked Harvey around 40,000 queries for their day-to-day client work.","sourceUrl":"https://www.aoshearman.com/en/news/ao-announces-exclusive-launch-partnership-with-harvey"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.aoshearman.com/en/news/ao-announces-exclusive-launch-partnership-with-harvey","title":"A&O announces exclusive launch partnership with Harvey","publisher":"A&O Shearman","date":"2023-02-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"allen-and-overy-harvey-legal-assistant"},{"title":"Allianz Partners: AI assisted travel claims turnaround","useCases":["travel-insurance-claims-and-assistance-agent","claims-triage-and-straight-through-processing"],"organization":{"name":"Allianz Partners","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[],"summary":"Allianz Partners' director of partnerships told a US travel advisor consortium in April 2026 that AI assistance now handles 65 to 70% of all its claims. She said this has cut claims turnaround time from about 14 days, the average for the consortium's member agencies, to three to four days, with some claims turned around in a matter of hours.","stage":"scaled","year":2026,"channels":[],"languages":[],"metrics":[{"kpi":"automation-rate","value":70,"unit":"percent","qualifier":"up-to","period":"Share of claims using AI assistance, stated at Signature Travel Network's Horizon Club, April 2026","claimant":"organization","quote":"We are right now using AI assistance for 65 to 70% of all of our claims, which has brought our claims turnaround time down from about 14 days, which was the average for Signature partners, to three to four days,","sourceUrl":"https://latteluxurynews.com/2026/04/20/ai-cuts-down-allianz-travel-insurance-claim-time-by-more-than-half/"},{"kpi":"cycle-time-days","value":4,"unit":"days","qualifier":"up-to","period":"Claims turnaround time with AI assistance, for Signature Travel Network member agencies","baseline":"About 14 days before AI assistance, the average for Signature Travel Network partners","claimant":"organization","quote":"We are right now using AI assistance for 65 to 70% of all of our claims, which has brought our claims turnaround time down from about 14 days, which was the average for Signature partners, to three to four days,","sourceUrl":"https://latteluxurynews.com/2026/04/20/ai-cuts-down-allianz-travel-insurance-claim-time-by-more-than-half/"}],"outcomeDisclosed":true,"sources":[{"url":"https://latteluxurynews.com/2026/04/20/ai-cuts-down-allianz-travel-insurance-claim-time-by-more-than-half/","title":"AI cuts down Allianz travel insurance claim time by more than half","publisher":"LATTE Australia","date":"2026-04-20"},{"url":"https://completeaitraining.com/news/allianz-cuts-travel-insurance-claim-processing-time-from-14/","title":"Allianz cuts travel insurance claim processing time from 14 days to 3 to 4 days using AI","publisher":"Complete AI Training","date":"2026-04-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"allianz-partners-ai-claims-turnaround"},{"title":"Ally Financial: generative AI for marketing content with regulatory review","useCases":["marketing-content-compliance-copilot"],"organization":{"name":"Ally Financial","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Ally","role":"in-house"}],"summary":"Ally ran a month long experiment in which a group of marketers used its in house Ally.ai platform, built on enterprise large language models, for tasks such as research, naming and first drafts of advertising copy, video scripts and social posts. In one example, an AI first draft of a blog article cut the time to create and edit it from four hours to one; the article was edited by Ally's content writers and still went through the bank's established regulatory review. Ally reported an average time saving, and says the largest reductions, of up to two to three weeks, came in early stages of the creative process such as research, first drafts and naming.","stage":"pilot","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":34,"unit":"percent","qualifier":"exact","period":"month long experiment","baseline":"time to produce creative campaigns and content without AI","claimant":"organization","quote":"Using the Ally.ai platform's large language model (LLM) chat and prompt functionality, a select group of marketers were able to reduce the time needed to produce creative campaigns and content by up to 2-3 weeks and reported an average time savings of 34%, compared to typical processes without AI.","sourceUrl":"https://media.ally.com/2023-11-16-Do-It-Right-with-AI-Ally-creators-experiment-with-generative-AI-in-marketing-test-case"}],"outcomeDisclosed":true,"sources":[{"url":"https://media.ally.com/2023-11-16-Do-It-Right-with-AI-Ally-creators-experiment-with-generative-AI-in-marketing-test-case","title":"'Do It Right' with AI: Ally creators experiment with generative AI in marketing test case","publisher":"Ally Financial","date":"2023-11-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"ally-financial-generative-ai-marketing-content"},{"title":"Amazon: generative AI tools that write product listings for selling partners","useCases":["product-content-and-catalog-enrichment","personalized-marketing-at-scale"],"organization":{"name":"Amazon","anonymized":false,"country":"US","region":"global","industry":"retail-and-ecommerce"},"vendors":[{"name":"Amazon Web Services","role":"platform"}],"summary":"Since the end of 2023 Amazon lets independent sellers create a product listing from a few words or a single image, and since March 2024 also from the URL of their own web page: generative AI on Amazon Bedrock drafts the title, bullet points, description and attributes, and the seller submits the draft, with Amazon encouraging a review first. Bulk creation from a spreadsheet followed, and Enhance My Listing, which Amazon said in May 2025 had begun rolling out in the US, suggests updates to existing listings based on shopping behaviour. In May 2025 Amazon reported that sellers accept the AI generated content with little to no edits about 90% of the time. The same post describes Amazon using generative AI to personalize product recommendation categories and product descriptions shown to customers on the website and in the shopping app, based on a customer's shopping activity; no outcome number is given for that side of the work.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":900000,"unit":"count","qualifier":"at-least","period":"selling partners who have used the listing tools, by May 2025","claimant":"organization","quote":"Now, more than 900,000 Amazon selling partners have embraced these tools, with sellers accepting AI-generated content with little to no edits approximately 90% of the time.","sourceUrl":"https://www.aboutamazon.com/news/innovation-at-amazon/amazon-generative-ai-seller-growth-shopping-experience"},{"kpi":"quality-score-uplift","value":40,"unit":"percent","qualifier":"exact","period":"overall listing quality of listings created with the tools, reported by May 2025","claimant":"organization","quote":"When sellers use our Gen AI tools to create listings, they see a 40% increase in overall listing quality, helping them create content that enhances customer engagement and boosts sales potential.","sourceUrl":"https://www.aboutamazon.com/news/innovation-at-amazon/amazon-generative-ai-seller-growth-shopping-experience"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.aboutamazon.com/news/innovation-at-amazon/amazon-generative-ai-seller-growth-shopping-experience","title":"Amazon sellers can now automatically improve product listings with our new Gen AI tool","publisher":"Amazon","date":"2025-05-08"},{"url":"https://www.aboutamazon.com/news/innovation-at-amazon/amazon-generative-ai-powered-product-listings","title":"Amazon selling partners can now access even more generative AI features to create high-quality product listings","publisher":"Amazon","date":"2024-03-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"amazon-generative-ai-listing-tools"},{"title":"Amazon: a code transformation agent helped migrate tens of thousands of production applications to Java 17","useCases":["legacy-code-modernization"],"organization":{"name":"Amazon","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Amazon","role":"in-house"}],"summary":"Amazon integrated the Java transformation capability of Amazon Q Developer into its internal systems and migrated tens of thousands of production applications from Java 8 or 11 to Java 17 with its assistance. AWS describes the product's transformation agents as analyzing source code, generating new code, testing it and executing the change once the customer approves; the sources do not describe how Amazon's own developers reviewed each upgrade. Amazon estimates more than 4,500 years of development work saved compared with manual upgrades, and annual savings from hosts it could remove after the faster Java 17 runtime. AWS has since extended the approach to .NET, VMware and mainframe workloads.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"cost-savings","value":260000000,"unit":"currency","currency":"USD","qualifier":"exact","period":"per year, estimated from hosts removed after the Java 17 upgrade","claimant":"organization","quote":"This effort saved more than 4,500 years of development work, compared to what it would have taken previously, and realized performance improvements of $260 million in annual cost savings.","sourceUrl":"https://press.aboutamazon.com/2024/12/new-amazon-q-developer-capabilities-accelerate-large-scale-transformations-of-legacy-workloads"}],"outcomeDisclosed":true,"sources":[{"url":"https://press.aboutamazon.com/2024/12/new-amazon-q-developer-capabilities-accelerate-large-scale-transformations-of-legacy-workloads","title":"New Amazon Q Developer Capabilities Accelerate Large-Scale Transformations of Legacy Workloads","publisher":"Amazon","date":"2024-12-03"},{"url":"https://aws.amazon.com/blogs/devops/amazon-q-developer-just-reached-a-260-million-dollar-milestone","title":"Amazon Q Developer just reached a $260 million dollar milestone","publisher":"AWS","date":"2024-08-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"amazon-java-upgrade-code-transformation"},{"title":"Amazon: Rufus and Alexa for Shopping conversational shopping assistant","useCases":["conversational-shopping-assistant"],"organization":{"name":"Amazon","anonymized":false,"country":"US","region":"global","industry":"retail-and-ecommerce"},"vendors":[{"name":"Amazon","role":"in-house"}],"summary":"Amazon launched Rufus in 2024 as a generative AI shopping assistant in its app, trained on the product catalog, customer reviews, community questions and answers and information from the web, to answer product questions, compare items and recommend products for an occasion or need. In 2026 it combined Rufus and Alexa+ into Alexa for Shopping, an agentic assistant that adds price history, price alerts and automated buying. Amazon reports over 350 million customers in twelve months, and says US customers who use Alexa for Shopping spend on average over 40% more per order than those who do not, a comparison between self selected groups rather than a controlled test.","stage":"scaled","year":2024,"channels":["mobile-app","web-chat"],"languages":[],"metrics":[{"kpi":"users-served","value":350000000,"unit":"count","qualifier":"at-least","period":"the 12 months to Q2 2026","claimant":"organization","quote":"Over 350 million customers have used it in the last 12 months, and engagement accelerated in Q2, with active users nearly doubling, and interactions up over 5x year-over-year.","sourceUrl":"https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-stores-growth-ai-shopping-q2-2026-earnings"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-stores-growth-ai-shopping-q2-2026-earnings","title":"Q2 earnings: CEO Andy Jassy on Amazon Stores growth, delivery speed, and AI shopping","publisher":"Amazon","date":"2026-07-31"},{"url":"https://www.aboutamazon.com/news/company-news/amazon-earnings-q2-2026-report","title":"Amazon Q2 2026 earnings report: Read the release","publisher":"Amazon","date":"2026-07-30"},{"url":"https://www.aboutamazon.com/news/retail/amazon-rufus","title":"Amazon Rufus AI experience comes to the Amazon Shopping app","publisher":"Amazon","date":"2024-02-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"amazon-rufus-and-alexa-for-shopping"},{"title":"American Addiction Centers: generative AI to shorten employee onboarding","useCases":["employee-onboarding-assistant"],"organization":{"name":"American Addiction Centers","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Google","role":"platform"}],"summary":"American Addiction Centers, a provider of addiction treatment, cut employee onboarding from three days to 12 hours with Gemini for Google Workspace. Its CIO called Gemini \"a key driver\" of that reduction in a Google Workspace recap of Google Cloud Next 2024, and Google Cloud repeats the figure in its list of customer use cases. The sources name a general productivity suite, not an assistant that guides new hires or runs the onboarding checklist, and they do not describe how the process was changed or whether the three days were working or calendar days.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"cycle-time-days","value":12,"unit":"hours","qualifier":"exact","baseline":"3 days of employee onboarding before","claimant":"organization","quote":"Gemini for Workspace was a key driver in reducing employee onboarding from 3 days to 12 hours.","sourceUrl":"https://workspace.google.com/blog/events/cloud-next-recap-20-ways-our-customers-outdo-themselves"}],"outcomeDisclosed":true,"sources":[{"url":"https://workspace.google.com/blog/events/cloud-next-recap-20-ways-our-customers-outdo-themselves","title":"How 20 of our customers outdo themselves with Google Workspace","publisher":"Google Workspace","date":"2024-04-16"},{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","archivedUrl":"https://web.archive.org/web/20241226092809/https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"american-addiction-centers-employee-onboarding"},{"title":"American Airlines: HEAT, a machine learning tool that reshapes hub schedules ahead of storms","useCases":["airline-operations-control-decision-support"],"organization":{"name":"American Airlines","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[{"name":"American Airlines","role":"in-house"}],"summary":"American Airlines built the Hub Efficiency Analytics Tool (HEAT) in house and has used it at its hubs since spring 2022. When severe weather is forecast, it weighs weather, load factors, customer connections, gate availability, air traffic control and crew constraints and shifts the departure and arrival times of flights at the hub, which operations center coordinators decide whether to apply. In a newsroom article from 2023, American says that since its initial deployment the year before, HEAT has prevented nearly 1,000 flight cancellations across its network. The same program includes intelligent gating at Dallas Fort Worth, which assigns the nearest available gate to arriving aircraft automatically.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://news.aa.com/news/news-details/2023/Meet-HEAT-Americans-tool-to-manage-through-summer-storms-OPS-OTH-07/default.aspx","title":"Meet HEAT: American's tool to manage through summer storms","publisher":"American Airlines Newsroom","archivedUrl":"https://web.archive.org/web/2026/https://news.aa.com/news/news-details/2023/Meet-HEAT-Americans-tool-to-manage-through-summer-storms-OPS-OTH-07/default.aspx"},{"url":"https://www.cio.com/article/406357/american-airlines-takes-flight-with-analytics-transformation.html","title":"American Airlines takes flight with analytics transformation","publisher":"CIO","date":"2022-09-07"},{"url":"https://crankyflier.com/2022/06/06/american-turns-on-heat-to-reduce-disruptions/","title":"American Turns on HEAT to Reduce Disruptions","publisher":"Cranky Flier","date":"2022-06-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"american-airlines-heat-hub-disruption-tool"},{"title":"Amtrak: AI cash forecasting to invest idle balances","useCases":["treasury-cash-flow-forecasting"],"organization":{"name":"Amtrak","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[{"name":"JPMorgan Chase","role":"platform"}],"summary":"Amtrak's treasury, which has to plan around the three large installments in which government funding arrives each year, went live with J.P. Morgan Payments Cash Flow Intelligence in May 2023 after testing it in beta. Separating daily card receipts, monthly partner receipts and infrequent federal receipts into their own categories improved projection accuracy, which let the team invest balances it had previously set aside as a buffer. No figures are published for the accuracy gain or the income.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.jpmorgan.com/insights/payments/data-intelligence/amtrak-cash-flow-forecasting","title":"Amtrak enhances cash forecasting using Cash Flow Intelligence","publisher":"J.P. Morgan","date":"2025-05-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"amtrak-cash-flow-intelligence"},{"title":"Ancine: AI extraction from digitized tax documents for accountability analysis","useCases":["intelligent-document-processing"],"organization":{"name":"Ancine","anonymized":false,"country":"BR","region":"latin-america","industry":"government"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Ancine, which Google Cloud describes as the Brazilian cinema industry regulator, uses Google Cloud AI to extract and structure data from digitized tax documents to automate the accountability analysis of subsidized projects. Google Cloud reports extraction accuracy above 90% and a tenfold increase in analysts' daily processing capacity.","stage":"production","year":2026,"channels":["api"],"languages":["pt"],"metrics":[{"kpi":"accuracy","value":90,"unit":"percent","qualifier":"at-least","period":"data extraction accuracy","claimant":"vendor","quote":"This AI implementation achieved over 90% data extraction accuracy, boosting analysts' daily processing capacity by 10x.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"},{"kpi":"productivity-gain","value":10,"unit":"multiplier","qualifier":"exact","period":"analysts' daily processing capacity","claimant":"vendor","quote":"This AI implementation achieved over 90% data extraction accuracy, boosting analysts' daily processing capacity by 10x.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"ancine-tax-document-extraction"},{"title":"ANS: agents that gather account context so sellers prioritize the right accounts","useCases":["outbound-sales-prospecting-agent"],"organization":{"name":"ANS","anonymized":false,"country":"GB","region":"europe","industry":"technology"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"ANS, a UK cloud, security and digital technology provider and Microsoft partner, uses Microsoft Copilot and agents in its selling process. Sellers ask an agent to gather and summarize information from several data sources, including past customer interactions, so they can decide which accounts and opportunities to focus on. Microsoft reports an expected improvement in closing ratio, which is a forecast and is not recorded as a result.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft","date":"2025-07-24"},{"url":"https://www.ans.co.uk/","title":"ANS: Leading Cloud, Security & Digital Technology Provider for AI","publisher":"ANS"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"ans-copilot-seller-account-agent"},{"title":"ANZ Bank: controlled experiment and rollout of GitHub Copilot to engineers","useCases":["developer-coding-assistant"],"organization":{"name":"ANZ","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"GitHub","role":"platform"}],"summary":"ANZ, which employs over 5,000 engineers, ran a six week experiment with GitHub Copilot from June to July 2023: after two weeks of preparation, over 100 participants were split at random into a control group and a Copilot group that solved the same algorithmic Python challenges. The Copilot group took markedly less time and produced code with fewer code smells and bugs, while the effect on security was inconclusive. By the time of writing about 1,000 engineers were using Copilot. The study was written by ANZ staff and published on arXiv.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":42.36,"unit":"percent","qualifier":"exact","period":"two week A/B test on algorithmic Python challenges, self reported time","baseline":"Mean time per challenge in the control group without Copilot","claimant":"organization","quote":"This study shows that Copilot improves the productivity of ANZ engineers by 42.36% on an average.","sourceUrl":"https://arxiv.org/pdf/2402.05636"}],"outcomeDisclosed":true,"sources":[{"url":"https://arxiv.org/abs/2402.05636","title":"The Impact of AI Tool on Engineering at ANZ Bank An Empirical Study on GitHub Copilot within Corporate Environment","publisher":"arXiv","date":"2024-02-08"},{"url":"https://arxiv.org/pdf/2402.05636","title":"The Impact of AI Tool on Engineering at ANZ Bank (full paper, PDF)","publisher":"arXiv"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"anz-github-copilot-study"},{"title":"ANZ, HSBC and Lloyds with Microsoft: AI agent proof of concept for letter of credit data","useCases":["trade-document-examination","trade-finance-crime-screening"],"organization":{"name":"ANZ, HSBC and Lloyds Banking Group","anonymized":false,"region":"global","industry":"banking"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Microsoft built a proof of concept with ANZ, HSBC and Lloyds, shown at Sibos 2025, in which an AI agent embedded in a corporate's ERP parses an incoming MT700 letter of credit, cross checks it against invoice and shipping data, flags discrepancies such as currency and amount, and sends structured data aligned to the ICC Key Trade Documents and Data Elements to the bank. The same agent answers treasury questions about compliance with the credit terms, and Microsoft says such agents can help flag references to sanctioned entities or ambiguous dual use goods descriptions. It is a demonstration, not a live service.","stage":"announced","year":2025,"channels":["internal-tools","api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/financial-services/2026/04/20/reimagining-trade-finance-with-ai-a-collaborative-proof-of-concept-from-microsoft-anz-hsbc-and-lloyds/","title":"Reimagining trade finance with AI: A collaborative proof of concept from Microsoft, ANZ, HSBC, and Lloyds","publisher":"Microsoft","date":"2026-04-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"anz-hsbc-lloyds-trade-finance-agent-proof-of-concept"},{"title":"Arch Capital: AI that brings past experience and submission data to underwriting decisions","useCases":["underwriting-risk-assessment-copilot"],"organization":{"name":"Arch Capital Group","anonymized":false,"country":"BM","region":"global","industry":"insurance"},"vendors":[{"name":"Arch Capital Group","role":"in-house"}],"summary":"Arch's 2024 annual report says it uses AI for catastrophe modelling and predictive analytics and, in its insurance operations, to provide more information about past experiences and submissions so that its professionals can make more data driven underwriting decisions. Every new generative AI technology proposed for use in its operations requires approval and is monitored closely. No outcome figures are disclosed.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/947484/000094748425000017/acgl-20241231.htm","title":"Arch Capital Group Ltd. Form 10-K for 2024","publisher":"Arch Capital Group via SEC EDGAR","date":"2025-02-27"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"arch-capital-ai-underwriting-insights"},{"title":"Art Basel: AI companion and Art Basel Lens artwork recognition for fair visitors","useCases":["ai-visitor-and-tour-guide"],"organization":{"name":"Art Basel","anonymized":false,"country":"CH","region":"global","industry":"media-and-entertainment"},"vendors":[{"name":"Microsoft (Azure AI Foundry, Azure OpenAI)","role":"platform"},{"name":"UIC Digital","role":"integrator"},{"name":"Valorem Reply","role":"integrator"}],"summary":"Art Basel's Companion app combines a conversational AI companion, grounded in gallery, artwork, dining and lodging information for its fairs in five cities, with the Art Basel Lens: a visitor photographs an artwork and receives artist and gallery details in about two seconds, matched by image embeddings against a vector index. Art Basel reports more engagement, return visits and time in the app, without publishing figures, and is exploring wayfinding and restaurant bookings.","stage":"production","year":2025,"channels":["mobile-app"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/24324-art-basel-azure-ai-foundry","title":"Art Basel bridges physical and digital worlds, drives engagement with Azure AI Foundry","publisher":"Microsoft Customer Stories","date":"2025-06-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"art-basel-companion-app-and-lens"},{"title":"Ashurst (now Ashurst Perkins Coie): firm wide trials of generative AI for legal drafting and research","useCases":["legal-research-and-drafting-assistant"],"organization":{"name":"Ashurst Perkins Coie","anonymized":false,"country":"GB","region":"global","industry":"professional-services"},"vendors":[],"summary":"Ashurst, which has since combined with Perkins Coie and now operates as Ashurst Perkins Coie, ran three global generative AI trials from November 2023 to March 2024 with 411 partners, lawyers and staff in 23 offices and 14 countries, and published the results in its report Vox PopulAI. In controlled experiments with small groups within those trials, it measured approximate time savings of 45% on first draft legal briefings, 59% on sector research reports and 80% on UK corporate filings that required extracting information from articles of association. In a blind study, an expert panel correctly identified all lawyer written outputs, but half of the AI generated outputs were either mistaken for human work or could not be classified. The trials used only publicly available data, not client data.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":45,"unit":"percent","qualifier":"approximately","period":"controlled experiments, November 2023 to March 2024","baseline":"time to create a first draft legal briefing","claimant":"organization","quote":"In our controlled experiments, we measured approximate time savings of 80% to draft UK corporate filings requiring review and extraction of information from company articles of association, 59% to draft industry/sector-specific research reports that require reviewing and extracting key information from public company filings (Form 10-Ks), and 45% on creating first draft legal briefings.","sourceUrl":"https://www.ashurstperkinscoie.com/en/insights/vox-populai-lessons-from-a-global-law-firms-exploration-of-generative-ai/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.ashurstperkinscoie.com/en/insights/vox-populai-lessons-from-a-global-law-firms-exploration-of-generative-ai/","title":"Vox PopulAI: Lessons from a global law firm's exploration of generative AI","publisher":"Ashurst Perkins Coie","date":"2024-06-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ashurst-generative-ai-legal-drafting-trials"},{"title":"Asset Living: EliseAI for leasing, collections, maintenance and renewals","useCases":["apartment-leasing-and-resident-service-agent"],"organization":{"name":"Asset Living","anonymized":false,"country":"US","region":"north-america","industry":"real-estate"},"vendors":[{"name":"EliseAI","role":"platform"}],"summary":"Asset Living, a Houston based property manager with more than 450,000 units, uses EliseAI's leasing, delinquency, maintenance, renewals and voice products to automate routine prospect and resident communications around the clock. In a joint announcement in September 2025 it reported more than 130,000 personalized payment reminders in the second quarter of 2025, a 600 basis point increase in on time rent payments, a 300 basis point increase in occupancy and 78.2 hours of incremental staff capacity per community per month. It is piloting AI guided tours and lease audits.","stage":"scaled","year":2025,"channels":["voice"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":130000,"unit":"count","qualifier":"at-least","period":"second quarter of 2025, personalized payment reminders sent","claimant":"organization","quote":"600 bps increase in on-time rent payments, enabled by over 130,000 personalized payment reminders in Q2 2025.","sourceUrl":"https://www.assetliving.com/blogs/asset-living-advances-operational-excellence-with-eliseai-partnership"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.assetliving.com/blogs/asset-living-advances-operational-excellence-with-eliseai-partnership","title":"Asset Living Advances Operational Excellence with EliseAI Partnership","publisher":"Asset Living","date":"2025-09-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"asset-living-eliseai-leasing-and-resident-communications"},{"title":"Assurant: AI claims fraud detection for the special investigations unit","useCases":["claims-fraud-detection"],"organization":{"name":"Assurant","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Shift Technology","role":"platform"}],"summary":"Assurant has used Shift Claims Fraud Detection since 2018 to flag suspicious claims to its special investigations unit, after a trial in which the system identified dozens of fraud cases. Alerts carry the context investigators need, the investigation software was adapted to the unit's workflow, and the models were refined for schemes involving specialty vehicles, extreme weather and multiple claims. Shift reports that the unit's case acceptance rate rose over four years, leading to more fraud mitigated, but gives no figures.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.shift-technology.com/resources/case-studies/assurants-focus-on-innovation-increases-stopped-fraud","title":"Assurant's Focus on Innovation Increases Stopped Fraud","publisher":"Shift Technology","date":"2022-12-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"assurant-claims-fraud-detection"},{"title":"Ateme: automated multilingual subtitle generation for broadcasters and streaming platforms","useCases":["audio-and-video-transcription-and-captioning"],"organization":{"name":"Ateme","anonymized":false,"country":"FR","region":"europe","industry":"media-and-entertainment"},"vendors":[{"name":"Google Cloud (Vertex AI, Gemini)","role":"platform"}],"summary":"Ateme, a French video compression and delivery company, added a subtitle step to its file transcoding platform: after transcoding, Gemini on Vertex AI transcribes the audio, spots timecodes and generates subtitles in the requested languages, and a script converts them to SRT for the client's workflow. Ateme says a job that took up to 15 hours of manual work per hour of video now takes minutes and costs less than a dollar per hour of content.","stage":"production","year":2025,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/ateme","title":"Ateme customer story | Google Cloud","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ateme-multilingual-subtitle-generation"},{"title":"Atlanticus: cash flow underwriting for marginal credit card declines","useCases":["alternative-data-credit-scoring"],"organization":{"name":"Atlanticus","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Nova Credit","role":"platform"},{"name":"Engine by Gen","role":"integrator"}],"summary":"Atlanticus, which runs credit card brands through bank partners for consumers overlooked by prime lenders, added consumer permissioned bank transaction data (Nova Credit Cash Atlas) to its decisions, with consent collected inside an embedded finance marketplace. The vendor reports that 15% of marginal declines, applicants that bureau data alone would have rejected, could be approved profitably with cash flow insights, with no deterioration in credit quality, and that the data is also used to set line sizes and pricing.","stage":"production","year":2025,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.novacredit.com/corporate-blog/case-study-how-atlanticus-unlocked-15-more-approvals-while-maintaining-risk","title":"Case Study: How Atlanticus Unlocked 15% More Approvals While Maintaining Risk Standards","publisher":"Nova Credit","archivedUrl":"https://web.archive.org/web/20260124014124/https://www.novacredit.com/corporate-blog/case-study-how-atlanticus-unlocked-15-more-approvals-while-maintaining-risk"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"atlanticus-cash-flow-underwriting"},{"title":"Audi: AI quality control of resistance spot welds in car body construction","useCases":["production-line-quality-inspection"],"organization":{"name":"Audi","anonymized":false,"country":"DE","region":"europe","industry":"automotive"},"vendors":[{"name":"Audi","role":"in-house"}],"summary":"After a successful pilot at its Neckarsulm site, Audi began rolling out an AI system for quality control of resistance spot welds in car body construction (project WPS Analytics). The AI analyses around 1.5 million spot welds on 300 vehicles each shift, where production staff used to check around 5,000 spot welds per vehicle with ultrasound based on random analyses, so employees can now focus on possible anomalies. Audi developed the process with the German Association for Quality (DGQ) and two Fraunhofer institutes so that it would hold up in audits and certification. It began installing the infrastructure at Audi Brussels, with Ingolstadt and the Volkswagen plant in Emden scheduled to follow, and is retraining the model for differences in weld settings.","stage":"production","year":2023,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"interactions-handled","value":1500000,"unit":"count","qualifier":"approximately","period":"spot welds analysed per shift at Neckarsulm, on 300 vehicles","claimant":"organization","quote":"Using artificial intelligence, Audi analyzes around 1.5 million spot welds on 300 vehicles each shift at its Neckarsulm site.","sourceUrl":"https://www.audi.com/en/press-releases/audi-begins-roll-out-of-artificial-intelligence-for-quality-control-of-spot-welds-15443"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.audi.com/en/press-releases/audi-begins-roll-out-of-artificial-intelligence-for-quality-control-of-spot-welds-15443","title":"Audi begins roll-out of artificial intelligence for quality control of spot welds","publisher":"Audi","date":"2023-06-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"audi-spot-weld-quality-analytics"},{"title":"Audibel: voice agent that captures and routes appointment calls for 400 hearing clinics","useCases":["patient-appointment-scheduling-and-reminders-agent"],"organization":{"name":"Audibel","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"PolyAI","role":"platform"}],"summary":"Audibel, a network of more than 400 hearing care clinics in the United States, put a PolyAI voice agent in front of its call centre, which received about 2,000 calls a day with 10 to 15 minute holds. The agent collects the caller's name, phone number and zip code, identifies the intent (such as scheduling an appointment), filters spam and hands over to a human with context. PolyAI reports 87% shorter wait times, abandonment down from 46% to 2% and appointment volume up 2% year on year. Automated booking and confirmations are named as the next phase, not yet live.","stage":"production","year":2025,"channels":["voice"],"languages":["en"],"metrics":[{"kpi":"response-time-reduction","value":87,"unit":"percent","qualifier":"exact","period":"call wait time","claimant":"vendor","quote":"With PolyAI, Audibel cut wait times by 87% and abandonment by 44%, while spam dropped 88%.","sourceUrl":"https://poly.ai/customers/audibel"}],"outcomeDisclosed":true,"sources":[{"url":"https://poly.ai/customers/audibel","title":"How Audibel reduced abandonment rates by 44% with PolyAI","publisher":"PolyAI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"audibel-voice-agent-appointment-calls"},{"title":"Austin Peay State University: text message chatbot for incoming students","useCases":["student-enrollment-and-services-assistant"],"organization":{"name":"Austin Peay State University","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[{"name":"Mainstay","role":"platform"}],"summary":"Austin Peay State University introduced a Mainstay text message chatbot, \"The Gov\", in 2017 to send incoming students orientation information and nudges, and later interactive surveys. Its first intent to enroll campaign in 2019 received a 40% response rate within a day, which Mainstay's case study says would have taken more than a month by paper survey, and showed staff which students still planned to attend. According to the case study (about 2020), the chatbot was also used for housing, advising and event updates.","stage":"production","year":2017,"channels":["sms"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://mainstay.com/case-study/austin-peay-state-university/","title":"Austin Peay State University cuts summer melt and sees record-breaking enrollment","publisher":"Mainstay"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"austin-peay-state-university-enrollment-chatbot"},{"title":"Australia Post: AI transaction monitoring and screening for its financial services with Napier AI","useCases":["aml-alert-triage"],"organization":{"name":"Australia Post","anonymized":false,"country":"AU","region":"asia-pacific","industry":"logistics-and-transportation"},"vendors":[{"name":"Napier AI","role":"platform"},{"name":"Microsoft Azure","role":"platform"}],"summary":"As Australia Post grew its financial services alongside postal services, it deployed Napier AI's platform on Microsoft Azure for transaction monitoring, client screening and behavioural analytics. The vendor reports large gains in false positive reduction and unusual activity detection, and suspicious activity information that helped dismantle a money laundering syndicate, but its case study credits these gains to rule configuration, the no code sandbox and rule calibration rather than to AI or ML alert scoring.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.napier.ai/case-study/australia-posts-digital-transformation-halves-false-positives","title":"Australia Post's digital transformation halves false positives","publisher":"Napier AI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"australia-post-napier-aml-monitoring"},{"title":"AvalonBay Communities: AI supported leasing, self guided tours and centralized customer care","useCases":["apartment-leasing-and-resident-service-agent"],"organization":{"name":"AvalonBay Communities","anonymized":false,"country":"US","region":"north-america","industry":"real-estate"},"vendors":[],"summary":"AvalonBay describes an operating model in which on site associates are supported by a centralized shared services organization and a technology platform that incorporates automation and AI. Its executives told Nareit that an AI assistant named Sidney answers prospects' common questions before self guided tours, that renewals are handled AI first with a central associate following up when the AI cannot answer, and that residents submit and track maintenance requests digitally. The company opened a second customer care center in San Antonio to extend leasing, renewal and service support; the article carries no date, and its wording (17 years since the first center opened in 2007, a San Antonio opening \"this spring\") places those details around 2024.","stage":"scaled","year":2025,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/915912/000091591226000004/avb-20251231.htm","title":"AvalonBay Communities Form 10-K for the fiscal year 2025","publisher":"AvalonBay Communities (SEC EDGAR)","date":"2026-02-27"},{"url":"https://www.reit.com/news/articles/avalonbay-integrates-ai-opens-second-customer-care-center-to-advance-centralization","title":"AvalonBay Integrates AI, Opens Second Customer Care Center to Advance Centralization","publisher":"Nareit","archivedUrl":"https://web.archive.org/web/20251031122208/https://www.reit.com/news/articles/avalonbay-integrates-ai-opens-second-customer-care-center-to-advance-centralization"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"avalonbay-ai-leasing-and-customer-care"},{"title":"Avanade: Security Copilot agents for identity and security investigations","useCases":["security-alert-triage-and-investigation"],"organization":{"name":"Avanade","anonymized":false,"country":"US","region":"global","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Avanade embedded Microsoft Security Copilot in its identity and security investigations, run with Microsoft Entra. After a five week pilot in which engineers turned recurring investigation patterns into prompt based workflows, it applied the same approach to security operations and built custom agents that consolidate signals, recommendations and playbook steps for analysts. Most of the reported gains are in identity incident handling, closer to identity and access support than to SOC alert triage: guest account investigations dropped from 10 minutes to 1 minute. A separate study of more than 4,100 security operations investigations reports gains in speed, quality and error rates.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"mttr-reduction","value":80,"unit":"percent","qualifier":"up-to","period":"identity incident resolution time","claimant":"organization","quote":"By turning expert knowledge into repeatable, AI-assisted workflows, we reduced resolution times by up to 80% and enabled the team to take on more work with greater consistency and confidence.","sourceUrl":"https://www.microsoft.com/en/customers/story/27277-avanade-microsoft-security-copilot"},{"kpi":"productivity-gain","value":70,"unit":"percent","qualifier":"exact","period":"speed and efficiency across more than 4,100 security operations investigations","claimant":"vendor","quote":"In a separate study of more than 4,100 security operations investigations, Avanade found a 70% improvement in speed and efficiency, a 7% improvement in investigation and documentation quality, and a 7% reduction in human error.","sourceUrl":"https://www.microsoft.com/en/customers/story/27277-avanade-microsoft-security-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/27277-avanade-microsoft-security-copilot","title":"Avanade cuts incident resolution time by up to 80% with Microsoft Security Copilot","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"avanade-security-copilot-identity-investigations"},{"title":"AXA Switzerland: real time claims fraud detection at first notice of loss","useCases":["claims-fraud-detection"],"organization":{"name":"AXA Switzerland","anonymized":false,"country":"CH","region":"europe","industry":"insurance"},"vendors":[{"name":"Shift Technology","role":"platform"}],"summary":"AXA Switzerland checks motor and property claims for fraud in real time at first notice of loss with Shift Claims Fraud Detection, using more than 100 fraud scenarios tuned to the Swiss market and its portfolio, and combining its own policy and claims data with external sources such as government records. Honest claims go straight to processing, suspicious ones go to an expert with the full context of the alert, and the models run again whenever new data is recorded on a claim. Shift reports that AXA has analysed more than 1 million claims and stopped over EUR 12 million in fraud.","stage":"scaled","year":2023,"channels":["api"],"languages":[],"metrics":[{"kpi":"interactions-handled","value":1000000,"unit":"count","qualifier":"at-least","period":"Claims analysed since deployment","claimant":"vendor","quote":"AXA has now analyzed more than 1 million claims with Shift, and stopped over €12M in fraud, freeing its teams to focus on customer satisfaction and achieve the goal of increasing its presence as #1 in the Swiss market.","sourceUrl":"https://www.shift-technology.com/resources/case-studies/axa-switzerland-insurance-fraud-detection-success"},{"kpi":"fraud-losses-prevented","value":12000000,"unit":"currency","currency":"EUR","qualifier":"at-least","period":"Fraud stopped since deployment, cumulative","claimant":"vendor","quote":"AXA has now analyzed more than 1 million claims with Shift, and stopped over €12M in fraud, freeing its teams to focus on customer satisfaction and achieve the goal of increasing its presence as #1 in the Swiss market.","sourceUrl":"https://www.shift-technology.com/resources/case-studies/axa-switzerland-insurance-fraud-detection-success"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.shift-technology.com/resources/case-studies/axa-switzerland-insurance-fraud-detection-success","title":"AXA Switzerland stops fraud in real-time to drive customer satisfaction","publisher":"Shift Technology","date":"2023-03-17"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"axa-switzerland-claims-fraud-detection"},{"title":"AXIS: AI classification and appetite matching of new business applications with Sixfold","useCases":["commercial-underwriting-submission-triage"],"organization":{"name":"AXIS Capital","anonymized":false,"country":"BM","region":"global","industry":"insurance"},"vendors":[{"name":"Sixfold","role":"platform"}],"summary":"AXIS underwriters spent much of each new submission classifying the applicant into the right industry and writing a snapshot of its operations. AXIS started Sixfold with an underwriter facing dashboard and, after the pilot, integrated it into its automated clearance process, where it applies industry codes and matches cases against AXIS risk appetite. The vendor reports more than 15,000 applications analyzed in the first month and an implementation of about six weeks. AXIS's 2025 annual report separately describes AI tools used to empower underwriters and an AI Underwriting Working Group that monitors AI activity affecting underwriting.","stage":"production","year":2024,"channels":["internal-tools","api"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":15000,"unit":"count","qualifier":"at-least","period":"applications analyzed in the first month","claimant":"vendor","quote":"In the first month, AXIS analyzed more than 15,000 applications using Sixfold’s AI.","sourceUrl":"https://www.sixfold.ai/case-study/axis"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sixfold.ai/case-study/axis","title":"AXIS | Sixfold Case Study","publisher":"Sixfold"},{"url":"https://www.sec.gov/Archives/edgar/data/1214816/000121481626000097/axs-20251231.htm","title":"AXIS Capital Holdings Limited Form 10-K for 2025","publisher":"AXIS Capital via SEC EDGAR","date":"2026-02-27"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"axis-sixfold-submission-classification"},{"title":"Aydem Energy: generative AI WhatsApp assistant for bills, meter readings and outages","useCases":["utility-billing-and-move-agent"],"organization":{"name":"Aydem Energy","anonymized":false,"country":"TR","region":"europe","industry":"energy-and-utilities"},"vendors":[{"name":"Softtech","role":"integrator"},{"name":"Microsoft","role":"model-provider"}],"summary":"Aydem Energy, which supplies electricity to around 6 million customers in Türkiye, built a WhatsApp assistant on Azure OpenAI with Softtech. It opens with data protection consent and an AI disclosure, matches the caller's phone number to CRM data, gives location specific outage updates, guides meter reading submission, checks outstanding bills and processes compensation claims, and escalates urgent or angry customers to a human. Aydem reports about 1,000 inquiries a day, 75% of them resolved without an agent, and it restricts the assistant to the knowledge relevant to each scenario to control cost.","stage":"production","year":2024,"channels":["whatsapp"],"languages":["tr"],"metrics":[{"kpi":"containment-rate","value":75,"unit":"percent","qualifier":"exact","period":"WhatsApp inquiries","claimant":"organization","quote":"Of the inquiries that come through WhatsApp, 75% are fully resolved by the digital assistant without the need for live agent support","sourceUrl":"https://www.microsoft.com/en/customers/story/20845-aydem-energy-azure-open-ai-service"},{"kpi":"interactions-handled","value":1000,"unit":"count","qualifier":"approximately","period":"per day","claimant":"vendor","quote":"Today, the digital assistant manages about 1,000 customer inquiries daily through WhatsApp, which represents 10% of total customers calling customer service.","sourceUrl":"https://www.microsoft.com/en/customers/story/20845-aydem-energy-azure-open-ai-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/20845-aydem-energy-azure-open-ai-service","title":"Aydem Energy manages multiple-fold seasonal call surges with an Azure OpenAI-powered digital assistant","publisher":"Microsoft","date":"2024-12-27"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"aydem-energy-whatsapp-assistant"},{"title":"Baltimore City 911: live transcripts, AI summaries, translation and automated QA","useCases":["emergency-call-triage-support","public-service-translation"],"organization":{"name":"Baltimore City 911 (Emergency Communications)","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Prepared","role":"platform"}],"summary":"Baltimore's emergency communications centre answers about 1.4 million 911 calls a year. Since partnering with Prepared in early 2022, call takers see a live transcript, AI summaries and highlighted key details on every call, which helps with addresses and callers who are hard to understand. Non English calls are transcribed and translated in real time, and for Spanish calls operators can dial in an automated voice translator instead of a third party interpreter. Automated QA now reviews every call, where a contractor previously reviewed roughly 30%. The case study also carries the unattributed line that the tool is \"about 98% accurate\", without saying what was measured or how, so it is not recorded as an accuracy figure.","stage":"scaled","year":2022,"channels":["voice","agent-desktop"],"languages":["en","es"],"metrics":[{"kpi":"quality-score-uplift","value":12,"unit":"percent","qualifier":"exact","period":"QA scores since automated QA was deployed; the QA method changed at the same time (sampling of about 30% of calls replaced by automated review of all calls), so before and after scores may not be comparable","claimant":"vendor","quote":"Since deploying Automated QA, Baltimore has seen a 12% improvement in QA scores.","sourceUrl":"https://www.prepared911.com/case-studies/baltimore-911-call-processing-assistive-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.prepared911.com/case-studies/baltimore-911-call-processing-assistive-ai","title":"Baltimore 911: Improving call processing efficiency with Assistive AI","publisher":"Prepared"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"baltimore-911-assistive-call-taking"},{"title":"Banco Bradesco: AILA generative AI assistant for the audit process","useCases":["internal-audit-copilot"],"organization":{"name":"Banco Bradesco","anonymized":false,"country":"BR","region":"latin-america","industry":"banking"},"vendors":[{"name":"Microsoft (Azure OpenAI)","role":"platform"}],"summary":"Bradesco built AILA, an assistant on Azure OpenAI that supports its audit process across planning, reporting, drafting and proofreading, and root cause analysis. Microsoft reports efficiency and time savings for each of those steps in a customer story round up, without stating the measurement method or period.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["pt"],"metrics":[{"kpi":"productivity-gain","value":65,"unit":"percent","qualifier":"exact","baseline":"audit planning efficiency before AILA","claimant":"vendor","quote":"With AILA, they achieved 65% more efficiency in audit planning, 55% less time in reporting, 50% less time writing/proofreading, and improved the identification of the root cause of problems by reducing the time required for this step by 20%.","sourceUrl":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/"},{"kpi":"handling-time-reduction","value":55,"unit":"percent","qualifier":"exact","baseline":"time spent in the audit reporting phase before AILA","claimant":"vendor","quote":"With AILA, they achieved 65% more efficiency in audit planning, 55% less time in reporting, 50% less time writing/proofreading, and improved the identification of the root cause of problems by reducing the time required for this step by 20%.","sourceUrl":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft Cloud Blog","date":"2025-07-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"banco-bradesco-aila-audit-assistant"},{"title":"Banco Supervielle: AI and automation for judicial orders on customer accounts","useCases":["account-servicing-execution"],"organization":{"name":"Banco Supervielle","anonymized":false,"country":"AR","region":"latin-america","industry":"banking"},"vendors":[{"name":"SS&C Blue Prism","role":"platform"}],"summary":"Banco Supervielle in Argentina automated the handling of judicial notifications, the court orders that require a bank to freeze, release or transfer funds on customer accounts. Digital workers that use natural language processing and machine learning connect to court systems, identify the type of order, perform the account checks, generate and submit the response letter and close the case. The vendor reports a 58% cut in processing time and full compliance with judicial deadlines. The bank also automated its loan disbursement process.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["es"],"metrics":[{"kpi":"processing-time-reduction","value":58,"unit":"percent","qualifier":"exact","period":"judicial notification processing, average response time from 12 to 5 minutes","claimant":"vendor","quote":"By combining automation and AI, the bank achieved a 58% reduction in processing time — cutting average response times from 12 minutes to just five — and increased case-handling capacity by 43% during a six-month period.","sourceUrl":"https://www.blueprism.com/resources/case-studies/banco-supervielle-legal-process-ai-automation/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.blueprism.com/resources/case-studies/banco-supervielle-legal-process-ai-automation/","title":"Banco Supervielle | Legal Process AI Automation Case Study | SS&C Blue Prism","publisher":"SS&C Blue Prism","date":"2025-12-29"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"banco-supervielle-judicial-notification-automation"},{"title":"Banestes: Gemini to speed up credit analysis and balance sheet reviews","useCases":["credit-memo-drafting-agent"],"organization":{"name":"Banestes","anonymized":false,"country":"BR","region":"latin-america","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Banestes, a Brazilian bank, used Gemini in Google Workspace to accelerate credit analysis by simplifying balance sheet reviews. Banestes CTO Vicente Lopes Duarte said the generative AI tool has made it easier to read balance sheet documents, helping credit analysis teams work faster. No outcome figures for credit analysis are published; the story's only figure, a 70% ticket reduction, is about AppSheet, a separate Workspace tool, not the credit analysis use.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["pt"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://workspace.google.com/intl/pt-BR/customers/banestes/","title":"Banestes: soluções do Google Workspace","publisher":"Google"},{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"banestes-gemini-credit-analysis"},{"title":"Bank of America: AI conversation simulators in The Academy","useCases":["conversation-roleplay-training"],"organization":{"name":"Bank of America","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[],"summary":"The Academy, Bank of America's onboarding, education and professional development organization, uses AI conversation simulators in which employees practise different types of client interactions and receive real time feedback. The bank reports more than one million simulations completed in 2024 and says many employees note that practising client conversations helps them deliver better and more consistent service. No proficiency or client outcome figures are published.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":1000000,"unit":"count","qualifier":"at-least","period":"simulations completed by employees in 2024","claimant":"organization","quote":"Employees completed over 1 million simulations last year, with many noting that practicing client conversations helps them deliver better and more consistent service.","sourceUrl":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html","title":"AI Adoption by BofA's Global Workforce Improves Productivity, Client Service","publisher":"Bank of America","date":"2025-04-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bank-of-america-academy-conversation-simulators"},{"title":"Bank of America: ask MERRILL and ask PRIVATE BANK knowledge assistants for advisers","useCases":["enterprise-knowledge-search","wealth-advisor-knowledge-assistant"],"organization":{"name":"Bank of America","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"Bank of America","role":"in-house"}],"summary":"Merrill and Bank of America Private Bank teams use ask MERRILL and ask PRIVATE BANK, built on the technology behind Erica, to curate the information they need for clients. For more complex requests, the chat can connect teams with experts at the bank. The bank reports more than 23 million interactions with the two tools in 2024.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":23000000,"unit":"count","qualifier":"at-least","period":"calendar year 2024","claimant":"organization","quote":"In 2024, there were more than 23 million interactions with ask MERRILL and ask PRIVATE BANK, an increase of 1 million over 2023, helping employees more proactively connect with clients about timely and relevant opportunities.","sourceUrl":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html","title":"AI Adoption by BofA's Global Workforce Improves Productivity, Client Service","publisher":"Bank of America","date":"2025-04-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bank-of-america-ask-merrill-and-ask-private-bank"},{"title":"Bank of America: CashPro Chat with Erica for business clients","useCases":["corporate-client-servicing-assistant"],"organization":{"name":"Bank of America","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Bank of America","role":"in-house"}],"summary":"Bank of America brought the AI behind its Erica assistant into CashPro Chat, the service assistant inside the CashPro platform that corporate and commercial clients use for payments, deposits, loans and trade. It finds transactions and account information, guides users through the platform and routes complex requests to specialised service teams. After the Erica integration, chat volume rose 41% on the 2023 weekly average while chats with a live agent fell 16%. In August 2025 the bank said 65% of CashPro clients use CashPro Chat and that Erica handles more than 40% of client interactions in it.","stage":"scaled","year":2023,"channels":["web-chat"],"languages":[],"metrics":[{"kpi":"contact-deflection","value":16,"unit":"percent","qualifier":"exact","period":"chats with a live agent since the Erica integration, while chat volume rose 41% compared to the 2023 weekly average","claimant":"organization","quote":"Since launch, chat volume increased by 41% compared to the 2023 weekly average. Meanwhile, chats with a live agent decreased by 16%.","sourceUrl":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2023/09/enhancements-to-bofa-s-cashpro--chat-create-greater-efficiencies.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2023/09/enhancements-to-bofa-s-cashpro--chat-create-greater-efficiencies.html","title":"Enhancements to BofA's CashPro Chat Create Greater Efficiencies for Business Clients","publisher":"Bank of America","date":"2023-09-18"},{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/08/a-decade-of-ai-innovation--bofa-s-virtual-assistant-erica-surpas.html","title":"A Decade of AI Innovation: BofA's Virtual Assistant Erica Surpasses 3 Billion Client Interactions","publisher":"Bank of America","date":"2025-08-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bank-of-america-cashpro-chat"},{"title":"Bank of America: machine learning cash forecasting in CashPro","useCases":["treasury-cash-flow-forecasting"],"organization":{"name":"Bank of America","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Bank of America","role":"in-house"}],"summary":"Bank of America launched a cash forecasting solution that uses machine learning in January 2022, offered to business clients in its CashPro platform as CashPro Forecasting. The bank reports fast adoption rather than accuracy: from the first half of 2022 to the first half of 2023, new client enrollments rose 141%, active users 105% and sign ins among those users 375%. The same release describes enhancements to CashPro Chat, a virtual service advisor in CashPro that now uses the same AI and machine learning capabilities as Erica, the bank's consumer assistant.","stage":"scaled","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2023/09/enhancements-to-bofa-s-cashpro--chat-create-greater-efficiencies.html","title":"Enhancements to BofA's CashPro Chat Create Greater Efficiencies for Business Clients","publisher":"Bank of America","date":"2023-09-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bank-of-america-cashpro-forecasting"},{"title":"Bank of America: generative AI coding assistant for software developers","useCases":["developer-coding-assistant"],"organization":{"name":"Bank of America","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[],"summary":"Bank of America software developers use a generative AI tool that helps them write and optimize code. The bank reported the efficiency gain in its April 2025 update on how its workforce uses AI, alongside its internal assistants and contact centre tools. The bank does not name the underlying model or vendor.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":20,"unit":"percent","qualifier":"at-least","claimant":"organization","quote":"Coding assistance – Bank of America software developers are using a GenAI-based tool to assist with code writing and optimization, through which they have experienced efficiency gains of over 20%.","sourceUrl":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html","title":"AI Adoption by BofA's Global Workforce Improves Productivity, Client Service","publisher":"Bank of America","date":"2025-04-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bank-of-america-coding-assistant"},{"title":"Bank of America: Erica for Employees, the internal IT and HR assistant","useCases":["it-service-desk-resolution-agent","hr-and-policy-assistant"],"organization":{"name":"Bank of America","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Bank of America","role":"in-house"}],"summary":"Bank of America launched Erica for Employees in 2020, building on its customer facing assistant, to give staff technology support such as mobile device password resets and device activation. In 2023 it was extended to HR topics such as where to review health benefits and how to find payroll and tax forms. The bank reports that most employees use it and that it has more than halved calls into the IT service desk.","stage":"scaled","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"employee-adoption","value":90,"unit":"percent","qualifier":"at-least","period":"as of April 2025","claimant":"organization","quote":"Today, over 90% of employees use Erica for Employees, with the virtual assistant having reduced calls into the IT service desk by more than 50%.","sourceUrl":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html"},{"kpi":"contact-deflection","value":50,"unit":"percent","qualifier":"at-least","period":"calls into the IT service desk, as of April 2025","claimant":"organization","quote":"Today, over 90% of employees use Erica for Employees, with the virtual assistant having reduced calls into the IT service desk by more than 50%.","sourceUrl":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/04/ai-adoption-by-bofa-s-global-workforce-improves-productivity--cl.html","title":"AI Adoption by BofA's Global Workforce Improves Productivity, Client Service","publisher":"Bank of America","date":"2025-04-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bank-of-america-erica-for-employees"},{"title":"Bank of America: Erica, a virtual financial assistant with proactive insights","useCases":["financial-wellbeing-coach","offers-and-rewards-agent","first-line-contact-centre-agent","proactive-outbound-engagement-agent","branch-and-appointment-booking-agent"],"organization":{"name":"Bank of America","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Bank of America","role":"in-house"}],"summary":"Erica, launched in 2018, is Bank of America's virtual financial assistant in its Mobile Banking app. Beyond answering questions it delivers proactive, personalized insights: BankAmeriDeals cash back deals based on the client's spending, where balances are trending over the next seven days and eligibility for the Preferred Rewards program. It also gives guidance on investment topics for Merrill clients and hands off to people by scheduling appointments. The bank reports that clients have received and interacted with more than 1.7 billion of these insights, and that most users find the information they need, which it links to lower call centre volume. Bank of America says Erica selects answers from a predefined set and does not use generative AI or large language models.","stage":"scaled","year":2025,"channels":["mobile-app"],"languages":["en"],"metrics":[{"kpi":"users-served","value":50000000,"unit":"count","qualifier":"approximately","period":"since launch in 2018, as of August 2025","claimant":"organization","quote":"assisting nearly 50 million users since launch, surpassing 3 billion client interactions, and now averaging more than 58 million interactions per month","sourceUrl":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/08/a-decade-of-ai-innovation--bofa-s-virtual-assistant-erica-surpas.html"},{"kpi":"interactions-handled","value":3000000000,"unit":"count","qualifier":"at-least","period":"client interactions since launch in 2018, as of August 2025","claimant":"organization","quote":"surpassing 3 billion client interactions","sourceUrl":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/08/a-decade-of-ai-innovation--bofa-s-virtual-assistant-erica-surpas.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2025/08/a-decade-of-ai-innovation--bofa-s-virtual-assistant-erica-surpas.html","title":"A Decade of AI Innovation: BofA's Virtual Assistant Erica Surpasses 3 Billion Client Interactions","publisher":"Bank of America","date":"2025-08-20"},{"url":"https://info.bankofamerica.com/en/digital-banking/erica","title":"Erica: Virtual Financial Assistant","publisher":"Bank of America"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bank-of-america-erica-virtual-assistant"},{"title":"Merrill and Bank of America Private Bank: AI Powered Meeting Journey","useCases":["client-briefing-and-call-report-copilot","client-meeting-notes-and-crm-update"],"organization":{"name":"Bank of America","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"Bank of America","role":"in-house"}],"summary":"Merrill Wealth Management and Bank of America Private Bank rolled out an AI meeting solution at full scale in March 2026. It consolidates client relationship insights and recent activity into meeting preparation material, takes notes in virtual meetings with client consent, and turns the decisions into a summary, tasks and documentation afterwards. The bank says the capability can save advisors up to four hours per meeting; it presents this as potential, not as a measured result, so it is not recorded as a metric here.","stage":"scaled","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.bankofamerica.com/content/newsroom/press-releases/2026/03/merrill-and-bank-of-america-private-bank-launch-ai-powered-meeti.html","title":"Merrill and Bank of America Private Bank Launch AI-Powered Meeting Journey","publisher":"Bank of America","date":"2026-03-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bank-of-america-merrill-ai-meeting-journey"},{"title":"Bank of Singapore: HELIOS agentic AI platform for wealth client onboarding","useCases":["source-of-wealth-diligence"],"organization":{"name":"Bank of Singapore","anonymized":false,"country":"SG","region":"asia-pacific","industry":"wealth-and-asset-management"},"vendors":[],"summary":"In 2026 Bank of Singapore began using HELIOS, an agentic AI platform that streamlines due diligence and credit risk profiles when onboarding high net worth and ultra high net worth clients. More than 100 relationship managers (about a quarter) in Singapore, Hong Kong and Dubai had begun using it. Its chief executive said roughly 50 clients had been fully onboarded through it. The bank aims to cut account opening from more than 30 business days to 15, which is a target rather than a result. The move follows a request from the Monetary Authority of Singapore to shorten account opening times for private bank clients.","stage":"pilot","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":50,"unit":"count","qualifier":"approximately","period":"clients fully onboarded at time of reporting","claimant":"organization","quote":"Jason Moo, Bank of Singapore's CEO, said that roughly 50 clients have been fully onboarded through the platform, taking the HNW and UHNW segments together.","sourceUrl":"https://globalbusinessoutlook.com/banking-and-finance/bank-of-singapore-dbs-take-ai-route-to-accelerate-wealth-client-onboarding/"}],"outcomeDisclosed":true,"sources":[{"url":"https://globalbusinessoutlook.com/banking-and-finance/bank-of-singapore-dbs-take-ai-route-to-accelerate-wealth-client-onboarding/","title":"Bank of Singapore, DBS take AI route to accelerate wealth client onboarding","publisher":"Global Business Outlook","date":"2026-07-31"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"bank-of-singapore-agentic-wealth-onboarding"},{"title":"Bank of Singapore: Source of Wealth Assistant (SOWA)","useCases":["source-of-wealth-diligence"],"organization":{"name":"Bank of Singapore","anonymized":false,"country":"SG","region":"asia-pacific","industry":"wealth-and-asset-management"},"vendors":[{"name":"OCBC","role":"in-house"}],"summary":"Bank of Singapore, the private bank of OCBC, rolled out an agentic AI tool that drafts source of wealth reports for know your customer due diligence. Relationship managers upload the client's documents (financial statements, tax notices, property valuations, corporate filings, payslips) and SOWA reviews them and generates a standardized report, checking plausibility against benchmarks such as salary and company revenue from Bank of Singapore and OCBC data. The relationship manager verifies and refines the draft before it goes to compliance. The bank says report writing time fell from 10 days to one hour, with fewer inconsistencies and omissions. The tool runs on the bank's private cloud.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"cycle-time-days","value":1,"unit":"hours","qualifier":"exact","period":"per source of wealth report","baseline":"10 days to write a source of wealth report before SOWA","claimant":"organization","quote":"Time taken to write the report has been shortened from 10 days to one hour, with greater accuracy and consistency.","sourceUrl":"https://www.bankofsingapore.com/media-releases/2025/bank-of-singapore-deploys-agentic-ai-tool-to-automate-writing-of-source-of-wealth-reports.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.bankofsingapore.com/media-releases/2025/bank-of-singapore-deploys-agentic-ai-tool-to-automate-writing-of-source-of-wealth-reports.html","title":"Bank of Singapore deploys agentic AI tool to automate writing of source of wealth reports","publisher":"Bank of Singapore","date":"2025-10-10"},{"url":"https://www.wealthbriefing.com/html/article.php/bank-of-singapore-deploys-ai-to-accelerate-source-of-wealth-verification","title":"Bank of Singapore Deploys AI To Accelerate Source Of Wealth Verification","publisher":"WealthBriefing","date":"2025-10-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bank-of-singapore-source-of-wealth-assistant"},{"title":"BARK: AI agent Scout for order tracking and shipping questions","useCases":["order-status-and-returns-agent"],"organization":{"name":"BARK","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Sierra","role":"platform"}],"summary":"BARK, the company behind BarkBox, ships tens of thousands of boxes a month and deployed an AI agent called Scout to take the high volume, repetitive conversations first: where is my order, what is my tracking number, when will it arrive, and basic account questions. Within the first year the agent handled roughly a quarter of all customer conversations and chat response times fell below two minutes, while emotional conversations, such as the loss of a pet, stay with the human team.","stage":"production","year":2026,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"customer-satisfaction","value":98,"unit":"percent","qualifier":"exact","period":"first year of the agent, customer service overall","claimant":"vendor","quote":"BARK maintained a 98% customer satisfaction rate, holding the same high bar as the Happy Team.","sourceUrl":"https://sierra.ai/customers/bark"}],"outcomeDisclosed":true,"sources":[{"url":"https://sierra.ai/customers/bark","title":"How BARK delivers tail-wagging customer experiences with AI","publisher":"Sierra","date":"2026-05-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"bark-scout-order-tracking-agent"},{"title":"Barnet Council: Ami chatbot for council tax, benefits and waste enquiries","useCases":["non-emergency-service-request-routing"],"organization":{"name":"London Borough of Barnet","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Conversations By Ami Ltd","role":"platform"},{"name":"Capita PLC","role":"integrator"}],"summary":"Barnet Council piloted Ami on its website to signpost residents 24/7 to the right page or service in a few areas: council tax, housing benefits, waste and recycling, schools and pest control. Ami only uses council approved content, can move the resident straight to the relevant page, watches for signs that a resident needs more support and, in office hours, connects them to an agent in the Amazon Connect contact centre that Capita runs for the council. The pilot expected about 30,000 chats over six months; no decisions are made about residents.","stage":"pilot","year":2025,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/barnet-council-ami-chatbot","title":"Barnet Council: Ami Chatbot","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-01-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"barnet-council-ami-chatbot"},{"title":"Bayer: natural language self service analytics with Snowflake Cortex Analyst","useCases":["governed-text-to-sql-analytics"],"organization":{"name":"Bayer","anonymized":false,"country":"DE","region":"europe","industry":"pharma-and-life-sciences"},"vendors":[{"name":"Snowflake","role":"platform"}],"summary":"Bayer uses Snowflake Cortex Analyst as the query generation service and a Streamlit chat interface to answer natural language questions over its enterprise data platform, alongside its existing dashboards. The first phase answered executive questions from sales vice presidents, such as the market share of a product last month, and it has since been extended to business unit analysts with row level data. No outcome figures are published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.snowflake.com/en/blog/cortex-analyst-ai-self-service-analytics/","title":"Cortex Analyst: Paving the Way to Self-Service Analytics with AI","publisher":"Snowflake","date":"2024-08-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"bayer-cortex-analyst-self-service-analytics"},{"title":"Bayer: Genpact pharmacovigilance AI for adverse event case processing","useCases":["adverse-event-case-intake"],"organization":{"name":"Bayer","anonymized":false,"country":"DE","region":"europe","industry":"pharma-and-life-sciences"},"vendors":[{"name":"Genpact","role":"platform"}],"summary":"In November 2018 Genpact announced a multi year agreement with Bayer under which its Pharmacovigilance Artificial Intelligence (PVAI) suite, which incorporates the Genpact Cora PharmacoVigilance software product, is applied to Bayer's existing pharmacovigilance database and IT systems. Bayer's head of pharmacovigilance said the partnership offered an opportunity to further increase the efficiency of its pharmacovigilance operating model and case processing, and Genpact said Bayer was among the first companies going live with the AI based Case Management module of the PVAI suite. Neither company has published outcome figures.","stage":"announced","year":2018,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://media.genpact.com/2018-11-01-Genpact-and-Bayer-to-Co-innovate-to-Leverage-Artificial-Intelligence-Capabilities-for-Patient-Safety","title":"Genpact and Bayer to Co-innovate to Leverage Artificial Intelligence Capabilities for Patient Safety","publisher":"Genpact","date":"2018-11-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"bayer-genpact-pharmacovigilance-ai"},{"title":"British Columbia Investment Management Corporation: automation cuts internal audit report writing time","useCases":["internal-audit-copilot"],"organization":{"name":"British Columbia Investment Management Corporation","anonymized":false,"country":"CA","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"Microsoft (Microsoft 365 Copilot and Azure)","role":"platform"}],"summary":"BCI uses Microsoft 365 Copilot and the Azure ecosystem to automate manual tasks across its operations. Among the outcomes Microsoft reports is less time spent writing internal audit reports; the source does not say which tool produced that result. The organization wide productivity figures on the same page are not specific to audit and are not recorded here.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":30,"unit":"percent","qualifier":"exact","baseline":"time spent writing internal audit reports before the automation","claimant":"vendor","quote":"The organization saved more than 2,300 person-hours through automation, reduced the time spent on writing internal audit reports by 30%, and saved a month of processing time to analyze 8,000 survey comments.","sourceUrl":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft Cloud Blog","date":"2025-07-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"bci-copilot-internal-audit-reports"},{"title":"BCU: AI document fraud detection in lending and account opening with Inscribe","useCases":["application-and-identity-fraud-detection"],"organization":{"name":"BCU","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Inscribe","role":"platform"}],"summary":"BCU, a US credit union, uses Inscribe's document fraud detection on loan documents and member account applications to find altered bank statements, pay stubs and other documents. Its investigators used the signals to uncover fraud rings, synthetic identities and reused document templates. The vendor reports USD 5.6 million in losses from altered documents prevented in the first nine months of 2025.","stage":"production","year":2025,"channels":["api","internal-tools"],"languages":["en"],"metrics":[{"kpi":"fraud-losses-prevented","value":5600000,"unit":"currency","currency":"USD","qualifier":"exact","period":"first nine months of 2025","claimant":"vendor","quote":"In the first nine months of 2025, BCU prevented $5.6 million in losses from altered documents and has now saved $80 million total!","sourceUrl":"https://www.inscribe.ai/customers/bcu"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.inscribe.ai/customers/bcu","title":"BCU Success Story","publisher":"Inscribe"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"bcu-inscribe-document-fraud-detection"},{"title":"Belastingdienst: rules based VAT signal model for large businesses, published in the Dutch algorithm register","useCases":["tax-compliance-risk-scoring"],"organization":{"name":"Belastingdienst","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[{"name":"In house (Belastingdienst)","role":"in-house"}],"summary":"Since 1 October 2024 the Dutch Tax and Customs Administration has supported staff of its Large Businesses directorate with a signal model that risk assesses VAT returns. The model applies business rules drawn from legislation, expertise and statistics, is explicitly not self learning, and sorts signals into priority categories that help staff decide when and by whom a signal is handled. The register says the model also takes decisions itself; where it cannot (more complex situations or a deviation in the return), a staff member intervenes. Samples are also drawn from returns the model does not select, and the rules are evaluated every year against those results. It is one of several VAT signal models the Belastingdienst has published in the national algorithm register.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/189378/62272663/signaalmodel-omzetbelasting-grote-ondernemingen-sob-go","title":"Signaalmodel Omzetbelasting Grote Ondernemingen (SOB GO), Algoritmeregister","publisher":"Algoritmeregister van de Nederlandse overheid"},{"url":"https://over-ons.belastingdienst.nl/onderwerpen/omgaan-met-gegevens/algoritmeregister/signaalmodel-omzetbelasting-grote-ondernemingen-sob-go/","title":"Signaalmodel Omzetbelasting Grote Ondernemingen (SOB GO)","publisher":"Belastingdienst","date":"2024-11-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"belastingdienst-vat-signal-model-large-businesses"},{"title":"Bell Canada: network AI Ops for detecting, ranking and resolving network issues","useCases":["network-fault-triage-copilot"],"organization":{"name":"Bell Canada","anonymized":false,"country":"CA","region":"north-america","industry":"telecommunications"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Google (Gemini models)","role":"model-provider"}],"summary":"Bell deployed a network AI Ops solution built on Google Cloud that correlates network data with customer experience to rank issues by their real impact, so a small fault on a busy site gets attention before a larger one on a quiet site. Custom machine learning models handle detection and prioritisation, a graph of network relationships estimates customer impact, and Gemini models support incident analysis, retrieval of historical context and access to vendor documentation. Bell says the approach has significantly improved mean time to resolution but does not give a figure specific to AI Ops.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.bce.ca/news-and-media/newsroom?article=Bell-Canada-launches-AI-powered-network-operations-solution-built-on-Google-Cloud","title":"Bell Canada launches AI-powered network operations solution built on Google Cloud","publisher":"BCE","date":"2025-02-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bell-canada-network-ai-ops"},{"title":"Bell: AI Suspicious Call Detection and spoofed call flagging","useCases":["spam-and-scam-call-blocking"],"organization":{"name":"Bell Canada","anonymized":false,"country":"CA","region":"north-america","industry":"telecommunications"},"vendors":[],"summary":"Bell launched AI powered Suspicious Call Detection in 2025 to block or label spam and fraudulent calls for its wireless customers. In July 2026 it added a new AI model that identifies and flags spoofed calls, which manipulate caller ID to impersonate trusted organizations or contacts, in real time. The enhancement rolls out automatically, with no opt in required, to customers on iOS and Android phones across Bell, Virgin Plus, Lucky Mobile, PC Mobile, No Name and Maxi. Bell says it was the first carrier in Canada to flag spoofed calls this way.","stage":"scaled","year":2025,"channels":["voice","mobile-app"],"languages":["en","fr"],"metrics":[{"kpi":"interactions-handled","value":540000000,"unit":"count","qualifier":"at-least","period":"2025 launch to July 2026, calls blocked or labelled","claimant":"organization","quote":"Since launching Suspicious Call Detection in 2025, Bell has analyzed more than 4.4 billion calls and blocked or labelled over 540 million suspicious or fraudulent calls.","sourceUrl":"https://www.bce.ca/news-and-media/newsroom?article=bell-first-carrier-in-canada-to-detect-and-protect-against-spoofed-calls-with-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.bce.ca/news-and-media/newsroom?article=bell-first-carrier-in-canada-to-detect-and-protect-against-spoofed-calls-with-ai","title":"Bell first carrier in Canada to detect and protect against spoofed calls with AI","publisher":"BCE","date":"2026-07-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bell-suspicious-call-detection"},{"title":"Best Buy: generative AI virtual assistant for support, deliveries and appointments","useCases":["branch-and-appointment-booking-agent","order-status-and-returns-agent"],"organization":{"name":"Best Buy","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Accenture","role":"integrator"}],"summary":"In April 2024 Best Buy announced, with Google Cloud and Accenture, a generative AI virtual assistant for BestBuy.com, its app and its customer support line, expected to launch in late summer 2024, to help customers troubleshoot product issues, change order delivery and scheduling, and manage subscriptions and memberships. Google Cloud reported in its April 2026 list that Best Buy now guides shoppers through technical specifications, issue resolution and appointment scheduling autonomously, using Agent Assist powered by Gemini Enterprise for Customer Experience. It is a retail example of the same job a bank has when it books a branch or specialist appointment. No outcome figures for the assistant were published.","stage":"production","year":2024,"channels":["web-chat","mobile-app","voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://corporate.bestbuy.com/2024/generative-ai-customer-support/","title":"How Best Buy is using generative AI to create better customer support experiences","publisher":"Best Buy","date":"2024-04-09"},{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"best-buy-appointment-scheduling-agent"},{"title":"BigPay: stopping money mule networks with Feedzai in Malaysia","useCases":["mule-network-detection"],"organization":{"name":"BigPay","anonymized":false,"country":"MY","region":"asia-pacific","industry":"payments"},"vendors":[{"name":"Feedzai","role":"platform"}],"summary":"BigPay, a Malaysian electronic money platform that was receiving over 1,000 reported mule cases a month, worked with Feedzai to turn mule patterns into detection rules on its existing Feedzai platform: alert logic for suspicious inbound payments, decline rules based on funding velocity and beneficiary risk, and rules that block cash out channels such as crypto, ATMs and remittances. The source attributes BigPay's result to this analyst built rule set and presents Feedzai's AI assisted alert prioritisation as the platform's next layer of defense, not as the cause of the outcome. The vendor reports that BigPay neutralised a new mule network surge within 72 hours and reduced overall mule activity by more than 90 percent within 60 days.","stage":"production","year":2025,"channels":["api","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.feedzai.com/customer-stories/bigpay-money-mules/","title":"BigPay Stops Money Mules with Feedzai","publisher":"Feedzai","date":"2025-02-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"bigpay-feedzai-mule-detection"},{"title":"BioCatch Trust Australia: behavioural intelligence sharing on receiving accounts across five banks","useCases":["mule-network-detection","real-time-fraud-scoring"],"organization":{"name":"ANZ, Commonwealth Bank, NAB, Suncorp Bank and Westpac (BioCatch Trust Australia)","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"BioCatch","role":"platform"}],"summary":"In November 2024 ANZ, Commonwealth Bank, NAB, Suncorp Bank and Westpac joined BioCatch Trust Australia, launched as a pilot of an interbank network that shares behavioural and device intelligence about receiving accounts, so the sending bank can review a payment to a likely mule account before money leaves. BioCatch reports that in the third quarter of 2025 the network analysed more than 180 million payments and revealed more than $60 million in attempted fraud (currency not stated in the release), and that two more institutions, including Macquarie Bank, have since joined.","stage":"production","year":2025,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":180000000,"unit":"count","qualifier":"at-least","period":"payments analysed in the third quarter of 2025","claimant":"vendor","quote":"in the third quarter of 2025 alone, analyzed more than 180 million payments totaling more than $330 billion, revealing more than $60 million in attempted fraud.","sourceUrl":"https://www.biocatch.com/press-release/aussie-intelligence-sharing-exposes-more-than-60-million-in-fraud-attempts-in-three-months"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.biocatch.com/press-release/aussie-intelligence-sharing-exposes-more-than-60-million-in-fraud-attempts-in-three-months","title":"Aussie intelligence-sharing exposes more than $60 million in fraud attempts in three months","publisher":"BioCatch","date":"2025-11-26"},{"url":"https://www.biocatch.com/press-release/biocatch-partners-australian-banks-fraud-scams-intelligence-sharing-network","title":"BioCatch partners with Australian banks on launch of fraud and scams intelligence-sharing network","publisher":"BioCatch","date":"2024-11-20"},{"url":"https://www.nab.com.au/news/technology-ai/nab-biocatch-trust","title":"NAB joins BioCatch Trust Australia to protect customers from scams and fraud","publisher":"NAB","date":"2024-11-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"biocatch-trust-australia-mule-intelligence"},{"title":"Bloomberg Connects: museum audio guides created with Gemini","useCases":["ai-visitor-and-tour-guide"],"organization":{"name":"Bloomberg Connects","anonymized":false,"country":"US","region":"global","industry":"media-and-entertainment"},"vendors":[{"name":"Google Cloud (Gemini)","role":"model-provider"}],"summary":"Bloomberg Connects uses Gemini to help create immersive audio guides, with the stated aim of making museums more accessible to visually impaired visitors. Google Cloud's listing gives no detail on scale, languages or results.","stage":"production","year":2024,"channels":["mobile-app"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2024-04-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"bloomberg-connects-gemini-audio-guides"},{"title":"BMW Group: AIQX camera, sensor and acoustic quality inspection in vehicle assembly","useCases":["production-line-quality-inspection"],"organization":{"name":"BMW Group","anonymized":false,"country":"DE","region":"europe","industry":"automotive"},"vendors":[{"name":"BMW Group","role":"in-house"}],"summary":"BMW built AIQX (Artificial Intelligence Quality Next), an in house platform that places cameras and sensors along the assembly line, analyses their data in real time and sends line workers immediate feedback on smart devices. It checks variants and completeness and flags anomalies, and at Plant Dingolfing a sub area called Acoustic Analytics listens to driving noises through microphones on the seats as the final check before handover. BMW describes these AI quality tools as part of the production master plan for its plants worldwide. Separately, in 2025 Plant Regensburg piloted GenAI4Q, developed with Datagon AI, which builds an individual final inspection catalogue for each of the roughly 1,400 vehicles built there each day.","stage":"production","year":2023,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.bmwgroup.com/en/news/general/2023/aiqx.html","title":"How AI is revolutionising production.","publisher":"BMW Group"},{"url":"https://www.press.bmwgroup.com/global/article/detail/T0449729EN/artificial-intelligence-as-a-quality-booster?language=en","title":"Artificial intelligence as a quality booster","publisher":"BMW Group PressClub","date":"2025-04-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bmw-group-aiqx-quality-inspection"},{"title":"BMW Group: Factory Genius, a generative AI assistant for troubleshooting production equipment","useCases":["plant-operator-and-maintenance-copilot"],"organization":{"name":"BMW Group","anonymized":false,"country":"DE","region":"europe","industry":"automotive"},"vendors":[{"name":"BMW Group","role":"in-house"}],"summary":"When production equipment fails in a BMW plant, maintenance staff can ask Factory Genius, an in house assistant that searches equipment manuals, quality data, internal fault reports, planning documents and daily updated shift logs, shows the top matches with links to the sources, summarises the maintenance instructions and answers follow up questions in a chat. It also translates, which helped at the new plant in Debrecen where manuals are often not available in Hungarian. It grew out of a 2024 pilot in the Dingolfing body shop and parallel work in Spartanburg and Rosslyn, and BMW says it can now be used globally through an internal platform, in its initial development phase.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["de","en","hu"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.press.bmwgroup.com/global/article/detail/T0451072EN/%E2%80%9Cjust-ask-factory-genius-%E2%80%9D:-how-ai-helps-maintain-manufacturing-equipment?language=en","title":"“Just ask Factory Genius!”: How AI helps maintain manufacturing equipment","publisher":"BMW Group PressClub","date":"2025-07-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bmw-group-factory-genius-maintenance-assistant"},{"title":"BMW: Intelligent Personal Assistant with Amazon Alexa+ large language model technology","useCases":["in-car-ai-voice-assistant"],"organization":{"name":"BMW Group","anonymized":false,"country":"DE","region":"europe","industry":"automotive"},"vendors":[{"name":"Amazon","role":"platform"}],"summary":"BMW rebuilt its Intelligent Personal Assistant on Amazon's Alexa Custom Assistant with Alexa+ large language model technology, so drivers can hold natural conversations, ask several questions in one sentence about vehicle functions or general knowledge, and have the assistant recognize context. BMW introduced the German language version in the new BMW iX3 from mid April 2026 production; earlier iX3 vehicles were due to receive it by software update from the end of May 2026. BMW plans other markets and further models on BMW Operating System 9 and X from the second half of 2026. The assistant suggests routines from daily usage patterns, and from July 2026 BMW Operating System X adds new options for creating routines.","stage":"production","year":2026,"channels":["voice"],"languages":["de"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.press.bmwgroup.com/global/article/detail/T0458171EN/bmw-model-updates-summer-2026?language=en","title":"BMW model updates summer 2026","publisher":"BMW Group PressClub"},{"url":"https://www.press.bmwgroup.com/global/article/detail/T0454477EN/a-milestone-for-human-vehicle-interaction-bmw-intelligent-personal-assistant-expanded-to-include-amazon-alexa-technology?language=en","title":"A Milestone for Human-Vehicle Interaction. BMW Intelligent Personal Assistant expanded to include Amazon Alexa + Technology.","publisher":"BMW Group PressClub"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bmw-intelligent-personal-assistant-alexa-plus"},{"title":"BNSF Railway: AI wheel inspection and wayside detector analysis","useCases":["freight-rail-rolling-stock-predictive-maintenance"],"organization":{"name":"BNSF Railway","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[{"name":"BNSF Railway","role":"in-house"}],"summary":"BNSF Railway's condition based maintenance team uses thermal sensors and machine vision at wayside detectors to monitor more than 1.5 million wheels in motion across its network. Machine vision cameras inspect wheel surfaces for cracks and defects, and AI algorithms analyse the resulting wayside detector readings to predict maintenance needs before a breakdown, rather than reacting to a failure in service.","stage":"scaled","year":2025,"channels":["internal-tools","api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.bnsf.com/news-media/railtalk/innovation/artificial-intelligence.html","title":"Eyes on AI: BNSF innovates to better serve our customers","publisher":"BNSF Railway","date":"2025-01-09"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"bnsf-ai-wheel-and-wayside-inspection"},{"title":"BNY: Eliza AI platform resolves client transaction inquiries faster","useCases":["settlement-fail-prediction-and-exception-management"],"organization":{"name":"BNY","anonymized":false,"country":"US","region":"north-america","industry":"capital-markets"},"vendors":[{"name":"BNY","role":"in-house"},{"name":"Microsoft","role":"platform"}],"summary":"BNY built its own AI platform, Eliza, on Microsoft Azure and Microsoft Foundry, and uses it across operations. In a Microsoft customer story, BNY's Head of AI Enablement says more than ten percent of the inquiries clients raise about BNY's transactions were resolved or assisted by AI, with eighty percent faster processing of those inquiries. Microsoft's summary calls them client settlement inquiries; BNY's own words are broader, so the link to settlement exception work is adjacent rather than direct. The same story describes a digital employee that repairs incomplete payment instructions and an agentic workflow for client onboarding research.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":80,"unit":"percent","qualifier":"exact","period":"client transaction inquiries resolved or assisted by AI","claimant":"organization","quote":"More than ten percent of these inquiries were resolved or assisted by AI and that has resulted in eighty percent faster processing of these inquiries.","sourceUrl":"https://www.microsoft.com/en/customers/story/27322-bny-microsoft-365-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/27322-bny-microsoft-365-copilot","title":"Frontier Firm BNY resolves client inquires 80% faster with Microsoft AI powered Eliza","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"bny-eliza-client-settlement-inquiries"},{"title":"BNY: Eliza AI platform speeds up client onboarding research","useCases":["business-onboarding-and-ubo-discovery"],"organization":{"name":"BNY","anonymized":false,"country":"US","region":"north-america","industry":"capital-markets"},"vendors":[{"name":"BNY","role":"in-house"},{"name":"Microsoft","role":"platform"}],"summary":"BNY built its own AI platform, Eliza, on Microsoft Azure and Microsoft Foundry. In a Microsoft customer story, Saed Shonnar, Head of AI Enablement at BNY, says twenty five percent of the bank's new client onboardings this year were assisted with AI, resulting in a twenty percent faster onboarding process on average. A second BNY executive describes the research element of onboarding, the document processing and decision making behind verifying a new client, as the common challenge the AI addresses. The same story describes a digital employee that repairs incomplete payment instructions and faster processing of client settlement inquiries, which are reported as separate use cases.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":25,"unit":"percent","qualifier":"exact","period":"new client onboardings this year","claimant":"organization","quote":"Twenty-five percent of all of our new onboardings were assisted with AI this year and that has resulted in a twenty percent faster onboarding process on average for these clients,","sourceUrl":"https://www.microsoft.com/en/customers/story/27322-bny-microsoft-365-copilot"},{"kpi":"processing-time-reduction","value":20,"unit":"percent","qualifier":"exact","period":"new client onboardings this year, average","claimant":"organization","quote":"Twenty-five percent of all of our new onboardings were assisted with AI this year and that has resulted in a twenty percent faster onboarding process on average for these clients,","sourceUrl":"https://www.microsoft.com/en/customers/story/27322-bny-microsoft-365-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/27322-bny-microsoft-365-copilot","title":"Frontier Firm BNY resolves client inquires 80% faster with Microsoft AI powered Eliza","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"bny-eliza-onboarding-research"},{"title":"BNY: Eliza digital employee repairs over 10% of payment instructions worldwide","useCases":["payment-investigations-and-exceptions"],"organization":{"name":"BNY","anonymized":false,"country":"US","region":"north-america","industry":"capital-markets"},"vendors":[{"name":"BNY","role":"in-house"},{"name":"Microsoft","role":"platform"}],"summary":"BNY built its own AI platform, Eliza, on Microsoft Azure and Microsoft Foundry. In a Microsoft customer story, a BNY executive describes a digital employee that repairs missing or incomplete payment instructions so the payment can proceed without a time consuming manual review. The same executive says that digital employee now handles over ten percent of BNY's payment repair issues worldwide. The same story describes an agentic workflow for client onboarding research and faster processing of client settlement inquiries, which are reported as separate use cases.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":10,"unit":"percent","qualifier":"at-least","period":"payment repair issues worldwide","claimant":"organization","quote":"Today, that digital employee handles over ten percent of our payment repair issues around the world.","sourceUrl":"https://www.microsoft.com/en/customers/story/27322-bny-microsoft-365-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/27322-bny-microsoft-365-copilot","title":"Frontier Firm BNY resolves client inquires 80% faster with Microsoft AI powered Eliza","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"bny-eliza-payment-repair-digital-employee"},{"title":"BNY Mellon: machine learning to predict US Treasury settlement fails, with Google Cloud","useCases":["settlement-fail-prediction-and-exception-management"],"organization":{"name":"BNY","anonymized":false,"country":"US","region":"north-america","industry":"capital-markets"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"BNY","role":"in-house"}],"summary":"On 4 February 2021 BNY, then branded BNY Mellon, announced a collaboration with Google Cloud to predict settlement failures in the US Treasury market, where it provides clearance and settlement. It trains models on millions of trades on Google Cloud's data analytics and machine learning services. BNY Mellon's Clearance and Collateral Management head described the aim as helping clients predict approximately 40% of settlement failures in Fed eligible securities with 90% accuracy. The release describes a solution in development; no production results were published.","stage":"announced","year":2021,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.bny.com/corporate/global/en/about-us/newsroom/company-news/bny-mellon-and-google-cloud-collaborate-to-help-transform-us-treasury-market-settlement-and-clearance-process.html","title":"BNY Mellon and Google Cloud Collaborate to Help Transform U.S. Treasury Market Settlement and Clearance Process","publisher":"BNY","date":"2021-02-04"},{"url":"https://www.googlecloudpresscorner.com/22021-02-04-BNY-Mellon-and-Google-Cloud-Collaborate-to-Help-Transform-U-S-Treasury-Market-Settlement-and-Clearance-Process","title":"BNY Mellon and Google Cloud Collaborate to Help Transform U.S. Treasury Market Settlement and Clearance Process","publisher":"Google Cloud","date":"2021-02-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bny-mellon-treasury-settlement-fail-prediction"},{"title":"Booking.com: AI Trip Support, AI Voice Support and AI Trip Planner","useCases":["travel-and-hotel-booking-concierge"],"organization":{"name":"Booking.com","anonymized":false,"country":"NL","region":"global","industry":"travel-and-hospitality"},"vendors":[],"summary":"Booking.com has moved from a standalone AI Trip Planner to AI features embedded across the booking journey. AI Trip Support is a first point of contact around the clock that answers questions about a property (for example parking) and helps travellers change reservations, with a handover to a human for complex cases. AI Voice Support lets travellers manage or cancel a booking by phone in their own words, and connects to a human agent with the context when needed. An earlier generation, the Booking Assistant chatbot, answered stay related questions from 2017.","stage":"production","year":2025,"channels":["mobile-app","web-chat","voice"],"languages":[],"metrics":[{"kpi":"automation-rate","value":30,"unit":"percent","qualifier":"exact","period":"December 2017, pilot version of the earlier generation Booking Assistant, English language bookings","claimant":"organization","quote":"The chatbot can currently respond to 30% of customers’ stay-related questions automatically in less than 5 minutes.","sourceUrl":"https://news.booking.com/bookingcom-expands-global-access-to-the-booking-assistant/"}],"outcomeDisclosed":true,"sources":[{"url":"https://news.booking.com/bookingcom-debuts-agentic-ai-innovations-adding-to-its-robust-suite-of-genai-tools-for-customers/","title":"Booking.com Debuts Agentic AI Innovations, Adding to its Robust Suite of GenAI Tools for Customers","publisher":"Booking.com","date":"2025-10-09"},{"url":"https://news.booking.com/bookingcom-expands-global-access-to-the-booking-assistant/","title":"Booking.com Expands Global Access to the Booking Assistant","publisher":"Booking.com","date":"2017-12-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"booking-com-ai-trip-support-and-voice"},{"title":"Boomi: synthetic data warehouse for employees building AI agents","useCases":["synthetic-test-data-generation"],"organization":{"name":"Boomi","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Tonic.ai","role":"platform"}],"summary":"Boomi, an integration platform company, gave employees in its citizen developer programme a synthetic copy of its Snowflake data warehouse, generated with Tonic Structural, so they can build and test AI agents and workflows without touching customer data. When a prototype is approved, the enterprise AI team points it at the production warehouse. Refreshes run automatically each weekend, and five agents were being rolled out when the story was published.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":2000,"unit":"count","qualifier":"approximately","period":"Employees with access to the synthetic environment","claimant":"vendor","quote":"With de-identified data from Tonic Structural, Boomi has opened its citizen developer program to approximately 2,000 people.","sourceUrl":"https://www.tonic.ai/case-study/boomi-synthetic-data-unblocks-ai-agent-innovation"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.tonic.ai/case-study/boomi-synthetic-data-unblocks-ai-agent-innovation","title":"How Boomi unblocked 2,000 citizen developers to build AI agents with Tonic Structural","publisher":"Tonic.ai","date":"2026-07-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"boomi-synthetic-data-environment"},{"title":"Bosch Digital: Gemini models for localizing marketing content across business units","useCases":["marketing-and-product-content-localization"],"organization":{"name":"Bosch Digital","anonymized":false,"region":"europe","industry":"manufacturing"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Google Cloud lists Bosch Digital, which it describes as one of Europe's leading technology and services companies, as using Gemini models to localize marketing content for different markets and demographics, with many business units having started. The entry mentions time and cost savings but gives no figures.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","archivedUrl":"https://web.archive.org/web/20250104231021/https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"bosch-gemini-marketing-localization"},{"title":"Bouygues Telecom: Iris answer agent for contact centre reps and store associates","useCases":["retail-store-and-kiosk-assistant"],"organization":{"name":"Bouygues Telecom","anonymized":false,"country":"FR","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Salesforce","role":"platform"}],"summary":"Bouygues Telecom built Iris, an employee facing agent on Salesforce Agentforce, over a rewritten knowledge base of more than 500 articles and a unified customer profile. Its 6,000 service reps ask Iris questions about billing, technical issues and promotions during calls, and the same answers are now available to associates in 500 retail stores. Before launch, Bouygues Telecom required Iris to exceed 90% accuracy or outperform a supervisor; Salesforce reports 95% on day one. Reps stay in control of what the customer hears.","stage":"scaled","year":2026,"channels":["agent-desktop","internal-tools"],"languages":["fr"],"metrics":[{"kpi":"accuracy","value":95,"unit":"percent","qualifier":"exact","period":"at launch, against a 90% threshold","claimant":"vendor","quote":"Iris cleared it on day one, reaching 95% accuracy.","sourceUrl":"https://www.salesforce.com/customer-stories/bouygues-telecom/agentic-service-faqs/"},{"kpi":"users-served","value":6000,"unit":"count","qualifier":"exact","period":"contact centre service reps using Iris","claimant":"vendor","quote":"Today, 90% of the 6,000 service reps who use Iris rate it four or five stars — a clear signal of trust in the answers it delivers.","sourceUrl":"https://www.salesforce.com/customer-stories/bouygues-telecom/agentic-service-faqs/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.salesforce.com/customer-stories/bouygues-telecom/agentic-service-faqs/","title":"Bouygues Telecom delivers answers in seconds with agentic service","publisher":"Salesforce"},{"url":"https://www.salesforce.com/customer-stories/bouygues-telecom/","title":"Bouygues Telecom serves more, faster as an Agentic Enterprise","publisher":"Salesforce"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"bouygues-telecom-iris-service-agent"},{"title":"Bowhead Specialty: AI underwriting workbench for casualty submissions with Kalepa","useCases":["underwriting-risk-assessment-copilot"],"organization":{"name":"Bowhead Specialty","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Kalepa","role":"platform"}],"summary":"Bowhead's casualty underwriters use Kalepa's AI underwriting platform to bring research, data and underwriting guidelines into one workspace, surface information missing from broker submissions and apply appetite consistently. Bowhead's head of casualty says the book profile has improved because the tool helps avoid heightened risk profiles that were not visible in the broker's submission; no figures are given. Bowhead's own 2025 annual report states that it does not currently use generative AI tools, so the platform should be read as data and analytics support rather than a generative assistant.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.kalepa.com/case-studies/driving-profitable-growth-bowhead-specialtys-results-with-kalepa","title":"Driving Profitable Growth: Bowhead Specialty's Results with Kalepa","publisher":"Kalepa"},{"url":"https://www.sec.gov/Archives/edgar/data/2002473/000162828026011089/bow-20251231.htm","title":"Bowhead Specialty Holdings Inc. Form 10-K for 2025","publisher":"Bowhead Specialty via SEC EDGAR","date":"2026-02-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"bowhead-specialty-kalepa-underwriting-workbench"},{"title":"BPI: BEA Chat, an always on digital assistant for general banking inquiries","useCases":["first-line-contact-centre-agent"],"organization":{"name":"Bank of the Philippine Islands","anonymized":false,"country":"PH","region":"asia-pacific","industry":"banking"},"vendors":[],"summary":"Bank of the Philippine Islands runs BEA Chat, a conversational AI assistant on its website and Facebook page that answers general inquiries around the clock, offers self service for common concerns, lets customers apply for products and track service requests, and escalates complex issues to live agents. It serves both logged in clients and guests, and the bank positions it as a channel for Filipinos working overseas who would otherwise need a local SIM card or international calls. No containment or cost outcome is disclosed.","stage":"production","year":2025,"channels":["web-chat","social-messaging"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.bpi.com.ph/about-bpi/news/bpi-enhances-bea-chat-with-smarter-ai-powered-features-for-customer-support","title":"BPI enhances BEA Chat with smarter, AI-powered features for customer support","publisher":"Bank of the Philippine Islands","date":"2025-08-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bpi-bea-chat-digital-assistant"},{"title":"British Gas: automated quality and regulatory assurance of voice and webchat","useCases":["call-quality-and-compliance-monitoring"],"organization":{"name":"British Gas","anonymized":false,"country":"GB","region":"europe","industry":"energy-and-utilities"},"vendors":[{"name":"CallMiner","role":"platform"}],"summary":"British Gas, which runs a 20 million contact operation across voice and webchat, used to assess quality and regulatory compliance manually, which limited volume and consistency. With CallMiner it now runs millions of automated assessments across voice and webchat against its Five Steps to Customer Excellence framework and its Ofgem regulatory checks, and feeds the results into agent coaching. The automated scores had to reach at least 80% agreement with human reviewers before going live, and agents and team leaders can challenge scores. The vendor reports that quality scores improved by about 10% and that several regulatory scores now meet or exceed target.","stage":"scaled","year":2026,"channels":["voice","web-chat"],"languages":["en"],"metrics":[{"kpi":"quality-score-uplift","value":10,"unit":"percent","qualifier":"approximately","period":"quality scores against the Five Steps to Customer Excellence framework","claimant":"vendor","quote":"Quality scores against the Five Steps to Customer Excellence framework have improved by approximately 10%, with a general upward trend.","sourceUrl":"https://callminer.com/customers/stories/british-gas-scales-quality-and-regulatory-assurance"}],"outcomeDisclosed":true,"sources":[{"url":"https://callminer.com/customers/stories/british-gas-scales-quality-and-regulatory-assurance","title":"British Gas Scales Quality & Regulatory Assurance from Thousands to Millions with CallMiner","publisher":"CallMiner"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"british-gas-quality-and-regulatory-assurance"},{"title":"BrowserStack: test case generation from user context in its test platform","useCases":["requirements-to-test-case-generation"],"organization":{"name":"BrowserStack","anonymized":false,"country":"IN","region":"asia-pacific","industry":"technology"},"vendors":[{"name":"Microsoft","role":"model-provider"}],"summary":"BrowserStack, a cloud testing platform, added Azure OpenAI based features that recommend and generate test cases from context the user provides, convert them into automated scripts and repair tests when the user interface changes. The figures on the page are product claims for the platform's customers in general, not a measured result at one named customer, and no baseline is given.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":70,"unit":"percent","qualifier":"up-to","period":"QA cycle time, product claim across customers","claimant":"vendor","quote":"The platform intelligently creates and maintains test cases, converting them into automated scripts, reducing QA cycle time by up to 70%.","sourceUrl":"https://www.microsoft.com/en-in/aifirstmovers/browserstack"},{"kpi":"accuracy","value":90,"unit":"percent","qualifier":"at-least","period":"Coverage accuracy of recommended test cases, product claim","claimant":"vendor","quote":"Leveraging Azure OpenAI, BrowserStack platform recommends test cases based on user-provided context, ensuring comprehensive coverage with 90%+ accuracy.","sourceUrl":"https://www.microsoft.com/en-in/aifirstmovers/browserstack"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en-in/aifirstmovers/browserstack","title":"Browserstack: Enhance testing with AI-driven productivity","publisher":"Microsoft India"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"browserstack-ai-test-case-generation"},{"title":"BT Group: machine learning cell sleep across EE mobile sites","useCases":["ran-energy-optimization"],"organization":{"name":"BT Group","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[],"summary":"After trials in each of the UK's home nations, BT Group rolled out cell sleep software to more than 19,500 EE mobile sites. It puts 4G capacity carriers to sleep during quiet periods that machine learning has predicted for each site, wakes them automatically at busy times and within seconds when traffic surges unexpectedly, and can use a deeper sleep mode overnight. The sleep functions come from the radio equipment suppliers; BT Group's site data drives the statistical algorithms that control them. BT Group published an expected annual energy saving rather than a measured result.","stage":"scaled","year":2024,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.bt.com/bt-group-rolls-out-energy-saving-cell-sleep-technology-to-ee-mobile-sites-nationwide/","title":"BT Group rolls-out energy-saving ‘cell sleep’ technology to EE mobile sites nationwide","publisher":"BT Group","date":"2024-06-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bt-ee-cell-sleep-energy-saving"},{"title":"BT: Enhanced Call Protect scam and spam screening on Digital Voice landlines","useCases":["spam-and-scam-call-blocking"],"organization":{"name":"BT Group","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Hiya","role":"platform"}],"summary":"BT's Digital Voice home phone service includes Enhanced Call Protect, an AI powered tool from Hiya that monitors incoming calls, diverts scam calls to a junk voicemail and shows a \"Nuisance?\" warning on the landline display for suspected spam, while showing the name of registered businesses. In its first four months it blocked more than 2.4 million scam calls and identified about 17.7 million spam calls. BT also runs an AI network level firewall against calls from abroad that use a UK number for scam purposes.","stage":"scaled","year":2024,"channels":["voice"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":2430000,"unit":"count","qualifier":"at-least","period":"May to early October 2024, scam calls blocked","claimant":"organization","quote":"BT’s new Enhanced Call Protect on Digital Voice has successfully blocked more than 2,430,000 scam and identified 17,700,000 spam calls to landlines since the new scam protection service from Hiya was introduced in May.","sourceUrl":"https://newsroom.bt.com/bts-new-home-phone-scam-protection-service-stops-201-million-scam-and-spam-attempts-in-first-4-months/"},{"kpi":"users-served","value":2500000,"unit":"count","qualifier":"exact","period":"October 2024","claimant":"organization","quote":"2.5 million BT customers already receive the new call vetting service, a benefit of migrating to Digital Voice.","sourceUrl":"https://newsroom.bt.com/bts-new-home-phone-scam-protection-service-stops-201-million-scam-and-spam-attempts-in-first-4-months/"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.bt.com/bts-new-home-phone-scam-protection-service-stops-201-million-scam-and-spam-attempts-in-first-4-months/","title":"BT’s new home phone scam protection service stops 20.1 million scam and spam attempts in first 4 months","publisher":"BT Group","date":"2024-10-03"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"bt-enhanced-call-protect"},{"title":"BT Group: EE virtual assistant Aimee with generative AI","useCases":["bill-explanation-and-billing-dispute-agent","first-line-contact-centre-agent"],"organization":{"name":"BT Group","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Sprinklr","role":"platform"}],"summary":"BT Group runs the EE virtual assistant Aimee on Sprinklr's customer experience platform, drawing on BT Group data for personalised answers. The platform lets BT Group use generative AI for EE and BT customers, for example in an Aimee journey that prepares customers for international travel and in billing support, where generative AI gives detailed explanations of billing charges. BT Group says Aimee handles up to 60,000 conversations a week, double the volume of two years earlier, that the travel journey halved the need for chat support, and that it stays model agnostic behind a private cloud instance with safeguards against attempts to make the AI misbehave.","stage":"scaled","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":60000,"unit":"count","qualifier":"up-to","period":"per week","claimant":"organization","quote":"EE virtual assistant Aimee now handles up to 60,000 customer conversations per week, with automation success rates on several types of customer journey now approaching 50%, freeing time for guides to focus on more complex issues","sourceUrl":"https://newsroom.bt.com/bt-group-leans-on-ai-to-transform-customer-service-experience/"},{"kpi":"automation-rate","value":50,"unit":"percent","qualifier":"up-to","period":"several types of customer journey","claimant":"organization","quote":"EE virtual assistant Aimee now handles up to 60,000 customer conversations per week, with automation success rates on several types of customer journey now approaching 50%, freeing time for guides to focus on more complex issues","sourceUrl":"https://newsroom.bt.com/bt-group-leans-on-ai-to-transform-customer-service-experience/"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.bt.com/bt-group-leans-on-ai-to-transform-customer-service-experience/","title":"BT Group leans on AI to transform customer service experience","publisher":"BT Group","date":"2024-12-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bt-group-ee-aimee-virtual-assistant"},{"title":"bunq: data analysis and machine learning in place of customer self reporting in due diligence","useCases":["dynamic-customer-risk-rating"],"organization":{"name":"bunq","anonymized":false,"country":"NL","region":"europe","industry":"banking"},"vendors":[{"name":"bunq","role":"in-house"}],"summary":"As described in a 2022 ruling, Dutch neobank bunq assigned new private customers a \"regular user profile\" derived from data analysis of its customer base, instead of asking each customer about the purpose of the account, and then monitored transaction behaviour to adjust that profile and raise the customer's risk scores when needed. bunq says it favours technology such as machine learning; the ruling mentions a machine learning model in transaction monitoring but does not show that machine learning produces the rating itself. In October 2022 the Dutch Trade and Industry Appeals Tribunal (CBb) ruled largely in bunq's favour in its dispute with De Nederlandsche Bank over this approach. In May 2025 De Nederlandsche Bank fined bunq EUR 2.6 million because it had not sufficiently followed up signals in four customer files it had itself identified as high risk; bunq has appealed, and the notice does not attribute the shortcomings to the data driven approach.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en","nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://press.bunq.com/219361-bunq-wins-appeal-against-dutch-central-bank-dnb/","title":"bunq wins appeal against Dutch Central Bank (DNB)","publisher":"bunq","date":"2022-10-18"},{"url":"https://uitspraken.rechtspraak.nl/details?id=ECLI:NL:CBB:2022:707","title":"ECLI:NL:CBB:2022:707, College van Beroep voor het bedrijfsleven, 18-10-2022, 21/323 en 21/1108","publisher":"College van Beroep voor het bedrijfsleven","date":"2022-10-18"},{"url":"https://www.dnb.nl/en/general-news/enforcement-measures-2025/fine-for-bunq-b-v-for-insufficient-customer-due-diligence/","title":"Fine for bunq B.V. for insufficient customer due diligence","publisher":"De Nederlandsche Bank","date":"2025-08-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"bunq-machine-learning-customer-due-diligence"},{"title":"UK Cabinet Office: Assist, a generative AI tool for government communicators","useCases":["civil-servant-drafting-copilot"],"organization":{"name":"Cabinet Office (Government Communication Service)","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Anthropic","role":"model-provider"},{"name":"Amazon Web Services (Amazon Bedrock)","role":"platform"}],"summary":"Assist is a bespoke generative AI tool for members of the UK government communications profession. It offers pre built prompts for typical communications tasks, drafts first versions, helps with brainstorming and review, and can ground its answers in Government Communications policies and standards through retrieval. It is not to be used for decisions, and users remain responsible for every output. It is in production and was open to about 8,500 members of the profession in July 2026. The transparency record reports that users save around 3 hours a week on average and that 98% of users find it useful in their role, both from user research.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"hours-saved","value":3,"unit":"hours","qualifier":"approximately","period":"per user per week, from user research","claimant":"organization","quote":"User research shows that Assist users on average save around 3 hours per week by using the tool.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/cabinet-office-assist"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/cabinet-office-assist","title":"Cabinet Office: Assist (algorithmic transparency record)","publisher":"GOV.UK","date":"2026-07-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"cabinet-office-assist-government-communications"},{"title":"Cammeby's International: AI HVAC optimization with BrainBox AI","useCases":["building-energy-optimization"],"organization":{"name":"Cammeby's International","anonymized":false,"country":"US","region":"north-america","industry":"real-estate"},"vendors":[{"name":"BrainBox AI","role":"platform"}],"summary":"Cammeby's International, a real estate investment company, deployed BrainBox AI's AI Control solution across a 32 storey, 386,315 square foot office building in New York City's financial district, built in 1983, controlling 251,104 square feet of air handling units, variable air valves, outdoor and exhaust fans, and the hot and chilled water systems. Over an 11 month period in 2023, the deployment cut HVAC related electricity consumption, cost and emissions without disrupting daily operations, according to the building's facility engineers.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"energy-savings","value":15.8,"unit":"percent","qualifier":"exact","period":"over an 11 month period in 2023","claimant":"vendor","quote":"Our AI Control solution drove a 15.8% reduction in HVAC-related electricity consumption, saving $42,951, and mitigating 37.14 tCO2eq.","sourceUrl":"https://brainboxai.com/en/case-studies/cammebys-achieves-15.8-reduction-in-hvac-energy-use-and-costs"},{"kpi":"cost-savings","value":42951,"unit":"currency","currency":"USD","qualifier":"exact","period":"over an 11 month period in 2023","claimant":"vendor","quote":"Our AI Control solution drove a 15.8% reduction in HVAC-related electricity consumption, saving $42,951, and mitigating 37.14 tCO2eq.","sourceUrl":"https://brainboxai.com/en/case-studies/cammebys-achieves-15.8-reduction-in-hvac-energy-use-and-costs"}],"outcomeDisclosed":true,"sources":[{"url":"https://brainboxai.com/en/case-studies/cammebys-achieves-15.8-reduction-in-hvac-energy-use-and-costs","title":"Cammeby's International achieves 15.8% reduction in HVAC energy use and costs","publisher":"BrainBox AI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"cammebys-international-brainbox-ai-hvac"},{"title":"Capital One: Eno assistant alerts on unexpected card charges","useCases":["fraud-alert-confirmation","proactive-outbound-engagement-agent"],"organization":{"name":"Capital One","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Capital One","role":"in-house"}],"summary":"Eno is Capital One's virtual assistant. Capital One says it helps protect card accounts by looking out for charges that might surprise the customer, and sends insights when it spots free trials and recurring charges, through text, email and app alerts. Capital One does not publish outcome figures for Eno on this page.","stage":"production","year":2026,"channels":["sms","email","mobile-app"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.capitalone.com/digital/tools/eno/","title":"Eno, your Capital One assistant","publisher":"Capital One"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"capital-one-eno-assistant"},{"title":"Care Quality Commission: risk categorisation to prioritise assessment of health and care services","useCases":["inspection-prioritization"],"organization":{"name":"Care Quality Commission","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[],"summary":"The Care Quality Commission, the health and social care regulator for England, gives each assessment service group of a registered location or provider a monthly risk category (medium, high or very high) that inspectors use to prioritise assessment activity. The category combines outputs of four sector risk models (adult social care, general practice, independent healthcare, urgent and emergency care), built on data such as statutory notifications, whistleblowing, safeguarding concerns and public feedback, with rules based on the age and level of the current rating. Inspectors see the main data drivers, local insight can override the score, and the output is tested against inspection outcomes. The record adds that a machine learning model to prioritise care home inspections is in development but not yet in production.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":46000,"unit":"count","qualifier":"approximately","period":"refreshed risk scores per month","claimant":"organization","quote":"We are currently producing approxiately 46000 refreshed scores each month.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/care-quality-commission-risk-categorisation"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/care-quality-commission-risk-categorisation","title":"Care Quality Commission: Risk Categorisation (algorithmic transparency record)","publisher":"GOV.UK","date":"2026-03-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"care-quality-commission-inspection-risk-categorisation"},{"title":"Carlsberg: supply chain training produced in house with AI video","useCases":["training-content-generation"],"organization":{"name":"Carlsberg Group","anonymized":false,"country":"DK","region":"global","industry":"manufacturing"},"vendors":[{"name":"Synthesia","role":"platform"}],"summary":"Since 2023, Carlsberg's Integrated Supply Chain Academy has built training in house with Synthesia instead of external video agencies, and use has spread to shop floor onboarding, one point safety lessons, procurement training and change management content. A document, usually a PDF, goes into the tool's AI Assistant, which scaffolds the structure; the video is built in a branded template, generated in English and translated for each market. Synthesia reports that more than 100 employees have built content in under two years, that three agency shoots a year (its conservative baseline, at least €30,000 in fees) no longer sit on Carlsberg's P&L, and that the second supplier once hired to translate each eLearning is no longer needed. These are vendor stated figures, not a measured saving net of licence costs.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":100,"unit":"count","qualifier":"at-least","period":"in under two years","baseline":"employees across the business who have built content in Synthesia (creators, not learners)","claimant":"vendor","quote":"In under two years, more than 100 employees from across the business have built content in Synthesia, with adoption still growing.","sourceUrl":"https://www.synthesia.io/case-studies/carlsberg"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.synthesia.io/case-studies/carlsberg","title":"Carlsberg takes supply chain training in-house with AI video","publisher":"Synthesia","date":"2026-09-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"carlsberg-ai-video-supply-chain-training"},{"title":"CarMax: AI voice agent on inbound sales calls","useCases":["inbound-lead-qualification-agent"],"organization":{"name":"CarMax","anonymized":false,"country":"US","region":"north-america","industry":"automotive"},"vendors":[{"name":"Sierra","role":"platform"}],"summary":"CarMax, the largest used car retailer in the United States, deployed AI voice agents on its inbound sales calls in 2026. The agent asks clarifying questions to understand what the caller needs, answers common questions such as store hours and vehicle availability, and hands the caller to the right associate faster, whatever the call volume or time zone. CarMax reports more calls resolved and fewer unresolved calls without giving figures, and plans appointment management for appraisals, browsing and test drives. It already runs a web virtual assistant, Skye, on CarMax.com.","stage":"production","year":2026,"channels":["voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://sierra.ai/customers/carmax","title":"CarMax teams with Sierra to enhance inbound sales call experience","publisher":"Sierra","date":"2026-08-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"carmax-inbound-sales-voice-agent"},{"title":"Case Status: AI performance review cycle with Windmill","useCases":["performance-review-drafting-agent"],"organization":{"name":"Case Status","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Windmill","role":"platform"}],"summary":"Case Status, a client experience platform for law firms, used Windmill's AI review agent to solve what Windmill calls the \"blank page problem\": managers and employees starting a review with an empty form and having to reconstruct months of work from memory. Windmill surfaces work context from Slack and other connected tools and generates a summary that becomes the starting point for every review, replacing a prior process built on Google Docs and forms.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":84,"unit":"percent","qualifier":"exact","period":"full review cycle, versus the prior Google Docs and forms process","claimant":"vendor","quote":"Case Status completed their entire performance review cycle in 84% less time compared to their previous process, while dramatically improving the experience.","sourceUrl":"https://gowindmill.com/customers/case-status"}],"outcomeDisclosed":true,"sources":[{"url":"https://gowindmill.com/customers/case-status","title":"How Case Status Ran Fast Performance Reviews That Solved the Blank Page Problem with Windmill","publisher":"Windmill","date":"2025-01-21"},{"url":"https://gowindmill.com/resources/lists/companies-using-ai-performance-management/","title":"How 6 Companies Use AI for Performance Management","publisher":"Windmill","date":"2026-05-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"case-status-ai-performance-review-cycle"},{"title":"Catchtable: personalized restaurant recommendations that lift reservation conversion","useCases":["personalized-marketing-at-scale"],"organization":{"name":"Catchtable","anonymized":false,"country":"KR","region":"asia-pacific","industry":"travel-and-hospitality"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Catchtable, a restaurant discovery and reservation app from South Korea, built personalized recommendation models on Vertex AI that combine language models with custom embeddings to read search intent and recommend restaurants in real time. Google Cloud reports a 30% increase in reservation conversion rates and a 150% increase in impressions per restaurant search.","stage":"production","year":2025,"channels":["mobile-app"],"languages":[],"metrics":[{"kpi":"conversion-rate-uplift","value":30,"unit":"percent","qualifier":"exact","period":"reservation conversion rate","claimant":"vendor","quote":"Catchtable uses Vertex AI and Kubeflow with GPU optimization to build personalized restaurant recommendation models, achieving a 30% increase in reservation conversion rates and a 150% increase in impressions per restaurant search.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"},{"url":"https://play.google.com/store/apps/details?id=kr.co.catchtable.global.catchtable_global&hl=en","title":"CATCHTABLE: Book Restaurants","publisher":"WAD Corp.","date":"2026-08-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"catchtable-personalized-recommendations"},{"title":"US Centers for Disease Control and Prevention: generative AI stance analysis of public comments on proposed rules","useCases":["public-consultation-response-analysis"],"organization":{"name":"Centers for Disease Control and Prevention","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"CDC reports in the 2025 federal AI use case inventory that it uses generative AI to analyse public comments on its proposed rules. For each comment the system records a stance (support, oppose or neutral), topics and sentiment, which regulatory analysts use when they review and summarise the feedback for the rulemaking record. The inventory lists the use case as deployed since July 2023. No outcome figures are published.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"cdc-public-comment-stance-analysis"},{"title":"US Centers for Disease Control and Prevention: SmartFind knowledge bot for partner mailbox replies","useCases":["support-knowledge-article-generation","email-and-ticket-reply-drafting"],"organization":{"name":"Centers for Disease Control and Prevention","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"CDC's National Center for Immunization and Respiratory Diseases runs SmartFind, an internal knowledge bot with a SharePoint component that helps program staff manage partner emails and lets mailbox managers use a shared knowledge base. The agency lists partner mailbox email management and knowledge base maintenance as the purpose. The bot matches free text questions to agency cleared answers and flags complex or unanswerable queries for manual review. Earlier public facing versions gave agency cleared answers to public and partner questions during the COVID-19 pandemic. No outcome figures are published.","stage":"production","year":2024,"channels":["email","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (HHS entry, NCIRD SmartFind ChatBots, Public and Internal)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"cdc-smartfind-knowledge-bot"},{"title":"Central Bank: automated QA across all contact centre calls","useCases":["call-quality-and-compliance-monitoring"],"organization":{"name":"Central Bank","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Observe.AI","role":"platform"}],"summary":"Central Bank, a group of community banks serving several states, mostly in the Midwest, runs a customer service centre with more than 3,000 interactions a day. It replaced manual QA sampling with Observe.AI's automated QA, searchable transcripts and AI tagging of call reasons, which also replaced manual disposition codes. The team went from evaluating a handful of calls a month to evaluating every call, and used the insight on agent behaviours to lower handling time.","stage":"scaled","year":2024,"channels":["voice"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":167000,"unit":"count","qualifier":"exact","period":"third quarter of 2024","baseline":"24 calls evaluated per quarter before automated QA","claimant":"vendor","quote":"In the third quarter of 2024, the team evaluated 167,000 calls, a jump from just eight per month or 24 per quarter before adopting Auto QA.","sourceUrl":"https://www.observe.ai/customers/central-bank"},{"kpi":"handling-time-reduction","value":5,"unit":"percent","qualifier":"up-to","claimant":"vendor","quote":"Identifying key agent behaviors has helped the CSC team develop data-driven plans and reduce average handle time by up to 5%, resulting in significant efficiency gains.","sourceUrl":"https://www.observe.ai/customers/central-bank"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.observe.ai/customers/central-bank","title":"Central Bank streamlines call handling and boosts efficiency with Post-Interaction AI","publisher":"Observe.AI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"central-bank-automated-call-quality"},{"title":"Central Insurance: AI subrogation detection","useCases":["subrogation-opportunity-detection"],"organization":{"name":"Central Insurance","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Shift Technology","role":"platform"}],"summary":"Central Insurance, a US insurer, added Shift Subrogation Detection in 2023 after three years of using Shift to detect claims fraud. The system reviews claims for signs that a third party is wholly or partly responsible, often on the first day of the claim, and gives the recovery team alerts with the facts, comparative negligence rules and state recovery laws for PIP and medical payments. Handlers still refer claims to subrogation; the AI catches the ones they miss. Its claims recovery supervisor reports more referrals and hours saved every week, but no figures were published.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.shift-technology.com/resources/news/central-insurance-expands-relationship-with-shift-technology","title":"Central Insurance Expands Relationship with Shift Technology","publisher":"Shift Technology","date":"2023-06-28"},{"url":"https://www.shift-technology.com/resources/reports-and-insights/subrogation-central-insurance","title":"Central Insurance: The benefits of AI in subrogation","publisher":"Shift Technology"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"central-insurance-subrogation-detection"},{"title":"CFTC: machine learning spoofing detection pilot, retired","useCases":["market-abuse-surveillance-triage"],"organization":{"name":"Commodity Futures Trading Commission","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The CFTC Division of Enforcement ran a pilot in 2023 that applied supervised and unsupervised machine learning to order message data to find spoofing patterns, trained with the division's existing expert based spoofing detection algorithms. The model gave a probability of spoofing behaviour for every trader in a given market on a given day. The 2024 federal AI inventory lists the project as retired in April 2024; no results were published.","stage":"paused","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2024-Federal-AI-Use-Case-Inventory","title":"2024 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI inventory (entry Spoofing Detection AI/ML Project)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"cftc-spoofing-detection-pilot"},{"title":"C.H. Robinson: Lean AI across the shipment lifecycle","useCases":["freight-dispatch-and-load-matching-agent"],"organization":{"name":"C.H. Robinson","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[{"name":"C.H. Robinson","role":"in-house"}],"summary":"C.H. Robinson, a global freight broker and third party logistics provider, says it has applied what it calls Lean AI across the shipment lifecycle since 2022: pricing, order entry, capacity sourcing, pickup and delivery appointments, freight tracking, document handling and invoicing. Generative AI reads incoming emails, classifies them and, for a truckload quote request, replies with a price; a separate system matches load and dock details to book touchless pickup and delivery appointments. The company reports the approach has automated millions of shipping tasks and raised productivity, while continuing to provide premium service.","stage":"scaled","year":2026,"channels":["api","internal-tools","email"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":40,"unit":"percent","qualifier":"at-least","period":"since 2022","claimant":"organization","quote":"Lean AI has increased C.H. Robinson's productivity by more than 40%, automated millions of shipping tasks, saved thousands of hours of work per day and lowered its cost to serve, while continuing to provide premium service.","sourceUrl":"https://www.chrobinson.com/en-us/about-us/newsroom/news/2026/ai-disruptor-ch-robinson/"},{"kpi":"interactions-handled","value":2000,"unit":"count","qualifier":"exact","period":"per day","claimant":"organization","quote":"While the technology is replying to 2,000 customer quote requests a day, it opens the door to automating other transactions shippers and carriers choose to do by email.","sourceUrl":"https://investor.chrobinson.com/News-and-Events/Press-Releases/press-release-details/2024/New-C.H.-Robinson-Technology-Breaks-a-Decades-Old-Barrier-to-Automation-in-the-Logistics-Industry/default.aspx"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.chrobinson.com/en-us/about-us/newsroom/news/2026/ai-disruptor-ch-robinson/","title":"With AI, C.H. Robinson is the disruptor","publisher":"C.H. Robinson","date":"2026-02-13"},{"url":"https://investor.chrobinson.com/News-and-Events/Press-Releases/press-release-details/2024/New-C.H.-Robinson-Technology-Breaks-a-Decades-Old-Barrier-to-Automation-in-the-Logistics-Industry/default.aspx","title":"New C.H. Robinson Technology Breaks a Decades-Old Barrier to Automation in the Logistics Industry","publisher":"C.H. Robinson","date":"2024-05-07","archivedUrl":"https://web.archive.org/web/20241016003806/https://investor.chrobinson.com/News-and-Events/Press-Releases/press-release-details/2024/New-C.H.-Robinson-Technology-Breaks-a-Decades-Old-Barrier-to-Automation-in-the-Logistics-Industry/default.aspx"},{"url":"https://www.truckingdive.com/news/ch-robinson-ai-automate-emailed-price-quotes-touchless-appointments/714925/","title":"How CH Robinson is using AI to automate logistics tasks","publisher":"Trucking Dive","date":"2024-05-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"ch-robinson-lean-ai-freight-automation"},{"title":"Chicago Transit Authority: Chat with CTA","useCases":["public-transit-passenger-information-agent"],"organization":{"name":"Chicago Transit Authority","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[{"name":"Google Public Sector","role":"platform"},{"name":"Quantiphi","role":"integrator"}],"summary":"The Chicago Transit Authority (CTA), the independent government agency that runs Chicago's buses and trains, worked with Google Public Sector and Quantiphi to build Chat with CTA, a multilingual virtual assistant on its website. Riders ask when their bus is coming and report issues; the chatbot answers in five languages (English, Spanish, Polish, Simplified Chinese and Filipino/Tagalog) and delivers detailed and timely reports to maintenance crews, including flagging urgent situations for fast follow up. CTA staff review over 250 incidents a week spanning buses, trains and train stations. Google Public Sector says the chatbot \"has grown CTA's customer service reach by over 63%\" and that \"since launch, there's been a 16% improvement in conversation completion\".","stage":"production","year":2025,"channels":["web-chat"],"languages":["en","es","pl","zh","tl"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://publicsector.google/ai/chicago-transit-authority-launches-a-multi-lingual-chatbot-for-more-a-more-seamless-commute/","title":"Chicago Transit Authority Connects with City: AI Chatbot Bridges Language Barriers and Empowers Riders","publisher":"Google Public Sector","archivedUrl":"https://web.archive.org/web/20250317123845/https://publicsector.google/ai/chicago-transit-authority-launches-a-multi-lingual-chatbot-for-more-a-more-seamless-commute/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"chicago-transit-authority-chat-with-cta"},{"title":"Chipotle: conversational AI hiring assistant Ava Cado for restaurant recruiting","useCases":["recruitment-screening-and-interview-scheduling"],"organization":{"name":"Chipotle Mexican Grill","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[{"name":"Paradox","role":"platform"}],"summary":"In October 2024 Chipotle began a phased rollout of Paradox's conversational hiring system to its more than 3,500 restaurants in North America and Europe, with completion planned that month. A virtual assistant named Ava Cado chats with candidates in English, Spanish, French and German, answers their questions, collects basic information, schedules interviews for hiring managers and sends offers to the candidates managers select. The managers keep the hiring decision. Chipotle said it expects time to hire to fall by as much as 75%; that is a target, not a reported result.","stage":"production","year":2024,"channels":[],"languages":["en","es","fr","de"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.chipotle.com/2024-10-22-CHIPOTLE-INTRODUCES-NEW-AI-HIRING-PLATFORM-TO-SUPPORT-ITS-ACCELERATED-GROWTH","title":"Chipotle introduces new AI hiring platform to support its accelerated growth","publisher":"Chipotle Mexican Grill","date":"2024-10-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"chipotle-ava-cado-conversational-hiring"},{"title":"Chronopost: Léonard autonomous AI agent for parcel tracking and delivery requests","useCases":["parcel-tracking-and-delivery-exception-agent"],"organization":{"name":"Chronopost","anonymized":false,"country":"FR","region":"europe","industry":"logistics-and-transportation"},"vendors":[{"name":"Chronopost Robot Factory","role":"in-house"}],"summary":"Chronopost, a French parcel carrier and business unit of the Geopost group, runs Léonard, an autonomous generative AI agent developed with its internal Robot Factory. It handles routine enquiries such as parcel tracking and delivery times, which together make up 40 per cent of incoming contacts, takes actions directly in operational systems and pre qualifies complex cases for customer advisors. It won a Gold CX Award for innovation in June 2026, when Geopost said the rollout would extend across all services from July 2026.","stage":"production","year":2026,"channels":[],"languages":[],"metrics":[{"kpi":"containment-rate","value":85,"unit":"percent","qualifier":"exact","period":"simple requests","claimant":"organization","quote":"Chronopost achieves 40 per cent faster response times and resolves 85 per cent of simple requests end to end with its autonomous agent, Léonard.","sourceUrl":"https://www.geopost.com/en/news/chronopost-wins-gold-cx-award-for-ai-driven-service/"},{"kpi":"response-time-reduction","value":40,"unit":"percent","qualifier":"exact","claimant":"organization","quote":"Chronopost achieves 40 per cent faster response times and resolves 85 per cent of simple requests end to end with its autonomous agent, Léonard.","sourceUrl":"https://www.geopost.com/en/news/chronopost-wins-gold-cx-award-for-ai-driven-service/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.geopost.com/en/news/chronopost-wins-gold-cx-award-for-ai-driven-service/","title":"CX Award Innovation: Chronopost Wins Gold for AI Agent","publisher":"Geopost","date":"2026-06-11"},{"url":"https://www.geopost.com/en/news/geopost-puts-ai-and-smart-infrastructure-at-the-heart-of-delivery-innovation-at-vivatech-2026/","title":"Geopost puts AI and smart infrastructure at the heart of delivery innovation at VivaTech 2026","publisher":"Geopost","date":"2026-06-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"chronopost-leonard-ai-customer-service-agent"},{"title":"CIMB Niaga: AI agents that help staff give proactive, life stage guidance","useCases":["next-best-action-for-advisors","goal-based-financial-planning-assistant"],"organization":{"name":"CIMB Niaga","anonymized":false,"country":"ID","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Artefact","role":"integrator"}],"summary":"CIMB Niaga, one of Indonesia's largest banks, built purpose built AI agents with its AI Center of Excellence and Artefact on Google Cloud. The agents help bank staff offer tailored advice and proactive guidance matched to a customer's financial goals and life stage. No outcome figures were published.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"cimb-niaga-proactive-guidance-agents"},{"title":"CircleCI: AI assistant in docs, product and support for pipeline developers","useCases":["developer-api-integration-assistant"],"organization":{"name":"CircleCI","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Kapa.ai","role":"platform"}],"summary":"CircleCI fed an AI assistant its documentation, API references, forum threads, internal support knowledge and release notes, and put it in the docs, inside the CircleCI app and in front of its support team. It also exposes its configuration schema and API reference to coding agents through a hosted MCP endpoint, so generated pipeline configuration follows CircleCI's actual syntax. CircleCI reports 28% faster support response times, and Kapa.ai reports a 10% increase in coverage of languages other than English, without naming them.","stage":"production","year":2026,"channels":["web-chat","agent-desktop","api"],"languages":["en"],"metrics":[{"kpi":"response-time-reduction","value":28,"unit":"percent","qualifier":"exact","period":"support response times","claimant":"organization","quote":"The 28% faster response times increase the value of our support packages.","sourceUrl":"https://www.kapa.ai/customer-examples/circleci"},{"kpi":"interactions-handled","value":32000,"unit":"count","qualifier":"at-least","period":"technical questions answered, period not stated","claimant":"vendor","quote":"32,000+ technical questions answered","sourceUrl":"https://www.kapa.ai/customer-examples/circleci"},{"kpi":"hours-saved","value":500,"unit":"hours","qualifier":"at-least","period":"per month","claimant":"vendor","quote":"500+ support hours saved every month","sourceUrl":"https://www.kapa.ai/customer-examples/circleci"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.kapa.ai/customer-examples/circleci","title":"How CircleCI improved support response times by 28% across docs, product, and support","publisher":"Kapa.ai"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"circleci-docs-ai-assistant"},{"title":"Citi: AI enhancements to the CitiDirect Commercial Banking onboarding and KYC process","useCases":["corporate-account-onboarding-orchestration"],"organization":{"name":"Citi","anonymized":false,"country":"US","region":"global","industry":"banking"},"vendors":[],"summary":"Citi added AI driven automation to CitiDirect Commercial Banking, its digital platform for mid sized corporate clients, covering data extraction, form filling and client query routing. The KYC process for renewals is now, in Citi's words, \"fully integrated\" with automated notifications and prefilled information, and a \"Digital Servicing Hub\" centralizes client queries, updates and document submission. Citi says the fully digitized onboarding process, with real time status updates, has significantly reduced onboarding turnaround times, though it gives no figure for that reduction. The platform now covers more than 57 percent of Citi's commercial banking client base and is live in the United States, Hong Kong, India, Singapore, the United Kingdom, Canada, Australia and Brazil.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.citigroup.com/global/news/press-release/2025/citi-global-roll-out-enhancements-citidirect-commercial-banking-platform","title":"Citi Continues Global Roll Out and Enhancements to CitiDirect Commercial Banking Platform","publisher":"Citigroup","archivedUrl":"http://web.archive.org/web/20250817111747/https://www.citigroup.com/global/news/press-release/2025/citi-global-roll-out-enhancements-citidirect-commercial-banking-platform"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"citi-citidirect-commercial-onboarding"},{"title":"Citi Wealth: AskWealth assistant and Advisor Insights","useCases":["wealth-advisor-knowledge-assistant","investment-research-summarization","next-best-action-for-advisors"],"organization":{"name":"Citi","anonymized":false,"country":"US","region":"global","industry":"wealth-and-asset-management"},"vendors":[{"name":"Citi","role":"in-house"}],"summary":"Citi Wealth launched two AI tools built by its Data, Analytics and Innovation team. AskWealth is a generative AI assistant that gives service teams, advisors and managers answers across the wealth business, so that advisors can reach market insights and research when clients ask questions; after a launch in Asia it became available to Citi Wealth colleagues worldwide. Advisor Insights is a dashboard of timely messages about market moves, portfolios and events, including Chief Investment Office insights, piloted with Citigold and Citi Private Client advisors in North America with a wider rollout planned for Q4 2025 and Q1 2026. Citi says the tools will save hours of time but published no figures.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.citigroup.com/global/news/press-release/2025/citi-wealth-launches-advisor-insights-askwealth","title":"Citi Wealth Launches \"Advisor Insights\" Pilot and \"AskWealth,\" AI-Driven \"Gamechangers\" for Client Communications","publisher":"Citigroup","date":"2025-08-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"citi-wealth-askwealth-and-advisor-insights"},{"title":"Citigroup: generative AI coding tools for 30,000 developers","useCases":["developer-coding-assistant"],"organization":{"name":"Citigroup","anonymized":false,"country":"US","region":"global","industry":"banking"},"vendors":[],"summary":"On Citi's fourth quarter 2024 earnings call, CEO Jane Fraser said the bank had armed 30,000 developers with AI tools to write code and launched two AI platforms for 143,000 colleagues. CIO Dive reported the same rollout as generative AI coding tools. No productivity or quality figures for the coding tools are disclosed, and the vendors are not named.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.citigroup.com/rcs/citigpa/storage/public/Earnings/Q42024/4Q24-Earnings-Transcript.pdf","title":"Citi Fourth Quarter 2024 Earnings Call (transcript)","publisher":"Citigroup","date":"2025-01-15"},{"url":"https://www.ciodive.com/news/bank-technology-generative-ai-coding-deepseek-accenture/738300/","title":"Banks fire up coding assistants as AI costs plummet","publisher":"CIO Dive","date":"2025-01-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"citigroup-developer-coding-tools"},{"title":"City of Kelowna: AI answers for the 311 line and permit questions","useCases":["non-emergency-service-request-routing"],"organization":{"name":"City of Kelowna","anonymized":false,"country":"CA","region":"north-america","industry":"government"},"vendors":[{"name":"Zammo.ai","role":"platform"},{"name":"Microsoft (Azure OpenAI Service)","role":"model-provider"}],"summary":"Kelowna, a city of nearly 150,000 in British Columbia, uses Zammo.ai on Azure to answer calls and chats to its 311 non emergency line about property taxes, landfill rules, utilities and snowplowing, the last using live GPS data from the plows. With a provincial housing grant it also built a quick reference tool that returns the zoning bylaws and documents for a given address, and plans to extend it so the assistant walks applicants through the permit process up to the point where a city planner takes over. The city is creating an online registry of the data sources the system uses.","stage":"production","year":2023,"channels":["voice","web-chat"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":80,"unit":"percent","qualifier":"exact","period":"snowplow schedule calls","claimant":"vendor","quote":"In 80 percent of cases, AI has already proven that it can deliver correct responses to people who call in about the snow.","sourceUrl":"https://www.microsoft.com/en/customers/story/1641166525350342888-city-of-kelowna-government-azure-open-ai-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1641166525350342888-city-of-kelowna-government-azure-open-ai-service","title":"The City of Kelowna increases and speeds access to its services with Azure AI","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"city-of-kelowna-311-ai-assistant"},{"title":"Clearstream: AI Settlement Prediction Tool for settlement fails and penalties","useCases":["settlement-fail-prediction-and-exception-management"],"organization":{"name":"Clearstream","anonymized":false,"country":"LU","region":"europe","industry":"capital-markets"},"vendors":[{"name":"Clearstream","role":"in-house"}],"summary":"Clearstream, the Luxembourg based international central securities depository of Deutsche Börse Group, launched an AI Settlement Prediction Tool for its clients in July 2022, together with a Settlement Dashboard. The tool estimates the likelihood that a specific instruction settles on time. The enhanced version released in July 2025 identifies potential failure drivers and at risk instructions up to four business days in advance and estimates potential penalty costs, to support clients' T+1 readiness. Clearstream's current product page adds that the tool names the three primary factors most likely to cause a fail and estimates daily penalty costs. Clients access it in the Xact Web Portal. Clearstream has not published measured outcomes.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.clearstream.com/clearstream-en/newsroom/220711-3153554","title":"Clearstream launches data solutions to predict settlement failures and to foster settlement efficiency","publisher":"Clearstream","date":"2022-07-11"},{"url":"https://www.clearstream.com/clearstream-en/newsroom/250728-4582598","title":"Clearstream Enhances its Settlement Prediction Tool to Manage Settlement Risk and Support T+1 Client Readiness","publisher":"Clearstream","date":"2025-07-28"},{"url":"https://www.clearstream.com/clearstream-en/res-library/connectivity/optimize-your-settlement-efficiency-with-our-predictive-data-services-3446132","title":"Optimize your Settlement Efficiency with our Predictive Data Services","publisher":"Clearstream"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"clearstream-settlement-prediction-tool"},{"title":"ClearTax: WhatsApp agent that helps gig workers file income tax returns","useCases":["tax-questions-and-filing-assistant"],"organization":{"name":"ClearTax","anonymized":false,"country":"IN","region":"asia-pacific","industry":"technology"},"vendors":[{"name":"Microsoft (Azure OpenAI Service)","role":"model-provider"}],"summary":"ClearTax, an Indian online tax filing platform, built a generative AI agent on WhatsApp, using Azure OpenAI models, for low income and blue collar workers who have tax deducted at source from their income but rarely file a return, and so miss refunds and lack the income proof lenders ask for. The agent explains taxes and filing, walks users through the filing workflow and supports nine Indian vernacular languages. It is a private filing service, not a tax authority channel.","stage":"production","year":2024,"channels":["whatsapp"],"languages":[],"metrics":[{"kpi":"users-served","value":200000,"unit":"count","qualifier":"at-least","period":"workers who filed their returns independently","claimant":"vendor","quote":"200,000+ blue-collar workers successfully filed ITRs independently","sourceUrl":"https://www.microsoft.com/en-in/aifirstmovers/cleartax"},{"kpi":"customer-savings","value":300000000,"unit":"currency","currency":"INR","qualifier":"approximately","period":"tax refunds claimed by users, cumulative","claimant":"vendor","quote":"300 million INR unlocked in tax refunds, improving access to credit opportunities","sourceUrl":"https://www.microsoft.com/en-in/aifirstmovers/cleartax"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en-in/aifirstmovers/cleartax","title":"Cleartax: Empowering gig-workers of India","publisher":"Microsoft India (AI First Movers)","archivedUrl":"https://web.archive.org/web/20250317093328/https://www.microsoft.com/en-in/aifirstmovers/cleartax"},{"url":"https://www.deccanherald.com/business/cleartax-microsoft-team-up-to-simplify-tax-filing-for-gig-workers-3167187","title":"ClearTax, Microsoft team up to simplify tax filing for gig workers","publisher":"Deccan Herald (PTI)","archivedUrl":"https://web.archive.org/web/20240828125952/https://www.deccanherald.com/business/cleartax-microsoft-team-up-to-simplify-tax-filing-for-gig-workers-3167187"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"cleartax-whatsapp-tax-filing-agent"},{"title":"Cleveland Clinic: language model prescreening for a polycythemia vera trial","useCases":["clinical-trial-patient-matching"],"organization":{"name":"Cleveland Clinic","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Dyania Health","role":"platform"}],"summary":"Cleveland Clinic researchers used Dyania Health's Synapsis platform, a medically trained language model system embedded in the electronic medical record, to prescreen patients for a phase 3 polycythemia vera trial. From 4.7 million active records it identified 28,200 patients with an oncology diagnosis in the past three years, narrowed them to 904 patients with polycythemia vera, assessed each against the trial's seven eligibility and 20 exclusion criteria within one week and found 22 eligible patients, all confirmed by research staff (100% positive predictive value). The usual workflow had prescreened nine patients and enrolled four over twelve months. Looking ahead, Cleveland Clinic and Dyania Health have announced a collaboration to integrate the platform across the health system's clinical research enterprise; Cleveland Clinic has also invested in Dyania Health.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":904,"unit":"count","qualifier":"exact","period":"patients assessed against the trial criteria in one week","claimant":"organization","quote":"“The AI tool completed full eligibility assessments on these 904 patients within one week, against the trial’s criteria and identified 22 eligible patients,” Dr. Gerds and his colleagues reported.","sourceUrl":"https://consultqd.clevelandclinic.org/ai-screening-platform-accelerates-trial-recruitment-in-polycythemia-vera"},{"kpi":"accuracy","value":100,"unit":"percent","qualifier":"exact","period":"of the 22 patients identified as eligible, confirmed by research staff (positive predictive value)","claimant":"organization","quote":"In a study presented at the 2025 American Society of Hematology (ASH) Annual Meeting, investigators reported that the Dyania Health’s Synapsis™ AI platform, an artificial intelligence tool, identified seven times more eligible patients for a polycythemia vera trial than standard workflows, while achieving 100% positive predictive value following research-staff verification.","sourceUrl":"https://consultqd.clevelandclinic.org/ai-screening-platform-accelerates-trial-recruitment-in-polycythemia-vera"}],"outcomeDisclosed":true,"sources":[{"url":"https://consultqd.clevelandclinic.org/ai-screening-platform-accelerates-trial-recruitment-in-polycythemia-vera","title":"AI Screening Platform Accelerates Trial Recruitment in Polycythemia Vera","publisher":"Cleveland Clinic Consult QD"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"cleveland-clinic-ai-trial-prescreening"},{"title":"Close Brothers Motor Finance: AI document fraud detection on finance applications with Resistant AI","useCases":["application-and-identity-fraud-detection"],"organization":{"name":"Close Brothers Motor Finance","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[{"name":"Resistant AI","role":"platform"}],"summary":"Close Brothers Motor Finance added Resistant AI's document forensics to its motor finance application process in August 2024, after a trial that surfaced 18 further fraud cases. Underwriters get a document fraud check, for example on company bank statements, in 12 seconds instead of a 15 minute manual assessment. The vendor reports fraud losses prevented, a high return on investment and faster application reviews.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"fraud-losses-prevented","value":800000,"unit":"currency","currency":"GBP","qualifier":"exact","period":"first 8 months after implementation","claimant":"vendor","quote":"£800,000 in fraud losses prevented in 8 months.","sourceUrl":"https://resistant.ai/case-studies/close-brothers"}],"outcomeDisclosed":true,"sources":[{"url":"https://resistant.ai/case-studies/close-brothers","title":"Close Brothers","publisher":"Resistant AI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"close-brothers-resistant-ai-document-fraud"},{"title":"CME Group: Gemini Code Assist for developers","useCases":["developer-coding-assistant"],"organization":{"name":"CME Group","anonymized":false,"country":"US","region":"north-america","industry":"capital-markets"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"CME Group, which operates the Chicago Mercantile Exchange, gave its developers Gemini Code Assist. Google Cloud reports that most developers using it say they gain at least 10.5 hours a month.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"hours-saved","value":10.5,"unit":"hours","qualifier":"at-least","period":"per developer per month, self reported","claimant":"vendor","quote":"CME Group, which operates the Chicago Mercantile Exchange, says most developers using Gemini Code Assist report a productivity gain of at least 10.5 hours a month.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2025-04-15","archivedUrl":"https://web.archive.org/web/20250415211336/https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"cme-group-gemini-code-assist"},{"title":"CMS: AI complaint analysis to find root causes, validated by experts (pilot)","useCases":["complaints-root-cause-analysis"],"organization":{"name":"Centers for Medicare and Medicaid Services","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"A team at the US Centers for Medicare and Medicaid Services is piloting AI that analyses a high volume of complaint cases to identify root causes and trends, maps them to the applicable regulatory citations, draws a sample for subject matter experts to validate and recommends next steps. The stated aim is to reduce repeat issues that delay benefits or access to care and to improve health plan compliance. The 2024 inventory describes an earlier proof of concept that used a large language model on complaint data used by the same office (OPOLE). No results are published.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (HHS/CMS/OPOLE entry Complaint Analysis)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI inventory (HHS entry Complaint Analysis POC)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"cms-complaint-root-cause-analysis"},{"title":"Centers for Medicare & Medicaid Services: anomaly detection on prescription drug cost data","useCases":["benefit-fraud-and-error-detection"],"organization":{"name":"Centers for Medicare & Medicaid Services","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The CMS Division of Payment Reconciliation monitors Prescription Drug Event (PDE) data from Medicare Part D for accuracy. It reported an AI project to identify outliers in that data so that errors can be corrected through outreach to the plans and improper payments, an overpayment or an underpayment of a government benefit, are prevented. The project was at the initiated stage in the 2024 federal inventory; no results are published.","stage":"announced","year":2021,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"cms-prescription-drug-cost-anomaly-detection"},{"title":"Centers for Medicare & Medicaid Services: WISeR model for technology assisted prior authorization","useCases":["health-prior-authorization-and-claims-adjudication"],"organization":{"name":"Centers for Medicare & Medicaid Services","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Cohere Health","role":"platform"},{"name":"Genzeon","role":"platform"},{"name":"Humata Health","role":"platform"},{"name":"Innovaccer","role":"platform"},{"name":"Virtix Health","role":"platform"},{"name":"Zyter","role":"platform"}],"summary":"In the WISeR model, the US Centers for Medicare & Medicaid Services works with technology companies that use AI and machine learning, together with clinical review, to review prior authorization requests and claims before payment for a selected set of Original Medicare services with a history of waste, fraud and abuse. It runs from January 1, 2026 to December 31, 2031 in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington. Every recommendation for non payment is decided by an appropriately licensed clinician. Six technology companies take part, one per state, and are paid a share of the averted spending, adjusted for performance measures that include provider experience. The model page reports no results yet.","stage":"production","year":2026,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.cms.gov/priorities/innovation/innovation-models/wiser","title":"WISeR (Wasteful and Inappropriate Service Reduction) Model","publisher":"Centers for Medicare & Medicaid Services"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"cms-wiser-prior-authorization-model"},{"title":"CNA: AI solutions for underwriting triage and submission responsiveness","useCases":["commercial-underwriting-submission-triage"],"organization":{"name":"CNA Financial","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"CNA Financial","role":"in-house"}],"summary":"In its fourth quarter 2025 earnings remarks, CNA said it had deployed a number of AI solutions across underwriting, claims and the back office over the past year and rolled out generative AI tools to every employee. Management reported faster triage, better submission responsiveness and measurable time savings, and framed further investment around risk selection, service quality and efficiency. No figures were disclosed.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/21175/000002117526000008/q42025ex994earningsremarks.htm","title":"CNA fourth quarter 2025 earnings remarks (Form 8-K, Exhibit 99.4)","publisher":"CNA Financial via SEC EDGAR","date":"2026-02-09"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"cna-ai-underwriting-triage"},{"title":"CNG Holdings: identity first application fraud prevention with SAS","useCases":["application-and-identity-fraud-detection"],"organization":{"name":"CNG Holdings","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"SAS","role":"platform"},{"name":"Microsoft Azure","role":"platform"}],"summary":"CNG Holdings, a US consumer lender with online lending and around 1,000 retail stores, runs SAS fraud decisioning with machine learning on Microsoft Azure to verify applicants' identities in real time and stop third party and synthetic identity fraud at application. SAS reports a steep drop in third party fraud within 90 days, a very low fraud false positive rate and lower fraud programme costs after CNG retired several older tools.","stage":"scaled","year":2023,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"cost-reduction","value":30,"unit":"percent","qualifier":"at-least","period":"fraud programme costs","claimant":"vendor","quote":"By replacing a disjointed patchwork of expensive and ineffective fraud tools with SAS’ integrated defenses, Cooney estimates that CNG cut its fraud program costs more than 30%.","sourceUrl":"https://www.sas.com/en_us/news/press-releases/2023/june/cng-holdings-zaps-fraud.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sas.com/en_us/news/press-releases/2023/june/cng-holdings-zaps-fraud.html","title":"CNG Holdings zaps third-party and synthetic fraud with 'identity-first' fraud prevention","publisher":"SAS","date":"2023-06-13"},{"url":"https://www.sas.com/en_us/customers/cng-holdings.html","title":"Identity management is key to preventing credit fraud","publisher":"SAS"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"cng-holdings-sas-identity-fraud"},{"title":"Coast: AI assisted fraud case management with Oscilar","useCases":["fraud-alert-triage"],"organization":{"name":"Coast","anonymized":false,"country":"US","region":"north-america","industry":"payments"},"vendors":[{"name":"Oscilar","role":"platform"}],"summary":"Coast, a US fleet and fuel card provider, uses Oscilar's risk platform for fraud decisioning and case management, including rule based routing and assignment of cases, a feedback loop, and generative AI features for case assignment and fraud analysis. The vendor reports that the time Coast's staff spend on manual reviews fell from 2 hours per person per day to under 30 minutes. This is time per person per day, not time per case: queues and auto assignment also let entry level case managers work independently from analysts, so part of the drop may be work moved between roles.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://oscilar.com/customers/coast","title":"How Coast Cut Manual Review Time by 75%","publisher":"Oscilar","archivedUrl":"https://web.archive.org/web/20241015125523/https://oscilar.com/customers/coast"},{"url":"https://www.coastpay.com","title":"Coast fleet and fuel cards","publisher":"Coast"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"coast-oscilar-fraud-case-review"},{"title":"Comcast: AI that groups outage alarms and finds the cause so customers can be notified","useCases":["network-outage-communication-agent"],"organization":{"name":"Comcast","anonymized":false,"country":"US","region":"north-america","industry":"telecommunications"},"vendors":[{"name":"Comcast","role":"in-house"}],"summary":"Comcast deployed a proprietary AI tool nationwide across its footprint after trials in the 2024 hurricane season. When many modems go offline it groups the individual alarms by location and time into one alert, then analyses device and network data to determine the cause, for example a commercial power outage. Comcast says this diagnosis lets it notify affected customers with recommendations such as contacting their local power provider, and that the tool streamlines the dispatch of technicians with the tools they need. Comcast does not say that the AI sends the notices itself or that customers hear before they call. Comcast says trials in the 2024 hurricane season raised storm recovery effectiveness by fifty percent in impacted regions.","stage":"scaled","year":2025,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://corporate.comcast.com/press/releases/comcast-leverages-ai-to-restore-service-faster-when-commercial-power-outages-impact-customers","title":"Comcast Leverages AI to Restore Service Faster When Commercial Power Outages Impact Customers","publisher":"Comcast","date":"2025-09-26","archivedUrl":"https://web.archive.org/web/20260517021049/https://corporate.comcast.com/press/releases/comcast-leverages-ai-to-restore-service-faster-when-commercial-power-outages-impact-customers"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"comcast-ai-outage-detection-and-notification"},{"title":"Comeen: multilingual subtitles for internal communication videos","useCases":["audio-and-video-transcription-and-captioning"],"organization":{"name":"Comeen","anonymized":false,"country":"FR","region":"europe","industry":"technology"},"vendors":[{"name":"Google Cloud (Vertex AI, Gemini, Speech to Text)","role":"platform"}],"summary":"Comeen, a workplace software company, launched automatic multilingual subtitles for videos shown on its clients' digital signage at the end of 2024. Employees upload a video in their usual tools and receive subtitles in 40 languages, produced by a multi step AI workflow that the company says makes them usable as is. It replaces a process that used several providers over several days.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/customers/comeen","title":"Comeen case study | Google Cloud","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"comeen-multilingual-video-subtitles"},{"title":"Commerzbank: AI agent documents client advisory calls","useCases":["client-meeting-notes-and-crm-update"],"organization":{"name":"Commerzbank","anonymized":false,"country":"DE","region":"europe","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Commerzbank implemented an AI agent on Gemini 1.5 Pro that automates the documentation of client calls, a manual task for its financial advisors. Google Cloud reports a significant reduction in processing time, which lets advisors spend more time with clients, but gives no figure.","stage":"production","year":2024,"channels":["voice","internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"commerzbank-client-call-documentation"},{"title":"Commonwealth Bank: Customer Engagement Engine for next best conversations","useCases":["offers-and-rewards-agent","financial-wellbeing-coach","loan-restructuring-recommendations","proactive-outbound-engagement-agent","personalized-marketing-at-scale"],"organization":{"name":"Commonwealth Bank of Australia","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Pegasystems","role":"platform"}],"summary":"Commonwealth Bank's Customer Engagement Engine (CEE), built on Pega Customer Decision Hub, suggests in real time the next best conversation to have with each customer, whether in the branch, on the phone, online or on a mobile device. Beyond suggesting conversations, the bank uses it to match customers to government benefits and rebates they may be missing (Benefits finder) and to reach customers hit by natural disasters with same day support such as a loan deferral. The same decisions feed digital channels and prompts for branch and contact centre staff.","stage":"scaled","year":2022,"channels":["mobile-app","agent-desktop"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.commbank.com.au/articles/newsroom/2022/06/CBA-artificial-intelligence-usages.html","title":"How artificial intelligence is changing the face of banking","publisher":"Commonwealth Bank of Australia","date":"2022-06-24"},{"url":"https://www.pega.com/customers/cba-marketing","title":"Delivering next best conversations with Pega","publisher":"Pegasystems"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"commonwealth-bank-customer-engagement-engine"},{"title":"Commonwealth Bank: AI orchestration for messaging service with human handoff","useCases":["account-and-card-servicing-agent","card-dispute-and-chargeback-intake","first-line-contact-centre-agent"],"organization":{"name":"Commonwealth Bank of Australia","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Commonwealth Bank built a central AI orchestration agent that reads the customer's intent and routes it to a conversational AI, retrieval over public content, a deterministic guarded path for regulated journeys such as fraud disputes, or a human specialist with the full context, on its messaging channel. It migrated nearly 700 chatbot topics and launched a generative AI banking chatbot in November 2024. Voice bots are a planned extension of the orchestration layer.","stage":"scaled","year":2024,"channels":[],"languages":["en"],"metrics":[{"kpi":"containment-rate","value":84.6,"unit":"percent","qualifier":"approximately","period":"May 2026, self service messaging","claimant":"vendor","quote":"In May 2026, approximately 84.6% of self-service messaging interactions were resolved end-to-end in the messaging channel.","sourceUrl":"https://news.microsoft.com/source/asia/features/how-commonwealth-bank-and-microsoft-are-reimagining-the-future-of-customer-service/"}],"outcomeDisclosed":true,"sources":[{"url":"https://news.microsoft.com/source/asia/features/how-commonwealth-bank-and-microsoft-are-reimagining-the-future-of-customer-service/","title":"How Commonwealth Bank and Microsoft are reimagining the future of customer service","publisher":"Microsoft Source Asia","date":"2026-07-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"commonwealth-bank-customer-service-orchestration"},{"title":"Commonwealth Bank: proactive scam warnings, in app transaction verification and a fraud detection agent","useCases":["scam-payment-interception","fraud-alert-confirmation","real-time-fraud-scoring"],"organization":{"name":"Commonwealth Bank of Australia","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Snowflake","role":"platform"}],"summary":"Commonwealth Bank combines several AI controls against scams and fraud. Its fraud systems monitor more than 80 million signals a day and the CommBank app sends proactive warning alerts on payments that look risky; NameCheck and Confirmation of Payee check payee details on first time payments. From August 2025 customers are asked to verify certain online card transactions in the app, in real time, before they are authorised. In April 2026 the bank described an agentic system that spots emerging fraud patterns and proposes new detection rules, which the fraud analytics team reviews and approves before they go live. The bank reports a 76% fall in customer scam losses since their peak without attributing it to any single tool, and says its fraud detection technology played a role in cutting fraud losses by over 20% in the first half of FY26.","stage":"scaled","year":2025,"channels":["mobile-app","api"],"languages":["en"],"metrics":[{"kpi":"fraud-loss-reduction","value":76,"unit":"percent","qualifier":"exact","period":"second half of FY25 versus first half of FY23 (the peak)","baseline":"customer scam losses at their peak in the first half of FY23","claimant":"organization","quote":"CommBank has seen a 76% drop in customer scam losses since peak (2H25 vs. 1H23)","sourceUrl":"https://www.commbank.com.au/articles/newsroom/2025/08/commbank-customer-scam-losses-fall-truyu.html"},{"kpi":"interactions-handled","value":40000,"unit":"count","qualifier":"at-least","period":"per day on average, proactive warning alerts in the CommBank app","claimant":"organization","quote":"Each day, CommBank processes more than 20 million payments on average and sends more than 40,000 proactive warning alerts on average to customers via the CommBank app.","sourceUrl":"https://www.commbank.com.au/articles/newsroom/2026/04/ai-agent-spots-fraud-in-real-time.html"},{"kpi":"fraud-loss-reduction","value":20,"unit":"percent","qualifier":"at-least","period":"first half of FY26 versus first half of FY25","claimant":"organization","quote":"The bank’s fraud detection technology has played a role in helping to reduce fraud losses by over 20% in the first half of the 2026 financial year compared to the first half of the 2025 financial year.","sourceUrl":"https://www.commbank.com.au/articles/newsroom/2026/04/ai-agent-spots-fraud-in-real-time.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.commbank.com.au/articles/newsroom/2025/08/commbank-customer-scam-losses-fall-truyu.html","title":"CBA sees customer scam losses fall by 76% and adds two new forms of armour to help keep customers safe","publisher":"Commonwealth Bank of Australia","date":"2025-08-11"},{"url":"https://www.commbank.com.au/articles/newsroom/2026/04/ai-agent-spots-fraud-in-real-time.html","title":"CommBank develops AI agent that spots new fraud and helps build defences","publisher":"Commonwealth Bank of Australia","date":"2026-04-24"},{"url":"https://www.commbank.com.au/articles/newsroom/2025/07/scam-protection-confirmation-of-payee.html","title":"Strengthening scam protection: Introducing Confirmation of Payee","publisher":"Commonwealth Bank of Australia","date":"2025-07-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"commonwealth-bank-scam-and-fraud-interventions"},{"title":"Comrade Trustee Services: AI reconciliation for a defence force pension fund","useCases":["ledger-and-payment-reconciliation"],"organization":{"name":"Comrade Trustee Services","anonymized":false,"country":"PG","region":"asia-pacific","industry":"wealth-and-asset-management"},"vendors":[{"name":"Smartstream","role":"platform"}],"summary":"Comrade Trustee Services, trustee of the Defence Force Retirement Benefit Fund in Papua New Guinea, went live with Smartstream's Air AI reconciliation platform, which it uses to reconcile multiple file types, including fixed length files and PDFs that need advanced matching logic. The article says Air replaced manual data collection and spreadsheet preparation, and the vendor says processing time fell from up to eight hours to under five minutes.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.assetservicingtimes.com/assetservicesnews/technologyarticle.php?article_id=18038","title":"Comrade Trustee Services goes live with Smartstream’s Air","publisher":"Asset Servicing Times","date":"2026-06-11"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"comrade-trustee-services-ai-reconciliation"},{"title":"Con Edison: enterprise data analytics and AMI operations with C3 AI","useCases":["smart-meter-analytics"],"organization":{"name":"Consolidated Edison (Con Edison)","anonymized":false,"country":"US","region":"north-america","industry":"energy-and-utilities"},"vendors":[{"name":"C3 AI","role":"platform"}],"summary":"Con Edison built an enterprise data analytics platform on C3 AI to run Advanced Metering Infrastructure operations for its 5.3 million meter smart meter rollout, aggregating two years of data from 13 source systems covering 5 million customer accounts and integrating over 180 billion rows of data a year across those systems. Two machine learning algorithms and 50 analytics identify deployment and installation issues and determine meter and network health, giving the utility a real time, prioritised view from an individual meter up to the whole system, in a 10 month project.","stage":"production","year":2019,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://c3.ai/customers/conedison/","title":"ConEdison","publisher":"C3 AI","archivedUrl":"https://web.archive.org/web/20190819211733/https://c3.ai/customers/conedison/"},{"url":"https://www.sec.gov/Archives/edgar/data/1577526/000162828024028786/ai-20240430.htm","title":"C3.ai, Inc. Form 10-K, fiscal year ended April 30, 2024","publisher":"C3.ai, Inc. (U.S. Securities and Exchange Commission, EDGAR)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"con-edison-c3-ai-smart-meter-analytics"},{"title":"Contentsquare: faster data incident detection and resolution","useCases":["data-quality-monitoring-agent"],"organization":{"name":"Contentsquare","anonymized":false,"country":"FR","region":"europe","industry":"technology"},"vendors":[{"name":"Monte Carlo","role":"platform"}],"summary":"Contentsquare, a digital experience analytics company, had too many manual data quality checks run by operations and data analysts, and still lacked visibility into data incidents before they reached stakeholders. It deployed Monte Carlo's end to end data observability platform to detect anomalies earlier and build a collaborative incident resolution workflow between the data team and the business. Monte Carlo reports that, within one month, Contentsquare saw faster detection and faster resolution of data incidents.","stage":"production","year":2023,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"mttr-reduction","value":17,"unit":"percent","qualifier":"exact","period":"in one month","baseline":"Time to detect a data incident before Monte Carlo","claimant":"vendor","quote":"Deploying Monte Carlo led to a 17% reduction in data incident detection time and a 16% reduction in time to resolution – in just one month.","sourceUrl":"https://www.montecarlodata.com/blog-how-contentsquare-reduced-time-to-data-incident-detection-by-17-percent-with-monte-carlo/"},{"kpi":"mttr-reduction","value":16,"unit":"percent","qualifier":"exact","period":"in one month","baseline":"Time to resolve a data incident before Monte Carlo","claimant":"vendor","quote":"Deploying Monte Carlo led to a 17% reduction in data incident detection time and a 16% reduction in time to resolution – in just one month.","sourceUrl":"https://www.montecarlodata.com/blog-how-contentsquare-reduced-time-to-data-incident-detection-by-17-percent-with-monte-carlo/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.montecarlodata.com/blog-how-contentsquare-reduced-time-to-data-incident-detection-by-17-percent-with-monte-carlo/","title":"How Contentsquare Reduced Time to Data Incident Detection by 17 Percent with Monte Carlo","publisher":"Monte Carlo"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"contentsquare-data-incident-detection"},{"title":"Continental Automotive: AI based requirements engineering for customer requirement documents","useCases":["requirements-to-test-case-generation"],"organization":{"name":"Continental AG","anonymized":false,"country":"DE","region":"europe","industry":"automotive"},"vendors":[{"name":"Microsoft","role":"platform"},{"name":"NTT DATA","role":"integrator"}],"summary":"Continental's Automotive division receives customer requirement documents that can run to hundreds of pages and up to 30,000 individual requirements, which engineers used to read, categorize as functional or non functional, check for safety and security relevance and check for contradictions and duplicates by hand. With Microsoft and NTT DATA it built a generative AI solution on Azure AI that scans the document, finds relevant sections and keywords, categorizes requirements and compares them with a catalogue of generic Continental features. The proof of concept was built in three months and the company plans to scale it; no measured saving is published.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/22437-continental-ag-azure-ai-services","title":"Continental relies on Microsoft Azure AI: A game changer in R&D requirements management","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"continental-ai-requirements-engineering"},{"title":"Copenhagen Emergency Medical Services: randomized trial of machine learning alerts for cardiac arrest on 112 calls","useCases":["emergency-call-triage-support"],"organization":{"name":"Copenhagen Emergency Medical Services","anonymized":false,"country":"DK","region":"europe","industry":"government"},"vendors":[],"summary":"From September 2018 to December 2019, a machine learning model listened to 112 emergency calls at Copenhagen EMS through speech recognition and flagged suspected out of hospital cardiac arrest. Because of downtime, it processed 169,049 of the 226,130 calls the service received (74.7%). It flagged 5,847 calls as suspected cardiac arrest, and the 5,242 eligible calls were randomized: dispatchers in one group saw an alert, the other group worked as usual. The model alone had higher sensitivity than dispatchers without alerts (85.0% against 77.5%) but a much lower positive predictive value (17.8% against 55.8%), and dispatchers with alerts did not recognize significantly more confirmed arrests (93.1% against 90.5%, P = .15). It shows that a model with higher sensitivity did not, in this trial, lead to better dispatcher recognition.","stage":"pilot","year":2018,"channels":["voice","agent-desktop"],"languages":["da"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2774598","title":"Effect of Machine Learning on Dispatcher Recognition of Out-of-Hospital Cardiac Arrest During Calls to Emergency Medical Services: A Randomized Clinical Trial","publisher":"JAMA Network Open"},{"url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC7788469/","title":"Effect of Machine Learning on Dispatcher Recognition of Out-of-Hospital Cardiac Arrest During Calls to Emergency Medical Services: A Randomized Clinical Trial (open access full text)","publisher":"PubMed Central"},{"url":"https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:33404620%20AND%20SRC:MED&resultType=core&format=xml","title":"Abstract record of the trial (PMID 33404620)","publisher":"Europe PMC"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"copenhagen-ems-cardiac-arrest-recognition-trial"},{"title":"Covéa: AI photo assessment of motor damage for its repairer network","useCases":["photo-based-damage-assessment"],"organization":{"name":"Covéa","anonymized":false,"country":"FR","region":"europe","industry":"insurance"},"vendors":[{"name":"Tractable","role":"platform"}],"summary":"Covéa, the French mutual group behind the MAAF, MMA and GMF brands, has used Tractable's AI since 2018 to analyse photos of vehicle damage in real time, so that its partner repairers receive instant assessments instead of waiting on administrative steps after an accident. In April 2026 the two renewed the partnership for three years. Covéa's head of the auto networks division says more than 160,000 claims have been finalized since the collaboration began.","stage":"production","year":2026,"channels":[],"languages":["fr"],"metrics":[{"kpi":"interactions-handled","value":160000,"unit":"count","qualifier":"at-least","period":"Claims finalized since the partnership began in 2018","claimant":"organization","quote":"Since the beginning of our collaboration, more than 160,000 claims have been finalized.","sourceUrl":"https://tractable.ai/covea-renewal-tractable/"}],"outcomeDisclosed":true,"sources":[{"url":"https://tractable.ai/covea-renewal-tractable/","title":"Covéa and Tractable renew their partnership to leverage AI in auto claims management in France","publisher":"Tractable","date":"2026-04-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"covea-ai-photo-damage-assessment"},{"title":"Crown Prosecution Service: Beam Notes transcripts and structured summaries of video evidence","useCases":["court-and-case-file-summarization"],"organization":{"name":"Crown Prosecution Service","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Beam","role":"platform"},{"name":"OpenAI","role":"model-provider"},{"name":"ElevenLabs","role":"model-provider"}],"summary":"Prosecutors in England and Wales review recorded video interviews as part of case preparation. Beam Notes lets them upload a recording and receive a time stamped transcript and a structured summary (key details such as names and dates, and a chronology of events) built on templates that the CPS co designed with the supplier. Prosecutors must still watch the original video, every summary is reviewed before anything is moved into case systems, users rate summaries in the app, and the tool gives no advice on charging or case outcomes. The transparency record says it will be used across all 14 CPS areas on about 11,000 cases a year, with data kept by the supplier for up to 30 days.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/the-crown-prosecution-service-beam-notes","title":"The Crown Prosecution Service: Beam Notes (algorithmic transparency record)","publisher":"GOV.UK","date":"2026-09-09"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"crown-prosecution-service-beam-notes-video-evidence-summaries"},{"title":"Crown Prosecution Service: Correspondence Drafting Tool for letters and emails from case data","useCases":["civil-servant-drafting-copilot"],"organization":{"name":"Crown Prosecution Service","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"NTT Data UK Limited","role":"integrator"},{"name":"OpenAI (via Microsoft Azure)","role":"model-provider"}],"summary":"The Crown Prosecution Service sends large volumes of letters and emails to people involved in prosecutions. Its Correspondence Drafting Tool pulls key information from the case management system into standard CPS templates and uses a large language model to summarise and pre populate the mandatory content, which authors then refine for the specific case. A built in workflow routes every letter through review and approval by trained staff before it is sent, and the tool plays no part in prosecution decisions. It replaced manual drafting in Word, where copying information by hand could introduce errors, and was in beta with 30 users.","stage":"pilot","year":2025,"channels":["internal-tools","email"],"languages":["en"],"metrics":[{"kpi":"users-served","value":30,"unit":"count","qualifier":"exact","period":"beta phase","claimant":"organization","quote":"Currently in beta phase, being used by small group of users (30 users) the number of users will increase as the tool is implemented.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/the-crown-prosecution-service-correspondence-drafting-tool"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/the-crown-prosecution-service-correspondence-drafting-tool","title":"The Crown Prosecution Service: Correspondence Drafting Tool (algorithmic transparency record)","publisher":"GOV.UK","date":"2025-08-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"crown-prosecution-service-correspondence-drafting-tool"},{"title":"Cushman & Wakefield: AI lease abstraction with Unframe","useCases":["lease-abstraction"],"organization":{"name":"Cushman & Wakefield","anonymized":false,"country":"US","region":"north-america","industry":"real-estate"},"vendors":[{"name":"Unframe","role":"platform"}],"summary":"Cushman & Wakefield evaluated multiple vendors and internal approaches for lease abstraction before selecting Unframe in a competitive tender. Unframe's case study reports that abstracting a single lease used to take six hours to three days; after deployment, the platform processes leases of any length, across multiple languages, and brokers can access lease insights in real time. What started as one use case has grown into more than 15 active Unframe projects across the business.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.unframe.ai/customer-stories/cushman-wakefield","title":"Cushman & Wakefield AI Lease Abstraction: Faster CRE Deals | Unframe AI","publisher":"Unframe"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"cushman-wakefield-unframe-lease-abstraction"},{"title":"Databricks: AI expense audit finds $483K in wasteful spend","useCases":["travel-and-expense-audit-agent"],"organization":{"name":"Databricks","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"AppZen","role":"platform"}],"summary":"Before AppZen, Databricks' four person global audit team relied on manager approval, which gave \"no visibility, no forensics, and no data analysis on receipts,\" plus a four eyes check where two auditors manually reviewed nearly 130,000 expense reports a year across more than 7,000 employees. AppZen's Expense Audit, layered onto the existing Emburse Chrome River expense system, automated that review for duplicates and policy risk. Custom AppStore models such as Double-Dip Detection catch employees who submit a meal expense while also being listed as an attendee on someone else's expense, something the team could not catch before AppZen.","stage":"production","year":2025,"channels":[],"languages":["en"],"metrics":[{"kpi":"hours-saved","value":3000,"unit":"hours","qualifier":"approximately","period":"per year","claimant":"vendor","quote":"In one year, Databricks saved nearly 3,000 manual auditor work hours.","sourceUrl":"https://www.appzen.com/databricks-improved-travel-expense-management/"},{"kpi":"automation-rate","value":72,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"Databricks identified $483K in wasteful spend and saved 3,000 auditor hours annually with AppZen, achieving 72% auto-approval for 9,000 employees.","sourceUrl":"https://www.appzen.com/databricks-improved-travel-expense-management/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.appzen.com/databricks-improved-travel-expense-management/","title":"Databricks: $483K Savings & 72% Auto-Approval","publisher":"AppZen","archivedUrl":"https://web.archive.org/web/20250118115135/https://www.appzen.com/databricks-improved-travel-expense-management/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"databricks-expense-audit-automation"},{"title":"Datasite: Redaction AI for M&A data rooms","useCases":["deal-sourcing-and-due-diligence-assistant"],"organization":{"name":"Datasite","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Datasite, a virtual data room provider for mergers and acquisitions, added Redaction AI to its Datasite Diligence application. It uses named entity recognition to find personal and sensitive data across the documents a seller prepares for due diligence, so bankers and lawyers on the sell side can redact in batches instead of one document at a time; machine translation helps them file documents in other languages. Microsoft reports that redaction time falls sharply.","stage":"production","year":2021,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"productivity-gain","value":80,"unit":"percent","qualifier":"up-to","period":"redaction step only (share of time and resources saved on redacting data room documents before due diligence), not the end to end deal cycle","baseline":"manual review and redaction document by document","claimant":"vendor","quote":"Now, the lawyers and investment bankers who coordinate sales can reduce redaction times by up to 80 percent, helping to move deals forward faster and support successful outcomes.","sourceUrl":"https://www.microsoft.com/en/customers/story/1379631359425784399-datasite-banking-capital-markets-azure-cognitive-services"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1379631359425784399-datasite-banking-capital-markets-azure-cognitive-services","title":"Datasite automates M&A and speeds redaction by 80%, saving customers valuable time with Azure Cognitive Services","publisher":"Microsoft","date":"2021-11-10"},{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft","date":"2025-07-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"datasite-redaction-ai"},{"title":"Day Knight & Associates: voice and SMS agents for consumer and healthcare debt","useCases":["collections-and-hardship-agent"],"organization":{"name":"Day Knight & Associates","anonymized":false,"country":"US","region":"north-america","industry":"professional-services"},"vendors":[{"name":"Skit.ai","role":"platform"}],"summary":"Day Knight & Associates, a Missouri agency founded in 2001 that collects healthcare and consumer debt, moved from outbound and inbound voice AI to a combined voice and SMS setup. The vendor reports that adding SMS reached consumers that voice alone missed and that, within a month of going multichannel, the cost of collecting a dollar fell from 22 cents to 8 cents; the agency's vice president of business development says collections doubled. It shows how channel choice, not only automation, drives results in collections.","stage":"production","year":2026,"channels":["voice","sms"],"languages":["en"],"metrics":[{"kpi":"cost-reduction","value":63,"unit":"percent","qualifier":"exact","period":"cost of collections, after moving to voice plus SMS","claimant":"vendor","quote":"See how Day Knight & Associates used AI for Debt Collections to double recoveries, reduce collection costs by 63%, and scale multichannel outreach with Voice AI and SMS.","sourceUrl":"https://skit.ai/resource/case-studies/22-cents-to-8-cents-how-skit-ai-cut-day-knights-cost-of-collections-by-63-and-doubled-recovery/"},{"kpi":"recovery-rate-uplift","value":100,"unit":"percent","qualifier":"approximately","period":"collections after moving to voice plus SMS","claimant":"organization","quote":"After adopting Skit.ai’s multichannel platform, we were able to double our collections and connectivity rate.","sourceUrl":"https://skit.ai/resource/case-studies/22-cents-to-8-cents-how-skit-ai-cut-day-knights-cost-of-collections-by-63-and-doubled-recovery/"}],"outcomeDisclosed":true,"sources":[{"url":"https://skit.ai/resource/case-studies/22-cents-to-8-cents-how-skit-ai-cut-day-knights-cost-of-collections-by-63-and-doubled-recovery/","title":"22¢ to 8¢, 2X recoveries: Day Knight & Associates Scales Smarter with Skit AI for Debt Collections","publisher":"Skit.ai","date":"2026-04-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"day-knight-associates-multichannel-collections"},{"title":"DBS: agentic AI that drafts credit memos for corporate bankers","useCases":["credit-memo-drafting-agent"],"organization":{"name":"DBS Bank","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"DBS Bank","role":"in-house"}],"summary":"DBS rolled out an agentic AI solution in which specialised agents handle more than 70 tasks to turn raw data (annual reports, industry research, internal records) into a review ready first draft of a credit memo for large and mid sized corporate clients. Relationship managers and credit risk managers iterate with the agents to reach the final memo. After a pilot with 150 users it reached about 1,500 employees globally in August 2026. DBS states a goal of cutting the time spent by at least 30%; that is a target, not a measured result.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":1500,"unit":"count","qualifier":"approximately","period":"employees globally, August 2026 rollout","claimant":"organization","quote":"After an initial pilot phase involving 150 participants, the capability has been rolled out to approximately 1,500 employees globally.","sourceUrl":"https://www.dbs.com/newsroom/DBS_scales_agentic_AI_to_transform_way_of_working_for_corporate_bankers_freeing_up_time_for_more_strategic_client_engagements"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.dbs.com/newsroom/DBS_scales_agentic_AI_to_transform_way_of_working_for_corporate_bankers_freeing_up_time_for_more_strategic_client_engagements","title":"DBS scales agentic AI to transform way of working for corporate bankers, freeing up time for more strategic client engagements","publisher":"DBS Bank","date":"2026-08-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dbs-agentic-credit-memo"},{"title":"DBS: CSO Assistant, a generative AI copilot for customer service officers","useCases":["live-agent-assist"],"organization":{"name":"DBS Bank","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"DBS Bank","role":"in-house"}],"summary":"DBS built CSO Assistant in house for its 500 customer service officers in Singapore, who handle queries from more than 250,000 customers a month. It combines a language model tuned to local languages and parlance with telephony and speech recognition: it transcribes the call in real time, searches the knowledge base live, then writes the call summary and prefills service request fields. Pilots began in October 2023; full rollout in Singapore was planned before the end of 2024, followed by Taiwan and Hong Kong. The 20% cut in call handling time in the release is an expectation, not a result.","stage":"pilot","year":2024,"channels":["voice","agent-desktop"],"languages":[],"metrics":[{"kpi":"accuracy","value":100,"unit":"percent","qualifier":"approximately","period":"pilot, October 2023 to July 2024","claimant":"organization","quote":"Based on data collected since pilots began in October 2023, CSO Assistant has demonstrated transcription and solutioning accuracy of nearly 100%, and when fully deployed, is expected to reduce call handling time by up to 20%.","sourceUrl":"https://www.dbs.com/newsroom/DBS_empowers_its_Customer_Service_Officers_with_Gen_AI_powered_virtual_assistant_to_reduce_toil_and_enhance_customer_experience"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.dbs.com/newsroom/DBS_empowers_its_Customer_Service_Officers_with_Gen_AI_powered_virtual_assistant_to_reduce_toil_and_enhance_customer_experience","title":"DBS empowers its Customer Service Officers with Gen AI powered virtual assistant to reduce toil and enhance customer experience","publisher":"DBS Bank","date":"2024-07-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"dbs-cso-assistant"},{"title":"DBS: generative and agentic AI in the DBS Joy and DBS digibot virtual assistants","useCases":["account-and-card-servicing-agent","offers-and-rewards-agent"],"organization":{"name":"DBS Bank","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"DBS Bank","role":"in-house"}],"summary":"DBS runs two generative AI virtual assistants on its own AI platforms: DBS digibot for individual customers in Singapore, Hong Kong and Taiwan, and DBS Joy for corporate and SME customers. In July 2026 DBS Joy became agentic in Singapore and now answers questions such as payment status and fees from the customer's own transaction and account data. DBS digibot answers card, refund, fee waiver and remittance questions today; DBS plans to add agentic tasks such as checking card usage, tracking reward points and blocking or replacing cards in the fourth quarter of 2026, for logged in customers only.","stage":"scaled","year":2026,"channels":["mobile-app","web-chat"],"languages":[],"metrics":[{"kpi":"containment-rate","value":90,"unit":"percent","qualifier":"approximately","period":"DBS digibot, first half of 2026, queries resolved without a follow up call","claimant":"organization","quote":"In the first half of 2026, DBS digibot successfully resolved nine in every 10 queries digitally, without customers needing to make a follow-up call.","sourceUrl":"https://www.dbs.com/newsroom/DBS_Gen_AI_enabled_virtual_assistants_reach_10_million_customers_and_go_agentic"},{"kpi":"contact-deflection","value":7,"unit":"percent","qualifier":"exact","period":"DBS Joy in Singapore, first six months of 2026, calls or emails to customer service","claimant":"organization","quote":"Active users increased by 61%, contributing to a 7% reduction in calls or emails to customer service.","sourceUrl":"https://www.dbs.com/newsroom/DBS_Gen_AI_enabled_virtual_assistants_reach_10_million_customers_and_go_agentic"},{"kpi":"customer-satisfaction-uplift","value":17,"unit":"percent","qualifier":"exact","period":"DBS Joy in Singapore, first six months of 2026","claimant":"organization","quote":"Customer satisfaction scores for DBS Joy rose by 17% over the same period.","sourceUrl":"https://www.dbs.com/newsroom/DBS_Gen_AI_enabled_virtual_assistants_reach_10_million_customers_and_go_agentic"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.dbs.com/newsroom/DBS_Gen_AI_enabled_virtual_assistants_reach_10_million_customers_and_go_agentic","title":"DBS' Gen AI-enabled virtual assistants reach 10 million customers and go agentic","publisher":"DBS Bank","date":"2026-07-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dbs-joy-and-digibot-virtual-assistants"},{"title":"DBS: generative AI DBS Joy assistant for corporate and SME clients","useCases":["corporate-client-servicing-assistant"],"organization":{"name":"DBS Bank","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"DBS Bank","role":"in-house"}],"summary":"DBS rolled out a generative AI version of DBS Joy, its virtual assistant for corporate clients, inside the IDEAL digital banking platform. It answers servicing questions around the clock from the bank's own knowledge base, passes complex requests to a service specialist who has a generative AI copilot, and its answers are reviewed afterwards by experienced customer service agents working as DBS Joy evaluators. DBS reports customer satisfaction scores improved by over 23% over the trial period. A later update made DBS Joy agentic (see the separate DBS Joy and digibot record).","stage":"scaled","year":2025,"channels":["web-chat"],"languages":[],"metrics":[{"kpi":"interactions-handled","value":120000,"unit":"count","qualifier":"at-least","period":"unique chats since the start of trials","claimant":"organization","quote":"Since early trials of the new features started in February, DBS Joy has managed over 120,000 unique chats and counting.","sourceUrl":"https://www.dbs.com/newsroom/DBS_rolls_out_Gen_AI_powered_chatbot_to_all_corporate_clients"},{"kpi":"users-served","value":4000,"unit":"count","qualifier":"approximately","period":"corporate clients per month","claimant":"organization","quote":"About 4,000 corporate clients, the vast majority of which are small and medium enterprises, now use the service every month.","sourceUrl":"https://www.dbs.com/newsroom/DBS_rolls_out_Gen_AI_powered_chatbot_to_all_corporate_clients"},{"kpi":"customer-satisfaction-uplift","value":23,"unit":"percent","qualifier":"at-least","period":"users of the virtual agent, from the start of trials in February to November 2025","claimant":"organization","quote":"In addition to quicker responses and shorter wait times, users of the virtual agent were also more satisfied with their experience, with customer satisfaction scores improving by over 23% in the same period.","sourceUrl":"https://www.dbs.com/newsroom/DBS_rolls_out_Gen_AI_powered_chatbot_to_all_corporate_clients"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.dbs.com/newsroom/DBS_rolls_out_Gen_AI_powered_chatbot_to_all_corporate_clients","title":"DBS rolls out Gen AI-powered chatbot to all corporate clients","publisher":"DBS Bank","date":"2025-11-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"dbs-joy-corporate-virtual-assistant"},{"title":"DBS and UOB: first live authenticated agentic transaction in Singapore with Mastercard Agent Pay","useCases":["agentic-payment-initiation"],"organization":{"name":"DBS Bank","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Mastercard","role":"platform"},{"name":"CardInfoLink","role":"integrator"}],"summary":"Mastercard completed its first live, authenticated agentic transaction in Singapore with DBS and UOB: an AI agent from CardInfoLink booked and paid for a ride to Changi Airport through the mobility provider hoppa. Each agent receives its own Mastercard Agentic Token, the consumer's consent is captured explicitly and the purchase is confirmed with Mastercard Payment Passkeys on a tokenized credential. DBS stresses safeguards, transparency and customer authorization, and UOB calls the collaboration a framework for agentic payments. It is a single milestone transaction, and no outcome figures are disclosed.","stage":"pilot","year":2026,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.mastercard.com/news/ap/en/newsroom/press-releases/en/2026/mastercard-delivers-its-first-live-agentic-transaction-in-singapore-with-dbs-and-uob/","title":"Mastercard delivers its first live agentic transaction in Singapore with DBS and UOB","publisher":"Mastercard Newsroom"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"dbs-uob-mastercard-agentic-transaction"},{"title":"DeAcero: agents that turn architectural blueprints into cost proposals","useCases":["sales-quote-and-estimate-generation"],"organization":{"name":"DeAcero","anonymized":false,"country":"MX","region":"latin-america","industry":"manufacturing"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"DeAcero, a Mexican steel producer, built agents on Google's Gemini Enterprise Agent Platform with multimodal models that analyze architectural blueprints in PDF format and turn them into detailed cost proposals for customers. Google Cloud reports that this sharply cut DeAcero's response time to customers but gives no figure.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"deacero-blueprint-cost-proposal-agents"},{"title":"Definity: call summaries and real time recommendations for contact centre staff","useCases":["live-agent-assist"],"organization":{"name":"Definity","anonymized":false,"country":"CA","region":"north-america","industry":"insurance"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Deloitte","role":"integrator"}],"summary":"Definity, the parent of several Canadian property and casualty insurers, worked with Deloitte to use Google's AI in its contact centre: summarizing calls, automating caller authentication, analyzing customer sentiment and giving team members real time recommendations. Calls are transcribed, passed through data loss prevention and summarized by language models, with the summaries stored in Salesforce. Definity says automated summaries cut three and a half minutes from each call within about a month; Google Cloud reports shorter call handling times and higher productivity overall.","stage":"production","year":2026,"channels":["voice","agent-desktop"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":20,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"Definity, with help from Google Cloud partner Deloitte, leverages Google’s AI capabilities to summarize calls, automate caller authentication, analyze customer sentiment, and provide real-time recommendations to contact center team members — reducing call handle times by 20% and boosting productivity by 15%.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"},{"kpi":"productivity-gain","value":15,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"Definity, with help from Google Cloud partner Deloitte, leverages Google’s AI capabilities to summarize calls, automate caller authentication, analyze customer sentiment, and provide real-time recommendations to contact center team members — reducing call handle times by 20% and boosting productivity by 15%.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"},{"kpi":"time-saved-per-task","value":3.5,"unit":"minutes","qualifier":"exact","period":"per call, within about a month of automating call summaries","baseline":"Staff previously spent three to five minutes writing call summary notes","claimant":"organization","quote":"Within about a month, we cut the time agents spend on each call by three and a half minutes.","sourceUrl":"https://cloud.google.com/customers/definity"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"},{"url":"https://cloud.google.com/customers/definity","title":"Definity: Modernizing call center experiences with AI","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"definity-contact-centre-agent-assist"},{"title":"Delaware County: automated Spanish text to voice translation on calls","useCases":["public-service-translation"],"organization":{"name":"Delaware County","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Prepared","role":"platform"}],"summary":"Delaware County (Chief of Communications Anthony Mignogna) uses an automated, real time text to voice translation tool to communicate directly with Spanish speaking callers. The vendor reports that the county and other agencies have significantly reduced call processing time for Spanish calls; the detailed figures sit behind a download form and are not recorded here.","stage":"production","year":2025,"channels":["voice"],"languages":["en","es"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.prepared911.com/case-studies/delaware-county-spanish-text-to-voice-translation","title":"Delaware County Streamlines Spanish Call Processing with Text-to-Voice Translation","publisher":"Prepared"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"delaware-county-911-spanish-voice-translation"},{"title":"Delta Air Lines: Delta Concierge AI assistant for disruptions, cancellations and bags","useCases":["flight-disruption-and-rebooking-agent"],"organization":{"name":"Delta Air Lines","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[],"summary":"Delta Concierge is an AI powered assistant inside the Delta app for SkyMiles Members. It authenticates the member, recognizes the nearest upcoming trip, explains what changed when a flight is disrupted and helps find alternative flights, cancels eligible flights and submits a refund request or issues an eCredit in real time, and summarizes bag status. It launched in beta to selected members from October 2025, was expanded in phases and became available to all SkyMiles Members in August 2026, with a handoff to Delta staff when it cannot help.","stage":"scaled","year":2025,"channels":["mobile-app"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://news.delta.com/more-control-fewer-taps-delta-concierge-expands-all-skymiles-members","title":"More control, fewer taps: Delta Concierge expands to all SkyMiles Members","publisher":"Delta News Hub","date":"2026-08-07"},{"url":"https://news.delta.com/smarter-journeys-start-here-delta-concierge-now-beta-rollout","title":"Smarter journeys start here: Delta Concierge now in beta rollout","publisher":"Delta News Hub","date":"2025-10-29"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"delta-concierge-ai-assistant"},{"title":"UK Department for Education: Correspondence Drafter for replies to external queries","useCases":["civil-servant-drafting-copilot"],"organization":{"name":"Department for Education","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Microsoft (Azure OpenAI Service)","role":"model-provider"}],"summary":"The Department for Education's correspondence teams paste an incoming query into the Correspondence Drafter, which retrieves the relevant passages from the department's core briefing packs of approved standard lines and drafts a reply in email form, with options to adjust tone and audience. Staff edit and quality assure every draft before sending. The department says the new process has been calculated to be 30 times quicker than searching briefing packs and copying standard lines by hand, which took about 30 minutes per reply, and expects the final phase to cover about 800 of the roughly 1,000 external queries a month that need a reply. The record describes a private beta that had not yet entered user testing, so no measured result is published.","stage":"pilot","year":2025,"channels":["internal-tools","email"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/dfe-correspondence-drafter","title":"DfE: Correspondence Drafter (algorithmic transparency record)","publisher":"GOV.UK","date":"2025-12-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"department-for-education-correspondence-drafter"},{"title":"Deutsche Bank: agentic AI for trade and communications surveillance with Google Cloud (reported)","useCases":["market-abuse-surveillance-triage"],"organization":{"name":"Deutsche Bank","anonymized":false,"country":"DE","region":"europe","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"According to Bloomberg reporting relayed by trade press in February 2026, Deutsche Bank is working with Google Cloud on AI agents that monitor trading, spot anomalies in orders, trades and market moves and flag them to a human compliance officer, and plans to use the same approach on the communications of client facing staff such as traders and salespeople. AI News adds that Goldman Sachs is exploring agentic surveillance too and that human compliance staff remain responsible for reviewing flagged cases. No bank statement or outcome figure was found.","stage":"announced","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.pymnts.com/news/artificial-intelligence/2026/deutsche-bank-google-build-ai-agents-patrol-trading/","title":"Deutsche Bank and Google Build AI Agents to Patrol Trading","publisher":"PYMNTS","date":"2026-02-25"},{"url":"https://www.artificialintelligence-news.com/news/goldman-sachs-and-deutsche-bank-test-agentic-ai-for-trade-surveillance/","title":"Goldman Sachs and Deutsche Bank test agentic AI for trade surveillance","publisher":"AI News","date":"2026-02-27"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"deutsche-bank-agentic-trade-surveillance"},{"title":"Deutsche Bank: DB Lumina research agent for analysts","useCases":["investment-research-summarization"],"organization":{"name":"Deutsche Bank","anonymized":false,"country":"DE","region":"europe","industry":"capital-markets"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Deutsche Bank","role":"in-house"}],"summary":"Deutsche Bank Research built DB Lumina, a research agent on Google Cloud and Gemini models that helps analysts ingest documents, summarize earnings releases and investor transcripts, answer questions with inline citations and edit notes, with guardrails, access control and audit logging. It went live in September 2024 and, when the bank described it in September 2025, was used by around 5,000 people across Deutsche Bank Research and divisions such as Investment Bank Origination and Advisory and Fixed Income and Currencies. Analysts report saving 30 to 45 minutes on earnings note templates and up to two hours on research reports and roadshow updates. The bank evaluates it with stable test sets, automated metrics and human review.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":5000,"unit":"count","qualifier":"approximately","period":"users at publication in September 2025","claimant":"organization","quote":"Currently, DB Lumina is already in the hands of around 5,000 users across Deutsche Bank Research, specifically in divisions like Investment Bank Origination & Advisory and Fixed Income & Currencies.","sourceUrl":"https://cloud.google.com/blog/topics/financial-services/deutsche-bank-delivers-ai-powered-financial-research-with-db-lumina"},{"kpi":"time-saved-per-task","value":120,"unit":"minutes","qualifier":"up-to","period":"per research report or roadshow update","claimant":"organization","quote":"Time savings: Analysts reported significant time savings, saving 30 to 45 minutes on preparing earnings note templates and up to two hours when writing research reports and roadshow updates.","sourceUrl":"https://cloud.google.com/blog/topics/financial-services/deutsche-bank-delivers-ai-powered-financial-research-with-db-lumina"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/blog/topics/financial-services/deutsche-bank-delivers-ai-powered-financial-research-with-db-lumina","title":"Deutsche Bank delivers AI-powered financial research with DB Lumina","publisher":"Google Cloud Blog","date":"2025-09-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"deutsche-bank-db-lumina-research"},{"title":"Deutsche Bank: agentic AI for Source of Wealth checks in private bank onboarding","useCases":["digital-onboarding-assistant","source-of-wealth-diligence"],"organization":{"name":"Deutsche Bank","anonymized":false,"country":"DE","region":"asia-pacific","industry":"banking"},"vendors":[],"summary":"Deutsche Bank Private Bank put an agentic AI solution live in its Singapore and Hong Kong booking centres that researches, documents and prepares Source of Wealth assessments, which the bank calls one of the most resource intensive parts of know your customer checks. It reads client documents alongside approved external data, flags gaps and inconsistencies, and hands the assessment to bank staff for review; relationship managers in Dubai use it for accounts booked in Singapore. The bank stresses that accountability stays with its people. No outcome figures are disclosed; the growth figures in the coverage are forecasts.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.fstech.co.uk/fst/Deutsche_Bank_Rolls_Out_Agentic_AI_To_Streamline_KYC_Onboarding_In_Private_Banking.php","title":"Deutsche Bank rolls out agentic AI to streamline KYC onboarding in private banking","publisher":"FStech","date":"2026-09-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"deutsche-bank-source-of-wealth-kyc-agent"},{"title":"Deutsche Bank: automated adverse media and PEP screening for new accounts and refreshes with WorkFusion","useCases":["pep-and-adverse-media-screening","perpetual-kyc"],"organization":{"name":"Deutsche Bank","anonymized":false,"country":"DE","region":"global","industry":"banking"},"vendors":[{"name":"WorkFusion","role":"platform"}],"summary":"Deutsche Bank used WorkFusion's AI automation for screening work in anti money laundering, including adverse media monitoring and PEP checks for new accounts and refresh screenings, which had required large teams to scan news reports manually. For the KYC programme as a whole, the vendor reports shorter handling times, about 25,000 cases handled per quarter and tens of thousands of hours saved each year.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":50,"unit":"percent","qualifier":"up-to","period":"range of 25 to 50%, across the whole KYC programme (screening and document processing)","claimant":"vendor","quote":"25–50% reduction in handling time","sourceUrl":"https://www.workfusion.com/customer-stories/deutsche-bank/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.workfusion.com/customer-stories/deutsche-bank/","title":"Deutsche Bank Customer Story","publisher":"WorkFusion","date":"2020-11-05"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"deutsche-bank-workfusion-screening-automation"},{"title":"Deutsche Telekom: RAN Guardian and MINDR agents for self healing network operations","useCases":["autonomous-network-operations","network-fault-triage-copilot","network-planning-and-capacity-optimization"],"organization":{"name":"Deutsche Telekom","anonymized":false,"country":"DE","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Google (Gemini models)","role":"model-provider"}],"summary":"Deutsche Telekom's RAN Guardian Agent, built with Gemini models on Google Cloud, went live in its German mobile network in November 2025. It is a multi agent system: one agent finds upcoming public events from public sources, another assesses whether nearby cells can carry the expected traffic and monitors them live, and a third executes corrective actions such as reallocating resources or adjusting configuration, documenting every action. It is being extended to the Czech Republic and Croatia. In February 2026 Deutsche Telekom announced MINDR, which applies the same approach end to end across radio, transport and core domains, with first production releases planned for later in 2026.","stage":"production","year":2025,"channels":["internal-tools","api"],"languages":[],"metrics":[{"kpi":"processing-time-reduction","value":95,"unit":"percent","qualifier":"at-least","period":"live operations, major events","baseline":"Time needed to manage major events before the agent (hours)","claimant":"organization","quote":"And in live operations it has reduced the time needed to manage major events from hours to around a minute, a more than 95% improvement.","sourceUrl":"https://www.telekom.com/en/newsroom/latest-updates/media-information/2026/2/mindr-ai-agents-in-telekom-network"},{"kpi":"interactions-handled","value":100,"unit":"count","qualifier":"at-least","period":"first month after launch, Christmas market events","claimant":"organization","quote":"Since its launch in November 2025, RAN Guardian Agent has autonomously triggered over 100 remediation actions at Christmas market events during its first month.","sourceUrl":"https://www.telekom.com/en/newsroom/latest-updates/media-information/2026/2/mindr-ai-agents-in-telekom-network"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.telekom.com/en/newsroom/latest-updates/media-information/2026/2/mindr-ai-agents-in-telekom-network","title":"Deutsche Telekom and Google Cloud Collaborate for Superior Network Experience with Agentic AI","publisher":"Deutsche Telekom","date":"2026-02-25"},{"url":"https://www.telekom.com/en/newsroom/latest-updates/media-information/2025/11/deutsche-telekom-ai-agents-for-mobile-network","title":"Deutsche Telekom: AI agents for mobile network","publisher":"Deutsche Telekom","date":"2025-11-11"},{"url":"https://www.telekom.com/en/newsroom/latest-updates/media-information/2025/2/agentic-ai-for-autonomous-networks","title":"Deutsche Telekom and Google Cloud Partner on Agentic AI for Autonomous Networks","publisher":"Deutsche Telekom","date":"2025-02-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"deutsche-telekom-ran-guardian-and-mindr-agents"},{"title":"Deutsche Telekom: digital human sales and service assistants","useCases":["retail-store-and-kiosk-assistant"],"organization":{"name":"Deutsche Telekom","anonymized":false,"country":"DE","region":"europe","industry":"telecommunications"},"vendors":[{"name":"UneeQ","role":"platform"}],"summary":"Deutsche Telekom uses a family of UneeQ digital humans that speak German and guide customers to products: Selena explains home broadband options after asking about the customer's home and lifestyle, Max answers customer questions on the Telekom website and app, and Mia works at events such as MWC. The vendor presents them as a way to give hesitant customers confidence in technical purchases. The case study shows conversion, cart abandonment and rating tiles without readable figures, so no metric is recorded. The case study does not describe screens in Telekom shops; the kiosk channel reflects Mia's use at events.","stage":"production","year":2024,"channels":["web-chat","mobile-app","kiosk"],"languages":["de"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.digitalhumans.com/case-studies/deutsche-telekom-case-study","title":"UneeQ Case Study, Deutsche Telekom","publisher":"UneeQ"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"deutsche-telekom-uneeq-digital-humans"},{"title":"Dubai Electricity and Water Authority: Rammas customer service chatbot","useCases":["utility-billing-and-move-agent"],"organization":{"name":"Dubai Electricity and Water Authority","anonymized":false,"country":"AE","region":"middle-east","industry":"energy-and-utilities"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"DEWA introduced its Rammas chatbot in 2017, which Microsoft describes as the first chatbot based on Microsoft AI at a government utility, and later integrated Azure OpenAI Service to make its answers more accurate and more natural. Microsoft reports that Rammas has responded to more than 9.2 million customer inquiries autonomously since launch, for a utility with more than 1.2 million customers. DEWA also runs AI based high water usage alerts, and is exploring voice conversations for customers.","stage":"scaled","year":2017,"channels":[],"languages":[],"metrics":[{"kpi":"interactions-handled","value":9200000,"unit":"count","qualifier":"at-least","period":"cumulative since 2017","claimant":"vendor","quote":"Since then, Rammas has responded to over 9.2 million customer inquiries autonomously.","sourceUrl":"https://www.microsoft.com/en/customers/story/1803548215767581643-dewa-azure-ai-services-government-en-united-arab-emirates"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1803548215767581643-dewa-azure-ai-services-government-en-united-arab-emirates","title":"DEWA pioneers the use of Azure AI Services in delivering utility services","publisher":"Microsoft","date":"2024-08-21"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"dewa-rammas-customer-chatbot"},{"title":"DHL Express: computer vision item identification for customs descriptions at booking","useCases":["customs-classification-and-declaration"],"organization":{"name":"DHL Express","anonymized":false,"country":"DE","region":"global","industry":"logistics-and-transportation"},"vendors":[],"summary":"In May 2026 DHL Express launched an AI feature in its international booking flow: the shipper photographs the item with a smartphone or other connected device, a server side computer vision model classifies it and proposes a structured, customs compliant item description within seconds, and the shipper reviews, edits or overrides it before submitting. DHL Express says it is live in eight markets (Canada, Germany, Hong Kong, the Netherlands, Singapore, South Africa, Spain and the United Arab Emirates) with a wider rollout planned through 2026. The announcement does not say who built the model, and no outcome figures were disclosed.","stage":"production","year":2026,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://group.dhl.com/en/media-relations/press-releases/2026/dhl-express-introduces-ai-powered-item-identification-for-international-shipping.html","title":"DHL Express introduces AI-powered item identification for international shipping","publisher":"DHL Group","date":"2026-05-07","archivedUrl":"https://web.archive.org/web/2026/https://group.dhl.com/en/media-relations/press-releases/2026/dhl-express-introduces-ai-powered-item-identification-for-international-shipping.html"},{"url":"https://www.parcelandpostaltechnologyinternational.com/news/cross-border/dhl-express-introduces-ai-tool-to-automate-customs-item-descriptions.html","title":"DHL Express introduces AI tool to automate customs item descriptions","publisher":"Parcel and Postal Technology International","date":"2026-05-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dhl-express-ai-item-identification-customs"},{"title":"DHS: turning statements in Congressional reports into trackable tasks (planned)","useCases":["supervisory-exam-response-assembly"],"organization":{"name":"U.S. Department of Homeland Security","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The DHS management directorate reports a pre deployment use case that converts plain text statements in Congressional reports into machine readable tasks that can be managed in Outlook, Jira or other project tracking software, and turns scanned financial tables into structured data. It corresponds to the commitment tracking step of oversight work: statements an overseer writes down become tasks that can be managed in project tracking software. Not yet live; no results published.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DHS-2342, JES and Appropriations Insight)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dhs-congressional-report-task-extraction"},{"title":"DHS Executive Secretariat: generative AI summaries of incoming inquiries and information requests","useCases":["supervisory-exam-response-assembly"],"organization":{"name":"U.S. Department of Homeland Security","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Since December 2024 the DHS Executive Secretariat has used generative AI to summarise incoming letters when a new work package is created in its correspondence tracking system, so analysts can act on and assign requests faster; the summaries flow into the system that tracks correspondence and information requests. The department states a future aim of drafting responses to those requests. It is the intake and tracking half of answering an external request, run by a government body rather than a bank.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DHS-2453, ESEC Inquiry (STORM) Summarization)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://www.dhs.gov/ai/use-case-inventory/mgmt","title":"DHS AI Use Case Inventory, Management Directorate (entry DHS-2453, ESEC Inquiry (STORM) Summarization)","publisher":"U.S. Department of Homeland Security"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dhs-executive-secretariat-inquiry-summarization"},{"title":"Discover Financial Services: patented SHAP based generation of adverse action reason codes","useCases":["adverse-action-explanations"],"organization":{"name":"Discover Financial Services","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Discover Financial Services","role":"in-house"}],"summary":"Discover Financial Services patented a framework that generates the adverse action reason codes a lender must give a declined credit applicant directly from a machine learning model. The system groups correlated input variables, scores each group with partial dependence plots and Shapley Additive Explanations, ranks the groups, and turns the top ranked groups into the reason codes sent to the applicant. The United States Patent and Trademark Office granted the patent in July 2024 on an application Discover filed in May 2020. Discover Financial Services merged into Capital One Financial Corporation in May 2025 (the patent assignment was recorded in July 2025), which is why current patent databases list Capital One as the assignee. No source discloses an error rate, approval volume or other outcome for the system.","stage":"announced","year":2024,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://patents.google.com/patent/US12050975B2/en","title":"US12050975B2: System and method for utilizing grouped partial dependence plots and Shapley additive explanations in the generation of adverse action reason codes","publisher":"United States Patent and Trademark Office","date":"2024-07-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"discover-financial-services-adverse-action-reason-codes"},{"title":"US Department of Justice: AI assisted search in Westlaw and Lexis, and a CoCounsel pilot","useCases":["legal-research-and-drafting-assistant"],"organization":{"name":"U.S. Department of Justice","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Thomson Reuters","role":"platform"},{"name":"LexisNexis","role":"platform"}],"summary":"The Department of Justice lists AI assisted legal research in Westlaw and LexisNexis as deployed department wide since 2019 and 2020. These features recommend case law, statutes, regulations and scholarly articles and in some circumstances give an overview of the law in response to a query; the inventory classifies them as NLP and classical machine learning, so they are AI assisted search and recommendation rather than a generative drafting assistant. The generative tool on record is a CoCounsel pilot at the Executive Office for United States Attorneys since March 2024, meant for document review and other tasks as a comparison point to other commercial alternatives. Its listed outputs are automated transcriptions, metadata tagging suggestions and object and facial recognition in media files, not legal research or drafting. No outcome figures are published.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entries DOJ-0105, DOJ-0109 and DOJ-0313)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"doj-ai-assisted-legal-research"},{"title":"US Department of Justice: AI features in eLitigation tools for review of voluminous electronic evidence","useCases":["ediscovery-and-disclosure-document-review"],"organization":{"name":"U.S. Department of Justice","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Relativity","role":"platform"},{"name":"Everlaw","role":"platform"},{"name":"Nuix","role":"platform"},{"name":"CloudNine Law","role":"platform"}],"summary":"Department of Justice components use commercial eLitigation platforms such as Relativity and Everlaw for investigations, litigation and FOIA and Privacy Act work, and these tools increasingly include AI. The department's inventory entry describes surfacing potentially discoverable material in large collections of emails, text messages and other records, locating potentially inculpatory or exculpatory evidence, and identifying material to disclose or withhold under legal rules and privileges. The entry is department wide and deployed since January 2024. It is formally classified as \"Presumed High-Impact, but Not High-impact\", because the output is not the principal basis for decisions with legal or similarly significant effect, while noting that the tools are used in high impact contexts and that individual uses vary. No outcome figures are published.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DOJ-0102, eLitigation Tools)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"doj-elitigation-ai-document-review"},{"title":"US Executive Office for Immigration Review: planned summaries of immigration court filings","useCases":["court-and-case-file-summarization"],"organization":{"name":"U.S. Department of Justice, Executive Office for Immigration Review","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Executive Office for Immigration Review, which runs the US immigration courts, lists a pre deployment use case for summarising court filings. Filings can run to hundreds of pages and are often poorly organised; the planned tool would summarise their contents with references to the source of the information, tab and label submission types, point adjudicators and legal support staff to where relevant content sits in the record, and summarise case law for training material. The stated aim is to let adjudicators spend their time on legal analysis and conclusions.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"doj-eoir-immigration-filing-summaries"},{"title":"US Department of Justice: AI vendor risk profiles for supply chain risk management","useCases":["vendor-due-diligence"],"organization":{"name":"U.S. Department of Justice","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Exiger","role":"platform"}],"summary":"Since September 2023 the Justice Management Division has used Exiger's DDIQ platform to run supply chain risk management assessments. The AI continuously ingests sources such as news articles, legal filings and public records and returns vendor profiles (corporate records, beneficial owners, sanctions lists, adverse media, litigation history, financial stability and supply chain relationships), risk dashboards and risk scores for foreign ownership, control or influence, reputational, criminal and regulatory issues and financial health. The department uses the output to decide whether to move forward with the acquisition of goods or services. No outcome figures are published.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entries DOJ-0174 and DOJ-0175, Exiger Supply Chain Risk Management)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI inventory (DOJ entries Exiger Supply Chain Risk Management)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"doj-exiger-supply-chain-risk-management"},{"title":"US Department of Justice: AI features in FOIA production tools for classification, redaction and deduplication","useCases":["freedom-of-information-request-processing"],"organization":{"name":"U.S. Department of Justice","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"FOIAXpress","role":"platform"},{"name":"Forum One","role":"integrator"},{"name":"Adobe","role":"platform"},{"name":"Polydelta","role":"integrator"}],"summary":"The Department of Justice reports a department wide entry for FOIA production tools, deployed in January 2025, that add AI to its FOIAXpress request processing: classifying documents, identifying sensitive or confidential information for redaction and removing duplicate documents. The outputs are classifications, recommendations and predictions, and the department rates the use as not high impact because it is not the principal basis for decisions with legal or significant effect. The expected benefits are faster processing, fewer human errors and more accurate, compliant processing; no measured outcome is published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"doj-foia-production-tools"},{"title":"US Department of Labor: note taking bot for meeting summaries and action items","useCases":["meeting-summarization-and-action-items"],"organization":{"name":"U.S. Department of Labor","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"The Department of Labor's Office of the Chief Information Officer runs a note taking bot that turns meeting transcripts into concise, searchable notes with a summary and action items, to reduce manual note taking and improve information sharing. It is listed as deployed since November 2024 and not high impact, and reports that it has no authority to operate (ATO); no outcome figures are published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DOL-32, Note Taking Bot)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dol-note-taking-bot"},{"title":"DOMCURA: Claimens voice agent for claim reporting","useCases":["claims-first-notice-of-loss-agent"],"organization":{"name":"DOMCURA","anonymized":false,"country":"DE","region":"europe","industry":"insurance"},"vendors":[{"name":"Parloa","role":"platform"},{"name":"Microsoft","role":"platform"}],"summary":"DOMCURA, a German underwriting agent, turned its claims chatbot Claimens into a phone based AI agent with Parloa that guides callers through the recurring steps of reporting a claim, matches the caller to the policy by recognising the policy number, and covers more than 20 types of damage claims that the DOMCURA team configured itself. It went from kickoff to live launch in three months. Parloa reports a 90% recognition rate for caller requests.","stage":"production","year":2023,"channels":["voice"],"languages":["de"],"metrics":[{"kpi":"accuracy","value":90,"unit":"percent","qualifier":"exact","period":"Recognition of caller requests","claimant":"vendor","quote":"90% recognition rate for requests","sourceUrl":"https://www.parloa.com/customers/the-evolution-from-chatbot-to-agent-claimens-makes-filling-claim-reports-easier-than-ever/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.parloa.com/customers/the-evolution-from-chatbot-to-agent-claimens-makes-filling-claim-reports-easier-than-ever/","title":"The evolution from chatbot to voice AI: “Claimens” makes filling claim reports easier than ever","publisher":"Parloa","date":"2025-07-16"},{"url":"https://web.archive.org/web/20230205195152/https://www.parloa.com/customers/domcura/","title":"The evolution from chatbot to phone bot: “Claimens” makes filling claim reports easier than ever (archived February 2023)","publisher":"Parloa (via Internet Archive)","date":"2023-02-05"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"domcura-claimens-voice-claims-reporting"},{"title":"Domino's: AI cash flow categorisation and forecasting for treasury","useCases":["treasury-cash-flow-forecasting"],"organization":{"name":"Domino's Pizza","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"JPMorgan Chase","role":"platform"}],"summary":"Domino's treasury team adopted J.P. Morgan Payments Cash Flow Intelligence to aggregate, categorise and reconcile cash flows across a franchise model of more than 20,500 stores in 90 markets and a securitised debt structure. The team runs weekly cash reviews and forecast updates in the tool and reports forecasts that aligned closely with its 2024 budget.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":90,"unit":"percent","qualifier":"up-to","period":"weekly manual data cleanup","claimant":"vendor","quote":"According to Domino’s Treasury Team Leader, Nancy Romain, the team’s weekly manual data cleanup efforts were reduced by up to 90%.","sourceUrl":"https://www.jpmorgan.com/insights/payments/data-intelligence/dominos-pizza-cash-flow-intelligence"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.jpmorgan.com/insights/payments/data-intelligence/dominos-pizza-cash-flow-intelligence","title":"Domino's Unlocks Efficiency With Cash Flow Intelligence","publisher":"J.P. Morgan","date":"2024-08-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"dominos-cash-flow-intelligence"},{"title":"DoorDash: automated quality evaluation across 19,000 support agents","useCases":["call-quality-and-compliance-monitoring"],"organization":{"name":"DoorDash","anonymized":false,"country":"US","region":"global","industry":"retail-and-ecommerce"},"vendors":[{"name":"Observe.AI","role":"platform"},{"name":"AWS","role":"platform"}],"summary":"DoorDash moved from manually reviewing a small sample of support interactions to automated evaluation of nearly all of them, reaching nearly 100% automated quality coverage across 19,000 frontline teammates in its own teams and BPO partners. It uses sentiment, comprehension and behavioural signals rather than only binary compliance checklists, and coaches from AI generated insights. Emerging problems that took days or weeks to surface are now seen in near real time.","stage":"scaled","year":2026,"channels":["voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.observe.ai/customers/doordash-scales-customer-centric-ai-across-19-000-agents-with-nearly-100-automated-quality-coverage","title":"DoorDash Scales Customer-Centric AI Across 19,000 Agents With Nearly 100% Automated Quality Coverage","publisher":"Observe.AI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"doordash-automated-quality-coverage"},{"title":"DPD Germany: Red, an AI chatbot for parcel status, redirection and shipping","useCases":["parcel-tracking-and-delivery-exception-agent"],"organization":{"name":"DPD Deutschland","anonymized":false,"country":"DE","region":"europe","industry":"logistics-and-transportation"},"vendors":[],"summary":"DPD Germany launched Red in October 2019, an AI controlled chatbot on dpd.de that lets parcel recipients check the status of a shipment on desktop and mobile, with a link to the service team when it cannot answer. Red later added redirection options and a deposit permission (leave the parcel without a signature), and from 2021 it also lets shippers and recipients book a parcel shipment. DPD Belgium runs a similar consignee chatbot, launched as Phil in 2020 and replaced by the self learning Ruby in 2022.","stage":"production","year":2019,"channels":["web-chat"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.dpd.com/de/en/news/digitaler-innovationsfuhrer-legt-nach-dpd-startet-ki-basierten-chatbot-2/","title":"Digital innovation leader adds new service: DPD launches AI-based chatbot","publisher":"DPD","date":"2019-10-23"},{"url":"https://www.dpd.com/de/en/news/weiterentwicklung-des-chatbots-red-kann-jetzt-auch-pakete-versenden/","title":"Further development of the chatbot: Red can now also send parcels","publisher":"DPD","date":"2021-02-03"},{"url":"https://www.dpd.com/be/en/news/meet-ruby-dpd-s-self-learning-chatbot/","title":"Meet Ruby, DPD's self-learning chatbot","publisher":"DPD Belgium","date":"2022-03-01"},{"url":"https://www.dpd.com/be/en/news/digital-innovation-leader-adds-new-service-dpd-belgium-belux-launches-consignee-chatbot/","title":"DPD Belgium launches consignee chatbot","publisher":"DPD Belgium","date":"2020-04-03"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dpd-germany-red-consignee-chatbot"},{"title":"DPD UK: AI element of the customer service chatbot disabled after it swore at a customer","useCases":["parcel-tracking-and-delivery-exception-agent"],"organization":{"name":"DPD UK","anonymized":false,"country":"GB","region":"europe","industry":"logistics-and-transportation"},"vendors":[],"summary":"DPD UK had used an AI element in its online customer service chat for several years. In January 2024, after a system update, a customer trying to trace a parcel got the chatbot to swear, call itself useless and criticise the company, and the screenshots spread widely on social media. DPD said an error occurred after the update and that the AI element was immediately disabled and being updated.","stage":"paused","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.bbc.com/news/technology-68025677","title":"DPD error caused chatbot to swear at customer","publisher":"BBC News","date":"2024-01-19"},{"url":"https://www.theguardian.com/technology/2024/jan/20/dpd-ai-chatbot-swears-calls-itself-useless-and-criticises-firm","title":"DPD AI chatbot swears, calls itself 'useless' and criticises delivery firm","publisher":"The Guardian","date":"2024-01-20"},{"url":"https://incidentdatabase.ai/cite/631/","title":"Incident 631: Chatbot for DPD Malfunctioned and Swore at Customers and Criticized Its Own Company","publisher":"AI Incident Database"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"dpd-uk-chatbot-ai-element-disabled"},{"title":"du: intent based autonomous network slicing on 5G Advanced","useCases":["autonomous-network-operations"],"organization":{"name":"du","anonymized":false,"country":"AE","region":"middle-east","industry":"telecommunications"},"vendors":[{"name":"Nokia","role":"platform"}],"summary":"du and Nokia implemented a 5G Advanced autonomous network slicing solution that continuously measures slice performance and adjusts radio policies by itself, so du can guarantee capacity or low latency for enterprise, event, gaming and broadcasting customers. Machine learning keeps an enterprise customer's capacity intent in all network conditions and enforces low latency slice policies for premium gamers in crowded areas. The companies present it as an industry first implementation; no operational results are published.","stage":"pilot","year":2025,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.nokia.com/newsroom/nokia-and-du-set-new-benchmark-in-5g-innovation-with-autonomous-network-slicing-in-industry-first/","title":"Nokia and du set new benchmark in 5G innovation with autonomous network slicing in industry first","publisher":"Nokia","date":"2025-12-03","archivedUrl":"https://web.archive.org/web/20260105214502/https://www.nokia.com/newsroom/nokia-and-du-set-new-benchmark-in-5g-innovation-with-autonomous-network-slicing-in-industry-first/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"du-nokia-autonomous-network-slicing"},{"title":"Duke Energy: central monitoring and diagnostics centre with predictive asset analytics for its generation fleet","useCases":["industrial-asset-predictive-maintenance"],"organization":{"name":"Duke Energy","anonymized":false,"country":"US","region":"north-america","industry":"energy-and-utilities"},"vendors":[{"name":"AVEVA","role":"platform"}],"summary":"Duke Energy runs a central Monitoring and Diagnostics (M&D) centre that watches coal, gas, combined cycle and other generating units in several US states with predictive asset analytics software; the AVEVA page names PRiSM Predictive Asset Analytics among its tools. Early warning notifications of equipment problems let a small team of experienced analysts alert the plants before a failure. AVEVA reports that the centre covers over 87% of Duke's generating fleet with over 11,000 models, and that a single early catch in 2016 avoided more than 34 million US dollars in cost had the problem gone undetected.","stage":"scaled","year":2016,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"cost-savings","value":34000000,"unit":"currency","currency":"USD","qualifier":"at-least","period":"a single early catch event in 2016, counterfactual avoided cost","claimant":"vendor","quote":"Savings of over $34 millions in a single early catch event in 2016.","sourceUrl":"https://www.aveva.com/en/perspectives/success-stories/duke-energy/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.aveva.com/en/perspectives/success-stories/duke-energy/","title":"Duke Energy Predictive Analytic Success Story","publisher":"AVEVA"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"duke-energy-monitoring-and-diagnostics-center"},{"title":"Dun & Bradstreet: generative AI email tool for sellers","useCases":["outbound-sales-prospecting-agent"],"organization":{"name":"Dun & Bradstreet","anonymized":false,"country":"US","region":"north-america","industry":"professional-services"},"vendors":[{"name":"Google Cloud","role":"model-provider"}],"summary":"Dun & Bradstreet, a business research and intelligence company, built an email generation tool with Google's Gemini models that helps its sellers write tailored, personalized messages to prospects and customers about its research services. No outcome has been published.","stage":"production","year":2024,"channels":["email","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"},{"url":"https://www.dnb.com/contact-us.html","title":"Contact Dun & Bradstreet for Support","publisher":"Dun & Bradstreet"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"dun-and-bradstreet-seller-email-generation"},{"title":"DVLA: natural language IVR and web chat bot in the contact centre","useCases":["citizen-information-assistant"],"organization":{"name":"Driver and Vehicle Licensing Agency","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Content Guru","role":"integrator"},{"name":"Google (Dialogflow)","role":"platform"}],"summary":"The Driver and Vehicle Licensing Agency (DVLA), which handles driving licence and vehicle enquiries such as vehicle tax, asks phone callers to say what their enquiry is about and routes them by intent to a recorded answer, an SMS link to a GOV.UK service or the right adviser. On web chat, a bot answers general enquiries with scripted responses and gathers diagnostic information before handing over to an adviser. Both run on Google Dialogflow inside the Content Guru Storm contact centre platform. The records state that neither uses generative AI to write answers (every response is configured by DVLA staff) and that neither makes significant decisions about customers. The natural language IVR cut the average time a caller spends in the IVR menus by half (90 seconds), a navigation time rather than an adviser handling time.","stage":"scaled","year":2025,"channels":["voice","web-chat"],"languages":["en"],"metrics":[{"kpi":"containment-rate","value":20,"unit":"percent","qualifier":"approximately","period":"web chat enquiries","claimant":"organization","quote":"Automation of around 20% of customer enquiries on the web chat channel through bot generated auto responses.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dvla-contact-centre-chatbot-service"},{"kpi":"users-served","value":300000,"unit":"count","qualifier":"approximately","period":"per month, web chat bot","claimant":"organization","quote":"Around 300k customers access the chat bot every month.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dvla-contact-centre-chatbot-service"},{"kpi":"time-saved-per-task","value":2,"unit":"minutes","qualifier":"approximately","period":"adviser chat handling time per handed over chat","baseline":"advisers asking for the diagnostic information themselves","claimant":"organization","quote":"Reduction in an advisor's chat handling time of around 2 minutes linked to chat bot data capture (minimising the advisor time asking for this).","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dvla-contact-centre-chatbot-service"},{"kpi":"users-served","value":900000,"unit":"count","qualifier":"approximately","period":"per month, natural language IVR","claimant":"organization","quote":"Around 900k customers access the IVR every month.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dvla-contact-centre-natural-language-ivr"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/dvla-contact-centre-chatbot-service","title":"DVLA: Contact Centre Chatbot service","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-12-16"},{"url":"https://www.gov.uk/algorithmic-transparency-records/dvla-contact-centre-natural-language-ivr","title":"DVLA: Contact Centre Natural Language IVR","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2026-04-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dvla-contact-centre-conversational-ai"},{"title":"Driver and Vehicle Standards Agency: MOT risk rating to prioritise visits to testing stations","useCases":["inspection-prioritization"],"organization":{"name":"Driver and Vehicle Standards Agency","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Kainos","role":"integrator"}],"summary":"DVSA approves MOT testers and testing stations (the record counts about 64,000 active testers and 23,000 active garages) and visits them to raise standards and detect deliberate or fraudulent testing. Its MOT risk rating, first built by Kainos and now run by DVSA, applies an outlier detection model (local outlier factor) to MOT test data and gives each tester and station a red, amber or green rating that is refreshed every month. Vehicle examiners see the rating in a Power BI app alongside other data and decide which sites to visit; the tool is not used for disciplinary decisions, which rest on evidence found during a visit. Previously visits were prioritised by time since the last visit.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":150,"unit":"count","qualifier":"approximately","period":"DVSA officers using it on a daily basis","claimant":"organization","quote":"Used on a daily basis by approx. 150 DVSA officers who will supervise visit and assess.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dvsa-mot-risk-rating"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/dvsa-mot-risk-rating","title":"DVSA: MOT Risk Rating (algorithmic transparency record)","publisher":"GOV.UK","date":"2025-02-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"dvsa-mot-garage-risk-rating"},{"title":"DWP: Conversational Platform on the benefits telephone lines","useCases":["benefits-eligibility-and-application-assistant"],"organization":{"name":"Department for Work and Pensions","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Omilia Natural Language Solutions","role":"platform"}],"summary":"Callers to the benefit telephone lines of the Department for Work and Pensions (DWP) are asked what they are calling about. Speech recognition and a natural language model then signpost them to GOV.UK, help them log in to their online account, answer some questions in the IVR (such as the next payment amount and date) or route them to the right adviser. The platform also runs identity and verification questions against the DWP trust hub by API. It makes no decisions; callers can always ask for a human.","stage":"scaled","year":2025,"channels":["voice"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":1000000,"unit":"count","qualifier":"approximately","period":"calls per month","claimant":"organization","quote":"At present approximately 1 million calls per month pass through the Conversational Platform.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dwp-conversational-platform"},{"kpi":"accuracy","value":97,"unit":"percent","qualifier":"exact","period":"speech recognition success","claimant":"organization","quote":"97% speech recognition success","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dwp-conversational-platform"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/dwp-conversational-platform","title":"DWP: Conversational Platform","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-12-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"dwp-conversational-platform"},{"title":"Department for Work and Pensions: machine learning risk model for Universal Credit advances","useCases":["benefit-fraud-and-error-detection","application-and-identity-fraud-detection"],"organization":{"name":"Department for Work and Pensions","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"In house (Integrated Risk and Intelligence Service)","role":"in-house"}],"summary":"DWP scores requests for Universal Credit advances in real time with a supervised machine learning classifier and refers the highest risk requests to a caseworker before payment. The caseworker is not told the referral came from the model, a random control group is referred alongside, and every decision to decline is made by a person and can be appealed. DWP's published effectiveness assessment for April 2025 to March 2026 finds the model 2.5 times more effective than random selection with a median payment delay of one day for approved referrals. It also finds that non UK nationals and several age bands were referred more often without a matching increase in confirmed fraud, and that referrals of couples were less often confirmed than those of single claimants; a retrained model was being tested.","stage":"scaled","year":2026,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"detection-rate-improvement","value":2.5,"unit":"multiplier","qualifier":"exact","period":"1 April 2025 to 31 March 2026","baseline":"Randomised control group sample of advances","claimant":"organization","quote":"The performance information for 2025 to 2026 demonstrates the model is 2.5 times more effective at identifying fraud risk than a randomised control group sample.","sourceUrl":"https://www.gov.uk/government/publications/effectiveness-assessment-including-statistical-analysis-of-the-universal-credit-advances-machine-learning-model-1-april-2025-to-31-march-2026/effectiveness-assessment-of-universal-credit-advances-model"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/government/publications/effectiveness-assessment-including-statistical-analysis-of-the-universal-credit-advances-machine-learning-model-1-april-2025-to-31-march-2026/effectiveness-assessment-of-universal-credit-advances-model","title":"Effectiveness Assessment of Universal Credit Advances Model","publisher":"Department for Work and Pensions","date":"2026-07-09"},{"url":"https://www.gov.uk/algorithmic-transparency-records/dwp-universal-credit-advances-model","title":"DWP: Universal Credit Advances Model (algorithmic transparency record)","publisher":"GOV.UK"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"dwp-universal-credit-advances-fraud-model"},{"title":"eBay: magical listing, generative AI that drafts listings from a photo","useCases":["product-content-and-catalog-enrichment"],"organization":{"name":"eBay","anonymized":false,"country":"US","region":"global","industry":"retail-and-ecommerce"},"vendors":[{"name":"eBay","role":"in-house"}],"summary":"eBay's magical listing tool writes item descriptions from known product attributes and, from a seller's photo, suggests the category and item specifics for the seller to review and approve. The description feature reached all sellers in eBay's top five markets by the end of 2023; a bulk version creates drafts from batches of photos, and in 2025 a simplified mobile flow starting from photos was rolled out in the US, UK and Germany. Early UK tests showed half as many steps to list.","stage":"scaled","year":2025,"channels":["mobile-app","internal-tools"],"languages":["en","de"],"metrics":[{"kpi":"users-served","value":10000000,"unit":"count","qualifier":"at-least","period":"sellers worldwide who have used any eBay AI feature (eBay AI overall, not only listing generation), by April 2025","claimant":"organization","quote":"Over 10 million sellers worldwide have already used eBay’s AI features, creating well over 100 million listings using AI and generating billions in gross merchandise volume (GMV).","sourceUrl":"https://innovation.ebayinc.com/stories/ebay-reduces-the-time-to-list-on-mobile-with-new-simplified-selling-tool-now-featuring-magical-listing-ai-technology/"},{"kpi":"interactions-handled","value":100000000,"unit":"count","qualifier":"at-least","period":"listings created with any eBay AI feature (eBay AI overall, not only listing generation), by April 2025","claimant":"organization","quote":"Over 10 million sellers worldwide have already used eBay’s AI features, creating well over 100 million listings using AI and generating billions in gross merchandise volume (GMV).","sourceUrl":"https://innovation.ebayinc.com/stories/ebay-reduces-the-time-to-list-on-mobile-with-new-simplified-selling-tool-now-featuring-magical-listing-ai-technology/"}],"outcomeDisclosed":true,"sources":[{"url":"https://innovation.ebayinc.com/stories/ebay-reduces-the-time-to-list-on-mobile-with-new-simplified-selling-tool-now-featuring-magical-listing-ai-technology/","title":"eBay Reduces the Time to List on Mobile With New Simplified Selling Tool, Now Featuring Magical Listing AI Technology","publisher":"eBay","date":"2025-04-09"},{"url":"https://innovation.ebayinc.com/stories/ebays-magical-listing-tool-wins-ai-breakthrough-award-for-best-overall-generative-ai-solution/","title":"eBay’s Magical Listing Tool Wins AI Breakthrough Award for ‘Best Overall Generative AI Solution’","publisher":"eBay"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ebay-magical-listing-generative-ai"},{"title":"Ecclesia Group: automatic matching and routing of claims correspondence","useCases":["correspondence-triage-and-routing"],"organization":{"name":"Ecclesia Group","anonymized":false,"country":"DE","region":"europe","industry":"insurance"},"vendors":[{"name":"ABBYY","role":"platform"}],"summary":"Ecclesia Group, a German insurance broker, uses ABBYY to process incoming claims correspondence. The platform extracts key data such as case numbers and licence plates from scanned documents, matches each document to the right record in the customer database and routes it to the responsible claims manager, replacing manual sorting. No outcome figures could be verified.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["de"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.abbyy.com/customer-stories/ecclesia-group-streamlines-correspondence-management-with-abbyy","title":"How Ecclesia Group Streamlined Insurance Claims Processing","publisher":"ABBYY"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ecclesia-group-claims-correspondence-routing"},{"title":"EDF: AI Bill Explainer that walks customers through their energy bill","useCases":["utility-billing-and-move-agent"],"organization":{"name":"EDF","anonymized":false,"country":"GB","region":"europe","industry":"energy-and-utilities"},"vendors":[],"summary":"EDF in the UK launched Bill Explainer, an AI powered tool in the customer account that breaks each bill down step by step with explanations tailored to the individual account. After a pilot made available to more than 58,000 customers, EDF reports that billing related contacts fell by 5.6% among customers who used it, and it announced a rollout to its 3.4 million residential and small business customers.","stage":"production","year":2026,"channels":[],"languages":["en"],"metrics":[{"kpi":"contact-deflection","value":5.6,"unit":"percent","qualifier":"exact","period":"billing related contacts among pilot customers who used the tool","claimant":"organization","quote":"Early results indicate a positive shift in how customers engage with their bills, with billing related contacts decreasing by 5.6% among those who used the tool.","sourceUrl":"https://www.edfenergy.com/media-centre/edf-harnesses-ai-help-customers-overcome-one-their-most-common-questions"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.edfenergy.com/media-centre/edf-harnesses-ai-help-customers-overcome-one-their-most-common-questions","title":"EDF harnesses AI to help customers overcome one of their most common questions","publisher":"EDF","date":"2026-05-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"edf-ai-bill-explainer"},{"title":"Elephant Insurance: AI subrogation detection","useCases":["subrogation-opportunity-detection"],"organization":{"name":"Elephant Insurance","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Shift Technology","role":"platform"}],"summary":"Elephant Insurance, a Virginia based auto insurer owned by Admiral Group, added Shift Subrogation Detection in November 2024 after using Shift for underwriting fraud since 2021 and claims fraud since 2020. Elephant's head of claims says the model helps the insurer find subrogation opportunities at scale and recommends handler actions to improve recovery, alongside the fraud detection already in place. No recovery figures were published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.elephant.com/newsroom/press-releases/elephant-insurance-to-expand-relationship-with-shift-technology","title":"Elephant Insurance to Expand Relationship with Shift Technology","publisher":"Elephant Insurance","date":"2024-11-13","archivedUrl":"https://web.archive.org/web/20241206005506/https://www.elephant.com/newsroom/press-releases/elephant-insurance-to-expand-relationship-with-shift-technology"},{"url":"https://www.shift-technology.com/resources/news/elephant-insurance-expands-relationship-with-shift-technology","title":"Elephant Insurance Expands Relationship with Shift Technology","publisher":"Shift Technology"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"elephant-insurance-subrogation-detection"},{"title":"Elon University: ElonGPT academic advising chatbot","useCases":["academic-advising-assistant"],"organization":{"name":"Elon University","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[],"summary":"Elon University runs ElonGPT, an AI chatbot published on its Office of Academic Advising site as a supplemental resource for general advising questions: what academic advising is, how to select courses for an upcoming semester, and general degree requirements. The university tells students to verify course to requirement mapping in the institution's own degree audit tool (My Progress in OnTrack) and to confirm anything the chatbot says against the academic catalog, their assigned advisor, or a professional advisor in the Office of Academic Advising.","stage":"production","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.elon.edu/u/academics/koenigsberger-learning-center/academic-advising/elongpt/","title":"ElonGPT | Academic Advising | Elon University","publisher":"Elon University","archivedUrl":"https://web.archive.org/web/20260209110026/https://www.elon.edu/u/academics/koenigsberger-learning-center/academic-advising/elongpt/"},{"url":"https://web.archive.org/web/20240713121348/https://www.elon.edu/u/academics/koenigsberger-learning-center/academic-advising/elongpt/","title":"ElonGPT | Academic Advising | Elon University (2024 Wayback capture, cited only for the launch year)","publisher":"Elon University","date":"2024-07-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"elon-university-elongpt-advising"},{"title":"Encova Insurance: AI document understanding for claims invoice intake","useCases":["correspondence-triage-and-routing"],"organization":{"name":"Encova Insurance","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"UiPath","role":"platform"}],"summary":"Encova Insurance, a US mutual insurer, replaced traditional OCR with UiPath Document Understanding in its claims invoice process, so incoming invoices are read, their data extracted and passed to automated processing, with exceptions fixed by staff in UiPath Action Center. Its solution architect says that traditional OCR got 40% of documents through without issues and that the new process has a 99% success rate, documents processed through without issues, which the vendor headline separately calls 99% accuracy. The vendor also reports that manual data entry time in the wider policy intake automation programme was cut by over 99% over the year.","stage":"scaled","year":2023,"channels":["internal-tools","api"],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":99,"unit":"percent","qualifier":"exact","period":"document understanding success rate on claims invoices, processed through without issues","baseline":"40% of documents through without issues with traditional OCR","claimant":"organization","quote":"With this new [UiPath] process, the success rate is 99%.","sourceUrl":"https://www.uipath.com/resources/automation-case-studies/encova-insurance-scales-automation"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.uipath.com/resources/automation-case-studies/encova-insurance-scales-automation","title":"Encova Insurance Scales Automation","publisher":"UiPath"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"encova-insurance-document-intake-automation"},{"title":"Enpal: machine learning roof analysis automates solar panel quotes","useCases":["sales-quote-and-estimate-generation"],"organization":{"name":"Enpal","anonymized":false,"country":"DE","region":"europe","industry":"energy-and-utilities"},"vendors":[{"name":"dida","role":"integrator"},{"name":"Google Cloud","role":"platform"}],"summary":"Enpal, a German solar energy company, worked with the AI firm dida to automate the roof assessment and panel sizing step of its solar quotes, which a salesperson had done by hand. A model trained on rooftop images from the Google Maps Platform detects the usable roof area and obstacles, projective geometry estimates the roof angle, and further steps calculate the number of panels and visualize their placement. dida says that during the six month build Enpal was able to manually adjust details such as the dimensions of a roof. The tool sizes the installation rather than setting prices. dida reports that the process now takes 15 minutes instead of 120 and that 150 Enpal employees use the tool.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"handling-time-reduction","value":87.5,"unit":"percent","qualifier":"exact","period":"staff time per solar quote, from 120 to 15 minutes","claimant":"vendor","quote":"Thanks to our solution, and the efficiency of building it with Google Cloud, what was once a manual process taking an Enpal salesperson 120 minutes to complete is now an automated process of just 15 minutes: a reduction of 87.5%.","sourceUrl":"https://cloud.google.com/blog/topics/customers/dida-creates-a-custom-ai-solution-with-google-cloud"},{"kpi":"users-served","value":150,"unit":"count","qualifier":"exact","period":"Enpal employees using the tool, four years after launch","claimant":"vendor","quote":"Four years on, this has risen to 150 Enpal employees, each saving 87.5% of their time, which they can now dedicate to other, more specialized tasks.","sourceUrl":"https://cloud.google.com/blog/topics/customers/dida-creates-a-custom-ai-solution-with-google-cloud"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/blog/topics/customers/dida-creates-a-custom-ai-solution-with-google-cloud","title":"How dida automates sales processes with mathematics and machine learning","publisher":"Google Cloud","date":"2024-06-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"enpal-solar-quote-automation"},{"title":"Entergy: satellite and AI vegetation management across its operating companies","useCases":["power-line-vegetation-management"],"organization":{"name":"Entergy","anonymized":false,"country":"US","region":"north-america","industry":"energy-and-utilities"},"vendors":[{"name":"AiDASH","role":"platform"}],"summary":"Entergy, which serves more than 3 million customers through operating companies in Arkansas, Louisiana, Mississippi and Texas, uses the AiDASH Intelligent Vegetation Management System, which analyses satellite imagery with AI to show where vegetation threatens its power lines. According to AiDASH, Entergy moved from a standardized five year trim cycle to risk based maintenance, tailoring intervals so that some areas may need trimming every two to three years while others can go as long as thirteen years. AiDASH reports that Entergy beat its vegetation reliability (Veg SAIFI) targets by over 30% in 2020 and 2021 on flat vegetation budgets, and quotes an Entergy vegetation management analyst saying all SAIFI targets were exceeded for all operating companies in the first year after IVMS was implemented and again in the second. A later AiDASH case study quotes Entergy's Vice President of Power Delivery Services saying vegetation impacts to customers improved by more than 20% within the first year.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.aidash.com/resource/entergys-journey-to-success-satellite-powered-vegetation-management/","title":"Entergy's Journey to Success: Satellite-Powered Vegetation Management","publisher":"AiDASH","date":"2023-10-17"},{"url":"https://www.aidash.com/resource/entergy-case-study/","title":"From Fixed Cycles to Risk-Based: How Entergy Improved Reliability by More Than 20%","publisher":"AiDASH","date":"2026-06-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"entergy-satellite-vegetation-management"},{"title":"US Environmental Protection Agency: risk scoring of large quantity hazardous waste generators for inspections","useCases":["inspection-prioritization"],"organization":{"name":"U.S. Environmental Protection Agency, Office of Enforcement and Compliance Assurance","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"University of Chicago Energy and Environment Lab","role":"integrator"}],"summary":"EPA's enforcement office reports in the 2025 federal AI use case inventory that it scores Large Quantity Generators of hazardous waste on a scale of 0 to 4 to support inspections under the Resource Conservation and Recovery Act (RCRA). The stated aims are reducing staff time and better identification of potential violators. The entry describes a classical machine learning model trained on historical compliance data, built in house together with the University of Chicago Energy and Environment Lab, marks it as high impact and lists it as deployed since October 2022. No outcome figures are published.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"epa-hazardous-waste-generator-inspection-risk-scoring"},{"title":"EQT: Motherbrain platform for deal sourcing and investment decisions","useCases":["deal-sourcing-and-due-diligence-assistant"],"organization":{"name":"EQT","anonymized":false,"country":"SE","region":"europe","industry":"wealth-and-asset-management"},"vendors":[{"name":"EQT Motherbrain","role":"in-house"}],"summary":"EQT, a global private markets investor, has run Motherbrain, its in house data and AI team and platform, since 2016. EQT says it uses Motherbrain across the firm to source deals and help investment teams make better informed decisions, for example by measuring the similarity between companies for competitor mapping, and a partner describes tools that rank potential targets by attractiveness across a range of criteria. The platform also captures lessons from past deals and tracks the deal pipeline, and EQT stresses that AI supports rather than replaces the dealmakers' judgment. No outcome figures have been published.","stage":"scaled","year":2016,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://eqtgroup.com/about/motherbrain","title":"A Powerful Synergy of AI and Human Expertise","publisher":"EQT"},{"url":"https://eqtgroup.com/news/eqt-s-ai-platform-motherbrain-pushes-the-boundaries-of-the-private-markets-with-novel-algorithm-for-better-decision-making","title":"EQT's AI platform Motherbrain pushes the boundaries of the private markets with novel algorithm for better decision making","publisher":"EQT","date":"2021-11-08"},{"url":"https://eqtgroup.com/thinq/technology/first-ai-native-private-equity-firm","title":"AI Promises to Make Private Equity Faster as Competition Heats Up","publisher":"ThinQ by EQT","date":"2025-10-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"eqt-motherbrain-deal-sourcing"},{"title":"Equinix: E-Bot resolves and routes IT tickets in Microsoft Teams","useCases":["it-service-desk-resolution-agent"],"organization":{"name":"Equinix","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Moveworks","role":"platform"}],"summary":"Equinix launched a Moveworks assistant, known internally as E-Bot, in April 2019. It resolves IT support issues end to end in Microsoft Teams and, for tickets it cannot finish, assigns them to the right one of thousands of IT assignment groups within 30 seconds. Moveworks contrasts this with the five hours a first line desk takes on average to read and route a ticket, and reports that E-Bot routes 82% of Equinix tickets automatically, which cut the average lifespan of all tickets by almost a third.","stage":"scaled","year":2019,"channels":["microsoft-teams"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":96,"unit":"percent","qualifier":"exact","period":"routing of tickets the assistant cannot resolve, about ten months after launch","claimant":"vendor","quote":"Instead of high-touch tickets requiring time-consuming agent attention, the Moveworks Triage Skill allows E-Bot to assign the tickets it can’t finish resolving to the correct subject matter experts — with the same 96% accuracy rate achieved by help desk agents.","sourceUrl":"https://www.moveworks.com/us/en/customers/equinix-disappears-it-queue-with-moveworks-triage-ticketing-system"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.moveworks.com/us/en/customers/equinix-disappears-it-queue-with-moveworks-triage-ticketing-system","title":"Equinix Makes IT Queues Disappear With AI Triage","publisher":"Moveworks"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"equinix-it-ticket-triage-assistant"},{"title":"Equity Residential: AI responses to customer inquiries in apartment leasing and service","useCases":["apartment-leasing-and-resident-service-agent"],"organization":{"name":"Equity Residential","anonymized":false,"country":"US","region":"north-america","industry":"real-estate"},"vendors":[],"summary":"Equity Residential, a large US apartment REIT, lists artificial intelligence responses to customer inquiries, self guided tours and enhanced service and maintenance management among its operating technology in its annual report, and says it has incorporated generative and/or agentic AI within its business. According to Multifamily Dive's report of the fourth quarter 2025 earnings call, the company's chief operating officer said its first round of centralization, automation and AI in parts of the leasing process had cut on site payroll by 15%, and that it expects more AI enabled applications and other automation, added over the next 18 months, to reduce on site payroll by a further 5% to 10% over the next several years (a forecast).","stage":"scaled","year":2025,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/906107/000119312526051433/eqr-20251231.htm","title":"Equity Residential Form 10-K for the fiscal year 2025","publisher":"Equity Residential (SEC EDGAR)","date":"2026-02-13"},{"url":"https://www.multifamilydive.com/news/equity-residential-2025-earnings-q4/811880/","title":"Equity Residential saw AI, automation bump in 2025","publisher":"Multifamily Dive","date":"2026-02-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"equity-residential-ai-leasing-and-resident-service"},{"title":"Estonia: Bürokratt, a shared virtual assistant for public sector organizations","useCases":["citizen-information-assistant"],"organization":{"name":"Estonian Information System Authority (RIA)","anonymized":false,"country":"EE","region":"europe","industry":"government"},"vendors":[{"name":"Estonian Information System Authority","role":"in-house"}],"summary":"Bürokratt is a virtual assistant for citizens that understands everyday Estonian and is available around the clock. Its development is managed by the Information System Authority, and it is built for local authorities and government agencies to manage enquiries, automate frequently asked questions and provide customer support. The authority lists about 20 organizations and websites that use it, including the state portal eesti.ee, the Estonian Tax and Customs Board, the Estonian Health Insurance Fund, the Police and Border Guard Board, the Consumer Protection and Technical Regulatory Authority and a legal information bot of the Ministry of Justice and Digital Affairs.","stage":"production","year":2026,"channels":["web-chat"],"languages":["et"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.ria.ee/en/state-information-system/artificial-intelligence/burokratt","title":"Bürokratt","publisher":"Estonian Information System Authority","date":"2026-05-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"estonian-information-system-authority-burokratt"},{"title":"Etisalat: AI next best action to retain at risk customers","useCases":["churn-prediction-and-retention-offers"],"organization":{"name":"Etisalat","anonymized":false,"country":"AE","region":"middle-east","industry":"telecommunications"},"vendors":[{"name":"Pega","role":"platform"}],"summary":"Etisalat in the UAE uses Pega Customer Decision Hub as its central decisioning engine, with predictive and adaptive models that identify each customer's context and orchestrate personalised next best actions across inbound, outbound and agent assisted channels, moving from a focus on sales to a focus on incremental value. During the COVID-19 pandemic it used the system to identify at risk customers and offer practical help such as a free VPN and its online collaboration platform. Pega reports a 15% reduction in churn and a 20% increase in renewals, and its executive quote describes a 20% year on year increase in incremental value; the churn and renewal figures first appeared in Pega's earlier case study on Etisalat's small and medium business (SMB) division, which used next best action to prioritise outbound sales calls and offers, not in connection with the COVID-19 work. All figures cover the whole decisioning programme, including sales, not retention alone.","stage":"scaled","year":2021,"channels":[],"languages":[],"metrics":[{"kpi":"churn-reduction","value":15,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"15% reduction in customer churn","sourceUrl":"https://www.pega.com/customers/etisalat-decision-hub"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.pega.com/customers/etisalat-decision-hub","title":"Etisalat revolutionizes customer experience with AI","publisher":"Pega"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"etisalat-next-best-action-retention"},{"title":"ETS: automated essay scoring engine in GRE Analytical Writing","useCases":["automated-scoring-of-written-responses"],"organization":{"name":"ETS","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[{"name":"ETS","role":"in-house"}],"summary":"ETS evaluates GRE Analytical Writing essays on a six point holistic scale, which includes a score from its own automated scoring engine, whose features ETS says are the result of nearly two decades of natural language processing research. All essays are also reviewed by trained analysts with essay similarity detection software and by experienced content experts. ETS cites a 2010 study (Attali, Bridgeman and Trapani) in which the engine agreed with a human rater on the GRE Issue and TOEFL Independent tasks more closely than two independent human raters agreed with each other. An ETS research report records that the engine was first implemented as a check score for the revised GRE General Test in August 2012.","stage":"scaled","year":2012,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.ets.org/gre/test-takers/general-test/scores/understand-scores.html","title":"Understanding GRE General Test Scores","publisher":"ETS"},{"url":"https://www.ets.org/erater/how.html","title":"How the e-rater Scoring Engine Works","publisher":"ETS"},{"url":"https://files.eric.ed.gov/fulltext/EJ1109327.pdf","title":"A Study of the Use of the e-rater Scoring Engine for the Analytical Writing Measure of the GRE revised General Test (ETS RR-14-24)","publisher":"ETS"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ets-gre-automated-essay-scoring"},{"title":"Etsy: Gemini enriches listing data and alt text across a 130 million item marketplace","useCases":["product-content-and-catalog-enrichment"],"organization":{"name":"Etsy","anonymized":false,"country":"US","region":"global","industry":"retail-and-ecommerce"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Etsy uses Gemini models with BigQuery and Dataflow to enrich data about more than 130 million items listed by more than 5 million sellers: classifying items, spotting items linked to emerging trends and generating better image alt text for listings. Etsy's engineering lead for search says the improved alt text increased visits from search engines by 5% and conversions by 3% for sellers.","stage":"production","year":2025,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/etsy-ai","title":"Etsy: Connecting nearly 90 million buyers with special items using gen AI and \"algotorial curation\"","publisher":"Google Cloud","archivedUrl":"http://web.archive.org/web/20250902172024/https://cloud.google.com/customers/etsy-ai"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"etsy-gemini-listing-enrichment"},{"title":"Euroclear: EasyFocus+ predictive analytics for matching and settlement exceptions","useCases":["settlement-fail-prediction-and-exception-management"],"organization":{"name":"Euroclear","anonymized":false,"country":"BE","region":"europe","industry":"capital-markets"},"vendors":[{"name":"Meritsoft (Cognizant)","role":"platform"},{"name":"Taskize","role":"platform"},{"name":"Microsoft","role":"platform"}],"summary":"In June 2025 Euroclear announced EasyFocus+, the next generation of its EasyFocus service, built with Meritsoft and Taskize and running on a Microsoft cloud. It uses predictive analytics to flag likely mismatches, identify root causes and resolve exceptions, and gives clients a single view of their settlement instructions across the Euroclear CSDs, which represent over 60% of EU settlement. The collaboration platform of Taskize, a member of the Euroclear group of companies, is embedded for routing and resolving the exceptions with counterparties. Euroclear positions it as support for the EU move to T+1 in October 2027.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.meritsoft.com/euroclear-teams-with-meritsoft-and-taskize-to-launch-next-generation-ai-service/","title":"Euroclear teams with Meritsoft and Taskize to launch next generation AI service","publisher":"Meritsoft","date":"2025-06-16"},{"url":"https://www.euroclear.com/newsandinsights/en/press/2025/mr-16-meritsoft-and-taskize-to-launch-next-generation-ai-service.html","title":"Euroclear teams with Meritsoft and Taskize to launch next generation AI service","publisher":"Euroclear","date":"2025-06-16","archivedUrl":"https://web.archive.org/web/2026/https://www.euroclear.com/newsandinsights/en/press/2025/mr-16-meritsoft-and-taskize-to-launch-next-generation-ai-service.html"},{"url":"https://www.euroclear.com/services/en/settlement/settlement-euroclear-bank/easyfocus.html","title":"Euroclear EasyFocus+","publisher":"Euroclear","archivedUrl":"https://web.archive.org/web/20260831082713/https://www.euroclear.com/services/en/settlement/settlement-euroclear-bank/easyfocus.html"},{"url":"https://www.marketsmedia.com/euroclear-updates-easyfocus-to-ease-t1-transition/","title":"Euroclear Updates EasyFocus+ to Ease T+1 Transition","publisher":"Markets Media"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"euroclear-easyfocus-plus-settlement-analytics"},{"title":"Europ Assistance: pricing model updates in one or two days instead of weeks","useCases":["insurance-pricing-and-actuarial-copilot"],"organization":{"name":"Europ Assistance","anonymized":false,"country":"FR","region":"europe","industry":"insurance"},"vendors":[{"name":"Akur8","role":"platform"}],"summary":"Europ Assistance adopted Akur8's cloud pricing platform, whose automated modelling cut the time spent running and updating pricing models. The vendor reports that work that took weeks now takes one or two days, that teams can reuse fitted models on new datasets, and that built in documentation lets stakeholders review and challenge the whole pricing process.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"cycle-time-days","value":2,"unit":"days","qualifier":"up-to","period":"pricing model execution and updates","baseline":"weeks before the platform","claimant":"vendor","quote":"As a result, what once took weeks in the pricing process can now be completed in just one or two days.","sourceUrl":"https://www.akur8.com/success-stories/from-weeks-to-one-day-how-europ-assistance-accelerated-pricing-with-akur8"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.akur8.com/success-stories/from-weeks-to-one-day-how-europ-assistance-accelerated-pricing-with-akur8","title":"From weeks to one day: how Europ Assistance accelerated pricing with Akur8.","publisher":"Akur8"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"europ-assistance-akur8-pricing"},{"title":"European Central Bank: Medusa for drafting and consistency checks of internal model assessment reports","useCases":["model-risk-validation-copilot"],"organization":{"name":"European Central Bank (ECB Banking Supervision)","anonymized":false,"region":"europe","industry":"government"},"vendors":[],"summary":"ECB Banking Supervision built Medusa, a suptech tool it describes as an AI application for intelligent consistency checks of these internal model assessment reports; its June 2023 overview of suptech tools lists Medusa as live, supporting the drafting and consistency checks of those reports. By October 2025 the ECB presented Medusa among its AI tools as a one stop shop for supervisory findings and measures, with smart search, reporting, visualisations and statistical analyses. The supervisors keep the judgement: the ECB stresses that its tools support and do not replace them. It is the supervisor's side of model validation, not a bank's own validation function.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.bankingsupervision.europa.eu/press/speeches/date/2023/html/ssm.sp230629~1b6d3ba3d7.en.pdf","title":"The SSM Digitalisation Blueprint","publisher":"European Central Bank","date":"2023-06-29"},{"url":"https://www.bankingsupervision.europa.eu/press/interviews/date/2024/html/ssm.in240226~c6f7fc9251.en.html","title":"From data to decisions: AI and supervision","publisher":"European Central Bank","date":"2024-02-26"},{"url":"https://www.bankingsupervision.europa.eu/press/speeches/date/2025/html/ssm.sp251014~5bc6e60334.en.html","title":"Artificial intelligence and supervision: innovation with caution","publisher":"European Central Bank","date":"2025-10-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"european-central-bank-medusa-internal-model-reports"},{"title":"European Commission: eTranslation for public administrations","useCases":["public-service-translation"],"organization":{"name":"European Commission","anonymized":false,"region":"europe","industry":"government"},"vendors":[{"name":"European Commission (Directorate General for Translation)","role":"in-house"}],"summary":"eTranslation is the European Commission's neural machine translation service, launched in 2017, trained on EU professional translation data and offered free to eligible users such as public administrations. It translates text and documents, offers an EU formal or general style, and can translate websites through an API. The Commission says most of its language tools cover all 24 official EU languages and several more, including Arabic, Chinese and Ukrainian, and that eTranslation translated 891 million pages in 2025, up from 19 million in its first year. It now sits alongside related AI tools for summaries and draft replies.","stage":"scaled","year":2017,"channels":["api","internal-tools"],"languages":["en","fr","de","nl","es","it","pl"],"metrics":[{"kpi":"interactions-handled","value":891000000,"unit":"count","qualifier":"exact","period":"pages translated in 2025","baseline":"19 million pages in its first year after the 2017 launch","claimant":"organization","quote":"In 2025, eTranslation translated 891 million pages.","sourceUrl":"https://translation.ec.europa.eu/language-data-and-ai-using-ai-break-down-language-barriers/ai-based-multilingual-services-using-eu-language-data-innovate_en"}],"outcomeDisclosed":true,"sources":[{"url":"https://commission.europa.eu/resources-partners/etranslation_en","title":"AI translation and language tools","publisher":"European Commission"},{"url":"https://translation.ec.europa.eu/language-data-and-ai-using-ai-break-down-language-barriers/ai-based-multilingual-services-using-eu-language-data-innovate_en","title":"AI-based multilingual services: using EU language data to innovate","publisher":"European Commission (Directorate General for Translation)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"european-commission-etranslation"},{"title":"Evri (formerly Hermes UK): Holly parcel chatbot and voice assistant","useCases":["parcel-tracking-and-delivery-exception-agent"],"organization":{"name":"Evri","anonymized":false,"country":"GB","region":"europe","industry":"logistics-and-transportation"},"vendors":[],"summary":"Hermes UK, now Evri, launched an AI chatbot called Holly in November 2018 to automate parcel tracking, signature questions, diversions and confirmation of delivery for recipients. In 2023 it added a voice assistant on a dedicated phone line that identifies the caller's active parcels from their phone number and routes them, with an automatic callback within 24 hours; the release says the phone line complements a recently introduced chat service. Evri's November 2024 release credits a new chatbot and the callback feature with a marked increase in consumer satisfaction. The sources do not say whether Holly itself still runs.","stage":"production","year":2018,"channels":["web-chat","voice"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":432000,"unit":"count","qualifier":"exact","period":"from the November 2018 launch to October 2019","claimant":"organization","quote":"Since the initial Chatbot launch in November, Holly has had a total of 432,000 chats resulting in a reduction of contacts to customer service operatives by 50%.","sourceUrl":"https://www.evri.com/press/feedback-shows-hermes-customers-focus-is-delivering"},{"kpi":"contact-deflection","value":50,"unit":"percent","qualifier":"exact","period":"contacts to customer service operatives, November 2018 to October 2019","claimant":"organization","quote":"Since the initial Chatbot launch in November, Holly has had a total of 432,000 chats resulting in a reduction of contacts to customer service operatives by 50%.","sourceUrl":"https://www.evri.com/press/feedback-shows-hermes-customers-focus-is-delivering"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.evri.com/press/feedback-shows-hermes-customers-focus-is-delivering","title":"Feedback Shows Hermes' Customers Focus is Delivering | Evri","publisher":"Evri","date":"2019-10-17"},{"url":"https://www.evri.com/press/evri-dials-up-new-automated-phone-line","title":"Evri dials up new automated phone line to quickly connect customers to UK advisors, as part of a £46m total investment | Evri","publisher":"Evri","date":"2023-11-10"},{"url":"https://www.evri.com/press/evri-delivers-record-year-and-significant-investment-in-customer-service","title":"Evri Delivers Record Year & Customer Service Investment | Evri","publisher":"Evri","date":"2024-11-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"evri-holly-parcel-chatbot-and-voice-assistant"},{"title":"Fannie Mae: Value Acceptance automated valuation for mortgage loans","useCases":["property-valuation-support"],"organization":{"name":"Fannie Mae","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[],"summary":"Fannie Mae's Desktop Underwriter can issue Value Acceptance, an offer to skip a traditional appraisal, using what Fannie Mae calls a robust data and modeling framework to confirm the validity of a property's value and sale price. Fannie Mae announced that, beginning in the first quarter of 2025, the eligible loan to value ratio for Value Acceptance on purchase loans for primary residences and second homes would increase from 80% to 90%, and it estimates that appraisal alternatives such as Value Acceptance and Value Acceptance + Property Data have saved mortgage borrowers more than $2.5 billion since early 2020.","stage":"scaled","year":2024,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"customer-savings","value":2500000000,"unit":"currency","currency":"USD","qualifier":"at-least","period":"since early 2020","claimant":"organization","quote":"Since early 2020, Fannie Mae estimates the use of appraisal alternatives such as Value Acceptance and Value Acceptance + Property Data on loans Fannie Mae has acquired saved mortgage borrowers more than $2.5 billion.","sourceUrl":"https://www.fanniemae.com/newsroom/fannie-mae-news/fannie-mae-announces-changes-appraisal-alternatives-requirements"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.fanniemae.com/newsroom/fannie-mae-news/fannie-mae-announces-changes-appraisal-alternatives-requirements","title":"Fannie Mae Announces Changes to Appraisal Alternatives Requirements","publisher":"Fannie Mae"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"fannie-mae-value-acceptance-appraisal-alternatives"},{"title":"FCDO: LLM triage of written consular enquiries","useCases":["citizen-information-assistant"],"organization":{"name":"Foreign, Commonwealth and Development Office","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Caution Your Blast Ltd","role":"integrator"},{"name":"Microsoft (Azure OpenAI Service, GPT-4 and GPT-3.5 Turbo)","role":"model-provider"},{"name":"Kainos","role":"integrator"}],"summary":"British nationals abroad who write to the FCDO first see a triage tool that matches their question to approved guidance templates, such as how to replace a lost passport or what documents are needed to marry abroad. Personal data is removed with Azure services, a model splits the message into component questions, an embedding search shortlists templates and a second model picks the best template identifier, so the user never sees generated text. Users can still send the written enquiry to staff, and the FCDO says it will review each such case against the tool's answer; urgent cases are pointed to the consular contact centre. The record lists the tool's phase as beta or pilot.","stage":"pilot","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":76,"unit":"percent","qualifier":"at-least","period":"offline test on batches of historic anonymised enquiries (76% to 81%)","baseline":"templates chosen by FCDO staff","claimant":"organization","quote":"Depending on the batch, the tool provided the correct response for 76% to 81% of enquiries.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/fcdo-consular-digital-triage-written-enquiries-llm"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/fcdo-consular-digital-triage-written-enquiries-llm","title":"FCDO: Consular Digital Triage, Written Enquiries LLM","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2024-12-17"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"fcdo-consular-enquiry-triage"},{"title":"US Food and Drug Administration: FRED tool that proposes FOIA redactions","useCases":["freedom-of-information-request-processing"],"organization":{"name":"U.S. Food and Drug Administration, Center for Drug Evaluation and Research","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"U.S. Food and Drug Administration","role":"in-house"},{"name":"Contractor teams (not named)","role":"integrator"}],"summary":"FDA's Center for Drug Evaluation and Research uses the FOIA Redaction (FRED) tool, a generative AI system built by FDA and contractor teams, to help FOIA staff redact records more efficiently and consistently, because redaction is time consuming and FOIA backlogs build up. The tool returns a PDF with boxes around the text it recommends redacting, each with a comment giving the redaction code. Its data are completed FDA Form 483 inspection records in their original and staff redacted versions, and every output needs human review and approval. Listed as deployed since May 2025; no outcome figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"fda-foia-redaction-tool"},{"title":"US Food and Drug Administration: AI extraction of data from IND safety reports into FAERS","useCases":["adverse-event-case-intake"],"organization":{"name":"U.S. Food and Drug Administration, Center for Drug Evaluation and Research","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"ThinkTrends","role":"platform"}],"summary":"FDA's Center for Drug Evaluation and Research reports in the 2025 federal AI use case inventory that it uses OCR and AI, through the commercial tool ThinkTrends, to extract data from the Investigational New Drug safety reports that sponsors send in. The extracted data is converted to the E2B(R2) format and ingested automatically into the FDA Adverse Event Reporting System. The inventory describes manual extraction of these reports as labor intensive and time consuming for regulatory staff, and states the aim as faster processing and regulatory action on adverse events reported in clinical trials. It lists the use as deployed since March 2025.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"fda-ind-safety-report-extraction"},{"title":"US Food and Drug Administration: chatbot for adverse event and product problem reports on the Safety Reporting Portal","useCases":["adverse-event-case-intake"],"organization":{"name":"U.S. Food and Drug Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Druid","role":"platform"}],"summary":"FDA reports in the 2025 federal AI use case inventory that a conversational assistant built on the commercial platform Druid helps people who report an adverse event or product problem through the Safety Reporting Portal. It answers questions from a knowledge base, routes the reporter to the right form for the product type, helps complete it and submits the report to the portal through an API, with the stated aims of better data integrity and faster form completion. The inventory lists it as deployed since March 2024; no measured results are published.","stage":"production","year":2024,"channels":["web-chat","api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"fda-safety-reporting-portal-chatbot"},{"title":"FDIC: AI extraction of invoice and contract data for reconciliation","useCases":["supplier-invoice-processing"],"organization":{"name":"Federal Deposit Insurance Corporation","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The FDIC, the US bank deposit insurer and supervisor, is developing AI that extracts data from invoice and contract PDFs and reconciles them, emailing oversight managers a spreadsheet of discrepancies and errors. A separate initiative in its Division of Finance plans AI monitoring of invoices for proper submission and duplicate payments. Both are listed as in development or initiated in the 2024 federal inventory; no results are published.","stage":"announced","year":2024,"channels":["email","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"fdic-invoice-and-contract-data-extraction"},{"title":"FDIC: plain language policy drafting assistant, initiated then retired","useCases":["policy-drafting-and-gap-analysis"],"organization":{"name":"Federal Deposit Insurance Corporation","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The FDIC reported in its 2024 AI inventory a planned assistant for its policy writers: it would check a draft policy against the Plain Language Writing Act for clarity, active voice, concision, jargon and acronyms, and redraft selected sections in plain language. In the 2025 inventory the entry is listed as retired. It is a useful signal that style and consistency checking of internal policy is an early candidate, and that not every initiative reaches production.","stage":"paused","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI inventory (FDIC entry Plain Language Policy Assistant)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (FDIC entry 3, Plain Language Policy Assistant, retired)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"fdic-plain-language-policy-assistant"},{"title":"FDIC: AI monitoring of invoices and payments for abnormalities (planned)","useCases":["continuous-controls-testing"],"organization":{"name":"Federal Deposit Insurance Corporation","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The FDIC's Division of Finance reports a use case in pre deployment that uses classical machine learning to monitor financials and invoices for proper submittal, duplicate payments and other abnormalities, described as an automated assist for its auditing and monitoring work. For each gap it would show why it was flagged and possible causes, in visual tables with drill downs to contract numbers and agency sections. Not yet live; no results published.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry FDIC 39, CFOO Transactional Data Analysis)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"fdic-transactional-data-monitoring"},{"title":"Federal Bureau of Prisons: Pathfinder suggests career pathways to employees","useCases":["internal-talent-marketplace-matching"],"organization":{"name":"Federal Bureau of Prisons","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Microsoft Azure","role":"platform"}],"summary":"The Federal Bureau of Prisons, part of the US Department of Justice, runs Pathfinder to help its employees with career pathways. The system generates assessments from what the user enters, scores them and offers the employee options for career pathways. The inventory lists it as deployed since September 2023 on Azure. It is narrower than a full talent marketplace (no project or mentor matching is described) and no outcome figures are published.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DOJ-0159, Pathfinder)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"federal-bureau-of-prisons-pathfinder-career-pathways"},{"title":"Federal Reserve Board: AI use case inventory and high impact review","useCases":["ai-model-inventory"],"organization":{"name":"Board of Governors of the Federal Reserve System","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Federal Reserve Board runs a central AI Program that collects every AI use case in Board work and in functions delegated to the Reserve Banks, checks each against the Board's AI policy, screens it for high impact characteristics and routes it to the matching governance path. Use cases sit in a common repository that supports reporting and ongoing tracking and is validated periodically. The 2025 public inventory records, per use case, the stage, purpose, vendor, data used, personal data involvement and high impact designation.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.federalreserve.gov/publications/files/compliance-plan-for-omb-memorandum-m-25-21-202509.pdf","title":"Compliance Plan for OMB Memorandum M-25-21","publisher":"Board of Governors of the Federal Reserve System"},{"url":"https://www.federalreserve.gov/AI-use-case-inventory-2025.htm","title":"AI Use Case Inventory 2025","publisher":"Board of Governors of the Federal Reserve System"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"federal-reserve-board-ai-use-case-inventory"},{"title":"Federal Reserve Board: Consumer Complaints Explorer topic modelling","useCases":["complaints-root-cause-analysis"],"organization":{"name":"Board of Governors of the Federal Reserve System","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Federal Reserve Board's Division of Consumer and Community Affairs has used an in house natural language processing tool since 2019 to sort large volumes of consumer complaint narratives into topics, so staff can analyse and respond to them. For each narrative it outputs a topic number, a fit score and the top five terms of that topic. The input is complaint data from the CFPB. It is a central bank analysing consumer complaints about financial companies from the CFPB database rather than a firm analysing its own complaints, but the method is the same clustering step a bank's root cause work starts from.","stage":"production","year":2019,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.federalreserve.gov/AI-use-case-inventory-2025.htm","title":"AI Use Case Inventory 2025","publisher":"Board of Governors of the Federal Reserve System"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"federal-reserve-board-consumer-complaints-explorer"},{"title":"Federal Reserve Board: Comment Review System for public comments on proposed rules","useCases":["public-consultation-response-analysis"],"organization":{"name":"Board of Governors of the Federal Reserve System","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Federal Reserve Board processes public comments on rulemakings, information collections and other proposals in its Comment Review System. The system uses traditional natural language processing for summaries, matching comments to lists of topics, entity identification and similarity matching, and flags duplicate and near duplicate comment letters. The Board states that all public comments are still reviewed in their entirety and that summaries only assist the review. The inventory lists it as deployed since July 2021; no outcome figures are published.","stage":"production","year":2021,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"federal-reserve-board-public-comment-review-system"},{"title":"Federal Reserve Board: machine learning checks on regulatory report data","useCases":["regulatory-report-assembly"],"organization":{"name":"Board of Governors of the Federal Reserve System","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Federal Reserve Board's Division of Supervision and Regulation uses models developed in house to check the data that reporting firms submit. In its Regulatory Data Analysis use case, in operation since September 2024, analysts receive predicted values at several percentile levels for each reporter to compare with the values it actually reported. A related use case, Decision Tree for Deposits Data (still in implementation and assessment), calculates set variables and filters them to flag potential outliers in the current reporting period. These are supervisor side checks that mirror the validation a bank can run on its own returns before filing. No outcome figures are published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"federal-reserve-board-regulatory-data-analysis"},{"title":"Federal Student Aid: Aidan virtual assistant on StudentAid.gov","useCases":["benefits-eligibility-and-application-assistant"],"organization":{"name":"Federal Student Aid (U.S. Department of Education)","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"Federal Student Aid, the office of the U.S. Department of Education that runs federal student financial aid, operates Aidan, a virtual assistant on StudentAid.gov that uses natural language processing to answer common financial aid questions and help customers find information about their own federal aid. The agency reports it in its AI use case inventory with usage figures for its first two years.","stage":"scaled","year":2025,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"users-served","value":2600000,"unit":"count","qualifier":"at-least","period":"unique customers in just over two years","claimant":"organization","quote":"In just over two years, Aidan has interacted with over 2.6 million unique customers, resulting in more than 11 million user messages.","sourceUrl":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv"},{"kpi":"interactions-handled","value":11000000,"unit":"count","qualifier":"at-least","period":"user messages in just over two years","claimant":"organization","quote":"In just over two years, Aidan has interacted with over 2.6 million unique customers, resulting in more than 11 million user messages.","sourceUrl":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv"}],"outcomeDisclosed":true,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"federal-student-aid-aidan-virtual-assistant"},{"title":"FEMA: machine translation of disaster survivors' documents for Individual Assistance","useCases":["public-service-translation","benefits-eligibility-and-application-assistant"],"organization":{"name":"Federal Emergency Management Agency","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"FEMA plans to translate the full text of non English documents that disaster survivors submit with their Individual Assistance applications, instead of relying on a contractor's summary of each document. The agency expects faster case processing and a drop in cost from about USD 40 per document to pennies. Original and translation will both be stored in the survivor's file, as substantiating documents that support assistance determinations.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"fema-individual-assistance-document-translation"},{"title":"FEMA: generative AI over spend plan data to answer data calls and information requests (planned)","useCases":["supervisory-exam-response-assembly"],"organization":{"name":"Federal Emergency Management Agency","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Microsoft","role":"model-provider"}],"summary":"FEMA reports a pre deployment tool that lets staff ask questions of spend plan and actual execution data in common language, using Azure OpenAI inside the agency's system boundary, so they can give rapid responses to data calls and requests for information and show leadership where budget was planned and spent. It illustrates the evidence retrieval step of answering an oversight request from governed internal data. Not yet live; no results published.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DHS-2709, Spend Plan Analysis GPT)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://www.dhs.gov/ai/use-case-inventory/fema","title":"DHS AI Use Case Inventory, FEMA (entry DHS-2709, Spend Plan Analysis GPT)","publisher":"U.S. Department of Homeland Security"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"fema-spend-plan-data-call-assistant"},{"title":"US Federal Housing Finance Agency: automated triage of user reported phishing emails","useCases":["security-alert-triage-and-investigation"],"organization":{"name":"Federal Housing Finance Agency","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"KnowBe4","role":"platform"}],"summary":"FHFA reports in the 2025 federal AI use case inventory that it uses KnowBe4 PhishER to manage the suspicious emails its staff report. Machine learning classifies each email as spam, phishing or malicious and sends the response automatically, so a cybersecurity specialist does not have to analyse hundreds of reported emails by hand. It is a machine learning classifier rather than a generative AI agent, so it covers the classification step of this use case only. The inventory lists the use case as deployed since April 2021 with risk reduction as its benefit, and says the tool reduces the time to respond to staff, but gives no figures.","stage":"production","year":2021,"channels":["email","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"fhfa-phishing-email-identification"},{"title":"Figure: AI chatbots for home equity lending","useCases":["conversational-loan-application-intake","home-loan-assistant-and-prequalification"],"organization":{"name":"Figure","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"model-provider"}],"summary":"Figure, a US fintech that offers home equity lines of credit, uses Gemini models to run chatbots that simplify and speed up the lending experience for consumers and for its own staff. No outcome figures were published.","stage":"production","year":2025,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"figure-lending-chatbots"},{"title":"Financial Conduct Authority: machine readable Intelligent Handbook","useCases":["regulatory-horizon-scanning"],"organization":{"name":"Financial Conduct Authority","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Corlytics","role":"platform"}],"summary":"The UK Financial Conduct Authority worked with Corlytics to turn its Handbook into a searchable, machine readable rulebook. A pilot that started in September 2016 added taxonomy tagging with a four eyes approval workflow. Corlytics then developed with the FCA a machine learning framework for auto tagging and classifying content, which the vendor describes as using a rules based approach, with users able to review, approve or reject tags under a full audit trail. After the pilot, Corlytics delivered taxonomy tagging for the remaining Handbook provisions; it describes the Handbook as over 18,000 provisions. The system went live on 10 May 2017, and according to the vendor it made regulatory obligations much easier to identify. It shows the regulator side of obligation mapping. No outcome figure was published.","stage":"production","year":2017,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.corlytics.com/case_studies/how-can-we-ensure-that-our-handbook-is-digitised-machine-readable-searchable-for-our-users/","title":"How can we ensure that our handbook is digitised, machine-readable & searchable for our users?","publisher":"Corlytics","date":"2024-09-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"financial-conduct-authority-intelligent-handbook"},{"title":"Financial Conduct Authority: synthetic data in the Digital Sandbox and the Synthetic Data Expert Group","useCases":["synthetic-test-data-generation"],"organization":{"name":"Financial Conduct Authority","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[],"summary":"The UK Financial Conduct Authority has built synthetic datasets since a 2020 DataSprint, gave participants in two Digital Sandbox pilots access to synthetic data, opened a permanent Digital Sandbox in August 2023 and released an authorised push payment fraud synthetic dataset in September 2023, so firms can build and test solutions without real customer data. Its Synthetic Data Expert Group published a March 2024 report with use cases on system testing, model validation and data sharing, including the trade offs between privacy, utility and fidelity.","stage":"production","year":2023,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"users-served","value":28,"unit":"count","qualifier":"exact","period":"First Digital Sandbox pilot, organisations given access to synthetic data","claimant":"organization","quote":"The first pilot, involving 28 organisations, underscored the value of synthetic data, emphasising the need for more referentially linked datasets and finer granularity.","sourceUrl":"https://www.fca.org.uk/publication/corporate/report-using-synthetic-data-in-financial-services.pdf"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.fca.org.uk/publication/corporate/report-using-synthetic-data-in-financial-services.pdf","title":"Using Synthetic Data in Financial Services","publisher":"Financial Conduct Authority, Synthetic Data Expert Group"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"financial-conduct-authority-synthetic-data-sandbox"},{"title":"Financial Times: Ask FT, answers for FT Professional subscribers from FT journalism","useCases":["publisher-archive-answer-engine"],"organization":{"name":"Financial Times","anonymized":false,"country":"GB","region":"europe","industry":"media-and-entertainment"},"vendors":[{"name":"Anthropic","role":"model-provider"}],"summary":"Ask FT is the Financial Times' first customer facing generative AI product: a search tool that answers questions using FT content only, with references to the articles used. It was tested internally by editorial and product teams, piloted with a few hundred FT Professional subscribers, then offered to larger client accounts, and became available to all FT Professional customers in April 2025. Users, mostly in finance, consulting and law, use it to prepare meetings and reports and tend to open the cited sources. The FT says the links to full articles encourage deeper reading that supports renewal of key accounts, and that the tool calls out when it lacks sufficient information for a credible answer.","stage":"production","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.ftstrategies.com/en-gb/insights/how-ask-ft-is-meeting-user-needs-one-year-in-with-our-generative-ai-feature","title":"Ask FT: Your direct route to insight","publisher":"FT Strategies"},{"url":"https://www.ftstrategies.com/en-gb/insights/from-experiment-to-impact-what-results-can-generative-ai-products-deliver-for-publishers","title":"From experiment to impact: What results can generative AI products deliver for publishers?","publisher":"FT Strategies"},{"url":"https://voicebot.ai/2024/03/25/financial-times-launches-generative-ai-chatbot-for-subscribers-powered-by-anthropics-claude-3-llm/","title":"Financial Times Launches Generative AI Chatbot for Subscribers Powered by Anthropic's Claude 3 LLM","publisher":"Voicebot.ai","date":"2024-03-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"financial-times-ask-ft"},{"title":"Finshark: AI assisted financial crime investigations and reporting with Lucinity's Luci agent","useCases":["suspicious-activity-report-drafting"],"organization":{"name":"Finshark","anonymized":false,"country":"SE","region":"europe","industry":"payments"},"vendors":[{"name":"Lucinity","role":"platform"}],"summary":"Swedish open banking and instant payments company Finshark uses Lucinity's case manager with the Luci AI copilot, which adds case summaries, report writing and customer research to investigations, together with Lucinity's regulatory reporting module. The case study says the administrative manual work in case investigations has been significantly reduced but gives no figures.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en","sv"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://lucinity.com/casestudy-finshark","title":"Lucinity and Finshark Case Study","publisher":"Lucinity"},{"url":"https://lucinity.com/blog/finshark-enhances-financial-crime-prevention-with-lucinity-ai-powered-platform","title":"Finshark Enhances Financial Crime Prevention with Lucinity's AI-Powered Platform","publisher":"Lucinity","date":"2024-10-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"finshark-lucinity-luci-investigations"},{"title":"First National Bank: Copilot for Sales for commercial bankers","useCases":["email-and-ticket-reply-drafting"],"organization":{"name":"First National Bank","anonymized":false,"country":"ZA","region":"africa","industry":"banking"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"First National Bank, a division of FirstRand, rolled out Microsoft Copilot for Sales to its bankers in December 2023, on top of a Dynamics 365 CRM that nearly all of its more than 3,500 bankers already used. The customer story focuses on drafting replies to commercial clients in Outlook that address each point in the client's message, which the banker reviews, edits and sends. Because bankers worried the drafts would sound robotic, the bank paired the rollout with a training programme with modules for each type of task.","stage":"scaled","year":2023,"channels":["email","internal-tools"],"languages":["en"],"metrics":[{"kpi":"employee-adoption","value":94,"unit":"percent","qualifier":"at-least","period":"commercial bankers","claimant":"vendor","quote":"More than 94% of commercial bankers now use Copilot to help them craft richer communications with their customers.","sourceUrl":"https://www.microsoft.com/en/customers/story/1761931588230983875-first-national-bank-dynamics-365-sales-banking-and-capital-markets-en-south-Africa"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1761931588230983875-first-national-bank-dynamics-365-sales-banking-and-capital-markets-en-south-Africa","title":"First National Bank enhances customer communications with Microsoft Copilot for Sales","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"first-national-bank-copilot-for-sales"},{"title":"FIS with Anthropic: Financial Crimes AI Agent in development at BMO and Amalgamated Bank","useCases":["aml-alert-triage","suspicious-activity-report-drafting"],"organization":{"name":"BMO and Amalgamated Bank","anonymized":false,"region":"north-america","industry":"banking"},"vendors":[{"name":"FIS","role":"platform"},{"name":"Anthropic","role":"model-provider"}],"summary":"FIS announced in May 2026 that it is building a Financial Crimes AI Agent with Anthropic that assembles evidence across a bank's core systems for anti money laundering alert and case investigations and supports suspicious activity report narratives. BMO and Amalgamated Bank are developing with the agent, and FIS plans general availability in the second half of 2026. The release states aims for investigation time and narrative quality but no measured results.","stage":"announced","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.fisglobal.com/about-us/media-room/press-release/2026/fis-brings-agentic-ai-to-banking-with-anthropic-starting-with-financial-crimes","title":"FIS Brings Agentic AI to Banking with Anthropic, Starting with Financial Crimes","publisher":"FIS","date":"2026-05-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"fis-financial-crimes-ai-agent"},{"title":"FNBO: agentic AI for enhanced due diligence and sanctions alerts with Nasdaq Verafin","useCases":["perpetual-kyc","sanctions-screening-adjudication"],"organization":{"name":"First National Bank of Omaha (FNBO)","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Nasdaq Verafin","role":"platform"}],"summary":"FNBO deployed Nasdaq Verafin's Agentic EDD Analyst and Agentic Sanctions Analyst, which remove manual information gathering across multiple systems for enhanced due diligence cases and sanctions alerts. The vendor reports that the bank spent 50% less time on these reviews and alerts and redirected investigator capacity to deeper analysis.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":50,"unit":"percent","qualifier":"exact","period":"per case, enhanced due diligence and sanctions alert reviews","claimant":"independent","quote":"At First National Bank of Omaha, AI agents have begun taking on some of the work of human financial crime investigators, reducing the time that people spend on each case by 50%, according to bank executives.","sourceUrl":"https://www.americanbanker.com/news/how-fnbo-uses-agentic-ai-to-investigate-financial-crime"}],"outcomeDisclosed":true,"sources":[{"url":"https://verafin.com/resource/fnbo-seizes-the-agentic-ai-advantage/","title":"FNBO Seizes the Agentic AI Advantage","publisher":"Nasdaq Verafin"},{"url":"https://www.americanbanker.com/news/how-fnbo-uses-agentic-ai-to-investigate-financial-crime","title":"How FNBO uses agentic AI to investigate financial crime","publisher":"American Banker","date":"2026-06-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"fnbo-verafin-agentic-edd-and-sanctions"},{"title":"Food Standards Agency: machine learning to help local authorities prioritise food hygiene inspections","useCases":["inspection-prioritization"],"organization":{"name":"Food Standards Agency","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Cognizant","role":"integrator"}],"summary":"After the pandemic, the number of food businesses awaiting their first hygiene inspection in England, Wales and Northern Ireland grew steadily. The Food Standards Agency built a LightGBM model, trained on its hygiene rating data, census data and open location data, that predicts whether a business awaiting inspection is likely to be compliant and what rating it would get, and offers the predictions to local authority officers as a table, a map and a download. Use is voluntary, the prediction must not replace or be used in isolation from the officer's judgment, and the agency applied fairness and explainability tooling during development. The transparency record describes an alpha pilot with local authorities from April 2022; no outcome figures are published. GOV.UK now lists the record's phase as Retired, and no pilot outcomes were ever published.","stage":"paused","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/food-standards-agency-food-hygiene-rating-scheme-ai","title":"Food Standards Agency: Food Hygiene Rating Scheme – AI (algorithmic transparency record)","publisher":"GOV.UK","date":"2024-02-29"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"food-standards-agency-food-hygiene-inspection-prioritisation"},{"title":"Foyer: AI analysis of damage photos for minor motor claims","useCases":["photo-based-damage-assessment"],"organization":{"name":"Foyer","anonymized":false,"country":"LU","region":"europe","industry":"insurance"},"vendors":[{"name":"Tractable","role":"platform"}],"summary":"Foyer, a Luxembourg insurer, announced in May 2026 that it has formalised a partnership with Tractable to identify damage automatically from photographs for minor motor incidents. The aim is immediate analysis of the damage the policyholder reports, faster handling of the simplest claims and fewer visits to the garage. No outcome figures were published.","stage":"announced","year":2026,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://tractable.ai/foyer-and-tractable/","title":"Foyer and Tractable: AI for motor claims management","publisher":"Tractable","date":"2026-05-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"foyer-ai-motor-claims-photos"},{"title":"Freshfields: Dynamic Due Diligence, a proprietary AI tool for legal reviews","useCases":["deal-sourcing-and-due-diligence-assistant"],"organization":{"name":"Freshfields","anonymized":false,"country":"GB","region":"europe","industry":"professional-services"},"vendors":[{"name":"Google Cloud","role":"model-provider"},{"name":"Freshfields Lab","role":"in-house"}],"summary":"Freshfields, a global law firm, built Dynamic Due Diligence (D3), a proprietary tool designed to enhance legal reviews and due diligence, and in 2025 announced that Google's Gemini models would power it. A year into the collaboration the firm reported that D3 is one of several bespoke Freshfields Lab platforms now running on Gemini, and a partner who co leads Freshfields Lab said teams and clients use Gemini daily across those platforms. The firm has not published a separate outcome for D3.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.freshfields.com/en/our-thinking/news/news-search/2025/04/freshfields-and-google-cloud-accelerate-legal-innovation-through-strategic-ai-collaboration2","title":"Freshfields and Google Cloud Accelerate Legal Innovation Through Strategic AI Collaboration","publisher":"Freshfields","date":"2025-04-08"},{"url":"https://www.freshfields.com/en/our-thinking/news/news-search/2026/04/freshfields-reports-google-cloud-collaboration-delivering-transformation-at-scale","title":"Freshfields Reports Google Cloud Collaboration Delivering Transformation at Scale","publisher":"Freshfields","date":"2026-04-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"freshfields-dynamic-due-diligence"},{"title":"FTC: AI classification and duplicate grouping of consumer fraud complaints","useCases":["complaints-root-cause-analysis"],"organization":{"name":"Federal Trade Commission","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Leidos","role":"integrator"}],"summary":"Since 2019 the US Federal Trade Commission has used AI on the complaints it receives through ReportFraud and other channels: one model classifies uncategorised complaints by product and service code, another groups duplicate complaints about the same issue so investigators can see which entities attract multiple reports, alongside graph analytics that connect complaints about the same company when it uses different names, phone numbers or aliases. It shows the two steps that make complaint themes countable: consistent categorisation and deduplication. No outcome figures are published.","stage":"production","year":2019,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entries FTC-0001, FTC-0002 and FTC-0005)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ftc-consumer-complaint-classification-and-grouping"},{"title":"Federal Trade Commission: active learning to prioritise documents in investigations and litigation","useCases":["ediscovery-and-disclosure-document-review"],"organization":{"name":"Federal Trade Commission","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Relativity","role":"platform"}],"summary":"The Federal Trade Commission's Bureau of Competition, Bureau of Consumer Protection and Office of the General Counsel use Relativity Active Learning in eDiscovery for consumer protection and competition investigations and litigation. The inventory entry gives the problem as manual document coding being very time consuming during legal review and the output as predicted pertinent documents. It lists the use as deployed since 2020 and not high impact. No outcome figures are published.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry FTC-0008, Relativity Active Learning)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ftc-relativity-active-learning-review"},{"title":"Galt Police Department: AI triage of non emergency calls and assistive 911 call taking","useCases":["non-emergency-service-request-routing","emergency-call-triage-support"],"organization":{"name":"Galt Police Department","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Prepared","role":"platform"}],"summary":"Galt Police Department serves 26,000 residents with eight dispatch staff, who handle nearly 30,000 calls a year, more than 73% of them non emergency. Since 2024 an AI agent from Prepared answers the ten digit non emergency line, works out what the caller needs, resolves it or routes it to the right resource, and transfers any genuine emergency immediately. On 911 calls, dispatchers get a live transcript, an AI summary and key details highlighted as the call happens. During a shooting in Galt, the agent handled the incoming non emergency calls in the background while dispatchers managed the response.","stage":"production","year":2024,"channels":["voice","agent-desktop"],"languages":["en"],"metrics":[{"kpi":"contact-deflection","value":73,"unit":"percent","qualifier":"exact","period":"share of call volume handled before reaching a dispatcher","claimant":"vendor","quote":"With 73% of call volume now handled before it reaches a dispatcher's headset, the calls that do come through are the ones that genuinely need a human.","sourceUrl":"https://www.prepared911.com/case-studies/galt-pd-assistive-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.prepared911.com/case-studies/galt-pd-assistive-ai","title":"Galt PD: Making the Unmanageable Manageable with Assistive AI","publisher":"Prepared"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"galt-police-department-non-emergency-call-triage"},{"title":"City of Amsterdam: planned generative AI summaries of legal advice on objections for the city's lawyers","useCases":["court-and-case-file-summarization"],"organization":{"name":"Gemeente Amsterdam","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[{"name":"In house (Gemeente Amsterdam)","role":"in-house"},{"name":"OpenAI (used via Microsoft Azure)","role":"model-provider"}],"summary":"The City of Amsterdam's legal department keeps an internal case library of its advice on objections (bezwaren) against municipal decisions. The city has registered a tool, built in house, that would use an OpenAI GPT model (gpt-3.5-turbo or gpt-4) through Azure to write a summary of the core of each existing advice, shown first in the library, so lawyers handling a new objection can judge more quickly whether an earlier advice is relevant. Users can report errors in a summary, which are then corrected; the register says summaries will be marked as made with generative AI and checked by sampling. The register entry is marked \"In gebruik\" (in use) with a start date of July 2022, but its method section says the algorithm still has to be developed and that the prompting approach is yet to be decided, so the tool is recorded here as announced. No outcome figures are published.","stage":"announced","year":2026,"channels":["internal-tools"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/gm0363/47842380/samenvatten-van-juridische-bezwaaradviezen","title":"Samenvatten van juridische bezwaaradviezen, Algoritmeregister","publisher":"Algoritmeregister van de Nederlandse overheid","date":"2026-07-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"gemeente-amsterdam-objection-advice-summaries"},{"title":"Gemeente Nissewaard: rules based eligibility check in online benefit applications","useCases":["benefits-eligibility-and-application-assistant"],"organization":{"name":"Gemeente Nissewaard","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[{"name":"Centric Netherlands BV","role":"platform"}],"summary":"Nissewaard runs an online application service for social assistance (bijstand), special assistance, support for the self employed and minimum income schemes. During the application a decision tree checks the data read in and the applicant's answers against the legal criteria and shows the outcome to the applicant; caseworkers can overrule it. The tool, supplied by Centric, has been in use since March 2017, and the register entry names the risk that applicants give up because it sets wrong expectations about the outcome. Entries for other municipalities say the same service is used by about 50 Dutch municipalities.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/gm1930/33327482/sociaal-domein-ediensten-voor-aanvragen","title":"Sociaal Domein: eDiensten voor aanvragen, Gemeente Nissewaard","publisher":"Algoritmeregister van de Nederlandse overheid","date":"2026-09-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"gemeente-nissewaard-benefit-application-eligibility-check"},{"title":"Gemeente Rotterdam: welfare reassessment risk model, stopped in 2022","useCases":["benefit-fraud-and-error-detection"],"organization":{"name":"Gemeente Rotterdam","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[{"name":"Accenture","role":"integrator"}],"summary":"From 2017 the City of Rotterdam used a machine learning model that gave each social assistance recipient a risk score between 0 and 1 for receiving benefits they were not, or no longer, entitled to, based on the outcomes of earlier eligibility reviews. High scores were one route to an invitation for a review interview, and an income consultant decided the outcome. The city classified the model as high risk and stopped using it in early 2022; its register entry states that a review found it was not currently possible to build a risk model that fits the city's policy. The register says the model processed no nationality, age or health data; journalists who obtained the model file reported 315 inputs, including age, gender and language skills, and found that it discriminated by ethnicity, age, gender and parenthood.","stage":"paused","year":2022,"channels":["internal-tools"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/gm0599/36585638/heronderzoeken-uitkeringsgerechtigden","title":"Heronderzoeken Uitkeringsgerechtigden, Algoritmeregister","publisher":"Algoritmeregister van de Nederlandse overheid"},{"url":"https://www.lighthousereports.com/investigation/suspicion-machines/","title":"Suspicion Machines","publisher":"Lighthouse Reports"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"gemeente-rotterdam-welfare-reassessment-risk-model"},{"title":"Gemeente Tilburg: Vragen.AI answers on the municipal website","useCases":["citizen-information-assistant"],"organization":{"name":"Gemeente Tilburg","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[{"name":"Swis (Vragen.ai)","role":"platform"}],"summary":"The municipality of Tilburg uses Vragen.AI, an AI search and answer function supplied by Swis, on its websites such as Tilburg Helpt, registered in the Dutch national algorithm register with a start date of September 2026. Answers come only from information already on the connected sites and always cite their source; personal data that residents type by mistake is detected and removed, and questions are not stored to help returning visitors or passed on to the AI companies. Staff do not watch conversations live but review the answers to improve the website. The register names the spreading of wrong information as a major risk. Several other Dutch municipalities (for example Doetinchem and Wijchen) register similar website chatbots grounded in their own pages.","stage":"production","year":2026,"channels":["web-chat"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/gm0855/44521279/vragenai","title":"Vragen.AI, Gemeente Tilburg","publisher":"Algoritmeregister van de Nederlandse overheid","date":"2026-09-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"gemeente-tilburg-vragen-ai"},{"title":"Generali France: automated, transparent pricing model building with Akur8","useCases":["insurance-pricing-and-actuarial-copilot"],"organization":{"name":"Generali France","anonymized":false,"country":"FR","region":"europe","industry":"insurance"},"vendors":[{"name":"Akur8","role":"platform"}],"summary":"Generali France's actuarial studies team uses Akur8, a pricing platform that automates the repetitive parts of building risk models while keeping the process transparent and auditable for actuaries. Its actuarial studies manager says modelling is five times faster and that the shared interface improved communication inside the team.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["fr"],"metrics":[{"kpi":"productivity-gain","value":5,"unit":"multiplier","qualifier":"exact","period":"speed of pricing model building","claimant":"organization","quote":"Modeling speed is 5x faster, while keeping a thoroughly transparent and auditable process.","sourceUrl":"https://www.akur8.com/resources/testimonials"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.akur8.com/resources/testimonials","title":"Best actuarial software: Discover Akur8 Customer Reviews","publisher":"Akur8"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"generali-france-akur8-pricing-models"},{"title":"Generali Global Corporate & Commercial: AI risk insights in the cyber underwriting workflow","useCases":["underwriting-risk-assessment-copilot","commercial-underwriting-submission-triage"],"organization":{"name":"Generali Global Corporate & Commercial","anonymized":false,"country":"IT","region":"global","industry":"insurance"},"vendors":[{"name":"Sixfold","role":"platform"}],"summary":"As its cyber book grew, Generali GC&C chose Sixfold as its first external AI partner and connected it to its cyber data sources and scoring system, so that all available risk information is structured in a dashboard for the underwriter against Generali's own guidelines. The vendor reports that over 90% of underwriters adopted the platform, that most cyber submissions now go through it with turnaround times for distribution cut by 50%, and that risk engineering reports take a few hours instead of about two days. Its Global Head of Operations and IT, Matthew Richardson, is quoted as saying that Sixfold's input is now required for every quote.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":50,"unit":"percent","qualifier":"exact","period":"turnaround for distribution channels on cyber submissions","claimant":"vendor","quote":"Today, most Cyber submissions are accelerated through the Sixfold solution, cutting turnaround times for our distribution channels by 50%.","sourceUrl":"https://www.sixfold.ai/case-study/generali-gc-c"},{"kpi":"employee-adoption","value":90,"unit":"percent","qualifier":"at-least","period":"cyber underwriters","claimant":"vendor","quote":"The results were immediate: over 90% of underwriters actively adopted the platform, reporting consistently high accuracy scores.","sourceUrl":"https://www.sixfold.ai/case-study/generali-gc-c"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sixfold.ai/case-study/generali-gc-c","title":"Generali GC&C | Sixfold Case Study","publisher":"Sixfold"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"generali-gcc-sixfold-cyber-underwriting"},{"title":"Georgia-Pacific: ChatGP, a generative AI assistant for machine operators that combines documents with live machine data","useCases":["plant-operator-and-maintenance-copilot"],"organization":{"name":"Georgia-Pacific","anonymized":false,"country":"US","region":"north-america","industry":"manufacturing"},"vendors":[{"name":"Amazon Web Services","role":"platform"},{"name":"Anthropic","role":"model-provider"}],"summary":"Georgia-Pacific built ChatGP with AWS Professional Services on Amazon Bedrock, using Anthropic's Claude, to give junior operators and maintenance technicians one place to ask about their machines. It answers from documents, maintenance records and Internet of Things sensor data streamed through Amazon Kinesis, so a question about a machine issue can draw on the machine's current state and recent trends and return step by step guidance tailored to that facility. The company also records conversations with experienced or retired experts and has a language model turn them into procedure documents. AWS reports that ChatGP reduced off quality production and machine downtime, and Georgia-Pacific planned to extend it to more facilities by the end of 2024.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://aws.amazon.com/solutions/case-studies/georgia-pacific-optimizes-operator-efficiency-case-study/","title":"Georgia-Pacific Optimizes Operator Efficiency Using Generative AI on AWS","publisher":"Amazon Web Services"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"georgia-pacific-chatgp-operator-assistant"},{"title":"Georgia-Pacific: predictive analytics on streamed equipment data to predict equipment failure 60 to 90 days ahead","useCases":["industrial-asset-predictive-maintenance"],"organization":{"name":"Georgia-Pacific","anonymized":false,"country":"US","region":"north-america","industry":"manufacturing"},"vendors":[{"name":"Amazon Web Services","role":"platform"}],"summary":"Georgia-Pacific, a pulp, paper and building products manufacturer, streams data from equipment at its North American facilities into an operations data lake on AWS and analyses it with an AWS based advanced analytics solution that includes Amazon SageMaker machine learning models. For selected assets the company can now predict equipment failure 60 to 90 days in advance, so it can plan equipment downtime instead of suffering unscheduled production stoppages. No outcome figure is published for the failure prediction; the same data platform also runs process optimization models for converting line speeds, which are outside this use case.","stage":"production","year":2019,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://aws.amazon.com/solutions/case-studies/georgia-pacific/","title":"Georgia-Pacific Optimizes Processes, Saves Millions of Dollars Yearly Using AWS","publisher":"Amazon Web Services"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"georgia-pacific-predictive-asset-analytics"},{"title":"Georgia State University: Pounce chatbot for incoming students and summer melt","useCases":["student-enrollment-and-services-assistant"],"organization":{"name":"Georgia State University","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[{"name":"AdmitHub (now Mainstay)","role":"platform"}],"summary":"Georgia State University combined a new student portal, which guides incoming students through the steps needed before the first day of classes (such as financial aid documents, immunization records, placement exams and class registration), with \"Pounce\", an AI enhanced chatbot that answers their questions around the clock by text message. The assistant vice president of undergraduate admissions said every interaction was tailored to the specific student's enrollment task. In its first summer (2016) Pounce delivered more than 200,000 answers and the university, with the portal and the chatbot together, reduced summer melt by 22 percent, an additional 324 students in class on the first day. Separately, in a randomized control trial the university saw a four percent overall decrease in the share of confirmed freshmen who did not enroll, and it says those gains came from the students who had access to Pounce. The same executive said the university would otherwise have needed 10 more full time staff to handle the volume of messaging.","stage":"scaled","year":2016,"channels":["sms"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":200000,"unit":"count","qualifier":"at-least","period":"first summer of implementation, 2016","claimant":"organization","quote":"In 2016, during the first summer of implementation, Pounce delivered more than 200,000 answers to questions asked by incoming freshmen, and the university reduced summer melt by 22 percent.","sourceUrl":"https://success.gsu.edu/reduction-of-summer-melt/"}],"outcomeDisclosed":true,"sources":[{"url":"https://success.gsu.edu/reduction-of-summer-melt/","title":"Reduction of Summer Melt","publisher":"Georgia State University","archivedUrl":"https://web.archive.org/web/20260514132005/https://success.gsu.edu/reduction-of-summer-melt/"},{"url":"https://mainstay.com/about/","title":"Mainstay: Our Story, Mission, Values, and Team","publisher":"Mainstay"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"georgia-state-university-pounce-enrollment-chatbot"},{"title":"General Insurance Association of Singapore: shared AI fraud analytics on travel and motor claims","useCases":["claims-fraud-detection","travel-insurance-claims-and-assistance-agent"],"organization":{"name":"General Insurance Association of Singapore","anonymized":false,"country":"SG","region":"asia-pacific","industry":"insurance"},"vendors":[{"name":"Shift Technology","role":"platform"}],"summary":"Member insurers of the General Insurance Association of Singapore analyse travel and motor claims together with Shift Technology's AI, which finds connections between people, providers and claims that look genuine when each insurer sees them alone. The data analytics initiative started in 2017 with 25 insurers; fraud alerts are issued to members and prompt joint investigations. The public page gives no outcome figures.","stage":"scaled","year":2017,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.shift-technology.com/resources/case-studies/power-of-the-collective-singapore-insurers-unite-to-fight-fraud","title":"Power of the collective: Singapore insurers unite to fight fraud","publisher":"Shift Technology","date":"2022-09-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"gia-singapore-industry-fraud-analytics"},{"title":"Ginnie Mae: machine learning to find exceptions in subledger transaction data","useCases":["ledger-and-payment-reconciliation"],"organization":{"name":"Ginnie Mae","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Ernst & Young","role":"integrator"}],"summary":"Ginnie Mae, part of the US Department of Housing and Urban Development, analyses the transaction data of its master subservicers every month. Since April 2021 it has used machine learning models, built in house and with Ernst & Young as contractor, to detect anomalies, data inconsistencies and exceptions in that data. It says early detection reduces manual adjustments to financial reporting, which saves cost and time. The 2025 federal inventory still lists the system as deployed. No outcome figures are published.","stage":"production","year":2021,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ginnie-mae-subledger-data-quality-machine-learning"},{"title":"GitHub Sponsors: AI generated chargeback evidence with Stripe Smart Disputes","useCases":["chargeback-and-representment"],"organization":{"name":"GitHub","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Stripe","role":"platform"}],"summary":"GitHub Sponsors, the platform through which people fund open source maintainers, uses Stripe Smart Disputes, which generates and submits evidence to contest chargebacks automatically. Before, the team reviewed disputes by hand and rarely contested them because gathering evidence took too long. The Stripe case study reports that the team now saves four to five hours of work a week and headlines an average reduction of 20 hours a month in time spent on disputes.","stage":"production","year":2025,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"hours-saved","value":20,"unit":"hours","qualifier":"exact","period":"per month, on average","claimant":"vendor","quote":"Time spent addressing disputes reduced by 20 hours per month, on average","sourceUrl":"https://stripe.com/gb/customers/github"}],"outcomeDisclosed":true,"sources":[{"url":"https://stripe.com/gb/customers/github","title":"Github case study | Stripe","publisher":"Stripe"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"github-sponsors-stripe-smart-disputes"},{"title":"GoHealth: AI role play training for licensed benefits consultants","useCases":["conversation-roleplay-training"],"organization":{"name":"GoHealth","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Second Nature","role":"platform"}],"summary":"GoHealth, a health insurance marketplace focused on Medicare, uses AI role play partners so licensed benefits consultants practise sales and compliance heavy conversations, both in onboarding and in ongoing training. For new hires, practice is built into the training weeks instead of a separate three week block of practice calls. After a pilot in the fourth quarter of 2022 GoHealth signed a long term contract. The vendor reports shorter onboarding, a higher sales conversion rate in the pilot and a higher trainee to trainer ratio.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"time-to-proficiency-reduction","value":55,"unit":"percent","qualifier":"exact","baseline":"onboarding cut from nine weeks to four weeks","claimant":"vendor","quote":"That’s a 55% time saving that allows new hires to start making effective calls five weeks earlier than before.","sourceUrl":"https://secondnature.ai/resources/gohealth-boosts-productivity-and-cuts-onboarding-time-with-second-nature/"},{"kpi":"conversion-rate-uplift","value":21,"unit":"percent","qualifier":"exact","period":"Q4 2022 pilot, sales rates 10 days before versus 10 days after about 34 minutes of AI practice","baseline":"sales rates in the 10 days before the pilot training","claimant":"vendor","quote":"Average 21% increase in sales conversions after 34 minutes of practice","sourceUrl":"https://secondnature.ai/resources/gohealth-boosts-productivity-and-cuts-onboarding-time-with-second-nature/"}],"outcomeDisclosed":true,"sources":[{"url":"https://secondnature.ai/resources/gohealth-boosts-productivity-and-cuts-onboarding-time-with-second-nature/","title":"GoHealth Cuts Onboarding 55% with AI Sales Training","publisher":"Second Nature"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"gohealth-ai-roleplay-sales-training"},{"title":"Gojob: Aglae assistant that prequalifies temporary work candidates by text message","useCases":["recruitment-screening-and-interview-scheduling"],"organization":{"name":"Gojob","anonymized":false,"country":"FR","region":"europe","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Gojob, a digital temporary employment agency, built Aglae on Azure OpenAI Service. When a job is published, Aglae searches its pool of 2 million profiles, holds text message conversations with candidates to prequalify them, answers their questions and redirects candidates who do not match; the recruiter steps in at the end. Staff check the conversations afterwards to catch bias, and the assistant hands over to a person when needed. Microsoft reports 1.5 million exchanges and a placement rate that rose from 60% to 95%.","stage":"scaled","year":2024,"channels":["sms"],"languages":[],"metrics":[{"kpi":"interactions-handled","value":1500000,"unit":"count","qualifier":"exact","period":"since go live","claimant":"vendor","quote":"1.5 million exchanges have taken place since the virtual assistant went live, and it takes Aglae no more than 15 minutes to close a conversation and pre-qualify the best candidate, compared with 30 days in classic employment agencies, and an average of five days for other temporary work agencies.","sourceUrl":"https://www.microsoft.com/en/customers/story/20838-gojob-azure-open-ai-service"},{"kpi":"cycle-time-days","value":15,"unit":"minutes","qualifier":"up-to","baseline":"an average of five days at other temporary work agencies (an industry comparison, not Gojob's own before)","claimant":"vendor","quote":"1.5 million exchanges have taken place since the virtual assistant went live, and it takes Aglae no more than 15 minutes to close a conversation and pre-qualify the best candidate, compared with 30 days in classic employment agencies, and an average of five days for other temporary work agencies.","sourceUrl":"https://www.microsoft.com/en/customers/story/20838-gojob-azure-open-ai-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/20838-gojob-azure-open-ai-service","title":"Gojob augments recruiters with Azure OpenAI Service","publisher":"Microsoft Customer Stories","archivedUrl":"https://web.archive.org/web/20250325212111/https://www.microsoft.com/en/customers/story/20838-gojob-azure-open-ai-service"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"gojob-aglae-candidate-prequalification"},{"title":"Golden 1 Credit Union: custom machine learning credit scorecard","useCases":["alternative-data-credit-scoring"],"organization":{"name":"Golden 1 Credit Union","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Zest AI","role":"platform"}],"summary":"Golden 1, a California credit union with about USD 21 billion in assets, built a custom machine learning credit scorecard with Zest AI, trained on its own members and on other Californians who resembled its membership. It launched on credit cards in December 2022 and extended to unsecured and auto loans. The CEO reported higher approvals overall and a 28% increase in approvals to protected classes of borrowers. The article does not say that the scorecard uses data from outside the credit file, so the record illustrates the machine learning half of this use case.","stage":"production","year":2024,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.bankingdive.com/news/golden-1-credit-union-zest-ai-partnership-28-percent-increase-protected-classes-bias-algorithm/709445/","title":"How Golden 1 used AI to find 'good risk'","publisher":"Banking Dive","date":"2024-03-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"golden-1-credit-union-ai-credit-scorecard"},{"title":"Google Chrome: AI agents that triage and fix security bugs","useCases":["software-vulnerability-remediation"],"organization":{"name":"Google","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Google","role":"in-house"}],"summary":"The Chrome security team uses Gemini based agents across the life of a security bug. An automated triage pipeline filters spam and duplicates, reproduces bugs, adds severity and routes them to the owner; fixing agents propose candidate patches that a critic agent reviews, and test writing agents add tests before a developer evaluates the fix. Chrome fixed 1,072 security bugs in milestones 149 and 150, more than the prior 23 milestones combined, and Google says LLMs now generate candidate fixes for most vulnerabilities. It estimates the triage automation saves hundreds of hours of developer time a month.","stage":"scaled","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://blog.google/security/chrome-stronger-with-every-update/","title":"Stronger with every update: How we’re making Chrome and the web safer in the AI Era","publisher":"Google (The Keyword)","date":"2026-07-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"google-chrome-ai-vulnerability-triage-and-fixing"},{"title":"Google: AI assisted internal code migrations cut engineering time by an estimated 50%","useCases":["legacy-code-modernization"],"organization":{"name":"Google","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Google","role":"in-house"}],"summary":"Google used an internal LLM based system to help engineers with large scale code migrations, such as changing identifier types from int32 to int64. Engineers doing the migrations estimated the total time spent was reduced by about 50%, and reported that 80% of the code changes in landed changelists were AI authored, with the rest written by humans.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"processing-time-reduction","value":50,"unit":"percent","qualifier":"approximately","claimant":"organization","quote":"The total time spent on the migration was reduced by an estimated 50% as reported by the engineers doing the migration.","sourceUrl":"https://research.google/blog/accelerating-code-migrations-with-ai/"}],"outcomeDisclosed":true,"sources":[{"url":"https://research.google/blog/accelerating-code-migrations-with-ai/","title":"Accelerating code migrations with AI","publisher":"Google Research","date":"2026-01-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"google-legacy-code-migration"},{"title":"Google: Gemini pipeline that drafts fixes for sanitizer bugs","useCases":["software-vulnerability-remediation"],"organization":{"name":"Google","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Google","role":"in-house"}],"summary":"Google's security engineering team built a pipeline that prompts Gemini to generate code fixes for bugs that sanitizers find during unit tests in C and C++, Java and Go code, such as uninitialised values, data races and buffer overflows. Every generated fix goes to a human reviewer before it lands. Google reports that the pipeline fixed 15% of these bugs, hundreds in total, and expects the rate to improve.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://research.google/pubs/ai-powered-patching-the-future-of-automated-vulnerability-fixes/","title":"AI-powered patching: the future of automated vulnerability fixes","publisher":"Google Research (Google Security Engineering Technical Report)"},{"url":"https://storage.googleapis.com/gweb-research2023-media/pubtools/7563.pdf","title":"AI-powered patching: the future of automated vulnerability fixes (full report, PDF)","publisher":"Google Security Engineering"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"google-sanitizer-bug-ai-patching"},{"title":"Google: SREs use Gemini CLI from page to postmortem","useCases":["aiops-incident-triage"],"organization":{"name":"Google","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Google (Gemini)","role":"in-house"}],"summary":"Google site reliability engineers use an agent in the Gemini CLI across an outage: reading the page, investigating, proposing mitigations, finding the root cause and drafting the postmortem. Every proposed change passes a policy layer (for example rules that need two person approval) and a forced human confirmation, and every proposal and approval is logged. No outcome figures are published.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/blog/topics/developers-practitioners/how-google-sres-use-gemini-cli-to-solve-real-world-outages","title":"How Google SREs Use Gemini CLI to Solve Real-World Outages","publisher":"Google Cloud Blog","date":"2026-01-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"google-sre-gemini-cli-incident-response"},{"title":"Government Digital Service: GOV.UK Chat in the GOV.UK app","useCases":["citizen-information-assistant"],"organization":{"name":"Government Digital Service","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Anthropic","role":"model-provider"},{"name":"Amazon Web Services","role":"platform"}],"summary":"GOV.UK Chat is a retrieval augmented generation assistant in the GOV.UK app that answers questions from GOV.UK guidance only, with links to the source pages under every answer. It retrieves from a vector index of roughly 100,000 GOV.UK pages, uses Claude Sonnet 4 on Amazon Bedrock in the EU, rejects questions that contain common personal data patterns, and runs every answer through a second model check on advice, language, tone and quality before the user sees it. The transparency record describes a limited test with up to 2,000 users over four weeks; it makes no decisions about users.","stage":"pilot","year":2025,"channels":["mobile-app"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/dsit-gov-dot-uk-chat","title":"DSIT: GOV.UK Chat","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-10-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"government-digital-service-gov-uk-chat"},{"title":"Government of the City of Buenos Aires: Boti citizen chatbot","useCases":["citizen-information-assistant"],"organization":{"name":"Government of the City of Buenos Aires","anonymized":false,"country":"AR","region":"latin-america","industry":"government"},"vendors":[{"name":"Microsoft (Azure OpenAI Service)","role":"platform"},{"name":"Pi Data Strategy & Consulting","role":"integrator"}],"summary":"Boti is the city's WhatsApp assistant for residents and visitors, launched on WhatsApp in 2019, which the city describes as the first government in the world to use WhatsApp as a contact channel with its citizens. Its services include appointments for procedures such as driver's licence renewal, public transport schedules, requests such as bulky waste collection and, during the pandemic, vaccination appointments and test results. In 2024 the city added a generative AI experience built on Azure OpenAI Service, scoped to tourism so the team could experiment without sensitive data. The city stresses tone (River Plate voseo, inclusive language) and a single central repository of government information as the basis for grounded answers.","stage":"scaled","year":2024,"channels":["whatsapp"],"languages":["es","en"],"metrics":[{"kpi":"interactions-handled","value":2000000,"unit":"count","qualifier":"at-least","period":"queries per month without human intervention","claimant":"vendor","quote":"Boti, the AI-powered chatbot, handles over 2 million queries per month without human intervention.","sourceUrl":"https://www.microsoft.com/en/customers/story/21596-government-of-the-city-of-buenos-aires-azure-open-ai-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/21596-government-of-the-city-of-buenos-aires-azure-open-ai-service","title":"Buenos Aires City: How generative AI is revolutionizing the lives of millions with Azure OpenAI Services","publisher":"Microsoft Customer Stories"},{"url":"https://buenosaires.gob.ar/boti","title":"Boti","publisher":"Gobierno de la Ciudad Autónoma de Buenos Aires"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"government-of-the-city-of-buenos-aires-boti"},{"title":"Great Ormond Street Hospital: London wide trial of an AI scribe across nine NHS sites","useCases":["ambient-clinical-documentation"],"organization":{"name":"Great Ormond Street Hospital for Children NHS Foundation Trust","anonymized":false,"country":"GB","region":"europe","industry":"healthcare"},"vendors":[{"name":"TORTUS","role":"platform"}],"summary":"An NHS England sponsored study led by the GOSH DRIVE innovation unit tested the TORTUS ambient scribe at nine London sites, including hospitals, GP practices, mental health services and ambulance teams, over more than 17,000 patient encounters. The tool transcribes the consultation and drafts a clinic note and letter that the clinician checks and edits before saving. Direct patient interaction time rose and appointments got shorter; in A&E at St George's University Hospital, clinicians saw more patients per shift. A rollout across GOSH outpatient settings was planned to follow.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":17000,"unit":"count","qualifier":"at-least","period":"evaluation across nine London NHS sites","claimant":"organization","quote":"Over 17,000 patient encounters were evaluated across a diverse range of sites including hospitals, GP practices, mental health services and ambulance teams.","sourceUrl":"https://www.gosh.nhs.uk/news/researchgosh-led-trial-of-ai-scribe-technology-shows-transformative-benefits-for-patients-and-clinicians-across-london/"},{"kpi":"handling-time-reduction","value":8.2,"unit":"percent","qualifier":"exact","period":"overall appointment length, trial sites","claimant":"organization","quote":"Results showed a 23.5% increase in direct patient interaction time during appointments, alongside an 8.2% reduction in overall appointment length when AI-scribes were used.","sourceUrl":"https://www.gosh.nhs.uk/news/researchgosh-led-trial-of-ai-scribe-technology-shows-transformative-benefits-for-patients-and-clinicians-across-london/"},{"kpi":"productivity-gain","value":13.4,"unit":"percent","qualifier":"exact","period":"A&E at St George's University Hospital, patients seen per shift","claimant":"organization","quote":"A&E saw particularly strong results, with a 13.4% increase in patients seen per shift.","sourceUrl":"https://www.gosh.nhs.uk/news/researchgosh-led-trial-of-ai-scribe-technology-shows-transformative-benefits-for-patients-and-clinicians-across-london/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gosh.nhs.uk/news/researchgosh-led-trial-of-ai-scribe-technology-shows-transformative-benefits-for-patients-and-clinicians-across-london/","title":"GOSH-led trial of AI-scribe technology shows 'transformative' benefits for patients and clinicians across London","publisher":"Great Ormond Street Hospital","date":"2025-09-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"great-ormond-street-hospital-ai-scribe-trial"},{"title":"GroupeActive: GAIA Propale agent drafts sales proposals from meeting notes","useCases":["rfp-and-proposal-response-drafting"],"organization":{"name":"GroupeActive","anonymized":false,"country":"FR","region":"europe","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"},{"name":"Witivio","role":"integrator"}],"summary":"GroupeActive, a French SME that supports a network of about 80 independent experts, built the \"GAIA Propale\" agent in Microsoft Copilot Studio with the partner Witivio. The agent turns the expert's customer meeting notes into several chapters of a structured sales proposal, which the expert reviews and a proofreading committee checks. In the pilot a proposal takes about two hours instead of a day, and the time from proposal to signed contract fell by a factor of four.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["fr"],"metrics":[{"kpi":"handling-time-reduction","value":75,"unit":"percent","qualifier":"exact","period":"pilot, average drafting time per sales proposal","claimant":"organization","quote":"Our members save an average of 75% on drafting time.","sourceUrl":"https://www.microsoft.com/en/customers/story/24754-groupeactive-microsoft-teams"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/24754-groupeactive-microsoft-teams","title":"GroupeActive and Microsoft reinvent sales proposals with Microsoft Copilot Studio","publisher":"Microsoft","date":"2025-07-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"groupeactive-gaia-propale-proposals"},{"title":"US General Services Administration: classifying federal transactions into the category management taxonomy","useCases":["procurement-spend-classification"],"organization":{"name":"U.S. General Services Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The General Services Administration runs an Acquisition Analytics capability that uses natural language processing to classify transactions within the Government-wide Category Management Taxonomy. The classification lets category managers see how obligations are distributed across categories and decide where to aggregate spend across agencies. The inventory lists it as deployed; no operational date or outcome figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (GSA entry \"Acquisition Analytics\")","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"gsa-acquisition-analytics-spend-classification"},{"title":"US General Services Administration: CALI proposal compliance checking tool, retired while in training","useCases":["procurement-contract-review"],"organization":{"name":"General Services Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Octo Consulting","role":"integrator"},{"name":"Amazon Web Services","role":"platform"}],"summary":"GSA's Contract Acquisition Lifecycle Intelligence (CALI) is a machine learning tool built to check vendor proposals for compliance in four areas (format, forms, representations and certifications, and requirements) to support source selection, with designated evaluation members reviewing the results. It was offered by Octo Consulting as a hosted service on AWS. The 2024 federal AI use case inventory describes it as still being trained on sample data, gives no implementation date, states that no testing in an operational environment had been done, and lists it as retired on 1 November 2024. No outcome and no reason for retirement are published, and nothing in the source shows it was used on live source selections; it is recorded here because stopped projects are part of the evidence.","stage":"paused","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated federal AI use case inventory (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"gsa-cali-proposal-evaluation"},{"title":"US General Services Administration: Solicitation Review Tool that screens ICT solicitations for compliance language","useCases":["procurement-contract-review"],"organization":{"name":"General Services Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"GSA's Solicitation Review Tool screens federal information and communications technology solicitations published on SAM.gov and flags those that may lack required compliance language, automating a first screening that would otherwise need extensive manual review by agencies. A related version that checks for Section 508 accessibility requirements was listed as pre deployment. The inventory lists the tool as deployed without further detail on benefits or outcomes.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"gsa-solicitation-review-tool"},{"title":"GXS Bank: ecosystem data from Grab and Singtel in FlexiLoan credit decisions","useCases":["alternative-data-credit-scoring"],"organization":{"name":"GXS Bank","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"FICO","role":"platform"}],"summary":"GXS Bank, a Singapore digital bank, decisions its FlexiLoan personal loan with user permissioned data from its ecosystem partners Grab and Singtel layered on top of credit bureau scores, to expand credit access to underserved users, such as people starting their careers and entrepreneurs with fluctuating incomes, who were previously overlooked by traditional banks. The FICO decision platform returns credit decisions in milliseconds, and the bank reports onboarding in under three minutes for the vast majority of approved applications. The platform was implemented in three months.","stage":"production","year":2024,"channels":["mobile-app","api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.theasianbanker.com/press-releases/gxs-bank-achieves-onboarding-efficiency-with-fico-platform","title":"GXS Bank achieves onboarding efficiency with FICO platform","publisher":"The Asian Banker (FICO press release)","date":"2024-05-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"gxs-bank-alternative-data-flexiloan"},{"title":"Hamilton County Schools: Khanmigo AI tutor in a two year randomized trial in 18 middle schools","useCases":["ai-tutor-for-students"],"organization":{"name":"Hamilton County Schools","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[{"name":"Khan Academy","role":"platform"}],"summary":"Philip Oreopoulos and Nina Low ran a two year cluster randomized trial in 18 Tennessee middle schools in Hamilton County in the 2024/25 and 2025/26 school years, in which randomly assigned students used Khan Academy with its AI tutor Khanmigo, configured to coach rather than give answers, during existing daily remedial maths sessions. Assignment raised maths achievement by 1.3 national percentile ranks per term, similar to Khan Academy practice without AI. Almost every student tried Khanmigo, but the median student messaged it on only a third of the days they practiced and in only 17 percent of the exercise sessions in which they made a mistake, and the messages students did send were mostly bare answers or clicks on suggested prompts. Chalkbeat reports that Khan Academy has since redesigned its interface to integrate Khanmigo better.","stage":"pilot","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://edworkingpapers.com/ai26-1551","title":"One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (EdWorkingPaper 26-1551)","publisher":"Annenberg Institute at Brown University"},{"url":"https://edworkingpapers.com/sites/default/files/ai26-1551.pdf","title":"One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (full paper, PDF)","publisher":"Annenberg Institute at Brown University"},{"url":"https://www.chalkbeat.org/2026/08/25/ai-tutoring-students-khanmigo-khan-academy-engagement-study/","title":"Students rarely engaged with Khan Academy's AI-powered tutor Khanmigo, study finds","publisher":"Chalkbeat","date":"2026-08-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"hamilton-county-schools-khanmigo-trial"},{"title":"Harvard University: the CS50 Duck, an AI tutor that guides instead of answering","useCases":["ai-tutor-for-students"],"organization":{"name":"Harvard University","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[{"name":"OpenAI","role":"model-provider"},{"name":"Harvard CS50","role":"in-house"}],"summary":"Since spring 2023 Harvard's introductory computer science course CS50 has run the \"CS50 Duck\", an AI tutor built on the ChatGPT API that is deliberately less helpful than a general chatbot: it asks more questions than it answers and avoids giving outright solutions. The course combines prompting with its own code that tries to evaluate each reply before the student sees it and sometimes rejects or retries it, and added a \"heart system\" that limits questions per period after some students asked around 200 questions. David Malan calls it a net positive but acknowledges the Duck still sometimes returns code despite instructions not to.","stage":"production","year":2023,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://news.harvard.edu/gazette/story/2026/09/taming-the-duck-for-starters/","title":"Taming the Duck, for starters","publisher":"Harvard Gazette","date":"2026-09-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"harvard-cs50-duck-ai-tutor"},{"title":"Harvard College: Student Compass academic advising chatbot","useCases":["academic-advising-assistant"],"organization":{"name":"Harvard College","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[{"name":"OpenAI","role":"model-provider"}],"summary":"Harvard College's Advising Programs Office built Student Compass, a ChatGPT Edu powered chatbot that answers undergraduate advising questions from a fixed set of policy documents and Faculty of Arts and Sciences websites, including the Student Handbook and departmental pages, and cites its sources. It reached incoming students over the summer of 2026 as they selected their first semester courses. The Harvard Crimson tested the tool, spoke with students and directors of undergraduate studies, and received statements from the program's Assistant Director: its own testing found the bot handled straightforward policy questions well but its limits showed on more nuanced concentration advising questions, and it cannot access live course search, Q reports, or some syllabi. Joseph K. Blitzstein, the Statistics department's director of undergraduate studies, separately tested it against real concentration advising questions and found it often gave wrong answers. In April 2026, Harvard's Advising Programs Office announced the tool as \"not meant to replace human advising\"; in a September 2026 statement to the Crimson, Assistant Director Brooks B. Lambert-Sluder called it a \"starting point\" rather than a replacement for human advising.","stage":"production","year":2026,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.thecrimson.com/article/2026/9/4/ai-advising-chatbot/","title":"Harvard Built an AI Academic Adviser. Here's How It Fared.","publisher":"The Harvard Crimson","date":"2026-09-04"},{"url":"https://www.thecrimson.com/article/2026/4/2/harvard-ai-chatbot-advising/","title":"Harvard College Plans AI Chatbot To Guide Students on Courses, Requirements Beginning This Summer","publisher":"The Harvard Crimson","date":"2026-04-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"harvard-student-compass-academic-advising"},{"title":"Hemominas: omnichannel chatbot for blood donor search and scheduling","useCases":["branch-and-appointment-booking-agent"],"organization":{"name":"Hemominas","anonymized":false,"country":"BR","region":"latin-america","industry":"healthcare"},"vendors":[{"name":"Xertica","role":"integrator"},{"name":"Google Cloud","role":"platform"}],"summary":"Hemominas, Brazil's largest blood bank, partnered with Xertica to develop an omnichannel chatbot for donor search and scheduling of donations. Google Cloud describes the expected impact in terms of potential lives saved, which is a projection, not a measured result, so the record is kept at the announced stage.","stage":"announced","year":2024,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"},{"url":"https://web.archive.org/web/20241003233844/https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"185 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"hemominas-donor-scheduling-chatbot"},{"title":"HHS Administration for Children and Families: AI review of documents against new directives","useCases":["policy-drafting-and-gap-analysis","regulatory-horizon-scanning"],"organization":{"name":"Administration for Children and Families","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Palantir","role":"platform"},{"name":"Credal","role":"platform"}],"summary":"In March 2025 the Administration for Children and Families, part of the US Department of Health and Human Services, deployed AI to review its existing grants, new grant applications and position descriptions for alignment with HHS Secretarial Directives related to recent executive orders. The AI produces an initial list of documents that may need revision (for grants, with an initial assessment and example passages); program office staff then review, justify and recommend. For position descriptions the agency states that AI was not used to make any final determinations. It is the gap detection half of policy change work: finding which existing documents a new requirement touches.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (HHS/ACF entries Document Review for Alignment with Executive Orders)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"hhs-acf-directive-alignment-document-review"},{"title":"US Administration for Children and Families: AI support for reviewing proposals and drafting technical evaluations","useCases":["procurement-contract-review"],"organization":{"name":"U.S. Department of Health and Human Services, Administration for Children and Families","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Credal","role":"platform"},{"name":"Ask Sage","role":"platform"},{"name":"Microsoft","role":"platform"}],"summary":"When the Administration for Children and Families receives many vendor responses to requests for information and proposals, review teams must write summarised comments on each response against pre established evaluation criteria. Since July 2025 evaluators use generative AI to find relevant passages in the proposals and to draft language for technical evaluation documents, for example pulling and formatting examples with page citations to support the evaluator's own assessment. The inventory states that AI does not make final determinations and that evaluators review all drafted language, revise it as needed and verify any cited excerpts. The systems listed are ACF Credal, Microsoft Copilot Chat and Ask Sage, with Ask Sage marked as decommissioned. No outcome figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"hhs-acf-proposal-review-drafting"},{"title":"Hippo: Clara, an AI assistant for first notice of loss and claims processing","useCases":["claims-first-notice-of-loss-agent"],"organization":{"name":"Hippo","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[],"summary":"Hippo, a US insurance platform that describes itself as a multi line carrier, reported in its first quarter 2026 investor update that it launched Clara in that quarter, an AI assistant for first notice of loss and end to end claims processing, as part of a wider move to agentic AI in claims, including capacity for catastrophe events. The presentation gives expectations, such as the share of homeowners claims it expects to be filed digitally, but no measured results yet.","stage":"production","year":2026,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1828105/000182810526000023/q126quarterlyinvestorupd.htm","title":"Hippo Holdings Inc. first quarter 2026 quarterly investor update (Form 8-K exhibit)","publisher":"Hippo Holdings Inc. (via SEC EDGAR)","date":"2026-04-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"hippo-clara-ai-claims-assistant"},{"title":"Hiscox: Microsoft 365 Copilot in claims handling","useCases":["claims-triage-and-straight-through-processing","outbound-notice-drafting"],"organization":{"name":"Hiscox","anonymized":false,"country":"GB","region":"europe","industry":"insurance"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Hiscox is rolling out Microsoft 365 Copilot to its more than 3,000 employees after a trial. A senior technical claims underwriter in the UK claims team uses it to identify and record the key information of a new claim, to summarise long expert medical evidence and legal advice, and to pull the progress of a claim from several emails and compose an update to a broker or customer. He says that recording a new claim now takes him as little as 10 minutes instead of up to an hour. This is one user's experience, not a measured program result.","stage":"production","year":2025,"channels":["internal-tools","email"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://ukstories.microsoft.com/features/how-ai-is-supercharging-hiscox-employees-to-do-what-theyre-great-at/","title":"How AI is ‘supercharging’ Hiscox employees","publisher":"Microsoft UK Stories","date":"2025-06-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"hiscox-copilot-claims-handling"},{"title":"Hiscox: generative AI lead underwriting model for sabotage and terrorism risks","useCases":["commercial-underwriting-submission-triage","underwriting-risk-assessment-copilot","insurance-renewal-and-retention"],"organization":{"name":"Hiscox","anonymized":false,"country":"GB","region":"europe","industry":"insurance"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Hiscox London Market combined its own Hiscox AI Laboratories (Hailo) with Google Cloud's Gemini model to automate lead underwriting from email submission to quote in its sabotage and terrorism line. In scope risks are assessed by the model and the process generates an email to the broker with pricing and other data completed, ready for underwriter review. After a December 2023 proof of concept, in which Hiscox said the manual extraction step can take up to three days and quotes could be produced within three minutes, the model went live in August 2024. It initially covers renewals of existing US and Canadian sabotage and terrorism risks, excluding the New York and Chicago metro areas.","stage":"production","year":2024,"channels":["email","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.hiscoxgroup.com/news/press-releases/2024/12-08-24","title":"Hiscox's generative AI-enhanced lead underwriting model enabled by Google Cloud goes live","publisher":"Hiscox Group","date":"2024-08-12"},{"url":"https://www.hiscoxgroup.com/news/press-releases/2023/12-12-23","title":"Hiscox and Google Cloud Collaborate on AI in lead underwriting for the London Market","publisher":"Hiscox Group","date":"2023-12-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"hiscox-generative-ai-lead-underwriting"},{"title":"HMRC: Ask HMRC online digital assistant with webchat escalation","useCases":["tax-questions-and-filing-assistant"],"organization":{"name":"HM Revenue and Customs","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Microsoft (formerly Nuance)","role":"platform"}],"summary":"HMRC's digital assistant answers tax questions typed in plain language, matching them to intents with natural language understanding and replying with non personalised answers that link to GOV.UK guidance. It covers 60 of HMRC's more than 120 taxes, asks the user to choose between likely meanings when unsure, and needs no login. Complex questions escalate to webchat with a human adviser, who sees the whole prior conversation and only continues after identity and verification checks.","stage":"scaled","year":2025,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":5480000,"unit":"count","qualifier":"at-least","period":"tax year 2024/25 to 6 March 2025","claimant":"organization","quote":"The digital assistant has had 5.48m interactions in the tax year 2024/25 (to date as at 6 March 2025).","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/hmrc-ask-hmrc-online"},{"kpi":"accuracy","value":83.03,"unit":"percent","qualifier":"exact","period":"NLU test set, March 2025, known intents","claimant":"organization","quote":"83.03% on known intents","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/hmrc-ask-hmrc-online"},{"kpi":"contact-deflection","value":20,"unit":"percent","qualifier":"exact","period":"webchat escalations to an adviser, tax year 2024/25 to 6 March 2025","baseline":"escalations in the 2023/24 tax year","claimant":"organization","quote":"For webchat, 909,000 users have escalated to speak to an adviser in the tax year 2024/25 (to date as at 6 March 2025). That is a 20% decrease on the 2023/24 tax year.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/hmrc-ask-hmrc-online"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/hmrc-ask-hmrc-online","title":"HMRC: Ask HMRC online","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-06-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"hmrc-ask-hmrc-digital-assistant"},{"title":"HM Revenue and Customs: VAT Return Analysis Tool with anomaly detection for compliance checks","useCases":["tax-compliance-risk-scoring"],"organization":{"name":"HM Revenue and Customs","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"In house (HMRC Data Science Analytics)","role":"in-house"}],"summary":"HMRC's VAT Return Analysis Tool brings a VAT trader's entity, ledger and return data for the most recent seven years into one interactive view for VAT officers, and uses a classical statistical model (seasonal trend decomposition with an interquartile range rule) to flag anomalous values in the return history. Officers use it to prepare and carry out compliance checks; the tool makes no decisions, and any assessment is made by an officer and can be appealed through the normal route. HMRC's transparency record says around 5,500 officers are licensed and the tool is used about 1,500 times a day.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":5500,"unit":"count","qualifier":"approximately","period":"licensed VAT officers","claimant":"organization","quote":"Around 5,500 officers have a license to use the tool as part of their VAT compliance work.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/hmrc-vat-return-analysis-tool"},{"kpi":"interactions-handled","value":1500,"unit":"count","qualifier":"approximately","period":"per day","claimant":"organization","quote":"There are ~1,500 daily uses of the tool.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/hmrc-vat-return-analysis-tool"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/hmrc-vat-return-analysis-tool","title":"HMRC: VAT Return Analysis Tool (algorithmic transparency record)","publisher":"GOV.UK","date":"2025-12-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"hmrc-vat-return-analysis-tool"},{"title":"Holland America Line: Anna digital concierge for cruise guests and travel advisors","useCases":["travel-and-hotel-booking-concierge"],"organization":{"name":"Holland America Line","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[{"name":"Microsoft","role":"platform"},{"name":"OpenAI","role":"model-provider"}],"summary":"Holland America Line built Anna, a generative AI digital concierge on its website for new and existing cruise guests and the travel advisors who book for them. The first release supports booking new cruises, adding products and services to existing bookings and general questions, and connects to the CRM and reservation data. It was rolled out in waves (contact centre agents, employees, then 5%, 50% and 100% of website visitors) and runs in the United States, with more markets and languages planned.","stage":"production","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/19787-holland-america-dataverse","title":"Holland America Line sees signs of more informed purchasing with Copilot Studio agent | Microsoft Customer Stories","publisher":"Microsoft","archivedUrl":"https://web.archive.org/web/20250119055636/https://www.microsoft.com/en/customers/story/19787-holland-america-dataverse"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"holland-america-line-anna-digital-concierge"},{"title":"Home Office: complexity routing of visitor visa applications (CARS Visits)","useCases":["immigration-and-visa-application-assistant"],"organization":{"name":"Home Office (Visa, Status and Information Services)","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[],"summary":"Since 25 April 2023 the Complexity Application Routing Solution for visits (CARS(V)) labels visitor visa applications under Appendix V as likely non complex or complex, so each goes to the right grade of decision maker, using business rules on the applicant's declared answers (such as a yes to any criminality question) plus matches against risk profiles and bulk data tables of information previously seen in applications that gave false information. It uses no machine learning, does not decide applications, and decision makers can reroute a case; its equality impact assessments allow direct discrimination on race or nationality where a Ministerial Authorisation supports it, new profiles are applied weekly, and the profiles and bulk tables are not set out in the record. The record gives around 2.5 million visit applications received in 2023 for scale, not a count routed by the tool; separately, the Home Office withdrew an earlier visa streaming algorithm in 2020 after a legal challenge.","stage":"scaled","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/home-office-complexity-application-routing-solution-visits","title":"Home Office: Complexity Application Routing Solution, Visits","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2024-12-17"},{"url":"https://www.gov.uk/government/publications/complexity-application-routing-solution-visits-caseworker-guidance/complexity-application-routing-solution-visits-carsv-accessible","title":"Complexity application routing solution (visits) (CARS(V)): caseworker guidance","publisher":"UK Visas and Immigration (GOV.UK)","date":"2023-05-17"},{"url":"https://www.foxglove.org.uk/2020/08/04/home-office-says-it-will-abandon-its-racist-visa-algorithm-after-we-sued-them/","title":"Home Office says it will abandon its racist visa algorithm, after we sued them","publisher":"Foxglove","date":"2020-08-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"home-office-visit-visa-complexity-routing"},{"title":"Hotel Swexan: Duetto revenue management for a one person team","useCases":["hotel-revenue-management-copilot"],"organization":{"name":"Hôtel Swexan","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[{"name":"Duetto","role":"platform"}],"summary":"Hôtel Swexan, a 134 room luxury property in Dallas with more than 20 room types, eight premium suites and five food and beverage outlets, runs its entire revenue management function with one person, Director of Revenue Jessica Schiele. Before Duetto, the hotel relied on a remote third party consultant with no visibility into daily operations, and rate changes were slow with no room for granular, segment level pricing. Duetto now forecasts occupancy, recommends demand driven rates by room type and segment, executes automated pricing strategies around the clock based on occupancy thresholds and segment logic Schiele defines, and scores group and event enquiries for profitability (room rate, food and beverage minimums, room rental and displacement) so any team member can quote a profitable group rate. Comparing 2025 with 2024, Duetto's case study reports double digit RevPAR growth in every segment.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"revenue-uplift","value":33,"unit":"percent","qualifier":"exact","period":"2025 versus 2024","baseline":"Total hotel RevPAR, 2024","claimant":"vendor","quote":"Total hotel RevPAR grew 33% year on year — a result that reflects more than favourable market conditions.","sourceUrl":"https://www.duettocloud.com/en-us/success-stories/hotel-swexan-drives-exceptional-revpar-growth-with-duetto"},{"kpi":"revenue-uplift","value":45,"unit":"percent","qualifier":"exact","period":"2025 versus 2024","baseline":"Suite RevPAR, 2024","claimant":"vendor","quote":"Suite RevPAR grew 45% year on year — occupancy up 29%, ADR up 13%.","sourceUrl":"https://www.duettocloud.com/en-us/success-stories/hotel-swexan-drives-exceptional-revpar-growth-with-duetto"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.duettocloud.com/en-us/success-stories/hotel-swexan-drives-exceptional-revpar-growth-with-duetto","title":"Hôtel Swexan drives exceptional RevPAR growth with Duetto","publisher":"Duetto"},{"url":"https://www.duettocloud.com/en-us/platform/gamechanger","title":"GameChanger: Predictive Analytics Software","publisher":"Duetto"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"hotel-swexan-duetto-revenue-management"},{"title":"Howard Brown Health: 24/7 multilingual voice agent that guides patients through scheduling in MyChart","useCases":["patient-appointment-scheduling-and-reminders-agent"],"organization":{"name":"Howard Brown Health","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"PolyAI","role":"platform"}],"summary":"Howard Brown Health, a federally qualified health center in Chicago, deployed a PolyAI voice agent named Alex that answers patient calls around the clock in several languages. Integrated with MyChart, it guides patients through scheduling appointments, test results and prescription refills, and escalates distressed callers to a person immediately. PolyAI reports 30% call containment against a 20% target, a 72% shorter average handle time for routine requests and a 4% increase in patient satisfaction. An Epic integration that lets patients create, reschedule and cancel appointments through the agent is described as the next phase.","stage":"production","year":2024,"channels":["voice"],"languages":[],"metrics":[{"kpi":"containment-rate","value":30,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"Initially aiming for a 20% call containment rate, PolyAI exceeded expectations by achieving 30%.","sourceUrl":"https://poly.ai/customers/howardbrownhealth"},{"kpi":"handling-time-reduction","value":72,"unit":"percent","qualifier":"exact","period":"routine requests","claimant":"vendor","quote":"72% decrease in Average Handle Time for routine requests","sourceUrl":"https://poly.ai/customers/howardbrownhealth"},{"kpi":"customer-satisfaction-uplift","value":4,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"Patient satisfaction scores also saw an increase of 4%, primarily driven by the enhanced ease and efficiency with which patients could schedule appointments and access services.","sourceUrl":"https://poly.ai/customers/howardbrownhealth"}],"outcomeDisclosed":true,"sources":[{"url":"https://poly.ai/customers/howardbrownhealth","title":"How Howard Brown Health provides personalized patient experiences with PolyAI","publisher":"PolyAI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"howard-brown-health-patient-voice-agent"},{"title":"HRSA (US Department of Health and Human Services): AI Audit Resolution Assistant","useCases":["outbound-notice-drafting"],"organization":{"name":"Health Resources and Services Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Mindpetal","role":"integrator"},{"name":"UiPath (Automation Cloud Public Sector)","role":"platform"}],"summary":"HRSA's auditors faced a sharp rise in Single Audits linked to COVID era Provider Relief Fund payments. Its AI Audit Resolution Assistant (AIARA) puts the Single Audit documents assigned to HRSA in a vector database and uses retrieval augmented generation with a large language model to summarise findings and recommendations, answer auditors' questions and draft the Management Decision Letters that close findings out, while robotic process automation pulls data from the Federal Audit Clearinghouse into letter templates. HRSA reports in the 2025 federal inventory that the pilot, operational since July 2024, has processed and resolved 73 audits and saved an estimated 276 hours.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":73,"unit":"count","qualifier":"exact","period":"since launch in July 2024, as reported in the 2025 inventory","baseline":"Single Audits processed and resolved with the assistant","claimant":"organization","quote":"Since its launch, the AIARA has successfully processed and resolved 73 audits.","sourceUrl":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv"},{"kpi":"hours-saved","value":276,"unit":"hours","qualifier":"approximately","period":"since launch in July 2024, as reported in the 2025 inventory","baseline":"auditor hours on Single Audit resolution without the automation","claimant":"organization","quote":"Automation has resulted in an estimated total of 276 hours of work saved.","sourceUrl":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv"}],"outcomeDisclosed":true,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated federal AI use case inventory (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"hrsa-ai-audit-resolution-assistant"},{"title":"HRSA: Policy Assistant for first drafts of policy documents (initiated)","useCases":["policy-drafting-and-gap-analysis"],"organization":{"name":"Health Resources and Services Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Health Resources and Services Administration, part of the US Department of Health and Human Services, reported in 2024 a Policy Assistant that uses large language models to generate first drafts of key policy documents, funding notices and budget documents from example documents, style guides, key policy decisions and its internal knowledge base, plus an editing tool that checks drafts for inconsistencies and errors. The goal is less drafting time and better document quality. The entry was at the initiated stage and does not appear in the 2025 inventory; no results are published.","stage":"announced","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2024-Federal-AI-Use-Case-Inventory","title":"2024 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI inventory (HHS/HRSA entry Policy Assistant)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"hrsa-policy-document-drafting-assistant"},{"title":"HSBC: Dynamic Risk Assessment, machine learning AML monitoring built with Google Cloud","useCases":["aml-alert-triage"],"organization":{"name":"HSBC","anonymized":false,"country":"GB","region":"global","industry":"banking"},"vendors":[{"name":"Google Cloud (AML AI)","role":"platform"}],"summary":"HSBC replaced rule based transaction monitoring with Dynamic Risk Assessment in its key markets, a machine learning system co developed with Google Cloud that scores customers for money laundering risk from their full transaction and customer data, and routes the highest risk to investigators. HSBC piloted it in 2021 and says it checks about 980 million transactions a month for signs of financial crime. HSBC reports finding two to four times more financial crime with much greater accuracy, and cutting the processing time to analyse billions of transactions from several weeks to a few days.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"false-positive-reduction","value":60,"unit":"percent","qualifier":"exact","period":"compared with before, per HSBC","claimant":"organization","quote":"Now, we have 60% fewer false positive cases.","sourceUrl":"https://www.hsbc.com/news-and-views/views/hsbc-views/harnessing-the-power-of-ai-to-fight-financial-crime"},{"kpi":"alert-volume-reduction","value":60,"unit":"percent","qualifier":"at-least","period":"compared with rules based transaction monitoring","claimant":"vendor","quote":"In fact, HSBC saw alert volumes decrease by more than 60%.","sourceUrl":"https://www.googlecloudpresscorner.com/2023-06-21-Google-Cloud-Launches-AI-Powered-Anti-Money-Laundering-Product-for-Financial-Institutions"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.hsbc.com/news-and-views/views/hsbc-views/harnessing-the-power-of-ai-to-fight-financial-crime","title":"Harnessing the power of AI to fight financial crime","publisher":"HSBC","date":"2024-06-10"},{"url":"https://www.googlecloudpresscorner.com/2023-06-21-Google-Cloud-Launches-AI-Powered-Anti-Money-Laundering-Product-for-Financial-Institutions","title":"Google Cloud Launches AI-Powered Anti Money Laundering Product for Financial Institutions","publisher":"Google Cloud","date":"2023-06-21"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"hsbc-dynamic-risk-assessment"},{"title":"HSBC: name screening, adverse media and transaction screening alert automation with Silent Eight","useCases":["sanctions-screening-adjudication","pep-and-adverse-media-screening"],"organization":{"name":"HSBC","anonymized":false,"country":"GB","region":"global","industry":"banking"},"vendors":[{"name":"Silent Eight","role":"platform"}],"summary":"Silent Eight has supplied HSBC with automation for name screening and adverse media alerts, and in February 2024 the two expanded the partnership to automated alert closure for transaction screening, which investigates and resolves payment screening alerts in real time. No outcome figures are disclosed.","stage":"production","year":2024,"channels":["internal-tools","api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.prnewswire.com/news-releases/silent-eight-announces-expansion-of-partnership-with-hsbc-to-provide-transaction-screening-solutions-302067562.html","title":"Silent Eight Announces Expansion of Partnership with HSBC To Provide Transaction Screening Solutions","publisher":"Silent Eight via PR Newswire","date":"2024-02-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"hsbc-silent-eight-screening-automation"},{"title":"US Department of Housing and Urban Development: Voice of the Customer analytics on surveys, calls and chats","useCases":["customer-feedback-analysis"],"organization":{"name":"U.S. Department of Housing and Urban Development","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Medallia","role":"platform"},{"name":"Qualtrics","role":"platform"}],"summary":"The customer experience team in HUD's Office of the Chief Financial Officer runs a Voice of the Customer application that applies transcription, speech and text analytics to customer feedback surveys and to contact centre calls and chats. It produces dashboards that trend sentiment and identify the key drivers of customer sentiment and service delivery performance, which HUD uses to manage its programs and contact centre providers. The department lists it as deployed since August 2024; no outcome figures are published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry HUD-2024-004, Voice of the Customer)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"hud-voice-of-the-customer-analytics"},{"title":"Hughes Network Systems: automated auditing and insight on sales calls","useCases":["sales-call-coaching-and-crm-update"],"organization":{"name":"Hughes Network Systems","anonymized":false,"country":"US","region":"north-america","industry":"telecommunications"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Hughes, part of EchoStar, replaced manual auditing of sales calls, where auditors listened to hours of recordings, with an automated speech to text and generative AI system on Azure AI Foundry. It produces call insights and directives for sales agents across the whole call, and the team uses automated evaluation tools to check the quality and groundedness of the AI output. Microsoft reports that the cost of a sales call audit fell by 90%, from USD 26 to USD 2 per call hour.","stage":"production","year":2025,"channels":["voice","internal-tools"],"languages":["en"],"metrics":[{"kpi":"cost-reduction","value":90,"unit":"percent","qualifier":"exact","baseline":"USD 26 per hour for each call audited manually","claimant":"vendor","quote":"Collectively, these AI initiatives have boosted overall productivity by up to 25%, including automated sales call audit reductions of 90%, from $26 per hour for each call to just $2.","sourceUrl":"https://www.microsoft.com/en/customers/story/24300-hughes-azure-ai-foundry"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/24300-hughes-azure-ai-foundry","title":"EchoStar and Hughes save thousands of work hours, cut costs with Azure AI","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"hughes-sales-call-auditing"},{"title":"Human Managed: Gemini and Google Security Operations for security alert triage","useCases":["security-alert-triage-and-investigation"],"organization":{"name":"Human Managed","anonymized":false,"country":"SG","region":"asia-pacific","industry":"technology"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"CloudMile","role":"integrator"}],"summary":"Human Managed, a Singapore based security intelligence provider for large enterprises, uses Google Security Operations, Vertex AI and Gemini to analyse its customers' alerts and logs. Before, its specialists needed up to 30 minutes per alert to analyse the data and send a contextual alert to the customer; now statistical models and generative AI produce a confidence score with an explanation, and customers get prioritised alerts with remediation advice within 15 minutes.","stage":"production","year":2025,"channels":["internal-tools","api"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":97,"unit":"percent","qualifier":"exact","period":"time to triage an alert, from up to 30 minutes to under one minute","claimant":"vendor","quote":"With the system monitoring data and continuously learning about new threats, manual triage has been reduced by more than 60% and the time to triage alerts has been slashed to less than one minute – a 97% reduction that helps Human Managed's customers respond to security incidents much faster.","sourceUrl":"https://cloud.google.com/customers/human-managed"},{"kpi":"productivity-gain","value":60,"unit":"percent","qualifier":"at-least","period":"reduction in manual triage work","claimant":"vendor","quote":"With the system monitoring data and continuously learning about new threats, manual triage has been reduced by more than 60% and the time to triage alerts has been slashed to less than one minute – a 97% reduction that helps Human Managed's customers respond to security incidents much faster.","sourceUrl":"https://cloud.google.com/customers/human-managed"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/human-managed","title":"Human Managed uses Vertex AI and Google SecOps to triage security alerts 97% faster","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"human-managed-google-secops-alert-triage"},{"title":"Humber River Health: Command Centre outcomes","useCases":["hospital-bed-and-staff-capacity-command-center"],"organization":{"name":"Humber River Health","anonymized":false,"country":"CA","region":"north-america","industry":"healthcare"},"vendors":[{"name":"GE Healthcare","role":"platform"}],"summary":"Humber River Hospital, in Toronto, opened Canada's first hospital command centre in November 2017, built in collaboration with GE Healthcare Partners. The NASA style control room combines live data from more than 600 connected patient rooms with predictive analytics and machine learning to prioritize risk, predict spikes in emergency visits and coordinate bed turnaround, portering and diagnostics across the hospital. Different departments monitor customizable analytic tiles on a shared wall of displays, and the hospital has since extended the program from operational functions (Generation 1) to clinical alerting (Generation 2), and was moving forward with virtual care and home monitoring (Generation 3) as of 2022.","stage":"scaled","year":2017,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":34,"unit":"percent","qualifier":"exact","period":"since implementation, reported 2022","claimant":"organization","quote":"Humber also saw a decrease in wait times for inpatient diagnostics and emergency rooms, with a 34 per cent reduction in the average time a patient in the emergency department waited before being placed in a bed.","sourceUrl":"https://www.hrh.ca/2022/07/28/humber-river-hospitals-command-centre-and-generation-3/"},{"kpi":"processing-time-reduction","value":45,"unit":"percent","qualifier":"exact","period":"since implementation, reported 2022","claimant":"organization","quote":"In addition, Humber had a 45 per cent decrease in the time to clean inpatient beds with accurate bed planning.","sourceUrl":"https://www.hrh.ca/2022/07/28/humber-river-hospitals-command-centre-and-generation-3/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.hrh.ca/2022/07/28/humber-river-hospitals-command-centre-and-generation-3/","title":"Humber River Health's Command Centre – Outcomes and Generation 3","publisher":"Humber River Health","date":"2022-07-28"},{"url":"http://www.newswire.ca/news-releases/humber-river-hospital-breaking-new-ground-with-the-opening-of-canadas-first-hospital-command-centre-660975993.html","title":"Humber River Hospital Breaking New Ground with the Opening of Canada's First Hospital Command Centre","publisher":"Humber River Hospital","date":"2017-11-30"},{"url":"https://www.humbercommandcentre.ca","title":"Humber River Health Command Centre","publisher":"Humber River Health"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"humber-river-health-command-centre"},{"title":"HYPE: AI email handling and reply assistance for neobank customer service","useCases":["email-and-ticket-reply-drafting"],"organization":{"name":"HYPE","anonymized":false,"country":"IT","region":"europe","industry":"banking"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Italian neobank HYPE handled nearly 45% of its customer contacts by email and 40% by chat and WhatsApp when it moved to a new CRM. It built an agent that reads customer emails, segments them by subject and triggers an automatic reply, plus a voice agent that handles calls about common issues. Its human agents use Copilot in Dynamics 365 Customer Service for case and conversation summaries and email assistance. Microsoft reports that agents resolve WhatsApp conversations in half the time and that the custom agents cut human customer service intervention by 70% over a year.","stage":"production","year":2025,"channels":["email","whatsapp","agent-desktop"],"languages":["it"],"metrics":[{"kpi":"processing-time-reduction","value":50,"unit":"percent","qualifier":"approximately","period":"WhatsApp chat conversations resolved by human agents with Copilot summaries and email assistance (chat, not email or tickets)","claimant":"vendor","quote":"Copilot in Customer Service provides automated, AI-powered case and conversation summaries and email assistance, helping human agents resolve WhatsApp conversations in half the time.","sourceUrl":"https://www.microsoft.com/en/customers/story/19684-hype-dynamics-365-customer-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/19684-hype-dynamics-365-customer-service","title":"HYPE makes finance easier with Dynamics 365","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"hype-customer-service-email-and-chat-assist"},{"title":"Hyundai Card: AI agent system that produces the personalized 2025 annual card statement","useCases":["financial-wellbeing-coach"],"organization":{"name":"Hyundai Card","anonymized":false,"country":"KR","region":"asia-pacific","industry":"payments"},"vendors":[{"name":"Hyundai Card","role":"in-house"}],"summary":"Hyundai Card, a South Korean credit card issuer, has published an annual statement (연간명세서) in its app since 2021 that summarizes each member's yearly card spending. For the 2025 edition, released in January 2026, Hyundai Card applied an AI agent system it built itself, described as generative AI based on a large language model that runs a predesigned workflow. According to the company, the agent was used across the whole production run: analysing the payment data of 12.6 million members, generating a personalized message for each member and reviewing the results. The statement is a spending review with narrative, persona based insights and peer comparisons, not a legal account statement or tax document.","stage":"scaled","year":2026,"channels":["mobile-app"],"languages":["ko"],"metrics":[{"kpi":"interactions-handled","value":12600000,"unit":"count","qualifier":"exact","period":"members whose payment data the AI agent analysed for the 2025 annual statement, with a personalized message generated and reviewed per member","claimant":"organization","quote":"이번 연간명세서 제작 과정에서는 1260만 회원의 결제 데이터 분석, 회원별 개인화 메시지 생성, 결과 검수까지 전 과정에 AI 에이전트가 사용됐다.","sourceUrl":"https://newsroom.hyundaicard.com/front/board/AI%EA%B0%80-%EB%93%A4%EB%A0%A4%EC%A3%BC%EB%8A%94-2025%EB%85%84-%EC%86%8C%EB%B9%84-%EC%9D%B4%EC%95%BC%EA%B8%B0-%EC%97%B0%EA%B0%84%EB%AA%85%EC%84%B8%EC%84%9C%EC%97%90%EC%84%9C-%EB%A7%8C%EB%82%98%EB%B3%B4%EC%84%B8%EC%9A%94"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.hyundaicard.com/front/board/AI%EA%B0%80-%EB%93%A4%EB%A0%A4%EC%A3%BC%EB%8A%94-2025%EB%85%84-%EC%86%8C%EB%B9%84-%EC%9D%B4%EC%95%BC%EA%B8%B0-%EC%97%B0%EA%B0%84%EB%AA%85%EC%84%B8%EC%84%9C%EC%97%90%EC%84%9C-%EB%A7%8C%EB%82%98%EB%B3%B4%EC%84%B8%EC%9A%94","title":"AI가 들려주는 2025년 소비 이야기, 연간명세서에서 만나보세요 (Hyundai Card opens the 2025 annual statement)","publisher":"Hyundai Card","date":"2026-01-15"},{"url":"https://www.kbanker.co.kr/news/articleView.html?idxno=223593","title":"현대카드, AI 에이전트 적용 '연간명세서 2025' 출시","publisher":"대한금융신문 (Korea Financial Newspaper)","date":"2026-01-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"hyundai-card-ai-annual-statement"},{"title":"IBM: AskHR, the virtual HR assistant for employees and managers","useCases":["hr-and-policy-assistant"],"organization":{"name":"IBM","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"IBM","role":"in-house"}],"summary":"IBM's internal virtual agent AskHR automates more than 80 HR tasks, from payslip and sickness policy questions to job verification letters and vacation requests, and lets managers start transfers and organization changes in SAP SuccessFactors. AskHR has been refined since 2016; IBM added watsonx Orchestrate for generative and agentic automation in 2025. IBM reports high containment, fewer tickets and lower HR operating cost from the assistant over that longer period.","stage":"scaled","year":2016,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"containment-rate","value":94,"unit":"percent","qualifier":"exact","period":"common questions","claimant":"organization","quote":"AskHR also achieved a 94% containment rate of common questions, has led to a 75% reduction in support tickets raised since 2016, and created more than 11.5 million employee interactions in 2024 alone.","sourceUrl":"https://www.ibm.com/case-studies/ibm-askhr"},{"kpi":"interactions-handled","value":11500000,"unit":"count","qualifier":"at-least","period":"calendar year 2024","claimant":"organization","quote":"AskHR also achieved a 94% containment rate of common questions, has led to a 75% reduction in support tickets raised since 2016, and created more than 11.5 million employee interactions in 2024 alone.","sourceUrl":"https://www.ibm.com/case-studies/ibm-askhr"},{"kpi":"contact-deflection","value":75,"unit":"percent","qualifier":"exact","period":"HR support tickets raised, since 2016","claimant":"organization","quote":"AskHR also achieved a 94% containment rate of common questions, has led to a 75% reduction in support tickets raised since 2016, and created more than 11.5 million employee interactions in 2024 alone.","sourceUrl":"https://www.ibm.com/case-studies/ibm-askhr"},{"kpi":"cost-reduction","value":40,"unit":"percent","qualifier":"exact","period":"over four years","baseline":"HR team operational costs four years earlier","claimant":"organization","quote":"The AI agent helped contribute to a 40% reduction in the HR team’s operational costs over the past four years.","sourceUrl":"https://www.ibm.com/case-studies/ibm-askhr"},{"kpi":"employee-adoption","value":99,"unit":"percent","qualifier":"exact","period":"managers","claimant":"organization","quote":"The adoption of AskHR has reached 99% among managers.","sourceUrl":"https://www.ibm.com/case-studies/ibm-askhr"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.ibm.com/case-studies/ibm-askhr","title":"IBM AskHR","publisher":"IBM","date":"2025-08-21"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"ibm-askhr"},{"title":"IBM: AskIT, the internal IT support assistant","useCases":["it-service-desk-resolution-agent"],"organization":{"name":"IBM","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"IBM","role":"in-house"}],"summary":"IBM's CIO organization receives around 785,000 IT support tickets a year covering device setup, password resets, VPN problems and replacements. It launched AskIT, built on watsonx Assistant after an analysis of over 300,000 support tickets and trained on the 80% of IT issues the company faces most often. It covers more than 200 support topics in more than 40 languages for over 280,000 employees. In its first four months more than 133,000 employees used it.","stage":"scaled","year":2023,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"containment-rate","value":75,"unit":"percent","qualifier":"at-least","period":"first four months after release","claimant":"organization","quote":"Of the queries that were submitted, over 75% were resolved by the new assistant itself.","sourceUrl":"https://www.ibm.com/case-studies/cio-watsonx-askit"},{"kpi":"users-served","value":133000,"unit":"count","qualifier":"at-least","period":"first four months after release","claimant":"organization","quote":"In the four months since AskIT’s release, over 133,000 IBM employees used the tool at least once.","sourceUrl":"https://www.ibm.com/case-studies/cio-watsonx-askit"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.ibm.com/case-studies/cio-watsonx-askit","title":"Using AI to deliver a digital first employee experience","publisher":"IBM"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"ibm-askit-service-desk-assistant"},{"title":"US Immigration and Customs Enforcement: generative AI assisted resume screening","useCases":["recruitment-screening-and-interview-scheduling"],"organization":{"name":"U.S. Immigration and Customs Enforcement","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"AIS","role":"integrator"},{"name":"OpenAI","role":"model-provider"}],"summary":"ICE uses a generative AI tool, based on OpenAI's GPT-4, that compares each candidate's resume with the job requirements and returns a numerical score, a colour coded scoring group, related experience and missing experience, with a separate group for resumes it could not score. The aim is to apply the same criteria to every resume and shorten time to hire; human reviewed resumes are compared with the tool's output for validation. DHS classifies it as high impact and lists its impact assessment, independent review, monitoring and appeal process as still in progress.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DHS-2556, AI-Assisted Resume Screening Tool)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"ice-ai-assisted-resume-screening"},{"title":"Industrialized Construction Group: Microsoft 365 Copilot for proposal writing","useCases":["rfp-and-proposal-response-drafting"],"organization":{"name":"Industrialized Construction Group","anonymized":false,"country":"US","region":"north-america","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Industrialized Construction Group (ICG), a five person construction consultancy, uses Microsoft 365 Copilot to turn historical proposals into new proposals from a template instead of rewriting them for every customer, alongside marketing, onboarding and reporting tasks. Microsoft reports that Copilot reduced ICG's proposal response time by 80%.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":80,"unit":"percent","qualifier":"exact","period":"proposal response time","claimant":"vendor","quote":"Already, Copilot has helped ICG reduce proposal response time by 80%, so the team can focus on customers instead of paperwork.","sourceUrl":"https://www.microsoft.com/en/customers/story/23620-icg-microsoft-365-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23620-icg-microsoft-365-copilot","title":"ICG cuts proposal response time by 80% with Microsoft 365 Copilot","publisher":"Microsoft","date":"2025-04-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"icg-copilot-proposal-drafting"},{"title":"ICICI Lombard: claims copilot for health claim adjudicators","useCases":["health-prior-authorization-and-claims-adjudication"],"organization":{"name":"ICICI Lombard","anonymized":false,"country":"IN","region":"asia-pacific","industry":"insurance"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"ICICI Lombard built a claims copilot for its health claim adjudicators with Azure OCR, Azure AI Document Intelligence, Azure OpenAI and an in house model. It structures discharge summaries, lab reports and bills into diagnosis, clinical presentation, history, treatment and investigations, and compares the treatment with National Health Authority and disease treatment guidelines, so the adjudicator reads a summary instead of 20 or more pages. Microsoft reports that the time to process a single health claim fell by over 50%.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":50,"unit":"percent","qualifier":"at-least","period":"Time for an adjudicator to process a single health claim","claimant":"vendor","quote":"This solution has reduced the time for claims adjudicators to process a single health claim by over 50%.","sourceUrl":"https://www.microsoft.com/en-in/aifirstmovers/icici-lombard"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en-in/aifirstmovers/icici-lombard","title":"ICICI Lombard: Increasing productivity using a claims copilot","publisher":"Microsoft India","archivedUrl":"https://web.archive.org/web/20250317095025/https://www.microsoft.com/en-in/aifirstmovers/icici-lombard"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"icici-lombard-health-claims-copilot"},{"title":"Incore Bank: agentic AI proof of concept for business customer onboarding with Kyndryl and Google Cloud","useCases":["business-onboarding-and-ubo-discovery"],"organization":{"name":"Incore Bank","anonymized":false,"country":"CH","region":"europe","industry":"banking"},"vendors":[{"name":"Kyndryl","role":"integrator"},{"name":"Google Cloud","role":"platform"}],"summary":"Incore Bank, a Swiss bank that serves other banks, financial intermediaries and corporates rather than retail customers, completed a proof of concept with Kyndryl and Google Cloud that applies agentic AI, built on Kyndryl's Agentic AI Framework and Google's Gemini models, to the know your customer checks it runs on prospective and existing business clients. Several AI agents extract and validate customer information from documents, internal systems and external sources, identify risk factors, produce an explainable risk score and create an auditable decision record for compliance staff to review. Kyndryl reports the proof of concept reached up to 99 percent accuracy extracting data from onboarding documents; the further claim that the approach could cut onboarding time from months to days is stated as a demonstrated potential, not a measured result.","stage":"announced","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"accuracy","value":99,"unit":"percent","qualifier":"up-to","period":"automated extraction of data from customer onboarding documentation, during the proof of concept","claimant":"vendor","quote":"During the proof of concept, the solution achieved up to 99% accuracy in the automated extraction of data from customer onboarding documentation and demonstrated potential to reduce onboarding time from months to days.","sourceUrl":"https://www.kyndryl.com/in/en/about-us/news/2026/08/agentic-ai-incore-bank"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.kyndryl.com/in/en/about-us/news/2026/08/agentic-ai-incore-bank","title":"Kyndryl and Google Cloud advance agentic AI at Incore Bank","publisher":"Kyndryl","date":"2026-08-31"},{"url":"https://thepaypers.com/fraud-and-fincrime/news/kyndryl-incore-bank-google-cloud-test-agentic-ai-onboarding","title":"Kyndryl, Incore Bank, Google Cloud test agentic AI onboarding","publisher":"The Paypers","date":"2026-09-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"incore-bank-agentic-kyc-onboarding"},{"title":"Indosat Ooredoo Hutchison: AI driven energy efficiency across its radio network","useCases":["ran-energy-optimization"],"organization":{"name":"Indosat Ooredoo Hutchison","anonymized":false,"country":"ID","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Nokia","role":"platform"}],"summary":"After a pilot in the live network, Indosat Ooredoo Hutchison deployed Nokia Energy Efficiency, part of Nokia's Autonomous Networks portfolio, across its entire Nokia radio access network footprint in Sumatra, Kalimantan, Central and East Java. The software uses AI and machine learning on real time traffic patterns to adjust or shut idle radio equipment during low demand and includes thermal management to cut cooling energy. It is multi vendor and delivered as a service. No measured savings are published.","stage":"production","year":2025,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.nokia.com/newsroom/indosat-ooredoo-hutchison-and-nokia-partner-to-reduce-energy-demand-and-support-ai-powered-sustainable-operations/","title":"Indosat Ooredoo Hutchison and Nokia partner to reduce energy demand and support AI-powered, sustainable operations","publisher":"Nokia","date":"2025-07-07","archivedUrl":"https://web.archive.org/web/20260103112934/https://www.nokia.com/newsroom/indosat-ooredoo-hutchison-and-nokia-partner-to-reduce-energy-demand-and-support-ai-powered-sustainable-operations/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"indosat-ooredoo-hutchison-nokia-energy-efficiency"},{"title":"ING: agentic card payment in production with Worldline and Mastercard","useCases":["agentic-payment-initiation"],"organization":{"name":"ING","anonymized":false,"country":"NL","region":"europe","industry":"banking"},"vendors":[{"name":"Worldline","role":"platform"},{"name":"Mastercard","role":"platform"}],"summary":"Worldline, ING and Mastercard completed a live agentic card payment in production between an ING cardholder and a merchant in the Netherlands, on infrastructure that also runs in Belgium. A merchant's AI agent found concert tickets within a defined budget, presented a selection and paid only after the shopper gave explicit approval. The transaction carries identifiers that mark it as agentic, so ING as issuer keeps control through authentication and authorisation. The parties call it a pilot and name recurring transactions and delegated purchases within predefined parameters as next use cases. No outcome figures are disclosed.","stage":"pilot","year":2026,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.mastercard.com/news/europe/en/newsroom/press-releases/en/2026/worldline-ing-and-mastercard-complete-a-live-end-to-end-european-agentic-payment-in-production/","title":"Worldline, ING and Mastercard complete a live end-to-end European agentic payment in production","publisher":"Mastercard Newsroom"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ing-worldline-agentic-payment-in-production"},{"title":"Ingka Group (IKEA): Billie chatbot for customer enquiries","useCases":["first-line-contact-centre-agent"],"organization":{"name":"Ingka Group","anonymized":false,"country":"SE","region":"global","industry":"retail-and-ecommerce"},"vendors":[],"summary":"Ingka Group, the largest IKEA retailer, rolled out the AI chatbot Billie in its 2021 financial year to answer simpler customer enquiries around the clock. Between 2021 and 2023 Billie resolved about 47% of the enquiries it received. With the chatbot taking simpler enquiries, Ingka reskilled 8,500 call centre staff for remote interior design and remote selling, and sales through its remote customer meeting points reached EUR 1.3 billion in its 2022 financial year.","stage":"scaled","year":2021,"channels":[],"languages":[],"metrics":[{"kpi":"containment-rate","value":47,"unit":"percent","qualifier":"approximately","period":"2021 to 2023","claimant":"organization","quote":"Since the rollout of the solution in FY21[1], Billie has continued to provide value, and from 2021 to 2023 it resolved approximately 47% of customer enquiries it received, which translates to 3,2 million interactions solved by the chatbot and nearly EUR 13 million in savings thus far.","sourceUrl":"https://www.ingka.com/newsroom/ai-and-remote-selling-bring-ikea-design-expertise-to-the-many/"},{"kpi":"interactions-handled","value":3200000,"unit":"count","qualifier":"approximately","period":"resolved by the chatbot, 2021 to 2023","claimant":"organization","quote":"Since the rollout of the solution in FY21[1], Billie has continued to provide value, and from 2021 to 2023 it resolved approximately 47% of customer enquiries it received, which translates to 3,2 million interactions solved by the chatbot and nearly EUR 13 million in savings thus far.","sourceUrl":"https://www.ingka.com/newsroom/ai-and-remote-selling-bring-ikea-design-expertise-to-the-many/"},{"kpi":"cost-savings","value":13000000,"unit":"currency","currency":"EUR","qualifier":"approximately","period":"cumulative, 2021 to 2023","claimant":"organization","quote":"Since the rollout of the solution in FY21[1], Billie has continued to provide value, and from 2021 to 2023 it resolved approximately 47% of customer enquiries it received, which translates to 3,2 million interactions solved by the chatbot and nearly EUR 13 million in savings thus far.","sourceUrl":"https://www.ingka.com/newsroom/ai-and-remote-selling-bring-ikea-design-expertise-to-the-many/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.ingka.com/newsroom/ai-and-remote-selling-bring-ikea-design-expertise-to-the-many/","title":"AI and Remote Selling bring IKEA design expertise to the many","publisher":"Ingka Group","date":"2023-06-29"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"ingka-group-billie-customer-service-chatbot"},{"title":"Insilico Medicine: Pharma.AI discovery platform","useCases":["ai-drug-discovery-platform"],"organization":{"name":"Insilico Medicine","anonymized":false,"region":"global","industry":"pharma-and-life-sciences"},"vendors":[],"summary":"Insilico Medicine runs an end to end AI platform, Pharma.AI, that combines target discovery (PandaOmics), generative molecule design (Chemistry42) and translational and clinical support tools to move programs from a biological hypothesis to a nominated preclinical candidate. Its most advanced program, rentosertib (formerly ISM001-055 / INS018_055), a TNIK inhibitor for idiopathic pulmonary fibrosis whose target and molecule were both identified and designed with this platform, entered a Phase III clinical trial in July 2026 after a randomized Phase IIa trial published in Nature Medicine showed a dose dependent lung function signal. Beyond rentosertib, the company reports 31 preclinical candidate nominations from its pipeline, 13 of which received IND clearance and 8 of which have reached ongoing Phase I trials.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":60,"unit":"percent","qualifier":"approximately","period":"as stated in Insilico's June 2025 Nature Medicine publication release (22 nominated candidates, 2021 to 2024); the July 2026 release quoted above restates the same 12 to 18 month range without an updated candidate count","baseline":"traditional early stage drug discovery, typically 2.5 to 4 years to preclinical candidate nomination","claimant":"organization","quote":"While traditional early-stage drug discovery typically takes 2.5 to 4 years, Insilico has consistently reached preclinical candidate (PCC) nomination in an average of just 12 to 18 months, with only 60 to 200 molecules synthesized and tested per program.","sourceUrl":"https://insilico.com/news/xmjsn4l091-insilico-initiates-phase-iii-clinical-tr"}],"outcomeDisclosed":true,"sources":[{"url":"https://insilico.com/news/xmjsn4l091-insilico-initiates-phase-iii-clinical-tr","title":"Insilico Initiates Phase III Clinical Trial for Rentosertib, Its AI-Empowered TNIK Inhibitor for Idiopathic Pulmonary Fibrosis","publisher":"Insilico Medicine","date":"2026-07-07"},{"url":"https://insilico.com/news/tnrecuxsc1-insilico-announces-nature-medicine-publi","title":"Insilico Announces Nature Medicine Publication of Phase IIa Results of Rentosertib","publisher":"Insilico Medicine","date":"2025-06-03"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"insilico-medicine-pharma-ai-platform"},{"title":"UK Intellectual Property Office: AI check before filing a trade mark application","useCases":["permit-and-licence-application-processing"],"organization":{"name":"Intellectual Property Office","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Deloitte","role":"integrator"}],"summary":"Before filing, applicants can use the IPO's free \"Check if you could register your trade mark\" tool. Using text and image embeddings, it matches the goods and services the applicant enters to approved terms, searches for similar earlier marks (including logos) and flags elements that may not be acceptable, such as offensive words or protected symbols. It needs no personal details. Previously, applications were filed without meeting essential criteria and were rejected automatically. The tool gives indicative guidance only: every full application is still assessed and examined by the IPO. The tool receives an average 3,000 visits per month, and the IPO says it supports about 20% of filings.","stage":"production","year":2025,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/ipo-check-if-you-could-register-your-trade-mark-tool","title":"IPO: Check if you could register your trade mark tool","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-12-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"intellectual-property-office-trade-mark-pre-application-check"},{"title":"US Department of the Interior: AI internal controls testing across grant awards","useCases":["continuous-controls-testing"],"organization":{"name":"U.S. Department of the Interior","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Interior Department's Office of Grants Management built an in house set of AI tools, operational since April 2024, for three reviews of financial assistance awards that had been manual, inconsistent across bureaus and labour intensive: project descriptions, pre award eligibility validations in SAM.gov and budget submissions. The department says it needed a way to conduct internal controls testing, eligibility checks and budget reviews at scale. The tools produce automated scoring, flags for risks or inconsistencies, cross walks between budget documents and audit ready records aligned with internal control requirements. The 2024 inventory files the project description and SAM.gov work under internal controls testing, including a large language model that scores SAM.gov documents against the award date and a planned Azure app, to be built with Microsoft, that would let bureau staff test project descriptions themselves.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":29000,"unit":"count","qualifier":"at-least","period":"per year, financial assistance actions","claimant":"organization","quote":"Together, these outputs streamline oversight, strengthen regulatory compliance, and create a consistent, defensible documentation trail for more than 29,000 annual financial assistance actions.","sourceUrl":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv"}],"outcomeDisclosed":true,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DOI-0180, PGM Grants Utility Tool)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI inventory (Interior entries on internal controls testing)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"interior-department-grants-internal-controls-testing"},{"title":"Internal Revenue Service: AI Contract Document Toolbox for drafting and reviewing procurement documents","useCases":["procurement-contract-review"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"IRS procurement staff must produce extensive documentation for every contract file, and drafting and quality review were largely manual and slow, with some errors missed. Since July 2025 the IRS AI Contract Document Toolbox gives them a generative AI chat that drafts first versions of procurement documents, summarises and rewrites documents and extracts data, and that can be instructed to apply lessons learned and new procurement policy goals and to detect common past errors. Staff refine the drafts and vet documents together with the AI; the inventory states that the tool offers recommendations but does not approve contracts, and that agency officials make the final decisions. It records the use case as presumed high impact but determined not high impact. No outcome figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"irs-ai-contract-document-toolbox"},{"title":"Internal Revenue Service: AI voiceover generation for eLearning courses","useCases":["training-content-generation"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"IRS training developers used to record course narration with employees and microphone kits; with return to office mandates that was no longer possible. Since May 2024 they enter narration scripts into a web based AI voice tool that returns audio files for import into eLearning authoring applications. Scripts contain no personal or taxpayer information and use fictitious names and addresses. The IRS reports better quality, a wider choice of voices and much faster generation and revision of voiceovers, but publishes no figures.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry TREAS-IRS-65, AI Voiceover Generation for eLearning Development)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"irs-ai-voiceover-elearning"},{"title":"Internal Revenue Service: machine learning to select large partnership returns for examination","useCases":["tax-compliance-risk-scoring"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"In house (IRS data science and tax enforcement teams)","role":"in-house"}],"summary":"In September 2023 the IRS announced that it was expanding its Large Partnership Compliance programme, which it described as a pilot leveraging AI, to additional large partnerships. It said the returns had been selected with the help of AI by data scientists and tax enforcement experts who applied machine learning to identify compliance risk in partnership tax, general income tax and accounting, and international tax. The IRS said it would open examinations of 75 of the largest partnerships, each with more than USD 10 billion in assets on average, a segment that had seen little examination coverage. The IRS also said AI would help improve case selection so that fewer taxpayers face audits that end with no change. No outcome of the AI selected examinations is published in the release.","stage":"pilot","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.irs.gov/newsroom/irs-announces-sweeping-effort-to-restore-fairness-to-tax-system-with-inflation-reduction-act-funding-new-compliance-efforts","title":"IRS announces sweeping effort to restore fairness to tax system with Inflation Reduction Act funding (IR-2023-166)","publisher":"Internal Revenue Service","date":"2023-09-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"irs-large-partnership-audit-selection"},{"title":"IRS: neural machine translation for taxpayer content and case work","useCases":["public-service-translation","tax-questions-and-filing-assistant"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Amazon Web Services (Amazon Translate)","role":"platform"},{"name":"SYSTRAN","role":"platform"}],"summary":"The IRS Linguistic Policy, Tools and Services team uses a cloud machine translation application on AWS, with the IRS Publication 850 glossary of English and Spanish tax terms, to translate text and files between English and Spanish, Chinese, Korean and Vietnamese and speed up responses to taxpayers. Separately, staff use SYSTRAN neural translation, augmented with a domain dictionary and translation memories, to triage non English documents for relevance to case work and as a starting point for manual translation. Both appear in the federal AI inventory as in operation.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en","es","zh","ko","vi"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"irs-machine-translation"},{"title":"Internal Revenue Service: generative AI summaries of what each contract buys, for category management","useCases":["procurement-spend-classification"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"Since April 2025 the IRS runs an analytics hub for contract documents and contract spending data in which generative AI reads contract PDFs and writes a summary of the products or services bought under each agency contract into a table. The aim is better category management and a better view of what the agency buys. Outputs go to contracting officials for review, and the inventory classifies the use as not high impact because it is not the principal basis for significant decisions. No outcome figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry TREAS-IRS-61, Procurement Data Transparency, Reporting, & Decision Tracking)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"irs-procurement-contract-spend-summaries"},{"title":"US Internal Revenue Service: generative AI resolution notes and knowledge articles at the IT service desk","useCases":["support-knowledge-article-generation"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The IRS User and Network Services IT service desk runs a generative AI pilot, described as a limited production challenge, that summarizes incident case notes for warm handoffs, writes resolution notes from the actions taken, and generates complete knowledge base articles from incident and case records. The agency expects the effort to feed its existing knowledge review and publication processes, save time on handoffs, shorten the mean time to restore and support more self service, but publishes no results.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry TREAS-IRS-42, Ticket Management Generative AI Pilot)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"irs-service-desk-knowledge-article-generation"},{"title":"US Internal Revenue Service: Synthetic Data Engine for testing tax processing systems","useCases":["synthetic-test-data-generation"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The IRS runs an AI based synthetic data generator that builds a large population of synthetic people, households and businesses, ages them over time and correlates them with the socioeconomic patterns of US taxpayers. It outputs synthetic individual and business tax returns for several tax years, plus the reference files that seed test systems, so tax processing systems can be tested, including simulated fraud cases, without exposing taxpayer information. Automated checks run on every schema version to catch anomalies in the generated returns. No outcome figures are published.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry TREAS-85, Synthetic Data Engine)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"irs-synthetic-data-engine"},{"title":"IRS: voice bots and chatbots on taxpayer phone lines and IRS.gov","useCases":["tax-questions-and-filing-assistant"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Nuance (Microsoft)","role":"platform"},{"name":"eGain","role":"platform"}],"summary":"Since 2021 the IRS has put intent based voice bots on many toll free lines: payment plans and balance due (with authentication so taxpayers can set up or change a payment plan), Where's My Refund and amended return status, notice clarifications, Economic Impact Payments and the Advance Child Tax Credit. On IRS.gov, chatbots answer FAQs on refunds, identity theft, payments and relief. The inventory stresses that answers are not generated: the model classifies the question and returns content approved by the business owner, or routes the call to a live assistor.","stage":"scaled","year":2022,"channels":["voice","web-chat"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":3000000,"unit":"count","qualifier":"at-least","period":"calls answered by voice bots, cumulative to June 2022","claimant":"organization","quote":"To date, the voice bots have answered over 3 million calls.","sourceUrl":"https://www.irs.gov/newsroom/irs-expands-voice-bot-options-for-faster-service-less-wait-time"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.irs.gov/newsroom/irs-expands-voice-bot-options-for-faster-service-less-wait-time","title":"IRS expands voice bot options for faster service, less wait time (IR-2022-127)","publisher":"Internal Revenue Service","date":"2022-06-17"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"irs-taxpayer-voicebots-and-chatbots"},{"title":"IRS: machine learning to predict contractor performance risk (pilot)","useCases":["vendor-due-diligence"],"organization":{"name":"Internal Revenue Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The IRS reports a pilot, with a 2019 operational date, that researches supervised learning methods to assess contractor responsibility and predict whether a prospective vendor would perform successfully, identifying vendors at heightened risk of poor performance or non compliance. The models are trained on contractor and contract data from SAM.gov and USASpending.gov, and the risk assessments are delivered as spreadsheets or dashboards. The AI only recommends: contracting decisions go through several layers of review by agency officials. The seminal research and model training were done by IRS and US Navy personnel. It remains a pilot and no outcome figures are published.","stage":"pilot","year":2019,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry TREAS-IRS-9, Vendor Risk Analytics)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"irs-vendor-risk-analytics"},{"title":"Japan Exchange Group: AI in preliminary investigations of unfair trading","useCases":["market-abuse-surveillance-triage"],"organization":{"name":"Japan Exchange Group","anonymized":false,"country":"JP","region":"asia-pacific","industry":"capital-markets"},"vendors":[{"name":"NEC Corporation","role":"platform"},{"name":"Hitachi","role":"platform"}],"summary":"Japan Exchange Regulation and the Tokyo Stock Exchange put two machine learning systems from NEC and Hitachi into their market surveillance operations on 19 March 2018. The systems were supplied with the knowledge surveillance staff had used to evaluate irregular trading, and help staff finish the preliminary investigation of orders flagged by the criteria based surveillance systems faster, so they can focus on detailed investigations. The decision whether to investigate further stays with surveillance personnel. No outcome figures were published.","stage":"production","year":2018,"channels":["internal-tools"],"languages":["ja"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.jpx.co.jp/english/corporate/news/news-releases/0060/20180319-01.html","title":"Introduction of Artificial Intelligence to Market Surveillance Operations","publisher":"Japan Exchange Group","date":"2018-03-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"japan-exchange-group-ai-market-surveillance"},{"title":"J.B. Hunt: agentic AI for freight execution with Overroute","useCases":["freight-dispatch-and-load-matching-agent"],"organization":{"name":"J.B. Hunt Transport Services","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[{"name":"Overroute","role":"platform"}],"summary":"J.B. Hunt spent a year working with Overroute to design an agentic AI platform for freight execution, and then put its AI agents to work across all of J.B. Hunt's business units, on millions of loads. The agents work inside operators' existing systems to automate the coordination work behind every load: they read live data, surface exceptions, and support operators in managing customer communications, rather than replacing the operators' tools.","stage":"production","year":2026,"channels":["internal-tools","api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.jbhunt.com/our-company/newsroom/2026/07/overroute-launches-to-streamline-freight","title":"Overroute Launches with J.B. Hunt To Streamline Freight Execution for Enterprise Carriers","publisher":"J.B. Hunt Transport Services"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"jb-hunt-overroute-agentic-freight-execution"},{"title":"JetBlue: AI virtual agent and digital messaging support","useCases":["flight-disruption-and-rebooking-agent","first-line-contact-centre-agent"],"organization":{"name":"JetBlue","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[{"name":"ASAPP","role":"platform"}],"summary":"JetBlue moved its customer support to an AI platform from late 2019, opening messaging channels (Apple Messages for Business, Google Business Messaging, web and app chat, WhatsApp) with Spanish language support, a virtual agent that resolves routine requests and AI assistance for the crewmembers who handle the rest. In a January 2026 conference session published by the vendor, a JetBlue customer support leader described weather disruptions, when passengers ask for their options, and said the conversations crewmembers now handle (rebooking, refunds, alternatives weeks away) are multifaceted, which is why the airline has looked at AI that orchestrates several workflows. The ASAPP speaker in the same session warned against reading containment gains without checking whether customers still have the option to escalate.","stage":"scaled","year":2019,"channels":["web-chat","mobile-app","whatsapp","social-messaging"],"languages":["en","es"],"metrics":[{"kpi":"containment-rate","value":45,"unit":"percent","qualifier":"exact","period":"May 2023, virtual agent","claimant":"vendor","quote":"The integration of virtual agent experiences facilitated streamlined interactions and contributed to a remarkable 36% year-over-year growth in containment, with a 45% containment rate achieved in May 2023.","sourceUrl":"https://www.asapp.com/case-studies/jetblue"},{"kpi":"hours-saved","value":73000,"unit":"hours","qualifier":"exact","period":"Q1 2023 only (one quarter, not annualized)","claimant":"vendor","quote":"In Q1 2023 alone, this AI-driven efficiency translated into significant savings of 73,000 workforce hours.","sourceUrl":"https://www.asapp.com/case-studies/jetblue"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.asapp.com/case-studies/jetblue","title":"ASAPP X JetBlue | ASAPP","publisher":"ASAPP"},{"url":"https://www.asapp.com/blog/what-airlines-like-jetblue-teach-us-about-building-agentic-customer-experience","title":"What airlines like JetBlue teach us about building agentic customer experience","publisher":"ASAPP","date":"2026-03-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"jetblue-asapp-digital-customer-support"},{"title":"Jio: WhatsApp assistant for acquisition, porting, plans and care","useCases":["order-to-activation-and-esim-onboarding-assistant","plan-upgrade-and-sales-assistant"],"organization":{"name":"Reliance Jio","anonymized":false,"country":"IN","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Haptik","role":"platform"}],"summary":"Jio runs a WhatsApp assistant built with Haptik with more than 900 intents. It covers the 5G customer lifecycle end to end, from lead generation and buying a 5G device to porting into Jio, choosing and buying plans and customer care, with separate journeys for prepaid and postpaid, and sends proactive top up and recharge reminders. Haptik reports that the channel acquires 8,000 new Jio Fiber and 5G customers a day.","stage":"scaled","year":2023,"channels":["whatsapp"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.haptik.ai/resources/case-study/jio-digital-life","title":"Jio Transforms CX with Haptik","publisher":"Haptik","archivedUrl":"https://web.archive.org/web/20230528064657/https://www.haptik.ai/resources/case-study/jio-digital-life"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"jio-whatsapp-customer-lifecycle"},{"title":"JLL: AI lease and LOI abstraction with Cadastral","useCases":["lease-abstraction"],"organization":{"name":"JLL (Jones Lang LaSalle)","anonymized":false,"country":"US","region":"north-america","industry":"real-estate"},"vendors":[{"name":"Cadastral","role":"platform"}],"summary":"JLL's leasing brokerage business replaced manual lease and letter of intent abstraction with Cadastral, an AI platform later acquired by the legal AI company Legora. Brokers use it to generate lease and LOI abstracts within seconds, answer questions about complex leases through an integrated chat feature, and compare draft LOIs to each other, instead of relying on manual review. Cadastral reports the deployment now produces thousands of abstracts a year and saves JLL hundreds of thousands of dollars annually, without giving an exact figure for either.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cadastral.ai/case-studies/jll","title":"LL Leasing: AI-Driven Lease & LOI Abstraction | Cadastral","publisher":"Cadastral","archivedUrl":"https://web.archive.org/web/20260208044726/https://cadastral.ai/case-studies/jll"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"jll-cadastral-lease-abstraction"},{"title":"Johns Hopkins Medicine: Judy Reitz Capacity Command Center","useCases":["hospital-bed-and-staff-capacity-command-center"],"organization":{"name":"Johns Hopkins Medicine","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"GE Healthcare","role":"platform"}],"summary":"Johns Hopkins Medicine and GE Healthcare built the Judy Reitz Capacity Command Center, opened in January 2016, where staff control bed assignments for all patients within The Johns Hopkins Hospital and also manage transfers to and from four Johns Hopkins Medicine member hospitals, from one control room. Staff from admitting, transport and referral intake sit together, watching software that predicts patient volumes by shift, day and week, one screen that forecasts bed occupancy rates by department, and another that shows incoming patient transfers, in place of the pen and paper, whiteboards and markers used before. The organization also reports a reduction in transfer delays out of the operating room after a procedure. More than 20 other institutions, including Duke Health and Yale New Haven Health, have since built similar centers after visiting.","stage":"scaled","year":2016,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":38,"unit":"percent","qualifier":"exact","period":"reported at the center's fifth anniversary, 2021","claimant":"organization","quote":"A patient is assigned a bed 38% faster (or 3.5 hours faster) after a decision is made to admit him or her from the emergency department.","sourceUrl":"https://www.hopkinsmedicine.org/news/articles/2021/03/capacity-command-center-celebrates-5-years-of-improving-patient-safety-access"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.hopkinsmedicine.org/news/articles/2021/03/capacity-command-center-celebrates-5-years-of-improving-patient-safety-access","title":"Capacity Command Center Celebrates 5 Years of Improving Patient Safety, Access","publisher":"Johns Hopkins Medicine","archivedUrl":"https://web.archive.org/web/2026/https://www.hopkinsmedicine.org/news/articles/2021/03/capacity-command-center-celebrates-5-years-of-improving-patient-safety-access"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"johns-hopkins-capacity-command-center"},{"title":"J.P. Morgan Payments: Cash Flow Intelligence for corporate treasurers","useCases":["treasury-cash-flow-forecasting"],"organization":{"name":"JPMorgan Chase","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"JPMorgan Chase","role":"in-house"}],"summary":"J.P. Morgan Payments offers Cash Flow Intelligence, a machine learning tool in its J.P. Morgan Access platform that categorises a corporate client's payment flows and produces cash forecasts. In a Bloomberg interview relayed by CTMfile, the bank's head of data and analytics for wholesale payments said that about a year after launch roughly 2,500 corporate clients used it free of charge, and Bloomberg reported that some had cut manual work in categorising and visualising payment flows by nearly 90%, while liquidity decisions stay with people. Separately, the bank built a prototype conversational analytics assistant that lets treasurers query their payments data in plain language.","stage":"scaled","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":2500,"unit":"count","qualifier":"approximately","period":"corporate clients, about a year after launch","claimant":"organization","quote":"He observed that since the AI tool was introduced about a year ago, approximately 2,500 JPMorgan corporate customers are currently using the product for free.","sourceUrl":"https://ctmfile.com/story/ai-driven-cashflow-tool-helps-corporate-clients-cut-manual-work-by-90"},{"kpi":"productivity-gain","value":90,"unit":"percent","qualifier":"approximately","period":"some corporate clients, manual work in categorising and visualising payment flows","claimant":"organization","quote":"Dubbed, Cash Flow Intelligence, the artificial intelligence (AI)-aided cashflow management tool launched by the largest US bank, JPMorgan Chase & Co. has helped some of its corporate clients vastly reduce their manual work by nearly 90%, as was reported last week by Bloomberg.","sourceUrl":"https://ctmfile.com/story/ai-driven-cashflow-tool-helps-corporate-clients-cut-manual-work-by-90"}],"outcomeDisclosed":true,"sources":[{"url":"https://ctmfile.com/story/ai-driven-cashflow-tool-helps-corporate-clients-cut-manual-work-by-90","title":"AI-driven cashflow tool helps corporate clients cut manual work by 90%","publisher":"CTMfile","date":"2024-03-13"},{"url":"https://www.jpmorgan.com/insights/payments/data-intelligence/genai-virtual-analytics-assistant-treasury","title":"Conversational Analytics, Virtual Assistants & GenAI in Corporate Treasury","publisher":"J.P. Morgan","date":"2023-12-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"jpmorgan-cash-flow-intelligence"},{"title":"JPMorgan Chase: AI generated marketing copy with Persado","useCases":["marketing-content-compliance-copilot"],"organization":{"name":"JPMorgan Chase","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Persado","role":"platform"}],"summary":"After a pilot on Card and Mortgage marketing that started in 2016, JPMorgan Chase signed a five year, enterprise wide agreement in 2019 to use Persado's AI to write copy for direct response campaigns in personal banking, home lending and wealth management and for digital advertising. The pilot used Persado's Message Machine, a marketing language knowledge base of more than one million tagged and scored words and phrases, and the announcement does not describe how copy passes the bank's marketing compliance review.","stage":"production","year":2019,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.persado.com/press-releases/jpmorgan-chase-announces-five-year-deal-with-persado-for-ai-powered-marketing-capabilities/","title":"JPMorgan Chase Announces Five-Year Deal with Persado For AI-Powered Marketing Capabilities","publisher":"Persado","date":"2019-07-30"},{"url":"https://www.marketingdive.com/news/jpmorgan-chase-inks-5-year-deal-to-generate-marketing-copy-via-ai/559836/","title":"JPMorgan Chase inks 5-year deal to generate marketing copy via AI","publisher":"Marketing Dive"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"jpmorgan-chase-ai-marketing-copy"},{"title":"JPMorgan Chase: AI and technology in KYC processing for commercial and investment banking clients","useCases":["perpetual-kyc"],"organization":{"name":"JPMorgan Chase","anonymized":false,"country":"US","region":"global","industry":"banking"},"vendors":[{"name":"JPMorgan Chase","role":"in-house"}],"summary":"At its 2025 Investor Day, JPMorgan Chase's Commercial and Investment Bank said it was using AI and technology across the client journey, including onboarding and know your customer processing, and reported a substantial fall in the unit cost of KYC since 2022. The disclosure covers KYC processing in general rather than event driven review specifically.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"cost-reduction","value":40,"unit":"percent","qualifier":"exact","period":"KYC unit cost, 2022 to 2025","claimant":"organization","quote":"In KYC, for instance, we've seen a 40% reduction in unit cost since 2022 due to AI and technology enhancements.","sourceUrl":"https://www.jpmorganchase.com/content/dam/jpmc/jpmorgan-chase-and-co/investor-relations/documents/events/2025/jpmc-2025-investor-day/cib.pdf"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.jpmorganchase.com/content/dam/jpmc/jpmorgan-chase-and-co/investor-relations/documents/events/2025/jpmc-2025-investor-day/cib.pdf","title":"2025 Investor Day, Commercial and Investment Bank transcript","publisher":"JPMorgan Chase","date":"2025-05-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"jpmorgan-chase-kyc-unit-cost"},{"title":"J.P. Morgan: Coach AI for private client advisers","useCases":["wealth-advisor-knowledge-assistant","next-best-action-for-advisors"],"organization":{"name":"JPMorgan Chase","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"JPMorgan Chase","role":"in-house"}],"summary":"J.P. Morgan's private client advisers use an internal generative AI tool, Coach AI, to find research and content for client conversations more quickly. The firm's asset and wealth management chief executive credited its AI tools, which pull clients' trading patterns and anticipate their questions, with helping advisers respond to clients during the April 2025 market sell off. Its asset and wealth management chief information officer said advisers find the right information up to 95% faster. The firm's aim of growing adviser client books by 50% over three to five years is a target, not a result.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"search-time-reduction","value":95,"unit":"percent","qualifier":"up-to","period":"time to find information for a client conversation","claimant":"organization","quote":"Our advisers are finding the right information up to 95% faster - which means they spend less time searching and more time engaging in meaningful conversations with clients","sourceUrl":"https://www.aol.com/news/jpmorgan-says-ai-helped-boost-170825172.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.aol.com/news/jpmorgan-says-ai-helped-boost-170825172.html","title":"JPMorgan says AI helped boost sales, add clients in market turmoil","publisher":"Reuters via AOL","date":"2025-05-05"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"jpmorgan-coach-ai-advisers"},{"title":"J.P. Morgan: AI payment validation screening that cuts account validation rejections","useCases":["payment-investigations-and-exceptions"],"organization":{"name":"JPMorgan Chase","anonymized":false,"country":"US","region":"global","industry":"banking"},"vendors":[],"summary":"J.P. Morgan says it uses AI powered large language models for payment validation screening, which it describes as speeding up processing by reducing false positives and enabling better queue management. In November 2023 the bank said it had used this for more than two years and that account validation rejection rates had fallen by 15 to 20 percent, alongside lower fraud and a better customer experience. Screening at validation works on preventing payment exceptions rather than on investigating payments already in trouble.","stage":"production","year":2023,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.jpmorgan.com/insights/payments/payments-optimization/ai-payments-efficiency-fraud-reduction","title":"AI Boosting Payments Efficiency & Cutting Fraud | J.P. Morgan","publisher":"J.P. Morgan","date":"2023-11-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"jpmorgan-payment-validation-screening"},{"title":"J.P. Morgan AI Research: synthetic financial datasets for AI research and development","useCases":["synthetic-test-data-generation"],"organization":{"name":"JPMorgan Chase","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"JPMorgan Chase","role":"in-house"}],"summary":"J.P. Morgan AI Research develops generators for realistic synthetic datasets in financial services and makes them available to researchers on request. The published sets cover anti money laundering customer traces, retail customer journeys, payments data for fraud detection, market order books, synthetic documents for layout recognition and simulated equity market data. Its documented method computes metrics on the real data, builds and optionally calibrates a generator (statistical or agent based simulation) and then compares the metrics of the synthetic and the real data. No business outcome figures are published.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.jpmorgan.com/technology/artificial-intelligence/initiatives/synthetic-data","title":"Synthetic Data","publisher":"JPMorganChase"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"jpmorgan-synthetic-data-research"},{"title":"Kaiser Permanente: ambient AI scribes for physicians and clinicians","useCases":["ambient-clinical-documentation"],"organization":{"name":"Kaiser Permanente","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Abridge","role":"platform"}],"summary":"Kaiser Permanente made an ambient documentation tool from Abridge available to doctors and other clinicians at its 40 hospitals and more than 600 medical offices in August 2024, after a year of testing. With the patient's consent, the tool listens to the visit and drafts the clinical note, which the clinician reviews before it enters the record. An analysis by The Permanente Medical Group in Northern California, published in NEJM Catalyst, found that the scribes saved the equivalent of 1,794 working days in one year, and that time savings were concentrated among the most frequent users. The tool does not make decisions or recommendations about care.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":7260,"unit":"count","qualifier":"exact","period":"The Permanente Medical Group, 63 week evaluation period","claimant":"organization","quote":"AI scribes were used by 7,260 Permanente physicians in more than 2.5 million patient encounters during the evaluation period.","sourceUrl":"https://permanente.org/analysis-ai-scribes-save-physicians-time-improve-patient-interactions-and-work-satisfaction/"},{"kpi":"interactions-handled","value":2500000,"unit":"count","qualifier":"at-least","period":"The Permanente Medical Group, 63 week evaluation period, patient encounters","claimant":"organization","quote":"AI scribes were used by 7,260 Permanente physicians in more than 2.5 million patient encounters during the evaluation period.","sourceUrl":"https://permanente.org/analysis-ai-scribes-save-physicians-time-improve-patient-interactions-and-work-satisfaction/"}],"outcomeDisclosed":true,"sources":[{"url":"https://permanente.org/analysis-ai-scribes-save-physicians-time-improve-patient-interactions-and-work-satisfaction/","title":"Analysis: AI scribes save physicians time, improve patient interactions and work satisfaction","publisher":"The Permanente Medical Group","date":"2025-04-07"},{"url":"https://about.kaiserpermanente.org/news/press-release-archive/kaiser-permanente-improves-member-experience-with-ai-enabled-clinical-technology","title":"Kaiser Permanente improves member experience with AI-enabled clinical technology","publisher":"Kaiser Permanente","date":"2024-08-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"kaiser-permanente-ambient-ai-scribes"},{"title":"KDDI: AI detection and automatic recovery of silent cell degradations","useCases":["predictive-network-maintenance","network-fault-triage-copilot","autonomous-network-operations"],"organization":{"name":"KDDI","anonymized":false,"country":"JP","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Nokia","role":"platform"}],"summary":"KDDI deployed Nokia's AVA Performance Degradation Detection and Resolution (PDDR) solution nationwide to monitor its 4G and 5G radio network around the clock. The model detects performance degradations that raise no alarm, so called silent cells, classifies the likely root cause, and hands recoverable cases to KDDI's own recovery system, which tries to fix them automatically. Recovered cells feed back into the training data. KDDI started on 4G in 2019 and extended the system to its 5G NSA network in 2021.","stage":"scaled","year":2022,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.nokia.com/newsroom/nokia-ava-pddr-solution-deployed-by-kddi-to-boost-network-quality/","title":"Nokia AVA PDDR solution deployed by KDDI to boost network quality","publisher":"Nokia","date":"2022-12-08","archivedUrl":"https://web.archive.org/web/20231206182419/https://www.nokia.com/about-us/news/releases/2022/12/08/nokia-ava-pddr-solution-deployed-by-kddi-to-boost-network-quality/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"kddi-nokia-performance-degradation-detection"},{"title":"Keurig Dr Pepper: automated cash application","useCases":["cash-application-and-remittance-matching"],"organization":{"name":"Keurig Dr Pepper","anonymized":false,"country":"US","region":"north-america","industry":"manufacturing"},"vendors":[{"name":"HighRadius","role":"platform"}],"summary":"Keurig Dr Pepper, named on the source page by its then name Dr Pepper Snapple Group, brought payments processing in house and deployed HighRadius's cash application software to replace manual remittance aggregation and posting across its accounts receivable operation. The system gives real time visibility into payment statuses, automates invoice matching and deductions coding, and captures remittance information from multiple payment formats. The vendor also quotes Colleen Zdrojewski, then Vice President of Financial Services at Dr Pepper Snapple Group, saying financial services costs declined by $2.5 million while volume, quality and productivity increased; the page's own meta description and About text frame this as part of an annual run rate saving from bringing the previously outsourced payments processing in house together with the software, not a figure attributable to the matching software alone. The page calls the saving \"Saved in One Year with AI\" in a stat box caption, but nowhere describes machine learning or an AI matching method, so that label is the vendor's marketing framing, not a technical claim this record can verify.","stage":"production","year":2021,"channels":[],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":98,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"98% Payments Auto-Applied by the System","sourceUrl":"https://www.highradius.com/resources/case-studies/keurig-dr-pepper/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.highradius.com/resources/case-studies/keurig-dr-pepper/","title":"KDP | Saving $2.5M with Cash App Automation","publisher":"HighRadius","archivedUrl":"https://web.archive.org/web/20230925105936/https://www.highradius.com/resources/case-studies/keurig-dr-pepper/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"keurig-dr-pepper-cash-application-automation"},{"title":"Kin Insurance: masked and subsetted test databases for developers and QA","useCases":["synthetic-test-data-generation"],"organization":{"name":"Kin Insurance","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Tonic.ai","role":"platform"}],"summary":"Kin Insurance, a digital home insurer, sealed its production database off from development and now gives engineers and QA only subsetted, masked copies generated with Tonic, including differential privacy to protect people who stand out in the data. Developers use the subsets to fix bugs and build features, and QA uses the same data to check that releases behave as they did in the sandbox. The vendor reports faster data access and fewer security concerns but no quantified outcome beyond a database that can be pulled down in an hour or less.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.tonic.ai/case-study/kin-insurance-accelerates-growth-with-faster-more-secure-data-access-courtesy-of-tonic","title":"Kin Insurance speeds growth with fast, secure data access from Tonic.ai","publisher":"Tonic.ai","date":"2022-07-05"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"kin-insurance-masked-test-data"},{"title":"Kingfisher: AI invoice capture in the accounts payable shared service centre","useCases":["supplier-invoice-processing"],"organization":{"name":"Kingfisher","anonymized":false,"country":"GB","region":"europe","industry":"retail-and-ecommerce"},"vendors":[{"name":"Rossum","role":"platform"},{"name":"SAP","role":"platform"}],"summary":"Kingfisher's global business services centre in Poland, which handles accounts payable for six European countries (about 40,000 invoices a month), went live with Rossum's AI data capture in January 2021. Invoices arrive in central email inboxes; Rossum routes them by country and invoice type, extracts and validates the data, checks for duplicates and passes it to robots that index the invoices in SAP, where accountants still handle exceptions. The vendor reports that 60% of invoices now pass the data extraction step without any manual intervention before they reach SAP (it does not report an end to end touchless rate), that average indexing time fell from 5 minutes to 25 seconds, and that 14 full time employees moved to other invoice processing tasks.","stage":"scaled","year":2021,"channels":["email","internal-tools"],"languages":[],"metrics":[{"kpi":"handling-time-reduction","value":90,"unit":"percent","qualifier":"approximately","period":"time to index an invoice into SAP (from about 5 minutes to 25 seconds on average)","claimant":"vendor","quote":"Kingfisher's GBS cuts SAP invoice indexing time by 90% with end-to-end AP automation","sourceUrl":"https://rossum.ai/customer-stories/customer-story-kingfisher/"},{"kpi":"productivity-gain","value":80,"unit":"percent","qualifier":"exact","period":"manual data entry work of the accounts payable accountants, with 79% of invoice fields read automatically","claimant":"organization","quote":"That means 80% less manual work for our accountants, so the team can focus on vendor queries, exceptions, and higher-value work instead of just typing data all day.","sourceUrl":"https://rossum.ai/customer-stories/customer-story-kingfisher/"}],"outcomeDisclosed":true,"sources":[{"url":"https://rossum.ai/customer-stories/customer-story-kingfisher/","title":"Kingfisher's GBS cuts SAP invoice indexing time by 90% with end-to-end AP automation","publisher":"Rossum"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"kingfisher-ap-invoice-capture"},{"title":"Kinsale Capital: AI driven submission routing and company wide AI tools for underwriting and actuarial teams","useCases":["commercial-underwriting-submission-triage","insurance-pricing-and-actuarial-copilot"],"organization":{"name":"Kinsale Capital Group","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Kinsale Capital Group","role":"in-house"}],"summary":"Kinsale, a US excess and surplus lines insurer that sources about 95% of its premium through wholesale brokers, told investors in January 2026 that AI driven routing improves the accuracy of submission routing and underwriter productivity, alongside an average submission clearance time of 9 minutes. Its 2025 annual report says it gave every employee an enterprise AI tool licence in 2025, that use is most prevalent in its IT, actuarial and analytical teams with selective use in underwriting, and that it also uses internally developed agents. Kinsale does not attribute the clearance time to AI, so no metric is recorded.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1669162/000166916226000004/investorday-1x8x2026.htm","title":"Kinsale Capital Group Investor Day presentation (Form 8-K, Exhibit 99.1)","publisher":"Kinsale Capital Group via SEC EDGAR","date":"2026-01-08"},{"url":"https://www.sec.gov/Archives/edgar/data/1669162/000166916226000015/knsl-20251231.htm","title":"Kinsale Capital Group Form 10-K for 2025","publisher":"Kinsale Capital Group via SEC EDGAR","date":"2026-02-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"kinsale-ai-submission-routing"},{"title":"Klarna: AI assistant as the first line of customer service","useCases":["first-line-contact-centre-agent","card-dispute-and-chargeback-intake","order-status-and-returns-agent"],"organization":{"name":"Klarna","anonymized":false,"country":"SE","region":"europe","industry":"payments"},"vendors":[{"name":"OpenAI","role":"model-provider"}],"summary":"Klarna announced in February 2024 that its AI assistant built on OpenAI models had been live globally for a month as the first line of its customer service, handling refunds, returns, payment issues, cancellations and disputes in more than 35 languages across 23 markets. In 2025 the company said it had gone too far in replacing people and began recruiting human agents again so that customers can always reach a person; the assistant still handles the majority of inquiries. The record is useful precisely because it shows both the gain and the correction.","stage":"scaled","year":2024,"channels":["mobile-app"],"languages":["en","ar","fr"],"metrics":[{"kpi":"interactions-handled","value":2300000,"unit":"count","qualifier":"exact","period":"first month after launch","claimant":"organization","quote":"The AI assistant has had 2.3 million conversations, two-thirds of Klarna’s customer service chats","sourceUrl":"https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/"},{"kpi":"response-time-reduction","value":82,"unit":"percent","qualifier":"exact","period":"since launch, as reported in 2025","baseline":"before the AI assistant","claimant":"organization","quote":"Since launch, response times have improved by 82%, and Klarna has seen a 25% drop in repeat issues.","sourceUrl":"https://www.customerexperiencedive.com/news/klarna-reinvests-human-talent-customer-service-AI-chatbot/747586/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/","title":"Klarna AI assistant handles two-thirds of customer service chats in its first month","publisher":"Klarna","date":"2024-02-27"},{"url":"https://www.customerexperiencedive.com/news/klarna-reinvests-human-talent-customer-service-AI-chatbot/747586/","title":"Klarna changes its AI tune and again recruits humans for customer service","publisher":"CX Dive","date":"2025-05-09"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"klarna-ai-assistant-customer-service"},{"title":"Klarna: generative AI for marketing copy and imagery","useCases":["marketing-content-compliance-copilot"],"organization":{"name":"Klarna","anonymized":false,"country":"SE","region":"europe","industry":"payments"},"vendors":[{"name":"OpenAI","role":"model-provider"},{"name":"Midjourney","role":"model-provider"},{"name":"Adobe","role":"model-provider"},{"name":"Klarna","role":"in-house"}],"summary":"Klarna uses generative AI across marketing: an in house copywriting tool (Copy Assistant) for most of its copy, and image generation tools for campaign imagery, while reducing spend on external agencies for translation, production, CRM and social. Klarna says the faster image cycle includes checks for brand consistency, image quality and legal compliance. It attributes a share of its sales and marketing savings to AI.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"cost-savings","value":10000000,"unit":"currency","currency":"USD","qualifier":"approximately","period":"annualized, as of Q1 2024","claimant":"organization","quote":"AI is responsible for 37% of the cost savings, or about $10 million on an annualized basis.","sourceUrl":"https://www.klarna.com/international/press/ai-helps-klarna-cut-marketing-agency-spend-by-25-and-run-more-campaigns/"},{"kpi":"cycle-time-days","value":7,"unit":"days","qualifier":"exact","baseline":"image development cycle of 6 weeks before generative AI","claimant":"organization","quote":"Increased Efficiency and Creativity: Generated over 1,000 images in the first three months of 2024 using genAI, reducing the image development cycle from 6 weeks to just 7 days.","sourceUrl":"https://www.klarna.com/international/press/ai-helps-klarna-cut-marketing-agency-spend-by-25-and-run-more-campaigns/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.klarna.com/international/press/ai-helps-klarna-cut-marketing-agency-spend-by-25-and-run-more-campaigns/","title":"AI helps Klarna cut marketing agency spend by 25% and run more campaigns","publisher":"Klarna","date":"2024-05-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"klarna-generative-ai-marketing-production"},{"title":"KPMG: onboarding agent that guides new team members","useCases":["employee-onboarding-assistant"],"organization":{"name":"KPMG","anonymized":false,"region":"global","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Microsoft reports that KPMG used Microsoft AI to develop a team member onboarding agent that guides new hires and gives them templates and historical references. Microsoft describes it as part of KPMG's AI strategy and says it is meant to speed up onboarding and reduce follow up calls by 20%; that figure is stated as an aim, with no period, baseline or measured result. The member firm, whether the agent is live and the number of users are not stated.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en-us/worklab/agents-of-change","title":"Agents of change","publisher":"Microsoft WorkLab","date":"2025-03-05"},{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft","date":"2025-07-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"kpmg-new-hire-onboarding-agent"},{"title":"Labelbox: AI coding agent and Snyk rescans to clear a static analysis backlog","useCases":["software-vulnerability-remediation"],"organization":{"name":"Labelbox","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Snyk","role":"platform"},{"name":"Cursor","role":"platform"}],"summary":"Labelbox, a San Francisco AI data company of about 200 people, had a growing backlog of high severity static analysis findings. Its lead security engineer paired the Cursor coding agent with Snyk's MCP server: the agent pulled each finding, judged exploitability, proposed a fix and retried until a Snyk rescan passed, after which the fix was tested and went through QA. Snyk reports the backlog was cleared in two to three weeks; the engineer says it took a couple of weeks and estimates the work at a full calendar year with the old workflow.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://snyk.io/blog/from-two-years-to-two-weeks-how-labelbox-erased-its-security-debt-with-snyks/","title":"From Two Years to Two Weeks: How Labelbox Erased Its Security Debt with Snyk's AI-Accelerated Remediation","publisher":"Snyk","date":"2025-09-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"labelbox-snyk-ai-sast-backlog-remediation"},{"title":"LAQO: Pavle, a generative AI digital assistant for policy and claims questions on WhatsApp","useCases":["insurance-policy-servicing-agent"],"organization":{"name":"LAQO","anonymized":false,"country":"HR","region":"europe","industry":"insurance"},"vendors":[{"name":"Infobip","role":"platform"},{"name":"Microsoft","role":"model-provider"}],"summary":"LAQO, Croatia's first fully digital insurance provider (part of Croatian Insurance), built Pavle with Infobip on Azure OpenAI Service to answer customers 24/7 on WhatsApp in Croatian. The assistant is limited to insurance claims and general information about LAQO to reduce the risk of misleading answers, guides customers through reporting an accident after confirming cover, and transfers complex queries to a live agent. The vendor reports that Pavle handles 30% of customer queries and that 90% of queries are resolved within three to five messages; LAQO's head of digital sales and customer support says the contact centre now spends 10 percent less effort.","stage":"production","year":2023,"channels":["whatsapp"],"languages":["hr"],"metrics":[{"kpi":"containment-rate","value":30,"unit":"percent","qualifier":"exact","period":"share of customer queries handled by the assistant","claimant":"vendor","quote":"Today, LAQOs digital assistant is handling 30 percent of customer queries, freeing LAQO’s agents to focus on complex cases and customer acquisition.","sourceUrl":"https://www.infobip.com/customer/laqo"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.infobip.com/customer/laqo","title":"LAQO Insurance elevates support with Infobip's Gen-AI and Azure OpenAI partnership","publisher":"Infobip","date":"2023-11-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"laqo-pavle-digital-assistant"},{"title":"Leeds City Council: Money Information Centre chatbot","useCases":["benefits-eligibility-and-application-assistant"],"organization":{"name":"Leeds City Council","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Kainos Software Ltd","role":"integrator"},{"name":"Amazon Web Services (Amazon Bedrock)","role":"platform"},{"name":"Anthropic (Claude 3 Sonnet)","role":"model-provider"}],"summary":"Leeds City Council piloted a retrieval augmented chatbot on its Money Information Centre website, a site with information on money and support services. It answers only from website content, with Amazon Bedrock guardrails that keep it on topic, does not give financial advice and warns users that no human is on the other side. It does not hand over to a human agent; instead it gives the council's phone number and refers users who show distress to emergency helplines. An evaluation plan decides at the end of a six week pilot whether it delivers its benefits.","stage":"pilot","year":2025,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":83,"unit":"percent","qualifier":"exact","period":"chatbot responses scored at least 3 out of 5 in quality, in testing before launch with input from domain experts","claimant":"organization","quote":"On average, 83% of chatbot responses were scored at least a 3 out of 5 in quality.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/leeds-city-council-money-information-centre-chatbot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/leeds-city-council-money-information-centre-chatbot","title":"Leeds City Council: Money Information Centre Chatbot","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-06-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"leeds-city-council-money-information-centre-chatbot"},{"title":"Leeds City Council: Xylo Core for planning application validation and officer reports","useCases":["permit-and-licence-application-processing"],"organization":{"name":"Leeds City Council","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Xylo AI Ltd","role":"platform"}],"summary":"Leeds City Council's planning service is running a proof of concept pilot of Xylo Core, which ingests a planning application when an officer opens the case, redacts personal data, builds a tailored validation checklist, flags missing documents or wrong information, suggests the relevant site history, policies and constraints from GIS data, and drafts routine correspondence and report sections. Models from OpenAI, Anthropic and Google are called through APIs, with EU data residency and zero data retention. Officers must accept or reject every suggestion, sources and reasoning are shown for each, an audit log records what the officer kept, and the tool never recommends approval or refusal. The pilot starts with about ten officers on householder applications, the simplest type, with more types to be added once accuracy is proven.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":85,"unit":"percent","qualifier":"at-least","period":"initial testing across validation and policy recommendation prompts","claimant":"organization","quote":"Initial testing is delivering accuracy of 85%+ across the different prompts used for the different validation and policy recommendation workflows","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/leeds-city-council-xylo-core"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/leeds-city-council-xylo-core","title":"Leeds City Council: Xylo Core","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-11-27"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"leeds-city-council-xylo-planning-validation"},{"title":"Lemonade: AI Jim claims bot for first notice of loss and automated settlement","useCases":["claims-first-notice-of-loss-agent","claims-triage-and-straight-through-processing","claims-fraud-detection"],"organization":{"name":"Lemonade","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Lemonade","role":"in-house"}],"summary":"Lemonade's claims bot AI Jim takes the first notice of loss in a chat with the customer, pays or declines simple claims within seconds and assigns the claims it may not settle, or has concerns about, to human claims experts based on their specialty, workload and schedule. A separate system, Forensic Graph, uses machine learning to predict, detect and block fraud across the customer engagement. The annual report states that AI Jim took the first notice of loss without human intervention 96% of the time and that roughly 55% of claims were automated from start to finish, both as of December 31, 2025.","stage":"scaled","year":2025,"channels":["mobile-app","web-chat"],"languages":["en"],"metrics":[{"kpi":"containment-rate","value":96,"unit":"percent","qualifier":"exact","period":"First notice of loss taken without human intervention, as of December 31, 2025","claimant":"organization","quote":"AI Jim is our claims bot, and, as of December 31, 2025, 96% of the time, it is AI Jim that will take the first notice of loss from a Lemonade customer without human intervention","sourceUrl":"https://www.sec.gov/Archives/edgar/data/1691421/000169142126000016/lmnd-20251231.htm"},{"kpi":"automation-rate","value":55,"unit":"percent","qualifier":"approximately","period":"Share of claims automated end to end, as of December 31, 2025","claimant":"organization","quote":"As of December 31, 2025, roughly 55% of our claims were automated, resulting in instant or near-instant processing from start to finish.","sourceUrl":"https://www.sec.gov/Archives/edgar/data/1691421/000169142126000016/lmnd-20251231.htm"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1691421/000169142126000016/lmnd-20251231.htm","title":"Lemonade, Inc. Annual Report on Form 10-K for the fiscal year ended December 31, 2025","publisher":"U.S. Securities and Exchange Commission (EDGAR)","date":"2026-02-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"lemonade-ai-jim-claims-automation"},{"title":"Lemonade: AI Maya for quote and buy, CX.AI for policy service requests","useCases":["conversational-insurance-quote-and-buy","insurance-policy-servicing-agent"],"organization":{"name":"Lemonade","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Lemonade","role":"in-house"}],"summary":"Lemonade sells renters, homeowners, pet, car and life insurance through a chat with its bot AI Maya, which collects information, personalizes coverage, creates the quote and takes payment by asking a limited number of high impact questions. Its 2025 annual report says AI Maya and its APIs sell 98% of its policies. A second bot platform, CX.AI, resolves pre and post purchase requests such as coverage questions, adding a spouse, changing coverage amounts or payment methods and adding newly bought items, and handles over half of customer inquiries without human intervention.","stage":"scaled","year":2025,"channels":["mobile-app","web-chat"],"languages":["en"],"metrics":[{"kpi":"containment-rate","value":50,"unit":"percent","qualifier":"at-least","period":"customer inquiries handled by CX.AI without human intervention, as reported in the 10-K for 2025","claimant":"organization","quote":"Currently, over half of Lemonade’s customer inquiries are handled this way.","sourceUrl":"https://www.sec.gov/Archives/edgar/data/1691421/000169142126000016/lmnd-20251231.htm"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1691421/000169142126000016/lmnd-20251231.htm","title":"Lemonade, Inc. Form 10-K for 2025","publisher":"Lemonade via SEC EDGAR","date":"2026-02-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"lemonade-ai-maya-and-cx-ai"},{"title":"LinkedIn: SQL Bot text to SQL assistant in the DARWIN data platform","useCases":["governed-text-to-sql-analytics"],"organization":{"name":"LinkedIn","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"LinkedIn","role":"in-house"}],"summary":"LinkedIn's SQL Bot, built into its DARWIN data science platform, finds the right tables, writes the query and fixes errors so employees can answer data questions themselves; datasets keep their own access control lists, and the bot only supplies group credentials the user is entitled to. Domain experts certified and described hundreds of key tables, which improved retrieval, and the example queries come from notebooks certified by users and those meeting recency and reliability heuristics. The interface shows the retrieved tables and the query, and hundreds of employees across business units use it. In a user survey about 95% rated its query accuracy \"Passes\" or above (about 40% \"Very Good\" or \"Excellent\"); LinkedIn publishes no measured accuracy rate.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.linkedin.com/blog/engineering/ai/practical-text-to-sql-for-data-analytics","title":"Practical Text-to-SQL for Data Analytics","publisher":"LinkedIn Engineering","date":"2024-12-09"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"linkedin-sql-bot-text-to-sql"},{"title":"Lionbridge: GPT-4 on Azure OpenAI in translation and localization workflows","useCases":["marketing-and-product-content-localization"],"organization":{"name":"Lionbridge","anonymized":false,"country":"US","region":"global","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Lionbridge, a translation and localization provider with more than 6,500 employees, began building generative AI into its workflows with GPT-4 on Azure OpenAI in 2023, alongside its long standing use of machine translation. Employees use it to translate and localize content, build project glossaries and style guides, and flag sensitive content for human review before delivery. Within nine months the new workflows served hundreds of customers, and the company reports turnaround times up to 30% shorter.","stage":"scaled","year":2024,"channels":["internal-tools","api"],"languages":[],"metrics":[{"kpi":"processing-time-reduction","value":30,"unit":"percent","qualifier":"up-to","period":"project turnaround time","claimant":"organization","quote":"We’ve reduced turnaround times by up to 30% and cut days or hours off delivery.","sourceUrl":"https://www.microsoft.com/en/customers/story/1792260322207475324-lionbridge-technologies-azure-openai-service-other-en-united-states"},{"kpi":"users-served","value":500,"unit":"count","qualifier":"at-least","period":"client organizations using AI for content optimization, not only localization","claimant":"organization","quote":"We already have more than 500 customers using AI to help with content optimization, and we’re getting that content to market with high efficiency and quality","sourceUrl":"https://www.microsoft.com/en/customers/story/1792260322207475324-lionbridge-technologies-azure-openai-service-other-en-united-states"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1792260322207475324-lionbridge-technologies-azure-openai-service-other-en-united-states","title":"Lionbridge disrupts localization industry using Azure OpenAI Service and reduces turnaround times by up to 30%","publisher":"Microsoft","date":"2024-07-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"lionbridge-azure-openai-localization"},{"title":"Lloyds Banking Group: AI that classifies customer complaints","useCases":["complaints-handling-agent"],"organization":{"name":"Lloyds Banking Group","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[],"summary":"Lloyds Banking Group lists complaints handling and automation among the roughly 50 generative AI use cases it had live in 2025. In its 2025 results presentation the bank reports that complaint classification now takes 1 second instead of about 5 minutes. The bank attributes about GBP 50 million of P&L benefit in 2025 to its generative AI use cases as a whole and does not break out the share of the complaints use case.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"time-saved-per-task","value":5,"unit":"minutes","qualifier":"approximately","period":"in 2025","baseline":"About 5 minutes to classify a complaint before the AI use case; 1 second after","claimant":"organization","quote":"Outcome: Classification times reduced to 1 second (from c.5 mins)","sourceUrl":"https://web.archive.org/web/20260905205440id_/https://www.lloydsbankinggroup.com/assets/pdfs/investors/financial-performance/lloyds-banking-group-plc/2025/q4/2025-lbg-fy-presentation.pdf"}],"outcomeDisclosed":true,"sources":[{"url":"https://web.archive.org/web/20260905205440id_/https://www.lloydsbankinggroup.com/assets/pdfs/investors/financial-performance/lloyds-banking-group-plc/2025/q4/2025-lbg-fy-presentation.pdf","title":"2025 Results presentation (Internet Archive snapshot)","publisher":"Lloyds Banking Group","date":"2026-01-29"},{"url":"https://www.lloydsbankinggroup.com/assets/pdfs/investors/financial-performance/lloyds-banking-group-plc/2025/q4/2025-lbg-fy-presentation.pdf","title":"2025 Results presentation","publisher":"Lloyds Banking Group","date":"2026-01-29"},{"url":"https://www.lloydsbankinggroup.com/media/press-releases/2026/lloyds-banking-group/ai-driven-benefits-2026.html","title":"Lloyds Banking Group expects over £100 million in value from next‑generation AI in 2026","publisher":"Lloyds Banking Group","date":"2026-01-29","archivedUrl":"https://web.archive.org/web/20260129171626/https://www.lloydsbankinggroup.com/media/press-releases/2026/lloyds-banking-group/ai-driven-benefits-2026.html"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"lloyds-banking-group-ai-complaints-processing"},{"title":"Lloyds Banking Group: faster income verification in mortgage applications","useCases":["conversational-loan-application-intake"],"organization":{"name":"Lloyds Banking Group","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Lloyds Banking Group uses Vertex AI to scale machine learning work across more than 300 data scientists and AI developers. In the same entry Google Cloud reports that the bank cut income verification in mortgage applications from days to seconds and has put 18 generative AI systems into production. This is back office work at the application stage, not a customer facing assistant.","stage":"production","year":2025,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"lloyds-mortgage-income-verification"},{"title":"Loadsure: document classification and extraction for cargo insurance claims","useCases":["correspondence-triage-and-routing"],"organization":{"name":"Loadsure","anonymized":false,"country":"GB","region":"europe","industry":"insurance"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Loadsure, a London based insurtech for freight insurance, automated the intake of claim documents such as bills of lading, invoices and shipping documents with Google Cloud Document AI. Each incoming document is first classified and then sent to an extractor built for its type, which feeds the claims verification process; Gemini was later used for a similar extraction workflow elsewhere in the business. The blog post, written by Google Cloud and Loadsure staff, says work that took 30 to 60 minutes per claim now happens in near real time.","stage":"production","year":2024,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/blog/topics/financial-services/loadsure-data-drive-insurance-claims-AI-eliminates-manual-processing","title":"Can AI eliminate manual processing for insurance claims? Loadsure built a solution to find out","publisher":"Google Cloud Blog","date":"2024-11-05"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"loadsure-claims-document-classification"},{"title":"Loft: Gemini assistant for mortgage simulations over WhatsApp","useCases":["home-loan-assistant-and-prequalification"],"organization":{"name":"Loft","anonymized":false,"country":"BR","region":"latin-america","industry":"real-estate"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Loft, a real estate technology and financial services company active in Brazil and Mexico, moved its data to Google Cloud and built the Assistente Loft on Gemini. Brokers at the real estate agencies connected to Loft (about 9,000, according to Google Cloud) use it on WhatsApp, by text or voice, to compare home financing conditions from different banks in seconds, so a buyer knows their borrowing power before choosing a property. Google Cloud reports about 900 financing simulations a week on the company's WhatsApp channel.","stage":"production","year":2025,"channels":["whatsapp"],"languages":["pt"],"metrics":[{"kpi":"interactions-handled","value":900,"unit":"count","qualifier":"approximately","period":"per week, home financing simulations on WhatsApp","claimant":"vendor","quote":"Cerca de 900 simulações de financiamento por semana no WhatsApp e corretores de 9 mil imobiliárias conectados","sourceUrl":"https://cloud.google.com/customers/intl/pt-br/loft"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/intl/pt-br/loft","title":"Loft migra para a nuvem e adota IA para melhorar o dia a dia de corretores e clientes","publisher":"Google Cloud"},{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"loft-whatsapp-mortgage-simulations"},{"title":"Lowe's: Mylow home improvement virtual advisor","useCases":["conversational-shopping-assistant"],"organization":{"name":"Lowe's","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"OpenAI","role":"model-provider"}],"summary":"Lowe's launched Mylow in March 2025, a customer facing virtual advisor built with OpenAI that answers home improvement questions, gives project steps and links the project to product discovery, with recommendations that can be refined by budget and zip code. In May 2025 it rolled out Mylow Companion, built on the same foundation, to associates in more than 1,700 stores, so staff on the floor get the same product, project and inventory answers. No outcome figures were published.","stage":"production","year":2025,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://corporate.lowes.com/newsroom/press-releases/lowes-launches-first-ai-powered-home-improvement-virtual-advisor-03-05-25","title":"Lowe's Launches First AI-Powered Home Improvement Virtual Advisor","publisher":"Lowe's","date":"2025-03-05"},{"url":"https://corporate.lowes.com/newsroom/press-releases/lowes-deploys-first-scale-ai-assistant-retail-associates-05-05-25","title":"Lowe's deploys first at-scale AI assistant for retail associates","publisher":"Lowe's","date":"2025-05-05"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"lowes-mylow-virtual-advisor"},{"title":"Loyola University's Schreiber Center: AI HVAC optimization and automated emissions reduction","useCases":["building-energy-optimization"],"organization":{"name":"Loyola University Chicago","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[{"name":"BrainBox AI","role":"platform"},{"name":"WattTime","role":"platform"}],"summary":"Loyola University's Schreiber Center, a LEED Gold certified, 10 storey mixed use building that houses Loyola's Quinlan School of Business in Chicago, ran a year long proof of concept combining BrainBox AI's HVAC optimization with WattTime's Automated Emissions Reduction signal, which pre cools the building during low emissions events, when the local grid has surplus renewable energy, and lets it drift when the grid relies more on fossil fuels. The project was carried out with UC Berkeley's Center for the Built Environment, and BrainBox AI points to a fuller study in an ASHRAE guide on the role of grid interactivity in decarbonization.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"energy-savings","value":10,"unit":"percent","qualifier":"exact","period":"over the year long proof of concept","claimant":"vendor","quote":"we were able to achieve a 10% savings in HVAC-related energy and a 10% reduction in HVAC-related CO2e emissions through our AI for HVAC solution","sourceUrl":"https://brainboxai.com/en/case-studies/leveraging-ai-and-aer-for-sustainable-excellence-loyola-universitys-schreiber-center"}],"outcomeDisclosed":true,"sources":[{"url":"https://brainboxai.com/en/case-studies/leveraging-ai-and-aer-for-sustainable-excellence-loyola-universitys-schreiber-center","title":"Leveraging AI and AER for Sustainable Excellence: Loyola University's Schreiber Center","publisher":"BrainBox AI","archivedUrl":"https://web.archive.org/web/20240514171126/https://brainboxai.com/en/case-studies/leveraging-ai-and-aer-for-sustainable-excellence-loyola-universitys-schreiber-center"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"loyola-university-schreiber-center-brainbox-ai"},{"title":"LTIMindtree: natural language test case creation pushed into test management","useCases":["requirements-to-test-case-generation"],"organization":{"name":"LTIMindtree","anonymized":false,"country":"IN","region":"asia-pacific","industry":"technology"},"vendors":[{"name":"Tricentis","role":"platform"}],"summary":"LTIMindtree (LTM), a technology services company and Tricentis Tosca implementation partner, piloted Tricentis Agentic Test Creation, in which testers describe the test they need in plain English and the system drafts the test case and pushes it straight into the qTest test management tool, replacing manual spreadsheet uploads. The pilot ran in an SAP GUI staging environment; its first phase covered 10 to 12 test cases across three complexity tiers, with several testers running identical cases to check consistency. Separately, Google Cloud lists an LTM Video Intelligence Agent that converts recorded videos into BDD test cases, without published results.","stage":"pilot","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":67,"unit":"percent","qualifier":"exact","period":"Pilot, time to create a low complexity test case","baseline":"30 minutes per test case written manually","claimant":"vendor","quote":"Low complexity test cases dropped from 30 minutes to 10 minutes—a 67% reduction","sourceUrl":"https://www.tricentis.com/case-studies/ltm-accelerates-testing-transformation-agentic-ai"},{"kpi":"processing-time-reduction","value":83,"unit":"percent","qualifier":"up-to","period":"Pilot, time to create a high complexity test case","baseline":"2 hours per test case written manually","claimant":"vendor","quote":"High complexity test cases fell from 2 hours to 20-30 minutes—up to 83% time savings","sourceUrl":"https://www.tricentis.com/case-studies/ltm-accelerates-testing-transformation-agentic-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.tricentis.com/case-studies/ltm-accelerates-testing-transformation-agentic-ai","title":"LTM accelerates testing transformation with Tricentis agentic AI","publisher":"Tricentis","date":"2026-05-20"},{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ltimindtree-agentic-test-case-creation"},{"title":"Lufthansa Group: self service AI agents for rebookings, refunds and alternative flights","useCases":["flight-disruption-and-rebooking-agent","first-line-contact-centre-agent"],"organization":{"name":"Lufthansa Group","anonymized":false,"country":"DE","region":"europe","industry":"travel-and-hospitality"},"vendors":[{"name":"Cognigy (NiCE)","role":"platform"}],"summary":"During the pandemic, when passengers flooded call centres to change or cancel flights, Lufthansa Group replaced its in house chatbot with a conversational AI platform and built self service AI agents that manage rebookings, check alternative flights, give travel information and process refunds. The agents run on the airline websites and through SMS links that open a self service chat, with multilingual support and real time translation, and are used to absorb peaks such as strikes.","stage":"scaled","year":2020,"channels":["web-chat","sms"],"languages":[],"metrics":[{"kpi":"interactions-handled","value":16000000,"unit":"count","qualifier":"approximately","period":"per year","claimant":"vendor","quote":"By leveraging AI-driven Self-Service Agents, the airline managed to significantly increase its interaction capacity, handling about 16 million conversations throughout the year with AI, with peak days seeing up to 375,000 interactions.","sourceUrl":"https://www.cognigy.com/en/case-study/lufthansa"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.cognigy.com/en/case-study/lufthansa","title":"Lufthansa | NiCE Cognigy","publisher":"Cognigy"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"lufthansa-group-self-service-ai-agents"},{"title":"Lumen Technologies: Copilot summaries of past sales interactions and account research","useCases":["sales-call-coaching-and-crm-update","business-connectivity-quoting-and-service-assistant","outbound-sales-prospecting-agent"],"organization":{"name":"Lumen Technologies","anonymized":false,"country":"US","region":"north-america","industry":"telecommunications"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Lumen's sellers use Microsoft Copilot to summarize past sales interactions, gather recent news, identify business challenges and industry trends, and suggest next steps for an account. Microsoft reports that this work took up to four hours per seller and that Lumen cut it to 15 minutes in 2024. Lumen's projected annual value of USD 50 million is a projection and is not recorded as a result.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"time-saved-per-task","value":225,"unit":"minutes","qualifier":"up-to","period":"summarizing past interactions and researching an account, per seller","baseline":"up to four hours per seller before","claimant":"vendor","quote":"This process traditionally took up to four hours per seller. In 2024, Lumen reduced that time to just 15 minutes, projecting annual time savings worth USD50 million.","sourceUrl":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft","date":"2025-07-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"lumen-copilot-sales-account-research"},{"title":"M-DAQ Global: AI driven Know Your Business checks for business customer onboarding","useCases":["digital-onboarding-assistant","business-onboarding-and-ubo-discovery"],"organization":{"name":"M-DAQ Global","anonymized":false,"country":"SG","region":"asia-pacific","industry":"payments"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"M-DAQ Global, a fintech group headquartered in Singapore that specialises in foreign exchange and cross border payments, runs a Know Your Business compliance solution on Vertex AI and Google Kubernetes Engine that uses natural language processing to automate the verification work behind onboarding business customers. The vendor reports a productivity gain of 30 times and shorter onboarding times.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"productivity-gain","value":30,"unit":"multiplier","qualifier":"exact","period":"compliance tasks, as reported by the vendor","claimant":"vendor","quote":"The natural language processing-based system automates compliance tasks and improves productivity by 30 times, reducing onboarding times and eliminating manual bottlenecks in customer verification.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"},{"url":"https://www.m-daq.com/about-us","title":"About Us","publisher":"M-DAQ Global"},{"url":"https://cloud.google.com/customers/mdaq","title":"Improve productivity by 30x with Vertex AI automation and data analysis","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"m-daq-global-kyb-onboarding"},{"title":"Macquarie Bank: AI fraud protection alerts and self service search","useCases":["fraud-alert-confirmation"],"organization":{"name":"Macquarie Bank","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Macquarie Bank uses Google Cloud AI for proactive fraud protection and digital self service. Google Cloud reports that the bank cut false positive alerts for client protection and that its help centre search sent more users to self service. The source does not say which channels carry the alerts.","stage":"production","year":2025,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"false-positive-reduction","value":40,"unit":"percent","qualifier":"exact","period":"not stated","claimant":"vendor","quote":"Macquarie Bank uses Google Cloud AI to enable efficient and proactive fraud protection and digital self-service capabilities — their Help Centre Search directed 38% more users towards self-service and they reduced false positive alerts for client protection by 40%.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"macquarie-bank-fraud-protection-and-self-service"},{"title":"City of Madrid (Madrid Destino): VisitMadridGPT multilingual visitor assistant","useCases":["public-service-translation","citizen-information-assistant","ai-visitor-and-tour-guide"],"organization":{"name":"Madrid Destino","anonymized":false,"country":"ES","region":"europe","industry":"government"},"vendors":[{"name":"iUrban","role":"integrator"},{"name":"Microsoft (Azure OpenAI Service)","role":"model-provider"}],"summary":"Madrid Destino, the city's municipal tourism office, runs VisitMadridGPT, a virtual assistant that answers visitors in more than 95 languages from the city's official, expert curated tourism site, available when physical offices are closed. The city analyses the questions to find the most requested topics and adjust its website content. According to the story, Madrid attracted 10.6 million visitors in 2023.","stage":"production","year":2024,"channels":["web-chat"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://customers.microsoft.com/en-us/story/1831036907720807463-esmadrid-azure-openai-service-national-government-en-spain","title":"Transforming tourism in Madrid with Azure OpenAI Service","publisher":"Microsoft Customer Stories"},{"url":"https://www.eldiariodemadrid.es/articulo/madrid/visitmadridgpt-asistente-virtual-turistas-madrid/20240410141511074094.html","title":"VisitMadridGPT, el asistente virtual para los turistas en Madrid","publisher":"El Diario de Madrid","date":"2024-04-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"madrid-destino-visitmadridgpt"},{"title":"MAIF: one pricing platform for data preparation, modelling and geographic models","useCases":["insurance-pricing-and-actuarial-copilot"],"organization":{"name":"MAIF","anonymized":false,"country":"FR","region":"europe","industry":"insurance"},"vendors":[{"name":"Akur8","role":"platform"}],"summary":"MAIF, a French mutual insurer, moved its pricing workflow from separate SAS and Python tools into Akur8, where data preparation, model building and geographic modelling happen in one place. The vendor says its machine learning explores thousands of variable combinations in parallel to find the most predictive features while actuaries keep control of the final selection. MAIF's pricing teams now manage several hundred databases of up to 30 million rows and several thousand models and versions on the platform; no time saving is quantified.","stage":"scaled","year":2026,"channels":["internal-tools"],"languages":["fr"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.akur8.com/success-stories/how-maif-cut-modeling-time-and-built-thousands-of-models-with-akur8","title":"How MAIF cut modeling time and built thousands of models with Akur8","publisher":"Akur8"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"maif-akur8-pricing-models"},{"title":"Majid Al Futtaim Retail: generative AI analysis of customer feedback for Carrefour","useCases":["customer-feedback-analysis"],"organization":{"name":"Majid Al Futtaim Retail","anonymized":false,"country":"AE","region":"middle-east","industry":"retail-and-ecommerce"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Majid Al Futtaim Retail, which runs Carrefour in the Middle East, Africa and Central Asia, built a text analytics solution called \"Excellence\" on Azure OpenAI Service. Before it, the marketing team manually processed 60,000 to 70,000 customer responses a week. The solution captures customer emotion and categorizes feedback by aspects such as delivery, quality, hygiene and checkout queue times, generating actionable insights for improvement. Microsoft reports that feedback processing fell from seven days to three hours; the same story quotes the Chief Digital Officer as saying it now takes three to four minutes.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"cycle-time-days","value":3,"unit":"hours","qualifier":"exact","baseline":"seven days of manual feedback processing","claimant":"vendor","quote":"The company saved USD1 million annually, cut feedback processing time from seven days to three hours, and improved geographic targeting, boosting efficiency with AI-driven solutions.","sourceUrl":"https://www.microsoft.com/en/customers/story/21059-majid-al-futtaim-retail-azure-open-ai-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/21059-majid-al-futtaim-retail-azure-open-ai-service","title":"From seven days to three minutes: Majid Al Futtaim boosts customer centricity with Azure OpenAI Service","publisher":"Microsoft Customer Stories","date":"2025-01-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"majid-al-futtaim-customer-feedback-analysis"},{"title":"Majid Al Futtaim: Mastercard Agent Pay pilot for VOX Cinemas tickets in the UAE","useCases":["agentic-payment-initiation"],"organization":{"name":"Majid Al Futtaim","anonymized":false,"country":"AE","region":"middle-east","industry":"retail-and-ecommerce"},"vendors":[{"name":"Mastercard","role":"platform"}],"summary":"Mastercard launched Agent Pay in the UAE together with Majid Al Futtaim as a pilot in which AI assistants help users find products and complete purchases on their behalf, starting with movie tickets at VOX Cinemas. Mastercard frames the collaboration as a pathway to broader adoption of agentic commerce across the region. No transaction volumes or outcome figures are disclosed.","stage":"pilot","year":2025,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.mastercard.com/news/eemea/en/perspectives/en/2025/bringing-agentic-payments-to-life-in-the-uae/","title":"Bringing agentic payments to life in the UAE","publisher":"Mastercard Newsroom"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"majid-al-futtaim-mastercard-agent-pay-pilot"},{"title":"Manulife: GenAI sales enablement for agents, advisors and distribution partners","useCases":["insurance-broker-and-agent-assistant"],"organization":{"name":"Manulife","anonymized":false,"country":"CA","region":"global","industry":"insurance"},"vendors":[{"name":"Manulife","role":"in-house"}],"summary":"Manulife's annual reports describe a generative AI sales tool in Singapore and Japan that drafts personalized engagement strategies for agents from customer needs, demographics and transaction history. By 2025 it had deployed GenAI sales enablement across nine markets and all four operating segments, with engagement insights, automated email drafting and real time coaching; in Hong Kong it launched AI Sales Pro for agents, and in Indonesia, Singapore and Japan AI assistants give agents faster access to product and policy information. In U.S. Retirement (part of its Global Wealth and Asset Management segment), Manulife says an AI sales enablement solution reduced time spent on information searches and tripled the number of sales opportunities compared with 2024.","stage":"scaled","year":2025,"channels":["internal-tools","mobile-app"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1086888/000108688826000003/a2025annualmdareport.htm","title":"Manulife 2025 Annual Management's Discussion and Analysis (Form 40-F, Exhibit 99.2)","publisher":"Manulife via SEC EDGAR","date":"2026-02-11"},{"url":"https://www.sec.gov/Archives/edgar/data/1086888/000108688825000054/a2024annualmdareport.htm","title":"Manulife 2024 Annual Management's Discussion and Analysis (Form 40-F, Exhibit 99.2)","publisher":"Manulife via SEC EDGAR","date":"2025-02-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"manulife-genai-sales-enablement"},{"title":"Manulife: generative AI document summarization and preliminary assessments in life underwriting","useCases":["life-underwriting-medical-record-summarization"],"organization":{"name":"Manulife","anonymized":false,"country":"CA","region":"global","industry":"insurance"},"vendors":[{"name":"Manulife","role":"in-house"},{"name":"Munich Re","role":"platform"}],"summary":"Manulife's 2024 annual report says generative AI in Singapore automates document digitization and summarization in underwriting, improving the accuracy of underwriting decisions and reducing processing time for policy applications, and that in the US it expanded the use of electronic health records and used generative AI to automate preliminary underwriting assessments. In 2025 it partnered with Munich Re Life US on alitheia, an AI driven risk assessment platform, raising the instant underwriting decision eligibility limit from US$3 million to US$5 million. No time or accuracy figures are disclosed.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1086888/000108688825000054/a2024annualmdareport.htm","title":"Manulife 2024 Annual Management's Discussion and Analysis (Form 40-F, Exhibit 99.2)","publisher":"Manulife via SEC EDGAR","date":"2025-02-19"},{"url":"https://www.sec.gov/Archives/edgar/data/1086888/000108688826000003/a2025annualmdareport.htm","title":"Manulife 2025 Annual Management's Discussion and Analysis (Form 40-F, Exhibit 99.2)","publisher":"Manulife via SEC EDGAR","date":"2026-02-11"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"manulife-generative-ai-life-underwriting"},{"title":"Manulife: AI document processing for health and dental claims","useCases":["health-prior-authorization-and-claims-adjudication"],"organization":{"name":"Manulife","anonymized":false,"country":"CA","region":"north-america","industry":"insurance"},"vendors":[],"summary":"In its first quarter 2026 report to shareholders, Manulife says it enhanced online claims processing for its Affinity health and dental business in Canada with AI driven document processing for the majority of claims that were processed manually, which improved processing speed and paid customers faster. No figures were published.","stage":"production","year":2026,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1086888/000108688826000044/q12026reporttoshareholderl.htm","title":"Manulife Financial Corporation first quarter 2026 report to shareholders (Form 6-K)","publisher":"Manulife Financial Corporation (via SEC EDGAR)","date":"2026-05-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"manulife-health-dental-claims-document-ai"},{"title":"Mapbox: AI assistant across docs, app and support for developers","useCases":["developer-api-integration-assistant"],"organization":{"name":"Mapbox","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Kapa.ai","role":"platform"}],"summary":"Mapbox, whose maps and location APIs are used by millions of developers, connected an AI assistant to its public docs, SDKs and API references plus private support knowledge, refreshed weekly. The same assistant sits in the docs, a dedicated developer assistant page, the logged in account app, Discord and the support desk, where it drafts answers with sources for support engineers. Mapbox reports a 30% monthly reduction in support tickets, and Kapa.ai reports that 20% of questions are answered in languages other than English, without naming them.","stage":"scaled","year":2026,"channels":["web-chat","social-messaging","agent-desktop"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":175000,"unit":"count","qualifier":"at-least","period":"per year","claimant":"vendor","quote":"That consistency drives a 30% reduction in monthly support tickets from paid users, 175,000+ questions answered yearly (40,000+ support hours saved), and a 26% increase in questions asked to Kapa, with 20% answered in non-English.","sourceUrl":"https://www.kapa.ai/customer-examples/mapbox"},{"kpi":"hours-saved","value":40000,"unit":"hours","qualifier":"at-least","period":"per year","claimant":"vendor","quote":"That consistency drives a 30% reduction in monthly support tickets from paid users, 175,000+ questions answered yearly (40,000+ support hours saved), and a 26% increase in questions asked to Kapa, with 20% answered in non-English.","sourceUrl":"https://www.kapa.ai/customer-examples/mapbox"},{"kpi":"contact-deflection","value":30,"unit":"percent","qualifier":"exact","period":"monthly support tickets","claimant":"organization","quote":"We've seen fantastic results with Kapa.ai, recently achieving a 30% monthly reduction in support tickets and significant productivity gains for our Technical Support Engineers.","sourceUrl":"https://www.kapa.ai/customer-examples/mapbox"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.kapa.ai/customer-examples/mapbox","title":"How Mapbox reduced monthly support tickets by 30% across web, Discord, and support","publisher":"Kapa.ai"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"mapbox-docs-ai-assistant"},{"title":"Markel UK: AI supported triage and routing of broker submissions with Cytora","useCases":["commercial-underwriting-submission-triage"],"organization":{"name":"Markel","anonymized":false,"country":"GB","region":"europe","industry":"insurance"},"vendors":[{"name":"Cytora","role":"platform"}],"summary":"Before the change, the most senior underwriter in each Markel UK team triaged every incoming submission against appetite, and underwriters rekeyed risk data into several systems and pulled third party data by hand. With Cytora, broker submissions are digitized, enriched with external data, prioritized against Markel's underwriting and distribution strategy and routed to the right specialist as decision ready risks, with data flowing into the CRM and policy systems. The vendor reports a 113% productivity uplift (GWP per FTE) and a quote turnaround SLA for strategic partners cut from 24 hours to 2 hours.","stage":"production","year":2023,"channels":["email","internal-tools"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":113,"unit":"percent","qualifier":"exact","period":"gross written premium per underwriting FTE","claimant":"vendor","quote":"an uplift of 113% in productivity (GWP/FTE) in their underwriting teams","sourceUrl":"https://www.cytora.com/risk-flow-center/blog/case-study-markel-records-113-productivity-increase-in-its-underwriting-team-following-cytora-partnership"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.cytora.com/risk-flow-center/blog/case-study-markel-records-113-productivity-increase-in-its-underwriting-team-following-cytora-partnership","title":"Markel uses Cytora and achieves +100% productivity uplift to fuel growth","publisher":"Cytora","date":"2023-09-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"markel-cytora-submission-triage"},{"title":"Mashreq: name screening and adverse media alert adjudication with Silent Eight","useCases":["sanctions-screening-adjudication","pep-and-adverse-media-screening"],"organization":{"name":"Mashreq","anonymized":false,"country":"AE","region":"middle-east","industry":"banking"},"vendors":[{"name":"Silent Eight","role":"platform"}],"summary":"Mashreq selected Silent Eight in May 2024 to automate the adjudication of name screening and adverse media alerts related to sanctions and anti money laundering requirements. Under the plan, false positives are to be investigated and closed quickly and potential true positives escalated to Mashreq analysts. The announcement is a multi year partnership; no results are disclosed.","stage":"announced","year":2024,"channels":["internal-tools"],"languages":["en","ar"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.silenteight.com/blog/mashreq-partners-with-silent-eight-for-compliance-alert-adjudication","title":"Mashreq Partners with Silent Eight for Compliance Alert Adjudication","publisher":"Silent Eight","date":"2024-05-07"},{"url":"https://www.prnewswire.com/news-releases/mashreq-partners-with-silent-eight-for-compliance-alert-adjudication-302137038.html","title":"Mashreq Partners with Silent Eight for Compliance Alert Adjudication","publisher":"PR Newswire","date":"2024-05-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"mashreq-silent-eight-alert-adjudication"},{"title":"Mass General Brigham: CodaMetrix coding automation for professional services","useCases":["medical-coding-automation"],"organization":{"name":"Mass General Brigham","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"CodaMetrix","role":"platform"}],"summary":"CodaMetrix, a company spun out of Mass General Brigham, uses machine learning and natural language processing on the clinical record to translate clinical notes into procedure and diagnosis codes automatically and reduce the workload of human coders. In CodaMetrix's February 2023 funding release, Mass General Brigham's vice president of physician revenue cycle services cited \"a 70% reduction in manual labor\" and a 59% reduction in denials due to coding as \"our outcomes\", without saying which work, period or sites the figures cover. Mass General Brigham physician organizations also invested in the company.","stage":"production","year":2023,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":70,"unit":"percent","qualifier":"exact","period":"manual labor, scope and period not stated","claimant":"organization","quote":"Our outcomes — a 70% reduction in manual labor — 59% reduction in denials due to coding, and a significant increase in cost savings — is the proof.\" said Michael Mercurio, Vice President of Physician Revenue Cycle Services at Mass General Brigham.","sourceUrl":"https://www.prnewswire.com/news-releases/codametrix-closes-55m-series-a-to-autonomously-power-medical-coding-boost-health-system-revenue-cycles-301756940.html"},{"kpi":"error-reduction","value":59,"unit":"percent","qualifier":"exact","period":"claim denials due to coding","claimant":"organization","quote":"Our outcomes — a 70% reduction in manual labor — 59% reduction in denials due to coding, and a significant increase in cost savings — is the proof.\" said Michael Mercurio, Vice President of Physician Revenue Cycle Services at Mass General Brigham.","sourceUrl":"https://www.prnewswire.com/news-releases/codametrix-closes-55m-series-a-to-autonomously-power-medical-coding-boost-health-system-revenue-cycles-301756940.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.prnewswire.com/news-releases/codametrix-closes-55m-series-a-to-autonomously-power-medical-coding-boost-health-system-revenue-cycles-301756940.html","title":"CodaMetrix Closes $55M Series A to Autonomously Power Medical Coding, Boost Health System Revenue Cycles","publisher":"CodaMetrix (PR Newswire)","date":"2023-02-27"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"mass-general-brigham-codametrix-coding-automation"},{"title":"Massachusetts DESE: AI scoring of MCAS essays and the 2025 rescoring","useCases":["automated-scoring-of-written-responses"],"organization":{"name":"Massachusetts Department of Elementary and Secondary Education","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"Massachusetts scores MCAS essays with AI trained on human scored examples of each score point, with humans giving 10 percent of AI scored essays a second read. The problem surfaced in summer 2025, when preliminary results went to districts. In Lowell, the example NBC10 Boston reports, a teacher found that some of her third grade students' scores did not add up and the issue went to district leaders; district leaders notified DESE. The state's testing contractor, Cognia, found that roughly 1,400 essays (of about 750,000 MCAS essays statewide) had not received the correct scores, which DESE attributed to a temporary technical issue in the process; the essays were rescored, 145 districts were notified and district data was corrected in August. DESE points to the discrepancy period in which districts can report issues with preliminary results as a check on accuracy.","stage":"scaled","year":2024,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.nbcboston.com/investigations/ai-grading-massachusetts-mcas/3807392/","title":"'No rhyme or reason': AI grading issue affects hundreds of MCAS essays","publisher":"NBC Boston","date":"2025-09-11"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"massachusetts-dese-mcas-ai-essay-scoring"},{"title":"Master Trust Bank of Japan: AI data capture for inbound financial documents","useCases":["correspondence-triage-and-routing"],"organization":{"name":"The Master Trust Bank of Japan","anonymized":false,"country":"JP","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Rossum","role":"platform"}],"summary":"The Master Trust Bank of Japan, a trust bank specialising in asset servicing, uses Rossum's intelligent document processing to read inbound Japanese financial documents such as trade instructions, dividend notices and tax returns, and pass the extracted data to its own robotic process automation. The deployment grew to more than 90 document types used by 10 to 15 departments and 100,000 documents a year. The vendor reports that manual workload fell by 75% and that full processing now takes under 10 minutes instead of up to 1.5 hours per document.","stage":"scaled","year":2023,"channels":["internal-tools","api"],"languages":["ja"],"metrics":[{"kpi":"productivity-gain","value":75,"unit":"percent","qualifier":"exact","period":"manual data capture workload","claimant":"vendor","quote":"The Rossum IDP platform has reduced the manual workload by 75% and improved the time spent verifying and validating documents.","sourceUrl":"https://rossum.ai/customer-stories/master-trust-bank-of-japan/"},{"kpi":"interactions-handled","value":100000,"unit":"count","qualifier":"exact","period":"documents per year","claimant":"vendor","quote":"To date, Rossum has scaled to handle 100,000 documents per year for MTBJ.","sourceUrl":"https://rossum.ai/customer-stories/master-trust-bank-of-japan/"}],"outcomeDisclosed":true,"sources":[{"url":"https://rossum.ai/customer-stories/master-trust-bank-of-japan/","title":"Customer Story - Master Trust Bank of Japan - Rossum.ai","publisher":"Rossum"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"master-trust-bank-of-japan-financial-document-capture"},{"title":"Mastercard: AI interview scheduling for candidates","useCases":["recruitment-screening-and-interview-scheduling"],"organization":{"name":"Mastercard","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[],"summary":"Mastercard uses an AI tool to coordinate and reschedule candidate interviews with hiring managers, letting candidates complete scheduling when it suits them. The company says candidates now see their interviews scheduled nearly 90% faster. The same article also describes Unlocked, Mastercard's internal talent marketplace, which is covered by a separate evidence record.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":90,"unit":"percent","qualifier":"approximately","claimant":"organization","quote":"As a result, candidates now see their interviews getting scheduled nearly 90% faster","sourceUrl":"https://www.mastercard.com/news/perspectives/2024/at-the-inflection-of-ai-and-hr-how-we-re-equipping-employees-for-the-ai-era/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.mastercard.com/news/perspectives/2024/at-the-inflection-of-ai-and-hr-how-we-re-equipping-employees-for-the-ai-era/","title":"At the inflection of AI and HR: How we're equipping employees for the AI era","publisher":"Mastercard Newsroom","archivedUrl":"https://web.archive.org/web/20260616181554/https://www.mastercard.com/news/perspectives/2024/at-the-inflection-of-ai-and-hr-how-we-re-equipping-employees-for-the-ai-era"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"mastercard-ai-interview-scheduling"},{"title":"Mastercard: Consumer Fraud Risk scores for account to account payments in the UK","useCases":["scam-payment-interception","real-time-fraud-scoring"],"organization":{"name":"Mastercard","anonymized":false,"country":"GB","region":"europe","industry":"payments"},"vendors":[{"name":"Mastercard","role":"in-house"}],"summary":"Mastercard's Consumer Fraud Risk uses AI and its view of account to account payment flows to give UK banks a real time risk score on outgoing payments, so a bank can intervene before money reaches a scammer. Mastercard says it is live with 10 large UK banks, with NatWest among the first users. The only outcome it cites is a TSB extrapolation of what the UK could save if all banks matched TSB's performance, which is a projection, not a measured result.","stage":"production","year":2023,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.mastercard.com/news/press/2024/april/mastercard-transforms-the-fight-against-scams-with-latest-ai-tech/","title":"Mastercard transforms the fight against scams with latest AI tech","publisher":"Mastercard","date":"2024-04-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"mastercard-consumer-fraud-risk"},{"title":"Mastercard: Decision Intelligence Pro, generative AI in real time card transaction scoring","useCases":["real-time-fraud-scoring"],"organization":{"name":"Mastercard","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[{"name":"Mastercard","role":"in-house"}],"summary":"Mastercard's Decision Intelligence scores card transactions for fraud risk in real time on behalf of issuing banks, and Mastercard says it already helps banks score and approve 143 billion transactions a year. Decision Intelligence Pro adds generative AI techniques that assess the relationships between entities around a transaction and return an improved score in less than 50 milliseconds. Mastercard also published detection and false positive figures from its initial modelling and own analysis before launch; these are not measured production results and are not recorded as metrics.","stage":"announced","year":2024,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.mastercard.com/news/press/2024/february/mastercard-supercharges-consumer-protection-with-gen-ai/","title":"Mastercard supercharges consumer protection with gen AI","publisher":"Mastercard Newsroom","date":"2024-02-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"mastercard-decision-intelligence-pro"},{"title":"Mastercard: AI matching in the Unlocked internal talent marketplace","useCases":["internal-talent-marketplace-matching"],"organization":{"name":"Mastercard","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[],"summary":"Mastercard uses AI in Unlocked, its internal talent marketplace, to match employees to short term projects, volunteering, open roles, mentors and learning pathways, based on the skills they have and the skills they want to build. The company says 90% of its workforce is on the platform, with 500,000 project hours delivered, and that the skills data shows where it has gaps so it can plan learning paths or hiring. The same article describes an AI interview scheduling tool.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.mastercard.com/news/perspectives/2024/at-the-inflection-of-ai-and-hr-how-we-re-equipping-employees-for-the-ai-era/","title":"At the inflection of AI and HR: How we're equipping employees for the AI era","publisher":"Mastercard Newsroom","archivedUrl":"https://web.archive.org/web/20260616181554/https://www.mastercard.com/news/perspectives/2024/at-the-inflection-of-ai-and-hr-how-we-re-equipping-employees-for-the-ai-era"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"mastercard-unlocked-talent-marketplace"},{"title":"Mattel: generative AI classification of consumer feedback across reviews, social media and contact centre","useCases":["customer-feedback-analysis"],"organization":{"name":"Mattel","anonymized":false,"country":"US","region":"north-america","industry":"manufacturing"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Mattel built a feedback classification system on BigQuery, Vertex AI and Gemini that analyses millions of consumer feedback points from customer reviews, social media and the contact centre. Google Cloud reports that analysis time fell from a month to a single minute and that data processing capacity rose a hundredfold.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"cycle-time-days","value":1,"unit":"minutes","qualifier":"exact","baseline":"a month per analysis before the system","claimant":"vendor","quote":"The system analyzes millions of feedback points from a diverse range of sources (customer reviews, social media, contact center) in seconds — delivering a staggering 100x increase in data processing capacity and slashing analysis times from a month to a single minute.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"mattel-feedback-classification"},{"title":"Mercari US: agentic AI assistant for IT support","useCases":["it-service-desk-resolution-agent"],"organization":{"name":"Mercari US","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Moveworks","role":"platform"}],"summary":"Mercari US, the US arm of the online marketplace, deployed a Moveworks AI assistant in Slack in July 2021. It resolves IT issues such as password resets, email group changes, device troubleshooting and software provisioning. Moveworks reports that the assistant resolves most issues without the service desk and that most employees now go to the assistant first rather than messaging the IT team.","stage":"scaled","year":2021,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":74,"unit":"percent","qualifier":"at-least","period":"as reported in the undated case study","claimant":"vendor","quote":"Today, the agentic AI Assistant handles over 74% of issues completely autonomously.","sourceUrl":"https://www.moveworks.com/us/en/customers/mercari-reduced-it-ticket-volume-moveworks-conversational-ai"},{"kpi":"employee-adoption","value":94,"unit":"percent","qualifier":"exact","period":"employees who go to the assistant first, as reported in the undated case study","claimant":"vendor","quote":"As a result, the vast majority of employees — 94% — reach out to the Assistant first when they have questions instead of Slacking the IT team directly.","sourceUrl":"https://www.moveworks.com/us/en/customers/mercari-reduced-it-ticket-volume-moveworks-conversational-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.moveworks.com/us/en/customers/mercari-reduced-it-ticket-volume-moveworks-conversational-ai","title":"Mercari US Reduced IT Tickets By 74% With AI","publisher":"Moveworks"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"mercari-it-support-assistant"},{"title":"Mercedes-Benz: generative AI knowledge feature in the MBUX Voice Assistant and announced Gemini based conversational navigation","useCases":["in-car-ai-voice-assistant"],"organization":{"name":"Mercedes-Benz Group","anonymized":false,"country":"DE","region":"europe","industry":"automotive"},"vendors":[{"name":"Microsoft","role":"model-provider"},{"name":"Google Cloud","role":"platform"}],"summary":"After a US beta of ChatGPT in 2023, Mercedes-Benz brought a general knowledge function to series production vehicles in December 2024: the MBUX Voice Assistant runs a Microsoft Bing search and answers in natural language with ChatGPT through Azure OpenAI Service, keeps the dialogue for up to one hour for follow up questions, stores voice data anonymized in its own cloud and uses a risk assessment tool to reduce harmful answers. Mercedes-Benz announced it as a free update for over three million vehicles in German and English. In January 2025 it announced Gemini based conversational search for points of interest with Google Cloud's Automotive AI Agent, starting in the new CLA.","stage":"scaled","year":2024,"channels":["voice"],"languages":["de","en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.conceptcarz.com/a56010/human-like-conversations-mercedes-benz-enabled-mbux-voice-assistant-ai-driven-knowledge-feature.aspx","title":"Human-like conversations with your Mercedes-Benz: Enabled by MBUX Voice Assistant and AI-driven knowledge feature","publisher":"Mercedes-Benz (press release republished by conceptcarz.com)","date":"2024-12-18"},{"url":"https://www.googlecloudpresscorner.com/2025-01-13-Mercedes-Benz-and-Google-Partner-on-AI-powered-Conversational-Search-within-Navigation-Systems","title":"Mercedes-Benz and Google Partner on AI-powered Conversational Search within Navigation Systems","publisher":"Google Cloud and Mercedes-Benz","date":"2025-01-13"},{"url":"https://group.mercedes-benz.com/innovation/digitalisation/connectivity/car-voice-control-with-chatgpt.html","title":"Mercedes-Benz takes in-car voice control to a new level with ChatGPT","publisher":"Mercedes-Benz Group","date":"2023-06-16","archivedUrl":"https://web.archive.org/web/2026/https://group.mercedes-benz.com/innovation/digitalisation/connectivity/car-voice-control-with-chatgpt.html"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"mercedes-benz-mbux-generative-ai-voice-assistant"},{"title":"Merck: generative AI platform for first drafts of clinical study reports","useCases":["clinical-and-regulatory-document-drafting"],"organization":{"name":"Merck & Co.","anonymized":false,"country":"US","region":"north-america","industry":"pharma-and-life-sciences"},"vendors":[{"name":"Merck & Co.","role":"in-house"}],"summary":"Merck built an internal generative AI platform that combines table preprocessing with large language model authoring to produce first drafts of clinical study reports, under the oversight of qualified medical writers. Merck reports that first drafts now take three to four days instead of two to three weeks, that the time to a fully human reviewed first draft fell from an average of 180 hours to 80 hours, and that draft errors halved. The first live reports built on the platform were submitted in 2025, and Merck said it was scaling the platform across its late phase pipeline.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"error-reduction","value":50,"unit":"percent","qualifier":"exact","period":"CSR first drafts, across multiple studies","claimant":"organization","quote":"Increased the quality of CSR drafts – as measured by reducing the number of errors by 50% – in categories such as data, messaging, citations, terminology and typography.","sourceUrl":"https://www.merck.com/news/merck-expands-innovative-internal-generative-ai-solutions-helping-to-deliver-medicines-to-patients-faster/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.merck.com/news/merck-expands-innovative-internal-generative-ai-solutions-helping-to-deliver-medicines-to-patients-faster/","title":"Merck Expands Innovative Internal Generative AI Solutions Helping to Deliver Medicines to Patients Faster","publisher":"Merck & Co.","date":"2025-06-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"merck-clinical-study-report-generation"},{"title":"Merge: AI agents that research key accounts and draft tailored outreach for reps","useCases":["outbound-sales-prospecting-agent"],"organization":{"name":"Merge","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Clay","role":"platform"}],"summary":"Merge, which sells integration products into many B2B industries, used Clay to enrich and categorize more than 50,000 accounts in Salesforce with an industry and a suggested use case, so reps no longer research each account to find the angle. A separate AI agent follows more than 1,500 enterprise accounts every week for launches, partnerships and organizational changes and, when it finds a relevant update, sends the account owner an alert with the context and a drafted email ready for the rep to send. Clay reports that SDRs got more than 10 hours back every week, that response rates climbed 20% (enriched accounts compared with accounts not enriched) and that enterprise meetings booked rose 15%.","stage":"production","year":2026,"channels":["email","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.clay.com/customers/merge","title":"How Merge books 15% more enterprise meetings by researching 1,500 key accounts weekly with Clay","publisher":"Clay"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"merge-clay-account-research-outreach"},{"title":"Merkur Versicherung: daily synthetic copies of health insurance customer data","useCases":["synthetic-test-data-generation"],"organization":{"name":"Merkur Versicherung AG","anonymized":false,"country":"AT","region":"europe","industry":"insurance"},"vendors":[{"name":"MOSTLY AI","role":"platform"}],"summary":"The Merkur Innovation Lab, the innovation arm of the Austrian insurer Merkur Versicherung, runs an automated pipeline that extracts its active customer data (about 600,000 rows and 55 columns) from an Oracle database, has MOSTLY AI generate a synthetic version through a REST call and writes the result to a PostgreSQL database with Apache Airflow every day. The synthetic health data feeds internal analysis dashboards and is used to explore data sharing with third parties. The vendor reports that time to data fell from one month to one day.","stage":"production","year":2023,"channels":["api","internal-tools"],"languages":[],"metrics":[{"kpi":"cycle-time-days","value":1,"unit":"days","qualifier":"exact","period":"Time from data request to usable synthetic data","baseline":"About one month before the automated pipeline","claimant":"vendor","quote":"The end-to-end automated workflow has cut Merkur’s time-to-data from 1-month, to 1-day.","sourceUrl":"https://mostly.ai/blog/insurance-innovation-powered-by-synthetic-data"}],"outcomeDisclosed":true,"sources":[{"url":"https://mostly.ai/blog/insurance-innovation-powered-by-synthetic-data","title":"Insurance innovation: 3 use cases powered by synthetic data in health insurance","publisher":"MOSTLY AI","date":"2023-08-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"merkur-versicherung-synthetic-health-data"},{"title":"Meta: AI assisted root cause analysis for reliability investigations","useCases":["aiops-incident-triage"],"organization":{"name":"Meta","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Meta (Llama)","role":"in-house"}],"summary":"Meta's reliability investigation tooling uses a heuristic retriever (code ownership, the runtime code graph of impacted systems) to narrow thousands of recent code changes to a few hundred, then a fine tuned Llama 2 model ranks them to the five most likely root causes when an investigation is opened. Meta reports the result from backtesting on historical investigations in its web monorepo and stresses that responders must be able to verify the suggestions.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":42,"unit":"percent","qualifier":"exact","period":"backtesting on historical investigations, web monorepo","baseline":"root cause among the top five suggested code changes","claimant":"organization","quote":"Based on exhaustive backtesting, with historical investigations and the information available at their start, 42% of these investigations had the root cause in the top five suggested code changes.","sourceUrl":"https://engineering.fb.com/2024/06/24/data-infrastructure/leveraging-ai-for-efficient-incident-response/"}],"outcomeDisclosed":true,"sources":[{"url":"https://engineering.fb.com/2024/06/24/data-infrastructure/leveraging-ai-for-efficient-incident-response/","title":"Leveraging AI for efficient incident response","publisher":"Engineering at Meta","date":"2024-06-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"meta-ai-assisted-root-cause-analysis"},{"title":"Meta: TestGen-LLM had 73% of its generated unit tests accepted for production during internal test events","useCases":["developer-coding-assistant"],"organization":{"name":"Meta","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Meta","role":"in-house"}],"summary":"Meta built TestGen-LLM, a tool that uses large language models to draft and improve unit tests. During Instagram and Facebook test events, it improved 11.5% of all classes it was applied to, and 73% of its recommended test improvements were accepted by Meta software engineers for production deployment.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"accuracy","value":73,"unit":"percent","qualifier":"exact","claimant":"organization","quote":"During Meta's Instagram and Facebook test-a-thons, it improved 11.5% of all classes to which it was applied, with 73% of its recommendations being accepted for production deployment by Meta software engineers.","sourceUrl":"https://arxiv.org/abs/2402.09171"}],"outcomeDisclosed":true,"sources":[{"url":"https://arxiv.org/abs/2402.09171","title":"Automated Unit Test Improvement using Large Language Models at Meta","publisher":"Meta (arXiv, FSE 2024)","date":"2024-02-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"meta-testgen-llm-unit-tests"},{"title":"Microsoft Azure: Triangle multi agent incident triage","useCases":["aiops-incident-triage"],"organization":{"name":"Microsoft","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Microsoft","role":"in-house"}],"summary":"Azure uses the Triangle System to triage incidents with AI agents. In local triage, one agent per engineering team, built on the team's historical incidents and troubleshooting guides, accepts or rejects an incoming incident on the team's behalf and can recommend the team it should move to; a global triage layer coordinates the agents to route incidents. Local triage has been in production since mid 2024 and was live for six teams in January 2025, with more than 15 onboarding; Microsoft reports triage accuracy and a time to mitigate reduction for one team as initial results.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":90,"unit":"percent","qualifier":"exact","period":"initial results, as of January 2025","claimant":"organization","quote":"The initial results are promising, with agents achieving 90% accuracy and one team saw a reduction in their TTM of 38%, significantly reducing the impact to customers.","sourceUrl":"https://azure.microsoft.com/en-us/blog/optimizing-incident-management-with-aiops-using-the-triangle-system/"},{"kpi":"mttr-reduction","value":38,"unit":"percent","qualifier":"exact","period":"one team, initial results","baseline":"time to mitigate (TTM) before the agents","claimant":"organization","quote":"The initial results are promising, with agents achieving 90% accuracy and one team saw a reduction in their TTM of 38%, significantly reducing the impact to customers.","sourceUrl":"https://azure.microsoft.com/en-us/blog/optimizing-incident-management-with-aiops-using-the-triangle-system/"}],"outcomeDisclosed":true,"sources":[{"url":"https://azure.microsoft.com/en-us/blog/optimizing-incident-management-with-aiops-using-the-triangle-system/","title":"Optimizing incident management with AIOps using the Triangle System","publisher":"Microsoft Azure Blog","date":"2025-03-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"microsoft-azure-triangle-incident-triage"},{"title":"Microsoft: AI powered Proposal Resource Library for sellers answering RFPs and questionnaires","useCases":["rfp-and-proposal-response-drafting"],"organization":{"name":"Microsoft","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Responsive","role":"platform"}],"summary":"Microsoft's Proposal Center of Excellence has run a Proposal Resource Library on the Responsive platform since 2020. Sellers and experts across the worldwide sales organization use its AI recommendations to find vetted answers for proposals, RFPs, RFIs and security, legal and compliance assessments, searching more than 18,000 question and answer pairs that the proposal team's knowledge managers and technical experts across the company keep current. The vendor reports 18,000 users and, counted over a wider pool of more than 20,000 resources, more than 200,000 uses of AI answers.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":18000,"unit":"count","qualifier":"exact","period":"authenticated users of the library","claimant":"vendor","quote":"18K authenticated users leverage Responsive AI to quickly find proposal content and answers for security questionnaires, legal assessments, and highly technical bids","sourceUrl":"https://www.responsive.io/customer-stories/microsoft"},{"kpi":"interactions-handled","value":200000,"unit":"count","qualifier":"at-least","period":"uses of AI powered answers in proposals and assessments, cumulative","claimant":"vendor","quote":"The Field used AI-powered answers — drawn from over 20,000 resources — more than 200,000 times in sales proposals, RFPs, RFIs, and security, legal, and compliance assessments.","sourceUrl":"https://www.responsive.io/customer-stories/microsoft"},{"kpi":"time-saved-per-task","value":20,"unit":"minutes","qualifier":"exact","period":"per search for proposal content","claimant":"vendor","quote":"The Field saves 20 minutes per search for proposal content, totaling more than $17M worth of time spent on customer relationships and building pipeline instead of searching for content.","sourceUrl":"https://www.responsive.io/customer-stories/microsoft"},{"kpi":"hours-saved","value":93000,"unit":"hours","qualifier":"exact","period":"cumulative seller hours, period not stated","claimant":"vendor","quote":"Sellers gained 93K additional hours to spend on customer relationships and building pipeline, instead of searching for answers and proposal content.","sourceUrl":"https://www.responsive.io/customer-stories/microsoft"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.responsive.io/customer-stories/microsoft","title":"Microsoft - Customer Story | Responsive","publisher":"Responsive","date":"2024-11-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"microsoft-responsive-proposal-resource-library"},{"title":"Mid and South Essex NHS Foundation Trust: AI that predicts missed appointments and books backup slots","useCases":["patient-appointment-scheduling-and-reminders-agent"],"organization":{"name":"Mid and South Essex NHS Foundation Trust","anonymized":false,"country":"GB","region":"europe","industry":"healthcare"},"vendors":[{"name":"Deep Medical","role":"platform"}],"summary":"Mid and South Essex NHS Foundation Trust piloted software from Deep Medical that predicts which outpatient appointments are likely to be missed, using anonymised data and external factors such as weather, traffic and jobs. It offers patients more convenient times (for example evening and weekend slots for people who cannot take time off) and places intelligent backup bookings so clinical time is not lost. NHS England reports that the six month pilot cut non attendance by 30%, prevented 377 missed appointments and let an additional 1,910 patients be seen, and it announced a rollout to ten more trusts. The outcome is a reduction in missed appointments, which has no matching KPI in the taxonomy, so it is recorded in this summary rather than as a metric.","stage":"pilot","year":2024,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/","title":"NHS AI expansion to help tackle missed appointments and improve waiting times","publisher":"NHS England","date":"2024-03-14","archivedUrl":"https://web.archive.org/web/20250125233954/https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"mid-and-south-essex-missed-appointment-prediction"},{"title":"UK Ministry of Justice: Justice Transcribe meeting summaries for probation staff","useCases":["meeting-summarization-and-action-items"],"organization":{"name":"Ministry of Justice","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[],"summary":"Justice Transcribe is an AI transcription and meeting summarisation tool used by probation staff in England and Wales. The Ministry of Justice publishes transparency data on its use: between 7 October 2025 and 14 September 2026 more than 1.6 million meetings were summarised with it. Probation Workforce Transformation within HM Prison and Probation Service advised, as a broad operational assumption, about 10 minutes saved per meeting; the ministry itself labels the resulting hours figure illustrative, so it is not recorded as a result.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":1600000,"unit":"count","qualifier":"at-least","period":"meetings summarised, 7 October 2025 to 14 September 2026","claimant":"organization","quote":"Between 7 October 2025 and 14 September 2026, over 1,600,000 meetings were summarised using Justice Transcribe.","sourceUrl":"https://www.gov.uk/government/publications/justice-transcribe/justice-transcribe-data-7-october-2025-to-14-september-2026"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/government/publications/justice-transcribe/justice-transcribe-data-7-october-2025-to-14-september-2026","title":"Justice Transcribe data: 7 October 2025 to 14 September 2026","publisher":"Ministry of Justice (GOV.UK)","date":"2026-09-16"},{"url":"https://www.gov.uk/government/publications/justice-transcribe","title":"Justice Transcribe","publisher":"Ministry of Justice (GOV.UK)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"ministry-of-justice-justice-transcribe"},{"title":"Mizuho: generative AI for event detection and recovery in IT operations (proof of concept)","useCases":["aiops-incident-triage"],"organization":{"name":"Mizuho Financial Group","anonymized":false,"country":"JP","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"IBM","role":"platform"}],"summary":"Mizuho and IBM ran a three month proof of concept that added patterns likely to cause errors in incident response to generative AI on IBM watsonx and linked it to the application that supports event detection, so that operators flooded with messages during a disruption can find the cause faster. Accuracy was measured on actual data. Both firms said they planned to expand the proof of concept and apply it to production, and to use generative AI for incident management and failure analysis next.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"accuracy","value":98,"unit":"percent","qualifier":"exact","period":"three month trial","claimant":"vendor","quote":"The new solution demonstrated a 98% accuracy[1] in monitoring and responding to error messages during a three-month trial.","sourceUrl":"https://newsroom.ibm.com/2024-05-22-Mizuho-and-IBM-Unveil-Generative-AI-Initiative-to-Accelerate-Recovery-Time-in-Operations"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.ibm.com/2024-05-22-Mizuho-and-IBM-Unveil-Generative-AI-Initiative-to-Accelerate-Recovery-Time-in-Operations","title":"Mizuho and IBM Unveil Generative AI Initiative to Accelerate Recovery Time in Operations","publisher":"IBM Newsroom","date":"2024-05-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"mizuho-generative-ai-event-detection"},{"title":"Mobily: AI self service agents across eight messaging channels","useCases":["bill-explanation-and-billing-dispute-agent","first-line-contact-centre-agent","plan-upgrade-and-sales-assistant"],"organization":{"name":"Mobily","anonymized":false,"country":"SA","region":"middle-east","industry":"telecommunications"},"vendors":[{"name":"NiCE Cognigy","role":"platform"}],"summary":"Mobily deployed customer facing AI agents on eight channels, including WhatsApp, Twitter and Apple Business Chat, connected to its internal systems. The agents answer billing, balance and data usage questions, change subscriptions, sell add ons, take payments and recharges, and handle feedback and complaints, with a warm handover to a specialist who can take over or hand back. NiCE Cognigy reports that the first response time fell from 20 minutes to about 6 seconds. The deployment was already live in 2022, when the case study described it as conversational AI; the current version presents it as agentic AI.","stage":"scaled","year":2022,"channels":["whatsapp","social-messaging"],"languages":[],"metrics":[{"kpi":"response-time-reduction","value":99.5,"unit":"percent","qualifier":"exact","period":"first response time on messaging channels","baseline":"first response up to 20 minutes before","claimant":"vendor","quote":"An AI agent picks up any inquiry in around 6 seconds, reducing first response times significantly from the previous 20 minutes: a 99,5% improvement.","sourceUrl":"https://www.cognigy.com/en/case-study/mobily"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.cognigy.com/en/case-study/mobily","title":"Mobily: 99.5% faster response times with Agentic AI","publisher":"NiCE Cognigy","archivedUrl":"https://web.archive.org/web/20220413151348/https://www.cognigy.com/en/case-study/mobily"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"mobily-agentic-ai-self-service"},{"title":"MOGUL.sg: WhatsApp agent for property searches and viewing appointments","useCases":["branch-and-appointment-booking-agent"],"organization":{"name":"MOGUL.sg","anonymized":false,"country":"SG","region":"asia-pacific","industry":"real-estate"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"MOGUL.sg, a Singapore property platform, launched MAIA in February 2025: an AI agent on WhatsApp that searches listings and books viewing appointments, built with Vertex AI, Gemini and the Google Maps API. Location lookup plus booking in one conversation is the same pattern a bank needs for \"find my nearest branch and book me in\".","stage":"production","year":2025,"channels":["whatsapp"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"mogul-whatsapp-viewing-appointments"},{"title":"monday.com: AI assistant in developer docs and the API Playground","useCases":["developer-api-integration-assistant"],"organization":{"name":"monday.com","anonymized":false,"country":"IL","region":"middle-east","industry":"technology"},"vendors":[{"name":"Kapa.ai","role":"platform"}],"summary":"monday.com, whose developer ecosystem counts more than 100,000 customers building on its API, deployed an AI assistant in two places: an Ask AI widget in the developer documentation and an in product assistant inside the API Playground and developer center. It answers implementation and troubleshooting questions in real time for a global, multilingual developer base, where a slow answer risks a stalled integration. Kapa.ai reports that 10% of questions are answered in languages other than English, without naming them. Kapa.ai also claims over 50,000 support hours saved from repetitive questions, but this appears to be a modelled vendor estimate rather than a measured result, with no stated period.","stage":"production","year":2026,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":125000,"unit":"count","qualifier":"at-least","period":"per year","claimant":"vendor","quote":"125,000+ technical queries answered every year","sourceUrl":"https://www.kapa.ai/customer-examples/monday"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.kapa.ai/customer-examples/monday","title":"How Monday.com scaled AI chat to 100,000+ customers","publisher":"Kapa.ai"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"monday-com-developer-docs-assistant"},{"title":"Montgomery County, Maryland: Monty 2.0 constituent chatbot","useCases":["citizen-information-assistant","non-emergency-service-request-routing","public-service-translation"],"organization":{"name":"Montgomery County Government","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Zammo.ai","role":"platform"},{"name":"Microsoft (Azure OpenAI Service)","role":"model-provider"}],"summary":"Montgomery County first launched Monty to relieve its 311 hotline during the pandemic, with 20 topics, and retired it when demand fell. Monty 2.0, built with Zammo.ai on Azure OpenAI Service and Azure AI Search, answers questions on more than 3,000 topics, with automatic translation into 140 languages, from the county's own knowledge base, and uses the county's geographic data to give address specific answers such as trash pickup days. It went through a seven month beta with a constituent focus group before the full launch in late 2024.","stage":"production","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":20000,"unit":"count","qualifier":"at-least","period":"since the beta deployment","claimant":"vendor","quote":"Since its beta deployment, Monty 2.0 has facilitated more than 20,000 constituent conversations, achieving a 50% customer satisfaction rate and reducing unanswered queries from 35%–45% to just 10%–15%.","sourceUrl":"https://www.microsoft.com/en/customers/story/23066-montgomery-county-azure-open-ai-service"},{"kpi":"customer-satisfaction","value":50,"unit":"percent","qualifier":"exact","period":"since the beta deployment","claimant":"vendor","quote":"Since its beta deployment, Monty 2.0 has facilitated more than 20,000 constituent conversations, achieving a 50% customer satisfaction rate and reducing unanswered queries from 35%–45% to just 10%–15%.","sourceUrl":"https://www.microsoft.com/en/customers/story/23066-montgomery-county-azure-open-ai-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23066-montgomery-county-azure-open-ai-service","title":"Montgomery County revolutionizes constituent experiences with an AI chatbot powered by Microsoft Azure OpenAI Service","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"montgomery-county-monty-chatbot"},{"title":"Morgan Stanley: AI @ Morgan Stanley Assistant gives advisers access to the firm's intellectual capital","useCases":["enterprise-knowledge-search","wealth-advisor-knowledge-assistant"],"organization":{"name":"Morgan Stanley","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"OpenAI","role":"model-provider"}],"summary":"Morgan Stanley Wealth Management fully rolled out the AI @ Morgan Stanley Assistant in September 2023, a generative AI chatbot that gives Financial Advisors quick access to the firm's intellectual capital. The rollout followed the firm's March 2023 announcement of OpenAI as its strategic partner. In its June 2024 release the firm said that 98% of Financial Advisor teams had adopted it. It was followed by AI @ Morgan Stanley Debrief, which drafts meeting notes and follow up emails with client consent.","stage":"scaled","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch","title":"Launch of AI @ Morgan Stanley Debrief","publisher":"Morgan Stanley","date":"2024-06-26","archivedUrl":"https://web.archive.org/web/20260903124043/https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"morgan-stanley-ai-assistant-knowledge-search"},{"title":"Morgan Stanley: BlackRock Aladdin Wealth Auto Commentary in its Portfolio Risk Platform","useCases":["portfolio-reporting-and-commentary","portfolio-drift-monitoring-and-rebalancing","suitability-assessment-assistant"],"organization":{"name":"Morgan Stanley","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"BlackRock","role":"platform"}],"summary":"BlackRock announced on 2 October 2025 that Morgan Stanley Wealth Management's Portfolio Risk Platform would be the first to implement Auto Commentary, a generative AI feature of Aladdin Wealth, with advisors in the U.S. getting access from October. The tool combines Aladdin risk analytics, the firm's Chief Investment Office outlook and the client's holdings and investment preferences to draft concise insights for the advisor, highlighting issues such as overweights or misalignment with the client's objectives or the firm's market view. Trade press describes the output as bullet point insights inside a template, not full scripts or emails, so it supports the advisor's conversation rather than producing a finished client report. No outcome figures were published.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.blackrock.com/aladdin/discover/press-release/aladdin-wealth-launches-ai-enabled-commentary-tool-at-morgan-stanley","title":"Aladdin Wealth™ Launches AI-Enabled Commentary Tool for Wealth Advisors; Morgan Stanley's Portfolio Risk Platform First to Implement","publisher":"BlackRock","date":"2025-10-02"},{"url":"https://www.investmentnews.com/alternatives/blackrock-debuts-ai-powered-commentary-tool-for-advisors-lands-morgan-stanley-as-first-client/262370","title":"BlackRock debuts AI-powered commentary tool for advisors, lands Morgan Stanley as first client","publisher":"InvestmentNews","date":"2025-10-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"morgan-stanley-aladdin-auto-commentary"},{"title":"Morgan Stanley: AskResearchGPT for institutional sales, trading and banking staff","useCases":["investment-research-summarization"],"organization":{"name":"Morgan Stanley","anonymized":false,"country":"US","region":"north-america","industry":"capital-markets"},"vendors":[{"name":"OpenAI","role":"model-provider"}],"summary":"Morgan Stanley Research launched AskResearchGPT, a GPT-4 based assistant that lets investment banking, sales and trading and research staff search and summarize the firm's research (more than 70,000 proprietary reports a year), with hyperlinks to the source reports and a one click transfer of findings into an email draft that staff edit before sending to clients. It is available in the browser, Microsoft Teams and Outlook. Morgan Stanley's global director of research told CNBC that a salesperson needs one tenth of the time to answer the average client inquiry with the tool, and the bank said staff ask three times as many questions as with the traditional AI tool it had used since 2017.","stage":"production","year":2024,"channels":["internal-tools","microsoft-teams","email"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.morganstanley.com/press-releases/morgan-stanley-research-announces-askresearchgpt","title":"Morgan Stanley Research Announces AskResearchGPT","publisher":"Morgan Stanley","date":"2024-10-23","archivedUrl":"https://web.archive.org/web/2026/https://www.morganstanley.com/press-releases/morgan-stanley-research-announces-askresearchgpt"},{"url":"https://www.cnbc.com/2024/10/23/morgan-stanley-rolls-out-openai-powered-chatbot-for-wall-street-division.html","title":"AI on the trading floor: Morgan Stanley expands OpenAI-powered chatbot tools to Wall Street division","publisher":"CNBC","date":"2024-10-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"morgan-stanley-askresearchgpt"},{"title":"Morgan Stanley: AI @ Morgan Stanley Debrief meeting notes for financial advisors","useCases":["client-meeting-notes-and-crm-update"],"organization":{"name":"Morgan Stanley","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"OpenAI","role":"model-provider"}],"summary":"Morgan Stanley Wealth Management launched AI @ Morgan Stanley Debrief in June 2024. With client consent, the tool takes notes in client meetings, surfaces action items, summarizes the key points, drafts a follow up email for the advisor to edit and send at their discretion, and saves a note into Salesforce. The release quotes advisors on the time saved on note taking (one cites about half an hour per meeting) but gives no firm wide measurement.","stage":"production","year":2024,"channels":["internal-tools","email"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch","title":"Launch of AI @ Morgan Stanley Debrief","publisher":"Morgan Stanley","date":"2024-06-26","archivedUrl":"https://web.archive.org/web/2026/https://www.morganstanley.com/press-releases/ai-at-morgan-stanley-debrief-launch"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"morgan-stanley-debrief-meeting-notes"},{"title":"Morgan Stanley: DevGen.AI translates legacy code into modern specifications","useCases":["legacy-code-modernization"],"organization":{"name":"Morgan Stanley","anonymized":false,"country":"US","region":"global","industry":"capital-markets"},"vendors":[{"name":"Morgan Stanley","role":"in-house"},{"name":"OpenAI","role":"model-provider"}],"summary":"Morgan Stanley launched DevGen.AI in January 2025, an in house tool built on OpenAI's GPT models and trained on the languages in its own code base, including company specific ones. It turns code in older languages such as Perl into plain English specifications that developers then use to rewrite the code in modern languages. The firm keeps developers in the loop because the tool does not yet write the new code as well as a human, and said it would not cut its engineering workforce as a result.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"hours-saved","value":280000,"unit":"hours","qualifier":"approximately","period":"first five months after launch","claimant":"organization","quote":"Mike Pizzi, Morgan Stanley’s global head of technology and operations, told WSJ that in the five months since its launch, DevGen.AI has worked through nine million lines of code, saving the firm’s 15,000 developers roughly 280,000 hours of work.","sourceUrl":"https://www.entrepreneur.com/business-news/morgan-stanley-builds-ai-tool-that-fixes-major-coding-issue/492697"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.entrepreneur.com/business-news/morgan-stanley-builds-ai-tool-that-fixes-major-coding-issue/492697","title":"'Building It Ourselves': Morgan Stanley Created an AI Tool to Fix the Most Annoying Part of Coding","publisher":"Entrepreneur","date":"2025-06-03"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"morgan-stanley-devgen-ai"},{"title":"Morgan Stanley: Next Best Action engine for financial advisors","useCases":["next-best-action-for-advisors"],"organization":{"name":"Morgan Stanley","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"Morgan Stanley","role":"in-house"}],"summary":"Morgan Stanley Wealth Management built Next Best Action, an internal AI based engine that delivers timely, customized messages to clients and prospects, guided by the financial advisor. In March 2023, when it announced a strategic initiative with OpenAI to create a bespoke solution that its financial advisors would use, the firm listed it among its recent AI projects, alongside its Genome capability that uses data analytics and machine learning to personalize client communication. No outcome figures are published in that release.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai","title":"Key Milestone in Innovation Journey with OpenAI","publisher":"Morgan Stanley","date":"2023-03-14","archivedUrl":"https://web.archive.org/web/2026/https://www.morganstanley.com/press-releases/key-milestone-in-innovation-journey-with-openai"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"morgan-stanley-next-best-action"},{"title":"Morrisons: AI demand forecasting in place of manual store replenishment","useCases":["retail-demand-forecasting-and-replenishment"],"organization":{"name":"Morrisons","anonymized":false,"country":"GB","region":"europe","industry":"retail-and-ecommerce"},"vendors":[{"name":"Blue Yonder","role":"platform"},{"name":"Microsoft","role":"platform"}],"summary":"The UK grocer Morrisons replaced its manual model of stock replenishment with Blue Yonder's AI powered demand forecasting and replenishment solution, built on Microsoft Azure, which predicts customer demand and orders the right level of stock for its stores. Blue Yonder's customer page says it helped Morrisons increase shelf availability of more than 29,000 products in 130 categories across its 500 stores and headlines a 30% on shelf availability improvement. Technology Record, a publication produced with Microsoft's support, reported in January 2018 that the solution cut shelf gaps in Morrisons stores by 30% and stockholding in store by two to three days.","stage":"scaled","year":2018,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://blueyonder.com/customers/morrisons","title":"Morrisons Simplifies Fresh Food Clearance with Blue Yonder","publisher":"Blue Yonder"},{"url":"https://www.technologyrecord.com/article/morrisons-implements-blue-yonders-ai-stock-replenishment-technology","title":"Morrisons implements Blue Yonder's AI stock replenishment technology","publisher":"Technology Record","date":"2018-01-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"morrisons-blue-yonder-automated-replenishment"},{"title":"Mount Sinai: systemwide AI clinical trial matching for cancer patients","useCases":["clinical-trial-patient-matching"],"organization":{"name":"Mount Sinai Health System","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Triomics","role":"platform"}],"summary":"The Mount Sinai Tisch Cancer Center deployed PRISM, an oncology specific trial matching platform from Triomics built on its OncoLLM language model pipeline, across the Mount Sinai Health System in January 2026. The platform reviews patient records against trial protocols so that patients seen at other hospitals in the system, such as Mount Sinai Queens and Mount Sinai Brooklyn, have the same access to trials as those treated at The Mount Sinai Hospital. Mount Sinai says the aim is also to let clinicians focus on conversations with patients rather than manual chart review; the January 2026 release reports no outcome figures. Mount Sinai said it would evaluate outcomes and publish them later.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.mountsinai.org/about/newsroom/2026/mount-sinai-launches-ai-powered-clinical-trial-matching-platform-to-expand-access-to-cancer-research","title":"Mount Sinai Launches AI-Powered Clinical Trial-Matching Platform to Expand Access to Cancer Research","publisher":"Mount Sinai Health System","date":"2026-01-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"mount-sinai-oncology-trial-matching"},{"title":"MYbank: the 310 model for collateral free SME loans","useCases":["sme-cash-flow-underwriting"],"organization":{"name":"MYbank","anonymized":false,"country":"CN","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"MYbank","role":"in-house"}],"summary":"MYbank, the Chinese digital bank associated with Ant Group, lends to small and micro businesses with its \"310 model\": a collateral free business loan that takes under three minutes to apply for on a phone, under one second to approve and no human interaction. The bank says AI, including Ant Group's Bailing foundation model and a supply chain knowledge graph, informs its lending decisions. By the end of 2023 it had served over 53 million small and micro businesses, and over 72% of the 3 million new borrowers it added in 2023 had obtained a business loan from a bank for the first time. It also uses satellite imagery to estimate farm output and an AI conversational system to manage credit lines.","stage":"scaled","year":2024,"channels":["mobile-app","api"],"languages":["zh"],"metrics":[{"kpi":"users-served","value":53000000,"unit":"count","qualifier":"at-least","period":"small and micro businesses served, cumulative to the end of 2023","claimant":"organization","quote":"the bank has cumulatively served over 53 million small and micro-sized enterprises (SMEs) as of the end of 2023","sourceUrl":"https://aijourn.com/leveraging-ai-mybank-enables-financing-services-for-53-million-smes/"}],"outcomeDisclosed":true,"sources":[{"url":"https://aijourn.com/leveraging-ai-mybank-enables-financing-services-for-53-million-smes/","title":"Leveraging AI, MYbank Enables Financing Services for 53 Million SMEs","publisher":"MYbank press release via Business Wire, republished by The AI Journal","date":"2024-04-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"mybank-310-sme-lending"},{"title":"NAB: QuickBiz automated unsecured small business lending","useCases":["sme-cash-flow-underwriting"],"organization":{"name":"National Australia Bank","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Amazon Web Services","role":"platform"}],"summary":"NAB's QuickBiz platform decides unsecured small business loans and overdrafts online. In 2021 its product page said NAB reviews the applicant's cash flow, credit score and time in business, and that businesses using Xero, MYOB or QuickBooks can link their accounting data. iTnews reported that the platform uses machine learning to make decisions faster. In 2019 the general manager of digital and sales transformation in NAB's business and private bank said that 45 percent of NAB's small business lending accounts were being opened this way, and that credit decisions often came the same day.","stage":"scaled","year":2019,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.itnews.com.au/news/nab-watches-cloud-based-quickbiz-lending-process-gain-traction-530744","title":"NAB watches cloud-based QuickBiz lending process gain traction","publisher":"iTnews","date":"2019-09-09"},{"url":"https://www.nab.com.au/business/loans-and-finance/business-loans/nab-quickbiz-loan","title":"NAB QuickBiz unsecured business loan","publisher":"National Australia Bank","archivedUrl":"https://web.archive.org/web/20210119020618/https://www.nab.com.au/business/loans-and-finance/business-loans/nab-quickbiz-loan"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"nab-quickbiz-automated-sme-lending"},{"title":"NASA IV&V: generative AI review of requirements quality, traceability and test artifacts (planned)","useCases":["requirements-to-test-case-generation"],"organization":{"name":"National Aeronautics and Space Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"NASA's Independent Verification and Validation programme at Goddard reports two related pre deployment tools. One drafts analysis of software requirements (quality attributes, decomposition of functionality, upward and backward traceability) and the other assesses test cases, procedures and steps for completeness and consistency. Analysts filter the findings by severity, give feedback and turn accepted findings into draft issues. Development used synthetic or open data; the move to beta and production on premises with real mission data was planned from fiscal year 2026.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entries NASA-908, IV&V Requirements Quality & Traceability Analysis, and NASA-910, IV&V Test Analysis)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"nasa-ivv-requirements-and-test-analysis"},{"title":"Nasdaq: generative AI for market abuse alert triage in its surveillance platform","useCases":["market-abuse-surveillance-triage"],"organization":{"name":"Nasdaq","anonymized":false,"country":"US","region":"north-america","industry":"capital-markets"},"vendors":[{"name":"Amazon Web Services","role":"platform"}],"summary":"Nasdaq added a generative AI feature, built on Amazon Bedrock, to the market surveillance technology it runs for regulators and marketplaces. When an alert fires, the feature gathers and condenses the evidence an analyst needs for the initial assessment, such as a table of the company's regulatory filings, news summaries and sentiment and other mitigating or aggravating factors. The reported gain comes from proof of concept testing, in which analysts estimated the investigation time saved; Nasdaq said it planned to use the feature for its own US equity market surveillance.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":33,"unit":"percent","qualifier":"approximately","period":"proof of concept testing, estimated by surveillance analysts","claimant":"organization","quote":"During proof-of-concept testing, surveillance analysts estimated a 33% reduction in investigation time, with improved overall outcomes.","sourceUrl":"https://press.aboutamazon.com/aws/2024/5/nasdaq-to-enhance-global-market-surveillance-offering-with-generative-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://press.aboutamazon.com/aws/2024/5/nasdaq-to-enhance-global-market-surveillance-offering-with-generative-ai","title":"Nasdaq to Enhance Global Market Surveillance Offering with Generative AI","publisher":"Nasdaq (published on the Amazon press center)","date":"2024-05-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"nasdaq-market-surveillance-generative-ai"},{"title":"National Bank of Greece (Cyprus): four reconciliation systems consolidated on an AI enabled platform","useCases":["ledger-and-payment-reconciliation"],"organization":{"name":"National Bank of Greece (Cyprus)","anonymized":false,"country":"CY","region":"europe","industry":"banking"},"vendors":[{"name":"Smartstream","role":"platform"}],"summary":"National Bank of Greece in Cyprus consolidated four reconciliation systems into one on Smartstream's AI enabled Air platform (the Air Cash module), replacing both incumbent and standalone systems. The bank had a fragmented landscape that needed significant daily manual effort across systems and data formats. The platform matches groups of items at once and flags data quality issues in internal data and incoming bank statements. The project was completed in three months; no operational outcome figures were disclosed.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://thepaypers.com/fintech/news/national-bank-of-greece-in-cyprus-goes-live-with-smartstream-air","title":"National Bank of Greece in Cyprus goes live with Smartstream Air","publisher":"The Paypers","date":"2026-05-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"national-bank-of-greece-cyprus-ai-reconciliation"},{"title":"National Gallery Singapore: G(ai)le AI docent for museum visitors","useCases":["ai-visitor-and-tour-guide"],"organization":{"name":"National Gallery Singapore","anonymized":false,"country":"SG","region":"asia-pacific","industry":"government"},"vendors":[{"name":"Microsoft (Azure OpenAI)","role":"model-provider"},{"name":"NCS","role":"integrator"}],"summary":"National Gallery Singapore built G(ai)le, an AI docent that searches the Gallery's archives and explains artworks to visitors in conversational language, in English, Mandarin, Malay and Tamil. It adapts its stories to interests a visitor mentions, offers an audio only mode and an \"eyes up\" mode, and gives the Gallery a view of what visitors ask about. Staff also use it to draft versions of tour content for different audiences, which writers then refine.","stage":"production","year":2025,"channels":[],"languages":["en","zh","ms","ta"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/24629-national-gallery-singapore-azure-openai","title":"National Gallery Singapore’s new virtual guide lets visitors engage with art in new ways, powered by Azure OpenAI","publisher":"Microsoft Customer Stories","date":"2025-07-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"national-gallery-singapore-gaile-ai-docent"},{"title":"National Grid: satellite and AI based, condition driven vegetation management in Massachusetts","useCases":["power-line-vegetation-management"],"organization":{"name":"National Grid","anonymized":false,"country":"US","region":"north-america","industry":"energy-and-utilities"},"vendors":[{"name":"AiDASH","role":"platform"}],"summary":"National Grid began working with AiDASH in 2020 in Massachusetts, a service area of over 13,500 line miles and more than 1.3 million customers. According to AiDASH, the utility had been on a five year trim cycle and had deferred work for four years running rather than fund it; in August 2020 it ran a proof of concept on its entire Massachusetts footprint, and the first model run produced its FY2021 work plan. It adopted the Intelligent Vegetation Management System in 2021, which uses satellite imagery and AI to show vegetation conditions across the network, and moved to condition based trimming, with circuits now on cycles of four to seven years. AiDASH reports $1M in avoided cost from dropping manual field reviews and $2M in efficiencies in the first few years, without saying whether the two overlap. It reports three sets of reliability results: on Massachusetts circuits worked in FY2022 to 2025, average improvements of 22% in tree events, 29% in customers impacted and 43% in customer minutes interrupted, measured 12 months after each circuit is worked; in its case study a 30% decline in tree related events, 38% in customers interrupted and 55% in customer minutes interrupted in the year after circuits were pruned; and from a talk by National Grid's vegetation strategy manager at the NextGrid Alliance Summit 2025, decreases of 26.4%, 30.2% and 46.5% on the same three measures.","stage":"production","year":2021,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"cost-savings","value":2000000,"unit":"currency","currency":"USD","qualifier":"exact","period":"in the first few years of adopting IVMS","claimant":"vendor","quote":"$2M in efficiencies realized – in the first few years of adopting IVMS.","sourceUrl":"https://www.aidash.com/resource/national-grid-delivers-tangible-value-with-ivms/"},{"kpi":"cost-savings","value":1000000,"unit":"currency","currency":"USD","qualifier":"exact","period":"not stated","claimant":"vendor","quote":"$1M in avoided cost – with technology removing the need for time consuming manual processes like field reviews to determine if a circuit needs to be pruned.","sourceUrl":"https://www.aidash.com/resource/national-grid-delivers-tangible-value-with-ivms/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.aidash.com/resource/real-results-from-national-grids-vegetation-program/","title":"Real Results from National Grid's Vegetation Program","publisher":"AiDASH","date":"2025-10-14"},{"url":"https://www.aidash.com/resource/national-grid-delivers-tangible-value-with-ivms/","title":"National Grid delivers tangible value with IVMS","publisher":"AiDASH"},{"url":"https://www.aidash.com/resource/how-national-grid-cut-vegetation-related-impacts-by-up-to-43-per-cent/","title":"How National Grid Cut Vegetation-Related Impacts by Up to 43%","publisher":"AiDASH","date":"2026-09-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"national-grid-satellite-vegetation-management"},{"title":"NatWest: agentic AI that investigates complaints for a human handler, tested with the FCA","useCases":["complaints-handling-agent"],"organization":{"name":"NatWest Group","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[],"summary":"NatWest is testing an agentic AI system that investigates customer complaints across several data sources and presents a summarised view to a complaint handler, who approves it, with the aim of speeding up handling and resolution. The trial runs in the FCA's AI Live Testing environment, with performance tracked daily on task accuracy, coherence and hallucination, and every AI generated summary subject to human oversight. The bank plans to use production like environments before any move to live use. It is a pilot; no outcome figures are disclosed.","stage":"pilot","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.natwestgroup.com/news-and-insights/latest-stories/ai-and-data/2026/may/collaborating-with-the-financial-conduct-authority-on-testing-ag.html","title":"Collaborating with the Financial Conduct Authority on testing agentic AI","publisher":"NatWest Group","date":"2026-05-05"},{"url":"https://connect.cefpro.com/article/view/fca-warns-banks-as-agentic-ai-nears-consumer-rollout","title":"FCA Warns Banks as Agentic AI Nears Consumer Rollout","publisher":"CeFPro Connect"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"natwest-agentic-ai-complaints-handling"},{"title":"NatWest: Cora AI assistant as the front door for everyday queries, including ATM disputes","useCases":["first-line-contact-centre-agent","atm-and-self-service-device-assistance"],"organization":{"name":"NatWest Group","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[],"summary":"NatWest routes a wide range of everyday customer queries through Cora, its AI assistant in online banking and the mobile app, now with generative AI (Cora+). Customers whose ATM withdrawal did not pay out are sent to Cora with the phrase \"ATM dispute\" as the first step of the claim, and the assistant is available before login as well. NatWest says the generative AI version improved customer satisfaction and reduced how often a colleague has to step in, and in 2025 it began a collaboration with OpenAI to extend the assistant to more complex tasks.","stage":"scaled","year":2025,"channels":["mobile-app","web-chat"],"languages":["en"],"metrics":[{"kpi":"customer-satisfaction-uplift","value":150,"unit":"percent","qualifier":"exact","period":"Cora+ generative AI functionality","claimant":"organization","quote":"The GenAI functionality offered by Cora+ has shown a 150% improvement in customer satisfaction, while reducing the number of times a colleague needs to intervene.","sourceUrl":"https://www.natwestgroup.com/news-and-insights/news-room/press-releases/ai-and-data/2025/mar/natwest-open-ai-collaborate-to-accelerate-cutting-edge-ai-transf.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.natwestgroup.com/news-and-insights/news-room/press-releases/ai-and-data/2025/mar/natwest-open-ai-collaborate-to-accelerate-cutting-edge-ai-transf.html","title":"NatWest & OpenAI collaborate to accelerate cutting-edge AI transformation in support of bank-wide simplification and enhanced customer experience","publisher":"NatWest Group","date":"2025-03-20"},{"url":"https://www.natwest.com/support-centre/banking-near-me/withdrawals/i-have-used-an-atm-to-withdraw-money-and-i-didnt-receive-anything.html","title":"I have used an ATM to withdraw money and I didn't receive anything?","publisher":"NatWest"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"natwest-cora-ai-assistant"},{"title":"NatWest: real time machine learning fraud and scam detection with Featurespace","useCases":["real-time-fraud-scoring"],"organization":{"name":"NatWest Group","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[{"name":"Featurespace","role":"platform"}],"summary":"NatWest began working with Featurespace in 2019, when its incumbent fraud detection system was struggling to identify fraud and scams, and moved to Featurespace's real time platform with adaptive machine learning models as the first line of defence, deployed enterprise wide. Building on its results in authorised push payment scam detection, the bank extended the platform to real time debit card fraud detection with deep behavioural models and ensembled risk scores, integrated with SMS alerts that let customers approve or decline transactions. The vendor reports, citing NatWest data from 2025, a higher value of fraud and scams detected and fewer false positives on scams.","stage":"scaled","year":2019,"channels":["api","sms"],"languages":["en"],"metrics":[{"kpi":"detection-rate-improvement","value":135,"unit":"percent","qualifier":"exact","period":"value of scams detected (NatWest data, 2025)","claimant":"vendor","quote":"135%Improved value of scams detected","sourceUrl":"https://www.featurespace.com/case-studies/natwest"},{"kpi":"detection-rate-improvement","value":57,"unit":"percent","qualifier":"exact","period":"value of fraud detected (NatWest data, 2025)","claimant":"vendor","quote":"57%Improved value of fraud detected","sourceUrl":"https://www.featurespace.com/case-studies/natwest"},{"kpi":"false-positive-reduction","value":75,"unit":"percent","qualifier":"exact","period":"scam detection (NatWest data, 2025)","claimant":"vendor","quote":"75%Reduced false positives for scams","sourceUrl":"https://www.featurespace.com/case-studies/natwest"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.featurespace.com/case-studies/natwest","title":"NatWest case study","publisher":"Featurespace"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"natwest-featurespace-fraud-and-scam-detection"},{"title":"nbn: machine learning Tech Lab to decide between remote fixes and technician visits","useCases":["field-technician-copilot-and-dispatch"],"organization":{"name":"nbn","anonymized":false,"country":"AU","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"nbn (in house Tech Lab)","role":"in-house"}],"summary":"In a 2017 blog post, nbn described a Tech Lab that would use big data and machine learning, including survey data from consenting end users, to improve the experience of its access network. One stated goal was to help teams determine whether a fault can be fixed remotely or needs a field technician to visit; another was to spot trends in failed activations so problems can be anticipated before a technician arrives. The post describes the programme's aims and reports no results.","stage":"announced","year":2017,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.nbnco.com.au/blog/the-nbn-project/nbns-tech-labs-using-machine-learning-to-improve-network-experience","title":"nbn’s Tech Lab: improving network experience through machine learning","publisher":"nbn","date":"2017-09-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"nbn-tech-lab-fault-dispatch-prediction"},{"title":"National Credit Union Administration: machine learning validation of Call Report data","useCases":["regulatory-report-assembly"],"organization":{"name":"National Credit Union Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The NCUA, which supervises US federal credit unions, uses a machine learning model developed in house to improve the quality of the quarterly Call Report data that credit unions file. Its output is a list of potential data outliers for each credit union. It has been in operation since February 2023. It is the supervisor side of regulatory reporting, and shows the kind of outlier check a filer can run on its own data before it submits. No outcome figures are published.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"ncua-call-report-machine-learning-validation"},{"title":"Netherlands Labour Authority: random forest risk model to select asbestos removal jobs for inspection","useCases":["inspection-prioritization"],"organization":{"name":"Nederlandse Arbeidsinspectie","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[{"name":"In house (Nederlandse Arbeidsinspectie)","role":"in-house"}],"summary":"Asbestos removal jobs must be notified in advance, and the Netherlands Labour Authority inspects a share of them to protect workers and the surroundings. Its IPA risk model, a supervised random forest classifier trained on past notifications and inspection findings together with Chamber of Commerce and pseudonymised employment data, gives every notified removal a score for the likelihood that it is done incorrectly, and inspectors combine the score with other information to choose which jobs to inspect. Model based inspections are only part of the programme: in principle every certified company is inspected at least once every three years, some inspections follow reports from citizens and other regulators, and some jobs are chosen at random, with their results used to improve the model's reliability. In use since September 2024.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/oorg12349/74441495/ipa-risicomodel","title":"IPA risicomodel, Algoritmeregister","publisher":"Algoritmeregister van de Nederlandse overheid","date":"2024-12-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"nederlandse-arbeidsinspectie-asbestos-removal-risk-model"},{"title":"Netflix: a foundation model for personalized recommendation","useCases":["content-recommendation-and-personalization"],"organization":{"name":"Netflix, Inc.","anonymized":false,"country":"US","region":"north-america","industry":"media-and-entertainment"},"vendors":[],"summary":"Netflix built a foundation model that learns members' preferences from their comprehensive interaction history in one place, tokenizing user actions the way text is tokenized for a large language model, and shares those learned preferences with other models through embeddings or through fine tuning, instead of each model learning from scratch. Netflix says it sees \"promising results from downstream integrations\" of the model, without disclosing whether it serves members' recommendations directly, or how many surfaces or how much volume it covers; live deployment status beyond those integrations is not disclosed. Netflix's own research team, with one academic coauthor, separately published a causal study of the value of personalization in its recommender system: replacing the current recommender with a simpler popularity based ranking would cut member engagement by 12%, most of it from effective targeting rather than just showing content to more people.","stage":"pilot","year":2025,"channels":["mobile-app","api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://netflixtechblog.com/foundation-model-for-personalized-recommendation-1a0bd8e02d39","title":"Foundation Model for Personalized Recommendation","publisher":"Netflix Technology Blog","archivedUrl":"https://web.archive.org/web/2026/https://netflixtechblog.com/foundation-model-for-personalized-recommendation-1a0bd8e02d39"},{"url":"https://arxiv.org/abs/2511.07280","title":"The Value of Personalized Recommendations: Evidence from Netflix","publisher":"arXiv","date":"2025-11-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"netflix-foundation-model-recommendation"},{"title":"Newcastle City Council: AI routing, agent prompts and call analytics in the contact centre","useCases":["non-emergency-service-request-routing"],"organization":{"name":"Newcastle City Council","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Amazon Web Services (Amazon Connect, Amazon Q in Connect, Contact Lens)","role":"platform"},{"name":"PwC","role":"integrator"}],"summary":"Newcastle City Council replaced legacy telephony with Amazon Connect, which routes resident calls and chats to the right team; Amazon Lex may also power menus for self service and triage. During live contacts, Amazon Q in Connect suggests approved knowledge and responses to agents; Contact Lens transcribes calls and surfaces topics and quality signals for supervisors. The council states that all outputs are advisory and do not decide eligibility, enforcement or case outcomes.","stage":"production","year":2026,"channels":["voice","web-chat","agent-desktop"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/newcastle-city-council-aws-contact-centre-services-amazon-q-and-contact-lens","title":"Newcastle City Council: AWS Contact Centre Services (Amazon Q and Contact Lens)","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2026-08-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"newcastle-city-council-contact-centre-ai"},{"title":"Nexo: AI agents that write alert narratives and propose dispositions with Unit21","useCases":["suspicious-activity-report-drafting","aml-alert-triage"],"organization":{"name":"Nexo","anonymized":false,"region":"europe","industry":"payments"},"vendors":[{"name":"Unit21","role":"platform"}],"summary":"Digital asset services company Nexo uses Unit21's transaction monitoring and case management with AI agents that automate alert narratives and dispositions. Analysts work in a supervisory role, verifying the AI generated output, investigating anomalies and applying judgment. The vendor reports that a majority of alert reviews are now automated, with further automation projected.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":57,"unit":"percent","qualifier":"exact","period":"share of alert reviews automated","claimant":"vendor","quote":"By automating alert narratives and dispositions, Unit21’s AI Agents have enabled Nexo to achieve 57% automation in alert reviews, with projections to reach up to 80% as the models continue to evolve.","sourceUrl":"https://www.unit21.ai/customers/nexo"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.unit21.ai/customers/nexo","title":"Nexo Case Study","publisher":"Unit21"},{"url":"https://baytobaynews.com/daily-state-news/stories/unit21-awarded-two-2026-datos-impact-awards-for-ai-innovation-cryptodigital-asset-aml-innovation,346520","title":"Unit21 Awarded Two 2026 Datos Impact Awards for AI Innovation & Crypto/Digital Asset AML Innovation","publisher":"Business Wire (via Bay to Bay News)","date":"2026-09-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"nexo-unit21-ai-alert-narratives"},{"title":"Next: AI agent for arranging returns across brands and markets","useCases":["order-status-and-returns-agent"],"organization":{"name":"Next","anonymized":false,"country":"GB","region":"europe","industry":"retail-and-ecommerce"},"vendors":[{"name":"Sierra","role":"platform"}],"summary":"Next, the British fashion retailer that operates across 83 countries and also serves customers of brands such as Gap, Victoria's Secret and Fat Face, launched an AI agent in six weeks across two use cases. The agent handles arrange return queries, matches the customer to the order without asking for an order number and manages identity verification. It adapts to regional preferences, so Next can add languages and processes as it grows, and it meets customers across chat, voice and WhatsApp. No outcome figures were published.","stage":"production","year":2026,"channels":["web-chat","voice","whatsapp"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://sierra.ai/customers/next","title":"How Next transforms global customer service with AI","publisher":"Sierra","date":"2026-02-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"next-returns-ai-agent"},{"title":"NJ Transit: Navvie trip planning chatbot","useCases":["public-transit-passenger-information-agent"],"organization":{"name":"NJ Transit","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[],"summary":"NJ Transit launched Navvie, its first AI powered chatbot, alongside a redesigned website; Mass Transit reported this by early September 2026. Navvie is available around the clock to help riders plan trips and get schedules, alerts and transfer information. NJ Transit describes it as a pilot: results are being analysed before a decision on integrating it into the mobile app, and the agency separately issued a request for information for a larger, unified real time customer communications platform.","stage":"pilot","year":2026,"channels":["web-chat"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.masstransitmag.com/management/news/55402724/new-jersey-transit-nj-transit-nj-transit-launches-new-website-ai-chatbot-rfi-to-improve-customer-communications","title":"NJ Transit launches new website, AI chatbot, RFI to improve customer communications","publisher":"Mass Transit","date":"2026-09-03"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"nj-transit-navvie-chatbot"},{"title":"Nomad eSIM: generative AI help for support agents answering trouble tickets","useCases":["email-and-ticket-reply-drafting"],"organization":{"name":"Nomad eSIM","anonymized":false,"country":"US","region":"north-america","industry":"telecommunications"},"vendors":[{"name":"Google","role":"platform"}],"summary":"Nomad eSIM, a LotusFlare brand used by international travellers, gives its customer support agents Gemini in Google Workspace to respond to trouble tickets more efficiently. Google Cloud reports higher customer satisfaction from faster support responses, but gives no figure.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"},{"url":"https://lotusflare.com/contact-us/","title":"Contact us (headquarters of LotusFlare, Inc.)","publisher":"LotusFlare"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"nomad-esim-support-ticket-replies"},{"title":"Norfolk Southern: Wheel Integrity System","useCases":["freight-rail-rolling-stock-predictive-maintenance"],"organization":{"name":"Norfolk Southern","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[{"name":"Norfolk Southern","role":"in-house"},{"name":"Georgia Tech Research Institute","role":"integrator"}],"summary":"Norfolk Southern built a standalone Wheel Integrity System with its own Data Science and AI team, developed with integration support from the Georgia Tech Research Institute. Six synchronized cameras capture about 55 high resolution images per wheel as trains pass at up to 70 mph, and AI algorithms analyse the images to detect subtle defects that are difficult for the human eye to identify consistently. The first site went live near Chicago on November 24, 2025. The new system follows the railroad's existing Digital Train Inspection (DTI) portals, a separate, earlier system that scans entire trains; the DTI portals had already identified and removed from service over 50 wheels with issues since January 2025. Unlike DTI, the new system zeroes in on wheels specifically, and it has pinpointed a vendor wheel casting flaw that triggered an industry recall and seven confirmed defects across North America.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.norfolksouthern.com/en/newsroom/story-yard/introducing-the-wheel-integrity-system--ns--latest-safety-revolution","title":"Introducing the Wheel Integrity System: NS' latest safety revolution","publisher":"Norfolk Southern","date":"2026-01-29"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"norfolk-southern-wheel-integrity-system"},{"title":"Novo Nordisk: NovoScribe for clinical study reports and regulatory documentation","useCases":["clinical-and-regulatory-document-drafting"],"organization":{"name":"Novo Nordisk","anonymized":false,"country":"DK","region":"europe","industry":"pharma-and-life-sciences"},"vendors":[{"name":"Anthropic","role":"model-provider"},{"name":"Amazon Web Services","role":"platform"},{"name":"MongoDB","role":"platform"}],"summary":"Novo Nordisk built NovoScribe, a documentation platform that combines retrieval augmented generation over expert approved text with case specific variables to draft clinical study reports. It runs on Amazon Bedrock and MongoDB Atlas with Claude models, and has been extended to device verification protocols and patient materials. Anthropic's case study quotes Novo Nordisk saying writing times on clinical study reports fell by 90%, with drafts going to people for review and approval; the company aims to extend it to full Common Technical Documents.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":90,"unit":"percent","qualifier":"exact","period":"writing time per clinical study report","claimant":"organization","quote":"“Claude has helped us cut writing times on CSRs by 90% so we can get documentation directly into human hands for review and approval,” said Waheed Jowiya, Digitalization Strategy Director at Novo Nordisk.","sourceUrl":"https://claude.com/customers/novo-nordisk"}],"outcomeDisclosed":true,"sources":[{"url":"https://claude.com/customers/novo-nordisk","title":"Novo Nordisk accelerates clinical documentation and drug development with Claude","publisher":"Anthropic"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"novo-nordisk-novoscribe-clinical-documentation"},{"title":"US National Science Foundation: ServiceNow Now Assist drafting responses, work notes and knowledge articles","useCases":["support-knowledge-article-generation"],"organization":{"name":"U.S. National Science Foundation","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"ServiceNow","role":"platform"}],"summary":"NSF's Office of Information and Resource Management uses ServiceNow's generative AI, Now Assist, on a FedRAMP High platform to generate content for its service operation, including responses, work notes and knowledge base articles, alongside recommendations and chatbots. It shows the common pattern of knowledge article drafting switched on inside an existing service management platform rather than built separately. No outcome figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (NSF entry 7, ServiceNow GenAI Now Assist)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"nsf-servicenow-now-assist-knowledge-articles"},{"title":"Nsure.com: Friendly John copilot for payments, renewal offers and discount requests","useCases":["insurance-policy-servicing-agent","insurance-renewal-and-retention"],"organization":{"name":"Nsure.com","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Microsoft","role":"platform"},{"name":"Netwise","role":"integrator"}],"summary":"Nsure.com is a Florida based digital insurance agency that lets consumers compare home and auto quotes from more than 50 insurers and buy online. Generative AI in Power Automate cut its service representatives' manual processing time by more than 60%, for example by triaging the shared inboxes and either preparing an automated response or routing each email to an agent. It replaced a third party chatbot with a Copilot Studio copilot, Friendly John, that helps customers submit payments, review renewal offers and request discounts, with an interactive voice response option and after hours support. Its VP of AI and Automation says it handles around 60% of customer questions, and the company plans to use copilots for new policy sales and cross selling.","stage":"production","year":2024,"channels":["web-chat","voice","email","sms"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1728829430186194098-nsure-power-platform-insurance-usa","title":"Digital insurance agency, Nsure.com, reduces manual processing time by 60% using generative AI and Power Automate","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"nsure-friendly-john-copilot"},{"title":"NTT DOCOMO: AI radio capacity planning for its 5G rollout","useCases":["network-planning-and-capacity-optimization"],"organization":{"name":"NTT DOCOMO","anonymized":false,"country":"JP","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Nokia","role":"platform"}],"summary":"NTT DOCOMO deployed Nokia's AI radio frequency capacity planning software, customised to its requirements, to support its 5G rollout. The software predicts the capacity of 4G cells from base station performance data and simulates the best candidate locations for 5G cells and radio hardware to meet the capacity needed in an area, helping DOCOMO see where congestion is starting and where to plan upgrades. No results are published.","stage":"production","year":2022,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.nokia.com/newsroom/nokia-deploys-ava-ai-software-to-help-ntt-docomo-enhance-5g-network-planning/","title":"Nokia deploys AVA AI software to help NTT DOCOMO enhance 5G network planning","publisher":"Nokia","date":"2022-09-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"ntt-docomo-nokia-ai-capacity-planning"},{"title":"Netherlands Food and Consumer Product Safety Authority: compliance model to select pig farms for welfare inspections","useCases":["inspection-prioritization"],"organization":{"name":"Nederlandse Voedsel- en Warenautoriteit (NVWA)","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[{"name":"In house (NVWA)","role":"in-house"}],"summary":"The NVWA, the Dutch authority that checks among other things whether pig farmers care for their animals properly, predicts for every pig farm the chance that it does not comply with animal welfare rules. So far the model has been rebuilt each time it is used, comparing several supervised machine learning techniques on past inspection results and registry data and choosing the best predictor on a held out test set. People set how many farms go on the inspection list and check the list by hand, and every farm on it gets a normal inspection. The authority keeps inspecting randomly selected farms to test whether the model finds more problems, compares the selected farms with the whole population to spot farm types that are always picked or always skipped, and makes sure inspectors never know for certain whether the model selected a farm. In use since March 2022; a similar model covers dairy cattle welfare.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/oorg10102/49428458/nalevingsmodel-varkenswelzijn","title":"Nalevingsmodel Varkenswelzijn, Algoritmeregister","publisher":"Algoritmeregister van de Nederlandse overheid","date":"2026-05-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"nvwa-pig-welfare-compliance-model"},{"title":"O2 Telefónica Germany: AI energy saving on its Nokia radio network","useCases":["ran-energy-optimization"],"organization":{"name":"O2 Telefónica Germany","anonymized":false,"country":"DE","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Nokia","role":"platform"}],"summary":"O2 Telefónica Germany chose Nokia's AVA for Energy software, delivered as a service, for the parts of its radio network built on Nokia equipment. The software monitors traffic patterns and throttles back resources such as base stations during low usage, while monitoring quality so that customers do not notice the change. In its test the operator switched off unused radio resources automatically and saw significant savings, but no figure specific to O2 Telefónica Germany is published.","stage":"production","year":2023,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.nokia.com/newsroom/nokia-ava-for-energy-saas-chosen-by-o2-telefonica-germany-to-curb-energy-use-mwc23/","title":"Nokia AVA for Energy SaaS chosen by O2 Telefónica Germany to curb energy use #MWC23","publisher":"Nokia","date":"2023-02-24","archivedUrl":"https://web.archive.org/web/20250124103611/https://www.nokia.com/about-us/news/releases/2023/02/24/nokia-ava-for-energy-saas-chosen-by-o2-telefonica-germany-to-curb-energy-use-mwc23/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"o2-telefonica-germany-nokia-energy-saas"},{"title":"OakNorth Bank: data driven SME underwriting and continuous borrower monitoring","useCases":["sme-cash-flow-underwriting","credit-early-warning-monitoring"],"organization":{"name":"OakNorth Bank","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[{"name":"OakNorth","role":"in-house"}],"summary":"OakNorth Bank, a UK lender to small and mid sized businesses, underwrites with human credit officers supported by systems that pull in and analyse public and alternative data, and monitors each borrower continuously against a peer group in the same sector and geography rather than waiting for audited financials every six months. By late 2020 it had lent GBP 4.6 billion to 750 businesses since 2016 and sold the same software to other banks. The published figures describe the lending book, not a measured effect of the AI.","stage":"scaled","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.euromoney.com/article/27sic7y97uvu96j2fuc5c/fintech/smbc-uses-oaknorths-credit-intelligence-software-to-grow-lending/","title":"SMBC uses OakNorth's credit intelligence software to grow lending","publisher":"Euromoney","date":"2020-11-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"oaknorth-bank-continuous-credit-monitoring"},{"title":"OCBC and Bank of Singapore: HELIOS agentic AI for customer due diligence in private banking","useCases":["pep-and-adverse-media-screening","perpetual-kyc"],"organization":{"name":"OCBC","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"OCBC","role":"in-house"}],"summary":"OCBC launched HELIOS in July 2026, an agentic AI platform that gathers intelligence on prospective private banking clients and completes most of the customer due diligence before a relationship manager engages them. OCBC says private banking accounts can now be opened in 15 business days, against an industry median of about six weeks, while relationship managers and review teams keep accountability for judgment and decisions. OCBC plans to extend HELIOS to ongoing monitoring of customer activity to detect changes in risk profiles. Bank of Singapore relationship managers use it in Singapore, Hong Kong and Dubai, with the rollout due to finish in the third quarter of 2026.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.ocbc.com/group/media/release/2026/ocbc-harnesses-agentic-ai-to-quicken-onboarding-customers.page","title":"OCBC harnesses agentic AI to sharpen and quicken onboarding of wealthy customers","publisher":"OCBC","date":"2026-07-29"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ocbc-helios-agentic-due-diligence"},{"title":"Octopus Energy: Arlo, an AI assistant that answers routine customer emails","useCases":["utility-billing-and-move-agent"],"organization":{"name":"Octopus Energy","anonymized":false,"country":"GB","region":"europe","industry":"energy-and-utilities"},"vendors":[{"name":"Kraken","role":"platform"}],"summary":"Octopus Energy and Kraken built Arlo, an AI assistant that answers straightforward customer emails about tariff renewals, payment dates and account details. Every AI written email is labelled, the customer can ask for a human at any time, and vulnerable customers, sensitive cases and complex complaints always go to the human team. In a three month UK trial Arlo handled around 8,000 emails a week (4% of customer emails) and scored 76% customer satisfaction against 72% for comparable human replies; Octopus then started a wider rollout while human experts keep reviewing its messages.","stage":"production","year":2026,"channels":["email"],"languages":["en"],"metrics":[{"kpi":"customer-satisfaction","value":76,"unit":"percent","qualifier":"exact","period":"three month trial, versus 72% for comparable human replies","claimant":"organization","quote":"Arlo achieved a 76% customer satisfaction score, beating comparable responses from human advisors, which scored 72%.","sourceUrl":"https://octopus.energy/press/more-news-press-releases/octopus-energy-s-ai-trial-wins-customer-approval/"},{"kpi":"interactions-handled","value":8000,"unit":"count","qualifier":"approximately","period":"per week during the trial","claimant":"organization","quote":"During the three-month trial, Arlo handled around 8,000 emails a week","sourceUrl":"https://octopus.energy/press/more-news-press-releases/octopus-energy-s-ai-trial-wins-customer-approval/"}],"outcomeDisclosed":true,"sources":[{"url":"https://octopus.energy/press/more-news-press-releases/octopus-energy-s-ai-trial-wins-customer-approval/","title":"Octopus Energy's AI trial wins customer approval","publisher":"Octopus Energy","date":"2026-07-07"},{"url":"https://octopus.energy/blog/arlo-ai-assistant/","title":"How Arlo, our AI assistant, is helping our humans help you","publisher":"Octopus Energy","date":"2026-08-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"octopus-energy-arlo-email-assistant"},{"title":"The ODP Corporation: sales assistant that matches SKUs and drafts quotes for pricing bids","useCases":["sales-quote-and-estimate-generation"],"organization":{"name":"The ODP Corporation","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"The ODP Corporation, parent of Office Depot and ODP Business Solutions, built a sales assistant on Azure OpenAI and Azure AI Search that lets representatives generate quotes in natural language. Its SKU matching cross references large product lists in minutes instead of days and produces quote ready summaries and tables; in one anecdote a representative cross referenced 4,000 competitor items. Microsoft reports that pricing bids now take hours instead of one to two days, that tailored quotes save representatives five to eight hours a week, and that the assistant drives 20% more sales opportunities a quarter.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/24030-the-odp-corporation-azure-ai-foundry","title":"The ODP Corporation transforms HR, sales, and retail workflows with Azure AI app platform","publisher":"Microsoft","date":"2025-05-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"odp-corporation-sales-quote-assistant"},{"title":"US federal government: consolidated AI use case inventory","useCases":["ai-model-inventory"],"organization":{"name":"Office of Management and Budget","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"US federal agencies must inventory their AI use cases every year, submit the inventory to the Office of Management and Budget and publish the releasable part as machine readable data. OMB consolidates the agency inventories in a public repository with a fixed schema (purpose, stage, vendor, data, risk designation). The 2025 consolidation, as of 13 April 2026, lists 3,611 individually reported AI use cases, 445 of them high impact, plus separately consolidated commercial off the shelf AI uses; the 2024 consolidation listed 2,133 use cases from 41 agency submissions.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"ai-systems-inventoried","value":3611,"unit":"count","qualifier":"exact","period":"2025 consolidated federal inventory, individually reported use cases in all stages, as of 13 April 2026","claimant":"organization","quote":"3,611 individually-reported AI use cases (all stages of development)","sourceUrl":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory"}],"outcomeDisclosed":true,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://github.com/ombegov/2024-Federal-AI-Use-Case-Inventory","title":"2024 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"},{"url":"https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf","title":"M-25-21: Accelerating Federal Use of AI through Innovation, Governance, and Public Trust","publisher":"Office of Management and Budget","date":"2025-04-03"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"omb-federal-ai-use-case-inventory"},{"title":"One Stop: machine learning store forecasts for fresh and weather driven products","useCases":["retail-demand-forecasting-and-replenishment"],"organization":{"name":"One Stop","anonymized":false,"country":"GB","region":"europe","industry":"retail-and-ecommerce"},"vendors":[{"name":"RELEX Solutions","role":"platform"}],"summary":"One Stop, the Tesco owned convenience chain with more than 900 company and franchise stores in Great Britain, moved store and distribution center forecasting and replenishment to RELEX in 2019; RELEX reports that this first rollout raised store availability by 1.9 percentage points and cut fresh spoilage value by 4%. One Stop then added machine learning forecasting to handle short shelf life lines, weather driven demand such as ice and cannibalization between promoted products. RELEX reports that within four months forecast accuracy rose by 3.17 percentage points at product and week level and 1.82 points at product, store and week level, and One Stop's Head of Supply Chain says availability of ultra fresh products with under three days of shelf life rose 8.5% with no corresponding rise in spoilage.","stage":"scaled","year":2022,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.relexsolutions.com/resources/case-one-stop/","title":"Case study: One Stop","publisher":"RELEX Solutions","date":"2022-04-11"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"one-stop-relex-machine-learning-forecasting"},{"title":"Openreach: Crystal Ball and Ask Me Anything for delayed fibre installations","useCases":["field-technician-copilot-and-dispatch"],"organization":{"name":"Openreach","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"CXone Proactive AI Agent","role":"platform"}],"summary":"When an Openreach engineer finds a problem during a full fibre installation, a prediction tool called Crystal Ball predicts what type of work is likely to be needed and whether it will take more or less than ten days. That prediction drives a clear text message update to the customer, typically within 24 hours. A generative AI capability, Ask Me Anything, lets customers ask questions about their installation in their own words. Both are part of Openreach's work with the CXone Proactive AI Agent platform. Openreach says the tools help prevent more than 3,000 cancelled orders a month.","stage":"production","year":2026,"channels":["sms"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.openreach.com/news/openreach-ai-tools-help-cut-cancelled-full-fibre-orders-by-three-thousand-a-month/","title":"Openreach AI tools help cut cancelled full fibre orders by three thousand a month","publisher":"Openreach","date":"2026-09-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"openreach-crystal-ball-installation-prediction"},{"title":"Oper Credits: AI document verification for mortgage applications","useCases":["conversational-loan-application-intake"],"organization":{"name":"Oper Credits","anonymized":false,"country":"BE","region":"europe","industry":"technology"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Oper Credits, a Belgian mortgage digitisation company that serves about 20 banks in six countries, uses Vertex AI to automate document verification that used to take several hours of manual work. According to Google Cloud, only 30 to 40% of loan applications in Belgium are complete and compliant on first submission, and most are returned for missing or incorrect information. The company aims to raise that to 90%; the aim is a target, not a reported result.","stage":"production","year":2025,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"oper-credits-mortgage-document-verification"},{"title":"Oportun: from sample based QA to monitoring every call","useCases":["call-quality-and-compliance-monitoring","live-agent-assist"],"organization":{"name":"Oportun","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Cresta","role":"platform"}],"summary":"Oportun, a US consumer lender, replaced manual, sample based QA with Cresta's AI quality management across all calls, combined with real time guidance for agents. Coaching now focuses on the behaviours that drive performance, visible across every call, instead of a small sample reviewed weeks later. No quantified outcome is published.","stage":"scaled","year":2024,"channels":["voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.cresta.com/customers/oportun","title":"How Oportun moved from sample-based QA to 100% interaction monitoring with Cresta","publisher":"Cresta"},{"url":"https://web.archive.org/web/20240423210739/https://cresta.com/customers/oportun/","title":"How Oportun transformed QM and reduced workload by 50% with Cresta","publisher":"Cresta"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"oportun-ai-quality-management"},{"title":"Orange: AI alarm correlation and anomaly detection in the network operations centre","useCases":["network-fault-triage-copilot","predictive-network-maintenance"],"organization":{"name":"Orange","anonymized":false,"country":"FR","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Augtera Networks","role":"platform"}],"summary":"After a two year production trial on the French backbone, Orange Global Network and an SD-WAN network, Orange added the Augtera Network AI platform to its NOC tools. Topology based auto correlation groups alarms so operations experts see far fewer of them, and anomaly detection on metrics and logs flags weak signals so incidents can be handled before customers notice. Orange and Augtera say the correlation will cut the daily number of alarms the NOC has to address by 70%, presented as the expected effect of the rollout rather than a measured result. The integration was due to start in April 2024 in Orange Global Networks, an IP network with thousands of routers in 800 points of presence across 100 countries, with full rollout planned by the end of 2024.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.orange.com/orange-introduces-augtera-network-ai-platform-to-offer-best-in-class-quality-of-service-and-customer-experience/","title":"Orange Introduces Augtera Network AI Platform to offer best-in-class quality of service and customer experience","publisher":"Orange","date":"2024-04-11"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"orange-augtera-noc-alarm-correlation"},{"title":"Orange France: Mon Assistant IA for sales advisors and the Sharlie voice assistant","useCases":["plan-upgrade-and-sales-assistant"],"organization":{"name":"Orange France","anonymized":false,"country":"FR","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Verint","role":"platform"},{"name":"Microsoft","role":"platform"},{"name":"ILLUIN Technology","role":"platform"}],"summary":"Orange France deployed Mon Assistant IA (MAIA) with Verint to 3,000 Orange sales advisors, launching it in early December 2025. During calls it understands the conversation, detects customer needs, retrieves relevant information and summarises the exchange to update the customer file; the advisor validates the proposed responses. Orange also announced Sharlie, a speech to speech voice assistant for its digital brand Sosh built with Microsoft and ILLUIN Technology on the ILLUIN Dialogue and Microsoft Foundry platforms, with capacity for over 3 million conversations a year once deployed.","stage":"scaled","year":2025,"channels":["agent-desktop","voice"],"languages":["fr"],"metrics":[{"kpi":"interactions-handled","value":1000000,"unit":"count","qualifier":"approximately","period":"conversations supported per month, as of March 2026","claimant":"organization","quote":"Launched in early December 2025, Mon Assistant IA already supports nearly 1 million conversations per month.","sourceUrl":"https://newsroom.orange.com/orange-france-launches-two-new-artificial-intelligence-services-to-enhance-customer-relations/"}],"outcomeDisclosed":true,"sources":[{"url":"https://newsroom.orange.com/orange-france-launches-two-new-artificial-intelligence-services-to-enhance-customer-relations/","title":"Orange France launches two new artificial intelligence services to enhance customer relations","publisher":"Orange","date":"2026-03-17"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"orange-france-maia-advisor-assistant"},{"title":"Origin Bank: agentic AI for enhanced due diligence reviews with Nasdaq Verafin","useCases":["perpetual-kyc"],"organization":{"name":"Origin Bank","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Nasdaq Verafin","role":"platform"}],"summary":"Origin Bank, a US bank with about USD 10 billion in assets, uses Nasdaq Verafin's agentic AI workforce for its enhanced due diligence reviews of high risk customers. Nasdaq Verafin's Digital EDD Analyst automates the bank's periodic EDD review process, closing low risk cases itself and escalating the rest. The vendor reports a large increase in the number of EDD reviews completed.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://verafin.com/resource/origin-bank-proves-the-business-case-for-the-agentic-ai-workforce/","title":"Origin Bank Proves the Business Case for the Agentic AI Workforce","publisher":"Nasdaq Verafin"},{"url":"https://verafin.com/news/nasdaq-verafin-announces-launch-of-its-agentic-ai-workforce-delivering-a-step-change-in-aml-compliance-efficiency/","title":"Nasdaq Verafin Announces Launch of its Agentic AI Workforce","publisher":"Nasdaq Verafin","date":"2025-07-21"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"origin-bank-verafin-agentic-edd-reviews"},{"title":"Oyster: automated research, enrichment and tailored messaging for intent based outbound","useCases":["outbound-sales-prospecting-agent"],"organization":{"name":"Oyster","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Clay","role":"platform"}],"summary":"Oyster, a global employment platform, ran an intent based outbound program that depended on manual account research and enrichment across G2, Clearbit, Salesforce, HubSpot and other tools. Its marketing operations team automated research, qualification, enrichment and message segmentation in Clay, generating content tailored to each intent signal and syncing accounts to the right BDR and email tool while keeping the CRM as the system of record. BDR leaders wanted the automation to support strategic account selection and personalized outreach, and the workflows were designed to safeguard those internal processes and rules of engagement. The vendor reports that Petra Hajal, who built the workflows, estimates each representative now saves approximately 40 hours per month.","stage":"production","year":2023,"channels":["email","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.clay.com/customers/oyster","title":"How Oyster uses Clay to run intent-based outbound campaigns, saving 40hrs/month per sales rep","publisher":"Clay"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"oyster-intent-based-outbound-automation"},{"title":"Pacers Sports & Entertainment: custom speech model for live arena and app captions","useCases":["audio-and-video-transcription-and-captioning"],"organization":{"name":"Pacers Sports & Entertainment","anonymized":false,"country":"US","region":"north-america","industry":"media-and-entertainment"},"vendors":[{"name":"Microsoft (Azure AI Foundry, Azure AI Speech)","role":"platform"}],"summary":"Pacers Sports & Entertainment trained a custom speech model on hundreds of hours of its own game broadcasts, with lists of player, coach and official names, to caption live announcers for fans who are deaf, hard of hearing or do not speak English, on arena screens and in its mobile apps. English came first, then Spanish, and Microsoft reports the team is adding 12 more languages. It is the live variant of the job: captions are delivered in real time, and Microsoft reports built in moderation filters that help keep inappropriate or misinterpreted content off the screen. The system now captions Pacers, Fever and All Star games at Gainbridge Fieldhouse.","stage":"production","year":2025,"channels":["mobile-app","kiosk"],"languages":["en","es"],"metrics":[{"kpi":"error-reduction","value":87,"unit":"percent","qualifier":"exact","baseline":"Transcription error rate of the speech to text model before it was tuned to the Pacers broadcast style","claimant":"vendor","quote":"By tailoring the model to the Pacers’ broadcast style, they reduced the speech-to-text transcription error rate by 87%—a level of precision that made it possible to extend captioning from mobile apps to arena screens with confidence.","sourceUrl":"https://www.microsoft.com/en/customers/story/23957-pacers-sports-and-entertainment-azure-ai-foundry"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23957-pacers-sports-and-entertainment-azure-ai-foundry","title":"Indiana Pacers use Azure AI Foundry to create first live arena captioning service","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"pacers-sports-and-entertainment-live-arena-captions"},{"title":"Paragon Insurance Group: automated submission ingestion, clearance and prioritization with Kalepa","useCases":["commercial-underwriting-submission-triage"],"organization":{"name":"Paragon Insurance Group","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Kalepa","role":"platform"}],"summary":"Paragon runs about 25 specialty programs, some of which receive upward of 50,000 submissions a year, and its underwriters could review only about 30% of incoming submissions. With Kalepa, every submission is ingested, cleared and ranked by fit and likelihood to bind within minutes, with research from news, legal filings and third party data attached. Paragon's CTO says extraction is now around 98 to 99% accurate, better than the former manual operations team, and its E&S president says the quote to bind ratio doubled within the first year.","stage":"scaled","year":2025,"channels":["email","internal-tools"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":98,"unit":"percent","qualifier":"approximately","period":"submission data extraction, after full inbox rollout","baseline":"manual extraction by an operations team","claimant":"organization","quote":"We're somewhere around 98 to 99% accurate now - even more accurate than when we had an operations team doing this manually.","sourceUrl":"https://www.kalepa.com/case-studies/paragon-doubled-quote-to-bind-rate"},{"kpi":"conversion-rate-uplift","value":2,"unit":"multiplier","qualifier":"exact","period":"quote to bind ratio, first year","claimant":"organization","quote":"We're seeing a better quote-to-bind ratio. In the past year it has doubled from what it was before.","sourceUrl":"https://www.kalepa.com/case-studies/paragon-doubled-quote-to-bind-rate"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.kalepa.com/case-studies/paragon-doubled-quote-to-bind-rate","title":"How Paragon Doubled Its Quote-to-Bind Rate and Achieved 99% Submission Accuracy with Kalepa","publisher":"Kalepa","date":"2026-03-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"paragon-kalepa-submission-triage"},{"title":"Patelco Credit Union: open banking credit score pilot","useCases":["alternative-data-credit-scoring"],"organization":{"name":"Patelco Credit Union","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"VantageScore","role":"platform"}],"summary":"Patelco Credit Union tested VantageScore 4plus, a score that combines credit file data with consumer permissioned open banking (bank account cash flow) data, on its own portfolio. In the pilot, 12% of subprime and 15% of near prime members moved to higher credit tiers, and predictive power in originations improved by 4.8% over VantageScore 3.0. It was a portfolio test, not a production rollout. The larger figures in the press release headline (33% and 41%) come from the second pilot at Michigan State University Federal Credit Union, not from Patelco.","stage":"pilot","year":2025,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.prnewswire.com/news-releases/vantagescore-4plus-pilots-find-33-of-subprime-and-41-of-near-prime-consumers-moved-to-higher-credit-tiers-by-adding-open-banking-data-302487924.html","title":"VantageScore 4plus Pilots Find 33% of Subprime and 41% of Near Prime Consumers Moved to Higher Credit Tiers by Adding Open Banking Data","publisher":"VantageScore via PR Newswire","date":"2025-06-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"patelco-credit-union-open-banking-score-pilot"},{"title":"PatientPoint: AI triage and remediation assistance for application security findings","useCases":["software-vulnerability-remediation"],"organization":{"name":"PatientPoint","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Checkmarx","role":"platform"}],"summary":"PatientPoint, a US healthcare company, used Checkmarx Triage Assist and Remediation Assist when its leadership asked the application security team to remediate vulnerabilities in a short period, while developers using AI to write code were adding findings faster than before. Checkmarx says the tools filtered out false positives before they reached developers and generated fix guidance the team could review; PatientPoint's application security engineer calls the tools very accurate and says they identified false positives and left room for human review. No figures are given.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://checkmarx.com/resources/on-point-fixes-how-patientpoint-outpaced-its-own-vulnerability-backlog/","title":"On-Point Fixes: How PatientPoint Outpaced Its Own Vulnerability Backlog","publisher":"Checkmarx","date":"2026-07-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"patientpoint-checkmarx-ai-triage-and-remediation"},{"title":"Patterson Dental: deidentified, production like test data for performance testing","useCases":["synthetic-test-data-generation"],"organization":{"name":"Patterson Dental","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Tonic.ai","role":"platform"}],"summary":"Patterson Dental, a division of Patterson Companies, uses Tonic Structural to generate deidentified, production like data for performance and functional testing of its dental practice platforms, so protected health information stays out of developer workflows. The vendor reports that test data preparation fell from 2.5 hours to 35 minutes per dataset and that performance testing grew from one practice to between 15 and 25 practices a day, with up to seven development teams using the data.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":75,"unit":"percent","qualifier":"exact","period":"Test data generation time per dataset","baseline":"2.5 hours per dataset, prepared manually from production","claimant":"vendor","quote":"With Tonic Structural, the company achieved immediate and measurable improvements in their testing workflows, reducing test data generation time by 75% and cutting it down from 2.5 hours to just 35 minutes.","sourceUrl":"https://www.tonic.ai/case-study/patterson-test-data-better-software-for-thousands"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.tonic.ai/case-study/patterson-test-data-better-software-for-thousands","title":"Patterson cuts test data time by 75% with Tonic.ai, delivering better software for thousands worldwide","publisher":"Tonic.ai","date":"2025-03-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"patterson-dental-deidentified-test-data"},{"title":"Pay.UK and Visa: AI fraud scoring pilot on UK account to account payments","useCases":["real-time-fraud-scoring"],"organization":{"name":"Pay.UK","anonymized":false,"country":"GB","region":"europe","industry":"payments"},"vendors":[{"name":"Visa","role":"platform"}],"summary":"In a pilot with Pay.UK, which runs the UK's retail payment operations, Visa applied AI risk scoring to billions of historic UK account to account transactions covering 12 months and more than half of annual volume. It identified 54% of the fraudulent transactions that had already passed through the banks' own fraud detection systems. The pilot was retrospective, on historical data, and on the same day Visa made the capability available to UK banks as a real time service, Visa Protect for A2A Payments.","stage":"pilot","year":2024,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.visa.co.uk/about-visa/newsroom/press-releases.3326480.html","title":"Visa's new AI tool for Faster Payments could help save UK over £330m a year on fraud and APP scams","publisher":"Visa UK","date":"2024-05-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"pay-uk-visa-account-to-account-fraud-pilot"},{"title":"Payoneer: AI document forensics in customer onboarding with Resistant AI","useCases":["application-and-identity-fraud-detection"],"organization":{"name":"Payoneer","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[{"name":"Resistant AI","role":"platform"}],"summary":"Cross border payments platform Payoneer added Resistant AI's document forensics to its intelligent document processing in onboarding, to detect fake documents and serial fraud attempts while keeping onboarding fast. Resistant AI says it now supports more than 82% of Payoneer's document fraud decision making, with only edge cases escalated for manual review.","stage":"scaled","year":2023,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://resistant.ai/case-studies/payoneer","title":"Payoneer","publisher":"Resistant AI","archivedUrl":"https://web.archive.org/web/20230205042853/https://resistant.ai/case-studies/payoneer/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"payoneer-resistant-ai-document-forensics"},{"title":"PayPal: Instant Buy checkout inside Perplexity's shopping agent","useCases":["agentic-payment-initiation"],"organization":{"name":"PayPal","anonymized":false,"country":"US","region":"north-america","industry":"payments"},"vendors":[{"name":"Perplexity","role":"platform"}],"summary":"In November 2025 PayPal enabled U.S. users of Perplexity's shopping experience to check out with PayPal without leaving the answer engine, from merchants such as Abercrombie & Fitch, Ashley Furniture, Fabletics, Adorama and Newegg. PayPal's agentic commerce services sync merchant catalogs and let merchants accept agent originated payments with their existing PayPal setup, while the retailer stays merchant of record and transactions pass PayPal's fraud screening, identity verification and purchase protection. No volumes or conversion figures are disclosed.","stage":"production","year":2025,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.paypal-corp.com/2025-11-PayPal-and-Perplexity-Launch-Instant-Buy","title":"PayPal and Perplexity Launch Instant Buy Ahead of Black Friday","publisher":"PayPal Newsroom","date":"2025-11-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"paypal-perplexity-instant-buy"},{"title":"PBGC: generative AI for IT security and privacy control assessment (planned)","useCases":["continuous-controls-testing"],"organization":{"name":"Pension Benefit Guaranty Corporation","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Pension Benefit Guaranty Corporation reports a use case in pre deployment that applies AI to IT security and privacy control assessments: evaluating controls against federal cybersecurity and privacy guidelines, working through the supporting evidence, generating findings and drafting control implementation statements. The stated aims are a higher volume of control assessments, less manual work and shorter control tailoring and implementation times. It is not yet live and no results are published.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry PBGC - 11, IT Security and Privacy Control Assessment)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"pbgc-security-and-privacy-control-assessment"},{"title":"Pegasus Airlines: FlyBot generative AI virtual assistant","useCases":["flight-disruption-and-rebooking-agent","first-line-contact-centre-agent"],"organization":{"name":"Pegasus Airlines","anonymized":false,"country":"TR","region":"europe","industry":"travel-and-hospitality"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Pegasus Airlines retrained FlyBot, the virtual assistant on its website, with Azure OpenAI and integrated it with internal systems, so customers can ask about flights, flight rules, baggage allowances and claims and reissue tickets in the same conversation. The airline reports that satisfaction with the virtual assistant doubled after the change.","stage":"production","year":2025,"channels":["web-chat"],"languages":[],"metrics":[{"kpi":"customer-satisfaction-uplift","value":100,"unit":"percent","qualifier":"exact","claimant":"organization","quote":"“Since we integrated Azure AI Services into our FlyBot, customer satisfaction rates for our virtual assistant have doubled,” points out Bora.","sourceUrl":"https://www.microsoft.com/en/customers/story/23194-pegasus-airlines-azure-ai-services"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23194-pegasus-airlines-azure-ai-services","title":"Pegasus Airlines transforms bookings and services with Azure OpenAI, doubling customer satisfaction scores","publisher":"Microsoft"},{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft","date":"2025-07-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"pegasus-airlines-flybot-virtual-assistant"},{"title":"Pegatron: visual AI agent that checks manual assembly steps and synthetic defect images for inspection models","useCases":["production-line-quality-inspection"],"organization":{"name":"Pegatron","anonymized":false,"country":"TW","region":"asia-pacific","industry":"manufacturing"},"vendors":[{"name":"NVIDIA","role":"platform"},{"name":"Pegatron","role":"in-house"}],"summary":"Pegatron, an electronics manufacturer with 24 sites, built an Assembly Guiding Agent on NVIDIA's video search and summarization blueprint that watches manual assembly in real time, spots missed steps such as a forgotten screw and alerts the worker, who can replay the clip and ask the agent questions. It also generates synthetic defect images from CAD drawings and historical data to train its visual inspection models, because real defect images are scarce on high yield lines. NVIDIA reports lower defect rates and labour cost per line for the assembly agent. The page is undated; the assembly agent dates from 2025, while the synthetic defect image work may be later.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"error-reduction","value":67,"unit":"percent","qualifier":"exact","period":"assembly lines using the Assembly Guiding Agent","claimant":"vendor","quote":"By augmenting the assembly process with this AI agent, Pegatron is seeing a 7% reduction in labor costs per assembly line and a 67% decrease in defect rates.","sourceUrl":"https://www.nvidia.com/en-us/case-studies/pegatron-scales-factory-operations-with-visual-ai-digital-twins/"},{"kpi":"cost-reduction","value":7,"unit":"percent","qualifier":"exact","period":"labour cost per assembly line","claimant":"vendor","quote":"By augmenting the assembly process with this AI agent, Pegatron is seeing a 7% reduction in labor costs per assembly line and a 67% decrease in defect rates.","sourceUrl":"https://www.nvidia.com/en-us/case-studies/pegatron-scales-factory-operations-with-visual-ai-digital-twins/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.nvidia.com/en-us/case-studies/pegatron-scales-factory-operations-with-visual-ai-digital-twins/","title":"Pegatron Scales Factory Operations With Visual AI Agents, and Digital Twins","publisher":"NVIDIA"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"pegatron-visual-ai-assembly-inspection"},{"title":"Pfizer: AI platform for adverse event case intake in drug safety","useCases":["adverse-event-case-intake"],"organization":{"name":"Pfizer","anonymized":false,"country":"US","region":"north-america","industry":"pharma-and-life-sciences"},"vendors":[],"summary":"Pfizer's Worldwide Safety organization, which processed about 1.4 million adverse events in 2019, worked with industry experts to build an AI platform for the repetitive intake steps of adverse event case processing. In its first phase the model makes basic intake decisions, such as whether a report is a valid case and whether it is fatal or life threatening. The Drug Safety Unit in Rome was the first location to use it in live operations, and Pfizer described the aim as freeing safety professionals for signal detection and investigation rather than replacing them. No outcome figures are published.","stage":"production","year":2020,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.pfizer.com/news/articles/ai-drug-safety-building-elusive-%E2%80%98loch-ness-monster%E2%80%99-reporting-tools","title":"AI in Drug Safety: Building the Elusive 'Loch Ness Monster' of Reporting Tools","publisher":"Pfizer","date":"2020-04-20","archivedUrl":"https://web.archive.org/web/2026/https://www.pfizer.com/news/articles/ai-drug-safety-building-elusive-%E2%80%98loch-ness-monster%E2%80%99-reporting-tools"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"pfizer-adverse-event-case-intake"},{"title":"PG&E: generative voice agent for outage and billing calls","useCases":["utility-billing-and-move-agent"],"organization":{"name":"Pacific Gas and Electric Company","anonymized":false,"country":"US","region":"north-america","industry":"energy-and-utilities"},"vendors":[{"name":"PolyAI","role":"platform"}],"summary":"PG&E, which receives about 16 million calls a year with sharp peaks during storms and outages, deployed a PolyAI voice agent named Peggy. It authenticates customers, gives location based outage updates, answers billing questions and FAQs in English and Spanish and texts links for self service, with integrations into Oracle, Cisco and in house systems. PolyAI reports 67% overall containment, 35,000 labour hours saved and a 22% increase in CSAT on outage calls. Start and stop service, appointment setting and expanded billing were named as the next use cases, not yet reported as live.","stage":"scaled","year":2024,"channels":["voice","sms"],"languages":["en","es"],"metrics":[{"kpi":"containment-rate","value":67,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"35,000 labor hours have been saved by the PolyAI agent, increasing CSAT by 22% and achieving 67% containment, 6% higher than the legacy IVR","sourceUrl":"https://poly.ai/customers/pge"},{"kpi":"hours-saved","value":35000,"unit":"hours","qualifier":"exact","claimant":"vendor","quote":"35,000 labor hours have been saved by the PolyAI agent, increasing CSAT by 22% and achieving 67% containment, 6% higher than the legacy IVR","sourceUrl":"https://poly.ai/customers/pge"},{"kpi":"customer-satisfaction-uplift","value":22,"unit":"percent","qualifier":"exact","period":"outage calls","claimant":"vendor","quote":"35,000 labor hours have been saved by the PolyAI agent, increasing CSAT by 22% and achieving 67% containment, 6% higher than the legacy IVR","sourceUrl":"https://poly.ai/customers/pge"}],"outcomeDisclosed":true,"sources":[{"url":"https://poly.ai/customers/pge","title":"How Pacific Gas and Electric saved 35,000 labor hours with PolyAI","publisher":"PolyAI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"pge-polyai-voice-agent-billing-and-outages"},{"title":"PNC: OakNorth credit monitoring to read pandemic impact across the loan book","useCases":["credit-early-warning-monitoring"],"organization":{"name":"PNC Financial Services","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"OakNorth","role":"platform"}],"summary":"In mid 2020 PNC, a large US regional bank, took OakNorth's credit monitoring system to understand the impact of the pandemic across its loan portfolios. The system models each borrower against sector and local peers with frequently updated data, such as reviews, footfall and pricing, instead of relying on lagging audited financials. OakNorth said it delivered the system within a week of the first conversation and that such monitoring often reveals a subset of loans where borrowers who are still current might well be heading for difficulty. PNC published no outcome figures.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.euromoney.com/article/27sic7y97uvu96j2fuc5c/fintech/smbc-uses-oaknorths-credit-intelligence-software-to-grow-lending/","title":"SMBC uses OakNorth's credit intelligence software to grow lending","publisher":"Euromoney","date":"2020-11-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"pnc-oaknorth-portfolio-monitoring"},{"title":"PostNL: renewed chatbot and conversational AI in customer service","useCases":["parcel-tracking-and-delivery-exception-agent"],"organization":{"name":"PostNL","anonymized":false,"country":"NL","region":"europe","industry":"logistics-and-transportation"},"vendors":[],"summary":"PostNL, the Dutch postal and parcel company, reports in its 2025 annual report that it renewed its customer service chatbot and introduced conversational AI to support service agents, to give faster and more consistent answers to common questions such as \"where is my parcel?\", and that it migrated to a new integrated conversational platform. The report names the chatbot Daan, available around the clock, and states PostNL's ambition to become an AI first organisation.","stage":"production","year":2025,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://annualreport.postnl.nl/app/uploads/2026/02/PostNL-Annual-Report-2025.pdf","title":"PostNL Annual Report 2025","publisher":"PostNL"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"postnl-chatbot-and-conversational-ai-customer-service"},{"title":"Priceline: Penny agentic AI travel assistant","useCases":["travel-and-hotel-booking-concierge"],"organization":{"name":"Priceline","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[{"name":"Anthropic","role":"model-provider"},{"name":"Google Cloud","role":"model-provider"},{"name":"OpenAI","role":"model-provider"},{"name":"Priceline","role":"in-house"}],"summary":"Priceline put its AI assistant Penny in front of customers first at checkout and in customer care. In June 2026 it announced a fully agentic version that takes a trip idea, compares hotels, flights and rental cars on a live map with real time inventory and deals, and books without leaving the conversation. Penny runs as more than ten specialized agents on Priceline's own AI stack, with Claude models for conversational reasoning and planning, and Google Cloud and OpenAI supporting search and voice capabilities.","stage":"production","year":2026,"channels":["mobile-app","web-chat"],"languages":[],"metrics":[{"kpi":"time-saved-per-task","value":10,"unit":"minutes","qualifier":"approximately","period":"per trip, Penny users versus customers who called support","claimant":"organization","quote":"Priceline has estimated that travelers who used Penny saved an average of nearly ten minutes per trip compared with those who called customer support.","sourceUrl":"https://press.priceline.com/pricelines-penny-goes-fully-agentic/"}],"outcomeDisclosed":true,"sources":[{"url":"https://press.priceline.com/pricelines-penny-goes-fully-agentic/","title":"Priceline's Penny Goes Fully Agentic","publisher":"Priceline Press Center","date":"2026-06-03"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"priceline-penny-agentic-travel-assistant"},{"title":"Progressive: digital claims journey with a generative AI assistant in claims messaging","useCases":["claims-first-notice-of-loss-agent"],"organization":{"name":"Progressive","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[],"summary":"In its 2025 letter to shareholders, Progressive says it implemented digital claims capabilities in 2025 that let customers interact from the first notice of loss through investigation, damage assessment and repair. The same program gave claims employees a new text and email communication platform that includes a customer facing generative AI assistant for automated tasks, information retrieval and tailored follow up actions. No outcome figures were published.","stage":"production","year":2025,"channels":["sms","email"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/80661/000008066126000086/pgr-20251231exhibit99.htm","title":"The Progressive Corporation 2025 Annual Report, Letter to Shareholders (Exhibit 99 to Form 10-K)","publisher":"The Progressive Corporation (via SEC EDGAR)","date":"2026-03-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"progressive-digital-claims-and-generative-ai-assistant"},{"title":"Province of North Holland: document analysis and redaction support for open government requests","useCases":["freedom-of-information-request-processing"],"organization":{"name":"Provincie Noord-Holland","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[{"name":"ZyLAB (Reveal)","role":"platform"}],"summary":"The Province of North Holland uses ZyLAB to handle large requests under the Dutch Open Government Act (Woo). After staff draw up a search plan and collect the potentially relevant documents, the platform makes the set searchable, including text recognition for scanned documents, helps staff judge relevance and, on instruction, produces a trial redacted version: generic rules recognise items such as phone and citizen service numbers, while names need their own individual rules, and staff switch the rules on themselves. Staff check every document to avoid too much or too little redaction, and what is released is agreed between the responsible staff and lawyers. In use since October 2021.","stage":"production","year":2021,"channels":["internal-tools"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/pv27/86188119/ondersteuning-openbaarmakingsverzoeken","title":"Ondersteuning Openbaarmakingsverzoeken, Algoritmeregister","publisher":"Algoritmeregister van de Nederlandse overheid"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"provincie-noord-holland-woo-request-support"},{"title":"Prudential plc: AI lead validation talkbot, PruAction and a health AI chatbot for agents","useCases":["insurance-broker-and-agent-assistant"],"organization":{"name":"Prudential plc","anonymized":false,"country":"HK","region":"asia-pacific","industry":"insurance"},"vendors":[{"name":"Prudential plc","role":"in-house"}],"summary":"Prudential uses AI across its Asian agency force. An AI talkbot validates and enriches leads before they reach agents in Singapore and the Philippines; for the Philippines, launched in the second half of 2024, Prudential reported an initial result of 98 per cent of the talkbot's validated leads being adopted by agents for follow up. In 2025 it launched PruAction in Singapore, a generative AI performance management platform with real time insights for agents, and introduced a Health AI chatbot in Singapore that helps agents access information more quickly. Prudential lists these tools among its agent productivity initiatives but does not attribute its productivity figures to them.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1116578/000110465925025883/tm2429688d4_6k.htm","title":"Prudential plc 2024 full year results (Form 6-K)","publisher":"Prudential plc via SEC EDGAR","date":"2025-03-20"},{"url":"https://www.sec.gov/Archives/edgar/data/1116578/000162828026019027/fullyearprelimreport.htm","title":"Prudential plc 2025 full year results (Form 6-K)","publisher":"Prudential plc via SEC EDGAR","date":"2026-03-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"prudential-plc-ai-agency-tools"},{"title":"Prudential plc: MedScreen+ AI underwriting tool for health cover in Hong Kong","useCases":["life-underwriting-medical-record-summarization"],"organization":{"name":"Prudential plc","anonymized":false,"country":"HK","region":"asia-pacific","industry":"insurance"},"vendors":[{"name":"Prudential plc","role":"in-house"}],"summary":"In its 2025 full year results, Prudential said it continues to enhance its health underwriting with AI powered solutions designed to increase underwriting automation and efficiency, and that it launched MedScreen+ in Hong Kong, an AI underwriting tool intended to provide underwriters with a faster, simpler and more transparent process and to support its financial consultants with instant, indicative underwriting results for customers. In its 2024 results it reported that around 74 per cent of new business policies were processed through auto underwriting capabilities. No outcome figures are given for MedScreen+ itself.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1116578/000162828026019027/fullyearprelimreport.htm","title":"Prudential plc 2025 full year results (Form 6-K)","publisher":"Prudential plc via SEC EDGAR","date":"2026-03-18"},{"url":"https://www.sec.gov/Archives/edgar/data/1116578/000110465925025883/tm2429688d4_6k.htm","title":"Prudential plc 2024 full year results (Form 6-K)","publisher":"Prudential plc via SEC EDGAR","date":"2025-03-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"prudential-plc-medscreen-ai-underwriting"},{"title":"Prysmian: AI cash flow forecasting for North American treasury","useCases":["treasury-cash-flow-forecasting"],"organization":{"name":"Prysmian","anonymized":false,"country":"IT","region":"north-america","industry":"manufacturing"},"vendors":[{"name":"JPMorgan Chase","role":"platform"}],"summary":"Prysmian, the cable manufacturer, used J.P. Morgan Payments Cash Flow Intelligence to automate cash visibility and forecasting for ten operating companies and thirteen bank accounts in North America, run by a treasury team of two. According to its treasurer, the forecast horizon grew from 30 to 91 days, manual reconciliation of more than 3,000 daily transactions was removed and answers to senior management questions came ten times faster.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"cost-savings","value":100000,"unit":"currency","currency":"USD","qualifier":"approximately","period":"per year, estimated labour cost","claimant":"vendor","quote":"Contessa reports that, post-implementation, Prysmian maintains a <1% error rate, saves $100,000 annually and reduced workload by 10 hours weekly","sourceUrl":"https://www.jpmorgan.com/insights/payments/data-intelligence/prysmian-ai-cash-flow-optimization"},{"kpi":"productivity-gain","value":50,"unit":"percent","qualifier":"approximately","period":"one treasury team member's manual forecasting and reconciliation time, about 10 hours a week","claimant":"vendor","quote":"Discover how Prysmian automated global cash flow forecasting, reduced manual work by 50% and saved $100K annually with J.P. Morgan Payments Cash Flow Intelligence.","sourceUrl":"https://www.jpmorgan.com/insights/payments/data-intelligence/prysmian-ai-cash-flow-optimization"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.jpmorgan.com/insights/payments/data-intelligence/prysmian-ai-cash-flow-optimization","title":"Prysmian's AI-Driven Cash Flow Optimization","publisher":"J.P. Morgan","date":"2025-06-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"prysmian-cash-flow-intelligence"},{"title":"Pupuk Indonesia: AI extraction from business documents with single person validation","useCases":["intelligent-document-processing"],"organization":{"name":"Pupuk Indonesia","anonymized":false,"country":"ID","region":"asia-pacific","industry":"manufacturing"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Devoteam","role":"integrator"}],"summary":"Pupuk Indonesia, which Google Cloud describes as Asia's largest fertilizer producer, automated its document processing workflows with Vision AI and Gemini, working with Devoteam. Google Cloud reports that data extraction time fell from 5 to 10 minutes to 40 to 70 seconds, and that one employee now validates the results.","stage":"production","year":2025,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"pupuk-indonesia-document-processing"},{"title":"Purpose Legal: generative AI issues review of 300,000 documents for a court deadline","useCases":["ediscovery-and-disclosure-document-review"],"organization":{"name":"Purpose Legal","anonymized":false,"country":"US","region":"north-america","industry":"professional-services"},"vendors":[{"name":"Relativity","role":"platform"}],"summary":"A law firm that took over a matter as new counsel had one week to review more than 300,000 documents for a court ordered production, with each produced document mapped to the requests for production. Its eDiscovery provider Purpose Legal ran Relativity aiR for Review, a large language model review tool, refining the prompt for ten issues with a firm partner on a sample of fewer than 500 documents before running the full set. The team validated the result with random precision and elusion samples and reports recall above 95%. The results are published by the vendor, and the time and cost savings are estimated against a traditional Active Learning or TAR 2.0 contract review that would have taken multiple weeks and was never run.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":85,"unit":"percent","qualifier":"exact","baseline":"project time compared with a traditional Active Learning or TAR 2.0 contract review that the team says would have taken multiple weeks; that review never ran, so this is an estimate against a counterfactual, not a measured comparison","claimant":"vendor","quote":"Used aiR for Review’s issues analysis to identify documents responsive to 10 key issues – resulting in an 85% reduction in project time.","sourceUrl":"https://www.relativity.com/resources/customers/purpose-legal-relativity-air/"},{"kpi":"hours-saved","value":4000,"unit":"hours","qualifier":"exact","period":"one matter","baseline":"reviewer hours compared with the same counterfactual Active Learning or TAR 2.0 contract review","claimant":"vendor","quote":"Faced with an aggressive deadline and limited budget, aiR for Review allowed the firm to easily complete the review with a skeleton team – reducing review time by 85%, or 4,000 hours.","sourceUrl":"https://www.relativity.com/resources/customers/purpose-legal-relativity-air/"},{"kpi":"cost-savings","value":70000,"unit":"currency","currency":"USD","qualifier":"at-least","period":"one matter","baseline":"compared with the same counterfactual contract review","claimant":"vendor","quote":"This resulted in cost savings of over $70,000 for the law firm, who were thrilled with how aiR for Review let them breathe easy in a tough spot.","sourceUrl":"https://www.relativity.com/resources/customers/purpose-legal-relativity-air/"},{"kpi":"interactions-handled","value":300000,"unit":"count","qualifier":"at-least","period":"seven days, one matter","claimant":"vendor","quote":"Over 300,000 documents were reviewed in just seven days, using only one project manager and a law firm partner to provide subject matter expertise.","sourceUrl":"https://www.relativity.com/resources/customers/purpose-legal-relativity-air/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.relativity.com/resources/customers/purpose-legal-relativity-air/","title":"Purpose Legal Slashes Thousands of Hours to Beat Impossible Deadline with Relativity aiR for Review","publisher":"Relativity","archivedUrl":"https://web.archive.org/web/20251006015118/https://www.relativity.com/resources/customers/purpose-legal-relativity-air/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"purpose-legal-generative-ai-issues-review"},{"title":"PZU: AI assessment of car damage from policyholder photos","useCases":["photo-based-damage-assessment"],"organization":{"name":"PZU","anonymized":false,"country":"PL","region":"europe","industry":"insurance"},"vendors":[{"name":"Tractable","role":"platform"}],"summary":"In March 2022 Tractable and PZU, Poland's largest insurer, announced that policyholders can submit smartphone photos of car damage when they report an accident. Tractable's AI assesses the damage and calculates the repair cost, the claim handler can share the result within minutes, and the policyholder can accept a cash settlement instead of waiting days. PZU had worked with Tractable since 2017 and already used its AI to check how body shops carry out repairs, processing several hundred thousand claims with its AI based tools. No outcome figures for the photo journey were published.","stage":"production","year":2022,"channels":[],"languages":["pl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://tractable.ai/pzu-is-first-polish-insurer-to-enable-its-customers-to-use-ai-to-assess-car-damage-and-settle-claims-in-minutes/","title":"PZU is first Polish insurer to enable its customers to use AI to assess car damage and settle claims in minutes","publisher":"Tractable","date":"2022-03-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"pzu-ai-car-damage-assessment"},{"title":"Quilter: Microsoft 365 Copilot for meeting notes and investment writing","useCases":["client-meeting-notes-and-crm-update","portfolio-reporting-and-commentary"],"organization":{"name":"Quilter","anonymized":false,"country":"GB","region":"europe","industry":"wealth-and-asset-management"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Quilter, a UK wealth manager, rolled out Microsoft 365 Copilot and names meetings and transcriptions as its biggest use case. Microsoft reports that Quilter estimates Copilot will save more than 13,000 hours per month of post call admin time; an investment manager at Quilter Cheviot builds that estimate from an assumed 45 minutes saved per client meeting across 174 investment managers doing about 100 meetings each. Both figures are projections, not measured savings, so neither is recorded as a metric. Quilter also tested turning a portfolio manager interview transcript into an investment commentary: about 15 minutes of prompting and half an hour of editing instead of a few days, which it describes as a one off test.","stage":"production","year":2025,"channels":["microsoft-teams","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23237-quilter-microsoft-365-copilot","title":"Quilter achieves fastest-ever tech ROI with Microsoft 365 Copilot","publisher":"Microsoft Customer Stories","archivedUrl":"https://web.archive.org/web/20250517152254/https://www.microsoft.com/en/customers/story/23237-quilter-microsoft-365-copilot"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"quilter-copilot-meeting-notes"},{"title":"Radisson Hotel Group: personalized advertising at scale with generative AI","useCases":["personalized-marketing-at-scale"],"organization":{"name":"Radisson Hotel Group","anonymized":false,"country":"BE","region":"europe","industry":"travel-and-hospitality"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Accenture","role":"integrator"}],"summary":"Radisson Hotel Group worked with Accenture to personalize its advertising at scale with Vertex AI and Gemini models, trained on extensive datasets stored in BigQuery. Google Cloud reports that ad team productivity rose by around 50% and that revenue from the AI powered campaigns rose by more than 20%. The comparison group for the revenue figure is not stated.","stage":"production","year":2024,"channels":[],"languages":[],"metrics":[{"kpi":"revenue-uplift","value":20,"unit":"percent","qualifier":"at-least","period":"revenue from AI powered campaigns","claimant":"vendor","quote":"By training them on extensive datasets stored in BigQuery, its ad teams saw productivity rise around 50% while revenue increased from AI-powered campaigns by more than 20%.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"},{"kpi":"productivity-gain","value":50,"unit":"percent","qualifier":"approximately","period":"advertising team productivity","claimant":"vendor","quote":"By training them on extensive datasets stored in BigQuery, its ad teams saw productivity rise around 50% while revenue increased from AI-powered campaigns by more than 20%.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"},{"url":"https://en.wikipedia.org/wiki/Radisson_Hotel_Group","title":"Radisson Hotel Group","publisher":"Wikipedia"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"radisson-hotel-group-personalized-advertising"},{"title":"Ratepay: payment screening and transaction monitoring with Hawk, AI features planned","useCases":["sanctions-screening-adjudication","aml-alert-triage"],"organization":{"name":"Ratepay","anonymized":false,"country":"DE","region":"europe","industry":"payments"},"vendors":[{"name":"Hawk","role":"platform"}],"summary":"Ratepay, a German provider of white label buy now pay later solutions and part of the Nexi Group, replaced its previous solution with Hawk's Payment Screening, which screens transactions in real time against global sanctions lists, and Hawk's AML Transaction Monitoring, with centralised case management for investigators and auditors. The vendor's story says Ratepay is now planning to add Hawk's AI technology for anomaly detection and false positive reduction, so the AI adjudication step this record is filed under is announced rather than live.","stage":"announced","year":2025,"channels":["internal-tools","api"],"languages":["en","de"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://hawk.ai/news-press/how-ratepay-scaling-bnpl-solutions-aml-screening-technology-hawk","title":"How Ratepay Is Scaling BNPL Solutions With AML & Screening Technology From Hawk","publisher":"Hawk","date":"2025-08-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ratepay-hawk-aml-screening"},{"title":"RBC: NOMI personal insights, cash flow forecasts and saving","useCases":["financial-wellbeing-coach"],"organization":{"name":"Royal Bank of Canada","anonymized":false,"country":"CA","region":"north-america","industry":"banking"},"vendors":[{"name":"Borealis AI","role":"in-house"}],"summary":"RBC's NOMI is a set of AI features in the RBC Mobile app and RBC Online Banking that give clients personalized insights about their money. NOMI Forecast, built with the bank's research centre Borealis AI, uses deep learning to show a seven day view of upcoming preauthorized payments and cash flow; NOMI Find and Save helps clients put money aside; NOMI Budgets tracks spending. RBC says clients using Find and Save have put aside more than CAD 3.6 billion.","stage":"scaled","year":2023,"channels":["mobile-app"],"languages":["en"],"metrics":[{"kpi":"users-served","value":900000,"unit":"count","qualifier":"at-least","period":"NOMI Forecast, September 2021 to April 2023","claimant":"organization","quote":"Since its launch in September 2021, more than 900,000 clients have used the feature.","sourceUrl":"https://www.newswire.ca/news-releases/rbc-wins-best-use-of-ai-for-customer-experience-for-nomi-forecast-882322906.html"},{"kpi":"interactions-handled","value":10000000,"unit":"count","qualifier":"at-least","period":"NOMI Forecast, 2021 to April 2023","claimant":"organization","quote":"The addition of NOMI Forecast has led to more than 10 million client interactions since 2021.","sourceUrl":"https://www.newswire.ca/news-releases/rbc-wins-best-use-of-ai-for-customer-experience-for-nomi-forecast-882322906.html"},{"kpi":"customer-savings","value":3600000000,"unit":"currency","currency":"CAD","qualifier":"at-least","period":"NOMI Find and Save, cumulative as of April 2023","claimant":"organization","quote":"Clients using NOMI Find & Save have put aside more than $3.6 billion in savings.","sourceUrl":"https://www.newswire.ca/news-releases/rbc-wins-best-use-of-ai-for-customer-experience-for-nomi-forecast-882322906.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.newswire.ca/news-releases/rbc-wins-best-use-of-ai-for-customer-experience-for-nomi-forecast-882322906.html","title":"RBC Wins Best Use of AI for Customer Experience for NOMI Forecast","publisher":"RBC Royal Bank via Cision","date":"2023-04-28"},{"url":"https://www.rbcroyalbank.com/mobile/feature/nomi/index.html","title":"NOMI","publisher":"RBC Royal Bank"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"rbc-nomi-personal-insights"},{"title":"Recursion Pharmaceuticals: Recursion OS drug design platform","useCases":["ai-drug-discovery-platform"],"organization":{"name":"Recursion Pharmaceuticals","anonymized":false,"country":"US","region":"north-america","industry":"pharma-and-life-sciences"},"vendors":[],"summary":"Recursion Pharmaceuticals, which completed its business combination with Exscientia in November 2024, runs an AI native operating system that combines phenomic screening with automated, precision small molecule chemistry to take programs from an initial hit to a development candidate. As of its February 2026 results, the company reports the platform has delivered more than 10 development candidates to date, including REC-617, a CDK7 inhibitor identified as lead candidate in under 11 months with 136 novel compounds synthesized, and REC-7735, a PI3Kα H1047R inhibitor precision designed with 242 compounds synthesized from first novel hit to REC-7735 in 10 months, now in IND enabling studies. Recursion's partnership with Sanofi has the potential for up to 15 AI designed small molecule programs, of which 5 or more span immunology and oncology, and had reached five progress based milestones as of the same results.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":60,"unit":"percent","qualifier":"approximately","period":"reported February 2026; average for advanced candidates delivered by the platform, per program","baseline":"industry average of over 2,500 compounds and 42 months per program","claimant":"organization","quote":"Advanced candidates have been delivered by synthesizing ~330 compounds per program in ~17 months, compared to industry averages of over 2,500 compounds and 42 months, respectively.","sourceUrl":"https://www.globenewswire.com/news-release/2026/02/25/3244408/0/en/recursion-reports-fourth-quarter-and-full-year-2025-financial-results-and-provides-business-update.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.globenewswire.com/news-release/2026/02/25/3244408/0/en/recursion-reports-fourth-quarter-and-full-year-2025-financial-results-and-provides-business-update.html","title":"Recursion Reports Fourth Quarter and Full Year 2025 Financial Results and Provides Business Update","publisher":"Recursion Pharmaceuticals, Inc.","date":"2026-02-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"recursion-pharmaceuticals-os-platform"},{"title":"Région Sud: automated document checks for jobseeker training grants","useCases":["benefits-eligibility-and-application-assistant"],"organization":{"name":"Région Provence-Alpes-Côte d'Azur (Région Sud)","anonymized":false,"country":"FR","region":"europe","industry":"government"},"vendors":[{"name":"Microsoft (Azure OpenAI Service)","role":"platform"},{"name":"Mistral AI (models on Azure)","role":"model-provider"},{"name":"Exakis Nelite","role":"integrator"}],"summary":"As part of its regional AI plan, Région Sud in southern France automated the verification of the supporting documents that jobseekers submit for skills training grants, and deployed a chatbot on Azure OpenAI Service and Mistral models that helps agents at its Allo Région call centre answer citizens; the administration receives 75,000 requests a year. The chatbot queries the region's own databases in a secure environment.","stage":"production","year":2025,"channels":["internal-tools","agent-desktop"],"languages":["fr"],"metrics":[{"kpi":"interactions-handled","value":30000,"unit":"count","qualifier":"at-least","period":"grant documents per year","claimant":"vendor","quote":"At the same time, they have automated the verification of documents required for jobseekers skill training grants: no fewer than 30,000 documents per year are now processed by the machine.","sourceUrl":"https://www.microsoft.com/en/customers/story/23393-region-sud-microsoft-365-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23393-region-sud-microsoft-365-copilot","title":"Région Sud and Microsoft co-build a smart 4.0 territory with Azure OpenAI Service and Microsoft 365 Copilot","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"region-sud-training-grant-document-verification"},{"title":"Reserve Bank of India: MuleHunter.AI mule account detection model for banks","useCases":["mule-network-detection"],"organization":{"name":"Reserve Bank Innovation Hub (Reserve Bank of India)","anonymized":false,"country":"IN","region":"asia-pacific","industry":"government"},"vendors":[{"name":"Reserve Bank Innovation Hub","role":"in-house"}],"summary":"The Reserve Bank Innovation Hub, a subsidiary of the Reserve Bank of India, built MuleHunter.AI, a machine learning model that helps banks detect mule accounts used to move fraud proceeds. The RBI announced the pilot with two large public sector banks in December 2024, and the Governor said in October 2025 that it had been scaled from about 5 banks to 21 banks, using system wide learning. The RBI has declined to disclose how many mule accounts it has identified.","stage":"scaled","year":2025,"channels":["api"],"languages":[],"metrics":[{"kpi":"users-served","value":21,"unit":"count","qualifier":"exact","period":"banks using the model, October 2025","claimant":"organization","quote":"MuleHunter.ai, developed by the Reserve Bank Innovation Hub has been scaled up from about 5 banks at the beginning of this year to 21 banks.","sourceUrl":"https://www.rbi.org.in/Scripts/BS_SpeechesView.aspx?Id=1525"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.rbi.org.in/Scripts/BS_SpeechesView.aspx?Id=1525","title":"Driving Inclusive and Sustainable Growth Through Digital Public Infrastructure and FinTech","publisher":"Reserve Bank of India","date":"2025-10-10"},{"url":"https://www.rbi.org.in/scripts/BS_PressReleaseDisplay.aspx?prid=59245","title":"Statement on Developmental and Regulatory Policies","publisher":"Reserve Bank of India","date":"2024-12-06"},{"url":"https://www.medianama.com/2025/12/223-rti-23-banks-mulehunter-mule-accounts/","title":"23 Banks Use Mulehunter.AI, RBI Won't Disclose Mule Data","publisher":"MediaNama","date":"2025-12-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"reserve-bank-innovation-hub-mulehunter-ai"},{"title":"ResMed: automated cash application across all business units","useCases":["cash-application-and-remittance-matching"],"organization":{"name":"ResMed","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"HighRadius","role":"platform"}],"summary":"ResMed, a global connected care company, deployed HighRadius's Customer-to-Cash Receivables Management suite to standardize accounts receivable operations across business units, posting cash automatically and auto coding deductions. A ResMed manager reported that the solution saved 50% of an analyst's time specifically on data aggregation, one sub task of cash application, not an overall productivity gain. HighRadius also reported a reduction in days sales outstanding of about 33 days within 10 months; the case study covers ResMed's broader Customer-to-Cash Receivables Management suite, so it is unclear how much of that reduction is attributable to cash application alone.","stage":"production","year":2021,"channels":[],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":96,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"96% Cash Posting Hit-Rate","sourceUrl":"https://www.highradius.com/resources/case-studies/resmed/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.highradius.com/resources/case-studies/resmed/","title":"ResMed | Customer-to-Cash Receivables Management","publisher":"HighRadius","archivedUrl":"https://web.archive.org/web/20230925062621/https://www.highradius.com/resources/case-studies/resmed/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"resmed-cash-application-automation"},{"title":"Revolut: AI card scam detection with an in app intervention flow","useCases":["scam-payment-interception","fraud-alert-confirmation","real-time-fraud-scoring"],"organization":{"name":"Revolut","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[{"name":"Revolut","role":"in-house"}],"summary":"In February 2024 Revolut launched a machine learning feature, built by its financial crime team, that estimates whether a card payment is part of a scam. When the risk is high it declines the payment, blocks similar payments and sends the customer through an in app intervention flow that asks about the payment, checks whether someone is guiding them, shows scam stories and offers a chat with a fraud specialist.","stage":"production","year":2024,"channels":["mobile-app"],"languages":["en"],"metrics":[{"kpi":"fraud-loss-reduction","value":30,"unit":"percent","qualifier":"exact","period":"since launch, fraud losses from card scams where money was sent for investment opportunities","claimant":"organization","quote":"Since the launch of the card scam detection feature, Revolut has observed a 30% reduction in the fraud losses resulting from card scams where money has been sent for investment opportunities.","sourceUrl":"https://www.revolut.com/en-US/news/revolut_launches_ai_feature_to_protect_customers_from_card_scams_and_break_the_scammers_spell/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.revolut.com/en-US/news/revolut_launches_ai_feature_to_protect_customers_from_card_scams_and_break_the_scammers_spell/","title":"Revolut launches AI feature to protect customers from card scams and break the scammers \"spell\"","publisher":"Revolut","date":"2024-02-15","archivedUrl":"https://web.archive.org/web/20250917122343/https://www.revolut.com/en-US/news/revolut_launches_ai_feature_to_protect_customers_from_card_scams_and_break_the_scammers_spell/"},{"url":"https://www.openbankingexpo.com/news/revolut-introduces-new-ai-powered-card-scam-detection-feature/","title":"Revolut introduces AI-powered card scam detection feature","publisher":"Open Banking Expo","date":"2024-02-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"revolut-card-scam-detection"},{"title":"Rho: AI drafted performance reviews with Windmill","useCases":["performance-review-drafting-agent"],"organization":{"name":"Rho","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Windmill","role":"platform"}],"summary":"Rho, a fintech platform for startups, runs its full performance review cycle on Windmill's AI review agent. The People team replaced a process where managers spent significant time gathering work artifacts and feedback across separate tools with a cycle that drafts each review from collected feedback and work history: self reviews complete in about 2.5 days, 360 reviews in 5 days, and the full self to manager cycle in 8 days, against an industry average that Windmill describes as several weeks.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":83,"unit":"percent","qualifier":"exact","period":"total hours spent on reviews","claimant":"vendor","quote":"Results: 83% reduction in total hours spent on reviews and 93% of employees preferred Windmill to the prior process.","sourceUrl":"https://gowindmill.com/resources/lists/companies-using-ai-performance-management/"}],"outcomeDisclosed":true,"sources":[{"url":"https://gowindmill.com/resources/lists/companies-using-ai-performance-management/","title":"How 6 Companies Use AI for Performance Management","publisher":"Windmill","date":"2026-05-07"},{"url":"https://gowindmill.com/customers/rho-perf","title":"Rho: Streamlining Performance Reviews at Scale with Windmill","publisher":"Windmill","date":"2025-11-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"rho-ai-performance-review-drafting"},{"title":"RIMC Hotels & Resorts Group: Duetto revenue management","useCases":["hotel-revenue-management-copilot"],"organization":{"name":"RIMC Hotels & Resorts Group","anonymized":false,"country":"DE","region":"europe","industry":"travel-and-hospitality"},"vendors":[{"name":"Duetto","role":"platform"}],"summary":"RIMC Hotels & Resorts Group, a hotel association headquartered in Hamburg with business, city and holiday hotels across several countries, replaced manual pricing, made through the property management system and time consuming for pricing and capacity control, with Duetto's cloud revenue management system in 2022. Duetto forecasts demand and recommends real time, demand driven room prices across the portfolio, in place of the group's previous manual adjustments. Head of Revenue Henning Möhn describes an increase in RevPAR at every Duetto hotel in the portfolio and a 28.44 percent RevPAR increase at the group's Polish property, plus daily time saved on pricing and monitoring per hotel.","stage":"production","year":2022,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"revenue-uplift","value":28.44,"unit":"percent","qualifier":"exact","claimant":"organization","quote":"We've seen a 28.44% increase in RevPAR for our Polish property and an overall RevPAR increase across our portfolio.","sourceUrl":"https://www.duettocloud.com/en-us/success-stories/duetto-drives-28-increase-in-revpar-for-rimc-hotels-resorts-group"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.duettocloud.com/en-us/success-stories/duetto-drives-28-increase-in-revpar-for-rimc-hotels-resorts-group","title":"28% increase in RevPAR for RIMC Hotels & Resorts Group","publisher":"Duetto"},{"url":"https://www.duettocloud.com/en-us/platform/gamechanger","title":"GameChanger: Predictive Analytics Software","publisher":"Duetto"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"rimc-hotels-resorts-duetto-revenue-management"},{"title":"Rio de Janeiro: 1746 citizen service chatbot for urban maintenance requests","useCases":["non-emergency-service-request-routing"],"organization":{"name":"Rio de Janeiro City Data Office (Escritório de Dados)","anonymized":false,"country":"BR","region":"latin-america","industry":"government"},"vendors":[{"name":"Google Cloud (Dialogflow)","role":"platform"}],"summary":"Rio de Janeiro's City Data Office (Escritório de Dados) uses Dialogflow to run the chatbot of 1746, the city's citizen service, which handles urban maintenance requests and municipal enquiries, the kind of work a 311 line does elsewhere. Google reports that it cut the citizen response time from 30 minutes to 5 minutes. The only public source found is a one paragraph entry in Google's customer list.","stage":"production","year":2025,"channels":[],"languages":["pt"],"metrics":[{"kpi":"interactions-handled","value":30000,"unit":"count","qualifier":"at-least","period":"conversations per month","claimant":"vendor","quote":"The conversational AI reduces citizen response time from 30 minutes to 5 minutes across more than 30,000 monthly conversations.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"rio-de-janeiro-1746-citizen-service-chatbot"},{"title":"Riverside County: C3 AI residential property appraisal","useCases":["property-valuation-support"],"organization":{"name":"Riverside County Assessor-County Clerk-Recorder","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"C3 AI","role":"platform"}],"summary":"Riverside County's Assessor-County Clerk-Recorder deployed the C3 AI Residential Property Appraisal application for around 460,000 single family homes and condominiums, replacing manual linear regression models used to automatically enroll eligible change of ownership transfers at their sale price. C3 AI describes this as Riverside's initial production deployment, delivered in under six months, meant to demonstrate the application's ability to improve staff efficiency and reduce the complexity of its modeling approach.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://c3.ai/customers/riverside-county-drives-40-increase-in-model-accuracy-for-property-appraisal/","title":"Riverside County Drives 40% Increase in Model Accuracy for Property Appraisal","publisher":"C3 AI","archivedUrl":"https://web.archive.org/web/20241203111911/https://c3.ai/customers/riverside-county-drives-40-increase-in-model-accuracy-for-property-appraisal/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"riverside-county-c3-ai-property-appraisal"},{"title":"Rivian: grounded, shareable knowledge base for frequently asked questions","useCases":["support-knowledge-article-generation"],"organization":{"name":"Rivian","anonymized":false,"country":"US","region":"north-america","industry":"automotive"},"vendors":[{"name":"Google","role":"platform"}],"summary":"Rivian uses NotebookLM to centralize answers to frequently asked questions from verified sources and share them as a knowledge base with interactive chat. Google Cloud reports that it reduced repetitive inquiries and saved employees time, without a figure. It is a simple example of turning scattered answers into shared, grounded knowledge.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"rivian-notebooklm-shared-knowledge-base"},{"title":"RMB: AI powered trade document checking on the Traydstream platform","useCases":["trade-document-examination"],"organization":{"name":"Rand Merchant Bank","anonymized":false,"country":"ZA","region":"africa","industry":"banking"},"vendors":[{"name":"Traydstream","role":"platform"}],"summary":"RMB (Rand Merchant Bank), which describes itself as a leading African corporate and investment bank, announced that it had gone live on Traydstream's AI enabled trade finance platform. The platform can digitise documents related to letters of credit, collections and open account transactions for automated document checking, clause matching and rules validation with machine learning and OCR; the releases do not say which of these RMB uses it for. RMB's head of trade described the checking of numerous unstructured trade documents as manual and extremely time consuming. No outcome figures are published.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.rmb.co.za/news/rmb-streamline-and-digitise-its-trade-finance-process","title":"RMB streamline and digitise its trade finance process","publisher":"RMB","date":"2022-09-20","archivedUrl":"https://web.archive.org/web/20220920143316/https://www.rmb.co.za/news/rmb-streamline-and-digitise-its-trade-finance-process"},{"url":"https://traydstream.com/news/ai-powered-platform-for-trade","title":"RMB goes live with Traydstream's AI powered platform for trade","publisher":"Traydstream","date":"2022-09-21"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"rmb-automated-trade-document-checking"},{"title":"Rocket Mortgage: AI Digital Assistant from first question to preapproval","useCases":["inbound-lead-qualification-agent","conversational-loan-application-intake"],"organization":{"name":"Rocket Mortgage","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Sierra","role":"platform"}],"summary":"Rocket Mortgage runs an AI Digital Assistant across chat and voice that takes prospective borrowers from first questions to preapproval: it answers questions, collects information, pulls credit, presents personalised rates and loan options and hands the client to a human banker. According to the vendor, the programme started as a proof of concept and has grown to more than 400,000 successful chat conversations and over one million outbound dials a month. Clients who start with the assistant close at three times the rate of those who do not. Sierra also reports that clients who use both the AI chat and a banker convert four times better, without stating the comparison group.","stage":"scaled","year":2025,"channels":["web-chat","voice"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":400000,"unit":"count","qualifier":"at-least","period":"successful chat conversations per month","claimant":"vendor","quote":"What started as a proof of concept in May has grown to more than 400,000 successful chat conversations and over one million outbound dials each month, and both are rising fast.","sourceUrl":"https://sierra.ai/customers/rocket-mortgage"}],"outcomeDisclosed":true,"sources":[{"url":"https://sierra.ai/customers/rocket-mortgage","title":"How Rocket Mortgage is reimagining the journey home with AI","publisher":"Sierra","date":"2025-10-27"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"rocket-mortgage-digital-assistant"},{"title":"Rogo: AI platform that builds company profiles and drafts investment memos for bankers","useCases":["deal-sourcing-and-due-diligence-assistant"],"organization":{"name":"Rogo","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Google Cloud","role":"model-provider"}],"summary":"Rogo, a New York AI company serving investment banks and private equity firms, combines a firm's own memos, research and files with external sources such as SEC filings, PitchBook, S&P Global, FactSet and Preqin, and automates workflows such as company profiles, competitive benchmarking, slide decks and investment memo drafts. Google Cloud reports that moving to Gemini 2.5 Flash cut hallucination rates in Rogo's evaluation, and counts thousands of bankers and analysts on the platform.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":6000,"unit":"count","qualifier":"at-least","period":"investment bankers and analysts on the platform","claimant":"vendor","quote":"Builds trust in the Rogo AI platform among 6,000+ investment bankers and analysts","sourceUrl":"https://cloud.google.com/customers/rogo"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/rogo","title":"Rogo: Enabling faster time-to-insights for financial services firms with agentic AI","publisher":"Google Cloud"},{"url":"https://rogo.ai/","title":"Rogo, AI for the most ambitious firms in finance","publisher":"Rogo"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"rogo-investment-banking-research-agents"},{"title":"RTVE: automatic metadata for the Documentary Archive","useCases":["media-archive-metadata-tagging"],"organization":{"name":"Radiotelevisión Española (RTVE)","anonymized":false,"country":"ES","region":"europe","industry":"media-and-entertainment"},"vendors":[{"name":"NexTReT","role":"integrator"}],"summary":"NexTReT describes an AI based automatic metadata service for RTVE's Documentary Archive, integrated into its ARCA document management system with an interface for documentalists to validate the results, to make decades of audiovisual heritage material searchable and reusable. The case study is undated and does not say when the service ran or whether it later ended. Separately, RTVE ran a 2025 tender for automatic archive metadata, reported by Panorama Audiovisual; see the verification note for why that tender is not folded into this record.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.panoramaaudiovisual.com/en/2025/09/04/crosspoint-and-amplify-will-manage-the-automatic-metadata-of-the-rtve-archive-using-ai/","title":"Crosspoint and Amplify will manage the automatic metadata of the RTVE Archive using AI","publisher":"Panorama Audiovisual","date":"2025-09-04"},{"url":"https://nextret.net/en/casos-de-exito/automatizacion-del-metadatado-audiovisual-en-rtve/","title":"Success Story: Audiovisual Metadata Automation at RTVE","publisher":"NexTReT"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"rtve-archive-metadata-automation"},{"title":"Safaricom: AI energy efficiency software across about 30,000 radio cells","useCases":["ran-energy-optimization"],"organization":{"name":"Safaricom","anonymized":false,"country":"KE","region":"africa","industry":"telecommunications"},"vendors":[{"name":"Nokia","role":"platform"}],"summary":"After a pilot, Safaricom Kenya rolled out Nokia's AVA Energy Efficiency software across approximately 30,000 5G, 4G and 3G cells. The software uses AI and machine learning to switch off idle and unused equipment automatically during low usage periods, together with Nokia's radio energy efficiency features, while maintaining network quality. The release gives planned energy cost savings, not measured results.","stage":"scaled","year":2023,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.nokia.com/newsroom/nokia-deploys-ava-energy-efficiency-for-safaricom-kenya-to-drive-network-energy-savings/","title":"Nokia deploys AVA Energy Efficiency for Safaricom Kenya to drive network energy savings","publisher":"Nokia","date":"2023-11-27","archivedUrl":"https://web.archive.org/web/20250809193611/https://www.nokia.com/newsroom/nokia-deploys-ava-energy-efficiency-for-safaricom-kenya-to-drive-network-energy-savings/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"safaricom-nokia-ava-energy-efficiency"},{"title":"Safe Rate: AI mortgage assistant for rate comparison and refinance quotes","useCases":["home-loan-assistant-and-prequalification"],"organization":{"name":"Safe Rate","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"model-provider"}],"summary":"Google Cloud describes Safe Rate as a digital mortgage lender that uses Gemini models to create an AI mortgage agent with chat features called \"Beat this Rate\" and \"Refinance Me\", which let borrowers compare rates and get a personalised quote in under 30 seconds. Safe Rate's own website now presents it as a US mortgage shopping service that ranks lenders, and says its AI assistant answers plain English questions about rates, lenders and local costs of ownership. Neither source reports outcome figures.","stage":"production","year":2025,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"},{"url":"https://www.saferate.com/","title":"Shop for a Mortgage on Your Terms","publisher":"Safe Rate"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"safe-rate-ai-mortgage-agent"},{"title":"SameDay Auto Finance: voice AI for early stage collections","useCases":["collections-and-hardship-agent"],"organization":{"name":"SameDay Auto Finance","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Skit.ai","role":"platform"}],"summary":"SameDay Auto Finance, a Dallas auto lender whose portfolio sits mostly in early delinquency, moved its early stage outreach to AI voice agents calling around the clock, with SMS for customers who do not answer calls, and redeployed its human agents to inbound returns, skip tracing and complex accounts. The rollout ran in four phases over a year. The vendor reports 43% higher collections and 75% lower collection call costs in the early delinquency buckets.","stage":"production","year":2026,"channels":["voice","sms"],"languages":["en"],"metrics":[{"kpi":"recovery-rate-uplift","value":43,"unit":"percent","qualifier":"exact","period":"early stage delinquency","claimant":"vendor","quote":"SameDay Auto Finance Achieves 43% Higher Collections and 75% Lower Call Costs in Early-DPD using AI for Collections.","sourceUrl":"https://skit.ai/resource/case-studies/from-legacy-tech-to-2x-ptp-inone-year/"},{"kpi":"cost-reduction","value":75,"unit":"percent","qualifier":"exact","period":"collection call cost, early stage delinquency","claimant":"vendor","quote":"SameDay Auto Finance Achieves 43% Higher Collections and 75% Lower Call Costs in Early-DPD using AI for Collections.","sourceUrl":"https://skit.ai/resource/case-studies/from-legacy-tech-to-2x-ptp-inone-year/"}],"outcomeDisclosed":true,"sources":[{"url":"https://skit.ai/resource/case-studies/from-legacy-tech-to-2x-ptp-inone-year/","title":"SameDay Auto Finance Achieves 43% Higher Collections and 75% Lower Call Costs","publisher":"Skit.ai","date":"2026-04-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"sameday-auto-finance-voice-collections"},{"title":"Sandvik Coromant: Copilot for Sales for meeting summaries, email and CRM capture","useCases":["sales-call-coaching-and-crm-update"],"organization":{"name":"Sandvik Coromant","anonymized":false,"country":"SE","region":"europe","industry":"manufacturing"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Sandvik Coromant, a supplier of cutting tools with about 8,000 staff, was an early adopter of Microsoft Copilot for Sales on top of Dynamics 365. Sellers use it to summarize email threads and Teams meetings, capture contact details into the CRM in one click, add email summaries as CRM notes, get a summary of next steps after each meeting and a suggested recap for the customer, and draft replies. The company reports that account managers save three minutes per transaction several times a day, and Microsoft reports 20 minutes a day saved on email summaries.","stage":"production","year":2024,"channels":["microsoft-teams","email","internal-tools"],"languages":["en"],"metrics":[{"kpi":"time-saved-per-task","value":3,"unit":"minutes","qualifier":"exact","period":"per transaction, several times a day per account manager","claimant":"organization","quote":"With everything on the side panel in Outlook, account managers save three minutes per transaction multiple times a day.","sourceUrl":"https://customers.microsoft.com/en-us/story/1785448033474736158-sandvik-coromant-microsoft-copilot-for-sales-discrete-manufacturing-en-sweden"}],"outcomeDisclosed":true,"sources":[{"url":"https://customers.microsoft.com/en-us/story/1785448033474736158-sandvik-coromant-microsoft-copilot-for-sales-discrete-manufacturing-en-sweden","title":"Sandvik Coromant hones sales experience with Microsoft Copilot for Sales","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"sandvik-coromant-copilot-for-sales"},{"title":"Banco Santander: first live payment executed by an AI agent in Europe, with Mastercard Agent Pay","useCases":["agentic-payment-initiation"],"organization":{"name":"Banco Santander","anonymized":false,"country":"ES","region":"europe","industry":"banking"},"vendors":[{"name":"Mastercard","role":"platform"},{"name":"PayOS","role":"integrator"}],"summary":"Banco Santander and Mastercard announced a live payment from start to finish that was initiated and executed by an AI agent, run in a controlled environment through Santander's live payments infrastructure using Mastercard Agent Pay, with PayOS orchestrating the transaction. The model lets agents pay on behalf of customers within predefined limits and permissions, and the aim was to validate the bank's operational and control framework under real conditions. The release calls it a pilot that is not a commercial rollout; Santander will now move into extended testing and scaling. No outcome figures are disclosed.","stage":"pilot","year":2026,"channels":[],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.mastercard.com/news/europe/en/newsroom/press-releases/en/2026/santander-and-mastercard-complete-europe-s-first-live-end-to-end-payment-executed-by-an-ai-agent/","title":"Santander and Mastercard complete Europe's first live end-to-end payment executed by an AI agent","publisher":"Mastercard Newsroom"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"santander-mastercard-agent-pay-live-payment"},{"title":"Santander UK: automated adverse media screening in digital onboarding with ComplyAdvantage","useCases":["pep-and-adverse-media-screening"],"organization":{"name":"Santander UK","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[{"name":"ComplyAdvantage","role":"platform"}],"summary":"Santander UK used ComplyAdvantage's adverse media screening, delivered through an API, as part of a digital onboarding proposition for corporate and SME customers, and screens every entity linked to an onboarding case. The vendor reports that the onboarding cycle fell from 12 days to 2 days on average; the figure covers the whole onboarding process, not the screening step alone.","stage":"production","year":2019,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://complyadvantage.com/customer-stories/santander-case-study/","title":"Santander: KYC/AML Case Study","publisher":"ComplyAdvantage","date":"2021-02-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"santander-uk-complyadvantage-onboarding-screening"},{"title":"Save the Children: AI due diligence reports on corporate donors with Xapien","useCases":["pep-and-adverse-media-screening"],"organization":{"name":"Save the Children","anonymized":false,"country":"GB","region":"europe","industry":"cross-industry"},"vendors":[{"name":"Xapien","role":"platform"}],"summary":"Save the Children uses Xapien, an AI supported due diligence platform that produces a report on a prospective donor and surfaces areas of concern early in the report, to vet corporate donors for alignment with its values and for reputational risk. The vendor reports that review times fell by more than 60%, with reports completed in as little as 15 minutes rather than over an afternoon, so the team can vet more donors. The platform is one part of a wider, human led review. It shows the same adverse media job outside banking.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":60,"unit":"percent","qualifier":"at-least","period":"analyst review time per corporate donor","claimant":"vendor","quote":"Save the Children uses Xapien to accelerate corporate donor due diligence, cutting review times by over 60%.","sourceUrl":"https://xapien.com/case-studies/save-the-children-supporting-childrens-wellbeing-globally-one-xapien-report-at-a-time/"}],"outcomeDisclosed":true,"sources":[{"url":"https://xapien.com/case-studies/save-the-children-supporting-childrens-wellbeing-globally-one-xapien-report-at-a-time/","title":"Save the Children cuts donor due diligence time by 60%","publisher":"Xapien","date":"2026-03-11"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"save-the-children-xapien-donor-due-diligence"},{"title":"SBF Group: AI classification of customer feedback and NPS survey answers","useCases":["customer-feedback-analysis"],"organization":{"name":"SBF Group","anonymized":false,"country":"BR","region":"latin-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"SBF Group (Grupo SBF), the Brazilian sporting goods retailer behind Centauro and Fisia, the official Nike distributor in Brazil, uses Google Cloud AI to analyse customer feedback and customer satisfaction (NPS) forms. Google Cloud reports that the solution eliminated manual data analysis work, cut annual costs by about 95%, raised feedback classification accuracy from 16% to 84% and made it possible to process daily feedback that previously went unanalysed.","stage":"production","year":2026,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"cost-reduction","value":95,"unit":"percent","qualifier":"approximately","period":"per year (the cost base is not specified)","claimant":"vendor","quote":"The solution reduced annual costs by approximately 95% and eliminated manual data analysis work, in addition to increasing the accuracy rate in feedback classification from 16% to 84%, allowing daily processing of information that was previously not analyzed.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"},{"kpi":"accuracy","value":84,"unit":"percent","qualifier":"exact","baseline":"16% classification accuracy before","claimant":"vendor","quote":"The solution reduced annual costs by approximately 95% and eliminated manual data analysis work, in addition to increasing the accuracy rate in feedback classification from 16% to 84%, allowing daily processing of information that was previously not analyzed.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"},{"url":"https://www.gruposbf.com.br/sobre-nos","title":"Sobre nós","publisher":"Grupo SBF"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"sbf-group-nps-feedback-analysis"},{"title":"Schneider Electric: Open Talent Market matches employees to gigs, roles and mentors","useCases":["internal-talent-marketplace-matching"],"organization":{"name":"Schneider Electric","anonymized":false,"country":"FR","region":"global","industry":"manufacturing"},"vendors":[{"name":"Gloat","role":"platform"}],"summary":"Schneider Electric's internal surveys, as reported by Gloat, showed that nearly half of departing employees cited a lack of internal growth opportunities as their main reason for leaving. The company launched Open Talent Market, an AI talent marketplace, with a pilot in HR and a global launch in April 2020. Employees build a profile with their skills and aspirations and receive recommendations for part time projects, internal positions and mentors; managers posting projects see employees whose skills fit. Gloat reports that more than 2,300 employees began to explore new roles within the first two months, alongside an adoption rate above 60% whose base it does not state, and cumulative figures of more than 360,000 unlocked hours and over USD 15 million in productivity gains and reduced recruitment costs.","stage":"scaled","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":2300,"unit":"count","qualifier":"at-least","period":"first two months after launch","baseline":"employees who began to explore new roles through the platform; the same sentence gives an adoption rate above 60% without saying what it is a share of, and 2,300 is about 1.5% of the 155,000 workforce, so the percentage is not recorded","claimant":"vendor","quote":"Within the first two months of launch, the platform achieved an adoption rate that surpassed 60%, enabling more than 2,300 employees to begin to explore new roles within the business.","sourceUrl":"https://gloat.com/resources/customer-stories-2/how-schneider-electric-increased-employee-retention/"},{"kpi":"cost-savings","value":15000000,"unit":"currency","currency":"USD","qualifier":"at-least","period":"cumulative since the April 2020 launch, end date not stated","baseline":"productivity gains and reduced recruitment costs combined, as stated in the vendor's story; not an annual figure","claimant":"vendor","quote":"To date, Schneider Electric’s talent marketplace has unlocked more than 360,000 hours and created a savings of over $15,000,000 in productivity gains and reduced recruitment costs.","sourceUrl":"https://gloat.com/resources/customer-stories-2/how-schneider-electric-increased-employee-retention/"}],"outcomeDisclosed":true,"sources":[{"url":"https://gloat.com/resources/customer-stories-2/how-schneider-electric-increased-employee-retention/","title":"How Schneider Electric increased employee retention","publisher":"Gloat"},{"url":"https://www.cio.com/article/651553/schneider-electric-leverages-ai-to-help-develop-employees-careers.html","title":"Schneider Electric leverages AI to help develop employees' careers","publisher":"CIO"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"schneider-electric-open-talent-market"},{"title":"Scotiabank: AI agents that assemble the Client Insight Report for payments clients","useCases":["client-briefing-and-call-report-copilot"],"organization":{"name":"Scotiabank","anonymized":false,"country":"CA","region":"north-america","industry":"banking"},"vendors":[{"name":"Microsoft","role":"platform"},{"name":"EY","role":"integrator"}],"summary":"Scotiabank's Global Transaction Banking business prototyped a team of five specialised AI agents that transform, reconcile and explain a client's payment data and assemble the Client Insight Report, an analysis of what was processed, what failed and what remediation was needed, which the bank uses to discuss with the client the value it delivers. It is a payments analytics report that feeds client conversations rather than a full meeting briefing pack. The report used to be a manual, high touch service for a select set of clients; the bank says the work that took weeks now takes seconds, which unlocks the ability to serve all clients. The prototype was built with EY on Microsoft's Copilot platform in under three months.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://gtb.scotiabank.com/en/global-transaction-banking/resources/insights/article.insights.how-ai-agents-are-transforming-scotiabank-s-payment-operations.html","title":"How AI Agents are Transforming Scotiabank's Payment Operations","publisher":"Scotiabank Global Transaction Banking"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"scotiabank-client-insight-report-agents"},{"title":"Scotiabank: AI adverse media monitoring for anti money laundering with WorkFusion","useCases":["pep-and-adverse-media-screening"],"organization":{"name":"Scotiabank","anonymized":false,"country":"CA","region":"north-america","industry":"banking"},"vendors":[{"name":"WorkFusion","role":"platform"}],"summary":"Scotiabank automated its adverse media monitoring (negative news search) for anti money laundering with WorkFusion, applying the vendor's intelligent automation to the analysis and disposition of adverse media. The vendor reports a sharp fall in false positives, wider media search coverage (30 articles per name instead of 20) and the equivalent of more than a hundred compliance analysts freed for other work.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"false-positive-reduction","value":95,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"95% reduction in false positives","sourceUrl":"https://www.workfusion.com/customer-stories/scotiabank/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.workfusion.com/customer-stories/scotiabank/","title":"Scotiabank Customer Story","publisher":"WorkFusion","date":"2020-11-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"scotiabank-workfusion-adverse-media-monitoring"},{"title":"SeatGeek: ML anomaly detection and field level lineage for data quality","useCases":["data-quality-monitoring-agent"],"organization":{"name":"SeatGeek","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Monte Carlo","role":"platform"}],"summary":"SeatGeek's data platform and analytics teams were losing full days root causing data anomalies that business users noticed first, averaging about 10 internal data downtime issues a month. They adopted Monte Carlo's ML enabled anomaly detection and field level lineage tracking to catch problems before they reached business users. Monte Carlo reports that SeatGeek reduced data incidents per month from 10 to 0 in the second quarter after enabling the platform at scale, and cut the resource drain from root cause analysis by half.","stage":"production","year":2022,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"productivity-gain","value":50,"unit":"percent","qualifier":"exact","period":"since implementing Monte Carlo at scale","baseline":"Resource drain from root cause analysis before Monte Carlo","claimant":"vendor","quote":"Reduced resource drain from root-cause analysis by 50% and improved efficiency across all data teams","sourceUrl":"https://www.montecarlodata.com/blog-how-seatgeek-reduced-data-incidents-to-zero-with-data-observability/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.montecarlodata.com/blog-how-seatgeek-reduced-data-incidents-to-zero-with-data-observability/","title":"How SeatGeek Reduced Data Incidents to Zero with Data Observability","publisher":"Monte Carlo"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"seatgeek-data-quality-observability"},{"title":"SEB: quoted at the launch of Feedzai's Farol fraud agent","useCases":["fraud-alert-triage"],"organization":{"name":"SEB","anonymized":false,"region":"europe","industry":"banking"},"vendors":[{"name":"Feedzai","role":"platform"}],"summary":"Feedzai launched Farol in September 2026, an AI agent embedded in its fraud platform that retrieves and summarises alert data for investigators and supports fraud strategy work such as rule suggestions. SEB, described in the release as a northern European financial services group, is quoted at launch through its fraud prevention business owner on using a single interface for data retrieval, insight generation and rule suggestions. The release does not say how or how widely SEB uses Farol, and no SEB specific results are disclosed.","stage":"announced","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.prnewswire.com/news-releases/as-banks-pivot-to-agentic-ai-feedzai-unveils-farol-to-transform-fraud-analysis-and-cut-investigation-times-302888211.html","title":"As Banks Pivot to Agentic AI, Feedzai Unveils Farol to Transform Fraud Analysis and Cut Investigation Times","publisher":"Feedzai via PR Newswire","date":"2026-09-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"seb-feedzai-farol-fraud-agent"},{"title":"SEB: AI agent suggests responses and summarizes calls in wealth management","useCases":["live-agent-assist","client-meeting-notes-and-crm-update"],"organization":{"name":"SEB","anonymized":false,"country":"SE","region":"europe","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Bain & Company","role":"integrator"}],"summary":"SEB, a Nordic corporate bank, worked with Bain & Company to build an AI agent on Google Cloud for its wealth management division. The agent suggests responses during conversations with customers and generates call summaries afterwards. Google Cloud reports a 15% efficiency gain.","stage":"production","year":2025,"channels":["agent-desktop"],"languages":[],"metrics":[{"kpi":"productivity-gain","value":15,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"The agent, built with Google Cloud, enhances end-customer conversations with suggested responses and generates call summaries, helping to increase efficiency by 15%.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"seb-wealth-advisor-agent-assist"},{"title":"US Securities and Exchange Commission: CoCounsel pilot and AI features in legal research services","useCases":["legal-research-and-drafting-assistant"],"organization":{"name":"U.S. Securities and Exchange Commission","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Thomson Reuters","role":"platform"},{"name":"LexisNexis","role":"platform"}],"summary":"The SEC's Division of Enforcement, Division of Examinations, Office of the General Counsel and Office of the Chief Data Officer are piloting Westlaw CoCounsel, a generative AI assistant, since October 2025 for legal research, document analysis and document production; its outputs are document summaries, draft documents and answers to user selected questions. Separately, the agency has used AI features in Lexis+ and Westlaw Precision for legal research since January 2020, and Enforcement uses Westlaw Quick Check to cross check SEC documents against case law. No outcome figures are published.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entries SEC-44 Westlaw Co-Counsel, SEC-62, SEC-63 and SEC-35)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"sec-cocounsel-legal-research-pilot"},{"title":"US SEC: AI that surfaces accounts trading ahead of material price moves","useCases":["market-abuse-surveillance-triage"],"organization":{"name":"U.S. Securities and Exchange Commission","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Aretec","role":"integrator"}],"summary":"The SEC Division of Enforcement uses a classical machine learning tool, operational since April 2018, to identify accounts whose trading came in advance of material equity price moves and that warrant further investigation. The output is a list of leads for enforcement staff, who decide what to investigate. The entry appears in the 2025 US federal AI use case inventory; no outcome figures are published.","stage":"production","year":2018,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry SEC-34, Single Event Insider Trading Analysis)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"sec-single-event-insider-trading-analysis"},{"title":"Sedgwick: Sidekick Agent for claims examiner guidance","useCases":["claims-triage-and-straight-through-processing"],"organization":{"name":"Sedgwick","anonymized":false,"country":"US","region":"global","industry":"insurance"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Sedgwick, a global claims administrator, integrated Sidekick Agent into the workflows and screens of its own claims management systems. Built on Azure OpenAI Service and Azure AI Document Intelligence, it gives examiners claim insights, the day's top priorities, forecasts of anticipated claim trajectories, analytics on claim durations and reserves, and guidance on the next steps in the claim life cycle, and supports quality assurance for consistency and compliance. It follows an earlier version of Sidekick built on ChatGPT technology. No outcome figures were published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sedgwick.com/press-release/sedgwick-optimizes-claim-workflows-with-ai-application-sidekick-and-microsoft-integration/","title":"Sedgwick optimizes claim workflows with AI application Sidekick and Microsoft integration","publisher":"Sedgwick","date":"2025-04-29"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"sedgwick-sidekick-agent-claims-guidance"},{"title":"SEP2: custom Gemini agents for security alert triage in a managed SOC","useCases":["security-alert-triage-and-investigation"],"organization":{"name":"SEP2","anonymized":false,"country":"GB","region":"europe","industry":"technology"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"SEP2, a UK managed security provider, built triage and threat intelligence agents on Google's Gemini Enterprise Agent Platform alongside Google Security Operations. When an alert fires, the agents gather data from firewalls, endpoints and threat feeds in about a minute, a task that took analysts up to 20 minutes, and hand a standardised summary to the human team, who review it and authorise remediation. Engineers also use Gemini to write detection rules from plain language.","stage":"production","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":20,"unit":"multiplier","qualifier":"exact","period":"gathering the data for a new security alert before human review, from up to 20 minutes to one minute","claimant":"vendor","quote":"20x faster to triage new security alerts","sourceUrl":"https://cloud.google.com/customers/sep2"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/sep2","title":"SEP2 triages cybersecurity threats 20x faster with Gemini Enterprise Agent Platform","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"sep2-gemini-security-triage-agents"},{"title":"Serious Fraud Office: AI document review for privilege and disclosure in criminal cases","useCases":["ediscovery-and-disclosure-document-review"],"organization":{"name":"Serious Fraud Office","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"OpenText","role":"platform"}],"summary":"The UK Serious Fraud Office first used an AI tool in its Rolls-Royce investigation to screen about 30 million documents for material potentially covered by legal professional privilege, work that independent barristers had previously done by hand. From April 2018 it made AI document review available to all new cases and adopted OpenText Axcelerate, which groups documents by subject, builds timelines and removes duplicates; at launch the SFO said it would eventually be able to sift for relevancy. The 2024 inspection by HM Crown Prosecution Service Inspectorate describes Axcelerate's artificial intelligence and machine learning features and how document reviewers tag material for relevancy in it. It warned that there remains a risk that some staff are not confident using the system, found that a number of staff were not using it to its full potential and that many viewed its training as inadequate, and warned that searching millions of documents is not an exact science. It also found that disclosure problems in an earlier case were compounded by a misunderstanding of how searches worked in the SFO's previous document review system, which is no longer used.","stage":"scaled","year":2018,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":30000000,"unit":"count","qualifier":"approximately","period":"Rolls-Royce investigation pilot, privilege screening","claimant":"organization","quote":"It enabled the estimated 30 million documents provided by the company to be analysed for material potentially covered by Legal Professional Privilege.","sourceUrl":"https://www.wired-gov.net/wg/news.nsf/articles/AI+powered+RoboLawyer+helps+step+up+the+SFOs+fight+against+economic+crime+11042018162000?open="}],"outcomeDisclosed":true,"sources":[{"url":"https://www.wired-gov.net/wg/news.nsf/articles/AI+powered+RoboLawyer+helps+step+up+the+SFOs+fight+against+economic+crime+11042018162000?open=","title":"AI powered 'Robo-Lawyer' helps step up the SFO's fight against economic crime (official press release)","publisher":"Serious Fraud Office, republished by Wired-Gov","date":"2018-04-11"},{"url":"https://www.sfo.gov.uk/2018/04/10/ai-powered-robo-lawyer-helps-step-up-the-sfos-fight-against-economic-crime/","title":"AI powered 'Robo-Lawyer' helps step up the SFO's fight against economic crime","publisher":"Serious Fraud Office","date":"2018-04-10","archivedUrl":"https://web.archive.org/web/20240225110751/https://www.sfo.gov.uk/2018/04/10/ai-powered-robo-lawyer-helps-step-up-the-sfos-fight-against-economic-crime/"},{"url":"https://cdn.websitebuilder.service.justice.gov.uk/uploads/sites/24/2024/08/SFO-Disclosure-Report-2.pdf","title":"Serious Fraud Office: Disclosure. An inspection of the handling and management of disclosure in the Serious Fraud Office","publisher":"HM Crown Prosecution Service Inspectorate"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"serious-fraud-office-ai-document-review"},{"title":"Sheba Medical Center: AI triage for intracranial hemorrhage","useCases":["radiology-worklist-triage"],"organization":{"name":"Sheba Medical Center","anonymized":false,"country":"IL","region":"middle-east","industry":"healthcare"},"vendors":[{"name":"Aidoc","role":"platform"}],"summary":"Sheba Medical Center, where Aidoc originated, has embedded Aidoc's AI platform across its emergency and radiology workflows to flag urgent findings, including intracerebral hemorrhage, large vessel occlusion stroke and pulmonary embolism, directly on images in real time, and alert the treating physician on desktop and mobile. Sheba's own account of the deployment cites a peer reviewed clinical study that found the integration of Aidoc into its emergency workflow was associated with a 30% reduction in mortality for patients with intracerebral hemorrhage, earlier treatment initiation, improved discharge outcomes and fewer unnecessary ICU stays. Sheba's own site frames Aidoc as one part of its wider Smart Hospital program, and lists a separate initiative, Project K, an AI powered emergency room, as a related case study; the page does not describe Project K as an extension of Aidoc's triage.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://sheba-global.com/project/aidoc/","title":"Aidoc: Real-Time AI-Powered Radiology at Sheba","publisher":"Sheba Medical Center","archivedUrl":"https://web.archive.org/web/20251225075441/https://sheba-global.com/project/aidoc/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"sheba-medical-center-aidoc-triage"},{"title":"Sheffield Children's NHS Foundation Trust: AI predictor that targets extra reminders and transport support","useCases":["patient-appointment-scheduling-and-reminders-agent","outbound-reminder-and-confirmation-agent"],"organization":{"name":"Sheffield Children's NHS Foundation Trust","anonymized":false,"country":"GB","region":"europe","industry":"healthcare"},"vendors":[{"name":"Alder Hey Innovation","role":"platform"}],"summary":"Sheffield Children's piloted an AI Predictor developed by Alder Hey Innovation that estimates which children are likely to miss (\"was not brought\") an appointment, using markers that include health inequalities. Families with a predicted risk of 50% or more received an extra text reminder with an offer of support the day before; families at 85% or more were contacted and offered funded transport or a rebooking. NHS England reports that 53,800 texts were sent in the first 12 months, that recorded non attendance came in well below the expected benchmark (almost 200 more attended appointments a month), and that in a 13 week period 152 families had transport arranged and 129 appointments were rebooked.","stage":"pilot","year":2023,"channels":["sms"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/","title":"NHS AI expansion to help tackle missed appointments and improve waiting times","publisher":"NHS England","date":"2024-03-14","archivedUrl":"https://web.archive.org/web/20250125233954/https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"sheffield-childrens-ai-attendance-predictor"},{"title":"Shell: AI predictive maintenance on the C3 AI platform scaled to 10,000 pieces of equipment","useCases":["industrial-asset-predictive-maintenance"],"organization":{"name":"Shell","anonymized":false,"country":"GB","region":"global","industry":"energy-and-utilities"},"vendors":[{"name":"C3 AI","role":"platform"}],"summary":"Shell runs a predictive maintenance programme built on the C3 AI platform. Machine learning models flag equipment degradation and likely failures early so operators can intervene before unplanned downtime, production interruptions or safety and environmental risks; C3 AI names control valves, pumps and compressors among the monitored equipment. In a March 2022 C3 AI press release, Shell's Dan Jeavons said that \"Monitoring 10,000 pieces of critical equipment\" with AI predictive maintenance was a target Shell had set for 2021 and achieved. The wording \"more than 10,000\", the asset scope and the technical figures in the release are C3 AI's. In its own TechXplorer Digest article, Shell describes the rollout asset by asset (a Dutch refinery in 2020, where the models flagged 65 control valves in need of repair that traditional methods would have missed, then Singapore, the USA and Canada in early 2021) and a process in which, for most assets, a remote engineer vets each alert before it reaches the asset engineers.","stage":"scaled","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://c3.ai/shell-achieves-major-milestone-scales-artificial-intelligence-predictive-maintenance-to-10000-pieces-of-equipment-using-c3-ai/","title":"Shell Achieves Major Milestone: Scales Artificial Intelligence Predictive Maintenance to 10,000 Pieces of Equipment Using C3 AI","publisher":"C3 AI","date":"2022-03-08"},{"url":"https://www.shell.com/what-we-do/technology-and-innovation/shell-techxplorer-digest/shell-techxplorer-digest-2020/_jcr_content/root/main/section/list_copy_copy_copy/list_item_copy_98181_819446707/links/item0.stream/1669888451651/dabc9c17a2c9a00d39cb4f442e75d667920c8562/the-shell-journey-towards-global-predictive-maintenance-velthuis.pdf","title":"The Shell journey towards global predictive maintenance","publisher":"Shell"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"shell-c3-ai-predictive-maintenance"},{"title":"Shift4: AI transaction monitoring replacing static rules with ThetaRay","useCases":["aml-alert-triage"],"organization":{"name":"Shift4","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[{"name":"ThetaRay","role":"platform"}],"summary":"Payments company Shift4 selected ThetaRay's AI transaction monitoring platform in late 2023 to replace static threshold rules, and completed its European deployment in the first quarter of 2024. The vendor reports a much lower false positive rate and more productive alerts that lead to investigations, with explainable risk scores for the compliance team.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"false-positive-reduction","value":86,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"How Shift4 slashed false positives by 86% and reclaimed analyst bandwidth across $200B+ in annual volume","sourceUrl":"https://thetaray.com/customer-stories/how-shift4-increased-productive-alerts-by-70-with-cognitive-ai-transaction-monitoring/"}],"outcomeDisclosed":true,"sources":[{"url":"https://thetaray.com/customer-stories/how-shift4-increased-productive-alerts-by-70-with-cognitive-ai-transaction-monitoring/","title":"How Shift4 Increased Productive Alerts by 70%","publisher":"ThetaRay","date":"2026-08-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"shift4-thetaray-aml-transaction-monitoring"},{"title":"SIGNAL IDUNA: Co SI knowledge assistant for health insurance service agents","useCases":["enterprise-knowledge-search","live-agent-assist"],"organization":{"name":"SIGNAL IDUNA","anonymized":false,"country":"DE","region":"europe","industry":"insurance"},"vendors":[{"name":"Google Cloud","role":"platform"},{"name":"Boston Consulting Group","role":"integrator"},{"name":"Deloitte","role":"integrator"}],"summary":"SIGNAL IDUNA, a German insurer, built Co SI with Google Cloud, BCG and Deloitte: a knowledge assistant that helps customer service agents answer complex health insurance questions. Google Cloud reports that for less experienced agents, information searches are 30% faster and inquiries that previously needed further escalation dropped from 27% to 3%.","stage":"production","year":2025,"channels":["agent-desktop"],"languages":["de"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","archivedUrl":"https://web.archive.org/web/20251027121348/https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"signal-iduna-co-si-knowledge-assistant"},{"title":"SimCorp: Wealth Vision and Wealth Lens rebalancing scores","useCases":["portfolio-drift-monitoring-and-rebalancing"],"organization":{"name":"SimCorp","anonymized":false,"country":"DK","region":"europe","industry":"technology"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"SimCorp, an investment management software provider that uses Azure Machine Learning as its main AI platform and Semantic Kernel to build its AI solutions, describes Wealth Vision: it gives a portfolio manager a list of portfolios to rebalance, and its Wealth Lens tool scores each portfolio between 0.00 and 1.00, where 1.00 marks a prime candidate for rebalancing. This is a vendor product description on a Microsoft blog; no named client deployment or outcome is published.","stage":"announced","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://devblogs.microsoft.com/semantic-kernel/customer-case-study-simcorps-ai-journey-with-semantic-kernel/","title":"Customer Case Study: SimCorp's AI Journey with Semantic Kernel","publisher":"Microsoft Developer Blogs","date":"2024-07-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"D","id":"simcorp-wealth-lens-rebalancing"},{"title":"Singtel: agentic AI assistant Shirley for care, roaming and sales","useCases":["device-and-connectivity-troubleshooting-agent","order-to-activation-and-esim-onboarding-assistant","plan-upgrade-and-sales-assistant"],"organization":{"name":"Singtel","anonymized":false,"country":"SG","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Sierra","role":"platform"}],"summary":"In a partnership announced on 4 March 2026, Singtel upgraded the customer care assistant Shirley of Singtel Singapore with Sierra's agentic AI, starting with a pilot that went live in under ten weeks. Shirley verifies customer details, resolves mobile and home troubleshooting, completes roaming sign ups and understands local expressions including Singlish; customers purchased more than 200 roaming add ons independently. In its first six weeks it handled over 70,000 cases. Singtel Singapore also says it will deploy voice AI agents for outbound sales calls within defined compliance and governance standards.","stage":"production","year":2026,"channels":["web-chat","voice"],"languages":["en"],"metrics":[{"kpi":"containment-rate","value":73,"unit":"percent","qualifier":"exact","period":"mobile and home troubleshooting cases, initial results after launch","claimant":"organization","quote":"73% of mobile and home troubleshooting cases were resolved without requiring a Customer Care officer.","sourceUrl":"https://sierra.ai/customers/singtel"},{"kpi":"automation-rate","value":76,"unit":"percent","qualifier":"exact","period":"roaming sign up requests, initial results after launch","claimant":"organization","quote":"76% of roaming sign-up requests were completed successfully without requiring a Customer Care officer.","sourceUrl":"https://sierra.ai/customers/singtel"},{"kpi":"interactions-handled","value":70000,"unit":"count","qualifier":"at-least","period":"first six weeks after launch","claimant":"organization","quote":"Singtel went live in less than 10 weeks, and in the first six weeks since launch, ”Shirley” handled over 70,000 customer cases involving high volume requests in areas such as mobile issues and roaming services.","sourceUrl":"https://sierra.ai/customers/singtel"}],"outcomeDisclosed":true,"sources":[{"url":"https://sierra.ai/customers/singtel","title":"Singtel Group partners with Sierra to transform customer engagement with AI","publisher":"Sierra"},{"url":"https://www.singtel.com/about-us/media-centre/news-releases/singtel-group-partners-sierra-to-transform-custome-engagement-with-ai","title":"Singtel Group partners Sierra to transform customer engagement with AI","publisher":"Singtel","date":"2026-03-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"singtel-shirley-agentic-ai-agent"},{"title":"Skyward Specialty: AI submission preprocessing and risk summaries across six business units","useCases":["underwriting-risk-assessment-copilot","commercial-underwriting-submission-triage"],"organization":{"name":"Skyward Specialty Insurance Group","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Sixfold","role":"platform"}],"summary":"Skyward Specialty announced in December 2025 that its partnership with Sixfold was entering its second year. The platform preprocesses submissions and generates recommendations on prioritization, appetite alignment and risk summarization and assessment, while underwriters stay in the loop to apply their judgment. The platform is live across six business units and more than 10 product lines, with an average deployment timeline of 8 to 10 weeks. The company presents the partnership as a step toward fully AI powered underwriting across its US property and casualty lines; no outcome figures were disclosed.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://skywardinsurance.com/press-releases/skyward-specialty-and-sixfold-partner-to-advance-ai-powered-underwriting/","title":"Skyward Specialty and Sixfold Partner to Advance AI-Powered Underwriting","publisher":"Skyward Specialty Insurance Group","date":"2025-12-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"skyward-specialty-sixfold-ai-underwriting"},{"title":"Southern California Gas Company: digital energy efficiency reports with Bidgely","useCases":["smart-meter-analytics"],"organization":{"name":"Southern California Gas Company (SoCalGas)","anonymized":false,"country":"US","region":"north-america","industry":"energy-and-utilities"},"vendors":[{"name":"Bidgely","role":"platform"}],"summary":"Southern California Gas Company (SoCalGas) worked with Bidgely on a digital only home energy report programme for medium consumption residential gas customers, a segment that traditional paper based home energy reports, which target high consumption customers, do not reach. The reports were built on AMI meter disaggregation, and the programme exceeded its savings goal, saving 565,000 therms by December 2020, with more than 405,000 customers receiving the reports digitally at a 50 percent open rate.","stage":"production","year":2020,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.bidgely.com/resources/southern-california-gas-company-case-study-with-bidgely/","title":"SoCalGas Case Study: Delivering Energy Efficiency","publisher":"Bidgely","archivedUrl":"https://web.archive.org/web/20250115180354/https://www.bidgely.com/resources/southern-california-gas-company-case-study-with-bidgely/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"socalgas-bidgely-energy-efficiency"},{"title":"SoftBank: AI agent for small and medium business sales inquiries","useCases":["business-connectivity-quoting-and-service-assistant"],"organization":{"name":"SoftBank Corp.","anonymized":false,"country":"JP","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Sierra","role":"platform"}],"summary":"SoftBank's Customer Growth Division, which sells to Japan's small and medium sized businesses, put a conversational AI agent on its corporate website that answers questions about products, services, pricing and contracts. What it cannot resolve goes through an inquiry form to a sales representative. SoftBank says the self resolution rate, measured by the platform's own AI assessment, rose from about 50% at launch to 70%; Sierra puts the volume at about 100 inquiries a day.","stage":"production","year":2026,"channels":["web-chat"],"languages":["ja"],"metrics":[{"kpi":"containment-rate","value":70,"unit":"percent","qualifier":"exact","period":"website inquiries from business customers","baseline":"about 50% at launch","claimant":"organization","quote":"SoftBank's corporate business covers a wide range of services, and the self-resolution rate* — which started at around 50% at launch — has now risen to 70%, thanks to iterative improvements to help the AI agent better understand customer intent and the relevant service.","sourceUrl":"https://sierra.ai/customers/softbank"},{"kpi":"interactions-handled","value":100,"unit":"count","qualifier":"approximately","period":"inquiries per day","claimant":"vendor","quote":"Currently, the AI agent handles approximately 100 inquiries per day, with 70% resulting in customers finding the information they need and resolving their inquiries on their own.","sourceUrl":"https://sierra.ai/customers/softbank"}],"outcomeDisclosed":true,"sources":[{"url":"https://sierra.ai/customers/softbank","title":"How SoftBank's Customer Growth Division delivers fast, high-quality customer service with AI","publisher":"Sierra","date":"2026-07-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"softbank-smb-sales-ai-agent"},{"title":"Softcat: organisation wide Microsoft 365 Copilot use for meeting summaries and follow ups","useCases":["meeting-summarization-and-action-items"],"organization":{"name":"Softcat","anonymized":false,"country":"GB","region":"europe","industry":"technology"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Softcat, the largest Microsoft Solutions Partner in the UK, widened its Microsoft 365 Copilot rollout to 1,500 people. One of its top sellers queries meeting transcripts to pull out information and list actions, so he sends follow ups faster and no longer takes notes during customer meetings; its IT change manager estimates that a third of users use it daily for tasks such as email and meeting summaries. Microsoft reports that 85% of licensed users use it regularly.","stage":"scaled","year":2024,"channels":["microsoft-teams","internal-tools"],"languages":["en"],"metrics":[{"kpi":"employee-adoption","value":85,"unit":"percent","qualifier":"exact","period":"licensed users using Copilot regularly","claimant":"vendor","quote":"Eighty-five percent of licenced Softcat users are using Microsoft 365 Copilot regularly.","sourceUrl":"https://www.microsoft.com/en/customers/story/19912-softcat-microsoft-365-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/19912-softcat-microsoft-365-copilot","title":"Softcat leads Microsoft 365 Copilot adoption to deliver efficiencies and quality improvements","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"softcat-copilot-meeting-summaries"},{"title":"Spotify: Discover Weekly personalized playlist","useCases":["content-recommendation-and-personalization"],"organization":{"name":"Spotify","anonymized":false,"country":"SE","region":"europe","industry":"media-and-entertainment"},"vendors":[],"summary":"Discover Weekly, \"Spotify's first personalized playlist\", updates every Monday with songs and artists \"handpicked just for them\". Spotify's own support documentation lists Discover Weekly among its personalized playlists, which are \"created by Spotify's algorithms that look at factors like what the person is listening to and when, which songs they're adding to their playlists, the listening habits of people who have similar tastes, and much more\". Ten years after launch, Spotify says the playlist has driven \"more than 100 billion tracks streamed\" and \"ignites more than 56 million new artist discoveries\" every week, \"with 77% coming from emerging artists\". These are cumulative and weekly volume figures Spotify discloses about the whole feature, not a measured before and after effect. In 2025 Spotify added up to five genre options, \"personalized based on your listening history\", that generate a fresh 30 track playlist \"inspired by your selection\".","stage":"scaled","year":2025,"channels":["mobile-app"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://newsroom.spotify.com/2025-06-30/discover-weekly-turns-10-celebrating-100-billion-tracks-streamed-and-a-decade-of-personalized-discovery/","title":"Discover Weekly Turns 10: Celebrating 100 Billion+ Tracks Streamed and a Decade of Personalized Discovery","publisher":"Spotify Newsroom","date":"2025-06-30"},{"url":"https://support.spotify.com/us/artists/article/types-of-spotify-playlists/","title":"Types of Spotify playlists","publisher":"Spotify Support"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"spotify-discover-weekly-personalization"},{"title":"Square Enix: AI optimized, personalized player emails","useCases":["personalized-marketing-at-scale"],"organization":{"name":"Square Enix","anonymized":false,"country":"JP","region":"asia-pacific","industry":"media-and-entertainment"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Square Enix uses customer data to develop AI optimized marketing assets and send players personalized emails matched to their preferences. Google Cloud reports a 20% increase in email opens and a 10% increase in retention.","stage":"production","year":2024,"channels":["email"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"},{"url":"https://www.hd.square-enix.com/eng/company/","title":"Company","publisher":"Square Enix Holdings"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"square-enix-personalized-emails"},{"title":"US Social Security Administration: Customer Insight Tool for survey text analytics","useCases":["customer-feedback-analysis"],"organization":{"name":"U.S. Social Security Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Medallia","role":"platform"}],"summary":"SSA's customer survey system includes an AI text analytics capability that reads the free text of survey answers and returns sentiment ratings, categorized themes, suggested actions, customer effort indicators and machine translation. The agency uses it to spot emerging trends early and to see the positive and negative drivers in customer interactions. It is listed as deployed since August 2021; no outcome figures are published.","stage":"production","year":2021,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (SSA entry, Customer Insight Tool)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"ssa-customer-insight-survey-text-analytics"},{"title":"SS&C GIDS and RS: AI agents that draft complaint closing letters for human review","useCases":["outbound-notice-drafting"],"organization":{"name":"SS&C GIDS and RS","anonymized":false,"country":"GB","region":"europe","industry":"wealth-and-asset-management"},"vendors":[{"name":"SS&C Blue Prism","role":"in-house"}],"summary":"SS&C Global Investor and Distribution Solutions and Retirement Solutions, which runs customer service for asset managers, insurers and wealth managers under FCA rules, rebuilt its complaints process with AI agents. After a human investigator records the findings, the agents use the case notes to draft the closing letter that summarises them; an employee checks the letter before the agents send it. SS&C Blue Prism, a business in the same SS&C group, reports that complaint cycle times fell by 25%.","stage":"production","year":2026,"channels":["email","internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":25,"unit":"percent","qualifier":"exact","period":"complaint cycle time, whole process","claimant":"vendor","quote":"Cycle times have been reduced by 25%, so customers get a follow-up letter more quickly.","sourceUrl":"https://www.blueprism.com/resources/case-studies/ssnc-gids-agentic-agents-ai-customer-service/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.blueprism.com/resources/case-studies/ssnc-gids-agentic-agents-ai-customer-service/","title":"SS&C GIDS & RS | Agentic AI Case Study in Customer Service | SS&C Blue Prism","publisher":"SS&C Blue Prism","date":"2026-01-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ssc-gids-complaint-closing-letters"},{"title":"SS&C GIDS: in house LLM that drafts customer letters","useCases":["outbound-notice-drafting"],"organization":{"name":"SS&C GIDS","anonymized":false,"region":"global","industry":"wealth-and-asset-management"},"vendors":[{"name":"SS&C Blue Prism","role":"in-house"}],"summary":"SS&C GIDS handles customer communications for asset managers and other financial institutions, such as instructions on selling assets and responses to changes of address, broker or customer ID. After an employee investigates a request and records comments in a template, a digital worker validates the case and prompts an in house large language model, which generates a personalised letter; quality control reviews and adjusts it before it is sent. SS&C Blue Prism, a business in the same group, reports that these communications are now produced three times faster than with the manual process.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.blueprism.com/resources/case-studies/ssc-gids-ai-customer-communications/","title":"SS&C GIDS | BPM, IA & AI for Customer Communications | SS&C Blue Prism","publisher":"SS&C Blue Prism","date":"2024-12-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ssc-gids-llm-customer-letters"},{"title":"SS&C Technologies: generative AI document agents for loan servicing and address changes","useCases":["account-servicing-execution"],"organization":{"name":"SS&C Technologies","anonymized":false,"country":"US","region":"global","industry":"wealth-and-asset-management"},"vendors":[{"name":"SS&C Blue Prism","role":"in-house"}],"summary":"SS&C Technologies, which provides software and services to wealth and asset management firms and runs fund administration, linked its robotic process automation digital workers to a secure, proprietary large language model so they can read unstructured documents. For loan credit agreements the digital workers ask the model for key terms, validate the answers and enter them into SS&C GoLoans, routing discrepancies to an employee. It reports that credit agreements are now processed in six minutes, 95% faster than by hand, and it uses the same approach for address changes and collateral margin call agreements.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":95,"unit":"percent","qualifier":"exact","period":"loan credit agreement processing, now six minutes","baseline":"manual review of two hours per document","claimant":"vendor","quote":"These digital workers, also known as document agents, now process loan credit agreements in just six minutes — 95% faster than the manual process.","sourceUrl":"https://www.blueprism.com/resources/case-studies/ssc-ai-lending-loan-unstructured-data/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.blueprism.com/resources/case-studies/ssc-ai-lending-loan-unstructured-data/","title":"SS&C Tech | Automation & AI for Unstructured Loan Lending Data | SS&C Blue Prism","publisher":"SS&C Blue Prism","date":"2024-09-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ssc-technologies-generative-ai-document-agents"},{"title":"St. Luke's University Health Network: Security Copilot agents for phishing alert triage","useCases":["security-alert-triage-and-investigation"],"organization":{"name":"St. Luke's University Health Network","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"St. Luke's University Health Network, a US hospital network with more than 23,000 employees, runs Microsoft Security Copilot across Defender and Sentinel. Its Security Alert Triage Agent (formerly the Phishing Triage Agent) reads user reported emails, decides whether each is a genuine phishing attempt or a false alarm, explains its verdict in plain text and closes false positives on its own, so analysts can move to threat hunting. Copilot also drafts incident reports that analysts edit and escalate.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"hours-saved","value":200,"unit":"hours","qualifier":"approximately","period":"per month, phishing alert triage","claimant":"organization","quote":"It’s saving us nearly 200 hours monthly by autonomously handling and closing thousands of false positive alerts.","sourceUrl":"https://www.microsoft.com/en/customers/story/25330-st-lukes-university-health-network-microsoft-security-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/25330-st-lukes-university-health-network-microsoft-security-copilot","title":"St. Luke’s saves nearly 200 hours monthly with AI-powered Security Copilot agents","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"st-lukes-security-alert-triage-agent"},{"title":"Standard Bank Group: automated letter of credit document checking at Stanbic Bank Uganda","useCases":["trade-document-examination","trade-finance-crime-screening"],"organization":{"name":"Stanbic Bank Uganda","anonymized":false,"country":"UG","region":"africa","industry":"banking"},"vendors":[{"name":"Traydstream","role":"platform"}],"summary":"Stanbic Bank Uganda, part of Standard Bank Group, signed an agreement to implement Traydstream's platform to digitise the manual vetting of letters of credit for discrepancies, after trade document processing on it over the previous few months. The platform digitises the documents, checks them against trade rules and adds an aggregated compliance module; the group presented it as faster processing with more thorough trade checks and more transparent transactions. No figures are published.","stage":"pilot","year":2021,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://traydstream.com/traydstream-and-standard-bank-group-are-pleased-to-announce-achieving-a-major-milestone-with-its-automated-trade-document-checking-solution-roll-out/","title":"Traydstream and Standard Bank Group are pleased to announce achieving a major milestone with its automated trade document checking solution roll out","publisher":"Traydstream","date":"2021-04-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"standard-bank-automated-trade-document-checking"},{"title":"Standard Chartered: Client Insights and automated call reports for corporate bankers","useCases":["client-briefing-and-call-report-copilot"],"organization":{"name":"Standard Chartered","anonymized":false,"country":"GB","region":"global","industry":"banking"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Standard Chartered's Corporate and Investment Banking division began moving its client relationship management onto one Dynamics 365 platform in 2021, to improve client engagement for the division's clients in the 67 countries in which it operates. The programme was meant to make the platform's more than 6,000 users more effective and productive. The platform combines news feeds and internal information with AI in a Client Insights feature that surfaces opportunities for proactive engagement, and makes automated call reports that document client engagement easy to create and share in Outlook; the story does not say that AI drafts those call reports. It says the bank plans to adopt further AI capabilities through Microsoft Copilot for Sales next.","stage":"scaled","year":2021,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1762958431882021291-standard-chartered-bank-dynamics-365-sales-banking-and-capital-markets-en-united-kingdom","title":"Standard Chartered sets new standard in innovation with Dynamics 365","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"standard-chartered-cib-client-insights-crm"},{"title":"Standard Chartered: Instabase machine learning for corporate client onboarding and KYC","useCases":["corporate-account-onboarding-orchestration"],"organization":{"name":"Standard Chartered","anonymized":false,"country":"GB","region":"global","industry":"banking"},"vendors":[{"name":"Instabase","role":"platform"}],"summary":"Standard Chartered partnered with Instabase to automate client onboarding, credit documentation and know your customer checks in Corporate and Institutional Banking. The machine learning and natural language processing solution reads documents and unstructured forms, and automates client due diligence by sourcing sanctions and adverse media information from public and private registries, so cases are processed automatically overnight instead of by staff copying and pasting from registries by hand. The bank says average client onboarding time in that business fell from 41 to 8 days since 2015, and expects the Instabase solution to bring it down further. The solution went live in Singapore, India and the United Kingdom, with Bangladesh among its first markets and further rollout planned.","stage":"production","year":2018,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"cycle-time-days","value":8,"unit":"days","qualifier":"exact","period":"average onboarding time in Corporate and Institutional Banking, at the time of the 2018 announcement","baseline":"41 days, in 2015","claimant":"organization","quote":"In Corporate & Institutional Banking (CIB), average client onboarding times have dramatically reduced from 41 to 8 days since 2015, and will continue to fall with the use of Instabase to digitise the bank.","sourceUrl":"https://www.sc.com/en/press-release/were-transforming-client-experience-with-smart-technologies/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sc.com/en/press-release/were-transforming-client-experience-with-smart-technologies/","title":"We're transforming client experience with smart technologies","publisher":"Standard Chartered","date":"2018-11-07","archivedUrl":"http://web.archive.org/web/20260514130352/https://www.sc.com/en/press-release/were-transforming-client-experience-with-smart-technologies/"},{"url":"https://www.bankingtech.com/2018/11/standard-chartered-teams-with-instabase-for-clever-kyc/","title":"Standard Chartered teams with Instabase for clever KYC","publisher":"FinTech Futures (Banking Technology)","date":"2018-11-08","archivedUrl":"http://web.archive.org/web/20190605181832/https://www.bankingtech.com/2018/11/standard-chartered-teams-with-instabase-for-clever-kyc/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"standard-chartered-instabase-client-onboarding"},{"title":"Standard Chartered: LLM as a judge testing of a generative AI email drafting tool with PwC","useCases":["model-risk-validation-copilot"],"organization":{"name":"Standard Chartered","anonymized":false,"country":"GB","region":"global","industry":"banking"},"vendors":[{"name":"PwC","role":"integrator"}],"summary":"In the AI Verify Foundation's Global AI Assurance Pilot (February to May 2025), PwC acted as independent tester of a generative AI tool Standard Chartered built to draft personalised client emails for wealth relationship managers, a tool that was itself still in an internal pilot. PwC turned the requirements in the tool's system prompts into a structured checklist and ran batch tests on synthetic client profiles and edge cases, using an LLM as a judge to find hallucinations and contradictions against the input data and to score completeness, coherence, engagement and internal compliance, alongside NLP similarity metrics for robustness. Human subject matter experts scored a subset of drafts on the same framework to check that the automated judge was calibrated. The published case study reports the method and the effort involved, not the test results, which it says are confidential.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://assurance.aiverifyfoundation.sg/report/pilot-participants-and-use-cases/","title":"Pilot participants and use cases","publisher":"AI Verify Foundation"},{"url":"https://assurance.aiverifyfoundation.sg/wp-content/uploads/2025/05/StanChart-X-PwC-Solutions.pdf","title":"Wealth Relationship Manager Client Engagement Mail: Standard Chartered Bank x PwC","publisher":"AI Verify Foundation"},{"url":"https://aiverifyfoundation.sg/ai-assurance-pilot/","title":"Global AI Assurance Pilot","publisher":"AI Verify Foundation"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"standard-chartered-pwc-llm-judge-genai-validation"},{"title":"Standard Chartered: machine learning screening optimisation with Silent Eight","useCases":["sanctions-screening-adjudication"],"organization":{"name":"Standard Chartered","anonymized":false,"country":"GB","region":"global","industry":"banking"},"vendors":[{"name":"Silent Eight","role":"platform"}],"summary":"Standard Chartered announced a partnership with Silent Eight to give its financial crime compliance teams machine learning and natural language processing for name screening. The system recommends whether a screening alert is a true or false match and explains the recommendation in a plain English narrative for the analyst. The announcement describes aims rather than results.","stage":"announced","year":2018,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sc.com/en/press-release/weve-partnered-with-regulatory-technology-firm-silent-eight/","title":"We've partnered with Regulatory Technology firm Silent Eight","publisher":"Standard Chartered","date":"2018-07-09"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"standard-chartered-silent-eight-screening"},{"title":"Starling Bank: Starling Assistant and its smart tools for money management","useCases":["financial-wellbeing-coach"],"organization":{"name":"Starling Bank","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[{"name":"Google","role":"model-provider"}],"summary":"Starling Assistant is an agentic AI assistant in the Starling app that responds to text and voice, analyses spending patterns, creates savings Spaces and sets up transfers on the customer's behalf. In August 2026 Starling added \"smart tools\" to it and said new ones would follow every week for the rest of 2026 and at least monthly after that. The launch set includes a tax saver that sweeps a share of the transactions a small business picks into a Space, a Making Tax Digital guide, a spending quiz and a student budget planner. A rainy day saver, which works out with the customer how much they can realistically save and sets up transfers into a savings Space, was announced as a follow up tool. The assistant is built on Google's Gemini models. No outcome figures are disclosed.","stage":"production","year":2026,"channels":["mobile-app"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.starlingbank.com/news/starling-to-release-weekly-smart-tools","title":"Starling to release weekly 'smart tools' that supercharge money management for millions of Brits","publisher":"Starling Bank","date":"2026-08-20"},{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"starling-assistant-agentic-financial-assistant"},{"title":"Starling Bank: Scam Intelligence agent inside the banking app","useCases":["scam-payment-interception"],"organization":{"name":"Starling Bank","anonymized":false,"country":"GB","region":"europe","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"model-provider"}],"summary":"Starling launched Scam Intelligence in October 2025, letting customers upload marketplace ads and messages so a Gemini based model can flag signs of a purchase scam before they pay. In June 2026 the feature became an agent inside Starling Assistant, available to its five million customers: when a customer describes a planned transfer that looks like a romance, investment or other scam, the assistant asks probing questions, gives its view and suggests a call with the support team. Use is opt in and data stays in the bank's cloud environment.","stage":"production","year":2025,"channels":["mobile-app"],"languages":["en"],"metrics":[{"kpi":"detection-rate-improvement","value":300,"unit":"percent","qualifier":"exact","period":"since launch (October 2025), rate at which customers cancel marketplace payments; period and baseline not stated","baseline":"the rate at which customers cancelled marketplace payments before Scam Intelligence","claimant":"vendor","quote":"Scam Intelligence has already increased the rate at which customers cancel marketplace payments by 300%.","sourceUrl":"https://cloud.google.com/customers/starling"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.starlingbank.com/news/scam-intelligence-launch/","title":"Starling launches UK-first AI tool to combat scams","publisher":"Starling Bank","date":"2025-10-27"},{"url":"https://www.starlingbank.com/news/new-ai-feature-detects-romance-scammers/","title":"New AI feature detects romance scammers, investment heists and deepfake phishing attempts","publisher":"Starling Bank","date":"2026-06-24"},{"url":"https://cloud.google.com/customers/starling","title":"Starling case study","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"starling-bank-scam-intelligence"},{"title":"State Bank of India: machine learning models to identify income leakage","useCases":["fee-and-interest-leakage-detection"],"organization":{"name":"State Bank of India","anonymized":false,"country":"IN","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"State Bank of India","role":"in-house"}],"summary":"State Bank of India's Analytics Department builds machine learning models in house, and the bank's 2019-20 annual report lists models to identify income leakage among them, next to fraud, early warning and lead models. An article by an SBI chief manager in the journal of the Indian Institute of Banking and Finance (October to December 2021), citing the bank's Analytics Department, reports processing fees and facility fees recovered through this work in fiscal years 2018-19 and 2019-20. Neither source explains how the models work, how many accounts they cover or whether customers who were overcharged were also identified.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://sbi.bank.in/corporate/AR1920/download_center/english/11-4.2-Information%20Technology.pdf","title":"Annual Report 2019-20, Directors' Report: Information Technology","publisher":"State Bank of India","date":"2020-06-23"},{"url":"https://www.iibf.org.in/documents/BankQuest/Articles/5.Role%20of%20Aritificial%20Intelligence%20and%20Analytics%20in%20Banking%20-%20Kommana%20V%20Ganesh%20Kumar.pdf","title":"Role of Artificial Intelligence & Analytics in Banking","publisher":"Indian Institute of Banking & Finance (Bank Quest)","archivedUrl":"https://web.archive.org/web/20230314164128/http://www.iibf.org.in/documents/BankQuest/Articles/5.Role%20of%20Aritificial%20Intelligence%20and%20Analytics%20in%20Banking%20-%20Kommana%20V%20Ganesh%20Kumar.pdf"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"state-bank-of-india-income-leakage-models"},{"title":"stc: AI powered cognitive SON for autonomous radio optimisation","useCases":["network-planning-and-capacity-optimization","autonomous-network-operations"],"organization":{"name":"stc Group","anonymized":false,"country":"SA","region":"middle-east","industry":"telecommunications"},"vendors":[{"name":"Nokia","role":"platform"}],"summary":"Nokia deployed its MantaRay Cognitive SON, an AI powered feature of its self organizing network platform, in stc's commercial network in Saudi Arabia for the first time. The system optimises radio parameters autonomously. Nokia reports that during a period of high traffic it processed more than 10,000 actions, raised the utilisation rate of loaded cells by about 30 percent and average user throughput by 10 percent while traffic rose 40 percent, and that it reduced manual work. The results are stated by the vendor.","stage":"production","year":2024,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.nokia.com/newsroom/nokia-and-stc-group-optimize-network-with-ai-powered-mantaray-cognitive-son-solution-in-saudi-arabia/","title":"Nokia and stc Group optimize network with AI-powered MantaRay Cognitive SON solution in Saudi Arabia","publisher":"Nokia","date":"2024-07-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"stc-nokia-cognitive-son"},{"title":"Stripe: Radar fraud scoring and the Payments Foundation Model","useCases":["real-time-fraud-scoring"],"organization":{"name":"Stripe","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[{"name":"Stripe","role":"in-house"}],"summary":"Stripe's Radar scores payments on its network for fraud in real time and has learned from more than a decade of Stripe data. In May 2025 Stripe described a Payments Foundation Model trained on tens of billions of transactions that turns each payment into an embedding used for real time predictions; on sophisticated card testing attacks against large users, Stripe says detection rose from 59% to 97% overnight. A new multihead model now triggers step up authentication for risky payments below the block threshold.","stage":"scaled","year":2025,"channels":["api"],"languages":[],"metrics":[{"kpi":"fraud-loss-reduction","value":30,"unit":"percent","qualifier":"at-least","period":"early users of intelligent 3DS interventions, eligible transactions","claimant":"organization","quote":"Backed by a new multihead model and decisioning layer, early users have seen an over 30% reduction in fraud on eligible transactions, representing one of the largest ever improvements to Radar.","sourceUrl":"https://stripe.com/blog/using-ai-optimize-payments-performance-payments-intelligence-suite"}],"outcomeDisclosed":true,"sources":[{"url":"https://stripe.com/blog/using-ai-optimize-payments-performance-payments-intelligence-suite","title":"Using AI to optimize payments performance with the Payments Intelligence Suite","publisher":"Stripe","date":"2025-05-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"stripe-radar-payments-foundation-model"},{"title":"SMBC: licensing OakNorth's credit intelligence software for lending and monitoring","useCases":["sme-cash-flow-underwriting","credit-early-warning-monitoring"],"organization":{"name":"Sumitomo Mitsui Banking Corporation","anonymized":false,"country":"JP","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"OakNorth","role":"platform"}],"summary":"In November 2020 Sumitomo Mitsui Banking Corporation licensed the credit underwriting and monitoring software built by OakNorth, a UK SME lender, and invested USD 30 million in OakNorth equity. The software pulls in public and alternative data and compares each borrower with sector and local peers, so lenders can underwrite businesses and watch them continuously rather than waiting for periodic audited financials. SMBC's group CFO said the alliance would bring more sophistication to its corporate lending platforms and that the group was harnessing AI through big data and machine learning across its strategic markets in Southeast Asia, such as Indonesia. No outcome figures were published.","stage":"announced","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.euromoney.com/article/27sic7y97uvu96j2fuc5c/fintech/smbc-uses-oaknorths-credit-intelligence-software-to-grow-lending/","title":"SMBC uses OakNorth's credit intelligence software to grow lending","publisher":"Euromoney","date":"2020-11-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"sumitomo-mitsui-banking-corporation-oaknorth-credit-intelligence"},{"title":"Sun & Ski Sports: AI agent Sunny for order status, returns and product advice","useCases":["order-status-and-returns-agent","conversational-shopping-assistant"],"organization":{"name":"Sun & Ski Sports","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Sierra","role":"platform"}],"summary":"Sun & Ski Sports, a Texas based outdoor retailer with a strongly seasonal business, started its AI agent Sunny on basic returns and order status questions and then extended it to expert product advice on skis, boards, boots and bindings on its product pages. Sierra, the vendor, reports higher satisfaction on conversations the agent handles than on those transferred to humans, higher conversion for shoppers who engage with it, and a winter season without hiring temporary service staff.","stage":"production","year":2025,"channels":["web-chat"],"languages":["en"],"metrics":[{"kpi":"customer-satisfaction","value":90,"unit":"percent","qualifier":"exact","period":"conversations handled by the agent, as reported in October 2025","baseline":"68% for conversations transferred to human agents","claimant":"vendor","quote":"Sunny achieves 90% customer satisfaction compared to 68% for conversations transferred to human agents.","sourceUrl":"https://sierra.ai/customers/sun-and-ski-sports"},{"kpi":"customer-satisfaction-uplift","value":50,"unit":"percent","qualifier":"exact","period":"as reported in October 2025, three years after the CMO joined in 2022","claimant":"vendor","quote":"Three years later, Sunny, their AI agent, has improved CSAT by 50% and tripled product page conversion rates","sourceUrl":"https://sierra.ai/customers/sun-and-ski-sports"},{"kpi":"conversion-rate-uplift","value":3,"unit":"multiplier","qualifier":"exact","baseline":"shoppers who do not engage with the agent","claimant":"vendor","quote":"Customers who engage with Sunny convert at triple the rate of those who don't.","sourceUrl":"https://sierra.ai/customers/sun-and-ski-sports"}],"outcomeDisclosed":true,"sources":[{"url":"https://sierra.ai/customers/sun-and-ski-sports","title":"How Sun & Ski's AI agent \"Sunny\" turns approachability into sales","publisher":"Sierra","date":"2025-10-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"sun-and-ski-sports-sunny-ai-agent"},{"title":"Sun Life: AI in individual insurance underwriting, contact centre chat and advisor tools","useCases":["insurance-policy-servicing-agent","insurance-broker-and-agent-assistant"],"organization":{"name":"Sun Life","anonymized":false,"country":"CA","region":"north-america","industry":"insurance"},"vendors":[{"name":"Sun Life","role":"in-house"}],"summary":"Sun Life's 2025 annual report says AI tools cut median response times for individual insurance applications in a target segment by close to half, including a 50% increase in straight through underwriting. In its Client Contact Centre, generative AI raised chatbot containment by 16 percentage points year over year, and advisors use a generative AI Notes Assistant Tool; in Hong Kong it launched Advisory Buddy, a GenAI chatbot inside its Advisor Workbench. In Malaysia almost two thirds of clients received automated underwriting decisions within two hours.","stage":"production","year":2025,"channels":["internal-tools","web-chat"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":50,"unit":"percent","qualifier":"approximately","period":"median response time for individual insurance applications, target segment, 2025","claimant":"organization","quote":"Leveraged AI tools to do more for our Clients including reduced median response times for individual insurance applications for a target segment by close to half (including increasing straight-through underwriting by 50%), increased chatbot containment rate (up 16 percentage points year-over-year) with generative AI tools in the Client Contact Centre, and enhanced advisor productivity using a generative AI-powered Notes Assistant Tool.","sourceUrl":"https://www.sec.gov/Archives/edgar/data/1097362/000109736226000010/a2025q4slfmdalive.htm"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1097362/000109736226000010/a2025q4slfmdalive.htm","title":"Sun Life Financial Inc. Management's Discussion and Analysis for 2025 (Form 40-F, Exhibit 99.1)","publisher":"Sun Life via SEC EDGAR","date":"2026-02-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"sun-life-ai-underwriting-and-client-service"},{"title":"Supremo Tribunal Federal (Brazil): Maria generative AI platform for case reports and headnotes","useCases":["court-and-case-file-summarization"],"organization":{"name":"Supremo Tribunal Federal","anonymized":false,"country":"BR","region":"latin-america","industry":"government"},"vendors":[],"summary":"Brazil's Federal Supreme Court runs Maria, a generative AI platform inside its STF Digital environment that supports court staff in analysing cases and producing documents. It started by generating headnotes (ementas) in the National Council of Justice's standard format, case reports (relatórios) for extraordinary appeals and questionnaires for initial petitions in constitutional complaints (reclamações), and has been extended to more case classes, grammatical review and a unified search of related precedents. The court stresses that Maria never acts autonomously: staff review, adapt and decide whether to use each report or headnote. It builds on earlier tools, Victor (2018) for triaging extraordinary appeals and VitorIA (2023) for grouping similar cases. The court is deploying open source language models on servers that will soon be available in its own data center. No outcome figures are given in the article.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["pt"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://noticias.stf.jus.br/postsnoticias/stf-amplia-uso-de-inteligencia-artificial-em-apoio-a-atividade-jurisdicional/","title":"STF amplia uso de inteligência artificial em apoio à atividade jurisdicional","publisher":"Supremo Tribunal Federal","date":"2025-09-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"supremo-tribunal-federal-maria-case-reports"},{"title":"SUSE: multilingual AI SDR agent for inbound discovery and meeting booking","useCases":["inbound-lead-qualification-agent"],"organization":{"name":"SUSE","anonymized":false,"region":"europe","industry":"technology"},"vendors":[{"name":"Qualified","role":"platform"}],"summary":"SUSE, the open source software company, moved to ungated content and lost the form data that told it who was on its website. It deployed an AI SDR agent that engages every visitor, asks discovery questions adapted to developers, architects or CIOs, recognises returning accounts through Salesforce and intent data, answers in the visitor's language and follows up by email. The vendor reports that the agent turns 70% of qualified conversations into booked meetings and that its email follow up influenced more than USD 10 million in pipeline.","stage":"production","year":2024,"channels":["web-chat","email"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.qualified.com/customers/suse","title":"SUSE turns 70% of qualified conversations into meetings with Piper the AI SDR Agent","publisher":"Qualified"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"suse-ai-sdr-inbound-qualification"},{"title":"SVT (Sveriges Television): automated transcription and closed captions for regional news","useCases":["audio-and-video-transcription-and-captioning"],"organization":{"name":"Sveriges Television (SVT)","anonymized":false,"country":"SE","region":"europe","industry":"media-and-entertainment"},"vendors":[{"name":"Microsoft (Azure AI Speech, Speech Studio)","role":"platform"}],"summary":"SVT, Sweden's public broadcaster, transcribes its video content and generates closed captions automatically with Azure speech services, in production since 2021. The broadcaster says it could not caption its local news otherwise, because it publishes for 21 regional stations at the same time several times a day, and that feedback from viewers with hearing loss is mostly positive.","stage":"scaled","year":2021,"channels":["api"],"languages":["sv"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1533261355300735842-sveriges-television-ab-media-entertainment-azure","title":"Swedish Television improves public news accessibility with AI","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"sveriges-television-automated-closed-captions"},{"title":"Swarovski: Génie generative AI portal speeds up campaign localization","useCases":["marketing-and-product-content-localization","personalized-marketing-at-scale"],"organization":{"name":"Swarovski","anonymized":false,"country":"AT","region":"europe","industry":"retail-and-ecommerce"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Swarovski, which sells in more than 140 markets, launched Génie in 2023, a generative AI portal on Vertex AI and Gemini, on top of a BigQuery data foundation consolidated with the partner CloudSufi. More than 1,000 employees use it for tasks including content translation into over 20 languages, creative asset generation and testing visuals and descriptions for different regions. Google Cloud reports that campaign localization became 10 times faster through AI assisted translation and asset adaptation, and that Génie's AI personalized email campaigns see 17% higher open rates and 7% higher click through rates. Every AI application is evaluated against an internal ethics and risk model.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"processing-time-reduction","value":10,"unit":"multiplier","qualifier":"exact","period":"speed of campaign localization","claimant":"vendor","quote":"Campaign localization is 10x faster, thanks to AI-assisted translation and asset adaptation","sourceUrl":"https://cloud.google.com/customers/swarovski"},{"kpi":"users-served","value":1000,"unit":"count","qualifier":"at-least","period":"employees using the Génie portal","claimant":"vendor","quote":"Over 1,000 employees now utilize Génie for tasks such as contract review, content translation into over 20 languages, creative digital asset generation, and campaign inspirations, and product design cost estimation.","sourceUrl":"https://cloud.google.com/customers/swarovski"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/swarovski","title":"Letting data shine bright: How Swarovski personalizes luxury with Google Cloud","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"swarovski-genie-campaign-localization"},{"title":"SWISS and Lufthansa Group: Operations Decision Support Suite for aircraft rotation and passenger recovery","useCases":["airline-operations-control-decision-support"],"organization":{"name":"Swiss International Air Lines","anonymized":false,"country":"CH","region":"europe","industry":"travel-and-hospitality"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"SWISS, part of Lufthansa Group, developed the Operations Decision Support Suite (OPSD) with Google Cloud as a layer above its operational systems. It replicates the state of the operation minute by minute with crew, passenger, rotation and technical data in one place, and uses optimization and machine learning to propose scenarios, such as swapping aircraft between flights or managing missed connections, that operations controllers approve before they take effect. The first use cases were rotation planning and passenger management. Lufthansa Group says OPSD lays the foundation for optimizing across the different operational dimensions, and eventually across airline borders, with the goal of rolling it out to other Lufthansa Group airlines.","stage":"production","year":2022,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"cost-savings","value":1000000,"unit":"currency","currency":"CHF","qualifier":"at-least","period":"first three and a half months of the rotation optimization feature","claimant":"organization","quote":"In three and a half months, this has generated more than a million Swiss Francs in savings.","sourceUrl":"https://cloud.google.com/customers/swiss"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/customers/swiss","title":"SWISS: Making air travel more sustainable and improving operational quality with Google Cloud","publisher":"Google Cloud"},{"url":"https://innovation-runway.lufthansagroup.com/en/focus-areas-projects/projects/ops-suite.html","title":"Operations Decision Support Suite (OPSD)","publisher":"Lufthansa Group Innovation Runway"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"swiss-lufthansa-opsd-operations-decision-support"},{"title":"T-Mobile: PromoGenius app and product agent for retail and care staff","useCases":["retail-store-and-kiosk-assistant","plan-upgrade-and-sales-assistant"],"organization":{"name":"T-Mobile","anonymized":false,"country":"US","region":"north-america","industry":"telecommunications"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"T-Mobile built PromoGenius on Power Apps to give retail and call centre representatives one place for current promotions, discounts and trade in values, used on iPads on the shop floor. An agent built in Copilot Studio reads more than 20 device makers' websites, answers technical questions in natural language during a customer conversation and builds comparison tables that can be shown to the customer. Microsoft reports over 83,000 unique users and 500,000 launches a month for the app, which supports all T-Mobile retail stores and call centres.","stage":"scaled","year":2025,"channels":["internal-tools","kiosk"],"languages":["en"],"metrics":[{"kpi":"users-served","value":83000,"unit":"count","qualifier":"at-least","period":"unique users, retail and call centre staff","claimant":"vendor","quote":"The app, called PromoGenius, is the second most popular app at T-Mobile, supporting all T-Mobile retail outlets and call centers, with over 83,000 unique users and 500,000 launches a month.","sourceUrl":"https://www.microsoft.com/en/customers/story/23087-t-mobile-usa-microsoft-copilot-studio"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23087-t-mobile-usa-microsoft-copilot-studio","title":"T-Mobile drives more effective customer conversations with Microsoft Power Apps and Copilot Studio","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"t-mobile-promogenius-retail-agent"},{"title":"Takeda: AI expense audit across 63 countries","useCases":["travel-and-expense-audit-agent"],"organization":{"name":"Takeda","anonymized":false,"country":"JP","region":"asia-pacific","industry":"pharma-and-life-sciences"},"vendors":[{"name":"AppZen","role":"platform"}],"summary":"Takeda deployed AppZen's Expense Audit solution to review 100% of employee expense reports across its global operations, replacing a manual process that, at times, flagged as much as 70 to 100% of expenses for audit based on trigger rules yet still left the company vulnerable to duplicates and employee spend leakage. The AI models include translation capability to handle the language differences across Takeda's European and Asian operations, and auditors now focus their attention on high risk expenses.","stage":"scaled","year":2022,"channels":[],"languages":[],"metrics":[{"kpi":"automation-rate","value":63,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"Takeda achieved 63% auto-approval across 63 countries, processing 400K expense audits annually while saving 4,000 auditor hours quarterly with AppZen.","sourceUrl":"https://www.appzen.com/resources/case-studies/how-takeda-is-transforming-global-expense-auditing-with-ai"},{"kpi":"interactions-handled","value":400000,"unit":"count","qualifier":"approximately","period":"per year","claimant":"vendor","quote":"Takeda achieved 63% auto-approval across 63 countries, processing 400K expense audits annually while saving 4,000 auditor hours quarterly with AppZen.","sourceUrl":"https://www.appzen.com/resources/case-studies/how-takeda-is-transforming-global-expense-auditing-with-ai"},{"kpi":"hours-saved","value":4000,"unit":"hours","qualifier":"approximately","period":"per quarter","claimant":"vendor","quote":"Takeda achieved 63% auto-approval across 63 countries, processing 400K expense audits annually while saving 4,000 auditor hours quarterly with AppZen.","sourceUrl":"https://www.appzen.com/resources/case-studies/how-takeda-is-transforming-global-expense-auditing-with-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.appzen.com/resources/case-studies/how-takeda-is-transforming-global-expense-auditing-with-ai","title":"Takeda Saves 4,000 Auditor Hours Quarterly","publisher":"AppZen","archivedUrl":"https://web.archive.org/web/20220702200424/https://www.appzen.com/resources/case-studies/how-takeda-is-transforming-global-expense-auditing-with-ai?hsLang=en"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"takeda-expense-audit-automation"},{"title":"TD Bank: AI powered observability for proactive incident detection","useCases":["aiops-incident-triage"],"organization":{"name":"TD Bank","anonymized":false,"country":"CA","region":"north-america","industry":"banking"},"vendors":[{"name":"Dynatrace","role":"platform"}],"summary":"TD Bank is consolidating roughly ten monitoring tools across its clouds into one observability platform whose AI identifies the root cause of emerging issues across its hybrid, multicloud estate, so teams can respond to transaction failures before users are affected. The customer story does not describe the AI as generative, and the outcome figures come from the vendor.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"detection-rate-improvement","value":25,"unit":"percent","qualifier":"exact","baseline":"incidents identified proactively, before the platform","claimant":"vendor","quote":"As a result, TD Bank is identifying 25% more incidents proactively and responding to them 20% faster.","sourceUrl":"https://www.dynatrace.com/customers/td-bank/"},{"kpi":"mttr-reduction","value":20,"unit":"percent","qualifier":"exact","baseline":"time to respond to and resolve IT incidents before the platform, vendor phrasing \"responding to them 20% faster\"","claimant":"vendor","quote":"As a result, TD Bank is identifying 25% more incidents proactively and responding to them 20% faster.","sourceUrl":"https://www.dynatrace.com/customers/td-bank/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.dynatrace.com/customers/td-bank/","title":"TD Bank customer story","publisher":"Dynatrace","date":"2024-04-29","archivedUrl":"https://web.archive.org/web/20240523164055/https://www.dynatrace.com/customers/td-bank/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"td-bank-aiops-observability"},{"title":"Tekmetric: AI merchant underwriting and monitoring for Tekmetric Payments","useCases":["merchant-underwriting-and-risk-monitoring"],"organization":{"name":"Tekmetric","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Coris","role":"platform"},{"name":"Stripe","role":"platform"}],"summary":"Tekmetric, a cloud based auto repair shop management platform, offers Tekmetric Payments on Stripe Connect. It replaced underwriting and monitoring based on static credit reports, spreadsheets and manual emails with the Coris risk platform: every active merchant is scored daily, each alert gets an AI generated recommendation and summary, and onboarding approvals and dispute reminders are sent to merchants automatically. Coris reports 70% fewer manual reviews and faster onboarding.","stage":"production","year":2025,"channels":["internal-tools","email"],"languages":["en"],"metrics":[{"kpi":"alert-volume-reduction","value":70,"unit":"percent","qualifier":"exact","period":"manual merchant risk reviews after go live","claimant":"vendor","quote":"Discover how Tekmetric reduced manual risk reviews by 70% and accelerated onboarding with Coris’ AI-driven workflows and real-time monitoring.","sourceUrl":"https://www.coris.ai/customer/how-tekmetric-automated-merchant-risk-and-cut-manual-reviews-by-70"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.coris.ai/customer/how-tekmetric-automated-merchant-risk-and-cut-manual-reviews-by-70","title":"How Tekmetric Automated Merchant Risk and Cut Manual Reviews by 70%","publisher":"Coris","archivedUrl":"https://web.archive.org/web/20251014133241/https://www.coris.ai/customer/how-tekmetric-automated-merchant-risk-and-cut-manual-reviews-by-70"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"tekmetric-coris-merchant-risk-automation"},{"title":"Telefónica: AI and machine learning to steer radio power saving features","useCases":["ran-energy-optimization"],"organization":{"name":"Telefónica","anonymized":false,"country":"ES","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Ericsson","role":"platform"}],"summary":"Telefónica has activated power saving features in its mobile networks for more than a decade. The early ones for 2G and 3G used static parameters; the current ones for 4G and 5G use AI and machine learning to predict traffic, set thresholds, shut down cells in low traffic hours and check quality. The platforms were first tested in O2 Germany in 2021. Telefónica Spain was the first operator to test Ericsson's Radio Deep Sleep Mode at a 5G site in Madrid, where the company reports savings of up to 8% of the site's 24 hour consumption and up to 26% in low traffic hours.","stage":"production","year":2022,"channels":["api"],"languages":[],"metrics":[{"kpi":"energy-savings","value":8,"unit":"percent","qualifier":"up-to","period":"total 24 hour consumption of one 5G test site in Madrid, Radio Deep Sleep Mode","claimant":"organization","quote":"Supported by Artificial Intelligence and Machine Learning algorithms, the company achieved savings of up to 8%, considering the site’s total 24-hour consumption, and up to 26% in low traffic hours.","sourceUrl":"https://www.telefonica.com/en/communication-room/press-room/telefonica-drives-energy-consumption-optimisation-through-solutions-based-on-artificial-intelligence-and-machine-learning/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.telefonica.com/en/communication-room/press-room/telefonica-drives-energy-consumption-optimisation-through-solutions-based-on-artificial-intelligence-and-machine-learning/","title":"Telefónica drives energy consumption optimisation through solutions based on Artificial Intelligence and Machine Learning","publisher":"Telefónica","date":"2022-03-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"telefonica-ai-radio-power-saving-features"},{"title":"Telefónica España: big data and AI for network anomaly detection and optimization","useCases":["network-planning-and-capacity-optimization","predictive-network-maintenance"],"organization":{"name":"Telefónica España","anonymized":false,"country":"ES","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Telefónica España built a network data platform on Microsoft Azure (Azure Data Explorer, Azure Databricks and Power BI) to store and analyse the large volumes of data its 4G and 5G mobile network produces. The team uses it for anomaly detection, to address issues before they affect customers, and for automated network optimization. Telefónica says the project is live with several use cases deployed and that results have been very positive; Microsoft's summary adds substantial savings in operating costs. No figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/21150-telefonica-group-spain-azure-ai-and-machine-learning","title":"Telefónica España's transformation with Microsoft Azure: Enhancing network performance through big data and AI","publisher":"Microsoft","date":"2025-02-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"telefonica-espana-network-analytics-optimization"},{"title":"Telefónica: GenIA assistant in a device marketplace for businesses and public bodies","useCases":["business-connectivity-quoting-and-service-assistant"],"organization":{"name":"Telefónica España","anonymized":false,"country":"ES","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Telefónica","role":"in-house"}],"summary":"In June 2026 Telefónica España launched a marketplace for companies and government agencies to select and buy workplace devices from a catalogue of 160 models. Its GenIA virtual assistant lets buyers describe needs in natural language, compare alternatives, get personalised recommendations and resolve technical questions in real time, alongside a productivity calculator. No outcome figures were published.","stage":"production","year":2026,"channels":["web-chat"],"languages":["es"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.telefonica.com/en/communication-room/press-room/telefonica-is-using-ai-to-optimize-equipment-selection-for-businesses-and-government-agencies/","title":"Telefónica is using AI to optimize equipment selection for businesses and government agencies","publisher":"Telefónica","date":"2026-06-17"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"telefonica-genia-b2b-device-marketplace"},{"title":"Telenet: AI decisioning for next best action, churn and upgrades","useCases":["churn-prediction-and-retention-offers","plan-upgrade-and-sales-assistant"],"organization":{"name":"Telenet","anonymized":false,"country":"BE","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Pega","role":"platform"}],"summary":"Telenet, a provider of connectivity and entertainment services in Belgium, uses Pega's AI based Customer Decision Hub as a single decisioning system that responds to customer signals in real time, anticipates how behaviour may change and proposes the next best action, such as personalised upgrades and solutions, with the stated goals of reducing churn and raising offer acceptance. Pega reports a 20% reduction in churn, a 75% increase in offer acceptance and a 33% increase in cross sell. These figures cover the whole decisioning programme, including upgrades and cross sell, not retention alone.","stage":"scaled","year":2023,"channels":[],"languages":[],"metrics":[{"kpi":"churn-reduction","value":20,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"20% reduction in churn","sourceUrl":"https://www.pega.com/customers/telenet-customer-decision-hub"},{"kpi":"conversion-rate-uplift","value":75,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"75% increase in offer acceptance","sourceUrl":"https://www.pega.com/customers/telenet-customer-decision-hub"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.pega.com/customers/telenet-customer-decision-hub","title":"Anticipating customer needs with AI-powered decisioning","publisher":"Pega"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"telenet-next-best-action-decisioning"},{"title":"Telkomsel: Veronika virtual assistant on Azure OpenAI","useCases":["first-line-contact-centre-agent"],"organization":{"name":"Telkomsel","anonymized":false,"country":"ID","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Telkomsel, with more than 159 million mobile subscribers, rebuilt its Veronika virtual assistant on Azure OpenAI to handle routine customer questions in Bahasa Indonesia with the local accents and expressions its first chatbot could not follow, so that human agents could focus on complex issues. Telkomsel's chief information officer says the architecture can handle up to 5 million transactions a month. Telkomsel's chief marketing officer says customer self service interactions rose from 19% to 45% after the launch.","stage":"scaled","year":2024,"channels":[],"languages":["id"],"metrics":[{"kpi":"automation-rate","value":45,"unit":"percent","qualifier":"exact","period":"share of customer interactions that were self service, as Telkomsel describes it","baseline":"19% before the generative AI Veronika","claimant":"organization","quote":"“I’m thrilled to share that since introducing Veronika, we’ve seen a leap in customer self-service interactions from 19 percent to 45 percent,” enthuses Heng.","sourceUrl":"https://customers.microsoft.com/en-us/story/1739790055278570755-telkomsel-azure-openai-service-telecommunications-en-indonesia"}],"outcomeDisclosed":true,"sources":[{"url":"https://customers.microsoft.com/en-us/story/1739790055278570755-telkomsel-azure-openai-service-telecommunications-en-indonesia","title":"Telkomsel customer support teams speak volumes, with help from copilot based on Azure OpenAI Service","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"telkomsel-veronika-virtual-assistant"},{"title":"Telstra and CommBank: Scam Indicator and Fraud Indicator built on mobile network intelligence","useCases":["telecom-fraud-detection","application-and-identity-fraud-detection"],"organization":{"name":"Telstra","anonymized":false,"country":"AU","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Quantium Telstra","role":"in-house"}],"summary":"Quantium Telstra built two services in collaboration with Commonwealth Bank. Scam Indicator detects and intercepts suspected scam calls to bank customers in real time and was later extended to landlines. Fraud Indicator, live from early 2025, securely shares intelligence about unusual mobile service usage so the bank can spot fraudsters opening accounts with a phone number they control. Telstra describes the Scam and Fraud Indicator as using AI and says it has safeguarded thousands of customers and prevented millions of dollars in fraud since 2023; the expected gain in detection of fraudulent accounts was published as a forecast.","stage":"production","year":2025,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.telstra.com.au/exchange/telstra-and-commbank-expand-collaboration-to-increase-fraud-dete","title":"Telstra and CommBank expand collaboration to increase fraud detection rates","publisher":"Telstra","date":"2025-02-10"},{"url":"https://www.telstra.com.au/exchange/telstra-s-ai-transformation--strategy--partnerships-and-real-wor","title":"Telstra's AI transformation: strategy, partnerships and real-world results","publisher":"Telstra","date":"2026-04-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"telstra-quantium-fraud-indicator"},{"title":"Telstra: network level blocking of scam, spoofed and Wangiri calls","useCases":["spam-and-scam-call-blocking","telecom-fraud-detection"],"organization":{"name":"Telstra","anonymized":false,"country":"AU","region":"asia-pacific","industry":"telecommunications"},"vendors":[],"summary":"As part of its Cleaner Pipes initiative, Telstra blocks suspected scam calls in its network before they reach customers. Upgrades in 2021 made blocking more aggressive, improved detection of Wangiri one ring calls from international premium numbers and of spoofed calls that pretend to come from local numbers or trusted brands, and doubled the monthly volume blocked within four months. Telstra says it keeps evolving its algorithms and detection methods and takes care not to block genuine calls. From December 2024 it added Telstra Scam Protect, an in house network feature that warns customers on screen about calls that look spoofed, arrive from overseas while showing a local number, or come from a number with a suspicious calling pattern. Its Scam Protect article (published March 2025, updated May 2026) reports blocking more than 11 million scam calls a month on average and Scam Protect warnings on an average of 12 million calls a month.","stage":"scaled","year":2021,"channels":["voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.telstra.com.au/exchange/were-now-blocking-over-13-million-scam-calls-a-month","title":"We're now blocking over 13 million scam calls a month","publisher":"Telstra","date":"2021-06-15"},{"url":"https://www.telstra.com.au/exchange/suspicious-phone-calls--what-telstra-is-doing-to-raise-the-alarm","title":"Suspicious phone calls: what Telstra is doing to raise the alarm","publisher":"Telstra","date":"2025-03-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"telstra-scam-call-blocking"},{"title":"Telstra: agentic AI proof of concept for self healing telco cloud operations","useCases":["autonomous-network-operations","network-fault-triage-copilot"],"organization":{"name":"Telstra","anonymized":false,"country":"AU","region":"asia-pacific","industry":"telecommunications"},"vendors":[{"name":"Red Hat","role":"platform"},{"name":"Dell Technologies","role":"platform"},{"name":"Cisco","role":"platform"}],"summary":"In a proof of concept in a live telco cloud environment, Telstra showed an agentic AI capability that detected an unplanned infrastructure outage and resolved it autonomously by moving critical network applications to healthy hardware in minutes rather than hours. Using the Model Context Protocol with retrieval augmented generation, AI agents connect data from multiple vendor platforms into one view and give teams context aware recommendations to speed up fault resolution. Telstra presents it as groundwork for self healing, self optimising operations, not yet as a production service.","stage":"pilot","year":2026,"channels":["internal-tools","api"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.telstra.com.au/exchange/telstra-advanced-autonomous-networks-ambition-through-breakthrou","title":"Telstra advanced autonomous networks ambition through breakthrough collaboration with Red Hat, Dell Technologies and Cisco","publisher":"Telstra","date":"2026-03-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"telstra-self-healing-network-proof-of-concept"},{"title":"Telstra: SIM swap and port out risk ratings for banks","useCases":["telecom-fraud-detection"],"organization":{"name":"Telstra","anonymized":false,"country":"AU","region":"asia-pacific","industry":"telecommunications"},"vendors":[],"summary":"In January 2022 Telstra said it had started working with organisations in the banking industry and would provide a risk rating, a number on a risk scale, when a banking organisation asks whether a mobile service used as a form of identity has had a recent SIM swap or port out. Banks and credit unions would request it when a customer makes a transfer, especially to a new recipient, and use it to ask for more information rather than to block the customer automatically. Telstra said it was considering applying the technology in retail, insurance, transport and logistics, social networking and online gaming. No results are published, and no later Telstra source cited here confirms how widely the rating is used today.","stage":"announced","year":2022,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.telstra.com.au/exchange/what-are-sim-swaps-and-porting-fraud--and-how-are-we-working-to-","title":"What are SIM swaps and porting fraud, and how are we working to stop it?","publisher":"Telstra","date":"2022-01-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"telstra-sim-swap-and-porting-risk-signal"},{"title":"Telstra: SmartFix proactive fixes in network operations","useCases":["predictive-network-maintenance","autonomous-network-operations"],"organization":{"name":"Telstra","anonymized":false,"country":"AU","region":"asia-pacific","industry":"telecommunications"},"vendors":[],"summary":"Telstra's SmartFix system is embedded in its network operations and automatically fixes many issues before customers notice a problem. Telstra reports the number of proactive actions it performed in FY25 and says they prevented nearly 1 million support calls. The blog post is part of Telstra's description of its wider AI program and gives no detail on the models or the types of fixes.","stage":"scaled","year":2025,"channels":[],"languages":[],"metrics":[{"kpi":"interactions-handled","value":2500000,"unit":"count","qualifier":"exact","period":"FY25, proactive actions","claimant":"organization","quote":"It automatically fixes many issues before customers notice a problem – in FY25 it performed 2.5 million proactive actions, preventing nearly 1 million support calls by resolving issues in advance.","sourceUrl":"https://www.telstra.com.au/exchange/telstra-s-ai-transformation--strategy--partnerships-and-real-wor"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.telstra.com.au/exchange/telstra-s-ai-transformation--strategy--partnerships-and-real-wor","title":"Telstra's AI transformation: strategy, partnerships and real-world results","publisher":"Telstra","date":"2026-04-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"telstra-smartfix-proactive-network-fixes"},{"title":"Texas Education Agency: hybrid automated scoring of STAAR written responses","useCases":["automated-scoring-of-written-responses"],"organization":{"name":"Texas Education Agency","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Texas Education Agency scores the short and extended constructed responses on the English language STAAR tests with a hybrid model: an automated scoring engine gives every response its first score, and at least 25 percent of responses per grade and subject are routed to trained human raters to monitor the engine. Responses with condition codes (for example blank, off topic, another language or vocabulary unlike the training data) or low confidence also go to humans, and a human score is always the score of record. The engine must agree with human raters as often as humans agree with each other before use; STAAR Spanish responses are scored only by humans. TEA says it used the hybrid approach to score all constructed responses in the December 2023 administration, after a study that rescored spring 2023 responses with the engine. TEA calls it an automated scoring engine and says it differs from AI that teaches itself: it is programmed on about 3,000 human scored field test responses per item.","stage":"scaled","year":2023,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://tea.texas.gov/data-reports/staar/scoring-process-staar-constructed-response-1.pdf","title":"Scoring Process for STAAR Constructed Responses","publisher":"Texas Education Agency, Student Assessment Division"},{"url":"https://tea.texas.gov/student-assessment/testing/hybrid-scoring-key-questions.pdf","title":"Hybrid Scoring Key Questions","publisher":"Texas Education Agency","archivedUrl":"https://web.archive.org/web/20240512091507/https://tea.texas.gov/student-assessment/testing/hybrid-scoring-key-questions.pdf"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"texas-education-agency-staar-automated-scoring"},{"title":"Textron Aviation: TAMI, a generative AI assistant for aircraft maintenance technicians","useCases":["plant-operator-and-maintenance-copilot"],"organization":{"name":"Textron Aviation","anonymized":false,"country":"US","region":"north-america","industry":"manufacturing"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Textron Aviation, maker of Cessna and Beechcraft aircraft, built TAMI (Textron Aviation Maintenance Intelligence) on Azure OpenAI Service so that technicians in its service centres can query more than 60,000 pages of maintenance documentation for over 50 aircraft models in natural language, in several languages. Microsoft reports that troubleshooting that took up to 20 minutes now takes one to two minutes. A CIO.com profile of the company's CIO describes a pay as you go proof of concept in which senior mechanics' test questions were answered correctly 19 times out of 20, followed by a rollout to more than 1,500 mechanics across global service centres.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"time-saved-per-task","value":18,"unit":"minutes","qualifier":"up-to","baseline":"Up to 20 minutes per troubleshooting search before TAMI; one to two minutes after","claimant":"vendor","quote":"Troubleshooting that previously took up to 20 minutes can now be accomplished in one to two minutes.","sourceUrl":"https://www.microsoft.com/en/customers/story/23024-textron-aviation-azure-open-ai-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23024-textron-aviation-azure-open-ai-service","title":"Textron Aviation enhances maintenance efficiency with Azure AI","publisher":"Microsoft"},{"url":"https://www.cio.com/article/4025048/textron-takes-flight-with-gen-ai.html","title":"Textron takes flight with gen AI","publisher":"CIO.com","date":"2025-07-21"},{"url":"https://www.metisstrategy.com/textron-takes-flight-with-gen-ai/","title":"Textron takes flight with Gen AI","publisher":"Metis Strategy","date":"2025-07-21"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"textron-aviation-tami-maintenance-assistant"},{"title":"TIME: the TIME AI Agent answers readers from more than a century of TIME reporting","useCases":["publisher-archive-answer-engine"],"organization":{"name":"TIME","anonymized":false,"country":"US","region":"north-america","industry":"media-and-entertainment"},"vendors":[{"name":"Scale AI","role":"integrator"}],"summary":"TIME first tested a conversational AI toolbar on its Person of the Year coverage in December 2024, and chose not to let that version use content from outside TIME. In November 2025 it launched the TIME AI Agent, built with Scale AI, which lets readers ask questions, get summaries in text or audio, translate into 13 languages, and search more than a century of TIME reporting with semantic and hybrid search. TIME says attribution and citation are preserved in every interaction and that the system was red team tested. Digital Content Next reports that the agent is now trained on more than three quarters of a million pieces of TIME content, some digitized from PDFs of old magazines, that new articles are indexed in near real time, and that users who engage with the agent are more likely to return and spend more time on the site than those who do not.","stage":"scaled","year":2025,"channels":["web-chat"],"languages":["en","fr","es","de","it","pt","ja","ko","zh","hi","he","ar","ru"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://time.com/7332572/the-story-behind-the-time-ai-agent/","title":"The Story Behind the TIME AI Agent","publisher":"TIME","date":"2025-11-10","archivedUrl":"https://web.archive.org/web/20251110183151/https://time.com/7332572/the-story-behind-the-time-ai-agent/"},{"url":"https://digitalcontentnext.org/blog/2026/04/16/inside-times-rollout-of-its-timeai-interactive-agent/","title":"Inside TIME's rollout of its TIMEAI interactive agent","publisher":"Digital Content Next","date":"2026-04-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"time-ai-agent-archive-answers"},{"title":"Together Credit Union: voice agent that answers the phone line first","useCases":["first-line-contact-centre-agent"],"organization":{"name":"Together Credit Union","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Posh","role":"platform"}],"summary":"Together Credit Union runs a voice agent on its inbound phone line that answers member questions, resolves routine requests and hands the rest to a person, drawing on one knowledge base that the branch, contact centre and chat channels also use. The vendor reports that the agent now contains the majority of inbound calls, 12 points more than at the start, that after hours escalations to an outsourced contact centre fell, and that member satisfaction held above 94% during the rollout.","stage":"production","year":2025,"channels":["voice"],"languages":["en"],"metrics":[{"kpi":"containment-rate","value":66,"unit":"percent","qualifier":"exact","period":"total inbound calls, as reported by the vendor","baseline":"12 points lower at the start of the rollout","claimant":"vendor","quote":"Heather, Together CU's Posh-powered voice agent, now contains 66% of total inbound calls","sourceUrl":"https://www.posh.ai/client-stories/client-story-together-credit-union"},{"kpi":"contact-deflection","value":14,"unit":"percent","qualifier":"exact","period":"after hours escalations to the outsourced contact centre","claimant":"vendor","quote":"After-hours escalations to their third-party contact center dropped 14%.","sourceUrl":"https://www.posh.ai/client-stories/client-story-together-credit-union"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.posh.ai/client-stories/client-story-together-credit-union","title":"How Together Credit Union Built One Source of Truth Across Every Channel","publisher":"Posh"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"together-credit-union-voice-agent"},{"title":"Tokio Marine & Nichido Fire: AI review of damage photos, estimates and suspicious claims","useCases":["claims-triage-and-straight-through-processing","claims-fraud-detection","photo-based-damage-assessment"],"organization":{"name":"Tokio Marine & Nichido Fire Insurance","anonymized":false,"country":"JP","region":"asia-pacific","industry":"insurance"},"vendors":[{"name":"Shift Technology","role":"platform"}],"summary":"Tokio Marine & Nichido Fire Insurance uses Shift Technology's claims intake and claims fraud detection solutions, extended with generative AI that extracts data from structured and unstructured sources such as images and documents. The system highlights the points handlers should check for consistency across estimates, damage photos and claim statements, which makes reviews more efficient and more standardized, including during the surge of claims after large disasters, and it helps detect suspicious claims, which tend to rise after such events. No figures were published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["ja"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.shift-technology.com/resources/case-studies/tokiomarine_casestudy","title":"Case Study: Tokyo Marine & Nichido Fire Insurance Co., Ltd.","publisher":"Shift Technology"},{"url":"https://www.shift-technology.com/resources/press/tokio-marine-deploys-shift-technologys-gen-ai-for-claims-fraud-detection","title":"Tokio Marine Deploys Shift Technology’s Gen AI for Claims, Fraud Detection","publisher":"Shift Technology","date":"2025-09-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"tokio-marine-nichido-shift-claims-review"},{"title":"Toyota Motor Europe: generative AI documents a legacy mainframe warranty application","useCases":["legacy-code-modernization"],"organization":{"name":"Toyota Motor Europe","anonymized":false,"country":"BE","region":"europe","industry":"automotive"},"vendors":[{"name":"AWS","role":"platform"},{"name":"Anthropic","role":"model-provider"},{"name":"Deloitte","role":"integrator"}],"summary":"Toyota Motor Europe runs more than 70 custom applications on legacy mainframe and AS400 platforms, several written in a proprietary language with little documentation and a shrinking pool of experts. With Deloitte and the AWS Generative AI Innovation Center it built a proof of concept on Amazon Bedrock that generates technical documentation, business documentation and process flows from the source code of a warranty handling application of over 1.3 million lines. The remaining experts reviewed a sample against the source and confirmed its accuracy. The proof of concept covered 2 of the 10 modules; AWS reports that it has since led to a production rollout.","stage":"production","year":2026,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://aws.amazon.com/blogs/industries/accelerating-mainframe-modernization-how-toyota-motor-europe-tme-uses-amazon-bedrock-to-automate-legacy-code-documentation","title":"Accelerating mainframe modernization: How Toyota Motor Europe (TME) uses Amazon Bedrock to automate legacy code documentation","publisher":"AWS","date":"2026-03-04"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"toyota-motor-europe-legacy-code-documentation"},{"title":"Trace3: Microsoft Copilot for first pass resume assessment and meeting highlights","useCases":["recruitment-screening-and-interview-scheduling","meeting-summarization-and-action-items"],"organization":{"name":"Trace3","anonymized":false,"country":"US","region":"north-america","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"At technology consultancy Trace3, HR managers use Microsoft Copilot for an initial assessment of resumes, so they review submissions faster and respond to applicants within a couple of days instead of the several weeks it could take before. Its practice director for Azure uses it for highlights of Teams meetings and long email chains and for first drafts; colleagues use it for a broad range of tasks. The outcome is described qualitatively, with no measured figure.","stage":"production","year":2024,"channels":["microsoft-teams","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1790119689031635867-trace3-microsoft-365-professional-services-en-united-states","title":"Trace3 expands the realm of clients' possibilities with Windows 11 Pro and Microsoft Copilot","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"trace3-microsoft-copilot-recruiting-and-meetings"},{"title":"Travelers: generative AI voice agent for first notice of loss and straight through claims","useCases":["claims-first-notice-of-loss-agent","claims-triage-and-straight-through-processing"],"organization":{"name":"Travelers","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[],"summary":"In its 2025 annual report to shareholders, Travelers says it launched a natural language generative AI voice agent that takes first notice of loss by phone, used first for auto damage claims and planned to expand to more lines of business and claim interactions. The same letter reports that more than half of all claims are eligible for straight through digital processing, which customers choose about two thirds of the time, and that another 15% of claims are handled with advanced digital tools. No outcome figures for the voice agent itself were published; the company describes early adoption and feedback as positive.","stage":"production","year":2026,"channels":["voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/86312/000110465926040152/tm2611214d1_ars.pdf","title":"The Travelers Companies, Inc. 2025 Annual Report to Shareholders","publisher":"The Travelers Companies, Inc. (via SEC EDGAR)","date":"2026-04-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"travelers-generative-ai-fnol-voice-agent"},{"title":"Travelers: generative AI classification of policy service emails","useCases":["correspondence-triage-and-routing"],"organization":{"name":"Travelers","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Amazon Web Services","role":"platform"},{"name":"Anthropic","role":"model-provider"}],"summary":"Travelers receives millions of emails a year from agents and customers asking for policy service. With AWS it built a classifier that reads each email and its attachments (Amazon Textract turns PDF attachments into text) and assigns one of 13 service categories with a prompted foundation model (Anthropic's Claude) on Amazon Bedrock; the post says the classifier powers an automation system for these requests but does not confirm that it runs in production. Initial testing without prompt engineering gave 68% accuracy; prompt engineering, condensed categories and better instructions raised it to 91%. The ground truth set held over 4,000 labelled emails.","stage":"announced","year":2023,"channels":["email","api"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":91,"unit":"percent","qualifier":"exact","period":"evaluation accuracy after prompt engineering (testing, not production)","baseline":"68% in initial testing without prompt engineering","claimant":"vendor","quote":"After using a variety of techniques with Anthropic’s Claude v2, such as prompt engineering, condensing categories, adjusting document processing process, and improving instructions, accuracy increased to 91%.","sourceUrl":"https://aws.amazon.com/blogs/machine-learning/how-travelers-insurance-classified-emails-with-amazon-bedrock-and-prompt-engineering/"}],"outcomeDisclosed":true,"sources":[{"url":"https://aws.amazon.com/blogs/machine-learning/how-travelers-insurance-classified-emails-with-amazon-bedrock-and-prompt-engineering/","title":"How Travelers Insurance classified emails with Amazon Bedrock and prompt engineering","publisher":"AWS Artificial Intelligence Blog","date":"2025-01-31"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"travelers-policy-service-email-classification"},{"title":"Trip.com: TripGenie AI travel assistant","useCases":["travel-and-hotel-booking-concierge"],"organization":{"name":"Trip.com","anonymized":false,"country":"SG","region":"asia-pacific","industry":"travel-and-hospitality"},"vendors":[],"summary":"TripGenie is the AI travel assistant in the Trip.com app and website. It helps travellers find inspiration, compare hotels and book hotels, flights and attractions, answers pre and post sales service questions, and during the trip offers menu help, live translation and questions about images. After three years of use, Trip.com reports that nearly 60% of TripGenie interactions are booking related, that service questions are about a quarter of interactions, and that TripGenie assisted order volume grew by around 400% year on year.","stage":"scaled","year":2023,"channels":["mobile-app","web-chat"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.prnewswire.com/news-releases/three-years-of-tripgenie-how-travellers-around-the-world-are-using-ai-differently-302713190.html","title":"Three Years of TripGenie: How Travellers Around the World are Using AI Differently","publisher":"Trip.com via PR Newswire","date":"2026-03-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"trip-com-tripgenie-ai-travel-assistant"},{"title":"US Transportation Security Administration: AI summaries and recommended replies for AskTSA agents","useCases":["email-and-ticket-reply-drafting"],"organization":{"name":"Transportation Security Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Sprinklr","role":"platform"}],"summary":"AskTSA answers traveller questions in writing through X, Facebook Messenger, Apple Messages and text message. Since December 2024 TSA has used generative AI in the AskTSA customer service process to summarize why a traveller is getting in touch, recommend replies tailored to that summary, categorize incoming inquiries and report on where the virtual assistant falls short. The recommended replies assist the human agents; the entry does not describe automatic sending. The department classifies it as not high impact; no outcome figures are published.","stage":"production","year":2024,"channels":["social-messaging","sms","agent-desktop"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DHS-2604, AskTSA)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://www.tsa.gov/contact/customer-service","title":"Customer Service","publisher":"Transportation Security Administration"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"tsa-asktsa-response-assist"},{"title":"Turing: AI drafted replies to HR support tickets","useCases":["hr-and-policy-assistant","email-and-ticket-reply-drafting"],"organization":{"name":"Turing","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Google","role":"platform"}],"summary":"Turing (listed by Google Cloud as Turing Enterprises), an AI company headquartered in San Francisco, built a custom AI model trained on its internal knowledge that drafts replies to HR support tickets. Google Cloud reports a one third cut in ticket processing time after two days of development. A plan to automate 60% of its 52,000 annual HR tickets with Gemini Gems is a target, not a result.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":33,"unit":"percent","qualifier":"exact","claimant":"vendor","quote":"Turing also built a custom AI model trained on internal knowledge to draft replies to HR support tickets, reducing ticket processing time by 33% after two days of development.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"},{"url":"https://www.turing.com/company","title":"About Turing","publisher":"Turing"},{"url":"https://www.turing.com/terms-of-service","title":"Terms of Service","publisher":"Turing"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"turing-hr-ticket-reply-drafting"},{"title":"TÜV SÜD: Security Copilot for threat analysis in the SOC","useCases":["security-alert-triage-and-investigation"],"organization":{"name":"TÜV SÜD","anonymized":false,"country":"DE","region":"europe","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"TÜV SÜD, the German testing and certification group, protects about 28,000 employees with Microsoft Defender solutions, runs Microsoft Sentinel as its SIEM and joined the early adopter programme for Security Copilot. Its analysts use Copilot inside Defender to enrich alerts, investigate threats, start remediation and produce consistent investigation reports, and the company says new analysts become effective within months.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[{"kpi":"productivity-gain","value":60,"unit":"percent","qualifier":"approximately","period":"speed of threat analysis, reported as 60% to 70% faster","claimant":"organization","quote":"We analyze results about 60% to 70% faster with Security Copilot.","sourceUrl":"https://www.microsoft.com/en/customers/story/25045-tuv-sud-microsoft-security-copilot"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/25045-tuv-sud-microsoft-security-copilot","title":"TÜV SÜD anticipates the future confidently with Microsoft Defender, Security Copilot","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"tuv-sud-security-copilot-threat-analysis"},{"title":"Uber Freight: AI load recommendations for carriers","useCases":["freight-dispatch-and-load-matching-agent"],"organization":{"name":"Uber Freight","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[{"name":"Uber Freight","role":"in-house"}],"summary":"Uber Freight replaced the default search page in its digital freight brokerage with a recommendations system that ranks loads for each carrier, using candidate generation from saved, clicked and previously booked loads and an XGBoost ranking model with lead time, repeat lane and distance preference as features, instead of leaving carriers to search for loads themselves.","stage":"pilot","year":2023,"channels":["mobile-app"],"languages":["en"],"metrics":[{"kpi":"conversion-rate-uplift","value":12,"unit":"percent","qualifier":"exact","baseline":"Default search page, same carrier population, active users segment only","claimant":"organization","quote":"We tested our new recommendations system with a user-level A/B experiment, to positive results. Lifts occurred throughout the carrier booking funnel, most notably an increase of 12% in bookings for active users and an overall increase of 3% in bookings and 5% in clicks.","sourceUrl":"https://www.uberfreight.com/en-US/blog/better-load-matching-with-ai"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.uberfreight.com/en-US/blog/better-load-matching-with-ai","title":"Achieving Better Load Matching with AI","publisher":"Uber Freight","date":"2023-09-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"uber-freight-ai-load-recommendations"},{"title":"Uber: QueryGPT natural language to SQL","useCases":["governed-text-to-sql-analytics"],"organization":{"name":"Uber Technologies","anonymized":false,"country":"US","region":"global","industry":"technology"},"vendors":[{"name":"Uber","role":"in-house"}],"summary":"Uber's QueryGPT turns an English question into SQL against its data platform, which handles about 1.2 million interactive queries a month. It narrows the problem with curated \"workspaces\" of tables and sample queries per business domain (such as Mobility, Ads and Core Services), picks the relevant tables and columns with separate agents, and returns the generated SQL with an explanation so the user can check it. It was released to some Operations and Support teams first.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":300,"unit":"count","qualifier":"approximately","period":"daily active users during the limited release","claimant":"organization","quote":"With our limited release to some teams in Operations and Support, we are averaging about 300 daily active users, with about 78% saying that the generated queries have reduced the amount of time they would’ve spent writing it from scratch.","sourceUrl":"https://www.uber.com/us/en/blog/query-gpt/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.uber.com/us/en/blog/query-gpt/","title":"QueryGPT, Natural Language to SQL Using Generative AI","publisher":"Uber Engineering Blog","date":"2024-09-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"uber-querygpt-natural-language-to-sql"},{"title":"UBS: UBS Red smart assistants for client advisors","useCases":["wealth-advisor-knowledge-assistant","investment-research-summarization"],"organization":{"name":"UBS","anonymized":false,"country":"CH","region":"europe","industry":"wealth-and-asset-management"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"UBS built two domain specific assistants, together called UBS Red, on Azure AI Search and Azure OpenAI Service to give client advisors fast, multilingual access to the bank's investment advice and product content during client work. UBS digitized about 60,000 investment advice and product documents into a queryable knowledge base, which it says saves considerable time in meeting preparation and research. Within 10 months the wider Azure OpenAI footprint reached key wealth, banking and operations divisions in the Switzerland, Hong Kong and Singapore booking centres. No usage or time saving figure specific to UBS Red is published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/19796-ubs-azure","title":"UBS and Microsoft unite: Co-creating the future of banking with Azure AI","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ubs-red-client-advisor-assistants"},{"title":"UBS: STAAT Insights engine for US financial advisors","useCases":["next-best-action-for-advisors"],"organization":{"name":"UBS","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"UBS","role":"in-house"}],"summary":"UBS's US wealth management business runs STAAT Insights, a machine learning engine from its Smart Technologies and Advanced Analytics Team (STAAT) that surfaces client opportunities and alerts to financial advisors, such as shifting liquidity needs from a maturing CD, a concentrated stock position or a life event, and sends pre meeting client briefings with suggested talking points. In a December 2025 interview its chief data and analytics officer said 80% of US advisors actively use the engine. Two time saving figures UBS has published are about its AI tools in general, not STAAT Insights alone, so they are not recorded as metrics here: an advisor recruiting page says some advisors using UBS AI have saved three to four hours per client meeting, and the same executive conservatively estimated that US advisors save 10,000 hours a month by using AI to prepare for client meetings.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"employee-adoption","value":80,"unit":"percent","qualifier":"exact","period":"US advisors actively using STAAT Insights","claimant":"organization","quote":"Let me give you some stats here: 80% of them are actively using that STAAT Insights engine.","sourceUrl":"https://www.financial-planning.com/news/ubs-turns-to-ai-to-gain-wallet-share-find-new-clients"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.ubs.com/us/en/wealth-management/financial-advisor-experience/articles/ai-for-financial-advisors.html","title":"UBS AI for Financial Advisors: A New Standard in Growth","publisher":"UBS"},{"url":"https://www.financial-planning.com/news/ubs-turns-to-ai-to-gain-wallet-share-find-new-clients","title":"UBS turns to AI to gain wallet share, find new clients","publisher":"Financial Planning","date":"2025-12-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ubs-staat-insights-for-advisors"},{"title":"University Hospitals Coventry and Warwickshire: AI process mining to time appointment reminders","useCases":["patient-appointment-scheduling-and-reminders-agent","outbound-reminder-and-confirmation-agent"],"organization":{"name":"University Hospitals Coventry and Warwickshire NHS Trust","anonymized":false,"country":"GB","region":"europe","industry":"healthcare"},"vendors":[{"name":"IBM","role":"integrator"},{"name":"Celonis","role":"platform"}],"summary":"University Hospitals Coventry and Warwickshire used AI based process mining with IBM and Celonis to study missed appointments, which were more common among patients with high deprivation scores. It found a spike in last minute cancellations after two SMS reminders and moved to a reminder 14 days before the appointment with a second one four days before, so patients could cancel early and the slot could be rebooked. NHS England reports that missed appointments in this subset of patients fell from 10% to 4%. The AI analysed the process; the reminders themselves are standard text messages.","stage":"pilot","year":2024,"channels":["sms","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/","title":"NHS AI expansion to help tackle missed appointments and improve waiting times","publisher":"NHS England","date":"2024-03-14","archivedUrl":"https://web.archive.org/web/20250125233954/https://www.england.nhs.uk/2024/03/nhs-ai-expansion-to-help-tackle-missed-appointments-and-improve-waiting-times/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"uhcw-process-mining-appointment-reminders"},{"title":"UK government i.AI: Redbox assistant for summarising and redrafting official documents","useCases":["civil-servant-drafting-copilot"],"organization":{"name":"Department for Science, Innovation and Technology (Incubator for Artificial Intelligence)","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Incubator for Artificial Intelligence (i.AI)","role":"in-house"}],"summary":"Redbox was an assistant built by the UK government's Incubator for Artificial Intelligence (i.AI) that let civil servants chat with large language models, with or without their own documents, for work up to OFFICIAL SENSITIVE, and choose between models. It returned summaries, redrafts or translations of material users provided and did not search the internet. In beta in the Cabinet Office, No 10 and DSIT it had about 2,000 users in February 2025, growing by approximately 150 a week. i.AI later decided on a controlled shutdown: tools such as Microsoft Copilot, whose chat became freely available to many departments, offered similar functionality, and the Cabinet Office moved to an enterprise Gemini option, so the service was run only until the end of 2025. In its lessons learned post i.AI reports that around 70% of interactions were general requests such as drafting emails and brainstorming. The code is open source, and the Department for Business and Trade built its own adapted version.","stage":"paused","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":2000,"unit":"count","qualifier":"approximately","period":"February 2025, Cabinet Office, No 10 and DSIT","claimant":"organization","quote":"As of February 2025 Redbox is used by 2,000 Civil Servants across the Cabinet Office, No10 and DSIT.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/dsit-redbox"}],"outcomeDisclosed":true,"sources":[{"url":"https://ai.gov.uk/blogs/redbox-reflections-5-key-lessons-from-building-and-sunsetting-our-government-ai-chatbot/","title":"Redbox Reflections: 5 Key Lessons from Building (and Sunsetting) Our Government AI Chatbot","publisher":"Incubator for Artificial Intelligence (i.AI)","date":"2025-10-23"},{"url":"https://github.com/i-dot-ai/redbox","title":"i-dot-ai/redbox (source code repository, archived)","publisher":"Incubator for Artificial Intelligence (GitHub)"},{"url":"https://www.gov.uk/algorithmic-transparency-records/dsit-redbox","title":"DSIT: Redbox (algorithmic transparency record)","publisher":"GOV.UK","date":"2025-04-28"},{"url":"https://www.gov.uk/algorithmic-transparency-records/dbt-redbox","title":"DBT: Redbox (algorithmic transparency record)","publisher":"GOV.UK","date":"2025-10-16"},{"url":"https://github.com/uktrade/redbox","title":"uktrade/redbox (source code repository)","publisher":"Department for Business and Trade (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"uk-cabinet-office-redbox-civil-service-assistant"},{"title":"UK Department for Transport: Consultation Analysis Tool, evaluated with The Alan Turing Institute","useCases":["public-consultation-response-analysis"],"organization":{"name":"Department for Transport","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Department for Transport AI and Data Science team","role":"in-house"},{"name":"The Alan Turing Institute","role":"integrator"}],"summary":"The Department for Transport runs around 55 consultations a year and co developed a Consultation Analysis Tool (CAT) with The Alan Turing Institute. An ensemble of large language models proposes themes and rare \"golden insights\", humans validate the themes in a structured review of a random sample, and the models then map every response to the validated themes. The published evaluation compares the tool with human coded datasets in blind and live settings, tests for accuracy differences across demographic groups (no evidence of systematic bias on the three questions analysed; small differences by ethnicity, under 3 percentage points and favouring minority groups, were of low practical significance) and estimates that the tool would save around 50 to 70 percent of the cost and time of a notional medium sized consultation (a modelled estimate, not a measured result).","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":200000,"unit":"count","qualifier":"approximately","period":"live pilots to date (December 2025 report)","claimant":"organization","quote":"The CAT has now been piloted on multiple live consultations, analysing 200,000 responses (exceeding 8 million words) to date.","sourceUrl":"https://assets.publishing.service.gov.uk/media/696f6641011505255b2d4203/ai-consultation-analysis-tool-CAT-evaluation.pdf"},{"kpi":"hours-saved","value":15000,"unit":"hours","qualifier":"approximately","period":"cumulative over four consultation and call for evidence or ideas projects to date (December 2025 report), modelled counterfactual","baseline":"Modelled scenario in which all responses are analysed manually by humans","claimant":"organization","quote":"even with this investment, the CAT has roughly saved 15,000 hours of work to date compared to a scenario where all responses for all consultations are rigorously analysed by humans manually.","sourceUrl":"https://assets.publishing.service.gov.uk/media/696f6641011505255b2d4203/ai-consultation-analysis-tool-CAT-evaluation.pdf"},{"kpi":"cost-savings","value":500000,"unit":"currency","currency":"GBP","qualifier":"approximately","period":"cumulative over four consultation and call for evidence or ideas projects to date (December 2025 report), modelled counterfactual","baseline":"Modelled scenario in which all responses are analysed manually by humans","claimant":"organization","quote":"the CAT has analysed 200,000 responses and over 8 million words, saving an estimated £0.5 million compared to a scenario where all responses for all consultations were rigorously analysed by humans.","sourceUrl":"https://assets.publishing.service.gov.uk/media/696f6641011505255b2d4203/ai-consultation-analysis-tool-CAT-evaluation.pdf"},{"kpi":"accuracy","value":92,"unit":"percent","qualifier":"at-least","period":"theme mapping, raw agreement with human coders in blind and live evaluations","claimant":"organization","quote":"The CAT-vs-human inter-rater reliability (IRR), using metrics commonly employed in qualitative research to assess how consistently two or more researchers analyse the same data, achieved over 92% overall raw agreement in both our blind and non-blind evaluation designs.","sourceUrl":"https://assets.publishing.service.gov.uk/media/696f6641011505255b2d4203/ai-consultation-analysis-tool-CAT-evaluation.pdf"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/government/publications/ai-consultation-analysis-tool-evaluation","title":"AI Consultation Analysis Tool evaluation","publisher":"Department for Transport","date":"2025-12-23"},{"url":"https://assets.publishing.service.gov.uk/media/696f6641011505255b2d4203/ai-consultation-analysis-tool-CAT-evaluation.pdf","title":"AI Consultation Analysis Tool v1.0 evaluation (PDF)","publisher":"Department for Transport"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"uk-department-for-transport-consultation-analysis-tool"},{"title":"UK Government: cross government Microsoft 365 Copilot experiment with 20,000 government employees","useCases":["meeting-summarization-and-action-items","civil-servant-drafting-copilot"],"organization":{"name":"Government Digital Service","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"The Government Digital Service ran a trial of Microsoft 365 Copilot with 20,000 employees across UK government from 30 September to 31 December 2024. Copilot in Teams was the most used application throughout, with adoption peaking at 71%, and one participant named summarising meeting notes among the tasks where it helped. Participants estimated an average saving of 26 minutes a day across all tasks; that figure is self reported, covers every Copilot use and is not recorded as a meeting metric. The report also notes weaker results on complex, nuanced or context heavy work, and concerns that Copilot relied on external sources without built in verification.","stage":"pilot","year":2024,"channels":["microsoft-teams","internal-tools"],"languages":["en"],"metrics":[{"kpi":"employee-adoption","value":71,"unit":"percent","qualifier":"up-to","period":"peak share of trial users using Copilot in Teams, October to December 2024","claimant":"organization","quote":"Teams was the most popular tool for M365 Copilot and remained dominant throughout the experiment with a maximum adoption of 71%.","sourceUrl":"https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html","title":"Microsoft 365 Copilot Experiment: Cross-Government Findings Report","publisher":"Government Digital Service (GOV.UK)","date":"2025-06-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"uk-government-m365-copilot-experiment"},{"title":"UK government i.AI: Consult for public consultation analysis, first used live by the Scottish Government","useCases":["public-consultation-response-analysis"],"organization":{"name":"Department for Science, Innovation and Technology (Incubator for Artificial Intelligence)","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Incubator for Artificial Intelligence (i.AI)","role":"in-house"}],"summary":"Consult is a generative AI tool built by the UK government's Incubator for Artificial Intelligence as part of the Humphrey suite. It proposes themes for each open question of a consultation, maps every response to those themes and shows the result in a dashboard that officials review and correct. Its first live use was on a Scottish Government consultation about non surgical cosmetic procedures, where it analysed more than 2,000 responses while officials also reviewed every response by hand; the government reported an F1 score of 0.76 in this first live evaluation. In July 2026 access was managed through a waitlist ahead of a cross government rollout planned for 2027.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":2000,"unit":"count","qualifier":"at-least","period":"first live consultation (Scottish Government, six open questions)","claimant":"organization","quote":"Reviewing comments from over 2,000 consultation responses using generative AI, Consult identified key themes that feedback fell into across each of six qualitative questions.","sourceUrl":"https://www.gov.uk/government/news/government-built-humphrey-ai-tool-reviews-responses-to-consultation-for-first-time-in-bid-to-save-millions"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/government/news/government-built-humphrey-ai-tool-reviews-responses-to-consultation-for-first-time-in-bid-to-save-millions","title":"Government-built \"Humphrey\" AI tool reviews responses to consultation for first time, in bid to save millions","publisher":"Department for Science, Innovation and Technology","date":"2025-05-14"},{"url":"https://www.gov.uk/algorithmic-transparency-records/dsit-consult","title":"DSIT: Consult (algorithmic transparency record)","publisher":"Department for Science, Innovation and Technology","date":"2025-05-14"},{"url":"https://www.gov.uk/government/publications/consult-register-your-interest/consult-ai-tool","title":"Consult: AI tool","publisher":"Department for Science, Innovation and Technology","date":"2026-07-22"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"uk-incubator-for-ai-consult-consultation-analysis"},{"title":"UK Valuation Office Agency: automated valuation model for the 2028 Wales Council Tax revaluation","useCases":["property-valuation-support"],"organization":{"name":"Valuation Office Agency","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[],"summary":"The UK's Valuation Office Agency built its own Automated Valuation Model, without an external supplier, to give a first pass value, the closest comparable sales, a reliability score and a Council Tax band for the vast majority of Wales's 1.5 million domestic properties ahead of the 2028 Council Tax revaluation. Valuers focus their time on the batches of properties the model flags as least reliable or closest to a band boundary, and the International Association of Assessing Officers reviewed the model's development process and reported confidence in its quality.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/valuation-office-agency-automated-valuation-model","title":"Valuation Office Agency: Automated Valuation Model","publisher":"GOV.UK","date":"2025-08-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"uk-voa-wales-automated-valuation-model"},{"title":"Ulta Beauty: agentic checkout in Google AI Mode and the Gemini app","useCases":["agentic-payment-initiation"],"organization":{"name":"Ulta Beauty","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Google","role":"platform"}],"summary":"Google Cloud lists Ulta Beauty as rolling out agentic commerce inside AI Mode in Google Search and the Gemini app, where shoppers get Ulta Beauty product recommendations, compare options and complete checkout for eligible purchases within Google's conversational interfaces. The entry was added in the April 2026 edition of Google's list and describes a rollout, not results. No outcome figures are disclosed.","stage":"announced","year":2026,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"1,302 real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"ulta-beauty-agentic-checkout-google-ai-mode"},{"title":"Unifonic: generative AI for sales outreach content and conversation insights","useCases":["outbound-sales-prospecting-agent"],"organization":{"name":"Unifonic","anonymized":false,"country":"SA","region":"middle-east","industry":"technology"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Unifonic, a Saudi customer communications platform, uses Microsoft 365 Copilot in its sales and marketing teams to analyze conversations across platforms, draft proposals, email campaigns and social media content, and summarize the action points of recorded customer meetings so a sales representative can email the customer with next steps right after. Microsoft reports that this led to a 20% increase in total sales outreach volume. The deployment is a general productivity assistant used for outreach, not a dedicated prospecting agent.","stage":"production","year":2025,"channels":["email","internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft","date":"2025-07-24"},{"url":"https://www.microsoft.com/en/customers/story/23690-unifonic-microsoft-365-e5","title":"Unifonic boosts productivity and accelerates customer engagement using Microsoft 365 Copilot","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"unifonic-copilot-sales-outreach"},{"title":"Unilever: AI inventory and assurance process with Holistic AI","useCases":["ai-model-inventory"],"organization":{"name":"Unilever","anonymized":false,"country":"GB","region":"europe","industry":"manufacturing"},"vendors":[{"name":"Holistic AI","role":"platform"}],"summary":"Unilever built an AI assurance process in which every new AI application, broadly defined to include any prediction or automation, is registered, triaged and rated red, amber or green for effectiveness and ethical risk before it goes into production. Holistic AI co created the inventory and risk management platform that the data ethics team uses to track submissions, completeness and risk ratings across a decentralised global business, with a growing share of assessments mapped to the EU AI Act. No quantified outcome is published in a form that can be quoted as a sentence.","stage":"scaled","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://sloanreview.mit.edu/article/ai-ethics-at-unilever-from-policy-to-process","title":"AI Ethics at Unilever: From Policy to Process","publisher":"MIT Sloan Management Review"},{"url":"https://uploads-ssl.webflow.com/6305e5d52c28356b4fe71bac/63dce08cf171bb4803af691c_Holistic-AI-Case-Study-Unilever.pdf","title":"Case Study: AI Inventory and Risk Management for Unilever","publisher":"Holistic AI"},{"url":"https://www.holisticai.com/how-we-helped-unilever","title":"How We Helped Unilever","publisher":"Holistic AI"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"unilever-ai-inventory-and-assurance"},{"title":"Unilever: Flex Experiences AI platform matches employees with internal projects","useCases":["internal-talent-marketplace-matching"],"organization":{"name":"Unilever","anonymized":false,"country":"GB","region":"global","industry":"manufacturing"},"vendors":[{"name":"Gloat","role":"platform"}],"summary":"Unilever's Flex Experiences platform uses AI to match employees with project opportunities across the business, so that people can work on projects for part of their time and build new skills while project leaders find the expertise they need. By 2020 it had been rolled out to more than 60,000 employees in more than 100 countries, and during the pandemic lockdowns of 2020 teams used it to staff urgent work quickly, such as an information and analytics (I&A) squad. Unilever publishes no outcome figures on the page cited.","stage":"scaled","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.unilever.com/news/news-search/2020/an-exciting-new-normal-for-flexible-working/","title":"An exciting new normal for flexible working","publisher":"Unilever","archivedUrl":"https://web.archive.org/web/2026/https://www.unilever.com/news/news-search/2020/an-exciting-new-normal-for-flexible-working/"},{"url":"https://gloat.com/blog/unilever-ai-retain-develop-engage-talent-innermobility-gloat/","title":"How Unilever implements AI to engage talents","publisher":"Gloat"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"unilever-flex-experiences-talent-marketplace"},{"title":"UniSuper: automated file notes for financial advisers","useCases":["client-meeting-notes-and-crm-update"],"organization":{"name":"UniSuper","anonymized":false,"country":"AU","region":"asia-pacific","industry":"wealth-and-asset-management"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"UniSuper, an Australian superannuation fund, uses Microsoft 365 Copilot to produce file notes that summarise the key details of each adviser conversation with a member held on Microsoft Teams. Microsoft reports advisers save roughly 30 minutes per client interaction; the fund also says the notes give it better visibility of interaction quality. The annual hours and extra members advised are projections and are not recorded.","stage":"production","year":2024,"channels":["microsoft-teams"],"languages":["en"],"metrics":[{"kpi":"time-saved-per-task","value":30,"unit":"minutes","qualifier":"approximately","period":"per client interaction","claimant":"vendor","quote":"Advisors are saving roughly 30 minutes per client interaction on Microsoft Teams by using automated, bespoke file notes that summarise key details of each conversation.","sourceUrl":"https://news.microsoft.com/source/asia/features/super-thinking-how-unisuper-is-using-microsoft-365-copilot-to-deliver-faster-and-better-outcomes-for-members/"}],"outcomeDisclosed":true,"sources":[{"url":"https://news.microsoft.com/source/asia/features/super-thinking-how-unisuper-is-using-microsoft-365-copilot-to-deliver-faster-and-better-outcomes-for-members/","title":"Super thinking: How UniSuper is using Microsoft 365 Copilot to deliver faster and better outcomes for members","publisher":"Microsoft Source Asia"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"unisuper-copilot-advice-file-notes"},{"title":"United Airlines: generative AI delay messages, automatic rebooking and ConnectionSaver","useCases":["flight-disruption-and-rebooking-agent"],"organization":{"name":"United Airlines","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[],"summary":"Customer service teams in United's network operations centre use generative AI to review flight data and write the text and email messages that explain why a flight is delayed or changed, including links to live radar maps during weather delays. When a flight is delayed or cancelled, United's self service tools automatically present personalized rebooking options, bag tracking and meal and hotel vouchers when eligible. Its AI powered ConnectionSaver tool identifies departing flights that can be held for connecting customers without delaying the on time arrival of those already on board; United says it has saved more than 3.3 million customer connections since launching in 2019. That tool works on the operation, not in conversations with passengers.","stage":"scaled","year":2024,"channels":["sms","email","mobile-app"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://united.mediaroom.com/2024-07-03-United-Now-Texts-Live-Radar-Maps-and-Uses-AI-to-Keep-Travelers-Informed-During-Weather-Delays","title":"United Now Texts Live Radar Maps and Uses AI to Keep Travelers Informed During Weather Delays","publisher":"United Airlines","date":"2024-07-03"},{"url":"https://united.mediaroom.com/2025-06-25-United-Mobile-App-Now-Gives-People-More-Information-About-Their-Connecting-Flight","title":"United Mobile App Now Gives People More Information About Their Connecting Flight","publisher":"United Airlines","date":"2025-06-25"},{"url":"https://united.mediaroom.com/2025-12-16-United-Adds-New-Features-to-Award-Winning-Mobile-App-Like-Virtual-Gate,-Club-Recommendation-Tool-and-Real-Time-Bag-Tracker","title":"United Adds New Features to Award-Winning Mobile App Like Virtual Gate, Club Recommendation Tool and Real-Time Bag Tracker","publisher":"United Airlines","date":"2025-12-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"united-airlines-disruption-messaging-and-rebooking"},{"title":"United Bank Limited: automated trade compliance screening against trade based money laundering","useCases":["trade-finance-crime-screening"],"organization":{"name":"United Bank Limited","anonymized":false,"country":"PK","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"LexisNexis Risk Solutions","role":"platform"}],"summary":"United Bank Limited in Pakistan selected an automated trade screening platform in 2020 to move its trade compliance checks from mostly manual work to automation and meet the Pakistan Single Window regulatory directive. One interface now runs customer sanctions screening and risk assesses the trading activity itself, including dual use goods identification, the countries involved and vessel history and tracking, with a full activity history for audits. The vendor's case study reports faster turnaround and fewer false positives but gives no figures, and it describes automated screening against sanctions, trade and vessel data, not AI.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://risk.lexisnexis.com/global/en/insights-resources/case-study/united-bank-limited-trade-compliance","title":"United Bank Limited Tackles TBML With Trade Compliance Solutions","publisher":"LexisNexis Risk Solutions"},{"url":"https://risk.lexisnexis.com/global/-/media/files/financial%20services/case%20study/lnrs-united-bank-limited-case-study-nxr16244-00-1123-en-us.pdf","title":"Tackling trade-based money laundering and improving transaction turnaround times with LexisNexis Firco Trade Compliance","publisher":"LexisNexis Risk Solutions"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"united-bank-limited-trade-compliance-screening"},{"title":"University of Utah: UGuide major exploration chatbot","useCases":["academic-advising-assistant"],"organization":{"name":"University of Utah","anonymized":false,"country":"US","region":"north-america","industry":"education"},"vendors":[],"summary":"The University of Utah's Academic Innovation + Intelligence Lab, part of the Office of Undergraduate Studies, built UGuide, an AI chatbot that helps students explore majors by consolidating data spread across the institution into one conversational interface. Students can investigate majors from angles such as career paths and scheduling, and the tool surfaces prompts based on what other students have asked. The university describes UGuide as being in the early stages of development, with piloting planned for upcoming semesters, distinct from the lab's other chatbot, UBot, a course specific virtual tutor already piloted in large gateway classes.","stage":"announced","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://us.utah.edu/news-and-updates/posts/ai-powered-chatbots.php","title":"AI-Powered Chatbots: Functional 4-Year Degree Plans","publisher":"University of Utah, Office of Undergraduate Studies","date":"2024-10-02"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"B","id":"university-of-utah-uguide-advising"},{"title":"UOB: LLM as a judge testing of an internal generative AI chatbot with PwC","useCases":["model-risk-validation-copilot"],"organization":{"name":"United Overseas Bank (UOB)","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"PwC","role":"integrator"}],"summary":"In the AI Verify Foundation's Global AI Assurance Pilot (February to May 2025), PwC tested UOB's internal retrieval augmented generation chatbot, which runs in production for selected staff on Meta Llama 3.1 and answers operational and domain questions from public company documents. The risk assessment focused on model risks. PwC combined rule based scoring for binary and multiple choice questions, embedding similarity for consistency across repeated runs, and LLM based checks of reasoning answers: an LLM split each answer into clauses, an LLM as a judge compared each clause with retrieved passages of the source document to flag contradictions (a clause with no supporting passage counted as a hallucination), and a judge listed the parts of each question left unanswered. Because the production infrastructure was shared with other use cases, outputs were generated manually in a sandbox; because of confidentiality, PwC used its own prompts and ground truths for ten companies, which UOB reviewed. The case study therefore treats the results as a proxy for the production tool and publishes none of them.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://assurance.aiverifyfoundation.sg/report/pilot-participants-and-use-cases/","title":"Pilot participants and use cases","publisher":"AI Verify Foundation"},{"url":"https://assurance.aiverifyfoundation.sg/wp-content/uploads/2025/05/Internal-GenAI-chatbot.pdf","title":"Internal GenAI Chatbot: UOB x PwC","publisher":"AI Verify Foundation"},{"url":"https://aiverifyfoundation.sg/ai-assurance-pilot/","title":"Global AI Assurance Pilot","publisher":"AI Verify Foundation"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"uob-pwc-llm-judge-genai-chatbot-testing"},{"title":"UOB: machine learning prioritisation of transaction monitoring alerts with Tookitaki","useCases":["aml-alert-triage"],"organization":{"name":"United Overseas Bank (UOB)","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Tookitaki","role":"platform"}],"summary":"UOB co developed a machine learning anti money laundering solution with Tookitaki that sorts transaction monitoring alerts into three priority tiers and identifies connected parties, so investigators focus on the cases most likely to be suspicious. UOB said in December 2020 that it was the first Singapore bank to apply AI to transaction monitoring and name screening at the same time, working through more than 5,700 alerts a month, and that the model complements its rules rather than replacing them.","stage":"production","year":2020,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"accuracy","value":96,"unit":"percent","qualifier":"exact","period":"true positive prediction rate of the high priority tier","claimant":"organization","quote":"Since its implementation, UOB’s new AI solution has proven an overall true positive prediction rate of 96 per cent in the ‘high priority’ category","sourceUrl":"https://www.uobgroup.com/web-resources/uobgroup/pdf/newsroom/2020/UOB-new-AI-money-laundering-solution.pdf"},{"kpi":"interactions-handled","value":5700,"unit":"count","qualifier":"at-least","period":"transaction alerts per month","claimant":"organization","quote":"UOB’s AI solution sieves through an average of more than 5,700 transaction alerts each month to flag cases that are more likely to be suspicious with an overall true positive prediction rate of 96 per cent","sourceUrl":"https://www.uobgroup.com/web-resources/uobgroup/pdf/newsroom/2020/UOB-new-AI-money-laundering-solution.pdf"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.uobgroup.com/web-resources/uobgroup/pdf/newsroom/2020/UOB-new-AI-money-laundering-solution.pdf","title":"UOB's new AI anti-money laundering solution helps the Bank cut through large volumes of transactions to pinpoint suspicious activities","publisher":"UOB","date":"2020-12-03"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"uob-tookitaki-aml-alert-prioritisation"},{"title":"UOB: machine learning pilot for name screening and transaction monitoring with Tookitaki","useCases":["sanctions-screening-adjudication"],"organization":{"name":"United Overseas Bank (UOB)","anonymized":false,"country":"SG","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Tookitaki","role":"platform"}],"summary":"In a six month pilot reported in August 2018, UOB tested Tookitaki's Anti-Money Laundering Suite, with machine learning features co created by the bank, on top of its rule based name screening (against internal and external watch lists) and transaction monitoring, to separate genuine risk from false positives with explainable outputs. UOB reported large false positive reductions on name screening alerts and said it would progressively roll the solution out to customer risk assessment and sanctions screening, which it treats as processes separate from name screening.","stage":"pilot","year":2018,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"false-positive-reduction","value":60,"unit":"percent","qualifier":"exact","period":"six month pilot, name screening alerts on individual names","claimant":"organization","quote":"For name screening alerts, there was a 60 per cent and 50 per cent reduction in false positives","sourceUrl":"https://www.uobgroup.com/web-resources/uobgroup/pdf/newsroom/2018/UOB-and-Tookitaki-strengthen-combat-against-money-laundering.pdf"},{"kpi":"false-positive-reduction","value":50,"unit":"percent","qualifier":"exact","period":"six month pilot, name screening alerts on corporate names","claimant":"organization","quote":"For name screening alerts, there was a 60 per cent and 50 per cent reduction in false positives","sourceUrl":"https://www.uobgroup.com/web-resources/uobgroup/pdf/newsroom/2018/UOB-and-Tookitaki-strengthen-combat-against-money-laundering.pdf"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.uobgroup.com/web-resources/uobgroup/pdf/newsroom/2018/UOB-and-Tookitaki-strengthen-combat-against-money-laundering.pdf","title":"UOB and Tookitaki strengthen combat against money laundering through co-created machine learning solution","publisher":"UOB","date":"2018-08-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"uob-tookitaki-name-screening-pilot"},{"title":"Uphold: pilot of an AI agent for alert review and regulatory filing preparation with Unit21","useCases":["suspicious-activity-report-drafting","aml-alert-triage"],"organization":{"name":"Uphold","anonymized":false,"country":"US","region":"north-america","industry":"payments"},"vendors":[{"name":"Unit21","role":"platform"}],"summary":"Crypto platform Uphold unified alerts, cases and regulatory filings with FinCEN and FINTRAC in Unit21, and piloted Unit21's AI agent to help analysts review alerts faster and more consistently. The vendor reports a drop in median alert review time from the pilot and predicts much faster suspicious transaction report preparation, which has not yet been measured.","stage":"pilot","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":44,"unit":"percent","qualifier":"exact","period":"median alert review time during the pilot","claimant":"vendor","quote":"Median alert review time has dropped by 44% thanks to a pilot of Unit21’s AI Agent, which helps analysts process alerts faster and more consistently.","sourceUrl":"https://www.unit21.ai/customers/uphold"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.unit21.ai/customers/uphold","title":"Uphold Case Study","publisher":"Unit21"},{"url":"https://baytobaynews.com/daily-state-news/stories/unit21-awarded-two-2026-datos-impact-awards-for-ai-innovation-cryptodigital-asset-aml-innovation,346520","title":"Unit21 Awarded Two 2026 Datos Impact Awards for AI Innovation & Crypto/Digital Asset AML Innovation","publisher":"Business Wire (via Bay to Bay News)","date":"2026-09-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"uphold-unit21-ai-agent-pilot"},{"title":"UPS: agentic AI in customs brokerage after the end of the US de minimis exemption","useCases":["customs-classification-and-declaration"],"organization":{"name":"United Parcel Service","anonymized":false,"country":"US","region":"north-america","industry":"logistics-and-transportation"},"vendors":[{"name":"UPS","role":"in-house"}],"summary":"UPS, one of the largest customs brokers, told investors on its third quarter 2025 earnings call that it enhanced its customs brokerage with agentic AI to cope with a tenfold rise in daily customs entries after the United States ended the de minimis exemption. Chief executive Carol Tomé said that in March 2025 about 21% of the 13,000 packages a day that needed a dutiable clearance were cleared without manual intervention, against 90% of 112,000 packages a day in September 2025. UPS also says it uses AI and machine learning to check Harmonized System classification codes, tariff rules and policies on imports.","stage":"scaled","year":2025,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":90,"unit":"percent","qualifier":"exact","period":"September 2025, US packages needing a dutiable clearance (112,000 a day)","baseline":"About 21% of 13,000 daily packages in March 2025","claimant":"organization","quote":"In September 2025, the carrier cleared 90% of 112,000 daily packages with no manual intervention.","sourceUrl":"https://www.supplychaindive.com/news/ups-ai-employee-upskilling-network-changes/816412/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.supplychaindive.com/news/ups-ai-employee-upskilling-network-changes/816412/","title":"How UPS is using AI, from shipper pricing to customs clearance","publisher":"Supply Chain Dive","date":"2026-04-06"},{"url":"https://equibles.com/stocks/ups/calls/2025-q3","title":"United Parcel Service Inc (UPS) Q3 2025 Earnings Call Transcript","publisher":"Equibles","date":"2025-10-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ups-agentic-ai-customs-brokerage"},{"title":"Upstart: alternative data and machine learning credit model under a CFPB No Action Letter","useCases":["alternative-data-credit-scoring"],"organization":{"name":"Upstart Network","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Upstart","role":"in-house"}],"summary":"Upstart underwrites and prices consumer loans with a machine learning model that adds alternative data, such as education and employment, to traditional credit data. In 2017 it received the CFPB's first No Action Letter for this model, and in 2019 the CFPB published the access to credit results Upstart reported under that letter: in simulations run against a hypothetical traditional model on the same applicant pool, Upstart's model approved 27% more applicants with 16% lower average APRs, across all tested race, ethnicity and sex segments. The CFPB states it did not separately replicate these simulations. An independent fair lending monitorship of the model published its initial report in April 2021.","stage":"scaled","year":2019,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.consumerfinance.gov/about-us/blog/update-credit-access-and-no-action-letter/","title":"An update on credit access and the Bureau's first No-Action Letter","publisher":"Consumer Financial Protection Bureau"},{"url":"https://www.relmanlaw.com/media/cases/1088_Upstart%20Initial%20Report%20-%20Final.pdf","title":"Fair Lending Monitorship of Upstart Network's Lending Model, initial report","publisher":"Relman Colfax PLLC","date":"2021-04-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"upstart-alternative-data-credit-model"},{"title":"U.S. Bank: generative AI Developer Assistant for API integration","useCases":["developer-api-integration-assistant"],"organization":{"name":"U.S. Bank","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[],"summary":"U.S. Bank added a generative AI Developer Assistant to its Developer Portal for developers at clients, software providers and aggregators who embed its treasury, payments and data services. It answers integration questions, recommends the right APIs, helps troubleshoot and generates sample code, and promotes practices such as account tokenization. The bank aims to cut average integration time by weeks by reducing technical consultations and support tickets; no measured result is published yet.","stage":"production","year":2026,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.usbank.com/about-us-bank/news-and-stories/article-library/genai-assistant-helps-companies-embed-us-bank-solutions-into-their-platforms.html","title":"GenAI assistant helps companies embed U.S. Bank solutions into their platforms","publisher":"U.S. Bank","date":"2026-01-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"us-bank-developer-assistant"},{"title":"US Department of Justice: consolidated 2025 AI use case inventory","useCases":["ai-model-inventory"],"organization":{"name":"U.S. Department of Justice","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Department of Justice consolidates the AI use cases of all its components into one annual inventory, reviewed by component representatives on its Emerging Technology Board with the Chief AI Officer. The 2025 inventory covers use cases in every stage from pre deployment to retired, combines similar, widely adopted AI use cases into single department wide entries and removes duplicate or mislabeled entries. It holds 315 entries, 30.7% more than the 2024 inventory, which DOJ attributes to closer collaboration between components to accelerate AI adoption.","stage":"scaled","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"ai-systems-inventoried","value":315,"unit":"count","qualifier":"exact","period":"2025 inventory","claimant":"organization","quote":"The 2025 AI Use Case Inventory includes 315 entries, a 30.7% increase from the 2024 inventory.","sourceUrl":"https://www.justice.gov/ai/ai-inventory"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.justice.gov/ai/ai-inventory","title":"AI Inventory","publisher":"U.S. Department of Justice"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"us-department-of-justice-ai-use-case-inventory"},{"title":"US Department of State: live AI interpretation at consular windows (LCALA)","useCases":["immigration-and-visa-application-assistant","public-service-translation"],"organization":{"name":"U.S. Department of State (Bureau of Consular Affairs)","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Consular Affairs is piloting Live Consular AI Language Augmentation (LCALA): real time transcription and neural machine translation of the spoken exchange at the visa interview window, with translated audio and on screen text and an optional short time stamped transcript. The visa interview pilot is flagged high impact in the federal inventory; its entry notes that interviews last about three minutes and that misunderstandings can force repeat questions, delays or uneven outcomes. A second pilot, for Overseas Citizens Services and American Citizens Services, supports calls and in person interactions and is described as assistive only, not a replacement for certified interpreters where they are required.","stage":"pilot","year":2025,"channels":["voice","internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"us-department-of-state-consular-ai-interpretation"},{"title":"US Department of the Interior: clustering and similarity tools for Freedom of Information Act requests","useCases":["freedom-of-information-request-processing"],"organization":{"name":"U.S. Department of the Interior, Office of the Solicitor","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"U.S. Department of the Interior","role":"in-house"}],"summary":"The Department of the Interior's Office of the Solicitor lists four deployed tools, developed in house, that group incoming FOIA requests: two clustering tools (one embedding based, one density based on term frequency, run in its document review platform), a semantic similarity score and a lexical similarity tool. They identify requests that ask for the same or similar records, including similar requests sent to several offices, so that the work can be coordinated and responses kept uniform instead of duplicated. The inventory gives operational dates of August and November 2023; no outcome figures are published.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"us-department-of-the-interior-foia-request-similarity-tools"},{"title":"US Department of Transportation: Public Comment Analyzer for regulations.gov dockets","useCases":["public-consultation-response-analysis"],"organization":{"name":"U.S. Department of Transportation, Office of the Secretary","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Office of the Secretary of Transportation reports a Public Comment Analyzer, deployed in February 2025, that uses a language model to categorise public comments by topic, detect their sentiment, generate summaries and provide daily updates on the comments to subject matter experts. The stated aim is to reduce the human effort of reading every comment and to let experts go from summaries to the underlying comments where needed. No measured outcome is published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"us-department-of-transportation-public-comment-analyzer"},{"title":"US Department of Veterans Affairs: Mail Automation Services for claims intake","useCases":["correspondence-triage-and-routing"],"organization":{"name":"U.S. Department of Veterans Affairs","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"IBM","role":"integrator"},{"name":"U.S. Department of Veterans Affairs","role":"in-house"}],"summary":"The Veterans Benefits Administration runs Mail Automation Services, an intake platform for claims material and other submissions from veterans and their representatives, mostly arriving through the Centralized Mail Portal. It combines form recognition, OCR, handwriting recognition and natural language processing to extract an average of 95 fields from more than 1,500 form layouts and to establish the claim and send each submission to the right business line. The agency reports that it ingests 25,000 to 40,000 packets a day; it has been in operation since 2020.","stage":"scaled","year":2020,"channels":["api","internal-tools"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":25000,"unit":"count","qualifier":"at-least","period":"packets per day (25,000 to 40,000)","claimant":"organization","quote":"The platform ingests and reviews 25k to 40k packets of information per day, primarily derived from the Centralized Mail Portal.","sourceUrl":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv"}],"outcomeDisclosed":true,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"},{"url":"https://github.com/ombegov/2024-Federal-AI-Use-Case-Inventory","title":"2024 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"us-department-of-veterans-affairs-mail-automation-services"},{"title":"US Fish and Wildlife Service: ePermits chatbot, application wizard and semantic search","useCases":["permit-and-licence-application-processing"],"organization":{"name":"U.S. Fish and Wildlife Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Fish and Wildlife Service is preparing AI for ePermits, its online permit system. A chatbot and application wizard will answer questions, give dynamic help on the application form, draw on a knowledge base and escalate to a human; a cognitive search layer adds semantic search, document summaries and natural language answers. The agency states that it does not use agentic AI in permit processing, that every AI assisted recommendation is reviewed and approved by a person, and that no permit decision rests on AI output alone.","stage":"announced","year":2025,"channels":["web-chat","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"us-fish-and-wildlife-service-epermits-assistant"},{"title":"US Immigration and Customs Enforcement: AI assisted review of emails for signs of compromise in the SOC","useCases":["security-alert-triage-and-investigation"],"organization":{"name":"U.S. Immigration and Customs Enforcement","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"ICE reports in the 2025 federal AI use case inventory that its security operations centre uses an AI Assisted Compromise Email Detector (AACED) to review a collection of emails exchanged with Microsoft under Emergency Directive 24-02. Named entity recognition flags personal data and keywords, and a chat interface lets analysts ask questions with an email as context, so they can find indicators of compromise faster. This was a review of a fixed set of emails tied to one directive rather than ongoing alert triage, and the inventory classifies it as classical machine learning rather than an agent, so it is a partial fit for this use case. The inventory lists it as deployed since June 2024 and gives no measured result.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"us-ice-soc-compromised-email-detector"},{"title":"US Immigration and Customs Enforcement: intelligent document processing for invoices and forms","useCases":["supplier-invoice-processing","intelligent-document-processing"],"organization":{"name":"U.S. Immigration and Customs Enforcement","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"UiPath","role":"platform"},{"name":"Microsoft","role":"platform"}],"summary":"Business units at ICE, part of the Department of Homeland Security, use an intelligent document processing platform (UiPath Suite and Azure AI Document Intelligence) with OCR and machine learning models to verify, extract and classify information from forms, automating repeatable work such as invoice processing and form entry validation. The agency lists it in operation since 2019 and says it saves staff significant time while improving data quality. No figures are published.","stage":"production","year":2019,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"us-immigration-and-customs-enforcement-intelligent-document-processing"},{"title":"ICE SEVP Response Center: SID voice assistant for student visa programme calls","useCases":["immigration-and-visa-application-assistant"],"organization":{"name":"U.S. Immigration and Customs Enforcement (Student and Exchange Visitor Program)","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"The Student and Exchange Visitor Program response center, which serves international students and school officials, is preparing SID (SEVIS Interactive Dialog), a voice assistant that answers routine caller questions through a deterministic question and answer workflow. When it cannot answer, it turns the caller over to an agent in the response center and creates a ticket with a caller transcript to reduce the burden on the agent. Separately, it captures the interaction and sends it through an API to the program's automated management system (SEVPAMS).","stage":"announced","year":2025,"channels":["voice"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"us-immigration-and-customs-enforcement-sevp-voice-assistant"},{"title":"US Marshals Service: AI presenters and narration for training videos (pilot)","useCases":["training-content-generation"],"organization":{"name":"U.S. Marshals Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"TechSmith","role":"platform"}],"summary":"Training content creators at the US Marshals Service, part of the Department of Justice, have piloted TechSmith Camtasia and Audiate since September 2025 to generate on screen presenters and realistic narration from text, and to produce voice audio in a range of voices and tones without recording actors. The goals are to shorten the time to release training material, support Section 508 accessibility and improve the online learning experience. No outcome figures are published.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entries DOJ-0319 Camtasia and DOJ-0330 Audiate)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"us-marshals-service-ai-training-video-pilot"},{"title":"US Treasury: machine learning in federal payment fraud prevention and check fraud recovery","useCases":["benefit-fraud-and-error-detection"],"organization":{"name":"U.S. Department of the Treasury, Bureau of the Fiscal Service","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"In house (Office of Payment Integrity)","role":"in-house"}],"summary":"The Office of Payment Integrity in Treasury's Bureau of the Fiscal Service added machine learning and risk based screening to how it checks federal payments. Treasury reports that these enhanced processes prevented and recovered over USD 4 billion in fraud and improper payments in fiscal year 2024, up from USD 652.7 million the year before. The release names machine learning only for identifying Treasury check fraud, which led to USD 1 billion in recovery; the USD 2.5 billion it reports from identifying and prioritising high risk transactions is not described as machine learning.","stage":"scaled","year":2024,"channels":[],"languages":["en"],"metrics":[{"kpi":"fraud-losses-prevented","value":1000000000,"unit":"currency","currency":"USD","qualifier":"exact","period":"fiscal year 2024 (October 2023 to September 2024)","claimant":"organization","quote":"Expediting the identification of Treasury check fraud with machine learning AI resulting in $1 billion in recovery.","sourceUrl":"https://home.treasury.gov/news/press-releases/jy2650"}],"outcomeDisclosed":true,"sources":[{"url":"https://home.treasury.gov/news/press-releases/jy2650","title":"Treasury Announces Enhanced Fraud Detection Processes, Including Machine Learning AI, Prevented and Recovered Over $4 Billion in Fiscal Year 2024","publisher":"U.S. Department of the Treasury","date":"2024-10-17"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"us-treasury-payment-integrity-machine-learning"},{"title":"US Citizenship and Immigration Services: intelligent document processing for I-539 applications","useCases":["intelligent-document-processing"],"organization":{"name":"U.S. Citizenship and Immigration Services","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Hyperscience","role":"platform"},{"name":"CGI Federal","role":"integrator"}],"summary":"Before this system, every page of an I-539 application (a request to extend or change nonimmigrant status) was scanned and stored as one document, which slowed adjudication and did not meet National Archives records standards. USCIS now uses an intelligent document processing tool to identify, classify and split each application into its component documents, such as the form itself, other USCIS forms, passports, driving licences, marriage certificates and bank statements. Pages the tool cannot identify go to a person. It is listed as deployed since November 2024; no outcome figures are published.","stage":"production","year":2024,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry DHS-2385, Intelligent Document Processing (IDP) for I-539 Form Digitization)","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://www.uscis.gov/i-539","title":"I-539, Application to Extend/Change Nonimmigrant Status","publisher":"U.S. Citizenship and Immigration Services"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"uscis-i-539-intelligent-document-processing"},{"title":"USCIS: generative AI form intake for myUSCIS and evidence classification in ELIS","useCases":["immigration-and-visa-application-assistant"],"organization":{"name":"U.S. Citizenship and Immigration Services","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Anthropic (Claude 3.7 Sonnet on Amazon Bedrock)","role":"model-provider"},{"name":"Amazon Web Services (Amazon Bedrock, Amazon Textract)","role":"platform"},{"name":"Analytica","role":"integrator"},{"name":"DV United","role":"integrator"},{"name":"SAIC","role":"integrator"}],"summary":"Two deployed USCIS systems prepare immigration applications for officers. PDF Intake lets people submit scanned forms online through myUSCIS instead of by mail: a generative model on Amazon Bedrock extracts the filled fields and checks them against form specific business rules, returning structured JSON to the case system. The ELIS Evidence Classifier tags scanned evidence pages (such as a marriage certificate or a border crossing card) so that adjudicators can jump to the document they need instead of scrolling through hundreds of pages. Their outputs are structured form data and evidence tags for the case system.","stage":"production","year":2025,"channels":["api","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"uscis-myuscis-pdf-intake-and-evidence-classification"},{"title":"USDA: AI screening in environmental permitting review","useCases":["permit-and-licence-application-processing"],"organization":{"name":"U.S. Department of Agriculture (Office of the Chief Information Officer)","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"USDA is piloting AI at several steps of its environmental review process for permits. At the evaluation stage the model predicts whether an application is a simple categorical exclusion (a project class that needs no detailed environmental assessment) or needs more human attention; AI is also used to process public comments. The aim is faster permit issuance with fewer staff hours before a project can break ground. It is developed in house. The use case was presumed high impact and then determined not to be, because the AI makes recommendations for authorized staff and no final decisions.","stage":"pilot","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (consolidated federal inventory data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"usda-environmental-permitting-review"},{"title":"USDA Forest Service: generative AI New Hire Experience assistant in the service CRM","useCases":["employee-onboarding-assistant"],"organization":{"name":"U.S. Department of Agriculture","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Salesforce","role":"platform"}],"summary":"The Forest Service, within USDA's Natural Resources and Environment mission area, runs a generative AI New Hire Experience capability in its Salesforce customer relationship manager. It gives users, new hires by its name, text based self help on human resources, business and finance processes and procedures. The inventory describes it as deployed with an operational date of January 2024, classifies it as not high impact and states that it uses no personal data; no outcome figures are published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry USDA-168, NRE FS Customer Relationship Manager New Hire Experience)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"usda-forest-service-new-hire-experience-assistant"},{"title":"US Department of Agriculture: machine learning classification of fire season incident purchases (pilot)","useCases":["procurement-spend-classification"],"organization":{"name":"U.S. Department of Agriculture","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[],"summary":"To plan for each fire season, USDA staff in Natural Resources and Environment categorized all of the previous year's incident related purchases by hand, which took a long time and could miss subtle purchasing patterns. An in house classical machine learning model, trained on 2023 item descriptions (such as protective work gloves) and validated on human labeled item categories from the same year, assigns purchases to common purchase categories identified by procurement staff in the pilot. The goal is to find purchasing patterns and cost savings opportunities, which the agency says could lead to quicker resource allocation to incident locations. The inventory lists it as a pilot since December 2024; no outcome figures are published.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry USDA-135, Yearly Incident Procurement Classification Report)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"usda-incident-procurement-classification"},{"title":"USDA: ProcureSight for market research and supplier responsibility checks","useCases":["vendor-due-diligence"],"organization":{"name":"U.S. Department of Agriculture","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"ProcureSight","role":"platform"}],"summary":"Since October 2024 USDA has used ProcureSight, a free AI search tool over public SAM.gov and USASpending data, to assist with market research and with responsibility determinations on prospective suppliers. The department expects higher procurement productivity from precise searches over public procurement data. No outcome figures are published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry USDA-097, ProcureSight)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"usda-procuresight-responsibility-determination"},{"title":"USTDA: AI assisted due diligence research on prospective partners","useCases":["vendor-due-diligence"],"organization":{"name":"U.S. Trade and Development Agency","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Exiger","role":"platform"}],"summary":"In its 2024 AI inventory the US Trade and Development Agency reports that, since May 2024, Exiger's analysts have used Exiger's DDIQ research software for due diligence on individuals and companies under consideration for partnership with the agency. The AI reviews and consolidates publicly available information such as news reports, and Exiger's due diligence and research professionals review, refine and analyse the result. It is a managed service model: the AI makes the research more efficient while analysts do the assessment. The entry does not appear in the 2025 inventory, and no outcome figures are published.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2024-Federal-AI-Use-Case-Inventory","title":"2024 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI inventory (USTDA entry Exiger Due Diligence IQ Research Software)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"ustda-exiger-partner-due-diligence"},{"title":"UWV: risk scan for culpable unemployment in unemployment benefit claims","useCases":["benefit-fraud-and-error-detection"],"organization":{"name":"Uitvoeringsinstituut Werknemersverzekeringen (UWV)","anonymized":false,"country":"NL","region":"europe","industry":"government"},"vendors":[],"summary":"The Dutch employee insurance agency UWV uses a risk scan on applications for unemployment benefit (WW) to signal cases where the applicant may have become unemployed through their own fault, which would remove the entitlement. The scan combines several data points, never a single characteristic, and uses no personal characteristics such as origin, gender or age; high risk applications are offered to staff for a fuller investigation, and staff decide. Specialists check the input data quality monthly and whether the development population is still representative, and an independent party reviews the work when the scan is developed further. Randomly chosen applications (30 percent, according to the register) are added to the scan's selection, so staff do not know which cases the scan flagged. UWV carried out a DPIA and an ethical impact assessment. In use since August 2022; no outcome figures are published.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["nl"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://algoritmes.overheid.nl/nl/algoritme/zb000117/15759870/risicoscan-verwijtbare-werkloosheid","title":"Risicoscan Verwijtbare Werkloosheid, Algoritmeregister","publisher":"Algoritmeregister van de Nederlandse overheid","date":"2025-12-15"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"uwv-unemployment-benefit-risk-scan"},{"title":"US Department of Veterans Affairs: e-VA digital assistant for appointment reminders and rescheduling (retired)","useCases":["outbound-reminder-and-confirmation-agent"],"organization":{"name":"US Department of Veterans Affairs, Veterans Benefits Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"SaraWorks","role":"platform"}],"summary":"The Veteran Readiness and Employment service of the Veterans Benefits Administration used e-VA, a digital assistant acquired from SaraWorks, to take routine follow up off vocational rehabilitation counselors. It created reminders for appointments, grades and receipts and sent routine messages by text or email; participants could reply to schedule and reschedule appointments, respond to reminders and send documents. The 2024 federal AI inventory lists it in operation since June 2020 and classifies it as both rights and safety impacting; the 2025 inventory lists it as retired. No results were published.","stage":"paused","year":2020,"channels":["sms","email"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2024-Federal-AI-Use-Case-Inventory","title":"2024 Federal AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI inventory (VA entry Electronic Virtual Assistant (e-VA))","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (VA entry electronic Virtual Assistant (e-VA), VA-24-3824)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"va-evirtual-assistant-appointment-reminders"},{"title":"Veterans Health Administration: generative AI categorizing purchase order lines for executive spend analysis","useCases":["procurement-spend-classification"],"organization":{"name":"Veterans Health Administration","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[],"summary":"The Veterans Health Administration uses generative AI to enrich purchase order data from its Integrated Funds Control, Accounting, and Procurement system (IFCAP) by assigning each line item to predefined spend categories that are more useful for analysis. The result feeds an executive dashboard with total spend, line items, categories and trends that can be filtered by region, budget object code, fund control point and vendor, so that network and national leaders can find areas to drive spending efficiencies. The inventory lists it as deployed and not high impact; no outcome figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry VA-25-364, Executive Spend Analysis)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"va-executive-spend-analysis"},{"title":"US Department of Veterans Affairs: test case generation for Salesforce CRM apps with VA GPT (planned)","useCases":["requirements-to-test-case-generation"],"organization":{"name":"US Department of Veterans Affairs, Office of Information and Technology","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Provar Manager","role":"platform"}],"summary":"The VA Office of Information and Technology reports a pre deployment generative AI use case that integrates Provar Manager with VA GPT to generate test cases automatically for the Salesforce CRM applications that support Department of Veterans Affairs services. The stated aim is less manual effort in writing tests and a closer match between requirements and deliverables. No outcome is published.","stage":"announced","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://github.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory","title":"2025 Federal Agency AI Use Case Inventory","publisher":"Office of Management and Budget (GitHub)"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry VA-25-5958, Integration of Provar Manager with VA GPT)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"va-provar-test-case-generation"},{"title":"Vanguard: Digital Advisor goal based portfolios with automatic rebalancing","useCases":["goal-based-financial-planning-assistant","portfolio-drift-monitoring-and-rebalancing","suitability-assessment-assistant"],"organization":{"name":"Vanguard","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"Vanguard","role":"in-house"}],"summary":"Vanguard Digital Advisor is an all digital advice service that gathers a client's goals, time horizon and risk tolerance, uses an algorithm to build a portfolio for them, and keeps it on track with ongoing monitoring. It rebalances when a portfolio drifts more than 5% from the recommended allocation and adjusts holdings when a client adds goals. Vanguard discloses limits of the automated assessment, for example that it does not assess the suitability of selling existing holdings, and it labels its projections and goal forecasts as hypothetical and educational, not guarantees. The service was live by 2020: a Vanguard page archived in December 2020 refers to Digital Advisor clients who enrolled before 1 October 2020.","stage":"production","year":2020,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://investor.vanguard.com/advice/digital-advisor","title":"Vanguard Digital Advisor","publisher":"Vanguard"},{"url":"https://web.archive.org/web/20201231135303/https://investor.vanguard.com/advice/digital-advisor","title":"Vanguard Digital Advisor (Wayback Machine capture of 31 December 2020)","publisher":"Vanguard","date":"2020-12-31"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"vanguard-digital-advisor"},{"title":"Veterans Benefits Administration: AI assistant builds eLearning for claims processors","useCases":["training-content-generation"],"organization":{"name":"Veterans Benefits Administration","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"Articulate","role":"platform"}],"summary":"The Veterans Benefits Administration uses the AI Assistant in Articulate 360 to create interactive eLearning for its claims processors more efficiently. The assistant generates text to voice audio, images, outlines, polls, quizzes and sounds for courses. The inventory entry says engaging eLearning reduces instructor burden and classroom time, that efficient training development reduces instructional design burden amid decreased hiring abilities, and that the time saved in developing training brings cost savings, without a figure. The inventory lists it as deployed; no outcome figures are published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 individually reported AI use cases (entry VA-25-1471, Articulate 360: AI Assistant)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"vba-articulate-ai-assistant-elearning"},{"title":"Veolia: AI invoice capture for a shared service centre serving 30 entities","useCases":["supplier-invoice-processing"],"organization":{"name":"Veolia","anonymized":false,"country":"FR","region":"europe","industry":"energy-and-utilities"},"vendors":[{"name":"Rossum","role":"platform"},{"name":"UiPath","role":"platform"},{"name":"InnovationPath","role":"integrator"}],"summary":"A Veolia shared service centre that posts supplier invoices for 30 group entities rebuilt the process around central email inboxes, a UiPath robot built by InnovationPath and Rossum's AI data capture, while moving 60,000 suppliers to paperless invoicing. The most technically skilled of the former data entry clerks now review extractions and handle exceptions as \"AI Associates\"; output goes as a standard EDI message to each entity's ERP. The vendor reports an eightfold speed up in the AI Associates' processing.","stage":"scaled","year":2022,"channels":["email","internal-tools"],"languages":[],"metrics":[{"kpi":"productivity-gain","value":87.5,"unit":"percent","qualifier":"up-to","period":"processing time savings of the AI Associates","claimant":"vendor","quote":"So far, the AI Associates were able to speed up their processing by 8x, with time-savings efficiency reaching up to 87.5%.","sourceUrl":"https://rossum.ai/customer-stories/veolia/"}],"outcomeDisclosed":true,"sources":[{"url":"https://rossum.ai/customer-stories/veolia/","title":"Customer Story - Veolia - Rossum.ai","publisher":"Rossum","archivedUrl":"https://web.archive.org/web/20220628152622/https://rossum.ai/customer-stories/veolia/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"veolia-ssc-invoice-processing"},{"title":"Verdantas: Copilot Studio agent that supports proposal development across a merged firm","useCases":["rfp-and-proposal-response-drafting"],"organization":{"name":"Verdantas","anonymized":false,"country":"US","region":"north-america","industry":"professional-services"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Verdantas, an environmental science, engineering and consulting firm of about 2,500 professionals formed through acquisitions, built a technical resource agent in Microsoft Copilot Studio on data unified in Microsoft Fabric. For proposal teams the agent summarizes RFPs, retrieves previous and similar proposals, surfaces market specific marketing material and finds staff with the right skills and licences. Multi agent orchestration made the agent answer up to twice as fast; no proposal outcome is disclosed.","stage":"production","year":2026,"channels":["microsoft-teams"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/26233-verdantas-microsoft-copilot-studio","title":"Verdantas builds agents for proposal development, contract management, and collaboration using Microsoft Copilot Studio","publisher":"Microsoft","date":"2026-03-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"verdantas-copilot-studio-proposal-agent"},{"title":"Verizon: AI powered Verizon Assistant and Customer Champion service model","useCases":["bill-explanation-and-billing-dispute-agent","plan-upgrade-and-sales-assistant","order-to-activation-and-esim-onboarding-assistant"],"organization":{"name":"Verizon","anonymized":false,"country":"US","region":"north-america","industry":"telecommunications"},"vendors":[],"summary":"In June 2025 Verizon announced a customer experience overhaul: a Customer Champion who owns complex issues end to end, drawing on Google Cloud AI including Gemini models, 24/7 live chat with human agents, and a new My Verizon app, which includes an AI powered Verizon Assistant, in which customers can become a customer, manage upgrades, add lines and ask billing questions. Verizon describes the assistant as voice enabled for mobile customers. Verizon's chief executive framed the programme as a way to build loyalty and improve retention. No outcome figures were published.","stage":"production","year":2025,"channels":["mobile-app","voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.verizon.com/about/news/verizon-launches-industry-leading-ai-powered-customer-experience","title":"Verizon, America's Most Reliable 5G Network, Launches Industry-Leading, AI Powered Customer Experience Innovations","publisher":"Verizon","date":"2025-06-24"},{"url":"https://www.verizon.com/about/customer-experience","title":"Verizon Customer Experience","publisher":"Verizon"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"verizon-ai-customer-experience-transformation"},{"title":"Verizon Business: AI guided selling engine for large accounts and SMBs","useCases":["business-connectivity-quoting-and-service-assistant"],"organization":{"name":"Verizon","anonymized":false,"country":"US","region":"north-america","industry":"telecommunications"},"vendors":[{"name":"Pega","role":"platform"}],"summary":"Verizon's business unit built a single Next Best X decisioning engine on Pega across five lines of business and more than 200 products and offers. It brings AI guided selling into the tools that front line sellers and business customers already use, across sales cycles that range from six month account based programmes to real time transactional sales for small businesses. Pega reports a 15% improvement in win rate, a doubling of the attach rate for value added services and a 10 to 20% reduction in handling time.","stage":"scaled","year":2022,"channels":["agent-desktop","internal-tools"],"languages":["en"],"metrics":[{"kpi":"conversion-rate-uplift","value":15,"unit":"percent","qualifier":"exact","period":"sales win rate","claimant":"vendor","quote":"15% improvement in win rate","sourceUrl":"https://www.pega.com/customers/verizon-customer-decision-hub"},{"kpi":"handling-time-reduction","value":20,"unit":"percent","qualifier":"up-to","claimant":"vendor","quote":"10-20% improvement in handling time reduction","sourceUrl":"https://www.pega.com/customers/verizon-customer-decision-hub"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.pega.com/customers/verizon-customer-decision-hub","title":"Verizon builds B2B-grade customer engagement platform","publisher":"Pega","date":"2022-08-12"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"verizon-business-next-best-x-engine"},{"title":"Verizon: machine learning to prevent fiber cuts from excavation","useCases":["predictive-network-maintenance"],"organization":{"name":"Verizon","anonymized":false,"country":"US","region":"north-america","industry":"telecommunications"},"vendors":[{"name":"Verizon (proprietary technology)","role":"in-house"}],"summary":"Verizon uses artificial intelligence and machine learning on the 811 call before you dig requests it receives to identify the excavations most likely to damage its underground fiber. The model weighs historical and current activity at the location and the past record of the excavator on site, and high risk digs trigger preventive steps such as extra communication with the excavator. The solution is integrated with Verizon's 811 system; Verizon describes the potential benefit but has not published a measured reduction in fiber cuts.","stage":"production","year":2024,"channels":[],"languages":[],"metrics":[{"kpi":"interactions-handled","value":10000000,"unit":"count","qualifier":"at-least","period":"per year, 811 dig requests screened","claimant":"organization","quote":"Verizon is utilizing advanced artificial intelligence (AI) and machine learning techniques to sort through over ten million 811 dig requests annually to identify high-risk excavations.","sourceUrl":"https://www.verizon.com/about/news/verizon-uses-ai-machine-learning-prevent-fiber-cuts"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.verizon.com/about/news/verizon-uses-ai-machine-learning-prevent-fiber-cuts","title":"Verizon uses AI & machine learning to prevent fiber cuts","publisher":"Verizon","date":"2024-08-07"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"verizon-fiber-cut-prevention"},{"title":"VERMEG: AI commitment management cuts on demand AWS costs","useCases":["cloud-cost-optimization-agent"],"organization":{"name":"VERMEG","anonymized":false,"country":"NL","region":"europe","industry":"technology"},"vendors":[{"name":"nOps","role":"platform"}],"summary":"VERMEG, a provider of software solutions to the worldwide financial services industry, used nOps to manage its AWS commitments (Reserved Instances) with a buyback guarantee across its accounts. nOps uses AI and analytics to monitor usage continuously and automatically adjusts those commitments to match it, instead of VERMEG locking in a fixed multi year forecast. Over ten months VERMEG cut its on demand AWS costs by more than 39% with no change to its own infrastructure or configuration, and gained continuous visibility into cloud spend.","stage":"production","year":2024,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.nops.io/casestudies/nops-customer-case-study-vermeg/","title":"Vermeg Partners with nOps to Reduce On-Demand Costs by More Than 39%","publisher":"nOps","archivedUrl":"https://web.archive.org/web/20240719222357/https://www.nops.io/casestudies/nops-customer-case-study-vermeg/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-28"},"grade":"C","id":"vermeg-cloud-cost-reduction"},{"title":"U.S. Department of Veterans Affairs: AI crisis call simulations for Veterans Crisis Line responders","useCases":["conversation-roleplay-training"],"organization":{"name":"U.S. Department of Veterans Affairs","anonymized":false,"country":"US","region":"north-america","industry":"government"},"vendors":[{"name":"ReflexAI","role":"platform"}],"summary":"The Veterans Crisis Line trains new crisis responders with ReflexAI simulations in which generative AI plays eight Veteran personas, each with its own configured motivation and crisis, so trainees can practise in a low risk setting before live calls. Each simulated call produces a scoring summary of strengths and areas of growth. The tools launched with the first full cohort of new trainees in May 2024, and VA's 2025 AI use case inventory lists the system as deployed and not high impact.","stage":"production","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://department.va.gov/ai/wp-content/uploads/sites/26/2026/04/VA-AI-Use-Case-Inventory-2025-Web-Compliance-Updates.xlsx","title":"VA AI Use Case Inventory 2025 (entry VA-24-2348, ReflexAI)","publisher":"U.S. Department of Veterans Affairs"},{"url":"https://news.va.gov/133911/ai-technology-is-helping-crisis-line-responders/","title":"AI technology is helping crisis line responders","publisher":"VA News","date":"2024-09-02"},{"url":"https://www.missiondaybreak.net/2024-team-spotlights/reflexai/","title":"ReflexAI, Mission Daybreak team spotlight","publisher":"Mission Daybreak (VA suicide prevention challenge)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"veterans-crisis-line-reflexai-training-simulations"},{"title":"Veterans Health Administration: Ambient Scribe for all primary care providers","useCases":["ambient-clinical-documentation"],"organization":{"name":"US Department of Veterans Affairs, Veterans Health Administration","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Abridge","role":"platform"},{"name":"Knowtex","role":"platform"}],"summary":"The 2025 federal AI use case inventory lists Abridge and Knowtex ambient scribe pilots at VA, both flagged as high impact. As of June 2026 VHA has deployed Ambient Scribe to all Patient Aligned Care Team primary care providers, including physicians, physician assistants and advanced practice nurses, after a phased rollout that began at 10 VA medical centers. With the Veteran's verbal consent the tool drafts the progress note from the conversation; the provider reviews and edits it before signing it into the record, and Veterans can opt out at any time, even mid visit. VA plans to extend it to selected outpatient specialty care. No outcome figures are published by VA.","stage":"scaled","year":2026,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://news.va.gov/148010/ambient-scribe-reimagining-va-clinic-experience/","title":"Ambient Scribe reimagines the VA clinic experience","publisher":"VA News","date":"2026-07-29"},{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"veterans-health-administration-ambient-scribe"},{"title":"Veterans Health Administration: computer assisted coding with automatic code suggestions","useCases":["medical-coding-automation"],"organization":{"name":"US Department of Veterans Affairs, Veterans Health Administration","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Solventum (formerly 3M)","role":"platform"}],"summary":"The Veterans Health Administration reports in the 2025 federal AI use case inventory that it uses the Solventum (formerly 3M) 360 Encompass computer assisted coding system to suggest ICD-10-CM, CPT and HCPCS codes from the clinical documentation of each encounter. Medical coders review, validate and select the codes, so the AI speeds up the coder rather than coding on its own. The inventory lists the use as deployed, classifies it as high impact and publishes no outcome figures.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2025-Federal-Agency-AI-Use-Case-Inventory/main/Data/2025_individually_reported_AI_use_cases.csv","title":"2025 federal agency AI use case inventory, individually reported use cases (raw data)","publisher":"Office of Management and Budget (GitHub)"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"veterans-health-administration-computer-assisted-coding"},{"title":"Virgin Media O2: AI voice agent for broadband fault calls","useCases":["device-and-connectivity-troubleshooting-agent"],"organization":{"name":"Virgin Media O2","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Sierra","role":"platform"}],"summary":"In September 2026 Virgin Media O2 began a gradual rollout of an AI voice agent, built with Sierra, that handles selected routine broadband fault calls, which Virgin Media O2 says are a small proportion of total call volumes. Human advisors stay available to every caller, calls are monitored by Virgin Media O2 teams, and customer feedback is used to judge performance before the agent is extended. The launch sits alongside a specialist team of more than 500 agents for complex and sensitive cases and the Lumi AI advisor tool. No results have been published yet.","stage":"pilot","year":2026,"channels":["voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://news.virginmediao2.co.uk/virgin-media-o2-launches-customer-service-ai-voice-agent-on-selected-calls-to-make-support-faster-and-easier-than-ever-before/","title":"Virgin Media O2 launches customer service AI voice agent on selected calls to make support faster and easier than ever before","publisher":"Virgin Media O2","date":"2026-09-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"virgin-media-o2-broadband-fault-voice-agent"},{"title":"Virgin Media O2: Call Defence AI scam and spam call labelling","useCases":["spam-and-scam-call-blocking"],"organization":{"name":"Virgin Media O2","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Hiya","role":"platform"}],"summary":"O2 launched Call Defence in November 2024 at no extra cost. Built with Hiya, it uses adaptive AI to analyse the behaviour of unknown numbers in real time and shows a warning label on the customer's screen for suspected scam or spam calls, while also blocking known fraudulent calls. It rolled out automatically on Android and on iOS 18 and later. By March 2026 it had labelled more than 1 billion calls, and O2 reports that calls labelled suspected scam are answered 42% less often and last 89% less time than unflagged calls.","stage":"scaled","year":2024,"channels":["voice","mobile-app"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":70000000,"unit":"count","qualifier":"approximately","period":"per month, calls labelled as suspected scam or spam (2026)","claimant":"organization","quote":"O2 first launched the service for its customers in November 2024, and today around 70 million calls every month are being labelled as suspected scam or spam.","sourceUrl":"https://news.virginmediao2.co.uk/ai-helps-virgin-media-o2-detect-and-flag-1-billion-suspected-scam-and-spam-calls-to-customers/"},{"kpi":"interactions-handled","value":1000000000,"unit":"count","qualifier":"at-least","period":"cumulative, November 2024 to March 2026","claimant":"organization","quote":"Virgin Media O2 has today reached a major milestone in its fight against fraudsters, after using AI to successfully label more than 1 billion suspected scam and spam calls to O2 customers.","sourceUrl":"https://news.virginmediao2.co.uk/ai-helps-virgin-media-o2-detect-and-flag-1-billion-suspected-scam-and-spam-calls-to-customers/"}],"outcomeDisclosed":true,"sources":[{"url":"https://news.virginmediao2.co.uk/ai-helps-virgin-media-o2-detect-and-flag-1-billion-suspected-scam-and-spam-calls-to-customers/","title":"AI helps Virgin Media O2 detect and flag 1 billion suspected scam and spam calls to customers","publisher":"Virgin Media O2","date":"2026-03-06"},{"url":"https://news.virginmediao2.co.uk/o2-launches-free-ai-powered-scam-call-detection-service-to-help-combat-fraud-and-nuisance-calls/","title":"O2 launches free AI-powered scam call detection service to help combat fraud and nuisance calls","publisher":"Virgin Media O2","date":"2024-11-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"virgin-media-o2-call-defence"},{"title":"Virgin Media O2: Daisy, an AI voice persona that wastes scammers' time","useCases":["spam-and-scam-call-blocking"],"organization":{"name":"Virgin Media O2","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[],"summary":"O2 created Daisy, a lifelike AI voice persona of an elderly woman, trained with help from the scambaiter Jim Browning. Daisy combines several AI models to listen and respond to scam callers in real time without human input, telling long stories and giving false details so fraudsters spend their time on her instead of real victims. O2 says Daisy has kept fraudsters on calls for 40 minutes at a time. It was launched as part of O2's Swerve the Scammers awareness campaign and is a disruption and awareness tool rather than a protection service for individual customers.","stage":"pilot","year":2024,"channels":["voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://news.virginmediao2.co.uk/o2-unveils-daisy-the-ai-granny-wasting-scammers-time/","title":"O2 unveils Daisy, the AI granny wasting scammers’ time","publisher":"Virgin Media O2","date":"2024-11-14"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"virgin-media-o2-daisy-ai-scambaiter"},{"title":"Virgin Media O2: Lumi AI prompts for care, telesales and retention advisors","useCases":["churn-prediction-and-retention-offers","plan-upgrade-and-sales-assistant"],"organization":{"name":"Virgin Media O2","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Virgin Media O2","role":"in-house"}],"summary":"Virgin Media O2 built its own tool, Lumi AI, that analyses a live conversation and prompts the advisor with resolutions that worked for similar customers and with the products and services most likely to interest this customer. In July 2025 it was in pilot with a cohort of advisors in care, telesales and retentions, with a wider rollout planned. Alongside it the operator uses an AI contact centre service from Amazon Web Services that routes callers by their stated reason, software that flags potentially vulnerable customers, and automatic call summaries. The retention effect of Lumi AI has not been published.","stage":"pilot","year":2025,"channels":["agent-desktop","voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://news.virginmediao2.co.uk/virgin-media-o2-launches-new-ai-tools-to-supercharge-customer-services-and-better-help-most-vulnerable-customers/","title":"Virgin Media O2 launches new AI tools to supercharge customer services and better help most vulnerable customers","publisher":"Virgin Media O2","date":"2025-07-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"virgin-media-o2-lumi-ai-advisor-assistant"},{"title":"Virgin Voyages: agents and generative media for personalized campaigns","useCases":["personalized-marketing-at-scale"],"organization":{"name":"Virgin Voyages","anonymized":false,"country":"US","region":"north-america","industry":"travel-and-hospitality"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Virgin Voyages uses Google's generative video and image models to create thousands of personalized advertisements and emails in its brand voice, and has deployed specialized agents that turn behavioural signals into personalized campaign actions. Google Cloud reports that campaign creation time fell by 40%.","stage":"production","year":2025,"channels":["email"],"languages":[],"metrics":[{"kpi":"processing-time-reduction","value":40,"unit":"percent","qualifier":"exact","period":"campaign creation time","claimant":"vendor","quote":"Virgin Voyages have already deployed more than 1,000 specialized agents to reduce campaign creation times by 40% by turning behavioral signals into personalized action.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","date":"2026-04-22"},{"url":"https://www.virginvoyages.com/terms-and-conditions","title":"Website Terms & Conditions","publisher":"Virgin Voyages"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"virgin-voyages-personalized-campaigns"},{"title":"Visa: Decision Manager machine learning fraud screening for merchants and acquirers","useCases":["real-time-fraud-scoring"],"organization":{"name":"Visa","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[{"name":"Visa","role":"in-house"}],"summary":"Decision Manager is Visa's machine learning fraud management platform for merchants and acquirers. It gives each transaction a risk score from 0 to 99 drawn from hundreds of real time data points and automates the accept, review or reject decision. Visa reports that almost all transactions it screened in 2023 were resolved automatically and that active users cut manual reviews, which is the triage workload for fraud teams.","stage":"scaled","year":2023,"channels":["api"],"languages":[],"metrics":[{"kpi":"automation-rate","value":98.7,"unit":"percent","qualifier":"exact","period":"2023","claimant":"organization","quote":"In 2023, Decision Manager screened 3.2 billion transactions and prevented an estimated $33 billion in potential fraud losses — with 98.7% of all transactions processed through Decision Manager resolved automatically by AI.","sourceUrl":"https://corporate.visa.com/en/solutions/visa-protect/insights/ai-fraud-detection.html"},{"kpi":"interactions-handled","value":3200000000,"unit":"count","qualifier":"exact","period":"2023","claimant":"organization","quote":"In 2023, Decision Manager screened 3.2 billion transactions and prevented an estimated $33 billion in potential fraud losses — with 98.7% of all transactions processed through Decision Manager resolved automatically by AI.","sourceUrl":"https://corporate.visa.com/en/solutions/visa-protect/insights/ai-fraud-detection.html"},{"kpi":"alert-volume-reduction","value":25,"unit":"percent","qualifier":"at-least","period":"active users, manual review reduction","claimant":"organization","quote":"For active users, Decision Manager has helped reduce manual reviews by 25% or more, freeing fraud teams to focus on complex or high-value cases rather than routine screening.","sourceUrl":"https://corporate.visa.com/en/solutions/visa-protect/insights/ai-fraud-detection.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://corporate.visa.com/en/solutions/visa-protect/insights/ai-fraud-detection.html","title":"AI solutions for fraud prevention and detection","publisher":"Visa"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"visa-decision-manager"},{"title":"Visa: AI dispute resolution services for issuers, acquirers and merchants","useCases":["card-dispute-and-chargeback-intake","chargeback-and-representment"],"organization":{"name":"Visa","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[{"name":"Visa","role":"in-house"}],"summary":"In April 2026 Visa announced six new and enhanced dispute resolution tools. For issuers and acquirers they include Dispute Intelligence (predictive models that support case by case decisions, generally available), Dispute Doc Analyzer (AI summaries of merchant documents for issuer analysts and auto populated questionnaires for acquirers) and Visa Dispute Case Manager, which unifies dispute workflows from intake to resolution. Merchant tools cover pre dispute handling, generative AI representment responses and Compelling Evidence 3.0 to reduce friendly fraud. Several tools are still in pilot or planned for late 2026; no outcome figures were disclosed.","stage":"production","year":2026,"channels":["api","agent-desktop"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22261.html","title":"Visa Unveils New Services to Modernize Dispute Resolution Process","publisher":"Visa","date":"2026-04-01"},{"url":"https://www.cnbc.com/2026/04/01/visa-ai-tools-dispute-management.html","title":"Visa launches new AI tools to manage the charge dispute process","publisher":"CNBC","date":"2026-04-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"visa-dispute-resolution-services"},{"title":"Visa: Intelligent Commerce pilots with agent initiated consumer and B2B purchases","useCases":["agentic-payment-initiation"],"organization":{"name":"Visa","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[],"summary":"Visa reported in December 2025 that hundreds of controlled, real world agent initiated transactions had been completed with partners in its Visa Intelligent Commerce programme. In the United States, closed beta pilots with agent enablers Skyfire, Nekuda, PayOS and Ramp executed consumer purchases and, in Ramp's case, corporate bill payments by card. In the UAE, Visa is working with Aldar so customers can use AI agents to pay recurring fees such as real estate service charges. In its June 2026 announcement with OpenAI, Visa says transactions will operate within user permissions such as spending limits, merchant categories or required approvals, on tokenized credentials. No outcome figures are disclosed.","stage":"pilot","year":2025,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.21961.html","title":"Visa and Partners Complete Secure AI Transactions, Setting the Stage for Mainstream Adoption in 2026","publisher":"Visa","date":"2025-12-18"},{"url":"https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22496.html","title":"Visa Partners with OpenAI to Power the Next Generation of AI Commerce","publisher":"Visa","date":"2026-06-10"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"visa-intelligent-commerce-agent-transactions"},{"title":"Visa: AI and machine learning to find merchants breaking network rules","useCases":["merchant-underwriting-and-risk-monitoring"],"organization":{"name":"Visa","anonymized":false,"country":"US","region":"global","industry":"payments"},"vendors":[{"name":"Visa","role":"in-house"}],"summary":"Visa monitors merchant activity across its network for illegal commerce and acts through the acquirer that holds the merchant relationship; acquirers must remove merchants that cannot comply with Visa's rules and applicable law. Visa says that, using its AI tools and machine learning models, it saw a fivefold increase in acquirer remediation and terminations for merchant noncompliance between 2020 and 2024. Visa also monitors acquirers that serve high risk categories to check that their controls work.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"detection-rate-improvement","value":5,"unit":"multiplier","qualifier":"exact","period":"2020 to 2024; counts acquirer remediation and termination actions, a proxy for detection rather than a measured detection rate","baseline":"acquirer remediation and terminations for merchant noncompliance in 2020","claimant":"organization","quote":"Using our AI tools and machine learning models, we have seen a 5x increase in acquirer remediation and terminations for merchant noncompliance between 2020 and 2024.","sourceUrl":"https://corporate.visa.com/en/about-visa/visa-network-integrity.html"}],"outcomeDisclosed":true,"sources":[{"url":"https://corporate.visa.com/en/about-visa/visa-network-integrity.html","title":"Visa Network Integrity","publisher":"Visa"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"visa-network-integrity-merchant-monitoring"},{"title":"VitalityHealth: automated QA of every advisor call with speech analytics","useCases":["call-quality-and-compliance-monitoring"],"organization":{"name":"VitalityHealth","anonymized":false,"country":"GB","region":"europe","industry":"insurance"},"vendors":[{"name":"CallMiner","role":"platform"},{"name":"Davies Consulting","role":"integrator"}],"summary":"VitalityHealth, a UK health insurer with more than 550 customer service advisors and over 1 million calls a year, moved from manual to automated quality assurance with CallMiner, run as a managed service by Davies Consulting. Every call is analysed and assessed in three areas: regulatory, service excellence (tone, empathy, how the call opened and closed) and process assurance, and results reach the coaching system within 24 hours. No quantified outcome is published.","stage":"scaled","year":2020,"channels":["voice"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://callminer.com/customers/stories/davies-consulting-callminer-partner-to-automate-quality-assurance-for-vitalityhealth","title":"Davies Consulting and CallMiner partner to automate quality assurance for VitalityHealth","publisher":"CallMiner"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"vitalityhealth-automated-quality-assurance"},{"title":"Vituity: AI assistant for IT and HR requests in a physician group","useCases":["it-service-desk-resolution-agent","hr-and-policy-assistant"],"organization":{"name":"Vituity","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Moveworks","role":"platform"}],"summary":"Vituity, a healthcare organization owned and led by a partnership of nearly 5,000 physicians, deployed a Moveworks AI assistant called Otto in Microsoft Teams in April 2020, connected to ServiceNow, Okta and internal knowledge. It started with IT workflows such as password resets, account provisioning and software access, then expanded into HR questions and other operational domains. The vendor reports that the average time to close issues fell by one full business day and that first line help desk capacity was freed.","stage":"scaled","year":2020,"channels":["microsoft-teams"],"languages":["en"],"metrics":[{"kpi":"productivity-gain","value":40,"unit":"percent","qualifier":"exact","period":"level 1 help desk capacity","claimant":"vendor","quote":"It now absorbs a significant share of routine IT and HR requests — including password resets, software access, HR questions, and account provisioning— freeing up 40% of level 1 help-desk capacity.","sourceUrl":"https://www.moveworks.com/us/en/customers/vituity-helps-physicians-with-moveworks-proactive-it-support"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.moveworks.com/us/en/customers/vituity-helps-physicians-with-moveworks-proactive-it-support","title":"Vituity helps physicians with proactive IT support","publisher":"Moveworks"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"vituity-it-and-hr-assistant"},{"title":"Vodafone Business: AI powered service management with ServiceNow","useCases":["business-connectivity-quoting-and-service-assistant"],"organization":{"name":"Vodafone Business","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"ServiceNow","role":"platform"}],"summary":"Vodafone Business and ServiceNow agreed a five year collaboration to run service management for business customers on ServiceNow's telecom service assurance products with agentic AI. Vodafone says it will give a single view of each customer's networks and applications, and that it can detect and fix service anomalies within minutes rather than hours, using AI and machine learning to predict, minimise and manage service interruptions, and to route customers to the right department first time across online, email and phone. Vodafone reports that an initial deployment in Ireland raised digital channel use by 45%, and says satisfaction levels rose fourfold there.","stage":"production","year":2025,"channels":["web-chat","email","voice","internal-tools"],"languages":["en"],"metrics":[{"kpi":"customer-satisfaction-uplift","value":4,"unit":"multiplier","qualifier":"exact","period":"initial deployment in Ireland","claimant":"organization","quote":"An initial deployment in Ireland led to a 45% rise in customers using digital channels, boosting satisfaction levels by 4x.","sourceUrl":"https://www.vodafone.com/news/newsroom/technology/vodafone-business-and-service-now-collaborate-to-enhance-the-customer-experience-with-ai-powered-service-automation"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.vodafone.com/news/newsroom/technology/vodafone-business-and-service-now-collaborate-to-enhance-the-customer-experience-with-ai-powered-service-automation","title":"Vodafone Business and ServiceNow collaborate to enhance the customer experience with AI-powered service automation","publisher":"Vodafone","date":"2025-04-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"vodafone-business-servicenow-service-automation"},{"title":"Vodafone Germany: TOBi on WhatsApp, Apple Business Chat and SMS","useCases":["first-line-contact-centre-agent","device-and-connectivity-troubleshooting-agent"],"organization":{"name":"Vodafone Germany","anonymized":false,"country":"DE","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Genesys","role":"platform"},{"name":"IBM","role":"model-provider"}],"summary":"Vodafone Germany moved its messaging channels into one central team and put the TOBi chatbot in front of every WhatsApp, Apple Business Chat and SMS conversation, with a handover to a human agent on the same screen. TOBi understands more than 230 intents and can classify the photos and screenshots customers send, such as a bill or a router with a flashing red light. Genesys reports that the share of inquiries resolved by AI (which it calls first contact success) rose from 16% at launch to 44%.","stage":"scaled","year":2020,"channels":["whatsapp","social-messaging","sms"],"languages":["de"],"metrics":[{"kpi":"containment-rate","value":44,"unit":"percent","qualifier":"exact","period":"share of messaging inquiries resolved by TOBi without a human","baseline":"16% at launch","claimant":"vendor","quote":"44% of inquiries now resolved by AI, up from 16%.","sourceUrl":"https://www.genesys.com/customer-stories/vodafone"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.genesys.com/customer-stories/vodafone","title":"Vodafone Germany unifies AI-based digital messaging to amplify the customer experience","publisher":"Genesys"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"vodafone-germany-tobi-messaging"},{"title":"Vodafone: machine learning trials in centralised self organizing networks","useCases":["network-planning-and-capacity-optimization"],"organization":{"name":"Vodafone","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Huawei","role":"platform"},{"name":"Cisco","role":"platform"}],"summary":"Vodafone ran early machine learning trials in centralised self organizing networks. In Germany, with Huawei, an algorithm found the optimal voice over LTE settings for 450 randomly chosen cells in four hours, a task Vodafone says would take an engineer around two and a half months. In Ireland, with Cisco, algorithms predicted where 3G traffic would peak in the following hour so the network could rebalance load between neighbouring cells; Vodafone reports an average 6 percent improvement in mobile download speed in initial results. Vodafone planned commercial use from its 2018/19 financial year.","stage":"pilot","year":2017,"channels":["api"],"languages":[],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://www.vodafone.com/news/newsroom/technology/ai-enabled-augmented-engineering-increases-network-optimisation","title":"AI enabled engineering increases network optimisation speed","publisher":"Vodafone","date":"2017-09-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"vodafone-machine-learning-son-trials"},{"title":"Vodafone: machine learning anomaly detection across its European mobile networks","useCases":["network-fault-triage-copilot","predictive-network-maintenance"],"organization":{"name":"Vodafone","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Nokia","role":"platform"},{"name":"Google Cloud","role":"platform"}],"summary":"Vodafone and Nokia jointly developed an Anomaly Detection Service, based on Nokia Bell Labs technology and running on Google Cloud, that detects and troubleshoots irregularities such as mobile site congestion, interference and unexpected latency before they affect customers. After an initial deployment on more than 60,000 4G cells in Italy, it was being rolled out across Vodafone's European network in July 2021, with all European markets planned by early 2022 and plans to apply it later to 5G and core networks. Vodafone expected around 80 percent of its anomalous mobile network issues and capacity demands to be detected and addressed automatically; no measured result was published.","stage":"production","year":2021,"channels":["internal-tools","api"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.nokia.com/newsroom/nokia-and-vodafone-harness-machine-learning-on-google-cloud-to-detect-network-anomalies/","title":"Nokia and Vodafone harness machine learning on Google Cloud to detect network anomalies","publisher":"Nokia","date":"2021-07-20","archivedUrl":"https://web.archive.org/web/2026/https://www.nokia.com/newsroom/nokia-and-vodafone-harness-machine-learning-on-google-cloud-to-detect-network-anomalies/"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"vodafone-nokia-network-anomaly-detection"},{"title":"Vodafone: Scam Signal network data service against impersonation fraud","useCases":["telecom-fraud-detection","scam-payment-interception"],"organization":{"name":"Vodafone","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[],"summary":"Vodafone Carrier Services launched Scam Signal, an API that analyses real time network data during a live bank transaction to detect social engineering behind authorised push payment fraud, so banks can stop fraudulent transfers as they happen. It sits in Vodafone's Identity Hub next to the SIM Swap and Number Verify APIs, which use CAMARA open standards. JT Group, working with FICO, was the first channel partner to offer it. In a three month pilot with a UK bank that Vodafone does not name, scam detection improved by 30%.","stage":"production","year":2024,"channels":["api"],"languages":[],"metrics":[{"kpi":"detection-rate-improvement","value":30,"unit":"percent","qualifier":"exact","period":"three month pilot with a UK bank","claimant":"organization","quote":"Scam detection using this service improved by 30% after only three months of a successful pilot with a leading UK bank.","sourceUrl":"https://www.vodafone.com/news/newsroom/technology/vodafone-business-launches-scam-signal-to-defend-against-impersonation-fraud"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.vodafone.com/news/newsroom/technology/vodafone-business-launches-scam-signal-to-defend-against-impersonation-fraud","title":"Vodafone Business launches scam signal to defend against impersonation fraud","publisher":"Vodafone","date":"2024-04-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"vodafone-scam-signal"},{"title":"Vodafone: SuperTOBi generative AI assistant and SuperAgent","useCases":["bill-explanation-and-billing-dispute-agent","first-line-contact-centre-agent"],"organization":{"name":"Vodafone","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Vodafone rebuilt its TOBi chatbot on Azure OpenAI as SuperTOBi, launched in Italy and Portugal, with Germany and Turkey announced to follow from July 2024 and other markets later that year. A companion SuperAgent helps human agents search the company knowledge base and, in Ireland, sends the human agent a summary of the online customer conversation so customers do not repeat themselves. Vodafone reports that initial tests at one of its call centres showed a 50% improvement in first time resolution of critical journeys such as complex billing inquiries, and that in Portugal first time resolution on appointment booking rose from 15% to 60%, with billing journeys being added next.","stage":"production","year":2024,"channels":["web-chat"],"languages":["it","pt"],"metrics":[{"kpi":"first-contact-resolution","value":60,"unit":"percent","qualifier":"exact","period":"appointment booking journey, Vodafone Portugal","baseline":"15% before SuperTOBi","claimant":"organization","quote":"As a result, the first-time resolution rate has increased from 15% to 60% and Vodafone’s online net promoter scores (where respondents are asked to rate their experience) improved by 14 points to 64 points – anything above 50 points is considered a strong result.","sourceUrl":"https://www.vodafone.com/news/newsroom/technology/meet-super-tobi-vodafone-s-new-generative-ai-virtual-assistant-now-serving-customers-in-multiple-countries"},{"kpi":"nps-change","value":14,"unit":"points","qualifier":"exact","period":"online NPS, Vodafone Portugal","claimant":"organization","quote":"As a result, the first-time resolution rate has increased from 15% to 60% and Vodafone’s online net promoter scores (where respondents are asked to rate their experience) improved by 14 points to 64 points – anything above 50 points is considered a strong result.","sourceUrl":"https://www.vodafone.com/news/newsroom/technology/meet-super-tobi-vodafone-s-new-generative-ai-virtual-assistant-now-serving-customers-in-multiple-countries"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.vodafone.com/news/newsroom/technology/meet-super-tobi-vodafone-s-new-generative-ai-virtual-assistant-now-serving-customers-in-multiple-countries","title":"Meet SuperTOBI, Vodafone's new Generative AI virtual assistant now serving customers in multiple countries","publisher":"Vodafone","date":"2024-07-04"},{"url":"https://www.vodafone.com/news/newsroom/technology/vodafone-supercharging-customer-experience-with-microsoft-s-gen-ai-tools","title":"Vodafone supercharging customer experience with Microsoft's GenAI tools","publisher":"Vodafone","date":"2024-05-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"vodafone-supertobi-generative-ai-assistant"},{"title":"Vodafone: TOBi virtual assistant across markets","useCases":["first-line-contact-centre-agent","device-and-connectivity-troubleshooting-agent","plan-upgrade-and-sales-assistant"],"organization":{"name":"Vodafone","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"TOBi is Vodafone's digital assistant on the website, the My Vodafone app, messaging and telephony, first launched in Italy and extended to 15 language versions. It handles billing questions, contract updates and simple technical troubleshooting, hands over to a live agent with a summary, and during the pandemic made the same sales offers as human agents, such as data boosts and upgrades. Microsoft reports that TOBi now fully resolves 70% of inquiries arriving through digital channels; an earlier Microsoft story quotes Vodafone on a 12% year on year fall in contacts to call centres after launch.","stage":"scaled","year":2024,"channels":["web-chat","mobile-app","social-messaging","voice"],"languages":["en","it"],"metrics":[{"kpi":"containment-rate","value":70,"unit":"percent","qualifier":"exact","period":"inquiries arriving through digital channels","claimant":"vendor","quote":"Currently, TOBi handles nearly 45 million customer calls a month, fully resolving 70% of customer inquiries coming through the company’s digital channels.","sourceUrl":"https://customers.microsoft.com/en-gb/story/1770174778560829849-vodafone-group-azure-telecommunications-en-united-kingdom"},{"kpi":"interactions-handled","value":45000000,"unit":"count","qualifier":"approximately","period":"per month","claimant":"vendor","quote":"Currently, TOBi handles nearly 45 million customer calls a month, fully resolving 70% of customer inquiries coming through the company’s digital channels.","sourceUrl":"https://customers.microsoft.com/en-gb/story/1770174778560829849-vodafone-group-azure-telecommunications-en-united-kingdom"},{"kpi":"contact-deflection","value":12,"unit":"percent","qualifier":"exact","period":"frequency of customer contacts to call centres, year over year after the TOBi launch","claimant":"organization","quote":"Since launching TOBi, we’ve reduced the frequency of customer contacts to call centers by 12 percent year-over-year","sourceUrl":"https://www.microsoft.com/en/customers/story/838350-vodafone-telecom-azure-cognitive-services"}],"outcomeDisclosed":true,"sources":[{"url":"https://customers.microsoft.com/en-gb/story/1770174778560829849-vodafone-group-azure-telecommunications-en-united-kingdom","title":"Vodafone amplifies call center innovation, customer service, and employee inclusion with Azure AI","publisher":"Microsoft"},{"url":"https://www.microsoft.com/en/customers/story/838350-vodafone-telecom-azure-cognitive-services","title":"Vodafone transforms its customer care strategy with digital assistant built on Azure Cognitive Services","publisher":"Microsoft"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"vodafone-tobi-virtual-assistant"},{"title":"Vodafone UK: always on next best action for retention","useCases":["churn-prediction-and-retention-offers"],"organization":{"name":"Vodafone UK","anonymized":false,"country":"GB","region":"europe","industry":"telecommunications"},"vendors":[{"name":"Pega","role":"platform"}],"summary":"Vodafone UK's Always on Marketing programme runs on Pega Customer Decision Hub. Customer interactions and events are processed as they happen and mapped to common intents, so one central system presents the next best action for each customer at every touchpoint instead of pushing products. Pega titles the case study as improving customer retention and says the programme was halfway through its transformation, but publishes no figures.","stage":"production","year":2023,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.pega.com/customers/vodafone-uk-customer-decision-hub","title":"Vodafone UK: Always-on marketing improves customer retention","publisher":"Pega"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"vodafone-uk-always-on-marketing"},{"title":"Volkswagen: ChatGPT in the IDA voice assistant across new and existing models","useCases":["in-car-ai-voice-assistant"],"organization":{"name":"Volkswagen","anonymized":false,"country":"DE","region":"europe","industry":"automotive"},"vendors":[{"name":"Cerence","role":"platform"},{"name":"OpenAI","role":"model-provider"}],"summary":"Volkswagen added ChatGPT to its IDA voice assistant through Cerence Chat Pro. IDA keeps handling vehicle functions such as infotainment, navigation and climate control; only questions its own system cannot answer are forwarded anonymously to ChatGPT, and the answer is read out in the familiar Volkswagen voice. ChatGPT gets no access to vehicle data, questions and answers are deleted immediately, and drivers can switch the online assistant off. It launched in 2024 in all new ID. models, the new Golf, Tiguan and Passat, in five languages, and Cerence reports that it reached cars already on the road by cloud update across Volkswagen, Cupra, Seat and Skoda.","stage":"scaled","year":2024,"channels":["voice"],"languages":["en","de","es","cs"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.volkswagen-group.com/en/articles/chatgpt-is-now-available-in-many-volkswagen-models-18464","title":"ChatGPT is now available in many Volkswagen models","publisher":"Volkswagen Group"},{"url":"https://investors.cerence.com/news-events/press-releases/detail/127/volkswagen-and-cerence-commence-roll-out-of-new-generative-ai-solutions-to-drivers","title":"Volkswagen and Cerence Commence Roll-Out of New Generative AI Solutions to Drivers","publisher":"Cerence","date":"2024-06-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"volkswagen-ida-voice-assistant-chatgpt"},{"title":"Volvo Group: AI document processing for invoices, credit notes and claims in service and financing","useCases":["intelligent-document-processing"],"organization":{"name":"Volvo Group","anonymized":false,"country":"SE","region":"europe","industry":"automotive"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Volvo Group's automation team built a document processing solution on Azure AI Document Intelligence for its service and financing businesses. It reads emails, digital and scanned PDFs and written bills, including stamps, photographs, handwritten notes over printed text and tables across pages, translates content between languages and outputs XML or CSV for the receiving division, with processing time and success rate tracked in a dashboard. Microsoft reports that the solution has saved 10,000 manual hours since launch, about 850 hours a month.","stage":"production","year":2023,"channels":["email","api"],"languages":[],"metrics":[{"kpi":"hours-saved","value":10000,"unit":"hours","qualifier":"exact","period":"since launch, about 850 hours per month","claimant":"vendor","quote":"Since launch, the company has saved 10,000 manual hours—about 850-plus manual hours per month.","sourceUrl":"https://www.microsoft.com/en/customers/story/1703814256939529124-volvo-group-automotive-azure-ai-services"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/1703814256939529124-volvo-group-automotive-azure-ai-services","title":"Volvo Group streamlines invoice and claims processing with Azure AI and AI Document Intelligence","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"volvo-group-document-intelligence"},{"title":"Walmart: AI forecasting that positions inventory across stores and fulfillment centers","useCases":["retail-demand-forecasting-and-replenishment"],"organization":{"name":"Walmart","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Walmart Global Tech","role":"in-house"}],"summary":"Walmart Global Tech describes how Walmart's supply chain systems use AI and forecasting models, drawing on signals such as historical sales, seasonality, local demand and weather, to decide which products are needed and where to position them across stores and fulfillment centers before customers order. Walmart says it combines AI foresight with human expertise to refine its demand forecasting, and once an order is placed the Walmart Fulfillment Engine takes over. Walmart gives no forecasting accuracy or inventory figures.","stage":"scaled","year":2025,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://tech.walmart.com/content/walmart-global-tech/en_us/blog/post/inside-the-ai-network-orchestrating-walmarts-fastest-holiday-deliveries-yet.html","title":"Inside the AI network orchestrating Walmart's fastest holiday deliveries yet","publisher":"Walmart Global Tech","date":"2025-12-08"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"walmart-ai-supply-chain-forecasting"},{"title":"Walmart: autonomous negotiation of supplier terms with tail end suppliers","useCases":["procurement-contract-review"],"organization":{"name":"Walmart","anonymized":false,"country":"US","region":"global","industry":"retail-and-ecommerce"},"vendors":[{"name":"Pactum","role":"platform"}],"summary":"Walmart uses a chatbot from Pactum to negotiate payment terms and price discounts with tail end suppliers, where buyers lack time to negotiate and around 20% of suppliers had signed standard terms that are often not negotiated. The HBR article describing it is co written by two sourcing leaders at Walmart International and two University of Arkansas professors. Walmart decides the acceptable negotiation trade offs and the bot negotiates and closes agreements within them; the article advises starting in indirect spend categories with pre approved suppliers and scaling by geography, category and use case. The authors report that, so far, the chatbot has closed agreements with 68% of suppliers approached. Pactum, the vendor, reports a 3% average gain across negotiations while extending payment terms by an average of 35 days. The first is a supplier agreement rate and the second a commercial gain on negotiated terms; neither measures how much of the process runs without human touch or what procurement costs to run.","stage":"production","year":2022,"channels":["web-chat","internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://hbr.org/2022/11/how-walmart-automated-supplier-negotiations","title":"How Walmart Automated Supplier Negotiations","publisher":"Harvard Business Review","date":"2022-11-08"},{"url":"https://pactum.com/clients","title":"Clients","publisher":"Pactum"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"walmart-autonomous-supplier-negotiation"},{"title":"Walmart: large language models that create and check product catalog data","useCases":["product-content-and-catalog-enrichment"],"organization":{"name":"Walmart","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[{"name":"Walmart Global Tech","role":"in-house"}],"summary":"Walmart uses several large language models to extract product attributes such as color, size and material from item descriptions and images and to create or improve catalog data. One model extracts the values and a second model, tuned on human validated labels, checks them; values for attributes above an accuracy threshold go into the catalog, and the rest are checked by the quality model, with specialists validating samples. Walmart's chief executive told investors in August 2024 that the work covered more than 850 million pieces of catalog data, and estimated that without generative AI the same work would have needed nearly 100 times the current headcount to finish in the same time.","stage":"scaled","year":2024,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":850000000,"unit":"count","qualifier":"at-least","period":"catalog data points created or improved, reported August 2024","claimant":"organization","quote":"We've used multiple large language models to accurately create or improve over 850 million pieces of data in the catalog.","sourceUrl":"https://corporate.walmart.com/content/dam/corporate/documents/newsroom/2024/08/15/walmart-releases-q2-fy25-earnings/corrected-walmart-inc-wmt-us-q2-2025-earnings-call-15-august-2024.pdf"}],"outcomeDisclosed":true,"sources":[{"url":"https://corporate.walmart.com/content/dam/corporate/documents/newsroom/2024/08/15/walmart-releases-q2-fy25-earnings/corrected-walmart-inc-wmt-us-q2-2025-earnings-call-15-august-2024.pdf","title":"Walmart Inc. Q2 FY25 earnings call, corrected transcript","publisher":"Walmart","date":"2024-08-15"},{"url":"https://tech.walmart.com/content/walmart-global-tech/en_us/blog/post/using-llms-to-manage-product-catalogs.html","title":"How Walmart uses LLMs to manage its massive product catalogs","publisher":"Walmart Global Tech","date":"2025-05-20"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"walmart-generative-ai-product-catalog"},{"title":"Walmart: Sparky generative AI shopping assistant","useCases":["conversational-shopping-assistant"],"organization":{"name":"Walmart","anonymized":false,"country":"US","region":"north-america","industry":"retail-and-ecommerce"},"vendors":[],"summary":"Walmart launched Sparky in June 2025 as an \"Ask Sparky\" button in its app across all categories. Sparky answers product questions, compares options, synthesizes reviews and recommends products for an occasion, and Walmart describes a roadmap toward reordering, service booking and multimodal input. It joins Walmart's earlier generative AI features for search, review summaries, product descriptions and comparisons. No outcome figures were published on the launch page.","stage":"scaled","year":2025,"channels":["mobile-app"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://corporate.walmart.com/news/2025/06/06/walmart-the-future-of-shopping-is-agentic-meet-sparky","title":"Walmart: The Future of Shopping Is Agentic. Meet Sparky.","publisher":"Walmart","date":"2025-06-06"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"walmart-sparky-shopping-assistant"},{"title":"Warner Bros. Discovery: AI captioning tool on Vertex AI","useCases":["audio-and-video-transcription-and-captioning"],"organization":{"name":"Warner Bros. Discovery","anonymized":false,"country":"US","region":"north-america","industry":"media-and-entertainment"},"vendors":[{"name":"Google Cloud (Vertex AI)","role":"platform"}],"summary":"Warner Bros. Discovery built an AI captioning tool on Vertex AI. Google Cloud reports that it delivered a 50% reduction in overall costs and cut the time to caption a file by 80% compared with manual captioning. No detail on languages, volumes or the review step is published.","stage":"production","year":2024,"channels":["api"],"languages":[],"metrics":[{"kpi":"cost-reduction","value":50,"unit":"percent","qualifier":"exact","baseline":"Overall costs before the AI captioning tool","claimant":"vendor","quote":"Warner Bros. Discovery built an AI captioning tool with Vertex AI, delivering a 50% reduction in overall costs and an 80% reduction in the time it takes to manually caption a file without the use of machine learning.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"},{"kpi":"processing-time-reduction","value":80,"unit":"percent","qualifier":"exact","baseline":"Manual captioning of a file without machine learning","claimant":"vendor","quote":"Warner Bros. Discovery built an AI captioning tool with Vertex AI, delivering a 50% reduction in overall costs and an 80% reduction in the time it takes to manually caption a file without the use of machine learning.","sourceUrl":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real world gen AI use cases from the world's leading organizations","publisher":"Google Cloud"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"warner-bros-discovery-ai-captioning-tool"},{"title":"The Washington Post: Ask The Post AI answers readers from its own reporting","useCases":["publisher-archive-answer-engine"],"organization":{"name":"The Washington Post","anonymized":false,"country":"US","region":"north-america","industry":"media-and-entertainment"},"vendors":[{"name":"The Washington Post","role":"in-house"}],"summary":"In November 2024 The Washington Post launched \"Ask The Post AI\", an experimental generative AI tool that answers readers' questions with summary answers and curated results drawn from articles its newsroom has published since 2016, ranked by relevance. To protect the integrity of the reporting, the tool does not serve an answer when it does not readily find a relevant article above a set threshold, even if one exists. It followed earlier experiments such as Climate Answers and AI article takeaways. In July 2026 Arc XP, the Post's technology business, launched a comparable answer layer, Ask The News, for other publishers.","stage":"production","year":2024,"channels":["web-chat"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.washingtonpost.com/pr/2024/11/07/washington-post-launches-ask-post-ai-new-search-experience/","title":"The Washington Post Launches \"Ask The Post AI,\" a New Search Experience","publisher":"The Washington Post","date":"2024-11-07","archivedUrl":"https://web.archive.org/web/20250724203918/https://www.washingtonpost.com/pr/2024/11/07/washington-post-launches-ask-post-ai-new-search-experience/"},{"url":"https://www.prnewswire.com/news-releases/publishers-are-losing-their-readers-to-ai-the-washington-posts-tech-arm-built-a-solution-302823034.html","title":"Publishers Are Losing Their Readers to AI. The Washington Post's Tech Arm Built a Solution","publisher":"Arc XP (PR Newswire)","date":"2026-07-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"washington-post-ask-the-post-ai"},{"title":"Waterdrop: Waterdrop Guardian AI suite for insurance sales, service, consultants and underwriting","useCases":["insurance-policy-servicing-agent","insurance-broker-and-agent-assistant"],"organization":{"name":"Waterdrop","anonymized":false,"country":"CN","region":"asia-pacific","industry":"insurance"},"vendors":[{"name":"Waterdrop","role":"in-house"}],"summary":"Waterdrop, a Chinese online insurance distribution and health services platform, runs a suite of AI applications called Waterdrop Guardian that either talk to users directly or support its online insurance consultants. In its second quarter 2025 results it reported that its AI Customer Service Agent resolved 60% of inquiries on first contact, that its Life Planner Copilot handled 300,000 product consultations for consultants, and that premiums facilitated by its AI Medical Insurance Expert rose 155% from the previous quarter (the release does not say whether this tool talks to users directly or supports consultants). It also launched KEYI.AI, a real time AI underwriting assistant for consultants. Its 2023 annual report describes an earlier LLM powered AI Insurance Consultant that was tested internally in medical insurance scenarios, an internal test rather than a sales deployment.","stage":"scaled","year":2025,"channels":["mobile-app","internal-tools"],"languages":["zh"],"metrics":[{"kpi":"first-contact-resolution","value":60,"unit":"percent","qualifier":"exact","period":"second quarter 2025, AI Customer Service Agent","claimant":"organization","quote":"‘AI Customer Service Agent’ resolved 60% of inquiries on first contact, enhancing user experience.","sourceUrl":"https://www.sec.gov/Archives/edgar/data/1823986/000110465925087282/tm2524976d1_ex99-1.htm"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sec.gov/Archives/edgar/data/1823986/000110465925087282/tm2524976d1_ex99-1.htm","title":"Waterdrop Inc. second quarter 2025 unaudited financial results (Form 6-K, Exhibit 99.1)","publisher":"Waterdrop via SEC EDGAR","date":"2025-09-04"},{"url":"https://www.sec.gov/Archives/edgar/data/1823986/000110465924051464/wdh-20231231x20f.htm","title":"Waterdrop Inc. annual report for 2023 (Form 20-F)","publisher":"Waterdrop via SEC EDGAR","date":"2024-04-25"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"waterdrop-guardian-ai-insurance-assistants"},{"title":"Weave: AI merchant risk triage and monitoring for Weave Payments","useCases":["merchant-underwriting-and-risk-monitoring"],"organization":{"name":"Weave Communications","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Coris","role":"platform"},{"name":"Stripe","role":"platform"}],"summary":"Weave, a customer experience and payments software company for dental, optometry, veterinary and other local healthcare practices, runs Weave Payments mainly on Stripe Connect. It replaced manual merchant reviews from spreadsheet exports with the Coris risk platform, which scores merchants and transactions, ranks an alert queue by risk and triggers workflows such as payout pauses and outreach. Coris reports that daily reviews fell from about 80 to 9 or 10 accounts, and that automated checks now cover nearly all merchants with meaningful volume, against about 3% of merchants reviewed manually before.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"alert-volume-reduction","value":89,"unit":"percent","qualifier":"approximately","period":"daily merchant reviews, within a week of switching primary triage to Coris","baseline":"about 80 accounts reviewed per day before Coris","claimant":"vendor","quote":"AI triage cut daily reviews by ~89% (from ~80/day to ~9–10/day), so analysts spend time where judgment matters most.","sourceUrl":"https://www.coris.ai/customer/how-weave-built-a-modern-ai-powered-risk-program-with-coris"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.coris.ai/customer/how-weave-built-a-modern-ai-powered-risk-program-with-coris","title":"How Weave Built a Modern, AI-Powered Risk Program with Coris","publisher":"Coris","archivedUrl":"https://web.archive.org/web/20251014131437/https://www.coris.ai/customer/how-weave-built-a-modern-ai-powered-risk-program-with-coris"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"weave-coris-merchant-risk-monitoring"},{"title":"Wells Fargo: patented adverse action methodology for machine learning credit risk models","useCases":["adverse-action-explanations"],"organization":{"name":"Wells Fargo","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Wells Fargo","role":"in-house"}],"summary":"Wells Fargo Bank patented a computer based credit evaluation system that pairs a machine learning credit risk model with an adverse action methodology: when the model denies an applicant, the system compares the applicant's characteristic values against anchor values taken from a top scoring population, calculates a replacement score for each characteristic, and ranks the characteristics to identify the principal adverse action factors for the denial. The United States Patent and Trademark Office granted the patent in October 2022 on an application Wells Fargo filed in October 2019. Separately, the trade publication Risk.net reported in August 2021 that a team of Wells Fargo researchers had begun deploying an explainability technique for its deep learning credit models, and a paper by six Wells Fargo model risk researchers proposed a related Shapley decomposition method for explaining adverse credit decisions. No outcome metric or notice volume is disclosed by any source.","stage":"pilot","year":2022,"channels":["api"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://patents.google.com/patent/US11475515B1/en","title":"US11475515B1: Adverse action methodology for credit risk models","publisher":"United States Patent and Trademark Office","date":"2022-10-18"},{"url":"https://www.risk.net/risk-management/7865541/wells-touts-new-explainability-technique-for-ai-credit-models","title":"Wells touts new explainability technique for AI credit models","publisher":"Risk.net","date":"2021-08-16"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"wells-fargo-adverse-action-methodology"},{"title":"Wells Fargo: retrieval tool for branch bankers on policies and procedures","useCases":["enterprise-knowledge-search"],"organization":{"name":"Wells Fargo","anonymized":false,"country":"US","region":"north-america","industry":"banking"},"vendors":[{"name":"Google Cloud","role":"platform"}],"summary":"Wells Fargo deployed a retrieval augmented tool for branch bankers that finds the relevant policies and procedures during customer interactions. Google Cloud reports that it reduced the workflow for query resolution by about 20%, without saying whether that means time, steps or effort. The bank uses reusable APIs on Apigee to scale generative AI across teams.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders","title":"Real-world gen AI use cases from the world's leading organizations","publisher":"Google Cloud","archivedUrl":"https://web.archive.org/web/20251027121348/https://cloud.google.com/transform/101-real-world-generative-ai-use-cases-from-industry-leaders"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"wells-fargo-branch-policy-retrieval"},{"title":"WellSpan Health: Ana, a generative AI voice agent for patient calls, scheduling and outreach","useCases":["patient-appointment-scheduling-and-reminders-agent","outbound-reminder-and-confirmation-agent"],"organization":{"name":"WellSpan Health","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Hippocratic AI","role":"platform"}],"summary":"WellSpan Health started with a Hippocratic AI voice agent that phones patients to close gaps in colorectal cancer screening, in English and Spanish, and supports low risk patients before and after a scheduled colonoscopy, with transcripts sent to clinicians and live transfer to a human where needed. By 2026 the agent, Ana, answered inbound calls and scheduled primary care appointments, and WellSpan announced an expanded partnership whose first new workflows call patients who missed imaging appointments. WellSpan reports that Ana manages more than 160,000 patient calls a month.","stage":"scaled","year":2024,"channels":["voice","web-chat"],"languages":["en","es"],"metrics":[{"kpi":"interactions-handled","value":160000,"unit":"count","qualifier":"at-least","period":"per month, patient calls","claimant":"organization","quote":"Ana, WellSpan’s generative AI agent developed in partnership with Hippocratic AI, currently manages more than 160,000 patient calls per month, engaging in more than 7,000 hours of conversation.","sourceUrl":"https://www.wellspan.org/articles/2026/07/30/13/05/web---hippocractic-ai-partnership-expansion"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.wellspan.org/articles/2026/07/30/13/05/web---hippocractic-ai-partnership-expansion","title":"WellSpan expands Hippocratic AI partnership to enhance operations and patient experience","publisher":"WellSpan Health","date":"2026-07-30"},{"url":"https://www.globenewswire.com/en/news-release/2024/09/26/2953949/0/en/WellSpan-One-of-the-First-Major-Health-Systems-in-the-World-to-Launch-Hippocratic-AI-s-Generative-AI-Healthcare-Agent.html","title":"WellSpan One of the First Major Health Systems in the World to Launch Hippocratic AI's Generative AI Healthcare Agent","publisher":"WellSpan Health (GlobeNewswire)","date":"2024-09-26"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"wellspan-hippocratic-ai-patient-voice-agent"},{"title":"West Berkshire Council: automated decisions on applications for a larger bin","useCases":["permit-and-licence-application-processing"],"organization":{"name":"West Berkshire Council","anonymized":false,"country":"GB","region":"europe","industry":"government"},"vendors":[{"name":"Goss Interactive","role":"platform"}],"summary":"West Berkshire Council assesses online applications for a larger household rubbish bin against the minimum criteria in its policy with a rules based tool built in house. Applications that miss the criteria are rejected and the applicant is told automatically; those that meet them go to the Waste Management team for review. In many cases staff time per application drops to recording the outcome, and every application and decision is stored so the team can scrutinise them. The tool was tested before launch to confirm it reached the same decisions as a person applying the same policy, production decisions are sampled periodically, and applicants can appeal to the waste team. It is a small but real example of automated decisions on a simple council permission.","stage":"production","year":2025,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"interactions-handled","value":42,"unit":"count","qualifier":"approximately","period":"decisions per month on average","claimant":"organization","quote":"An average of 42 decisions are made per month using the algorithmic tool.","sourceUrl":"https://www.gov.uk/algorithmic-transparency-records/west-berkshire-council-apply-for-a-larger-rubbish-bin"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.gov.uk/algorithmic-transparency-records/west-berkshire-council-apply-for-a-larger-rubbish-bin","title":"West Berkshire Council: Apply for a Larger Rubbish Bin","publisher":"GOV.UK (Algorithmic Transparency Recording Standard)","date":"2025-01-28"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"west-berkshire-council-larger-bin-application-assessment"},{"title":"Westpac: AI nudges in the mobile banking app, as assessed by Forrester","useCases":["financial-wellbeing-coach"],"organization":{"name":"Westpac","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[],"summary":"Forrester's 2025 review of Australian mobile banking apps, as reported by Mi3, ranked Westpac first for the third year running and credited AI features that nudge customers toward better financial decisions. The same coverage says that the big four Australian banks still fall short of what customers most want, including timely alerts and personalised guidance. This is an analyst assessment of the app, not a result published by the bank.","stage":"production","year":2025,"channels":["mobile-app"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.mi-3.com.au/01-10-2025/big-four-deliver-ai-nudges-and-cautious-cleverness-their-banking-apps-forrester","title":"Westpac blows app rivals away as Forrester rates Australia among world's best – but Big Four still missing key customer aspirations","publisher":"Mi3","date":"2025-10-01"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"D","id":"westpac-app-ai-nudges"},{"title":"Westpac: real time AI call assistant for scam conversations","useCases":["scam-payment-interception","fraud-alert-confirmation"],"organization":{"name":"Westpac","anonymized":false,"country":"AU","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Westpac","role":"in-house"}],"summary":"In May 2025 Westpac announced that it was piloting an AI call assistant with its specialist scam and fraud team. It transcribes live customer calls, flags indicators that the customer may be about to pay a scammer or is being coached in the background, and suggests questions for the banker. It sits alongside SaferPay (questions before high risk payments), SafeCall (verified calls through the app to resist spoofing), Westpac Verify (payee name mismatch warnings) and inbound payment detection. The bank reports early qualitative results only.","stage":"pilot","year":2025,"channels":["voice","agent-desktop"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.westpac.com.au/about-westpac/media/media-releases/2025/29-may/","title":"Using AI to put scammers out of business","publisher":"Westpac","date":"2025-05-29"},{"url":"https://www.westpac.com.au/news/making-news/2025/05/westpac-deploys-real-time-AI-to-take-on-scammers/","title":"Westpac deploys real-time AI to take on scammers","publisher":"Westpac Wire","date":"2025-05-29"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"westpac-scam-call-assistant"},{"title":"World Bank: generative AI tutor pilot in after school English classes in Edo, Nigeria","useCases":["ai-tutor-for-students"],"organization":{"name":"World Bank","anonymized":false,"country":"NG","region":"africa","industry":"education"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"In June and July 2024 a World Bank team, working with the Edo State education authorities, ran a six week after school programme in Edo, Nigeria, in which 800 first year senior secondary students used Microsoft Copilot as a tutor, mainly to learn English, with teacher support: teachers introduced each session's topic, suggested prompts and mentored students as they worked with the tool. In a randomized evaluation, participants outperformed their peers in English, AI knowledge and digital skills, and also did better in their end of year exams. The team reports learning gains of about 0.3 standard deviations, says the programme outperformed 80% of the interventions in a database of randomized evaluations in developing countries, and reports that the more sessions students attended, the greater their gains; girls, who started behind boys, seemed to gain even more.","stage":"pilot","year":2024,"channels":[],"languages":["en"],"metrics":[],"outcomeDisclosed":true,"sources":[{"url":"https://blogs.worldbank.org/en/education/From-chalkboards-to-chatbots-Transforming-learning-in-Nigeria","title":"From chalkboards to chatbots: Transforming learning in Nigeria, one prompt at a time","publisher":"World Bank Blogs","date":"2025-01-09"},{"url":"https://blogs.worldbank.org/en/education/From-chalkboards-to-chatbots-in-Nigeria","title":"From chalkboards to chatbots in Nigeria: 7 lessons to pioneer generative AI for education","publisher":"World Bank Blogs","date":"2024-09-18"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"world-bank-edo-nigeria-ai-tutor-pilot"},{"title":"World Food Programme: DARTS machine learning reconciliation of cash transfers","useCases":["ledger-and-payment-reconciliation"],"organization":{"name":"World Food Programme","anonymized":false,"region":"global","industry":"government"},"vendors":[],"summary":"DARTS (Data Assurance and Reconciliation Tool Simplified) is a web application that uses machine learning to help World Food Programme country offices apply controls to large cash transfer datasets and generate reconciliation reports, so that humanitarian cash assistance is paid out accurately and accountably. A 2024 US federal AI use case inventory entry by the USAID Bureau for Humanitarian Assistance, dated April 2024, listed it at the implementation and assessment stage and said it was funded through the WFP Innovation Accelerator. It does not appear in the 2025 inventory, so its current status is unknown. No outcome figures are published.","stage":"pilot","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://raw.githubusercontent.com/ombegov/2024-Federal-AI-Use-Case-Inventory/main/data/2024_consolidated_ai_inventory_raw_v2.csv","title":"2024 consolidated AI use case inventory (raw data, version 2)","publisher":"Office of Management and Budget (GitHub)","date":"2025-01-23"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"world-food-programme-darts-cash-transfer-reconciliation"},{"title":"XP Inc.: Microsoft 365 Copilot for the audit team","useCases":["internal-audit-copilot"],"organization":{"name":"XP Inc.","anonymized":false,"country":"BR","region":"latin-america","industry":"capital-markets"},"vendors":[{"name":"Microsoft (Microsoft 365 Copilot)","role":"platform"}],"summary":"XP Inc. uses Microsoft 365 Copilot to automate tasks across the company, and Microsoft reports a gain in audit team efficiency alongside total hours saved. The hours figure is company wide; how efficiency was measured is not published.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["pt"],"metrics":[{"kpi":"productivity-gain","value":30,"unit":"percent","qualifier":"exact","baseline":"audit team efficiency before Copilot","claimant":"vendor","quote":"XP Inc. uses leverages Microsoft 365 Copilot to automate tasks, significantly boosting productivity by saving more than 9,000 hours and increasing audit team efficiency by 30%.","sourceUrl":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en-us/microsoft-cloud/blog/2025/07/24/ai-powered-success-with-1000-stories-of-customer-transformation-and-innovation/","title":"AI-powered success, with more than 1,000 stories of customer transformation and innovation","publisher":"Microsoft Cloud Blog","date":"2025-07-24"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"xp-inc-copilot-audit-team"},{"title":"Yale Cancer Center: semiautomated clinical trial patient matching on OMOP data","useCases":["clinical-trial-patient-matching"],"organization":{"name":"Yale Cancer Center","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Yale Cancer Center","role":"in-house"}],"summary":"Yale Cancer Center built a clinical trial patient matching (CTPM) tool that combines rules with natural language processing over structured and unstructured record data standardized to the OMOP common data model. Validated first on one metastatic colorectal cancer trial, it was then implemented across 29 trials in several cancer specialties. Since September 2022 it has screened 98,348 patients, identified 825 eligible candidates and contributed to 117 enrollments, and it cut screening time per reviewed chart by 41%. The prescreening is semiautomated: research teams review the candidates the tool puts forward instead of reading every chart.","stage":"production","year":2022,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"handling-time-reduction","value":41,"unit":"percent","qualifier":"exact","period":"screening time per chart for patients who underwent review","claimant":"organization","quote":"Implementation reduced chart review workload 10-fold and screening time by 41% (3.1 to 1.8 minutes per chart) for those patients who did undergo review.","sourceUrl":"https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:41512229%20AND%20SRC:MED&resultType=core&format=json"},{"kpi":"interactions-handled","value":98348,"unit":"count","qualifier":"exact","period":"patients screened since September 2022, across 29 trials","claimant":"organization","quote":"Since September 2022, the system has screened 98,348 patients across 29 trials, identifying 825 eligible candidates and facilitating 117 patient enrollments with 9%-37% consent rates.","sourceUrl":"https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:41512229%20AND%20SRC:MED&resultType=core&format=json"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.ebi.ac.uk/europepmc/webservices/rest/search?query=EXT_ID:41512229%20AND%20SRC:MED&resultType=core&format=json","title":"Clinical Trial Patient Matching: A Real-Time, Common Data Model and Artificial Intelligence-Driven System for Semiautomated Patient Prescreening in Cancer Clinical Trials (abstract record)","publisher":"Europe PMC","date":"2026-01-09"},{"url":"https://ascopubs.org/doi/10.1200/CCI-25-00262","title":"Clinical Trial Patient Matching: A Real-Time, Common Data Model and Artificial Intelligence-Driven System for Semiautomated Patient Prescreening in Cancer Clinical Trials","publisher":"JCO Clinical Cancer Informatics"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"yale-cancer-center-clinical-trial-patient-matching"},{"title":"Yes Bank: RM Assist Chatbot (Ask Genie) for relationship managers","useCases":["wealth-advisor-knowledge-assistant"],"organization":{"name":"Yes Bank","anonymized":false,"country":"IN","region":"asia-pacific","industry":"banking"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Yes Bank developed RM Assist (Ask Genie), an internal chatbot on Azure OpenAI that gives relationship managers answers to customer queries from the bank's repository of product and policy documents and helps them prepare product pitches. Microsoft's page, which also names Power Apps Copilot, presents it as a way to improve first time resolution of customer interactions. The benefits are described as expectations; no live usage or outcome figures are published.","stage":"announced","year":2024,"channels":["internal-tools"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en-in/aifirstmovers/yes-bank","title":"Yes Bank: Answers to all questions","publisher":"Microsoft India"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"yes-bank-rm-assist-chatbot"},{"title":"YoungWilliams: Priya agent for SNAP, child support and Summer EBT enquiries","useCases":["benefits-eligibility-and-application-assistant"],"organization":{"name":"YoungWilliams","anonymized":false,"country":"US","region":"north-america","industry":"professional-services"},"vendors":[{"name":"Microsoft (Azure AI Foundry Agent Service)","role":"platform"}],"summary":"YoungWilliams, a US company, provides call centre and other services to government health and human services organizations. Its AI agent Priya answers public enquiries on child support services, the Supplemental Nutrition Assistance Program (SNAP) and Summer EBT, covering benefit applications, eligibility changes, reapplication and wait times, and supports human representatives by retrieving information, summarizing cases and clarifying policy. It retrieves live government data through Azure AI Search and Bing Search, asks clarifying questions instead of assuming, and enforces role based access to sensitive data.","stage":"production","year":2025,"channels":["web-chat","agent-desktop"],"languages":["en"],"metrics":[{"kpi":"response-time-reduction","value":99,"unit":"percent","qualifier":"exact","period":"initial response time, from almost four minutes to about three seconds","baseline":"human customer service representatives","claimant":"vendor","quote":"Priya was 99% faster in initial response time and delivered 45% more empathetic interactions","sourceUrl":"https://www.microsoft.com/en/customers/story/23958-young-williams-azure-ai-agent-service"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23958-young-williams-azure-ai-agent-service","title":"YoungWilliams cuts call center response time 99% with Azure AI Foundry Agent Service","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"youngwilliams-priya-benefits-agent"},{"title":"Your Health: autonomous medical coding across all service lines","useCases":["medical-coding-automation"],"organization":{"name":"Your Health","anonymized":false,"country":"US","region":"north-america","industry":"healthcare"},"vendors":[{"name":"Fathom","role":"platform"}],"summary":"Your Health, a senior focused primary and specialty care group in South Carolina and Georgia with about one million patient visits a year, went live with Fathom's autonomous coding in October 2025 across every service line and place of service, integrated with its athenahealth record system. In its own newsroom Your Health reports a 95.5% encounter level automation rate and coding accuracy up from 96.3% to 98.3% since implementation; Fathom's release adds more complete diagnosis capture that raised average risk adjustment scores. Neither source says how accuracy was measured or how the remaining encounters are handled.","stage":"scaled","year":2025,"channels":["api"],"languages":["en"],"metrics":[{"kpi":"automation-rate","value":95.5,"unit":"percent","qualifier":"exact","period":"encounter level, since implementation in October 2025","claimant":"organization","quote":"Since implementation in October 2025, we have achieved a 95.5% encounter-level automation rate and increased coding accuracy from 96.3% to 98.3%, reflecting meaningful gains in efficiency, precision, and scalability across the organization.","sourceUrl":"https://www.yourhealth.org/company-news-1/your-health-reaches-new-milestone-in-coding-accuracy-and-automation"},{"kpi":"accuracy","value":98.3,"unit":"percent","qualifier":"exact","period":"since implementation in October 2025","baseline":"96.3% under the prior coding approach","claimant":"organization","quote":"Since implementation in October 2025, we have achieved a 95.5% encounter-level automation rate and increased coding accuracy from 96.3% to 98.3%, reflecting meaningful gains in efficiency, precision, and scalability across the organization.","sourceUrl":"https://www.yourhealth.org/company-news-1/your-health-reaches-new-milestone-in-coding-accuracy-and-automation"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.yourhealth.org/company-news-1/your-health-reaches-new-milestone-in-coding-accuracy-and-automation","title":"Your Health Reaches New Milestone in Coding Accuracy and Automation","publisher":"Your Health","date":"2026-03-23"},{"url":"https://fathomhealth.com/insights/your-health-deploys-fathom-autonomous-medical-coding-to-achieve-95-5-automation-rate-at-98-3-accuracy-rate-across-all-service-lines","title":"Your Health deploys Fathom autonomous medical coding to achieve 95.5% automation rate at 98.3% accuracy rate across all service lines","publisher":"Fathom","date":"2026-03-19"},{"url":"https://finance.yahoo.com/sectors/healthcare/articles/health-deploys-fathom-autonomous-medical-140000178.html","title":"Your Health Deploys Fathom Autonomous Medical Coding to Achieve 95.5% Automation Rate at 98.3% Accuracy Rate Across All Service Lines","publisher":"Business Wire (via Yahoo Finance)","date":"2026-03-19"},{"url":"https://www.businesswire.com/news/home/20260319818217/en/Your-Health-Deploys-Fathom-Autonomous-Medical-Coding-to-Achieve-95.5-Automation-Rate-at-98.3-Accuracy-Rate-Across-All-Service-Lines","title":"Your Health Deploys Fathom Autonomous Medical Coding to Achieve 95.5 Automation Rate at 98.3 Accuracy Rate Across All Service Lines","publisher":"Business Wire","date":"2026-03-19"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"B","id":"your-health-autonomous-medical-coding"},{"title":"Zalando: AI powered fashion assistant in 25 markets","useCases":["conversational-shopping-assistant"],"organization":{"name":"Zalando","anonymized":false,"country":"DE","region":"europe","industry":"retail-and-ecommerce"},"vendors":[{"name":"OpenAI","role":"model-provider"},{"name":"Zalando","role":"in-house"}],"summary":"Zalando opened a beta of its assistant to logged in customers in Germany, Austria, the United Kingdom and Ireland by November 2023, after testing it internally and with selected customers. Since October 2024 it gives logged in customers fashion advice in their local language across all 25 Zalando markets, using Zalando's own models and OpenAI's large language models. It interprets context such as occasion, location and weather (\"what should I wear to my dad's 60th birthday in November in Barcelona?\"), and an update in March 2025 connected it to customer accounts and shopping history and made it aware of the page the customer is browsing. In a pilot of the personalised version Zalando observed 40% more high value interactions, such as likes and add to cart actions.","stage":"scaled","year":2023,"channels":["web-chat","mobile-app"],"languages":[],"metrics":[{"kpi":"users-served","value":2000000,"unit":"count","qualifier":"at-least","period":"cumulative since launch, as reported in March 2025","claimant":"organization","quote":"So far, over 2 million customers have used it to get inspired and find items they love.","sourceUrl":"https://corporate.zalando.com/en/technology/more-personal-and-smarter-zalando-assistant-enhanced-capabilities-inspire-customers"}],"outcomeDisclosed":true,"sources":[{"url":"https://corporate.zalando.com/en/technology/more-personal-and-smarter-zalando-assistant-enhanced-capabilities-inspire-customers","title":"More personal and smarter, the Zalando Assistant with enhanced capabilities to inspire customers","publisher":"Zalando","date":"2025-03-27"},{"url":"https://corporate.zalando.com/en/technology/zalando-brings-its-ai-powered-assistant-all-markets-and-adds-four-new-cities-its-trend","title":"Zalando brings its AI-powered assistant to all markets and adds four new cities to its Trend Spotter","publisher":"Zalando","date":"2024-10-01"},{"url":"https://corporate.zalando.com/en/technology/how-zalando-co-creating-its-new-ai-powered-assistant-together-customers","title":"How Zalando is co-creating its new AI-powered assistant together with customers","publisher":"Zalando","date":"2023-11-30"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"zalando-ai-fashion-assistant"},{"title":"ZLS Zoll und Logistikservice: AI prepared customs declarations with Digicust","useCases":["customs-classification-and-declaration"],"organization":{"name":"ZLS Zoll und Logistikservice GmbH","anonymized":false,"country":"DE","region":"europe","industry":"logistics-and-transportation"},"vendors":[{"name":"Digicust","role":"platform"}],"summary":"ZLS, a customs clearance and warehousing firm in Neuhaus am Inn, Germany, founded in 2018 with a focus on trade with Turkey, uses Digicust's AI platform to take over the manual, error prone data entry for its customs clearances. An earlier German version of the case study names the product as Dexter IDP and says it fills customs declarations from different data sources and assigns tariff numbers from the goods description. Digicust reports that the processing time per extensive clearance fell from three to four hours to 10 to 15 minutes, with costs down by up to 70% and fewer entry errors.","stage":"production","year":2025,"channels":["internal-tools"],"languages":["de"],"metrics":[{"kpi":"handling-time-reduction","value":90,"unit":"percent","qualifier":"exact","period":"per extensive customs clearance, from 3 to 4 hours to 10 to 15 minutes of processing","claimant":"vendor","quote":"How ZLS Logistik Service achieved a 90% reduction in processing time and 70% cost reduction through Digicust's AI-powered automation solution","sourceUrl":"https://digicust.com/en/case-studies/zls/"},{"kpi":"cost-reduction","value":70,"unit":"percent","qualifier":"up-to","claimant":"vendor","quote":"Up to 70% reduction in costs through automation, allowing ZLS to improve their competitive position in the market.","sourceUrl":"https://digicust.com/en/case-studies/zls/"}],"outcomeDisclosed":true,"sources":[{"url":"https://digicust.com/en/case-studies/zls/","title":"ZLS Case Study","publisher":"Digicust"},{"url":"https://www.digicust.com/de/testimonials/zls/","title":"Fallstudie: Optimierung der Zollprozesse bei ZLS","publisher":"Digicust","archivedUrl":"https://web.archive.org/web/20250623085110/https://www.digicust.com/de/testimonials/zls/"},{"url":"https://zls-logistik.de/","title":"ZLS Logistik","publisher":"ZLS Zoll und Logistikservice GmbH"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"zls-digicust-customs-declaration-automation"},{"title":"Zoom: AI avatar videos for sales enablement training","useCases":["training-content-generation"],"organization":{"name":"Zoom","anonymized":false,"country":"US","region":"north-america","industry":"technology"},"vendors":[{"name":"Synthesia","role":"platform"}],"summary":"Zoom's instructional designers train more than 1,000 salespeople on selling its products. When training material changed, whole videos had to be recorded again, with subject matter experts spending a day in front of a camera for about 15 minutes of usable footage. The team now produces AI avatar videos with Synthesia and builds them into interactive modules in Rise 360 and Storyline. Synthesia reports 90% time savings on video creation, more than 200 micro videos from one designer in about six months, 15 to 20 hours a month freed for Zoom's subject matter experts who no longer record themselves (the page does not say whether this is per expert or in total), and monthly cost savings of $1,000 to $1,500 per employee previously spent on creating training videos.","stage":"production","year":2023,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"processing-time-reduction","value":90,"unit":"percent","qualifier":"exact","baseline":"time to create a training video, from the page's headline figure; the body states it as \"90% faster than before, producing content in less than an hour\"","claimant":"vendor","quote":"90% time savings","sourceUrl":"https://www.synthesia.io/case-studies/zoom"},{"kpi":"hours-saved","value":15,"unit":"hours","qualifier":"at-least","period":"per month","baseline":"15 to 20 hours a month that Zoom's subject matter experts no longer spend recording (the page does not say whether this is per expert or in total)","claimant":"vendor","quote":"Time Saved for SMEs: Zoom's subject matter experts no longer need to record themselves, freeing up 15-20 hours each month to work on their actual job.","sourceUrl":"https://www.synthesia.io/case-studies/zoom"},{"kpi":"cost-savings","value":1500,"unit":"currency","currency":"USD","qualifier":"up-to","period":"per month, per employee","baseline":"$1,000 to $1,500 a month per instructional designer or expert previously spent on creating training videos; vendor stated, method not given, not net of licence costs","claimant":"vendor","quote":"Enhanced Productivity: Thanks to AI video, both IDs and SMEs can work more efficiently. This results in monthly cost savings of $1,000 - $1,500 per employee previously spent on creating training videos.","sourceUrl":"https://www.synthesia.io/case-studies/zoom"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.synthesia.io/case-studies/zoom","title":"How Zoom accelerated training video production by 90%","publisher":"Synthesia","date":"2023-09-13"}],"verification":{"level":"source-verified","checkedAt":"2026-09-27"},"grade":"C","id":"zoom-ai-video-sales-training"},{"title":"Zoya: Shariah compliance screening of stocks and funds at retail scale","useCases":["shariah-compliance-screening"],"organization":{"name":"Zoya","anonymized":false,"country":"US","region":"north-america","industry":"wealth-and-asset-management"},"vendors":[{"name":"Zoya","role":"in-house"}],"summary":"Zoya, a halal investing app, publishes Shariah compliance reports for more than 60,000 stocks and screens ETFs and mutual funds, then monitors holdings and alerts users when a stock's compliance status changes. It shows how screening against published Shariah criteria works at retail scale. The app does not describe its screening as generative AI, and it does not replace a Shariah board's review of a bank's own contracts.","stage":"scaled","year":2026,"channels":["mobile-app"],"languages":["en"],"metrics":[{"kpi":"users-served","value":400000,"unit":"count","qualifier":"at-least","period":"investors","claimant":"organization","quote":"Trusted by 400,000+ Investors","sourceUrl":"https://zoya.finance/"}],"outcomeDisclosed":true,"sources":[{"url":"https://zoya.finance/","title":"Zoya -","publisher":"Zoya"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"B","id":"zoya-shariah-stock-screening"},{"title":"Zurich Insurance Group: Copilot for Sales to keep commercial insurance CRM data current","useCases":["sales-call-coaching-and-crm-update","insurance-broker-and-agent-assistant","insurance-renewal-and-retention"],"organization":{"name":"Zurich Insurance Group","anonymized":false,"country":"CH","region":"europe","industry":"insurance"},"vendors":[{"name":"Microsoft","role":"platform"}],"summary":"Zurich's commercial insurance teams manage more than 100,000 active opportunities in Dynamics 365, and switching applications to copy updates from email into the CRM left data at risk of going stale. With Microsoft 365 Copilot for Sales, 300 users create and update contacts and link emails to opportunities from Outlook, get summaries of relationships and long email threads, and draft emails. Zurich estimates about 14,000 hours saved over the next year; that is an estimate, not a measured result. The story also says user feedback indicates the tool improves Zurich's sales and retention ratios, without figures.","stage":"production","year":2025,"channels":["email","microsoft-teams","internal-tools"],"languages":["en"],"metrics":[{"kpi":"users-served","value":300,"unit":"count","qualifier":"exact","claimant":"vendor","quote":"Copilot for Sales has quickly become a critical productivity tool for Zurich’s 300 Copilot users, who find even more benefits as they continue to work with it in Outlook.","sourceUrl":"https://www.microsoft.com/en/customers/story/23452-zurich-schweiz-microsoft-365-copilot-for-sales"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.microsoft.com/en/customers/story/23452-zurich-schweiz-microsoft-365-copilot-for-sales","title":"Zurich Insurance enhances customer relationships with Dynamics 365 and Microsoft 365 Copilot for Sales","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"zurich-copilot-for-sales-crm-updates"},{"title":"Zurich Insurance (Hong Kong): WhatsApp service agent that triages policy and claim questions","useCases":["insurance-policy-servicing-agent"],"organization":{"name":"Zurich Insurance (Hong Kong)","anonymized":false,"country":"HK","region":"asia-pacific","industry":"insurance"},"vendors":[{"name":"Microsoft","role":"platform"},{"name":"Twilio","role":"platform"}],"summary":"Zurich Insurance (Hong Kong) added WhatsApp to its contact centre with Dynamics 365 Contact Center and an agent built in Copilot Studio that captures preliminary information such as policy numbers before escalating to live staff, who then already know what the customer needs. Its head of customer services management says the bot might be able to answer simple questions. A second agent automates motor claim status updates from external surveyors to customers by SMS or email. The company reports lower call and email volumes and staff handling up to two chats at once, and was piloting Copilot to search the knowledge base during live chats. No containment figure is published for the AI agent itself.","stage":"production","year":2024,"channels":["whatsapp","sms","email","agent-desktop"],"languages":[],"metrics":[],"outcomeDisclosed":false,"sources":[{"url":"https://www.microsoft.com/en/customers/story/24082-zurich-dynamics-365-contact-center","title":"Zurich Insurance (Hong Kong) transforms customer service and communications with Dynamics 365","publisher":"Microsoft Customer Stories"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"zurich-hong-kong-whatsapp-service-agent"},{"title":"Zurich North America: AI drafted underwriting narratives for middle market submissions","useCases":["underwriting-risk-assessment-copilot"],"organization":{"name":"Zurich North America","anonymized":false,"country":"US","region":"north-america","industry":"insurance"},"vendors":[{"name":"Sixfold","role":"platform"}],"summary":"Zurich North America's U.S. Middle Market underwriters had to comb through hundreds of pages of exposures, operations, loss runs and supplemental forms per submission and then write a compliant underwriting narrative. After Sixfold was selected as one of nine winners of the Zurich Innovation Championship, the team rolled out a tool that gives underwriters an AI generated first draft of the narrative in Zurich's appetite, format and tone, with accuracy tracking and feedback sessions with underwriters. The vendor reports an average of 2 hours saved per submission and expansion from four offices to dozens within six months; an underwriter describes the tool surfacing a major exposure that changed a risk assessment.","stage":"scaled","year":2024,"channels":["internal-tools"],"languages":["en"],"metrics":[{"kpi":"time-saved-per-task","value":120,"unit":"minutes","qualifier":"exact","period":"per submission, as reported by underwriters","claimant":"vendor","quote":"Underwriters reported saving an average of 2 hours per submission","sourceUrl":"https://www.sixfold.ai/case-study/zurich"}],"outcomeDisclosed":true,"sources":[{"url":"https://www.sixfold.ai/case-study/zurich","title":"Zurich North America | Sixfold Case Study","publisher":"Sixfold"}],"verification":{"level":"source-verified","checkedAt":"2026-09-26"},"grade":"C","id":"zurich-north-america-sixfold-underwriting-narratives"}]}