What problem does it solve?
Routine servicing requests arrive in a constant stream in retail banking: where is my transaction, send me a statement, my card is lost, raise my limit. Each request is simple, but together they fill queues, push up waiting times at the moments customers are most anxious (a lost card, a payment that did not arrive) and take human agents away from the conversations that need judgment, such as hardship, fraud victims and complaints.
Many first generation banking chatbots were rule based: the CFPB describes them as using decision tree logic or a database of keywords to trigger preset, limited responses. A preset answer can explain a procedure, but the customer still has to finish the task somewhere else. The step change is an agent that is authenticated, can act in the core banking and card systems within strict limits, and knows when to stop and hand over. DBS describes a similar shift: its DBS Joy assistant used to give customers instructions on how to find the information they needed, and now answers from their own transaction and account data.
- 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.Chatbots in consumer finance (2023)
How does it work?
- Understand the request. The agent detects the intent ("freeze my card", "why was I charged twice") in the customer's own words and language, on any channel.
- Authenticate proportionally. Information requests need a logged in session; actions that move money or change a card need step up authentication, such as an app confirmation or voice biometrics on the phone.
- Act through approved tools. The agent calls a small allow list of banking APIs (card block, replacement order, statement request, limit change within set bounds) and confirms the result back to the customer.
- Answer policy questions from approved content. Fees, product terms and procedures come from retrieval over the bank's own documents, so answers are grounded and current.
- Hand over well. Vulnerability signals, complaints, disputes and anything outside the allow list go to a human specialist with a summary of the conversation, so the customer never repeats themselves.
- Audience
- Customer facing
- Autonomy
- Supervised agent
- Adoption
- Mainstream
- Channels
- Mobile app, Web chat, WhatsApp, Phone and voice
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Containment rate | Too few to pool | 84.6% to 90% | 2 | 1 organization, 1 vendor |
| Contact deflection | Too few to pool | 7% | 1 | 1 organization |
| Satisfaction uplift | Too few to pool | 17% | 1 | 1 organization |
Value drivers: Lower cost to serve, Customer experience, Inclusion and access, Employee productivity.
Indicative value
A retail bank with 1 million digitally active customers
USD 720,000 to USD 8.6 million
Human handled contact cost avoided per year
How this is calculated
Formula: customers * contactsPerCustomer * servicingShare * containment * costPerContact. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Digitally active customers customers, customers | 1,000,000 | 1,000,000 | The reference bank. |
| Assisted contacts per customer per year contactsPerCustomer, contacts per customer per year | 2 | 4 | Editorial assumption for a digitally active retail bank. Replace with your own contact volume. |
| Share of contacts that are routine servicing servicingShare, fraction of contacts | 0.4 | 0.6 | Editorial assumption, replace with the share from your own contact reason report. |
| Share of servicing contacts the agent resolves containment, fraction of servicing contacts | 0.3 | 0.6 | 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. |
| Cost of a human handled contact costPerContact, USD per contact | 3 | 6 | Editorial assumption for a blended chat and phone contact. Replace with your own fully loaded cost. |
What it leaves out: 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.
Market estimates (analyst estimates, not deployments)
- McKinsey estimates that generative AI could add between USD 200 billion and USD 340 billion in value to global banking each year. Capturing the full value of generative AI in banking (2023)
Who already uses it?
2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
DBS Bank
Singapore · Banking · 2026
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.
- Containment rate: about 90%, DBS digibot, first half of 2026, queries resolved without a follow up call
"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."
Claimed by: organization - Contact deflection: 7%, DBS Joy in Singapore, first six months of 2026, calls or emails to customer service
"Active users increased by 61%, contributing to a 7% reduction in calls or emails to customer service."
Claimed by: organization - Satisfaction uplift: 17%, DBS Joy in Singapore, first six months of 2026
"Customer satisfaction scores for DBS Joy rose by 17% over the same period."
Claimed by: organization
Commonwealth Bank of Australia
Australia · Banking · 2024
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.
- Containment rate: about 84.6%, May 2026, self service messaging
"In May 2026, approximately 84.6% of self-service messaging interactions were resolved end-to-end in the messaging channel."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- 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
Systems to integrate
- 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
Complexity: Medium
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.
- 1
Pick the first intents by volume and risk
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.
- 2
Define the action allow list
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.
- 3
Ground the answers
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.
- 4
Design the handover
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.
- 5
Test before customers do
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.
- 6
Launch in one channel, then widen
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
KPIs to instrument
- 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
Human in the loop
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.
Common failure modes
- Fluent but wrong policy answers
- Answers drawn from outdated or unapproved content. Prevent with document ownership, review dates and refusal when retrieval finds nothing.
- Containment that is really abandonment
- Customers give up rather than get helped, which looks like containment in the dashboard. Count repeat contacts and measure satisfaction per intent.
- Handover without context
- The customer has to start again with a human, which is worse than no assistant. Pass the summary and the authentication state.
- Scope creep into risky actions
- New actions are added without their own risk review. Treat every new action as a change with sign off.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
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.
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). People must be informed that they are interacting with an AI system unless this is obvious from the context.
- Chatbots in consumer finance (Consumer Financial Protection Bureau, North America). 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 to put in place
- 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
Frequently asked questions
- What share of servicing requests can an AI agent resolve?
- 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.
- Is a banking servicing chatbot high risk under the EU AI Act?
- 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.
- Which requests should stay with humans?
- Disputes, suspected fraud or scams, hardship and financial difficulty, complaints and any conversation where the customer shows signs of vulnerability.
How to cite this page
Blits.ai AI Use Case Library, "AI agent for account and card servicing", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/account-and-card-servicing-agent. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published