What problem does it solve?
A payment that flows straight through needs no manual work, but every payment that falls out does. A missing or malformed field, a name that does not match the account, a sanctions hit, a duplicate or a customer asking "where is my payment" each opens a case. Operators then read MT and MX messages, look up the payment in several systems, write free text queries to correspondent banks and wait for an answer that may arrive by message, email or not at all.
Much of this work is reading and writing: free format messages such as the MT199, emails and customer queries, while the customer keeps asking for news. ISO 20022 defines structured messages for exceptions and investigations, such as the interbank payment cancellation request (camt.056) and its response (camt.029), the payment status request (pacs.028) and the investigation request and response (camt.110 and camt.111). The Committee on Payments and Market Infrastructures recommends that payment system operators and participants align with its harmonised ISO 20022 data requirements for cross border payments before the end of 2027, and expects correct account data to mean fewer exceptions and investigations. Structured data makes cases easier to classify; AI does the reading, drafting and chasing that remain.
Banks have started with the repair step. 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 around the world.
How does it work?
- Classify the exception. The agent reads the rejected or held payment, the error codes and any inbound query (camt.056 recall, camt.110 investigation request, camt.029 or camt.111 response, gpi tracker status, free text MT199 or email) and assigns the case type.
- Gather the facts. It pulls the payment's lifecycle from the payment hub, the tracker and the customer record, and retrieves the relevant scheme rules and internal procedures.
- Repair or propose. For repairable errors it proposes the corrected fields (for example a BIC derived from the IBAN) with its reasoning. For investigations it proposes the next action: request information, recall, return or beneficiary correction.
- Draft the messages. It drafts the structured investigation message to the counterparty and the plain language update to the customer or the relationship manager.
- Approve and chase. An operator approves any repair, recall or credit adjustment. The agent then sends, tracks deadlines, chases unanswered queries and closes the case with a full trail.
- Audience
- Back office
- Autonomy
- Supervised agent
- Adoption
- Emerging
- Channels
- Internal tools, API and system to system, Email
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 |
|---|---|---|---|---|
| Automation rate | Too few to pool | at least 10% | 1 | 1 organization |
Value drivers: Lower cost to serve, Speed and cycle time, Customer experience, Risk and loss reduction.
Indicative value
A regional bank handling 100,000 payment exception and investigation cases a year
USD 291,667 to USD 2.3 million
Investigation effort avoided per year
How this is calculated
Formula: cases * minutesPerCase / 60 * effortReduction * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Exception and investigation cases per year cases, cases per year | 100,000 | 100,000 | The reference bank. Replace with your own case volume. |
| Operator minutes per case today minutesPerCase, minutes per case | 20 | 45 | Editorial assumption covering reading, lookups, drafting and follow up. Replace with your own time study. |
| Share of operator time the AI removes effortReduction, fraction of time per case | 0.25 | 0.5 | Editorial assumption; the AI drafts and gathers, a human still approves money movement. |
| Fully loaded operations cost per hour costPerHour, USD per hour | 35 | 60 | Editorial assumption, replace with your own. |
What it leaves out: 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.
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.
JPMorgan Chase
United States · Banking · 2023
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.
No outcome disclosed.
BNY
United States · Capital markets · 2026
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.
- Automation rate: at least 10%, payment repair issues worldwide
"Today, that digital employee handles over ten percent of our payment repair issues around the world."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- 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
Systems to integrate
- 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
Complexity: High
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.
- 1
Map the case types
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.
- 2
Automate the reading and the gathering first
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.
- 3
Add drafting with templates
Draft structured messages from the case data and free text only where the counterparty requires it, with operators approving every outgoing message at first.
- 4
Introduce straight through handling per case type
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.
- 5
Close the loop with the customer
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
KPIs to instrument
- 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
Human in the loop
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.
Common failure modes
- Wrong repair sent at scale
- A plausible but wrong field correction sends money to the wrong place. Keep human approval on repairs and validate against static data.
- Automated chasing that annoys counterparties
- Duplicate or badly timed chasers damage correspondent relationships. Respect agreed response windows and deduplicate.
- Case file and customer story drift apart
- The customer is told something the case does not support. Generate customer updates only from the case status.
What are the risks and rules?
EU AI Act
Minimal risk
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.
Rules that apply
Guidance
- Harmonised ISO 20022 data requirements for enhancing cross border payments (updated report) (Committee on Payments and Market Infrastructures, Global). 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 to put in place
- 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
Frequently asked questions
- Does ISO 20022 remove the need for AI in payment investigations?
- 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.
- Can an AI agent recall or repair a payment on its own?
- 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.
- What results have banks published?
- 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.
How to cite this page
Blits.ai AI Use Case Library, "AI for payment investigations and exceptions", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/payment-investigations-and-exceptions. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published