No. Reimbursing a fraud victim turns on judgements about what they knew, what they were told and how vulnerable they were, so a named person at the bank should decide. An AI agent can take the report, gather evidence, test the case against the bank's rules and draft a decision. Paying, refusing and partial awards stay with that person.
Why is reimbursement a judgement, not a calculation?
Some fraud cases are clear: a card was used by someone else, and the customer did nothing wrong. Others are not. In a scam, the customer usually made the payment themselves, persuaded by someone they trusted. Whether they should be reimbursed may depend on whether the bank gave a clear warning, whether the customer ignored obvious signs, and whether their circumstances made them less able to protect themselves.
Those questions need someone to weigh a frightened customer's account, a call recording and a timeline of messages. In some markets, rules set out when banks must reimburse scam victims, with exceptions for gross negligence and extra protection for vulnerable customers. Applying those exceptions is exactly the kind of judgement that should not be made by an agent.
Why does a refusal weigh more than a payment?
A wrong reimbursement costs the bank money. A wrong refusal costs a victim money they may have spent years saving, on top of the shock of the scam. They can complain and escalate, but the process is slow and stressful, and many give up. So the two errors are not equal, and the decision should sit with someone who can be held to account for it. The principle is set out in in fraud, an agent should contain, not conclude.
What may the agent do, and what stays with a person?
| Step | The agent | Stays with a person |
|---|---|---|
| Taking the report | Records the customer's account in their words | Nothing routine |
| Protecting the customer | Freezes cards and blocks further payments on request | Unblocking |
| Recovering funds | Sends recovery requests to the receiving bank under the bank's process | Nothing routine |
| Building the timeline | Assembles payments, messages, warnings shown and calls | Nothing |
| Checking vulnerability | Flags indicators from the account and the conversation | Deciding how they affect the outcome |
| Testing against the rules | Sets out each condition and the evidence for and against | Deciding |
| Drafting the decision | Drafts the outcome and reasons | Approving, amending or rejecting it |
| Paying | Prepares the payment | Releasing it |
The intake is covered in detail in whether an AI agent can handle a scam victim's report.
A worked example
Take a bank handling a report from a customer in her seventies. She transferred 12,000 to a "safe account" after a caller claiming to be from the bank's fraud team told her that her savings were at risk.
The agent records her account, blocks further payments and sends a recovery request to the receiving bank. It builds the timeline: the call, the transfer twenty minutes later, and the warning screen the app displayed before the payment, which she confirmed. It notes that the customer's age and the fact that she lives alone may be relevant to vulnerability under the bank's policy.
The agent drafts the case for reimbursement and the case against, with the evidence for each. It holds the case for the named fraud decision maker. The decision maker listens to the customer's call, reads the warning she saw, decides she was not in a position to recognise the scam despite the warning, and approves full reimbursement. A second person releases the payment.
What controls make it safe?
- No decision capability. Reimbursing and refusing are both outside the agent's bounds. It drafts.
- Both sides drafted. The draft must set out the evidence for reimbursing as well as against. A one-sided draft is sent back.
- Vulnerability as a flag, not a score. The agent raises indicators; it does not weigh them.
- Customer's words kept. The original report is stored verbatim beside any summary.
- Release held separately. The person who approves a reimbursement is not the one who releases the funds.
Detection is a different job, where agents can do more; see whether an AI agent can detect fraud in real time. The wider limits are in AI agents in banking: what they may and may not do.
Questions people ask
Can the agent reimburse clear-cut unauthorised card fraud?
Where the customer did not authorise the transaction and the bank's policy sets a clear rule, a bank may choose to let a proven capability refund small amounts at depth 3. Anything with doubt about authorisation stays with a person.
Can the agent tell the victim how likely they are to be reimbursed?
No. The agent can explain the process and the timescale. A prediction is heard as a promise and can cause real distress if it proves wrong.
Who reviews a refusal if the customer challenges it?
A different person from the one who decided. The agent assembles the full file, including its own draft, so the reviewer sees how the decision was reached.