Secure the account before anything else, then sort the claim: truly unauthorised, a scam the customer was tricked into, a merchant dispute, or a family member using the card. Clean unauthorised claims can be refunded within limits where your policy calls for prompt refunds. Any question of the customer's own involvement or liability goes to a named fraud analyst.
What is an unauthorised transaction claim, and who owns it?
An unauthorised transaction is one the customer did not make or consent to: a card used after theft, online banking accessed by someone else, a direct debit set up without a mandate. In several markets, the customer's payment provider must refund such transactions promptly unless it has reasonable grounds to suspect the customer acted fraudulently, and liability rules then decide who bears the loss. The detail varies by market; the agent works to your policy.
The hard part is the sorting. Customers describe several different situations as "I didn't authorise this": a payment they made under a scammer's instructions, a subscription they forgot, a purchase by their child. Each has a different route, different rules and a different answer. The head of fraud operations owns the process. Fraud analysts own liability decisions. For the wider frame, see our guide to AI agents in banking.
What does the agent prepare, and what does a person approve?
- Agent, depth 3: secures the account. Freezes the card, ends active online sessions, and starts the credential reset route. Evidence: protective actions with timestamps.
- Agent, depth 1: lists recent transactions with the customer. Which do they not recognise? Evidence: list and responses.
- Agent, depth 1: sorts each claim. Did the customer make the payment themselves, even if deceived? That is a scam, not an unauthorised transaction. Is it a merchant they know, charging wrongly? That is a dispute. Could someone in the household have used the card? Evidence: classification with the customer's answers.
- Agent, depth 1: gathers the authentication evidence. How each transaction was authenticated: chip and PIN, contactless, strong authentication in the app, device and location used, whether a one-time passcode was entered. These come from your systems and are signed. Evidence: authentication record per transaction.
- Agent, depth 1: checks for indicators. Transactions authenticated on the customer's usual device, previous claims, a pattern your transaction monitoring flagged or did not. Evidence: indicator list.
- Agent, depth 4 within limits: refunds clean claims. Where the claim is unauthorised, no indicators point to the customer, and the amount is within limits, it refunds and tells the customer. Evidence: refund verdict.
- Person, fraud analyst: decides the rest. Claims with indicators, high values, repeat claims, and any question of the customer's negligence or involvement. Evidence: decision and reasons.
- Agent, depth 2: drafts the decision letter for analyst decisions. Evidence: letter and sender.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Secure | Freezes card, ends sessions | 3 | Nothing |
| List | Reviews transactions with the customer | 1 | Nothing |
| Sort | Classifies each claim | 1 | Unclear cases |
| Evidence | Gathers authentication data | 1 | Nothing |
| Indicators | Checks for warning signs | 1 | Nothing |
| Refund | Refunds clean claims within limits | 4 | Claims with indicators or above limits |
| Liability | Nothing | none | Fraud analyst decides |
| Letter | Drafts the outcome | 2 | Analyst sends |
Which controls matter, and what evidence do they leave?
Secure first. A refund on an account still controlled by a fraudster funds the next theft. Evidence: protective actions precede the refund.
Correct routing. Scams go to the scam route; merchant disputes to the debt and billing dispute processes or chargebacks; unauthorised claims stay here. Evidence: routing per transaction.
No accusation. The agent never suggests the customer is lying. Indicators are recorded for the analyst, not put to the customer. Evidence: transcript sampling.
Authentication evidence is signed. What your systems recorded outweighs any account of events, either way. Evidence: provenance class on each item.
Refund limits and running totals. Evidence: totals on each refund verdict.
Direct debit claims checked against the mandate. A claim that a direct debit was never authorised is tested against the mandate record, not card authentication data. Evidence: mandate check on each such claim.
Liability by a person. Deciding that a customer was grossly careless, or involved, is a serious finding. Evidence: analyst identity on every such decision.
What should you measure?
- Time from claim to account secured.
- Time from claim to refund for clean claims, against your policy.
- Claims re-routed after initial sorting, which shows how well step 3 works.
- Refunded claims later found to involve the customer.
- Complaints about outcomes or tone.
Where does it usually go wrong?
Scams are refunded as unauthorised. The customer made the payment under deception; it belongs on the scam route with its own assessment.
Unauthorised claims are handled as scams. The reverse error: a card used by a thief is treated as if the customer chose to pay, and refund is delayed.
The account is not secured. The refund lands and the fraudster spends it.
Transactions are blocked instead of claimed. The agent tries to stop pending transactions it cannot stop. Whether it should try is covered in whether an AI agent can block a suspicious transaction.
Family use is mishandled. A child used a parent's card. The parent wants the money back and does not want the child treated as a criminal. A person handles this with care.
Questions people ask
Can the agent refund before the investigation is complete?
Where your policy requires prompt refunds for unauthorised transactions and no indicators point to the customer, yes, within limits. If the investigation later shows otherwise, recovering the refund is an analyst's decision.
What if the customer's own device was used?
That is an indicator, not a conclusion. The device may have been stolen or remotely controlled. The claim goes to an analyst.
Does this apply to purchases an agent makes on the customer's behalf?
That is a different problem: preventing your own agents from making purchases nobody authorised. It is covered in how to stop an AI agent from making unauthorised purchases.