How to handle payment returns with an AI agent

Match, credit, explain. The agent pairs each returned payment with the original, credits the funds back to the customer's account within bounds, and tells the customer why the payment came back and what to check before trying again. Unmatched returns, amounts reduced by exchange or charges, returns with fraud indicators and anything above limits go to a named payments analyst.

What is a payment return, and who owns the process?

A payment return is money coming back because the payment could not be completed: the receiving account was closed, the details were wrong, the receiving bank refused it, or a direct debit could not be collected. It is not a payment reversal, which undoes a transaction, usually before settlement, and not a recall, which asks the recipient's bank to send back money that did arrive.

Returns are mostly routine and mostly annoying. Customers notice the payment did not arrive, then find the money back in their account days later with no explanation. Business customers find half a salary run returned. Meanwhile, unmatched returns pile up in a suspense account. The head of payment operations owns the process; payments analysts own the exceptions. The wider frame is in our guide to AI agents in banking.

What does the agent prepare, and what does a person approve?

1. Agent, depth 1: reads the return. Scheme reason code, amount, currency, the original payment reference. Evidence: return record.
2. Agent, depth 1: matches it to the original. By reference first, then by amount, date and payee. Evidence: match result with confidence and the fields used.
3. Agent, depth 3: credits the customer where the match is exact and the amount equals the original. Evidence: credit verdict and reference.
4. Agent, depth 2: explains the return to the customer. The reason in plain words, what to check (account details, whether the payee has changed bank), and a warning not to resend until the details are confirmed. Evidence: message sent.
5. Agent, depth 1: flags reduced amounts. International returns can come back short after charges and exchange movements. Evidence: shortfall calculated.
6. Person, payments analyst: decides reduced amounts. Who bears the shortfall is a policy decision with customer impact. Evidence: decision and reason.
7. Person, payments analyst: works unmatched returns. The agent prepares candidates; the analyst confirms the right account. Evidence: candidate list and confirmed match.
8. Person, fraud: reviews returns with indicators. A return from an account closed for suspected fraud may mean the customer was being scammed. Evidence: referral and outcome.

Which controls matter, and what evidence do they leave?

Exact match before credit. A return credited to the wrong customer is two problems. Evidence: match fields on each credit verdict.

Amount equals original. Anything else is held. Evidence: amount comparison.

No resend by the agent. The agent never re-sends a returned payment, even to corrected details; the customer does that, as a new instruction. Evidence: refused verdicts on resend attempts.

Reason codes mapped to plain language. Approved wording per code. Evidence: wording version on each message.

Suspense ageing. Unmatched returns are worked within a set period. Evidence: ageing report.

Business customers told by batch. When a salary or supplier run returns several payments, the customer gets one message listing every returned line with its reason, not a stream of separate alerts. Evidence: batch message linked to each return.

Returns separated from reversals. Whether an agent may ever reverse a payment is a different and much harder question, covered in whether an AI agent can reverse a payment. Evidence: separate capabilities.

What should you measure?

• Returns matched automatically, against those needing an analyst.
• Time from return received to customer credited and told.
• Unmatched returns by age.
• Customers resending to the same wrong details. The explanation is not working if this is common.
• Returns referred to fraud, and how many were scams.

Where does it usually go wrong?

Silent credits. The money comes back with no message, and the customer resends to the same details.

Fuzzy matching credits the wrong account. Two returns of the same amount on the same day are swapped.

Shortfalls are absorbed without a decision. Small differences are written off, then larger ones.

Fraud signals are missed. A return because the receiving account was closed for fraud is a sign the customer may have been deceived into paying a criminal. Contact them.

Returned direct debits are misread. A collection returned unpaid is neither a complaint nor a scam; route it by its reason code to the right team.

Instant and batch returns are mixed. Instant payments have their own exception routes and time limits; see handling instant payment exceptions with an AI agent.

Where Pryme fits

In Pryme, matching, crediting and explaining are separate capabilities, each certified on your own past returns. A credit runs only when the match and amount checks pass at the moment of action; reduced amounts, unmatched returns and fraud indicators become held verdicts for a named analyst. Every return and decision is filed in the append-only record.