How to handle push payment scam reimbursement claims with an AI agent

The agent does the urgent and the laborious parts. It takes the report with care, stops further payments to the scammer, alerts the receiving bank within minutes, and assembles the evidence: payments, warnings shown, the customer's account of events, any vulnerability. A named claims assessor decides reimbursement every time, and the agent pays only what they approve.

What is the process, and who owns it?

In an authorised push payment scam, the customer is tricked into sending money themselves, to someone posing as a seller, an investment, a romantic partner or the customer's own bank. Because the customer authorised the payment, it is not an unauthorised transaction, and recovering it depends on the receiving bank acting fast. Some markets expect banks to reimburse scam victims in defined circumstances, with exceptions and time limits. The detail varies by market and changes over time, so the agent works to your current policy, never to its own reading of the rules.

The head of fraud operations owns the process. Named claims assessors own each decision. The case for keeping that decision with a person is made in scam reimbursement needs a person. The wider frame is in our guide to AI agents in banking.

What are the steps, and who acts at each one?

1. Agent, depth 2: takes the report. Calmly, without blame, in the customer's words. Many victims are distressed or embarrassed. How an agent should handle that first contact is covered in whether an AI agent can handle a scam victim's report. Evidence: the report.
2. Agent, depth 3: protects the account. Blocks further payments to the scammer's account, cancels pending ones, and checks for an ongoing scam: is the customer still in contact with the scammer, being told to move more money? Evidence: blocks applied and screening answers.
3. Agent, depth 3: alerts the receiving bank or banks through the agreed inter-bank route, with payment references and amounts. Minutes matter. Evidence: alert reference and time.
4. Agent, depth 1: assembles the payment record. Every payment in the scam, with dates, amounts, payees, and the warnings and name check results shown at the time. These come from your own systems and are signed. Evidence: payment list.
5. Agent, depth 1: gathers the customer's evidence. Messages, adverts, documents the scammer sent. These are external and are filed as such. Evidence: evidence index with provenance.
6. Agent, depth 1: records vulnerability factors the customer mentions or the account shows. Evidence: factors noted.
7. Agent, depth 2: drafts the assessment in the scam reimbursement claim assessment template: facts, timeline, policy questions and the evidence on each. It does not recommend an outcome. Evidence: draft.
8. Person, claims assessor: decides. Full, partial or no reimbursement, with reasons tied to policy. Evidence: decision.
9. Agent, depth 4, held: prepares the reimbursement payment for the assessor's approval, and drafts the decision letter. Evidence: held verdict, approver, letter.

Which controls matter, and what evidence do they leave?

Protection before questions. Blocks and alerts happen before the detailed interview. Evidence: timestamps.

No outcome from the agent. The assessment presents evidence against each policy question; the assessor answers them. Evidence: assessment template without an outcome field for the agent.

Signed and external evidence kept apart. The warnings your app showed are a fact from your records; a screenshot from the customer is evidence of what they saw. Evidence: provenance on each item.

Reimbursement payments always held. Evidence: held verdict and approver on every payment.

Updates while the customer waits. Assessments take time. The agent sends updates at set intervals so the customer is not left chasing. Evidence: update messages on the case.

Policy versioned. Decisions cite the policy in force on the date of the claim. Evidence: policy version on the decision.

Whether an agent can intervene before the money leaves is a separate question, answered in whether an AI agent can stop an authorised push payment scam.

What should you measure?

• Time from report to receiving bank alert. The number that most affects recovery.
• Funds frozen or recovered at receiving banks.
• Time from complete evidence to decision, against any limit your policy sets.
• Decisions changed on complaint or review.
• Repeat victims, who may need extra protection.
• Customers chasing for updates, which shows whether the update rhythm works.

Where does it usually go wrong?

The customer is interrogated. Questions framed as "why didn't you check?" deter reporting and harm vulnerable customers.

Alerts wait for the full story. The interview runs first and the alert goes out an hour later, after the funds have moved on.

The scam is still running. The customer is still talking to the scammer, who tells them the bank is the fraudster. Ask, and involve a specialist if so.

The agent's summary leans. Wording like "the customer ignored a warning" pre-judges a question the assessor must weigh. Report what was shown and what the customer says.

Recovery stops at the alert. Follow up with the receiving bank and record the outcome.