Can an AI agent approve a refund?

Yes, inside a set of conditions that all hold at once: the amount is within the business's limit, the refund goes back to the original payment method, the reason is confirmed by the business's own records, and the agent is certified for that refund type. Break any condition and the refund is held for a person, or refused.

Why is a refund different from most support tasks?

Because money leaves the business. A wrong answer about delivery times can be corrected in the next message. A wrong refund has to be recovered, often from a customer who now disputes owing it. That puts cash refunds at depth 4 in our scale, the level for moving money, even when the amounts are small.

Many businesses therefore start lower. The agent issues account credits or store vouchers at depth 3, which are easy to reverse, and drafts cash refunds at depth 2 for a person to release. As evidence builds on the business's own cases, cash refunds within a tight limit can move to the agent. The step from pilot to live decision-making is covered in how to approve a new AI agent for production.

What conditions decide it?

1. A limit set by the business. Per refund, per customer per period, and per day across all customers. How to set those is covered in how to set refund limits for an AI agent.
2. The original payment method. Money goes back the way it came. A request to send it elsewhere is a separate, higher-risk question, covered in whether an AI agent can send a refund to a different card or account.
3. A reason the records confirm. A returns scan, a cancelled order, a duplicate charge on the ledger. The customer's word alone is a claim, not a confirmation.
4. No abuse signals. Repeated "not received" claims, or a delivery address changed just before the order, send the request to a person.
5. A current certificate for this refund type, on the business's own past cases.

What may the agent do, and what stays with a person?

Mapping each refund type to a limit and a role is easier in a single table; a ready-made version is in the refund and credit limits matrix for support agents.

What does this look like in practice?

Take an online fashion retailer. Its agent is certified to issue refunds for returned items up to 150, once the warehouse has scanned the return.

A customer asks where the refund is for a jacket sent back last week. The warehouse scan shows the jacket received and passed inspection. The amount is 90, inside the limit. The agent issues the refund to the original card, tells the customer when it should appear, and files the decision with the scan and the rule. A second customer asks for a refund on a parcel the carrier marked as delivered to their door. The agent does not argue and does not refund. It collects the details, starts the lost-parcel check and holds the request for the claims team.

Which controls make it safe?

• Limits held outside the agent's instructions, enforced at the moment of action.
• Refund destination locked to the original payment method.
• Reasons checked against sealed records: order system, warehouse, ledger.
• Every refund filed with its evidence and the bound that decided.
• Sampled review of approvals, plus a watch on refund volume per customer.

In banking, refund requests on fees carry extra conduct duties; see how to handle fee refund requests on bank accounts with an AI agent.

Where Pryme fits

In Pryme the minimum chain decides each refund: role scope, certificate, sources, the customer's own limits and platform guardrails. The lowest bound wins, and the filed verdict names it, so you can see whether a held refund was held by your limit or by the agent's evidence. More on the approach is in the guide to AI customer support agents.