01When it can
- The refund tool rejects orders outside the window and amounts above the limit.
- Order facts (date, amount, earlier refunds) come from your system, not the customer.
- The agent has been tested with refund pushers and honest customers, many times.
- Every refund is logged with the conversation behind it.
02What should still go to a person
- Refunds above a set amount.
- Exceptions to policy, however good the story.
- A second refund on the same order, or store credit where it is not standard.
- Anything with signs of fraud: pressure, threats, claims that cannot be checked.
03Moving from approval to automatic
Start with the agent proposing and a person approving; automate within hard limits once tests show a held-firm rate whose likely range clears your threshold; keep re-testing on every change. This is the path up the Authority Ladder in the enterprise guide. How to test it: how to test an AI refund agent.
Questions
Is a refund limit in the prompt enough?
No. A limit written only in the agent’s instructions is a sentence a persistent customer can argue with. The refund tool itself should reject amounts and orders outside policy.
Doesn’t a person approving every refund remove the benefit?
At first, partly. Approval is a stage, not the end state: it shows which refunds the agent gets right, so routine ones can be automated within limits while unusual ones still reach a person.
See where your agent leaks money. Tested against your own rules, with the money at stake.