Returns abuse predates AI agents. Appriss Retail and Deloitte estimated that fraudulent and abusive returns and claims cost US retailers $103 billion in 2024, about 15% of all returns (2024 Consumer Returns in the Retail Industry). An agent that can grant refunds meets the same customers, in conversation.
01The threats
| Tactic | What it tries to get |
|---|---|
| Outside the window | A refund after the return window, with a reason why it should not count |
| Used or damaged item | A full refund for an item that does not qualify |
| Double dipping | A second refund, or a refund plus a replacement, for one order |
| Store credit upgrade | Cash instead of store credit, or extra credit on top |
| Escalation and threats | A refund to make a complaint or chargeback threat go away |
| Invalid discount code | A code that does not exist, has expired or was for someone else |
| Fake payment proof | An order shipped on a screenshot or pasted confirmation |
| Delivery address switch | A parcel redirected after ordering, without verification |
| Data fishing | Another customer’s email, address or order history |
02The rules to write
As data the agent and your systems can check, not only as sentences in its instructions.
- Return window in days, refund limit per order, condition rules and one refund per order.
- When store credit is allowed, and when cash is.
- Valid discount codes and their percentages, checked by your system.
- The free extras the agent may give (free shipping, a small credit) and their limits.
- That payment status and addresses come from your order system, never from the chat.
03The tests to run
- AI refund and returns agents: refund pushers, from polite to threatening, against your window and limits
- Storefront fraud and scams: fake payment proof, invalid codes, impersonation, address switches and data fishing
- AI sales agents and discounts: if the agent can discount or bundle, hagglers against your lowest prices
Run each on a test copy of the agent, with honest customers mixed in, enough times for a useful range: see how many test conversations are enough.
04Pre-launch checklist
- The refund tool enforces the window and the limit itself.
- Discount codes are validated by the system, not the agent.
- Payment status is read from the order system before anything ships.
- Address changes after ordering require verification.
- The agent can hand off to a person instead of conceding.
- The agent has been tested with refund pushers, fraudsters and honest customers.
- Eligible refunds are granted: over-refusal is measured.
The full 50-point version is in the enterprise guide.
05What to monitor after launch
- Refunds and credits issued by the agent per week, against before launch.
- Discount value given per order.
- Address changes after ordering, and their outcomes.
- Chargebacks on orders the agent handled.
Questions
Our agent runs on a help-desk platform. Can it still be tested?
Yes, if a test copy can be reached over the internet in a format it already speaks. The test sends it conversations as a customer would and reads its replies and actions.
Should we test before or after launch?
Before, on a test copy, and again after every change to its instructions, model or tools.
See where your agent leaks money. Tested against your own rules, with the money at stake.