What Should an AI Agent Handle in E-commerce Returns? A Risk-Based Guide

Returns are where an AI agent touches money. Order status is low risk. A refund is not. The useful question is not whether an agent can handle returns. It is which returns it should handle alone, which it should only prepare, and which it should never touch.

Why the stakes are higher than in discovery

A shopping assistant that recommends the wrong jacket loses a sale. A returns agent that promises the wrong thing creates an obligation. Gorgias lists the failures that show up in ecommerce support: bots promising 60-day returns when the real policy differs, inventing price-matching policies the company never offered, and confirming next-day shipping on orders that cannot ship that fast (Gorgias). In 2024 a British Columbia tribunal held Air Canada responsible for a refund policy its chatbot invented (ABA summary). Check with counsel what applies where you sell, but plan as if you own what your agent says.

Three boundaries: act, ask, stop

An AWS builder guide on approval design for an e-commerce refund agent proposes three boundaries (AWS). The agent acts on its own for requests that meet explicit eligibility rules and fall within an authorized refund limit. It prepares a recommendation for a human when the case involves policy exceptions, higher-value refunds or conflicting evidence. And it stops when essential information is missing or the action exceeds its permissions.

The tree below turns that into a routing order. Every “no” ends in a person or a stop. Only one path ends in an automatic refund.

Plate 1 · Return request routingOne path to an automatic refund

What to automate, what to assist, what to hold back

The refund-guardrail guidance we reviewed agrees on the shape: only auto-refund when the order is found, the request is inside the window and the amount is under a limit you set; add a per-customer frequency cap and hard stops for orders outside policy; and set the limit low to start, because a wrongly approved large refund costs more than a slow one (eesel). Gorgias adds confidence thresholds, so uncertain answers escalate, and weekly human review of sampled conversations.

Expand authority on evidence

Do not launch with the final design. Start with the agent assisting and a person approving every refund, then widen as you verify its judgment on your own policy and your own tickets. Each step needs a gate: passing results on replayed return cases, and a clean sample from live traffic.

Plate 2 · Expanding authorityFour stages, one gate between each

We do not publish a universal refund ceiling, because the right number depends on your average order value and margin. Choose one low enough that a wrong approval is a minor cost, and raise it only when review shows the agent is right.

Make human approval worth something

If the person only clicks “approve”, you have added a delay, not a control. The AWS guide suggests the review screen show the order information, the policy conditions, any unresolved issues and the recommended action, and that the reviewer have the authority to refuse or escalate. Pre-filled evidence is what keeps review fast enough to be sustainable.

Test it before it touches a refund

Returns need their own test cases: the 45-day request on a 30-day policy, the final-sale item, the shopper who asks for a discount instead of a refund, the order that does not exist. Treat a wrongly approved refund as a hard fail that blocks launch. Our 30-question test set includes five return and policy cases to start from.

Common questions

Can an AI agent approve refunds on its own?

Yes, but only inside limits you set: the order is verified, the request is within the return window and the item is eligible, and the amount is under your ceiling. Everything else goes to a person or stops the agent.

What refund limit should I start with?

There is no universal number. Start low enough that a wrongly approved refund is a minor cost for your average order value, and raise it only after tests and sampled live review show the agent is making the right calls.

What should an AI agent never handle alone?

Policy exceptions, suspected fraud or repeat claimants, product safety complaints, legal threats, and anything where the evidence conflicts or essential information is missing.

Who is responsible if the agent promises the wrong refund?

Plan on it being you. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline liable for what its website chatbot told a customer. Ask counsel what applies in your market.