How to handle returns requests with an AI agent

A standard return is a rules check the agent can run end to end. It reads the order, checks the item and date against your policy, and issues the return label at depth 3. Refunds follow warehouse inspection. Anything outside the policy, and any claim that goods are faulty, is drafted for a person.

What is the returns process, and who owns it?

Customers return goods because they changed their mind, the item did not fit, it arrived damaged, or it is faulty. Most retailers publish a returns policy setting the window, the condition goods must be in, and items that cannot be returned for hygiene, perishability or personalisation. Separately, consumer law in many markets gives buyers rights over faulty goods and some distance purchases, which a returns policy cannot reduce.

E-commerce or customer operations owns the returns policy. The warehouse owns receipt and inspection. Finance owns refunds. Fraud owns patterns of abuse. Support owns the conversation. The agent runs the standard path and stops at judgement, in line with our guide to the AI support agent that knows when to say no. This is about returning goods, not payment returns, where a bank sends a payment back.

What does the agent prepare, and what does a person approve?

1. Agent, depth 1: identifies the order and items, with enough verification to avoid exposing another customer's purchase.
2. Agent, depth 1: asks the reason: change of mind, wrong size, damaged on arrival, faulty, not as described. The reason decides the path.
3. Agent, depth 1: checks eligibility for change-of-mind returns: delivery date against the window, item category against exclusions, any condition requirements.
4. Agent, depth 3: issues the return authorisation and label for eligible standard returns, with instructions and the expected refund timing from your policy.
5. Agent, depth 1: routes fault and damage claims to the fault path, with photos attached as external evidence. Faulty goods carry rights a change-of-mind policy does not cover.
6. Agent, depth 2: drafts exceptions: a return just outside the window, a used item, a missing receipt, a gift return without the order details.
7. Person, support lead: decides exceptions within policy authority.
8. System and warehouse: inspect and trigger refunds under rules. Refunds outside those rules are held for a person.
9. Agent, depth 3: answers status questions about the return and refund from the system.

Which controls matter, and what evidence does each leave?

Policy version cited. Evidence: the policy version and rule applied to each decision.

Change of mind and fault kept apart. A customer with a faulty item should not be refused because the change-of-mind window has passed. Evidence: the reason recorded and the path taken.

Refunds follow inspection rules. The agent does not issue refunds. Evidence: refund events attributed to the system rule or a named person.

Exceptions are held. Evidence: held verdicts and approvals.

Patterns go to fraud, not to the customer. Frequent returners are routed for review without accusation. Evidence: referrals with the pattern found.

What should you measure?

• Returns authorised without a person, by reason.
• Time to authorise, and how average handle time changes for returns contacts. Lower is good only if repeat contacts do not rise.
• Exceptions drafted and approved.
• Return reasons by product. A product returned often as "not as described" needs its description fixed.
• Refund timing from receipt to payment.

Where does it usually go wrong?

Fault claims are treated as change of mind. The window has passed, so the agent refuses, though the customer has rights over a faulty item. Route by reason first.

Exclusions are applied too widely. A sealed hygiene item that is faulty is still returnable. Read exclusions with the reason.

Refund timing is promised. "You'll have your money in three days" ignores inspection and bank processing. Quote the policy period.

Returns and cancellations mix. An order not yet dispatched should be cancelled, not returned; see handling cancellation requests with an AI agent.

Customers who cannot post a parcel are stuck. People without transport or with disabilities may need a collection. Offer it where you can; can an AI agent handle accessibility requests? covers how.

Labels are issued for items that never arrive. A label is sent, the customer never posts the item, and nobody follows up, or the item is posted and lost on the way back. Track open authorisations and tell the customer what to keep as proof of posting.

Bundles are split wrongly. An item bought with a discount as part of a set is returned alone, and the refund ignores the bundle price. The refund rules, not the agent, should work out the amount, and the agent should explain it before the label is issued.

Gift returns go to the buyer. The refund reaches the person who paid, not the person returning. Explain this before the label is issued.

Where Pryme fits

In Pryme, issuing a return label is a depth 3 capability certified on your own past returns, while refunds sit at depth 4 and run only under your warehouse rules or a named person's approval. Customer photos arrive marked external, so they support a fault claim without deciding it.