Can an AI agent process a return?

Yes, for returns that fit the policy. An agent can check the order, the returns window and the item's eligibility, issue a return authorisation and label, and refund once the warehouse confirms the item arrived in acceptable condition. Disputes over condition, high-value items, returns outside the window and patterns that suggest return fraud go to a person.

Why does a return have two decisions, not one?

Authorising a return and refunding it are different steps with different risks. Authorising is cheap to reverse: if the item never comes back, nothing has been lost but a label. Refunding moves money, and recovering it later is slow and often fails. That refund is the point of no return in the process.

So the safe design lets the agent authorise freely within policy, but ties the refund to evidence from the warehouse. Some businesses refund at drop-off for trusted customers; that is a deliberate policy choice with its own limits, not a default.

What conditions decide it?

• The order exists in the business's records and belongs to the verified customer.
• The return is inside the window, counted from the delivery date the records hold.
• The item is eligible: not personalised, not hygiene-sealed and opened, not excluded by the policy.
• The refund waits for a scan and, where the policy requires, an inspection.
• The refund is within the agent's limit and goes to the original payment method.

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

Writing these steps down, with who acts at each one, is part of documenting an AI agent's process.

What does this look like in practice?

Take an outdoor clothing retailer with a 28-day returns window. A customer asks to return a waterproof jacket bought two weeks ago because it is too small.

The agent confirms the order and delivery date, sees the jacket is eligible, and issues a return label. Five days later the warehouse scans the parcel and the inspection passes. The agent refunds the 120 to the original card and tells the customer. A second customer returns walking boots under the same policy. The inspection notes heavy wear and mud. The agent does not refund. It records the inspection note and photos and holds the case for the returns team, who offer to send the boots back or refund a reduced amount under the policy's wear clause. The customer is told a person is reviewing it.

Which controls make it safe?

• Authorisation and refund as separate capabilities, each with its own depth.
• Refunds triggered only by warehouse scan and, where needed, inspection.
• Returns policy read from the policy library, including exclusions.
• Refund limits per item and per customer per period.
• Return frequency per customer monitored for patterns.

The general approach to support agents that can say no is in the guide to AI customer support agents.

Where Pryme fits

In Pryme the warehouse scan reaches the agent as a sealed fact from the system of record, while the customer's description of the item's condition arrives as declared. A refund can rest on the sealed scan; it cannot rest on the declaration alone. Authorisation and refund are certified separately, so each can sit at a different depth.