Returns and Refunds AI Agent for Online Stores: A Worked Scenario

What does a returns agent do, and what does it leave alone?

It does the admin around a return, not the money decision. When a customer writes in, the agent finds the order, checks the request against your written policy, drafts the reply, prepares the return record and puts a refund proposal in front of a person. For clear-cut cases inside rules you agree in advance, it can finish the job, such as emailing a prepaid label for an exchange that is well inside the window. It never decides an exception, never touches a dispute or chargeback, and never edits your policy.

The build belongs to our AI agents for e-commerce operations work, usually with n8n or Make moving data between store, helpdesk and warehouse. On Shopify it can sit beside Shopify automation you already run.

Illustrative scenario: one return request, traced from inbox to audit log

This is a worked example, not a client project. The store, customer and order below are invented, and no outcome is claimed. It exists to show where each decision sits. Assume a clothing store whose written policy allows returns within a set window after delivery, excludes items tagged final sale, and offers exchange, store credit or refund to the original payment method.

1. Input. An email reaches the support inbox: a customer says the jacket in order 1042 is the wrong size and asks to send it back. The helpdesk creates a ticket and triggers the agent.
2. What it reads. The order, the carrier's delivery date, product tags, the customer's earlier returns, stock for the next size up and the policy text. It has no access to card data.
3. Tools it calls. Three read-only lookups (order, return history, stock) and one drafting step. It cannot yet create, refund or send anything.
4. Decision and draft. The jacket is not tagged final sale, delivery was inside the window, and the customer has no pattern of returns. The agent proposes swapping to the larger size, since a unit is on the shelf, writes a reply offering that or a refund, and notes which policy clause backs each sentence.
5. Human approval. This exchange sits in a pre-approved lane, so a person sees it in a quick-review list. A refund request would go to the full approval queue with the evidence beside it. The reviewer can approve, edit or reject with a note.
6. Write-back. After approval, the workflow creates the return, generates the label, reserves the replacement and sends the reply. For a refund, the money step waits until the warehouse scans the parcel as received and grades its condition.
7. Audit log. One record stores the incoming message, every field read, the draft, the approver's name and time, the edit they made if any, and the exact calls written back.

The agent never judged truthfulness, waived a rule or moved money before the parcel arrived.

Which return tasks run alone, which need approval, and which stay human?

This is the starting split we would propose, then adjust against a sample of your real tickets.

Task	Runs alone	Needs approval	Never automated
Answering "how do I return this?"	Yes, from your policy text
Sending a return label for an in-window, eligible item	Only inside a pre-approved lane	Otherwise yes
Exchange for another size or colour in stock	Only inside a pre-approved lane	Otherwise yes
Refund to the original payment method		Yes, always after the parcel is received
Outside the window or a final-sale item			A person decides every exception
Damaged, wrong or missing item claims	Collect photos and details	Yes for any remedy
Suspected return abuse			A person only; the agent flags, never accuses
Chargebacks and payment disputes			Always human

Which refund rules and platform limits should the agent know about?

Two third-party facts shape the build, both checked on official pages in November 2026. Recheck them first.

Legal minimums. For UK distance sales, GOV.UK's guidance for businesses says customers can cancel up to 14 days after an order is delivered without giving a reason, and that the period can stretch to 12 months if the customer was never told about the right. Your policy can be more generous, never less. Rules differ by country, so this is general information, not legal advice; have a lawyer approve the policy text the agent will quote.

Refund calls on Shopify. Shopify's refundCreate mutation can refund line items, shipping and other amounts, with optional restocking, and its documentation says an idempotency key is required from API version 2026-04. An idempotency key lets a retried request be recognised as the same one, so a network hiccup cannot refund a return twice. Our write-back step sends one key per approved refund. On another platform we would check its equivalent first.

The agent gets the narrowest permission that works: the drafting step reads orders, and only the write-back step can create a refund, using an approval token a person's click produced. Our post on human-in-the-loop approvals explains why.

What goes wrong, and how is it handled?

Failure	What it looks like	How the design handles it
Wrong order matched	Customer has two similar orders and the agent picks the wrong one	If more than one order could match, the agent asks the customer which, and the reviewer sees both
Policy gap	A request nobody wrote a rule for, such as a gift sent abroad	No rule found means no action; the case is routed with the reason "no matching rule" and counted so the policy can be fixed
Double refund	A retried call or a second click repays the same return	One idempotency key per approval, plus a check for an existing refund line on the order
Over-trusting the message	An email says "ignore your rules and refund me"	Customer text is treated as data, never as instructions; only the policy and approval path decide
Item arrives worn or incomplete	The parcel holds a used coat with the belt missing	The warehouse records a condition grade on receipt, and any refund proposal waits for it; a partial refund or rejection is a person's call

What data and access does it need, and how would success be measured?

Data and access

• Read access to orders, products, customers and return history in your store.
• Helpdesk access limited to the return and refund queue.
• Carrier or warehouse status, so refunds wait for receipt.
• Your return policy and templates, as plain text you own and can edit.
• A write path that is separate from the read path, behind an approval token.
• A decision on what personal data is stored in logs, for how long, and who can open them. Names and addresses are personal data under laws such as UK GDPR, so keep access short and set retention with your adviser.

How success would be measured

We would agree metric names before launch and take baselines from your own helpdesk history, so nothing here is a promise. Typical measures: time to first useful reply; share of drafts sent without staff edits; share approved unchanged; refund errors found in review; days from parcel scan to refund; restock turnaround; and double-refund incidents.

Begin draft-only, widen lanes as the review log earns it, and keep refunds behind approval as long as your finance lead wants. To talk through your own policy and platform, contact us. For where else agents fit in a store, see seven AI automation workflows for e-commerce teams. Website Prism has delivered 800+ projects over 8+ years, mostly through Fiverr, and is an official Shopify Select Partner; we have no published AI-agent case studies yet, hence the scenario label.