Refund and return requests handled by an agent, inside your policy

Refund and return requests are rule-heavy and repetitive, which makes them tedious for people and tempting to automate badly. This playbook has an agent gather the facts, apply your written policy, and prepare an outcome, with money only moving inside limits you set.

01the problem and who owns it

Each request means opening the order system, checking purchase date, item condition claims, prior refunds, shipping status, and the policy that applied when the order was placed. Agents interpret policy differently, so two customers with the same situation get different answers, and finance discovers the inconsistency in the monthly numbers.

Support owns the conversation, operations owns the warehouse side, and finance owns the money. Nobody owns the end-to-end decision, which is why exceptions pile up.

02what the AI does, step by step

1. Understand the requestThe agent reads the message and any photos, identifies the order, and determines what the customer wants: refund, replacement, exchange, or store credit, and the reason given.
2. Gather the factsIt pulls order date, items, price paid, delivery confirmation, previous returns by this customer, and the policy version in effect at purchase from your commerce and shipping systems.
3. Apply the written policyPolicy rules are encoded as explicit logic (window, condition, final sale items, restocking fees). The model handles interpretation, such as matching a vague damage description to a policy reason, and records which rule it relied on.
4. Prepare the outcomeFor clear approvals under the auto-approve limit, it generates the return label or RMA and drafts the confirmation. For denials, it drafts a clear explanation citing the policy, which a person reviews.
5. Escalate exceptionsHigh-value orders, repeat refund patterns, goodwill cases, and anything outside policy go to a person with a summary and recommended action.
6. Close the loopWhen the return arrives or the refund posts, the ticket updates and the customer is notified. Reasons are tagged so product and operations see why things come back.

03systems it connects to

• Commerce platform. Shopify, BigCommerce, Magento, or your order management system.
• Payments. Stripe, Shopify Payments, or your processor, for refunds within approved limits.
• Returns and shipping. Returns platforms such as Loop or Narvar, or carrier APIs for labels and tracking.
• Helpdesk. Gorgias, Zendesk, or similar for the customer conversation.

04human checkpoints

• Auto-approve limits. Finance sets the dollar ceiling and eligible categories for automatic approval, reviewed quarterly.
• Every denial. A person approves denials before they are sent, since a wrong denial costs more than a slow one.
• Fraud patterns. Repeat-refund customers and mismatched addresses go to a reviewer, never auto-approved.

05what to measure

• Time to resolution. From request to refund issued or label sent.
• Policy consistency. Sampled decisions checked against the policy by a second reviewer.
• Exception rate. Share of requests needing a person, and the top reasons.
• Return reasons. Tagged reasons by product, shared with merchandising and quality.

06risks and guardrails

• Money moving without oversight. Keep refund permissions scoped to the limit, log every action, and alert on unusual volume in a day.
• Card data. The agent should never see full card numbers. Work through processor refund APIs so PCI scope does not expand.
• Consumer protection rules. Return and refund obligations vary by country and state. Have counsel confirm the encoded policy matches what you are legally required to honor.

07build vs buy

Returns platforms like Loop, Narvar, and AfterShip Returns run self-service portals with policy rules, and many helpdesks offer refund actions. For standard ecommerce, buy first.

Custom work fits when requests come through messy channels like email and phone, when policy has many exceptions tied to customer history, or when refunds touch systems a returns app does not integrate with, such as a custom ERP.

08related playbooks

Browse every customer support playbook or the full library.