Refunds and Returns

Use case

Refunds and returns

This worked guide shows how a fictional ecommerce business could translate a refund policy into AI agent permissions, approval steps and evidence requirements.

Fictional example: The limits below are illustrative, not industry standards. Adapt the amounts, time windows, approval roles and controls to your own policies, systems and legal obligations.

The sample policy

The downloadable example uses a fictional retailer, Northstar Outfitters. Its standard return window is 30 days from delivery for unworn resaleable items. The agent may issue a routine refund only after a carrier return scan and item condition check, up to $75, to the original payment method.

Exceptions require approval. A CX shift lead approves refunds from $75.01 to $250. A support operations manager approves refunds over $250 or policy-window exceptions. If no approver is available, the agent may send a holding message and offer policy-compliant options, but it may not execute the exception.

Permission levels used here

  • Inform: The agent may explain or retrieve information.
  • Recommend: The agent may propose an action, but another person or approved workflow must decide.
  • Act within limits: The agent may execute the action only when stated conditions are met.
  • Request approval: The agent must route the action before execution.
  • Not permitted: The agent must not propose or execute the action.

Routine actions

A routine action is a low-exception action the agent can complete because the business rule and system evidence match. In this example, the agent may explain eligibility, generate a return authorization and prepaid label when the item is eligible, and issue a refund up to $75 after the system confirms the return scan and item condition.

The agent may also recommend a refund amount. Recommendation is different from execution: the agent can calculate and propose an amount, but execution is allowed only if the refund remains inside the routine limits and required evidence exists.

Approval conditions

Approval is required when the action exceeds routine authority. In the example, triggers include refunds above $75, returns 31 to 45 days after delivery, shipping costs above $20, item condition uncertainty, damaged-in-transit claims, or goodwill refunds beyond one per customer per 180 days.

For more help deciding which actions can happen independently and which need approval, see How to Decide What Your AI Agent Can Do on Its Own.

The agent drafts the approval request. It does not execute the exception. The approval workflow records the approver, timestamp, decision, reason code and resulting customer-facing message.

When the request crosses a boundary, Designing the Handoff When an AI Agent Reaches Its Limits explains how to preserve the customer’s request and make responsibility clear while approval is pending.

Prohibited actions

Some actions should remain outside the agent’s authority. In this example, the agent may not override fraud flags, refund to a new payment method, or approve its own exception. Those actions require human review outside the agent workflow.

Business rules and technical controls

A business rule states the intended boundary: for example, “routine refunds are capped at $75.” A technical control enforces that boundary in the system: for example, a payment API rule that blocks refund execution above $75 unless an approved exception record exists.

Useful controls for this use case include return-window calculation, item-condition status, fraud-flag checks, refund amount limits, original-payment-method restrictions, approval workflow gating and immutable logs. The agent should not be trusted to provide its own account of success. The system should retain the evidence record.

Evidence requirements

Evidence retention is a system requirement. The system should retain the order ID, item status, delivery date, return scan, policy version, condition code, refund calculation inputs, approval request ID, approver identity, decision, reason code, transaction ID and final customer message.

Before granting refund authority, it also helps to plan how mistakes would be detected and corrected. See When an AI Agent Makes a Mistake: Design for Correction and Recovery.

Scenarios

  • Routine request: A customer returns an unworn $48 item 12 days after delivery. The system sees the return scan and clean condition. The agent may issue a refund to the original payment method.
  • Boundary request: A customer returns a $74.99 item on day 30. The agent may execute only if all evidence is clean and no exception is needed.
  • Exception request: A customer requests a $120 refund on day 38. The agent may draft an approval request, but a support operations manager must approve the return-window exception before execution.
  • Prohibited request: A customer asks for a refund to a different card, or the account has a fraud flag. The agent must not execute or propose a bypass. It routes the case to human review.

To see how unresolved payment destinations, conflicting records and unavailable approvers can be reviewed, explore the illustrative before-and-after permission plan review.

Use the worksheet

Read the completed example to understand the method, then use the blank worksheet to apply the same structure to your own policy and implementation context.