Playbook
How to automate returns safely
By the RoutMe team. Last updated .
To automate returns safely, let an AI assistant do the repeatable parts: identify the order, check eligibility against your written policy and the real purchase date, explain the steps, and route the request. Keep approval, refunds and exceptions in fixed rules or with a person, and keep the assistant read-only.
Break the job into parts
A return conversation has four parts, and they carry different risk.
| Part | Example | Automate? |
|---|---|---|
| Question | "How do I return this?" | Yes, from the policy |
| Eligibility | "Is it within the window?" | Yes, with rules and order data |
| Approval | "Can you make an exception?" | No, a person or a fixed rule |
| Action | Refund, label, restock | Through your own process, gated |
Step 1: write the policy where the assistant can read it
Put your return policy in plain language: windows, conditions, exclusions, who pays shipping, how refunds are issued. Add it as a knowledge source so the assistant quotes it. A URL source is a snapshot, so re-add the page when the policy changes. A policy that exists only in your agents' heads cannot be automated.
Step 2: compute eligibility in rules
The model should not do date arithmetic. In a RoutMe flow, a logic node compares the order date with your window and branches. The assistant then explains the outcome. This is the "agent proposes, flow decides" pattern: rules for what must never vary, the agent for the open-ended turn. See flows for the returns and refunds template.
Step 3: read the order, do not change it
An order lookup tool returns the line items, dates and delivery status. It should verify the customer first, as described in how an AI agent looks up order status safely. The chat assistant stays read-only, as explained in what read-only AI means. On Shopify, Admin API orders need the read_orders scope, which covers recent orders unless read_all_orders is also granted.
Step 4: route, do not refund
When a return is eligible, the assistant explains the steps and routes the request: a link to your returns portal, a message to your returns inbox, or a handoff to a person with the order attached. Issuing the refund happens in your own process, where your finance rules apply.
Step 5: define the human boundary
Pass these to a person immediately:
- Customers pushing back on an ineligible item
- Damaged, wrong or missing items
- High-value orders above a threshold you set
- Any mention of a payment dispute
The handoff should include the conversation, the order and the lookups, so nobody asks the customer again. See human handoff best practices.
What to test before launch
- An order inside the window and one outside it.
- A wrong order number and a wrong email.
- A customer who asks for an exception.
- A question your policy does not cover. The assistant should say it cannot confirm.
- A customer who asks for a person right away.
What can go wrong
| Risk | Control |
|---|---|
| Assistant invents policy | Knowledge source and an instruction to say when it is unsure |
| Wrong date math | Eligibility computed in flow rules |
| Customer pressure for refunds | No write access, and a handoff rule |
| Stale policy | Re-add changed pages, review quarterly |
Exchanges
Exchanges add stock to the question. Product availability can come from a read-only catalog tool, and sizing guidance from your own content. The assistant can show the available sizes as quick replies, and the exchange itself is routed to your process.
Launching in stages
Roll out returns automation in the order of risk. Start with questions only: the assistant explains the policy and the steps, quoted from your document, and nothing else. When that is stable, add the order lookup so it can say whether an item is within the window, with eligibility computed in flow rules. Only then connect the routing step that creates a request in your returns process. At each stage, read a sample of conversations and look at the handoffs, since they show where the policy is unclear. If a rule is hard to write down, that is a sign the policy needs a decision from a person, not more automation.
Sources
Explore the product: RoutMe for Shopify, solutions and the documentation.