Start with a bounded support decision
An order-status email looks simple until two customers share a surname, one order ships in two parcels, or the carrier has not received the package yet. A fluent reply can still give the wrong person the wrong information.
Start with a small intake record: message reference, received time, sender, requested order number, and issue category. AI can suggest a category, such as delivery status, damaged item, or cancellation. Classification decides which queue should inspect the request; it does not prove identity or authorize a refund. Zendesk documents this distinction between ticket classification and the workflows built around it. Zendesk — About intelligent triage
Next, verify the requester through the store’s approved process. Retrieve only the order details that person is allowed to see. If the order cannot be matched confidently, request clarification without revealing another customer’s information. Shopify’s order-status experience provides a useful reference: customers may need account access or additional identity checks to view details. Shopify — Order status page
For a verified match, assemble a factual reply brief: items, separate shipments, tracking references, latest available status, and when that information was retrieved. Missing tracking should remain missing. A shipping label should never become a claim that the parcel has arrived.
The agent’s review screen should show the draft beside its supporting facts. Address changes, refunds, delivery promises, and conflicting information go to a person who can make the decision. Record the approved version and the original message reference so rerunning the workflow does not create another outgoing reply.
Before considering wider use, test two similar customer names, a split shipment, missing tracking, a repeated message, and an unavailable order source. The acceptance criteria are concrete: no cross-customer disclosure, no invented shipment event, no automatic send, and a visible review item whenever verification or retrieval fails.
Make uncertainty useful to the agent
Suppose a customer writes, “My order has not arrived, and I also want to change the address.” This is one message with two different decisions. The workflow can collect shipment facts, while the address request waits for someone to confirm whether a change is still possible. Combining both into an automatic response could turn an unanswered question into an accidental commitment.
Give the reviewer a short checklist: correct requester, correct order, correct shipment, and a reply consistent with the store’s current policy. Show which facts are missing and whether the draft includes an assumption. The agent should be able to reject the draft without losing the original request or its place in the queue.
Ownership needs a rule too. If two agents open the same request, the active owner should be visible. If the customer sends a new message while a draft awaits approval, mark that draft for another review. Approval of an earlier version should not authorize a reply that ignores new information.
Decide what the pilot will prove
Keep the first pilot to one mailbox and order-status drafts. Agree how long requests can wait for review and who handles failed lookups. Use approved sample messages rather than importing a whole mailbox just to explore the idea.
Record why agents correct drafts: missing evidence, wrong order, inappropriate wording, or an exception the design does not cover. Review unresolved cases alongside accepted drafts. These observations help decide whether another category belongs in the workflow. A high draft count alone says little about whether the team can use the output safely.