Demo write-up

A support workflow run on six customer emails

This is a test run of the workflow we install, not a customer case study. Allbirds is not a customer and was not involved. We used their public refund and shipping policy pages because every Shopify store publishes these at /policies/. The orders, names and emails are invented.

What it did

• Order status, matching email: found order #58213, replied with the carrier, tracking link and estimated delivery date. Auto-sent.
• Return shipping cost: answered from the refund policy, including the return coverage option. Auto-sent.
• Worn shoes, return request: drafted a reply quoting the unworn and unwashed rule, then held it, because it declines the request.
• Unshipped order with a weekend deadline: drafted the status, held it, because a deadline needs a person.
• Order number from a different email address: gave no order details, asked the sender to reply from the order email, held for review.
• Complaint asking for compensation: acknowledged, asked for the order number and photos, made no decision, held for review.

The rules

Auto-send happens only for order status, returns or shipping questions, with confidence of at least 0.8, a reply grounded in the policy text or order data, a matching order email, and no refusal or deadline. Everything else waits for a person with a draft ready.

Screenshots

The review queue after the run, desktop view.

How it is built

A Python workflow that reads the store's /policies/refund-policy and /policies/shipping-policy pages, queries orders through the Shopify Admin GraphQL API with a read-only token, and uses a Google Gemini model to classify and draft. For a client we connect it to the support inbox and run it in draft-only mode first.