Flag damaged-item photos on return requests
When a customer attaches photos to a return request, a decision model sorts each photo by what it shows and marks where the damage is, so the returns person sees the clear cases first and the unclear ones flagged.
Uses image input, which TypeSafe AI labels a preview feature for Jev. Test it on your own sample images before relying on it.
An example pattern written by Agents Lookup. We have not yet found a public build of this exact job.
The use case
Return requests often come with a photo of the item, and someone has to look at each one before deciding what to offer. The decision Jev makes is only about what the photo shows: no visible damage, cosmetic damage, broken or not working, wrong item, or photo unclear. It also returns a Box around the damaged area so the reviewer can see it at a glance. Answers with a confidence of 0.85 or higher are labelled on the return; anything lower is labelled needs a closer look. Jev never approves a refund, exchange, or replacement. The returns person reads the request, looks at the photo and the label, and decides the outcome under your return policy.
Example workflow
- A new return request with photos arrives through your returns form or helpdesk, such as Gorgias or Zendesk.
- Make sends each photo to Jev with the image preview header and the order's product name as context.
- Jev answers a Choice question, what condition does this photo show, with the allowed answers no visible damage, cosmetic damage, broken or not working, wrong item, and photo unclear, and a Box question marking the damaged area.
- Code writes the label, the confidence, and the box coordinates back to the ticket as an internal note.
- Tickets with low confidence or photo unclear are tagged for a closer look and stay at the top of the returns queue.
- The returns person reviews each ticket, decides the outcome under your policy, and replies to the customer.
Tools
- Jev (TypeSafe AI)
- Shopify
- Gorgias
- Make
- Google Sheets
- Slack
Risks and guardrails
- Image input is a preview feature in Jev 1.14 and TypeSafe says it is not production-ready yet — guardrail: test it on a few hundred of your own photos against a person's labels before relying on it, and keep a person checking every answer until the results hold up.
- A customer could send an old or borrowed photo — guardrail: Jev only describes the photo; the reviewer checks it against the order and asks for more photos when unsure.
- The Box can point at a shadow or reflection rather than damage — guardrail: the box is shown as a hint on the image, never used to decide anything on its own.
- Labels could drift into acting as approvals — guardrail: no refund, credit, or label is ever issued from Jev's answer; the returns person takes every action.
Related recipe
Shopify Order-Exception Triage Agent: An agent reviews new Shopify orders for address problems, risk flags, stock gaps, split shipments, and customer notes, then sorts them, routes them, and drafts customer messages for a person to approve.
Want it set up for you? Request a quote.
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Shopify Dev MCP server (AI Toolkit) (official page): shopify.dev/docs/apps/build/ai-toolkit (checked 2026-11-09)
- Gorgias (official page): www.gorgias.com (checked 2026-11-09)
- Make MCP server (official page): developers.make.com/mcp-server (checked 2026-11-09)
- Slack MCP server (official page): docs.slack.dev/ai/slack-mcp-server (checked 2026-11-09)