Step 01
The customer uploads a photo
Your return policy decides which reasons require one. Out of the box that is damaged, quality, and wrong item — the reasons that cost you money when they are wrong.
AI photo inspection
Read the return photo before you refund it.
Returns leak money in three places: wrong-item swaps, damage that was never there, and empty boxes. WeaveCycle puts a vision model on the customer's photo — compared against the exact variant they ordered — while the refund is still a decision.
In the review queue
Utility jacket — Rust / L
product: mismatchcondition: Creason: unsupported84% conf.
Photo shows a blue jacket; the order was placed for Rust. Heavy wear on both cuffs.
⚠ AI flagged a product mismatch — consider rejecting this item.
Step 02
Two images go to the model
The catalog photo of the exact variant they ordered, and the photo they just uploaded. Comparing against the ordered variant is what makes a color swap detectable.
Step 03
You get three verdicts
A condition grade, whether the item is actually your product, and whether the photo supports the reason they gave. Each with a confidence score and a one-line summary.
Step 04
A human decides
The verdicts land in the review queue as badges. Red flags surface so you can reject one line and approve the rest of the return.
Three verdicts
Every inspected item comes back with the same three answers.
condition: Acondition: Bcondition: Ccondition: ungraded
How worn or damaged the item looks. The grade carries through to warehouse grading, so your team confirms a call instead of making one from scratch.
product: matchproduct: mismatchproduct: unclear
Whether this is the item they ordered. A different product — or the right product in a different color or variant — comes back as a mismatch.
reason: supportedreason: unsupportedreason: unclear
Whether the photo backs up the reason they selected. An item described as damaged that photographs clean is worth a second look before the money leaves.
Rules gate first · AI never overrides
The model advises. Your policy decides.
This is a boundary in the code, not a promise in the copy. Your return policy runs and settles the decision before any model is called, and the inspection writes only to its own fields.
Your rules decide, not the model
Eligibility, refund amount, and return labels come from your written policy applied by fixed rules. That runs first, and the AI writes to its own fields — it cannot widen the outcomes your policy allows.
Every call is on the record
Each inspection logs its provider, model, token count, latency, and status. Values carry machine-readable provenance under EU AI Act Article 50, and a rule-based fallback is recorded as not AI-generated.
Shoppers are told
The return portal carries an Article 50 disclosure explaining that a model reads the description and photo, that the refund decision is rule-based, and that names and email addresses are never sent to the model.
An outage is not an outage
With no vision provider reachable, the item simply stays ungraded for a person to inspect. A return submission never fails because inspection did.
Early access
See it run against your own returns.
Leave your email and we will get in touch when we can walk through inspection on your catalog and your return reasons.
Get started
Ready to catch the returns that cost you money?
Create a workspace to try inspection on a real return, or book a call to walk through it on your own catalog.