What happens when a return bot reads a claim that an item was not as described and answers with a list of reasons that include changed your mind? That is the report circulating among eBay sellers as an AI Return Assistant replaces the old dropdown of return reasons with a conversation. The rollout has been running quietly since at least September 29 and remains partial, so some buyers still see the legacy flow. As of November 17, 2026, the practical issue is narrower than whether the assistant is good: a claim coded as item not as described obliges the seller to accept the return and pay return shipping, while a remorse claim can be declined when the listing states no returns.
Two codes, two very different bills
The economics are one-sided on purpose. Under the Money Back Guarantee, a not-as-described return is one the seller must accept, and the seller pays return shipping. A remorse return can be refused if the listing says no returns, or the buyer can be asked to pay postage. So a miscoded claim lands on the side of the ledger that favours the seller, and that is exactly the trap: the buyer still believes the item was wrong, the feedback is still coming, and the platform can still rule on what happened rather than on the label a chatbot applied.
Seller complaints in public forums describe the second-order damage. Some report that a run of not-as-described claims pushed their metrics below standard, which in turn cut traffic. The claim type is treated as a service failure even when the underlying facts are disputed, so a coding error does not stay in the returns tab; it migrates into visibility.
The listing is the upstream fix
Because the assistant reads the listing, the durable fix sits there. Measurements in the body text rather than adjectives, close photographs of every flaw, and a straight-on shot of the size tag as it actually reads. Where a listing is ambiguous, the assistant has room to interpret, and interpretation tends to favour the buyer.
That connects returns to review content in a way sellers often miss. Customer review text that repeats the same complaint, a size that runs small, a colour that photographs differently, a missing part, are not merely reputation damage. They are evidence that the listing failed to state something a buyer needed. The same gap that produces a one-star review produces a not-as-described claim later, and fixing the listing reduces both.
What to log before you reply
Read the reason code and the buyer message separately, because they can disagree. Save the listing as it appeared at purchase, your pre-shipment photographs, packing photos where they exist, and the tracking record. If the words describe a real description fault while the code says the buyer changed their mind, handle it as a description fault anyway, then work the listing.
Keep the case inside the platform. Asking a buyer to email photos or to cancel and reopen later makes the record messier and weakens any later appeal. A short, factual reply that acknowledges the request and states what happens next through the platform is usually enough.
| Return reason | Who pays return shipping | Can the seller decline |
|---|---|---|
| Item not as described | Seller | No, under the Money Back Guarantee |
| Changed mind or wrong size ordered | Buyer, where policy allows | Yes, if the listing states no returns |
| Arrived damaged, not disclosed | Seller | No |
| Missing parts, not disclosed | Seller | No |
What to change this quarter
Start with the three listing fields that generate the most disputes: measurements, condition wording and what is in the box. Photograph the tag that states each one. Then review the last fifty returns and split them by what the buyer wrote rather than by the code the system applied; the gap between the two columns is your miscoding exposure.
Build a defect vocabulary
Group review text into defect types and count them by SKU. A SKU with a high density of one defect type is a listing or packaging problem, not a service problem, and it will keep producing claims until the cause is fixed. That ranking is more actionable than a star rating, which averages away the information a seller needs.
Finally, treat automation as a reason to be more explicit, not less. An assistant that reads the page rewards pages with fewer open questions. The sellers who benefit from the shift are not the ones who argue hardest after the claim arrives, but the ones whose listings leave nothing to argue about.
FAQ
Is the AI Return Assistant live for all eBay buyers?
No. Reporting describes a partial rollout running since at least September 29, so some buyers still see the legacy return-reason list.
Why does a miscoded claim cost the seller money?
Because a not-as-described return must be accepted with the seller paying return shipping, while a remorse return can be declined under a no-returns policy.
Can a seller appeal a claim that was coded incorrectly?
The platform can still review what happened rather than the label applied, so documenting the facts inside the case matters more than debating the code.
What is the fastest way to reduce these claims?
Fix the listing: real measurements, close photos of flaws, and an image of the size tag as printed. The assistant can only read what the page states.











