Vision lets AI Agent look at a damage photo and decide. The instructions telling it how to decide live in a settings panel, not your SOP library.
Gorgias shipped Vision for AI Agent. A customer sends a photo in chat or email, and AI Agent reads it. That includes deciding whether a damaged item qualifies for a refund.
Gorgias's own documentation states it plainly: customer uploads an image of a product, AI Agent evaluates refund eligibility based on your Guidance instructions. No agent opens the photo. No agent makes the call.
If you run refunds through Gorgias, this already changed how a decision gets made on your team. Most refund SOPs were never written with that decision in mind.
The procedure this changes. Written for a human triaging the ticket, checking the order, and applying policy. Vision skips straight to a decision.
Vision is a vision-model layer inside AI Agent. When a customer attaches a photo, AI Agent extracts what is in it, text, product details, visible damage, and folds that into its response.
Gorgias names four uses directly. Identifying missing or damaged items from a photo. Validating a return or refund against your Guidance instructions. Pulling an order number off a receipt screenshot. Matching a photo to a similar product in your catalog.
The second one is the SOP problem. A photo now triggers a refund decision without a human in the loop, governed entirely by whatever you typed into a Guidance field.
Guidance is a short instruction, something like if the image clearly shows a damaged item, offer a replacement or refund per our policy. Someone in your workspace wrote that sentence once, probably while setting up AI Agent, and has not opened it since.
That sentence is now a governing policy for every photo-based refund your store issues. It has no owner. No version history. No review date. Nobody outside your Gorgias admin can even see it. It is, structurally, exactly what this site has spent months describing as a shadow SOP: an unofficial procedure doing real work while the official one sits untouched.
The pattern this fits exactly. A Guidance field is a shadow SOP with better production values: it is still invisible, still unowned, and it is now making refund decisions.
Three concrete failure modes, not a hypothetical concern.
Cosmetic wear reads as damage. A shipping box with a dented corner is not the same as a broken product inside it. Vision reads the photo. It does not reliably tell packaging wear from a real defect on its own. Guidance written loosely, offer a refund if it looks damaged, does not draw that line either.
No dollar threshold. Most refund SOPs on this site use a real number: refunds under $75 approve outright, over $200 escalate to a named person. A Guidance instruction written in a settings field rarely carries that structure. Without it, AI Agent applies the same judgment to a $20 item and a $400 one.
Nobody reviews the pattern. A human agent who approves ten dubious refunds in a week gets noticed. An AI Agent approving the same ten, spread across a month, generates no signal at all unless someone is specifically looking.
Try it on one of your own procedures.
Record a process once, AI writes the structured SOP. 3 free SOPs, no credit card.
A vendor shipping a feature that changes how a decision gets made is trigger one of four: the tool changed. This is that trigger, arriving mid-August instead of on a review calendar.
Run this once now, then fold it into whatever cadence you already use for Gorgias.
| Check | What you're looking for |
|---|---|
| Pull the last 20 photo-based refunds | Filter tickets where AI Agent resolved a refund and a photo was attached |
| Open five at random | Does the photo actually support the decision made, not just resemble damage |
| Check the dollar range | Is anything above your normal one-touch threshold approved without review |
| Re-read your Guidance text cold | Would someone who has never seen your refund policy make the same call from that sentence alone |
The same audit habit, applied to the rule-based side of Gorgias. Run both on the same schedule.
Vision genuinely resolves tickets faster. A damage claim used to wait for an agent to open the photo, judge it, and reply. Now it closes in the time it takes AI Agent to read the image. For straightforward cases, that is a real improvement, not a risk.
The fix here is not disabling Vision. It is writing the Guidance sentence with the same care as any other refund rule, giving it a dollar threshold, and checking a sample of what it decided. That is maybe twenty minutes of setup. It closes a gap that otherwise sits open indefinitely, because nothing about a working feature ever prompts anyone to go check it.
Yes. Vision lets AI Agent read images customers send in chat or email, and Gorgias's own documentation states it evaluates refund and return eligibility based on your Guidance instructions, with no agent required to open the photo.
Guidance is a short instruction you write telling AI Agent how to handle a scenario, for example what to do when a customer's photo shows a damaged item. It functions as a policy, but it lives in a settings field rather than your documented SOP, with no owner, version history, or review date by default.
Pull the last twenty photo-based refunds AI Agent resolved, open five at random, and check whether the photo actually supports the decision. Check whether any approvals exceeded your normal one-touch dollar threshold without a human reviewing them. Reread your Guidance text as if you had never seen your refund policy, and see whether it still gives the same answer.
Yes, if one is not already in your Guidance instructions. A human-approved refund SOP typically has a threshold, refunds under a set amount approve outright, above it escalate to a named person. Vision-based approvals should follow the same rule rather than applying identical judgment to a $20 item and a $400 one.
3 free SOPs to start. No credit card required. See if drift detection keeps your docs honest.
Start for freeI built ReccordSOP after watching too many DTC ops teams lose months to undocumented workflows. These SOPs are battle-tested with Shopify operators running $1M to $50M brands.
Last reviewed August 15, 2026
A refund starts in your helpdesk and ends in your store. The seam between them is where refund procedures quietly break.
The Slack message pinned two years ago is the real SOP. The document nobody reads is the decoy.
Macros silently break. Customers notice before you do. Here's the audit that catches broken macros before CSAT slides.
We use essential cookies for sign-in and a small amount of analytics to improve the product. Privacy policy.