How AI is making fake return claims cheap, and what merchants can check first.
As you probably know, I’m in the process of transitioning my e-commerce business from focusing on human interactions to engagement that includes AI agents. Agents are becoming more active in the customer shopping experience: curating information, giving recommendations by matching requirements with information, and ultimately affecting final decisions (which could be the human’s or the agent’s, though I’d assume the vast majority for now are the former). But this piece is about a less welcome side of it: the refund scam.
What happened
Ever since we enabled an agentic storefront and structurally refreshed the format of the blogs we’ve written over the years (still relevant!) to encourage agent discovery and readability, we’ve been noticing a surge in organic traffic and conversion from both humans and bots. While we’re still observing the figures, it’s hard (and too early) to make a judgement to establish a causal relationship between the change in the make-up of the traffic and the actions we’ve taken.
And one undesirable side-effect, and for me the most off-putting part of running an e-commerce business, is the rise of scams and fraudulent cases (including the friendly ones) that has come with the increase in traffic, both paid and organic.
To be honest, I really dislike dealing with cases like these. But I still make a point of being on the front line with customers from time to time, because I need to understand exactly what went wrong with an order. Is it our problem? The supplier’s product? Logistics? Or something I don’t know of? I think it’s very important to keep doing this, so I don’t lose track.
But when customers hold us accountable for issues that aren’t properly evidenced, or can’t provide the information to support their case, and still demand a refund or return, the whole process becomes traumatic. Don’t get me wrong, most customers we’ve dealt with are super nice and sane humans. But the rise in exceptions is becoming hard to ignore.
And the highest cost isn’t just the time and effort spent investigating. It’s that it gets harder to pick out the genuine customers, the ones who have shown us support and trust and who we really don’t want to let down, and to give them the best help we can.
A case that stuck with me
In mid-August 2026, a customer claimed the product was unusable: door issues, a loose magnet, sticky residue. They wanted a full refund, no replacement, and their tone was pushy and discourteous in every message. When we asked for a video, they sent a clip, but then refused to return the used unit, even when we offered a free shipping label. They said they wouldn’t repackage it or go to the shipping store, that we’d already taken enough of their time, and that they’d already ordered elsewhere.
We were fortunate to work with a manufacturer who was really helpful. They looked into it with us and confirmed that the product was never shipped in a way that could leave residue like the one in the customer’s images.
In the end, we told the customer we wouldn’t compensate without a proper inspection, and that they were free to submit a chargeback if they wanted. They never wrote back, and no chargeback was ever filed.
But this is the sort of ultimatum I’d rather not have to give often. And all that time and effort is a cost nobody gets back.
Why the refund scam is growing
As the capability of AI grows, it becomes much more “cost-efficient” (in both time and money) for scammers to generate realistic “evidence” that’s hard for humans to tell apart without some form of help. Basically, technology has helped them achieve an outcome that far outweighs the effort they put into “generating” images, videos and messages that would previously have taken time and skill to make.
And on the merchant’s side, investigating is becoming more complex and more costly. Larger businesses can probably strike a balance by letting go and paying out some of the less obvious cases, to cut the wasted time and effort in their customer service operations. For SMEs and the growing number of one-person companies (OPCs), the loss in cash is painful, but it’s even harder to examine the evidence and handle the threats scammers often make. And enterprise-level software to weed out fraud might simply be too costly to afford.
All of this makes it really hard for e-commerce business owners to manage refunds and returns without hurting cashflow, reputation, store rating, inventory management and other related costs such as insurance. And the number of such cases might grow at an alarming rate. A few numbers that show this isn’t just me:
- Forter: AI-doctored refund images up over 15% in 2025 (WIRED; Forter)
- Signifyd: “item not as described” claims up 49% Jan–Apr 2026 vs 2025 (Signifyd State of Fraud 2026)
- Ravelin (State of Refund Abuse 2026): 65% of 6,200+ UK, German and French consumers say AI made false refund claims easier (Ravelin)
- NRF / Happy Returns: 9% of all US retail returns are fraudulent, out of $849.9B total returns in 2025 (NRF)
- WIRED covered AI damage photos on Chinese platforms; Boll & Branch and Bogg are already catching them and tightening verification (Modern Retail)
It’s something we have to plan for early.
So is there any way to detect fraudulent cases as an e-commerce operator?
Because of how painful it has been, I started researching the traits of recent e-commerce scams and how I can detect them. I’ve even gone as far as building our own system to ease detection, learning the patterns from fraudulent cases, combined with our own case patterns and Shopify’s scores, to judge how risky a return request is. This isn’t to assume orders are automatically fraudulent. It’s to help us focus our resources and time on customers with genuine needs, and to help them in a timely manner. And of course, to make sure we protect our business when necessary. The system is still being developed, and I’ll post updates later.
Here’s what we’ve identified so far, so you have a better idea when you receive requests that feel a bit off. I’ve ranked them from most to least reliable.
Strong signals: act on them
1. The delivery record contradicts the claim. If the carrier shows the parcel was delivered before the customer said it never arrived, that’s pretty telling. But if the delivery is dated after the claim, it actually backs the customer (it probably just arrived late).
2. Asking for a refund and a replacement at the same time. I can’t think of an honest reason to be made whole twice.
3. Refusing to return the item, while threatening a chargeback, legal action or bad reviews. Any one of these on its own is usually just frustration. Together, it reads more like a pressure script.
4. Recycled evidence. The exact same photo has turned up in another claim. You’ll only catch this if you keep a library of the evidence you’ve received.
5. Claims timed inside the card-dispute window. Claims made one to four months after shipping come with built-in leverage, because the customer can always go to their bank.
Real, but weigh carefully
6. Compliance theatre. Lots of “proof”, but never the one thing you actually asked for. Sending a lot isn’t the same as cooperating.
7. Customer history. A new account with a high-value first order and an instant claim is not the same as a two-year regular with one lost parcel.
8. Near-identical wording across unrelated claims. There’s research to support this one, but you need dozens of closed cases before you can trust it.
9. Hard payment failures (wrong CVV, hard address mismatch). Soft mismatches are normal for cross-border cards, so don’t read too much into those.
Looks convincing, mostly isn’t
10. Photo forensics: missing metadata, editing traces, AI-detector scores. Messaging apps strip metadata from genuine and fake photos alike, and detectors misfire on compressed images. The truth is in the behaviour and the timing, not the pixels.
How might you deal with these cases?
If you’re dealing with cases like these, here are a few steps worth trying.
Check first
- Pull the tracking record. Compare the delivery date with the date of their first “never arrived” message. If it was delivered before they complained, that’s a contradiction.
- Reverse-search their photos. Save the image and run it through Google Lens, TinEye or Yandex. This catches pictures lifted from other listings or from other people’s claims.
- Compare against your own history. Keep every evidence photo from every claim in one folder. If the same photo turns up under two different customers, that’s recycled evidence.
- Search their exact wording. Put one distinctive phrase in quotes and search for it. If identical complaints turn up on other stores, it’s a script.
- Look at the customer record. If it’s a repeat buyer with a clean history, lean generous. If it’s a first order, high value, with an immediate claim, verify before paying.
- Open your payment dashboard. A wrong CVV or a hard address-verification failure is a real signal.
- Note the timing. A claim filed one to four months after shipping, paired with threats, sits inside the bank-dispute window.
- Keep everything in writing, and clearly documented. You’ll need it for any dispute or chargeback.
Ask for evidence
- For a product defect, the best way to do it (as of now) is to ask for one continuous video showing the product, its shipping label, and a handwritten note with the order number plus a code you give the customer. This is the easiest, most effective way to verify a claim and detect AI-generated content.
- For delivery issues, always check whether your courier partner has evidence, like photos or a signed receipt. We’ve successfully identified some genuinely lost parcels this way, where the package had gone to the customer’s neighbour because of a wrong address.
- Always work with your manufacturer, suppliers or sourcing agents to verify claims (as we did in the case above). It’s the best form of support when you’re faced with customer threats, and it gives you evidence for chargeback cases.
The better way to do this going forward, I believe, is to use technology to beat technology. Some things are hard for humans to keep up with. But if we can translate what we need into a concrete solution that AI can help develop and operate, we’re definitely putting technology to good use. Less headache, happier customers, and more time for other good stuff.
