AI in ecommerce returns: how retailers catch return fraud without punishing good customers

Returns are one of the biggest costs in online retail. US retailers expected shoppers to return $849.9 billion of merchandise in 2025, or 15.8% of annual sales, and an estimated 19.3% of online sales. The same survey estimated that 9% of all returns are fraudulent ¹.

Retailers cannot simply make returns harder. In that survey, 82% of consumers said free returns were a major factor in where they shop ¹. Canada Post’s research with Canadian shoppers and merchants points the same way: an easy returns process supports sales and loyalty, while labour, transport, inspection and fraud make returns expensive ¹⁰.

That tension is why returns have become one of the busiest places for AI in commerce. In the same NRF and Happy Returns survey of large US merchants, 85% said they already use AI to detect or prevent return fraud ¹. This deep dive looks at what that AI actually does, what the early results show, and how ecommerce and direct-to-consumer (DTC) teams can use it without treating every customer as a suspect.

The business problem: one policy for every customer

Most online stores run one return policy for everyone: the same window, the same fee, the same refund timing. That is simple, but it is expensive in two directions.

It is too loose for the small group of people abusing it. Among retailers that track fraud, 71% reported more overstated return quantities, 65% more empty-box returns and 64% more decoy returns, where a cheaper or fake item is sent back in place of the real one ¹. Fraudsters now use AI too. Boll & Branch, a bedding brand that does about 80% of its business online, received a refund claim with a photo of "torn" sheets that carried an AI watermark. A review of recent tickets found a handful of other likely AI-made damage photos. When the team asked the customer to show the damage on a video call, the customer never replied ⁵. Bogg, a tote bag brand, received a second "defect" photo that was the customer’s own earlier photo with a rip added ⁵.

It is also too tight for the large majority who are honest. Riskified, a fraud-prevention vendor, reports an NRF figure that 68% of shoppers won’t come back after a refund takes more than five days ⁶. Blanket fixes, such as fees, shorter windows or holding every refund until the warehouse inspects the item, land hardest on good customers.

How AI changes the returns workflow

AI now shows up at four points in a typical online return.

1. Risk scoring when the return starts. A model gives each return a risk score based on the shopper’s history and the item, somewhat like a credit score. Happy Returns, a UPS-owned returns network, says signals include details such as the time between delivery and the start of the return ⁴. Low-risk returns move fast. High-risk returns are held for review before a refund is issued ³.

2. Checking what is in the box. Computer vision, meaning AI that reads images, compares the returned item with what was sold. Happy Returns’ Return Vision compares photos of returned items with the retailer’s catalog images to catch mismatches, such as an expensive dress returned as a cheap look-alike ³. In Reuters’ reporting, human auditors at hubs in California, Pennsylvania and Mississippi open flagged packages and photograph them, the AI compares the photos with what was expected, and people make the final decision ².

3. Reading claims and photos. For "item not received", "missing item" and damage claims, which often skip the warehouse entirely, retailers are adding checks on claim patterns, customer history and image details before approving a refund ⁵. The response at Boll & Branch and Bogg was a process change as much as a tool: ask for verification or require the product back before refunding suspicious claims, while still not requiring returns for most genuine defects ⁵.

4. Preventing the return in the first place. The cheapest return is the one that never happens. Zalando, a large European online fashion retailer, says its size and fit tools, which use machine learning to show size flags and recommendations across about 70% of its range, prevented 8% of size-related returns in 2025. Pilots of its virtual fitting room cut return rates by up to 40% ⁷.

What the early results show

The strongest public evidence so far comes from return networks and vendors, so it should be read with care.

• Happy Returns pilot. The tool was piloted from November 2025 with apparel brands including Everlane, Revolve and Under Armour ². Fewer than 1% of returns in the network were flagged as high risk, and about 10% of flagged items were confirmed as fraud, at an average of about $261 per fraudulent return ². CBS News, citing the company, put the confirmed share at roughly 13.5% of flagged returns ⁴. Supply Chain Dive reported that retailers see an average of $218 in prevented loss per flagged return, and that nearly all flagged returns at drop-off locations are verified as legitimate ³. The company planned a wider rollout in 2026 ³.
• Faster refunds for trusted customers. Riskified analysed more than 1 million refund claims from three large retailers in 2024 and classed nearly one in four refund dollars as abusive, using its own models. In a one-month pilot at one retailer, its risk-based returns feature instantly approved more than half of returns from loyal customers, those customers’ satisfaction scores rose more than 20%, and over 97% of early refunds went to people who did send the item back ⁶. These are vendor figures without independent audit.
• Fewer size returns. Zalando’s 8% reduction is company-reported and covers its size and fit work as a whole, not AI alone ⁷.

Two numbers matter most for planning. First, flagged returns are a tiny share of all returns, so the gain comes from focusing human review, not from blocking many customers. Second, most flagged returns are still legitimate ²³. A flag is a reason to look, not a verdict.

The limits and the risks

• Fraud estimates vary widely. NRF and Happy Returns estimate 9% of returns are fraudulent ¹. Trade press has cited 14% for the same partners ⁵, and Riskified’s one-in-four figure covers refund dollars on certain claim types, not all returns ⁶. Pick one definition and measure your own rate before buying anything.
• Some abuse is out of reach. Happy Returns’ tool does not address "wardrobing", where shoppers wear an item and then return it ².
• False positives cost customers. If most flagged returns turn out to be fine ³, an automatic denial would wrongly punish honest buyers, and those are the people who leave after a slow refund ⁶.
• Customer scores carry legal risk. In 2020, a shopper sued Sephora and The Retail Equation, a return-authorization service, alleging that customer data was shared without consent and used to score returns as fraudulent. The amended complaint added 12 more retailers and claims under the California Consumer Privacy Act (CCPA) ⁸. These were allegations, not findings, but they show how return scoring can be read as secret profiling.
• Quebec requires an explanation and a human review. Under Quebec’s private-sector privacy law, as amended by Law 25, a person whose personal information is used in a decision made exclusively by automated processing can be told the main factors behind it and has the right to have the decision reviewed by a person ⁹. A fully automatic return denial for a Quebec customer falls squarely in that zone.

Where human oversight stays essential

The working pattern in the evidence is "AI flags, people decide." Happy Returns’ own leaders describe the AI as a way to check a human decision ³, and Reuters reports that human teams make the final call on flagged packages ². Keep people in charge of:

• Every refund denial and account flag. A person should confirm before a customer loses a refund or is marked as a fraud risk.
• Claims based on photos alone. Ask for a video, a second photo or the item itself, as Boll & Branch and Bogg now do ⁵.
• Policy thresholds. Decide who sets the score cut-offs, how often they are reviewed and how a customer appeals.
• Data sharing. Know what customer data goes to a returns or fraud vendor, whether it is pooled across retailers and what your privacy notice says about it ⁸.

Practical next steps for ecommerce and DTC teams

1. Measure first. Split returns by reason, channel and claim type (return, damage, item not received). Estimate your own fraud rate and how long refunds take today.
2. Fix the biggest avoidable returns. If size and fit drive returns, start with better size data and recommendations before investing in fraud scoring ⁷.
3. Pilot risk-based treatment. Give low-risk customers faster refunds and route a small high-risk group to review. Track confirmed fraud, false positives, refund time and repeat purchases ²⁶.
4. Add a verification step for photo claims. Write standard wording for suspected manipulated images and a clear path to send the item back ⁵.
5. Write the appeal path. Tell customers how to contest a decision and make sure a person reviews it, which Quebec requires for fully automated decisions ⁹.
6. Check your contracts and privacy notices. Confirm what data your vendors use, whether they share it, and that your notice describes it ⁸.

Sources

1. National Retail Federation and Happy Returns (October 15, 2025). Consumers expected to return nearly $850 billion in merchandise in 2025. Surveys of 2,006 US consumers and 358 ecommerce professionals at merchants with more than $500 million in revenue. Happy Returns is a UPS-owned returns company.
2. Reuters, Lisa Baertlein and Alexandra Sarabia, via The Spokesman-Review (December 19, 2025). UPS company deploys AI to spot fakes amid surge in holiday returns. Pilot figures reported by the company.
3. Supply Chain Dive, Max Garland (January 7, 2026). UPS’ Happy Returns pilots AI tool to halt decoy returns fraud. Company-reported figures.
4. CBS News Los Angeles, Kristine Lazar (January 22, 2026). How retailers are using AI to stop return fraud. Company-reported figures.
5. Modern Retail (March 2, 2026). From Boll & Branch to Bogg, brands are battling a surge of AI-driven return fraud.
6. Riskified via Business Wire (January 12, 2026). Riskified analysis reveals 1 in 4 refund dollars is abusive. Vendor research: 1M+ claims from three retailers in 2024, classified by the vendor’s own models. Pilot at one retailer over one month.
7. Zalando (June 22, 2026). How Zalando uses technology to help customers find the right size. Company-reported.
8. Hunton Andrews Kurth, Retail Law Resource (August 17, 2020). Multiple retailers sued under CCPA for sharing data used to identify fraudulent returns. Law firm summary of a complaint, allegations only.
9. Ogletree Deakins (May 21, 2025). AI in job postings: what employers in Canada need to know. Law firm summary of section 12.1 of Quebec’s private-sector privacy act.
10. Canada Post. Rethinking returns: strategies to turn ecommerce returns challenges into opportunities. Based on surveys of Canadian online shoppers and merchants. Sample sizes not stated on the page.