TL;DR: Returns are one of the few ecommerce problems where AI has a credible, measurable story: size-and-fit recommendations and product-level return signals attack the single largest cause of apparel returns. But the headline "40% fewer returns" figures come almost entirely from vendors, the gains are concentrated in apparel and footwear, and most of the value depends on clean product and returns data that many retailers don't have yet. Treat returns AI as a data project first and a model project second.
Why returns became an AI problem
The scale is hard to ignore. The National Retail Federation and Happy Returns forecast that US consumers would return about $850 billion of merchandise in 2025, a 15.8% overall return rate, with roughly 19.3% of online sales coming back. The same report estimated that around 9% of all returns are fraudulent.
For online apparel, the numbers are worse. Industry sources consistently put size and fit at more than half of apparel returns, and "bracketing" — ordering several sizes with the plan of returning most of them — is a normalized shopping behavior. Statista data cited by returns-platform vendors suggests nearly half of consumers bracket when sizing is unclear.
Every one of those returns carries reverse logistics, inspection, repackaging, markdown risk and, increasingly, a shipping cost the retailer absorbs. For mid-market ecommerce brands operating on single-digit net margins, a few points of return rate is the difference between a profitable category and a subsidized one. That is why returns have moved from an operations nuisance to a board-level metric — and why AI vendors have piled in.
Where AI is genuinely delivering value today
Size and fit recommendation
This is the most mature and best-evidenced category. The approach is not exotic: models combine a shopper's purchase and return history, self-reported measurements or reference garments, and garment-level fit data to recommend a size. The underlying logic has academic support — a randomized field experiment published in Manufacturing & Service Operations Management found that giving online shoppers better fit information changed purchase and return behavior in measurable ways, and research teams at large fashion platforms (for example, Zalando's published work on "SizeFlags") have shown that flagging items that run small or large reduces size-related returns.
Vendor-reported results are larger — fit-tool providers claim that shoppers who use a size finder bracket dramatically less — but these figures come from self-selected users, who are already more engaged buyers. Expect real but smaller gains when you measure against a proper holdout group.
Product-level return intelligence
A quieter but often more valuable use case: using machine learning models to spot products with abnormal return rates and cluster the reasons. Classifying free-text return reasons, reviews and customer-service tickets ("fabric thinner than pictured", "sleeves short", "color differs from photo") lets merchandising teams fix the listing — or the product — rather than eat the returns. This is lower-glamour than virtual try-on, but it is cheaper, works outside apparel, and addresses the cause rather than the symptom.
Better product content
Returns often start with a listing that oversold or under-described the item. Generative AI is now widely used to draft richer product descriptions, fit notes and attribute data at scale. When it is grounded in accurate source data and reviewed by a human, it helps. When it is used to "enhance" imagery or invent attributes, it drives exactly the mismatch returns you are trying to prevent.
Returns fraud and policy abuse
With NRF estimating roughly 9% of returns as fraudulent, anomaly detection on return patterns — serial returners, wardrobing, receipt fraud, empty-box claims — is a legitimate AI application. It shares much of its infrastructure with AI-powered fraud detection at checkout, so retailers who already score orders for payment risk can often extend those pipelines to returns.
Where it's still overhyped or premature
Virtual try-on as a returns cure
Virtual try-on (VTO) has improved dramatically — Google expanded photo-based try-on in 2025, and large retailers such as Zara and ASOS have been piloting generative try-on experiences. Vendor and analyst-aggregated figures claim 25–40% return reductions. Be careful with those numbers:
- Most are vendor case studies, not independent controlled trials.
- Generative try-on is very good at showing how a garment looks and much weaker at showing how it fits. A photorealistic image of a shirt on your body is not evidence the chest measurement is right.
- Adoption is uneven: many shoppers never use the feature, so a large reduction among users can translate to a small reduction in the overall return rate.
VTO can help with "didn't like how it looked" returns, which are real. It is not a substitute for fit data.
"One model fixes everything" platforms
Returns have many causes — damage in transit, late delivery, wrong item picked, buyer's remorse, genuine defects. A tool focused on fit will not move a return rate driven by fulfillment errors. Before buying, decompose your returns by reason code; many retailers discover their biggest lever is a warehouse process, not a model.
Blanket automated refund decisions
Using AI to auto-deny returns or silently flag "bad customers" is technically easy and commercially risky. False positives hit loyal customers, and opaque policies generate complaints and chargebacks. Keep humans in the loop for anything that denies a customer money.
Realistic implementation risks — and what mitigates them
Dirty returns data. Free-text reasons, inconsistent reason codes across channels, and returns processed in a separate system make training and measurement unreliable. Mitigation: standardize reason codes, capture structured reasons at the point of return, and join returns back to orders and SKUs before you model anything.
Missing garment data. Fit models need garment measurements, not just size labels. Many brands — especially multi-brand retailers — don't have them. Mitigation: start with your top-returning SKUs and collect measurements for those first.
Integration cost. Fit widgets, returns portals, OMS, WMS and ecommerce platforms all need to share data. Without clean API integration, recommendations are stale and results can't be attributed. Mitigation: budget for integration explicitly — it is often the larger half of the project.
Measurement theater. Comparing users of a fit tool to non-users overstates impact. Mitigation: run a randomized holdout (a share of traffic never sees the tool) and measure net return rate and conversion together; a tool that lowers returns by discouraging purchases hasn't created value.
Customer trust and privacy. Body measurements and photos are sensitive. Under regimes such as GDPR or India's DPDP Act, you need clear consent, purpose limitation and retention rules. Mitigation: process images transiently where possible and avoid storing what you don't need.
How to evaluate whether your business is ready
Ask these questions before you sign a contract:
- What share of your returns is fit- or expectation-driven? If it's under a third, fit AI is not your first lever.
- Can you join a return to its order, SKU, customer and reason code in one query? If not, fix that first — every other step depends on it.
- Do you have garment-level measurements for your top 20% of SKUs by return volume?
- What is your return cost per unit, fully loaded? Without it, you can't build an ROI case or compare vendors.
- Can you run an A/B holdout on your storefront? If not, you won't know whether the tool worked.
- Who owns the outcome? Returns sit between merchandising, CX, operations and finance. Successful programs have a single owner with a return-rate target.
A retailer that can answer yes to most of these is ready to pilot. One that can't will get more value from a data and integration sprint than from any AI subscription.
Build, buy or combine?
For most mid-market retailers, the sensible pattern is to buy the commodity pieces (fit widgets, returns portals, generic VTO) and build the layer that is specific to you: a unified returns data model, reason-code classification tuned to your catalog, product-level return alerts for merchandisers, and the dashboards finance trusts. That custom layer is also what lets you swap vendors later without losing your history.
Sources
- National Retail Federation and Happy Returns, 2025 Retail Returns Landscape (nrf.com/research/2025-retail-returns-landscape)
- Gallino, S. and Moreno, A., "The Value of Fit Information in Online Retail: Evidence from a Randomized Field Experiment," Manufacturing & Service Operations Management (pubsonline.informs.org/doi/10.1287/msom.2017.0686)
- "SizeFlags: Reducing Size and Fit Related Returns in Fashion E-Commerce" (arxiv.org/abs/2106.03532)
- PYMNTS, "Retailers Bet on AI Fitting Rooms to Slash Costly Returns" (2026)
- True Fit and Redo, industry commentary on bracketing (vendor sources; figures treated as vendor-reported)
Conclusion: where Syslabs fits
Syslabs works with mid-market retailers on the unglamorous parts that decide whether returns AI works: unifying order, returns and product data, API integration between storefront, OMS and returns platforms, custom software for return-reason analytics, and machine learning models tuned to a specific catalog. We'd rather help you prove where your returns actually come from than sell you a model you can't measure.