How can AI reduce fashion returns? 12 levers ranked by likely impact
From size advice and fit flags to return prediction and try-on, a checklist of twelve AI levers for reducing online fashion returns, ranked by evidence and likely impact, with the data each needs.
KEY TAKEAWAYS Summary by the editors
- The AI levers with the strongest published evidence for reducing fashion returns are size and fit tools: size flags, personal size recommendations and measurement-based fit comparison.
- Zalando reports that size advice reduced size-related returns by 10% compared with items without it, and that pilots of its 3D virtual fitting room for jeans reduced return rates by up to 40%.
- Feeding return reasons back into product data and product development tackles the root cause, but its effect shows up over seasons rather than weeks.
- Return prediction at checkout and generative try-on are promising but have less public evidence of impact in fashion, so they should be tested with control groups.
- Every lever depends on clean data: SKU-level return records, structured return reasons and accurate graded garment measurements.
AI reduces fashion returns most reliably where it helps shoppers choose the right size and gives them an accurate picture of the product. The highest-impact levers, on current evidence, are size flags, size recommendations and measurement-based fit tools, followed by closing the loop from return reasons to product data. Return prediction, generative try-on and automated returns handling can add value, but have thinner public evidence and should be tested before scaling.
Why focus on returns now?
The US National Retail Federation estimated in October 2025 that US retail returns would reach $849.9 billion in 2025 and that online returns would equal 19.3% of online sales. Fashion is above that average. Zalando wrote in June 2026 that European online fashion return rates can reach about 50%, with size and fit accounting for up to half of those returns, and that jeans can reach 65%. Each return carries transport, handling, repackaging and markdown costs, plus emissions from extra shipments.
How are the 12 levers ranked?
The ranking below is an editorial assessment, not a measured league table. It weighs three factors: published evidence of impact in fashion, the share of returns the lever addresses (size and fit being the largest), and how directly the lever changes the purchase decision. Actual results depend on category, customer base and data quality.
| Rank | Lever | What it addresses | Evidence level | Key data needed |
|---|---|---|---|---|
| 1 | Personal size recommendation | Wrong size choice | Published retailer results | Purchase and return history, fit feedback, garment measurements |
| 2 | Size flags (runs small or large) | Systematic sizing deviations | Published retailer results | Return reasons by style, fitting model notes |
| 3 | Measurement-based 3D fit comparison | Size and fit for high-variance items such as jeans | Published pilot results | Body measurements, graded garment data |
| 4 | Return reasons fed into product development | Root causes in patterns and grading | Qualitative, slower to show | Structured reasons by SKU and size |
| 5 | Body measurement capture | Unknown body dimensions | Adoption data published | Consented photo or video capture |
| 6 | Product content accuracy checks | Not as described, colour, material | Qualitative | Images, attributes, return comments |
| 7 | Bracketing detection in the basket | Multiple sizes ordered to compare | Research evidence | Basket composition, customer history |
| 8 | Return prediction with nudges | High-risk baskets | Research evidence | Historical order lines with outcomes |
| 9 | Generative virtual try-on | Uncertainty about look and style | No published return data | Product images, shopper photo |
| 10 | Analysis of return comments with language models | Hidden quality and description issues | Qualitative | Free-text reasons, reviews, service tickets |
| 11 | Fraud and abuse detection | Wardrobing and fraudulent returns | Problem size published | Return history, item condition data |
| 12 | Returns routing and resale triage | Cost and value recovery after the return | Operational | Item condition, stock and demand data |
Which levers have the strongest evidence?
Size and fit tools lead because they address the largest return driver and have published results. Zalando says size flags tell customers whether an item runs smaller or larger than expected, using brand measurements, return reasons and findings from trained fitting models. It reports that size advice reduced size-related returns by 10% compared with items without it, that its size and fit solutions prevented 8% of size-related returns in 2025, and that pilots of its virtual fitting room for jeans, which uses body measurements to build a 3D avatar, reduced return rates by up to 40%. These are company-reported figures from one large retailer, but they are the most detailed public data available.
Body measurement capture is an enabler rather than a lever on its own. Zalando says more than 1.5 million customers have tried its tool, which predicts measurements from two photos or a video taken on a phone.
Which levers attack the root cause?
Size advice helps customers navigate inconsistent sizing; it does not fix the sizing. Levers 4, 6 and 10 aim at the cause. When return reasons show that a style consistently comes back as too small in the hips, the pattern or grading can be corrected for the next order or season. When language models summarise free-text return comments and reviews, they can surface issues such as misleading colours in product images or fabrics that pill. These levers rarely show quick wins, because changes take a production cycle to reach customers, but they reduce returns for every channel at once.
Where is the evidence thinner?
Return prediction has solid research behind it. A 2024 study on data from a German online retailer of occasion wear reached a balanced accuracy of up to 0.86 on future orders when customer history was available, but only 0.61 without it. Prediction is only useful if it triggers an intervention that changes behaviour, and published results for such interventions in fashion are scarce.
Generative try-on is spreading quickly. Google launched an AI try-on experiment in May 2025 that shows apparel listings on a shopper's own photo, using a model Google says accounts for how fabrics fold, stretch and drape. Google has not published return data for it, and image-based try-on does not use precise measurements, so its effect on size-related returns is unproven.
Fraud is real but a minority of the problem: the NRF estimated that 9% of US returns in 2025 were fraudulent. Detection tools protect margin but do not address the larger share of honest returns caused by fit and expectation gaps.
How should a fashion business prioritise?
- Measure first: calculate return rates by category, style and size, and the share of returns by reason.
- Fix return reason capture so size direction and “not as described” are recorded consistently.
- Deploy size flags and size recommendations in the highest-return categories.
- Set up a monthly loop from return data to product development and merchandising.
- Test return prediction, try-on and comment analysis with control groups before scaling.
- Track net impact: fewer returns, but also conversion, customer satisfaction and cost.
What data does every lever depend on?
- Order lines linked to return outcomes at SKU and size level.
- Structured return reasons, captured the same way in every channel.
- Accurate graded garment measurements and fit descriptions.
- Consent and privacy controls for body measurements and photos.
- A single owner for returns data across e-commerce, logistics and product teams.
Frequently asked questions
What is the most effective way to reduce fashion returns with AI?
On published evidence, size and fit tools: size recommendations, size flags that warn when items run small or large, and measurement-based fit comparison. Zalando reports return reductions from each of these, which address the largest single driver of fashion returns.
How much can AI reduce returns?
Results vary by category and data quality. Zalando reports that size advice cut size-related returns by 10% versus items without it, and that 3D fitting room pilots for jeans reduced return rates by up to 40%. Brands should test on their own data rather than rely on vendor estimates.
Does virtual try-on reduce returns?
Measurement-based 3D try-on has published pilot results. Generative image try-on, such as Google's, has no published return data yet, and because it does not use precise measurements, its effect on size-related returns is unproven.
What data do I need to start reducing returns with AI?
SKU and size-level return records, structured return reasons and accurate garment measurements per size. Without these, most AI levers cannot learn which products and sizes cause returns.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- NRF: 2025 Retail Returns Landscape
- Zalando Corporate: How Zalando uses technology to help customers find the right size
- Zalando Corporate: How Zalando leverages technology to help customers find the right size
- SCITEPRESS (ICAART 2024): Garment Returns Prediction for AI-Based Processing and Waste Reduction in E-Commerce
- Google Blog: AI Mode shopping and virtual try-on (I/O 2025)
