How can AI find the root cause of fashion returns?
Reason codes alone rarely explain why a style comes back. Combining them with review text and product data, AI can point to fixable causes such as wrong size charts, misleading imagery or fabric surprises.
KEY TAKEAWAYS Summary by the editors
- AI finds the root cause of fashion returns by joining three data sets: structured return reason codes, unstructured customer text (reviews, return comments, service chats) and product attributes from the PIM.
- Coresight Research found that the average US online apparel return rate reached 23.4 percent in 2025 and that nearly 70 percent of shoppers who returned clothing bought online cited size and fit as the reason.
- Large language models are now used to extract fit, style and fabric feedback from reviews at scale; Amazon described a seller tool that does this to help brands understand return reasons and spot size chart defects.
- A reason code such as 'too small' is a symptom; the root cause is usually a product data or product design issue, for example an incorrect size chart, a missing fabric stretch attribute or imagery that misrepresents colour.
- Root cause analysis only pays off if findings are routed to an owner (PIM, content, design or sourcing) and the effect on the return rate of the affected style is measured afterwards.
AI finds the root cause of returns by reading what customers say about a product (reason codes, reviews, return comments and service chats) and linking it to the product's own data, such as size chart, fabric, fit attributes and imagery. The output is not a better report of return rates but a short list of fixable causes per style, each with an owner. It works only where reason codes are captured consistently and product data is clean enough to compare like with like.
Why are reason codes not enough to explain fashion returns?
Most retailers ask for a reason when a customer returns an item, but the menu is short and the answers are blunt. According to a Coresight Research report covered by FashionUnited, the average US online apparel return rate reached 23.4 percent in 2025, and nearly 70 percent of shoppers who returned clothing bought online cited size and fit. That tells a merchandiser that fit is the problem in general, not why a specific trouser comes back.
The same study found that around 40 percent of US shoppers had abandoned an online apparel purchase because of confusing or missing product information. Returns and abandoned baskets therefore often share a cause: the product page did not tell the customer what they needed to know. Reason codes record the symptom ('too small', 'not as pictured'); the cause sits in the product data, the content or the garment itself.
Category level data shows how uneven the problem is. A report published in February 2026 by returns analytics firm Returnalyze found denim return rates above 51 percent early in the holiday season and dress return rates of up to 48.2 percent in the pre Black Friday window, while size bracketing (ordering several sizes to keep one) stayed between 5 and 9 percent across categories. A single average hides exactly the styles that need attention.
What data does AI need for returns root cause analysis?
The analysis combines structured and unstructured inputs. None of them is new; what AI changes is the ability to read the text at scale and connect it to product attributes.
| Data source | What it adds | Typical quality problem |
|---|---|---|
| Return reason codes | Volume and direction of the problem per style and size | Short menus, default answers, inconsistent codes across channels |
| Free text return comments | The customer's own words ('sleeves short', 'colour darker') | Often empty, multilingual, mixed with service complaints |
| Product reviews and ratings | Fit and fabric feedback from customers who kept the item too | Biased towards very happy or very unhappy customers |
| Service chats and emails | Detail on defects, sizing questions before purchase | Hard to link to a specific SKU without an order reference |
| PIM attributes and size charts | Measurements, fabric composition, fit type, stretch | Missing or wrong measurements, attributes not standardised |
| Imagery and copy | What the customer was shown and promised | Colour deviation, model size not stated, outdated descriptions |
The step that makes analysis possible is a reliable key between all of them: the SKU or style and colour code. If returns are booked against a generic article and reviews against a marketplace listing, the AI has nothing to join.
How does AI turn reviews and comments into root causes?
Language models are well suited to the first step: extracting structured signals from text. Amazon described in January 2024, as reported by The Paypers, a fit insights tool for selling partners that uses large language models to extract and aggregate customer feedback on fit, style and fabric, helping brands understand return reasons and identify size chart defects. The same announcement described the use of language models to extract product sizes from size charts and to remove duplicate information and auto correct missing or incorrect measurements.
A practical workflow for a brand or retailer looks like this:
- Extract. Classify each comment, review and chat into a fixed taxonomy: runs small or large, length, sleeve, shoulder, fabric feel, transparency, colour versus image, quality defect, delivery damage.
- Link. Attach every signal to style, colour and size, and add the PIM attributes (fit type, composition, stretch, measurements) plus the image set that was live at the time.
- Compare. Flag styles whose return rate or complaint mix deviates from comparable styles in the same category, fit and price band, not from the overall average.
- Explain. Ask the model to summarise the evidence for each flagged style with counts and example quotes, so a human can verify it in minutes.
- Route. Assign each cause to an owner: size chart and attributes to the PIM team, imagery and copy to content, pattern or grading to design, defects to sourcing and quality.
- Measure. Track the return rate of the fixed style against its own history and against similar styles after the change.
Which root causes can be fixed in product data and content?
Many causes do not require a new garment. A size chart that lists body measurements for one style and garment measurements for another, a missing note that a style runs small, no information on the model's height and size, or an image that renders a colour too bright can all be corrected in the PIM or the content system within days. Generative tools can help rewrite fit notes or flag copy that contradicts the attributes, but they should not invent measurements; those must come from the technical specification.
Other causes belong upstream. Repeated complaints about a narrow shoulder across several styles built on the same block point to a pattern issue for design. Pilling or seam complaints concentrated in one production lot point to a sourcing and quality issue. The analysis is most valuable when it makes these patterns visible across seasons, because a single style rarely has enough returns to prove a pattern on its own.
What are the limits and risks of AI returns analysis?
- Small samples. A style with a few dozen returns produces confident sounding summaries from thin evidence. Show counts next to every claim.
- Biased text. Reviews over represent strong opinions; return comments are often left blank. Combine sources rather than trusting one.
- Hallucinated causes. A model asked 'why' will always produce an answer. Require it to cite the comments it used, and let a human confirm before changes go live.
- Privacy. Service chats and comments can contain personal data. Strip or pseudonymise it before analysis and check the processing basis under GDPR.
- Fraud and policy abuse. The NRF estimated that 9 percent of all returns in 2025 were fraudulent. Such returns distort product signals and should be filtered out where they can be identified.
How should teams measure whether the fixes work?
Define the metric before the change: return rate by style and size, share of fit related reasons, and conversion on the product page, since better information can also raise sales. Compare the changed style with its own pre change period and with a control group of similar styles that were not touched, because seasonality and promotions move return rates on their own. The NRF projected total US retail returns of $849.9 billion for 2025 and estimated that 19.3 percent of online sales would be returned; at that scale even modest, verified improvements on high volume styles are worth the analytical effort, but only if each fix is tracked to an outcome rather than reported as an activity.
Frequently asked questions
What is the most common reason for clothing returns?
Size and fit. A Coresight Research report found that nearly 70 percent of US shoppers who returned clothing bought online cited size and fit as the reason. The underlying causes vary by style, from inaccurate size charts to fabrics that behave differently than expected.
Can AI read product reviews to reduce returns?
Yes, language models can classify review and comment text into fit, fabric and quality signals and link them to specific styles. Amazon has described a seller tool that aggregates fit, style and fabric feedback to help brands understand return reasons. The reduction in returns depends on acting on the findings.
What data do I need for returns root cause analysis?
At minimum, return reason codes and return comments linked to style, colour and size, product reviews, and PIM attributes such as measurements, composition and fit type. A consistent SKU key across systems is essential, otherwise text signals cannot be matched to products.
How do you know if a returns fix worked?
Compare the style's return rate and the share of fit related reasons before and after the change, and against similar styles that were not changed. This separates the effect of the fix from seasonality, promotions and changes in customer mix.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- FashionUnited: Sizing intelligence is strategic priority as brands prepare for AI-driven commerce
- NRF: 2025 Retail Returns Landscape
- The Paypers: Amazon Fashion uses AI to improve online clothing shopping
- Retail Dive (press release): Returnalyze report finds returns prevention strategy shields retailers
