Fashion e-commerce has a returns problem that the industry talks about in terms of logistics costs but rarely interrogates at the source. Return rates in apparel and footwear online retail run significantly higher than in other categories — commonly cited in the 20 to 40 percent range depending on the retailer's category mix, size range, and market — and the dominant industry response has been to optimise the returns process rather than reduce the underlying return rate. Better returns portals, faster refunds, pre-paid labels. These improve the consumer experience of returning things. They don't change the fundamental dynamic that sent the item back in the first place, which is almost always some variant of fit failure, colour expectation mismatch, or style ambiguity that the product description and a flat image couldn't resolve.
The interesting question is why this is still the case given how much progress has been made in computer vision, 3D body modelling, and fashion-specific attribute extraction over the last several years. Part of the answer is data structure. Fashion catalogues are notoriously difficult to work with at a machine learning level — the attributes that matter for fit (cut, fabric drape, size consistency across a range) are either not captured in structured metadata or captured inconsistently across brands. Size "L" in one brand is a different physical garment from size "L" in another, and the customer's body dimensions and wearing preferences are almost never in the retailer's data at the level of detail that would support genuine fit prediction. The data exists in scattered form — return reasons, customer reviews, sizing complaints in support tickets — but it's rarely assembled into a model that can feed back into product presentation at the point of discovery.
Dressipi, which we backed in 2022, approaches this from a fashion-specific attribute prediction angle — building models that understand garment attributes at the level of detail that actually drives fit and style match, and using those models to improve both product recommendation and product presentation for individual customers. The insight that underpins the work is that generic recommender models trained on purchase and view data learn correlations, not fashion logic. A purchase correlation doesn't tell you whether a customer bought a particular dress because they loved the silhouette or because it was the only item in their size on sale that day. Fashion-specific attribute extraction, combined with customer preference modelling at the style and fit level, allows for a qualitatively different kind of recommendation — one that can explain itself in terms the customer actually recognises.
We should be clear that AI-driven fit prediction isn't a solved problem. The accuracy of size recommendations varies substantially across body type distributions, and the cold-start problem for new customers who haven't provided sizing data or accumulated purchase history is a genuine limitation. The systems that work best are the ones where the retailer has a rich enough attribute structure in their catalogue to support meaningful inference, and where the customer base is large enough to generate reliable preference signals. That means the earliest high-value deployments tend to be at mid-to-large retailers with structured catalogues, not at niche boutiques with twenty SKUs.
The infrastructure moment we're watching is the convergence of AI-native product attribution — where AI fills in the attribute gaps that human cataloguing misses — with personalised search and recommendation. When a consumer can search for "a structured blazer that will work on someone with a shorter torso and broader shoulders" and get results that actually reflect that query at the garment attribute level, the conversion and returns dynamic in fashion e-commerce changes materially. That capability is closer to production than most people in the industry appreciate. The retailers and platforms that build toward it now will have a significant advantage in the markets where fashion e-commerce competition is most intense.