What AI Actually Does for Retail Returns

Returns are one of those topics where everyone in fashion ecommerce nods slowly when you bring them up. The gap between how the problem is discussed externally and how it actually feels operationally is considerable.

Externally: returns are a customer experience and sustainability challenge. Companies publish commitments and talk about circular fashion. Internally: returns are an operational problem with a large and poorly understood cost base. Items come back damaged, mislabelled, or out of season. Disposition decisions are made inconsistently. The true cost per return, including processing, restocking, remarking, markdown, and write-off, is often not accurately tracked at line level.

The numbers are significant even before you get to the UK specifically. The National Retail Federation reported that US retailers saw 16.9% of annual sales returned in 2024, rising to an estimated 19.3% in 2025. Processing a return costs retailers anywhere from 20 to 65% of the item's original value. For fashion, which has structurally higher return rates than general ecommerce, the economics are worse still.

The UK is a high-ecommerce, high-returns market. Online fashion return rates run meaningfully above the cross-category average, driven by sizing inconsistency across brands and by a customer base that has been conditioned, quite reasonably, to buy multiple sizes and return what doesn't fit.

AI is starting to address this at multiple stages. Not all of it is mature. Some of it is more interesting than the headlines suggest.

Prevention: The Hardest Part

The most valuable AI application in returns is the one nobody wants to discuss because it requires confronting an uncomfortable truth: a significant proportion of fashion returns are driven by sizing inconsistency, and that sizing inconsistency exists because clothing manufacturers have not standardised sizing in any meaningful way.

AI fit and sizing prediction tools use customer measurement data, purchase history, and return history to recommend the most likely-to-fit size. They address a symptom rather than the cause. But they are addressing a real symptom. Vendors report meaningful return rate reductions in category-specific pilots when these tools are implemented well and customer adoption is good, though independent verification of the figures they publish is limited and they should be treated as directional.

The caveat on customer adoption matters. These tools work best when customers provide accurate body measurement data, which requires friction and trust that not all customer bases will offer. They also improve over time as the model accumulates return and fit feedback data for individual customers.

Virtual try-on sits at the more visually appealing end of this space. It is more mature than it was two years ago. It still struggles with the gap between "the garment looks right on a representation of my body" and "this actually fits the way I expect." Most implementations are useful as supplementary signals rather than replacements for the physical try-before-you-buy need.

Processing: Where the Near-Term Gains Are Clearest

Once a return arrives in the warehouse, AI has a more tractable problem. The variability is lower, the data is structured, and the decisions are ones where pattern recognition is genuinely useful.

AI-powered returns processing identifies an item from image or barcode, matches it against purchase and product data, assesses likely resale value in its current condition, and routes it to the most appropriate disposition path: restock, remarking, outlet, resale, donation, or disposal. It does this faster and more consistently than manual processing. Larger operations are already deploying AI disposition engines and automated grading systems for exactly this purpose.

The underlying logic is sound: if you can make better disposition decisions faster, you recover more value from returned stock, reduce write-offs, and process higher volumes without adding headcount proportionally. The gains are real. They are also, notably, less visible to customers and less interesting to press releases, which is probably why this part of the story gets less coverage than virtual try-on.

This is where the connection to wider warehouse automation matters. AI returns processing does not operate in isolation; it sits inside a broader infrastructure of warehouse management systems, robotics, and operational data. Teams deploying AI for returns processing often find it accelerates the case for related investments. See also the warehouse automation story covered in what autonomous fulfilment actually looks like.

Predictive Analytics: The Interesting Frontier

The application I think is most underappreciated is predictive return likelihood at the point of purchase.

If you have enough return history for a customer, combined with product data, you can start to build reasonable predictions about which items in a basket are likely to come back and why. This does not mean refusing to sell to customers who return frequently; that path leads nowhere good. But it does mean you can intervene differently.

You can surface additional size guidance prominently for customers who have historically returned items due to fit. You can flag that a particular product has higher return rates in certain sizes and provide additional imagery or detail. You can build more accurate demand forecasts because you are modelling net sales (items sold minus items returned) rather than gross sales. That last point connects directly to the forecasting challenge covered in how AI is changing demand forecasting in retail.

None of this is revolutionary. All of it requires data infrastructure that many retailers do not have. The AI applications are only as good as the return data feeding them, and most retailers have return data that is fragmented across systems and inconsistently attributed.

The problem with most return reason codes is that customers select from a fixed list that bears an uncertain relationship to the actual reason. "Didn't suit me" and "wrong size" are both true and both useless as distinct categories without supporting data from the return inspection. The AI prediction problem is genuinely tractable; the data quality problem is not.

The Honest Challenge

The gap between what AI returns solutions can do in controlled conditions and what they can do in a real retail operation is still significant. Real-world complexity is not fully solvable algorithmically: items arriving back undeclared, reason codes bearing no relation to the actual reason, damage assessments requiring human judgement, the operational politics of deciding which items to actively prevent returns on.

What AI does well here is reduce inconsistency and scale decision-making. A returns processing system that makes the right disposition call 80% of the time and does it at ten times the speed of manual review is genuinely valuable, even if the remaining 20% still needs a person.

The operational change required to implement these systems properly is the harder problem: clean return data, consistent condition grading, integrated systems. The AI is often the last piece rather than the first. Vendors selling AI returns solutions rarely lead with that. Retailers who have implemented them usually learn it the hard way.

---

Data sources: NRF / Happy Returns 2024 Consumer Returns in the Retail Industry; Shopify UK: Ecommerce Returns; Shopify UK: Reverse Logistics Guide