Online retail has a number that everyone in the industry knows and nobody wants to say out loud: 19.3%. That is the share of online purchases that were returned in 2025, according to the US National Retail Federation, amounting to $849.9 billion across all retail sales. Most returned items never make it back onto shelves. The cost of processing the return often exceeds the value of the refund. For years the industry called this problem a "silent killer" of margins and treated it as an unavoidable cost of doing business online. A wave of AI startups is now arguing that generative AI has finally made virtual try-on technology good enough to change that calculation.
The idea of virtual try-on has been around since the early 2010s, when retailers first experimented with overlaying clothing images onto uploaded photos. The results were unconvincing and adoption was minimal. What has changed is the underlying model quality. Generative AI can now produce realistic, detailed simulations of how a specific garment will fit a specific body type, accounting for fabric drape, colour variation under different lighting, and proportion in ways that earlier computer vision approaches could not. Zara launched a virtual try-on feature at the end of 2025. Gap integrated checkout directly into Google's Gemini earlier this year. A cluster of startups is pitching the same capability to mid-market and independent retailers who lack the resources to build it in-house.
The business case is straightforward. If virtual try-on reduces a retailer's return rate by even a few percentage points, the margin impact can be significant. A retailer processing $500 million in annual online sales and returning 20% of those items is dealing with $100 million in returns; each percentage-point reduction in that rate is worth millions in recovered margin. The technology does not need to be perfect to be commercially valuable. It needs to be good enough that customers make better-informed purchasing decisions before they click buy.
There are reasonable questions about how broadly this translates. Virtual try-on works best for clothing, footwear, and accessories where fit and appearance are the primary drivers of returns. Electronics, homeware, and goods returned because they are defective or simply not as described are not helped by a better fitting room. And even within apparel, the technology needs sustained consumer trust to change purchasing behaviour at scale. Early adopters will use it; it is less clear whether the majority of shoppers will build a habit around it or treat it as a novelty.
The broader retail disruption story is larger than returns. Bloomberg noted earlier this year that retail is likely to be among the next industries significantly reshaped by AI, with applications ranging from demand forecasting and inventory management to personalised pricing and customer service automation. Virtual try-on is visible to consumers and easy to demonstrate; most of the other applications are operational and invisible. The combined effect, if adoption accelerates, is a retail industry that can do more with substantially fewer people in logistics, customer service, and inventory roles.
That is a pattern worth watching. The AI applications getting traction in retail are not creating new shopping experiences so much as removing friction from existing ones. Returns processing is expensive: eliminate the return and you eliminate the warehouse staff processing it, the customer service agent handling the query, and the transport involved. The savings are real. So is the displacement. The two tend to arrive together, and the industry coverage focuses on the savings.