How H&M uses AI: demand, allocation and digital twins

KEY TAKEAWAYS Summary by the editors

1. H&M has publicly described using AI and advanced analytics since at least 2018 to tailor store assortments and to make forecasting, quantification, allocation and pricing more precise.
2. In 2018 H&M had around 200 internal and external data scientists, analysts and engineers analysing store receipts, returns and loyalty-card data, according to Retail Dive.
3. In 2019 H&M's inventory stood at 18.6% of sales against a target of 12 to 14%, which was a central reason for its AI investment in quantification and allocation.
4. In March 2025 H&M announced generative AI digital twins of 30 consenting models, who retain control over how their replicas are used.
5. H&M has not published quantified results attributing sales, markdown or inventory improvements to AI.

H&M uses AI mainly to decide how much to buy and where to send it: since at least 2018 it has described models that analyse sales, returns and loyalty data to tailor store assortments and improve forecasting, quantification, allocation and pricing. Since 2025 it has also used generative AI to create digital twins of consenting models for campaign imagery. H&M has not published quantified results for either.

What has H&M built?

Public reporting describes two distinct strands. The first is analytical and sits in merchandising. In May 2018 Retail Dive reported that H&M was customising store assortments with algorithms that analyse store receipts, returns and loyalty-card data, with around 200 internal and external data scientists, analysts and engineers reviewing purchasing patterns across stores. In March 2019 then chief executive Karl-Johan Persson said H&M was investing heavily in AI to become more precise in forecasting trends, customer demand, quantification, allocation and pricing.

The second strand is generative and sits in marketing. In March 2025 H&M announced it would create AI digital twins of 30 models who had consented, including Mathilda Gvarliani. According to FashionUnited, the models retain control over the use of their digital replicas and are paid at rates comparable to a traditional shoot, and the first images were to carry watermarks identifying them as AI-generated. The first images, showcasing denim, went live on 2 July 2025.

Under chief executive Daniel Ervér, the stated ambition is for H&M to respond faster to consumer trends with AI, as Fortune reported in January 2025.

Why did H&M invest in AI?

The trigger was inventory. In 2018 H&M was trying to end a ten-quarter sales slump and reduce markdowns, Retail Dive reported. In the first quarter of 2019 its stock-in-trade equalled 18.6% of sales, against a target of 12 to 14%, and Persson said he believed the target would be reached but could not say when. For a business that buys large volumes months ahead, small improvements in buying depth and store allocation translate directly into less stock that has to be discounted.

The motive for digital twins is different: speed and flexibility in content production. H&M's chief creative officer Jörgen Andersson framed the project as exploring generative AI to amplify creativity and reimagine how fashion is showcased.

Read also

How Nike uses AI: direct-to-consumer data and generative product creation

How does it work (data, models, process)?

H&M has not published model architectures. What has been reported is the data. According to Retail Dive, the algorithms drew on:

• store receipts and returns;
• loyalty-card data from H&M's membership programme;
• data from around 5 billion visits to its stores and websites in the previous year;
• external sources such as blog posts and search engines;
• currency movements, which the algorithms took into account.

The best-documented application was store clustering. Testing the approach in Stockholm's Östermalm district showed a predominantly female customer base with strong demand for fashionable and higher-priced items; H&M adjusted the assortment, cut the number of products in the store by 40% and dropped most menswear there, and sales improved, according to the same report. Analysis: this is classic localised assortment planning, where models propose and merchandisers decide; H&M has not described how much of the final decision is automated.

What results has H&M reported?

H&M has not published figures that attribute sales, gross margin, markdown or inventory improvements to AI. The only outcome in the public record is qualitative: sales at the Östermalm test store improved after the data-driven assortment change. For digital twins, no efficiency or cost figures have been disclosed.

Group inventory figures cannot serve as a proxy, because many factors move them. In January 2025, for example, Fortune reported that H&M's inventory had been affected by disruption in the Red Sea and a late Black Friday, neither of which a forecasting model controls.

What are the limits and open questions?

The following is editorial analysis. First, attribution: without disclosed before-and-after results, it is impossible to say how much of H&M's inventory performance comes from AI rather than from range reductions, sourcing changes or market conditions. Second, data dependency: loyalty data is central to localisation, but it only covers members, so stores and markets with low membership are harder to model. Third, governance of generative content: H&M's consent, compensation and watermarking terms address the obvious ethical questions, but the project has drawn criticism over job displacement for photographers and production crews. Fourth, speed: responding faster to trends with AI only pays off if sourcing and production lead times allow a faster response.

A further question is continuity. The executives quoted in 2019 have since moved on, and the public record does not show how the allocation and quantification work has evolved under new leadership, beyond the stated aim of responding faster to trends.

Read also

AI for merchandisers and planners: a practical guide

What can other fashion companies learn?

1. Aim AI at a balance-sheet problem. H&M tied its programme to an explicit inventory-to-sales target.
2. Use loyalty and returns data, not sales alone, to understand what each store's customers actually want.
3. Test locally before scaling: a single-store pilot made the effect visible and explainable to merchandisers.
4. For generative imagery, settle consent, ownership, pay and labelling before production starts.
5. Publish or at least track outcome metrics internally; without them, AI investment becomes hard to defend.

Frequently asked questions

How does H&M use AI for allocation?

H&M has said it uses AI to make forecasting, quantification, allocation and pricing more precise. Reported data inputs include store receipts, returns and loyalty-card data, which help decide which products each store receives.

What are H&M's AI digital twins?

They are AI-generated replicas of real models used in campaign imagery. H&M announced digital twins of 30 consenting models in March 2025; the models retain control over their use and are paid at rates comparable to a traditional shoot.

Has AI reduced H&M's inventory?

H&M has not published figures attributing inventory changes to AI. Its inventory is also affected by external factors such as shipping disruption and timing of sales events, so group figures cannot be read as proof either way.

What data does H&M use for demand forecasting?

According to Retail Dive's 2018 report, H&M's algorithms used store receipts, returns, loyalty-card data, website and store visit data, and external signals from blogs and search engines, while accounting for currency movements.

GuideThe complete guide to AI in fashion merchandising and buyingRead the complete guide