Retail AI returns in 2026: which retailers put a number in writing

IKEA, Zalando and Walmart have published AI results. Kroger and Tesco describe AI programmes without a result figure. What that tells a mid-sized retailer.

Retailers that publish a result for AI in 2026 publish it for narrow jobs: Ingka Group's chatbot resolved about 47% of IKEA customer enquiries and saved nearly EUR 13 million, Zalando makes 90% of its on-site marketing content with AI, and Walmart says users of its Sparky assistant spend 40% more per order. Tesco and Kroger, describing much larger AI programmes this year, attached no outcome figure to AI at all.

That split is the most useful thing a mid-sized retailer can take from this year's announcements. The returns that companies are prepared to sign are the ones tied to one department and one number that existed before the project started. The programmes described as strategy do not come with a number, and there is no reason a 200-person retailer should expect its own version to.

What is in production, with a figure

Customer service deflection is the oldest and best documented. Ingka Group, the largest IKEA franchisee, said in June 2023 that its Billie chatbot resolved about 47% of the enquiries it received over roughly two years, around 3.2 million interactions, and that the company had reskilled 8,500 contact-centre co-workers for remote interior design and sales rather than cutting them. The figures are three years old, which says something in itself: this is a mature use, and nobody has needed to re-announce it.

Content production is the newest use with a signed number. Zalando reported in its 2025 results that AI-generated content went from near zero to 90% of its on-site marketing content in a year, with content output up 70%. Its own head of content solutions later narrowed that to the inspiration layer of the site: banners, trend pages and campaign teasers, not the product detail pages people buy from. The number is real. It is also smaller than most coverage of it.

Selling through AI assistants now has a figure in both directions, both from Walmart. When about 200,000 Walmart products could be bought inside ChatGPT, its executive vice president Daniel Danker said on the record that conversion was three times lower than for products that sent the shopper to Walmart's own site. Walmart moved the shopper back into its own checkout through its Sparky agent, and on its August earnings call chief executive John Furner said Sparky users were up 70% on a year earlier and spend 40% more per order than non-users. That is a comparison of people who chose to use the tool against people who did not, so it shows who uses Sparky more clearly than what Sparky causes.

What is described without a figure

Tesco's 2025/26 results, published on 16 April 2026, say its Save to Invest cost programme delivered about GBP 535m in the year and more than GBP 2.2bn over four years. The same document describes AI tools from its dunnhumby data science unit that adapt ranges to local tastes, and a large-scale trial of an AI assistant with about 280,000 colleagues, to reach customers later in the year. It does not attribute any part of the saving to AI.

Kroger set out its AI strategy on 22 September 2026, three days before this brief. The release covers demand forecasting, a shopping assistant, conversational search and AI literacy training for its 400,000-plus associates, and describes the goal as the right level of AI for the risk involved rather than autonomy everywhere. It gives no outcome metric for any of it.

Neither of these is evidence that the programmes are not working. Large grocers have forecast demand with statistical models for decades, and separating the AI part of a cost programme from the rest is genuinely hard. But it does mean that if your own vendor points to a large retailer as proof, you should ask which document the number is in. For most of this year's biggest announcements, there is not one.

Does this apply to a 200-person retailer?

Yes, and mainly as a filter. The uses that came with a signed number all attach to one existing cost line: contact-centre enquiries, photo and banner production, or orders through a channel you already sell in. Before buying anything, write down that line and last year's figure. If a proposal cannot name which line it moves, treat it as a strategy item, not a deployment.

What is theatre

Checkout inside someone else's chatbot, for now. Walmart's in-chat conversion figure is the only one a large retailer has published, and it was bad enough that the company took checkout back. For a smaller retailer the cheaper position is the one Shopify already offers: AI sales channels switched on by default, with the retailer remaining merchant of record, and a settings page that lets it switch them off. Keep product data clean and watch return rates from those channels before spending on anything bespoke.

Camera-based loss prevention that identifies people. The FTC banned Rite Aid from using facial recognition for surveillance for five years in December 2023, after a system it ran in hundreds of stores from 2012 to 2020 flagged customers as shoplifters, including matches against people enrolled for activity thousands of miles away. The order also made Rite Aid delete the images and any models built from them. That is the documented outcome for the category; we have not found a retailer that has published a loss-reduction figure for it that we can source.

Which vendors keep appearing

The same few names sit under most of the signed figures. OpenAI, as the channel for both Walmart's failed in-chat checkout and its replacement. Salesforce, whose customer story for Chile's Grupo Falabella, quoting named Falabella managers, reports a WhatsApp agent resolving 60% of inquiries on its own after a nine-week build. Shopify, which controls whether agents can buy from its merchants and sets that on by default. And in-house data teams: Zalando built its content pipeline and quality checks itself, and Tesco credits its own dunnhumby scientists. For a mid-sized retailer the pattern is simpler than it looks: the platform you already pay for is usually where the first useful AI arrives, switched on whether you asked for it or not.

What regulation is coming, and what is already here

Pricing is where retail's AI rules are furthest along. New York's Algorithmic Pricing Disclosure Act requires anyone showing a price set by an algorithm using a shopper's personal data to show the words THIS PRICE WAS SET BY AN ALGORITHM USING YOUR PERSONAL DATA beside it, with penalties of up to USD 1,000 per violation. The National Retail Federation sued; Judge Jed Rakoff dismissed the challenge on 8 October 2025, and the state has been enforcing it since November 2025. An appeal is at the Second Circuit. If your web shop changes prices by customer rather than by product, this already applies to your New York orders.

In Europe, the transparency duties in Article 50 of the AI Act have applied since 2 August 2026, and marketplaces are writing their own stricter rules on top. Zalando now requires invisible marking of AI-generated partner imagery by December 2026 and will reject model shots that were not photographed with a real person. A small brand selling through marketplaces will meet those platform rules before it meets a regulator.

The returns retailers are prepared to sign are tied to one department and one number that existed before the project began.

What this does not mean

It does not mean the unquantified programmes are failing, or that a retailer should wait for Tesco to publish a number before trying anything. It means the public evidence for AI returns in retail is narrower than the announcements suggest, and almost all of it comes from customer service, content and the checkout. Pricing, forecasting and loss prevention may well pay; nobody we can source has put that in writing this year.

A correction on our own pages: until today this retail hub carried figures and company names from its launch that we cannot trace to any primary source, including a markdown result, a returns-handling saving and deployment counts. We have removed them. Every figure on the hub now links to a document in this brief.