How Retail Brands Now Run on AI

The last time you bought something online and it arrived faster than you expected, or the return was refunded before you’d dropped the parcel, or the recommendation was suspiciously right, you were standing on top of a machine you never saw. The storefront looked the same as it did three years ago. Behind it, the work had been rebuilt.

That rebuild is happening at every size of retail and on every continent, and it has little to do with the robots and cashier-less stores that get the headlines. The real change is in the unglamorous operational core — writing the catalog, answering the customer, producing the campaign, pricing the inventory, translating the store — and the brands that rebuilt that core first are pulling away from the ones that didn’t.

The short version: The world’s retailers now run on AI across five operational layers — catalog and content, discovery, marketing and creative, customer service, and personalization — and the most instructive examples are not all American. This briefing tours real, sourced brands across Europe, Asia, the Americas, Africa, and Australia, shows where the documented results are strongest and where the claims should be read with caution, and ends with what “running on AI” actually means for a retail P&L.

Europe’s retail labs

European retailers have used AI less for spectacle than for two stubborn problems: the cost of content and the cost of returns. Zalando, the German fashion platform, built a generative shopping assistant that reached roughly ten million customer interactions in early 2026, up from six million in all of the prior year, and reported pilots in which assistant users returned goods at rates up to seven percent lower (Zalando, company-reported). In fashion, where returns can erase the margin on a sale, a seven-percent reduction is not a feature. It is the business.

On the content side, Spain’s Inditex — the owner of Zara — used AI to edit imagery shot with real, paid models and compressed its photo-production cycle from about eleven days to under forty-eight hours (FashionNetwork). Its compatriot Mango built a proprietary generative assistant in under nine months and co-created garments for more than twenty markets with AI in the design loop (WWD). Sweden’s H&M went further into the controversial: it created “digital twins” of about thirty models, who retain rights to their likenesses, to cut the cost and travel of campaign shoots — and labeled the images as AI, an honesty other brands skipped (Just-Style).

The most forward move came from grocery. In March 2026, France’s Carrefour became the first major European retailer to let customers shop inside ChatGPT — building baskets and meal plans and checking out without leaving the assistant — aimed squarely at the roughly twenty-six million French consumers already using it (The Connexion). That is the frontier the whole industry is watching: the store meeting the customer inside the AI rather than waiting for the customer to come to the store.

Asia’s scale experiments

If Europe used AI for efficiency, Asia used it for sheer volume. Japan’s Rakuten reported that AI cut zero-result searches on its fashion platform by 93.5% and lifted gross merchandise sales 5.3%, and credited AI with adding ¥10.5 billion to its 2024 operating income — a rare case of a retailer putting AI on the P&L line in an earnings disclosure (Rakuten, company-reported). Fast Retailing, the parent of Uniqlo, runs AI over more than thirty million annual customer-feedback items to iterate products and sharpen demand forecasting (analysis).

China is where the content numbers turn vertiginous. Alibaba’s AIGC tools generate on the order of two hundred million images and five million videos a month for merchants, and the company reports AI ad-bidding lifting return on investment by about twelve percent and AI-generated content lifting click-through roughly ten percent (Alizila, company-reported). JD.com’s AI digital-human livestream hosts helped more than five thousand brands lift off-peak conversion by thirty percent, and its AIGC tools cut the cost of a visual set by ninety percent and its production time from seven days to half a day (JD, company-reported). South Korea’s Coupang and Naver, meanwhile, built their personalization on home-grown models rather than US frontier labs — Naver’s shopping app runs on its own HyperCLOVA X (KED Global).

The Americas’ engine room

Latin America produced the deepest, most-quantified deployment of all, and because it surfaced in earnings rather than a press release, it is among the most credible. Mercado Libre, the region’s dominant marketplace, reported that its payments assistant handled more than nine million conversations in a single quarter with 87% resolved without a human; that its generative ad tools produced over ninety thousand creatives and lifted ad impressions 45%; and that its advertising business grew 67% year over year (PYMNTS, earnings-based). Brazil’s Magazine Luiza put generative AI behind its long-running “Lu” assistant and reported a 27% lift in return on ad spend from Google’s AI (Google / Magalu, vendor-reported).

The US anchors the volume. Amazon’s shopping assistant Rufus reached an estimated 250 million-plus users, with users reported as more than 60% more likely to complete a purchase on a trip where they used it (PYMNTS). And Walmart disclosed the single most quotable figure in retail: it used generative AI to create or improve more than 850 million pieces of catalog data, work it said would have required roughly one hundred times the headcount it actually used (Modern Retail). That is the clearest case anywhere of AI doing work that was not merely expensive but impossible at any realistic budget.

Africa and Australia: the leapfrog

The pattern reaches markets the coverage usually ignores. In Africa, Jumia tied AI chatbots and AI-driven workflows across logistics and marketing to a roughly 25% drop in site bounce rates and to its push toward profitability — reporting documented by Bloomberg and the regional press, not just the company (TechCabal). In Australia, Woolworths became the first national retailer on Google’s Gemini Enterprise, running a voicebot that cross-checks more than thirty thousand products and processes thousands of refunds and tracking requests a week (Computer Weekly).

The operator’s layer — and the small brand that proves it

The tools underneath much of this are increasingly the ones a single operator can run. Shopify’s merchant assistant Sidekick, which sits inside millions of small stores, is built on Claude, and Shopify reported AI-attributed orders across its platform rising roughly elevenfold over 2025 (Anthropic / Shopify, vendor-reported). Customer service for online retailers runs on the same layer: Tidio’s Claude-based agent automates the bulk of routine inquiries for the merchants that use it (Anthropic, vendor-reported).

Which is why the most important example on this page is the smallest. EZILDINHA, an independent Brazilian fashion label, runs the same five layers as a global retailer — catalog and content, search and AI-discovery, paid media on Google and Meta, customer communication — on an AI operating layer with a team of almost no one. The capability that Walmart applied to 850 million data points and Zalando to its returns problem is, in general-purpose form, available to a brand with five employees. The moat in retail was never the work. It was the cost of the work, and that cost has fallen by an order of magnitude.

How to read the numbers

A responsible reading requires naming where the figures come from. The strongest are the ones disclosed in earnings or audited by third parties — Mercado Libre’s ad growth, Rakuten’s P&L attribution, Walmart’s catalog figure, the Bloomberg-sourced Jumia reporting. The weakest are the platform self-reports from Alibaba and JD, which describe genuine capability but in numbers no outside party has audited. The honest summary is that the capability is real and proven across every region, the headline metrics range from rigorous to promotional, and any single one should be treated as a claim to verify rather than a fact to repeat. The brands winning are not the ones with the biggest press-release number. They are the ones with the operating discipline to apply the capability well.

What it means for a retail P&L

Strip the regional tour to its operating consequences and three lines on the income statement move. Content and catalog cost falls toward the marginal cost of compute, as Zara’s 48-hour cycle and Walmart’s 100x figure show. Service cost converts from headcount to capacity, as Tidio and Woolworths show — the savings best reinvested in the hard cases rather than pocketed. And the margin line moves through fewer returns and better matching, as Zalando’s seven-percent and Mercado Libre’s conversion gains show. The retailer that captures all three does not simply run cheaper. It runs at a scale and a responsiveness its headcount could never buy — which is exactly what an independent brand like EZILDINHA is doing to competitors many times its size.