AI in Retail: The Customer Has Already Switched

The customer switched first

Let us start with the fact that governs everything else, because it inverts the usual question. In most industries, the executive wonders whether customers are ready for AI; in retail, the customers got there first. By early 2026, 31% of French online shoppers use generative AI when buying online — rising to 49% among 15-24 year-olds (Fevad / Odoxa). The detail of the usage is even more instructive than the headline: 58% of these users turn to AI upstream of the purchase — comparing, shortlisting, researching — against 27% at the moment of purchase itself (Fevad / Odoxa). The switch does not happen at the checkout: it happens in the research phase, where it is decided in whose store the basket will end up.

The phenomenon is neither French nor marginal. Over the 2025 holiday season, Americans spent a record $257.8 billion online, and traffic to retail sites from generative-AI tools jumped 693% in one year (Adobe). The trend was already measurable a year earlier — AI-referred traffic up 1,200% between July 2024 and February 2025 — and its quality improves as it grows: visitors arriving from AI view 12% more pages, bounce 23% less, and their conversion gap has narrowed from -43% to -9% in eight months (Adobe). This is not curiosity traffic: it is intent traffic, learning to convert.

The monetary scale is now established. Over the same 2025 holidays, global online sales reached $1.29 trillion, of which $262 billion was attributed to AI and agents — 20% of all orders (Salesforce); Cyber Week alone accounted for $67 billion (Salesforce). At consumer level, a quarter of shoppers across twelve countries have already used generative-AI shopping tools, and a further 31% plan to (Capgemini). One-fifth of a holiday season's global online commerce passed through an AI: for a channel that did not exist three years ago, no precedent in the sector matches that speed of installation.

One point of honesty before going further: this customer-side switch is neither uniform nor unreserved. 71% of consumers say they are concerned about how these tools use their data (Capgemini), and most usage remains concentrated on research and comparison, not payment. But for a retail executive, the operational conclusion does not change: the first point of contact with a growing share of your customers is becoming an AI — the question is whether your catalog, your product pages and your inventory are readable by it, and what you do with the traffic it sends you.

The retailer, below its own national average

Facing this equipped customer, where does the retailer stand? The statistical answer is uncomfortable and deserves to be stated plainly. In France, 13% of trade-sector companies (10 employees or more) use at least one AI technology in 2025, against 18% across all sectors (INSEE). Retail sits below the national average — in the very country where a third of online shoppers already buy with AI. The gap also widens by size: 15% of trade companies with 10-49 employees are users, against 58% of those with 250 and more (INSEE) — a one-to-four ratio between the independent retailer and the large chain.

This lag is not a French exception. In the United States, about 14% of retail trade businesses report using AI, against 17 to 20% nationally, across survey waves from December 2025 to May 2026 (US Census Bureau). In the European Union, adoption across all branches reached 20.0% in 2025 — up from 13.5% a year earlier, with Denmark at 42% (Eurostat): the general movement is accelerating, and retail is joining it more slowly than other sectors, on both sides of the Atlantic. For the executive, the reading is twofold: the lag is real — and the competitive advantage is still up for grabs, precisely because most of the sector has not structured its approach.

The economic context makes this asymmetry costly. French retail comes out of a sluggish 2024 — sales up just 1.3% in volume, including only +0.4% for food — with employment down 10,000 jobs and business failures up 14.5% across the trade sector (INSEE). Meanwhile, the online channel keeps growing: €196.4 billion in 2025 (+7%), about 12% of retail trade, 42.2 million online buyers and more than 158,000 active merchant sites (Fevad). A sector with compressed margins, whose demand is migrating to channels where AI increasingly decides who captures the visit: that is the very definition of a strategic moment — and the sector's weight, 810,000 jobs and around €225 billion of activity for French food retail alone (FCD), says what is at stake.

One last contrast completes the picture, and it is telling: among established e-merchants, generative AI is already a standard — 94% of e-merchants in the Fevad panel use it, and about seven in ten see promise in agentic commerce (Fevad). The dividing line therefore does not run between "believers" and the rest: it runs between players born in the digital channel, who tool up by reflex, and physical and mid-sized retail, which is still watching. The rest of this report is addressed to that second group.

A third of customers already shop with AI; 13% of retailers use it. That gap is somebody's market — it might as well be yours.

Plenty of pilots, few income statements

Among large retailers, the problem is no longer starting — it is finishing. The survey of 52 global retail executives measures it with rare precision: 90% have experimented with generative AI, 64% have run pilots, only 26% are scaling it in their operations, and two executives out of fifty-two report an organization-wide rollout (McKinsey). The funnel does not eliminate at the idea stage: it eliminates at the passage from demonstrator to production — exactly where the results live.

The other available measures tell the same story from different angles. 75% of retail and consumer-goods executives call AI a top strategic priority, but only 16.5% can quantify a return, and enterprise-wide deployment remains in single digits (Deloitte). Globally and across sectors, nearly nine in ten organizations use AI in at least one function, but only 37% attribute any operating-profit impact to it (McKinsey). The paradox is not unique to retail — but retail adds an aggravating circumstance: while the retailer pilots, its customer has already moved to daily use.

Most striking of all, the money keeps flowing into the top of the funnel. 81% of the 1,500 retail and consumer-products executives surveyed by the IBM Institute for Business Value already use AI at least moderately, AI spending outside traditional IT budgets was expected to surge 52% in 2025, and the sector devotes an average of 3.32% of its revenue to AI (IBM); 82% of executives still plan to increase investment within twelve months (Deloitte). A sector spending over 3% of revenue on a subject whose return only 16.5% of its executives can measure: there, in two figures, is the room for improvement — it is not technological, it is methodological.

The horizon of expectation, finally, is closing in. 68% of global retail executives anticipate the adoption of agentic AI — agents that research, compare and order on the customer's behalf — within 12 to 24 months, and 67% expect to have AI-driven personalization capabilities within a year (Deloitte). In other words, the sector itself reckons the coming generation of tools will not grant two years of thinking time. The rest of this report therefore treats a single question: where to commit, and with what discipline, to be among the 26% — not the other 74.

Six use-case families, one common logic

The potential estimates set the order of magnitude: generative AI could deliver $240 billion to $390 billion in annual value for retailers — the equivalent of 1.2 to 1.9 margin points for the whole sector (McKinsey) — and $400 billion to $660 billion including consumer goods (McKinsey Global Institute). But a sector-level potential does not land on an income statement: it is captured use case by use case. The documented deployments of the past three years outline six families sharing three traits — the task is repeated at high frequency, its current cost is measurable, and the improvement translates directly into sales captured, hours saved or markdowns avoided. The point of reviewing them is not to collect them: it is to recognize the one that matches your situation.

Discovery and assisted selling form the family where everything moves fastest, because it is the one where the customer has already switched. Amazon said 300 million customers used its shopping assistant Rufus in 2025, with monthly active users up 115% and nearly $12 billion in incremental annualized sales attributed to it (Amazon). Walmart has consolidated its AI into four "super agents" — including Sparky, already live in its customer app (Walmart); Alibaba has placed its Qwen assistant at the heart of Taobao (Alibaba). And the effect reaches mid-sized merchants: on the Shopify platform, AI-driven traffic grew eightfold in a year and orders from AI-powered searches nearly thirteenfold (Shopify). In France, Carrefour became in March 2026 the first European retailer to let customers do their grocery shopping directly inside ChatGPT — recipe ideas, availability, basket — with checkout on its own site (Carrefour). The common lesson: the retailer that makes its offer readable by AIs captures the traffic the others lose.

Product content and creation are often the best economic entry point, for a structural reason: the raw material — the catalog — already exists, and the current cost of content production is known to the cent. Zalando produces about 70% of its editorial campaign imagery with generative AI, has cut content production costs by up to 90% and brought lead times down from six-eight weeks to days (Zalando). Cdiscount has enriched 17 million product pages with generative AI — titles, descriptions, attributes (Cdiscount). This quiet workstream conditions all the others: complete, structured product pages are exactly what the shopping assistants of the first family read.

Forecasting and waste reduction form the most profitable family in food retail, because it attacks two cost lines at once: the stockout, which costs the sale, and the surplus, which costs the margin. The cross-industry references bound the gain — forecast errors cut by 20 to 50%, lost sales and stockouts down by up to 65% (McKinsey). The field confirms it: OTTO runs about 50 AI applications in production and computes 7.5 billion individual forecasts per month to steer buying and logistics (OTTO); Albert Heijn predicts daily bread demand store by store and avoided 815,000 kg of waste in a single half-year, with automated dynamic markdowns at the end of the day (Albert Heijn); H&M combines AI forecasting with RFID to "produce what sells" (H&M Group). The prerequisite is always the same: a clean sales history, at store and SKU level.

Personalization and loyalty are the family with the biggest potential — and the tightest framework. The cross-industry reference puts the most frequent effect at 10 to 15% of additional revenue, and faster-growing companies drive 40% more of their revenue from personalization than their slower peers (McKinsey). The decisive asset is first-party data: Kroger's data subsidiary powers personalization from the data of more than 62 million US households (Kroger / 84.51°); in France, Fnac Darty is overhauling its CRM with conversational and predictive capabilities and has assembled some fifty experts on a joint marketing-AI platform (Fnac Darty). It is also the family where the law frames practice most tightly — we return to it in the next section, because a recent sanction has put a price on it.

The store and its operations answer the question every physical retailer asks: what does AI do for the sales associate? The best-documented answer comes from Target, which rolled out its "Store Companion" assistant — answering operational questions on handheld devices — to teams across nearly 2,000 stores in six months, from a 400-store pilot to the whole chain (Target). The case matters as much for its speed as for its object: a full-chain rollout in two quarters is possible when the use case is narrow, the tool simple and the gain immediate for the user.

Logistics and the warehouse, finally, are the family of volumes: the effect is proportional to the flows, which makes it a second-stage workstream for a mid-sized merchant and a first-stage one for a multi-site distributor. Amazon has deployed its millionth warehouse robot and cut its fleet's travel time by 10% with a dedicated AI model (Amazon); Ocado reaches 227 units picked per labor hour (+9.1% in a year), with more than a third of items at its most advanced site picked by robots (Ocado). The cross-industry references confirm the competitive gap: early adopters of AI-driven supply chains improved logistics costs by 15% and inventory levels by 35% versus slower competitors (McKinsey).

Six families, one common thread: product data is the foundation of everything. The retailer that structures its catalog serves its customers, its associates — and the AIs that decide the traffic.

The proof, in French

The classic objection — "these are American giants' results" — no longer holds. French retailers now document their own deployments, under the same rules as everywhere else: a narrow scope, prepared data, a measured result. Three recent cases suffice to show it.

Taken together, these cases sketch the profile of the deployment that succeeds at French scale: none starts with frontier technology, all start with the data asset that already existed — the catalog at Cdiscount, purchase history at Carrefour, the customer base at Fnac Darty. The entry step is not capital-intensive: it is organizational.

The regulatory rail: four rules that already frame your use cases

Retail has a particularity that cross-sector comparisons overlook: its most profitable AI use cases touch the consumer directly — and consumer law and data law already frame them, without waiting for AI-specific texts. Four rules in force, here taken from the French and EU framework, lay the rail every project must run on — and readers elsewhere will recognize close equivalents in their own consumer-protection and privacy regimes.

The first concerns personalized pricing: since May 2022, any trader must inform the consumer when the displayed price has been personalized on the basis of automated decision-making (Légifrance). The second concerns customer reviews: publishing reviews without verifying that they come from real buyers constitutes a misleading commercial practice, punishable by two years' imprisonment and a €300,000 fine (DGCCRF) — a direct point of attention in the age of AI-generated reviews. The third touches loyalty programs: France's data-protection authority clarified in late 2025 that the right to portability covers the detail of products purchased and the amount of promotions applied (CNIL) — the data feeding your personalization also belongs to your customer.

The fourth rule is not a text: it is a price. In late December 2025, the CNIL fined a retailer €3.5 million for transmitting to a social network, without valid consent, the contact details of more than 10.5 million loyalty-program members for targeted advertising (CNIL). The message for any executive committee is clear: AI-driven personalization is one of the sector's most profitable families, and for that very reason consent, customer information and impact assessment are not formalities to handle after the pilot — they are part of the initial costing.

A project's journey — and its four breaking points

Retail AI projects that stall resemble one another. Superimposing the field reports — the 26% that scale as much as the 74% that stay in pilot (McKinsey) — a typical journey and four breaking points emerge, almost always in the same places. Each is prevented by a decision that costs nothing, provided it is taken before launch.

1. Breaking point 1 — the shop window before the foundation. The retailer launches a visible customer assistant before structuring its product pages and inventory: the tool answers poorly, the team concludes "AI doesn't work". The successful cases do the opposite — Cdiscount enriched 17 million product pages, Carrefour prepared its foundation for three years before opening its groceries inside ChatGPT (Cdiscount; Carrefour). The countermeasure: start with product data, even though it is the least visible workstream.
2. Breaking point 2 — the pilot with no starting figure. The case is chosen for its demonstration effect, not its measurable cost; with no baseline, the pilot can neither succeed nor fail — it joins the 83.5% of sector executives unable to quantify a return (Deloitte). The countermeasure: choose from one of the six documented families, where your current cost — content-production hours, stockout rate, markdowns, service contacts — is easiest to quantify.
3. Breaking point 3 — personalization before compliance. Crossing loyalty data with advertising platforms comes before consent and impact assessment: the bill can reach €3.5 million (CNIL) — not counting trust, when 71% of consumers already say they are concerned about the use of their data (Capgemini). The countermeasure: build the compliance review — personalized pricing, reviews, loyalty — into the use case's costing, not into its correction.
4. Breaking point 4 — the pilot that lasts. No decision date was ever set; the pilot is renewed quarter after quarter, "promising" without ever being settled — while AI-driven traffic grows by hundreds of percent a year (Adobe). The countermeasure: put the decision date — industrialize or stop — on the executive committee's agenda before the pilot's first day. Target went from a 400-store pilot to the whole chain in six months: speed is a decision, not an accident (Target).

Each breaking point is prevented by a decision that costs nothing — provided it is taken before starting, not after.

What the executive must decide — and when

Neither a three-year transformation plan nor a prudent watch while agentic commerce settles: the trajectory that works, judging by the deployments carried through, holds in five decisions over ninety days — each with its indicator, each answering one of the breaking points above.

1. Choose a family, not a technology. Among the six documented families, take the one where your current cost is most measurable — product content, forecasting-markdown and customer service are the most frequent entry points. Indicator: a validated starting-cost figure (hours, stockout rate, markdowns, contacts).
2. Audit the data of the chosen case — not the whole company. Product pages, sales history or loyalty base: does it exist, is it clean, since when? Indicator: a written verdict within fifteen days.
3. Run the case's compliance review. Personalized pricing, reviews, loyalty data: which obligations apply, what consent, what customer information? Indicator: a written review attached to the costing — before signature.
4. Cost in total cost. License, integration, data preparation, change management, operation — the project's cost does not stop at the tool. Indicator: a budget where data preparation and team support are visible on page one.
5. Set the decision date and instrument measurement from day one. Industrialize or stop, at twelve or twenty-four weeks depending on the family; the ending figure is compared with the starting figure, on the same scope. Indicator: the date is on the executive committee's agenda, the before/after dashboard is live from week 1.

The third scenario — doing nothing — also has a price, even though it appears on no budget. Traffic and orders are migrating to AI interfaces at a rate measured in hundreds of percent a year (Adobe; Salesforce); the retailers structuring their product data today are the ones those interfaces will recommend tomorrow; and the sector itself anticipates shopping agents within 12 to 24 months (Deloitte). That is exactly the discipline separating the 26% that scale from the 74% that pilot without end: not more resources — more method.