AI in retail covers five workstreams that already deliver measured results in France. Customer service, product page visuals, sales associate training, shelf availability and sales reporting. The first two ship within weeks, without touching your core systems. The next three require integration work and a budget.
This article gives the public figures from three French retailers. It also gives the results of an audit Tandem ran at a brand selling online, and the launch order that limits risk. No figure appears here without its source and its date.
Where French retailers actually stand on AI
Adoption is already massive on the online side. In the 2026 e-commerce key figures published by Fevad on 1 July 2026, 94% of French e-merchants say they use generative artificial intelligence solutions. Nearly one online shopper in three already uses AI in their buying journey, mostly to search and compare before deciding. The market was worth 196.4 billion euros in 2025, up 7%, with 42.2 million buyers.
Physical stores are following, more slowly and with infrastructure budgets.
On 18 February 2026, Carrefour announced a partnership with Vusion to digitalise every French hypermarket and supermarket by 2030. The rollout covers electronic shelf labels, connected rails and Captana micro-cameras that scan aisles continuously. They detect out-of-stocks, price discrepancies and planogram errors. Alexandre Bompard, its CEO, frames this digitalisation as the indispensable foundation of the Carrefour 2030 commerce vision. Pilots had been running since June 2025.
Customer service, the first measurable opportunity
Customer service is the fastest AI retail workstream to quantify, because the volume already exists and is already logged. Tandem measured it during an audit at Moose Knuckles, a premium outerwear brand of around 350 employees. The analysis covered several thousand real tickets, split into about ten categories.
The result, more than 35% of demand is automatable in a first version. These are order tracking, returns, sizing and availability questions. Disputes, warranty and product defects stay with human agents, because they commit the brand. On the automated share, the cost of handling a ticket drops by two orders of magnitude.
That share matters mostly because of seasonality. At a winter brand, most revenue lands in a short window. At peak, a first reply can take several hours, and every hour of waiting weighs on conversion. Absorbing repetitive requests does not shrink the team. It gives back the time it spent answering the same question three hundred times.
The phone channel follows the same logic when traffic justifies it. Success conditions and the still fragile use cases are detailed in our AI voice agent guide.
Product visuals, where AI changes the scale
Product content is the second opportunity, and often the most underrated. On that same audit, the cycle from concept to a live product page reached three months. Yet a good half of the visuals were only colourway variations of one single shoot. Generating those variants from a real master shot takes the unit cost from around a hundred euros in studio to a few cents. The three-month cycle compresses into weeks.
I filmed an interview on exactly this topic with the CEO of Meero, the French AI visual company renamed Diffusely at the end of 2024. He shares field measurements. A property listing illustrated in 5 minutes instead of 12 to 24 hours with a photographer. A large German car dealership group putting around 600,000 euros of savings on a single year. His red line is clear, quality is never traded for volume, so a check stays in place on every batch. On the most complex products, he accepts paying a high price per image. On the remaining 95%, generation plus a review is enough.
The same reasoning applies to advertising and social content. I detailed the end-to-end method in our article on generating ads with AI, prompts and iterations included.
Training sales associates, the use case few retailers saw coming
The fastest growing in-store use case is not customer relations, it is team training. In late June 2026 Leroy Merlin presented Pocket Coach, a conversational AI that has its advisers rehearse against a simulated customer. The app sits on the work smartphone. The AI plays the hesitant customer, then switches roles and gives feedback on sales posture, questioning quality and add-on selling.
The pilot figures are public and they are serious. Eight stores, more than 300 employees, four months of use. The conversion rate rises by 15 points among users compared with non-users. Add-on sales gain 9%. Customer satisfaction climbs 21 Net Promoter Score points in the pilot stores. On the largest purchases, the average basket rises by around 90 euros.
Isma Hadjersi, sales director of Leroy Merlin France, credits those results to team posture rather than to the technology. She describes a skill that now lasts three to five years, and one-off training formats that no longer keep up. Two to three simulations a week are enough in her view.
This use case is not a 2026 novelty. I was already testing the mechanic back in 2024, with a voice assistant playing the prospect and returning its analysis after the call. What changed is voice quality and the ability to industrialise across thousands of sellers. The full programme is described in our article on AI training for sales teams.
Shelves and reporting, the workstreams that need integration
Shelf availability and sales reporting are the two AI retail workstreams that touch your systems, so your timeline and your budget. On the shelf, automatic detection of out-of-stocks and price gaps requires installed, connected hardware, as in the Carrefour rollout. The gain shows up in product availability, and therefore directly in revenue lost for want of stock on the shelf.
Reporting takes another route. Asking a sales question in plain language, on top of the existing data warehouse, removes the ticket to the analytics team and the answer the next day. On the Moose Knuckles audit, this workstream came out with a payback identified between four and seven months, on recovered working time alone. It does require a clean warehouse and shared metric definitions.
| Workstream | What changes | Prerequisite | First result |
|---|---|---|---|
| Customer service | Repetitive requests are absorbed | A usable ticket history | 4 to 8 weeks |
| Product visuals | Variants come from a single shoot | A quality master shot and a style guide | 4 to 8 weeks |
| Sales associate training | Every adviser rehearses on a simulated customer | Scenarios written with the field | 1 quarter |
| Shelf availability | Out-of-stocks and price gaps report themselves | Hardware installed in store | Several quarters |
| Sales reporting | Questions are asked in plain language | A clean data warehouse | 1 to 2 quarters |
Where to start without launching an AI agent project
The most common mistake is aiming straight at the autonomous agent. On our engagements, most of the return on investment lands before that. A well configured assistant delivers immediate productivity under human control. A simple workflow applies fixed rules with maximum reliability. A workflow enriched with AI handles unstructured data while keeping the decision framed. The agent only comes after, with more leverage and more uncertainty.
I summarised these four levels in a one-pager I shared on LinkedIn in February 2026. The key point still holds. Many retail projects stop at level two or three, and that is entirely enough.
The method Tandem applies in retail has three beats. A short framing on real volumes, a first workstream delivered within weeks, then extension once the number is confirmed. That is the content of our AI audit engagements, and the approach detailed on our AI for retail and distribution page.
Which AI retail use case should you launch first
Start with customer service if your request volume concentrates on seasonal peaks. Start with product visuals if your catalogue turns over fast and the studio is the bottleneck. Both workstreams share the same profile, a clear scope, a measurement available from week one and no dependency on an infrastructure project.
Keep shelf availability and reporting for the following year, once a first quantified result has convinced people internally. And refuse any project that cannot be measured. A retail use case without a baseline is a pilot that will end like the others, with a successful demo and no decision behind it.
