The question 1,106 Americans answered, and no French worker did

Between July 10 and July 19, 1,106 American workers watched an unusual question appear on screen. Not "do you use AI," which everyone has been asking for three years. This one asked: among these tasks, which ones did you or your team used to hand off to a contractor or a colleague, and are now handled largely or entirely by AI?

One in five ticked at least one box. That's the figure Epoch AI published on August 6, picked up since across the trade press, Europe included, always with the same careful note that these are Americans. What nobody adds is the rest of it. In Europe, that question has never been put to anyone.

A clean number, and a smaller one than reality

The survey ran with Ipsos, on its KnowledgePanel: a panel recruited by randomly sampling postal addresses, not a pool of volunteers who signed up online. The ten tasks tested come from O*NET, the US Labor Department's occupational database, chosen for their weight in actual employment. The results are weighted to represent the American workforce as a whole.

Here's a detail neither the publication nor its pickups flag: the questionnaire only shows a respondent tasks they already do with AI, five at most, drawn at random when there are more. The denominator, meanwhile, stays the entire workforce. So the counter adds up everybody, while the form only ever reaches people who already have a foot in the door.

Put another way, 20% is a floor. The real share is necessarily above it, and Epoch makes no claim to know it. Task by task, the reallocation stays modest: 7.1% for data analysis, 5.7% for reading work documents, 5.3% for record keeping.

France counts the kitchens, not the dishes

If you want to know what any of this looks like in your own office, you'll hit a closed door fast. French official statistics do measure AI at work, but they survey the company, never the person.

Insee has just published its 2025 numbers: 18% of French firms with ten or more employees say they use at least one AI technology, against 10% a year earlier and 6% in 2023. The European average sits at 20%. Among firms with 250 employees or more, it climbs to 58%.

And Insee itself takes the trouble to add a sentence nobody relays. Firms that use AI account for 59% of employment in the scope observed, which, the institute writes, "does not mean that 59% of French employees use AI directly." Counting the kitchens with an oven tells you nothing about who does the dishes. The statistician knows it, publishes it, and the number goes out into the wild regardless.

Three surveys, three scales set to three different units

That leaves the private surveys, and they do exist. Cegid had OpinionWay poll 509 French office workers in July 2025: 58% say they've already used AI at work, including 22% regularly. APEC surveyed 2,000 private-sector managers in March 2026: half use generative AI at least once a week, fifteen points more than a year before.

These numbers are useful, and they don't answer the American question. They measure usage, not substitution. Not one of them asks whether somebody stopped receiving work.

They also cover slices of the population, office workers here, managers there, rather than the workforce as a whole. Both run online, on quotas, while the American survey draws its respondents at random from postal addresses. Two methods, two populations, two questions: three reasons not to line the results up side by side.

Cegid, for its part, sells business management software, and its study report says outright that it wants to "feed a communications campaign." That invalidates nothing, it's worth flagging.

The Fondation Jean-Jaurès did the most honest work on the subject last December: a paper on French employees and AI, assembled by aggregating fifteen or so private polls, for lack of anything better. Its conclusion on non-managerial staff, who make up close to 80% of French employees, fits on one line: the available data "remain, unfortunately, more than limited." It's the same setup we already ran into at British employment tribunals: the rise in claims gets counted, the share that belongs to AI gets counted by nobody.

Europe's best instrument is pointed at the next room

Europe does have a tool the United States lacks. The European Working Conditions Survey, Eurofound's survey on working conditions, interviewed roughly 36,600 workers across 35 countries in the first half of 2024. Face to face, at home, 45 minutes per interview, on a random sample. On method alone, it beats any online panel.

Golo Henseke, a researcher at University College London, used it in late April to produce the broadest picture available of AI adoption in Europe: 12% of workers say they use generative AI for their job, with a huge spread from country to country, from under 3% to around 25%. France sits in the bottom half of that ranking, though the study never prints its rate anywhere but in a chart.

Except the survey's two consequence questions talk about "technology," not AI: has technology taken tasks away from you, has it created new ones. No colleague, no contractor, nobody at all. The camera is excellent, it's filming the room next door.

And when Henseke goes looking for a causal effect, he doesn't find one. The raw association is spectacular, the occupations that adopt the most are also the ones reporting the most change in their tasks. It collapses the moment you control for occupation and country. His reading: at this stage, AI slides into existing processes without people recognizing either a task lost or a task gained. The slicing of jobs into pieces may well be underway, it isn't showing up in the answers yet.

What we do know about French employees anyway

There is one European number on employees, and it's more uncomfortable than the one we were looking for. The global study by KPMG and the University of Melbourne, fielded in the winter of 2024-2025, 48,340 respondents across 47 countries, scientific direction handed to Melbourne Business School, contains an item few people quote.

Among French employees who use AI, 60% say they have already relied on a tool's output without assessing whether it was accurate. The global average is 66%. KPMG sells AI governance consulting, which makes it a stakeholder; the academic oversight dampens that bias without removing it. And its fieldwork is eighteen months older than Epoch's, an eternity on this subject.

So we don't know whether French employees have started handing the machine what they used to hand the desk next door. We do know that a good share of them have already turned in work they never checked.