Paris, a Tuesday morning in September. Victor, head of transformation at a services group, scrolls through a dashboard in front of his executive committee. The figure he shows looks great: 71 % of his people opened the AI assistant at least once this month. He smiles. Then the CFO asks the only question that matters: “Victor, how much did productivity actually move?” Silence in the room. Victor has no answer, because no one measured it.
That scene replays in almost every company I work with. It sums up France’s 2026 AI paradox: massive adoption, barely any harvest.
Two-thirds adopting, one-third gaining
The numbers come from a Banque de France survey run between February and April 2026, covering companies with at least 20 employees. More than two-thirds of them, 67 %, say they use generative AI tools (Banque de France). The barrier to entry has all but vanished: a ChatGPT subscription, a Copilot already baked into the office suite, and thousands of employees are equipped within days.
Then comes the wake-up call. Only one company in three, 31 %, reports a positive effect on productivity (ZDNet). Two in three handed out the tool and saw nothing move. The gap between the adoption rate and the gains rate is the real story. You do not close it by handing out more licenses.
Company size widens the divide further. In large groups, generative AI adoption reaches 83 %, against 64 % in SMEs of 20 to 49 employees (ZDNet). And the moment you leave generative AI for the more classic kind, the one that requires integrating structured data into the information system, the fracture grows: smaller firms stay well behind the large ones. That integration AI, less spectacular than a chatbot, is exactly what produces durable gains. The chatbot dazzles fast. The rebuilt back office pays off for years.
The sandbox trap
Why such a gap? Because usage stays locked in the sandbox. People write an email a little faster, summarize a meeting note, ask for a rewrite. Useful gestures, never structural ones. The workflow itself has not moved an inch. We laid 2026 technology on top of 2015 processes, then act surprised the machine does not race ahead.
The survey confirms it: half of the companies that have not taken the plunge blame their inertia on the absence of an identified use case (ZDNet). They look for what to automate and come up empty. That is the symptom of an organization waiting for the tool to name its own target. It never will. An AI assistant has no idea which bottleneck is choking your value chain. You do.
I write in my book that the invariant of innovation was never the technology (My book, chapter 14). Technology changes constantly, it creates fads, its only invariant is the acceleration of its own change. The real invariant is the human being: their fears and their working habits. A company that does not innovate is already dead, it just does not know it yet. And handing out a tool without rethinking the work is a box-ticking exercise.
The invariant was never the tool
Ask the wrong question and you get the wrong transformation. “How do I replace five people with AI?” breeds fear and defensive projects. “How can five people do the work of fifty with AI?” opens a different worksite. The first question bets on subtraction. The second bets on amplification, and amplification is what moves the lines.
Any use of AI that fails to answer one of two questions is a distraction: how do I raise the productivity and quality of what I deliver in my current role, and what can I do today that was impossible yesterday? The rest is tool noise.
Panic does the opposite work. The Stanford AI Index measured AI-related nervousness rising from 38 % to 52 % in a single year (Stanford). That fear ties directly to the survival instinct. When a leader deploys AI without saying clearly what it is for, the vacuum fills itself with the dread of being replaced. An anxious employee invents no new uses. They protect themselves. Psychological safety stays the first condition for someone to dare say “here is the task I hate, and here is how AI could take it over.”
The real decoupling: France versus Europe, not the United States
Here is a correction that changes the diagnosis. The comfortable story says France is falling behind the “American model.” The figures say otherwise. According to the French Treasury, 61 % of French companies used at least one advanced digital technology in 2025, against 78 % in the United States (French Treasury). The 17-point gap is real. But the European Union average sits at 77 %, one point off the Americans. France is not falling behind a distant American model. It is falling behind its direct neighbors, Germany and Spain included.
That nuance matters, because it shifts the responsibility. Comparing yourself to Silicon Valley giants licenses fatalism: they have the capital and the scale. Comparing yourself to a German manufacturer or a Spanish SME removes that excuse. The same Treasury note names the digital lag as the first factor in France’s productivity decoupling from the United States (French Treasury). The problem is within reach. It sits at home, inside our organizations.
Measure, or nothing
There remains the headline number, the one that makes committees dream. Everyone quotes the BCG-Harvard study: 758 consultants, 25 % faster on their tasks. The trap is believing that figure transfers to every job. Another study, covering roughly 25,000 workers, brings the real average time saved back to around 3 %. The gap between 25 % and 3 % is exactly where the disillusion of management is born.
The economist Daron Acemoglu, a Nobel laureate, set a guardrail back in April 2024: generative AI would add no more than 0.9 % to GDP over ten years, and about 0.5 % to total factor productivity (MIT). Read the figure carefully: it is a level effect accumulated over the decade, not a point of annual growth, and Acemoglu presents it as an upper bound. Without deep organizational restructuring, AI will deliver no macroeconomic miracle.
Same lesson at company scale: assumptions are highly dangerous and should be killed with experimentation as early as possible. Before you deploy, measure the starting point. During, measure real usage, not the number of licenses opened. After, measure the gain on the task and the sentiment of the people who lived it. Two questions per activity are enough: is the expected gain realistic, and is it verifiable? Without an answer, you are not steering a transformation. You are funding a fad.
The essentials
Remember three things:
- Adoption is not transformation. 67 % of companies use generative AI, only 31 % get a productivity gain from it. The rest are handing out licenses.
- The lock is human and organizational first. As long as workflows go unredesigned and fear goes untreated, the tool stays in the sandbox.
- No measurement, no steering. Set the zero point, track real usage, verify the gain task by task. An unmeasured gain is an imaginary gain.
France holds the tools in its hands. What it lacks is the nerve to rebuild its methods instead of dressing old processes in a coat of AI. Innovate or agonize, the choice is yours.
References
- (My book) https://philippeboulanger.com/book/
- (ZDNet) https://www.zdnet.fr/actualites/ia-en-entreprise-une-adoption-massive-en-france-mais-des-gains-de-productivite-qui-stagnent-502829.htm
- (Banque de France) https://www.banque-france.fr/fr/actualites/lia-sinstalle-dans-les-entreprises-francaises-une-diffusion-rapide-mais-des-gains-de-productivite
- (French Treasury) https://www.tresor.economie.gouv.fr/Articles/2026/09/01/le-numerique-principal-facteur-du-decrochage-de-la-productivite-francaise-par-rapport-aux-etats-unis
- (MIT) https://shapingwork.mit.edu/wp-content/uploads/2024/04/Acemoglu_Macroeconomics-of-AI_April-2024.pdf
- (Stanford) https://aiindex.stanford.edu
