FOUNDER · EVERY WEDNESDAY · SEPTEMBER 16, 2026
"We use AI" has stopped meaning anything.
In 2025, 20 percent of EU enterprises with ten or more employees used at least one AI technology, against 13.5 percent a year earlier and 8.1 percent in 2023. This is Eurostat, one survey, one definition, twenty-seven countries. Eight technologies are counted, from text analysis to autonomous robots, and nobody at Eurostat has an interest in the number going up or down.
That single line now covers two companies that have almost nothing in common. One has bought a generative assistant as a company licence, which anyone may open for whatever they like. The other runs five AI technologies inside its operations, two thirds of them built in house. Both answer yes to the same question, and both land in the same 20 percent.
Now look inside that single average. Denmark reports 42 percent. Romania reports 5.2 percent. Same survey, same year, same eight technologies, same single market: an eight-fold gap between two members of the same union. Whatever that spread measures, it is not whether a technology is available. It is available everywhere.
So the adoption rate is measuring something else, and the French statistics office put a name on it. Among French companies that do use AI, 18 percent cannot state a purpose for it. Among those using a single technology, usually a generative assistant, that figure rises to 28 percent. At the other end, the companies running five or more technologies are 10 percent of users and employ nearly half of everyone working in an AI-using firm.
One caveat I would rather state than hide: these are declarative surveys. Companies say what they use. Nobody went and measured it.
Read those numbers next to the 71 percent of non-users who say they see no use for AI, and a strange symmetry appears: a large group that has not started because it sees no purpose, and a smaller group that has started without one.
This is not a technology problem. It is an execution problem.
Kiwee.care recovers unpaid hospital bills. Three payers for a single medical procedure, hundreds of thousands of case files, and a recovery team working them by hand.
They did not start by buying a tool. They ran a data diagnostic with an expert, going through their own case history and interviewing the people who do the work. It produced five candidate uses: prioritising files, optimising send times, predicting the recovery rate, optimising reminder sequences, automating invoice processing.
They kept two, chosen on impact and technical feasibility. The first is in service: it reads the statements payers send as PDFs, structures the key fields, and the recovery team stops retyping them. The second, the reminder sequences, is still being built.
Notice what the AI was asked for. Not "help us with documents". One document type, one team, one measurable saving. Bpifrance's own account of a company it supports.
THE 3 ESSENTIALS FOR YOU
- The adoption rate is not a scoreboard you can place yourself on. Denmark and Romania sit in the same survey. Two companies inside the same 20 percent can be a year of work apart. Comparing yourself to a national average tells you nothing about what to do on Monday.
- A purpose is not a use case, and a licence is not a deployment. Eighty percent of users buy off-the-shelf software, which is fine, but the purchase is where most companies stop. Across the EU the single most common use is still analysing written language, at 11.8 percent of enterprises: the easy half of the work.
- The question that separates the two ends is small and awkward: what task, whose, measured how. If nobody in the room can answer it, you are in the group that declares no purpose, whatever your licence says.
What I would do in your shoes
Stop asking whether you use AI. Ask what it is used for, and write the answers down.
Take one hour with the three people who use it most. Ask each of them for the task, the frequency, and what it replaced. You will get one of two results, and both are useful. Either you get a list of real tasks with real hours behind them, and your first case is already identified by the people doing the work. Or you get nothing specific, which means your licence is paying for an option nobody has exercised.
Then pick one task from that list and give it a decision date. Not a start date. A date on which you compare a figure to the one you had before, and decide to industrialise or stop.
What we are building at YPrtnrs points the same way. An execution engine, Y, your AI chief of staff, that prepares, drafts and organises, then leaves the decision to you. I propose, you decide.
My inbox is open for the lists that come back empty, and for the ones that surprise you: arnaud@yprtnrs.com. I read everything, and I answer personally.
In the assessments we run with mid-market companies across Europe, those that get somewhere are rarely those with the most tools. They are those where somebody could name the task.
The case for waiting is not absurd. A company in the middle of an ERP migration, or one whose operating data sits in four systems that disagree, can reasonably decide this is next year's work.
What has expired is the comfort of the adoption rate. Being inside the 20 percent was reassuring while the figure was rare. It has tripled since 2023, and a number that everyone reaches stops telling anyone apart.
