A ten-person firm goes deep on AI as often as a five-thousand-person one.

In December 2025 the European Central Bank asked 5,067 companies across twelve euro area countries how they would describe their own use of AI. Around 70 percent reported some use. Broken down: 27 percent not using it, 33 percent using it very infrequently or in pilots, 31 percent moderately, 7 percent significantly, and 2 percent unsure. One survey, one question, one base, and a central bank has no interest in the number landing high or low.

The 7 percent is the figure worth sitting with. It is the share of companies that say AI is part of how the work gets done, rather than a project running beside the work.

Now split that 7 percent by company size, which is where the survey stops being comfortable. Among companies with fewer than ten employees, 8 percent report significant use. Among those above 250 employees, 7 percent. In between, 7 percent, every time. Meanwhile the share using no AI at all collapses across the same range, from 41 percent of the smallest to 13 percent of the largest.

Read those two lines together and something breaks. Size decides whether a company has started. Size does not decide whether it got anywhere. The paper's own authors put it plainly: intensive use stays limited to a core group of firms and is spread fairly evenly across every size class.

What I find interesting is that this rules out the explanation everybody reaches for first. Budget is precisely what separates a ten-person firm from a five-thousand-person one. If depth were bought, the large end of that chart would be thick with it, and it is not.

The companies themselves say the same thing when asked what is stopping them. Among those not using AI or using it only lightly, and that base matters because it includes a great many users, just 9 percent say the cost exceeds the expected benefit. A quarter name missing skills instead. Respondents could give two answers, so the shares do not add to a hundred, but the ranking is not close.

One caveat I would rather state than hide: this is a declarative survey with a four-point scale. Companies grade their own depth. Nobody went in and measured it.

What the ones who got through actually did

Look at the gesture rather than the spend.

Auto Trader is the largest car marketplace in Britain. It did not launch an AI programme next to the business. It put AI inside the single thing roughly 11,000 car dealers already did every morning: writing the advert for a vehicle, ordering its photographs, pulling out the points a buyer actually looks for. In twelve months that produced more than 2.5 million generated descriptions and image re-orders, across 85 percent of the retailers on the platform, rising to 98 percent among those on its customer portal. Its chief product officer describes the company's AI as foundational rather than experimental. These are Auto Trader's own figures, published by Auto Trader, so treat them as an order of magnitude.

Notice what made that possible, because it is not glamorous. The task already existed. It already repeated thousands of times a day. And somebody could already count it. A pilot placed on a task nobody counts has no way of proving itself, so it quietly becomes one of the 33 percent.

THE 3 ESSENTIALS FOR YOU

Depth does not arrive with scale, so waiting to be bigger changes nothing. The chart is flat from ten employees to five thousand. Whatever separates the 7 percent from the 33 percent, a company of any size can have it or lack it today.

Widening and deepening are two different projects, and only one of them shows up in a result. Britain makes this visible on a separate measure. The national statistics office reports that among businesses with ten or more employees, self-reported use of at least one AI technology rose from around 12 percent in late 2023 to around 35 percent by June 2026, while the average number of AI technologies per adopting business moved only from around 1.4 to 1.6. Among those adopters, 10 percent describe their use as extensive. Two surveys, two definitions, two countries: adoption widened, depth barely moved. The statisticians say so themselves, warning that their own measure treats every reported use as equivalent.

The gap compounds, and that is the part with a clock on it. In the same euro area survey, companies expecting to allocate the largest share of their total investment to AI over the next year are the ones already using it significantly, at 20 percent of investment. Those using it moderately expect 11 percent. Those not using it expect 4 percent. Depth is what makes the next investment worth making, which is why the two groups drift apart rather than converge.

What I would do in your shoes…

Before approving the next pilot, name the task.

Not the department, not the tool, not the ambition. One task, the way an operations person would describe it: the thing somebody sits down and does, that already happens hundreds of times a week, and for which a number already exists somewhere in the business. How many invoices. How many adverts. How many statements retyped. How many quotes drafted.

If the number exists, you have a before and you will have an after, and the decision at the end is arithmetic rather than opinion. If nobody can produce the number, that is the more useful answer of the two: it tells you the pilot was never going to be judged, and that a pilot nobody can judge does not fail, it just renews.

Then give that task a decision date. Not a launch date. A date on which you compare the two numbers and choose to industrialise or to 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 tasks that turn out to have no number attached, 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, the ones that get somewhere are rarely the ones that spent the most. They are the ones that started from a task that already had a number.

None of this makes experimenting wrong. A third of Europe is doing exactly that, and it is how anyone finds out what a technology is good at. A company still learning what AI does well has every reason to keep a few pilots running.

What has stopped working is treating depth as something the budget will eventually deliver. Repetition is what turns a pilot into a habit, and a habit is the only thing that survives a budget review.

---