Europe has caught up on adoption. It has not caught up on use

Two figures that look like they say the same thing

The European Investment Bank surveyed more than 12,000 firms in the EU and more than 800 in the US during the spring and summer of 2025. Thirty-seven percent of the European firms were using generative AI. Thirty-six percent of the American ones were.

That comparison is usually reported as the moment Europe caught up. It is accurate, and it is not the whole measurement.

Among the firms that actually use AI, eighty-one percent of the American ones apply it in more than two activities. Among the European ones the figure is fifty-five percent.

Breadth is decided after the purchase, in the months when one more function learns what the tool does with its work.

The gap cannot be explained by access to technology. The same models are sold on both sides of the Atlantic, at the same price, by the same providers. What differs is how many people in how many functions can do something with them.

Eurostat counts AI technologies and finds the same shape

The pattern turns up in a completely different kind of measurement as well.

Eurostat's survey of ICT usage in enterprises 2025 covers around 157,000 enterprises with at least ten employees across the EU27, collected during the first quarter of 2025. It is compulsory and harmonised through EU legislation, which makes it something other than a voluntary sample of management opinion. Twenty percent of enterprises used at least one of the AI technologies listed in the questionnaire. Thirteen percent used at least two. Eight point three percent used at least three, and Eurostat notes that using several technologies at the same time remains less common. The 2025 questionnaire added a category for tools that generate images, video and audio, which makes these figures a cross-section rather than a point on a trend line.

The two measures are not interchangeable. Eurostat counts how many AI technologies a firm uses, the European Investment Bank how many business activities. The numbers do not belong in one series. What can be compared is the shape. The share of firms that have started is one number. The share that has moved past the first step is a lower one, in both measurements. Two surveys, built on different methods and different populations, point at the same place, and it is not adoption.

What changed this summer, and what did not

The regulatory picture shifted at the same time, and the news has been read as a reason to wait.

Regulation (EU) 2026/1744 amended Article 4 in substance. The duty used to be to ensure a sufficient level of AI literacy. It is now to take measures "to promote the development of" that literacy, and the text adds that the duty does not require providers or deployers to guarantee any particular level in individual people.

The duty therefore stands. What went was the burden of proof. For anyone planning a programme that is a simplification: you need to show that you did something, not that every employee passed a test.

In the recitals to the same amendment the legislator then does something unusual. Competence should be "a strategic priority", it says, "irrespective of whether it is a legal obligation and irrespective of any penalties". A legal text rarely justifies itself on anything other than the law.

Anyone who read the amendment as permission to postpone read half the text. And the figures that open this article have nothing to do with compliance at all.

Breadth is a competence question, and competence cannot be bought at the speed required

Using AI in one activity takes a decision and a licence. Using it in five requires that five functions each have someone who knows what the technology can do with their own task. The second is a competence problem, and Europe's firms report it themselves as their heaviest constraint.

In the same survey, seventy-nine percent of firms name a shortage of skilled labour as a major investment barrier. That is not an answer about AI, it is an answer about investment in general, which makes it weightier rather than lighter.

The obvious way out is to recruit. Eurostat's figures argue against it. Among EU firms that recruited or tried to recruit ICT specialists, close to six in ten had difficulty filling the posts. Over the same period, a little over a fifth of firms trained their whole workforce in digital skills. The figures are for 2023 and are the most recent comparable ones.

The conclusion follows from the three figures together. The competence that produces breadth sits in the operating functions rather than in IT, the labour market does not supply it in time, and fewer than a quarter of firms build it themselves. The capacity is therefore where it has always been, in the people already on the payroll.

The people who would use the systems are asking for it

Demand for competence is not something that has to be sold downwards through the organisation. It is already there.

In ISACA's European poll of 561 IT and cybersecurity professionals, nearly three in four say staff are already using generative AI at work, up ten points in a year. Just under a third of organisations have a formal, comprehensive policy for it. Forty-two percent judge that they will need to increase their own AI skills within six months, and eighty-nine percent within two years.

Three things are true inside the organisation at once: use, demand for competence, and an absence of governance. That is an unusually favourable starting position for anyone who wants to act, because what is missing is the decision rather than the willingness.

What happens while the decision waits

Use does not wait. In KPMG and the University of Melbourne's study of more than 48,000 respondents across 47 countries, forty-seven percent of employees say they have received AI training. Almost half say they have used AI in ways that contravene their employer's policies. Sixty-six percent rely on what the system produces without evaluating whether it is right.

Abstaining does not remove the use. It removes the ability to shape it.

The three figures belong together. Use happens whatever has been decided, policies nobody was prepared for are not followed, and a person with no basis for judging an answer carries a risk that properly belongs to the organisation.

Where the effect actually comes from

Three things separate a programme that shows up in the business from one that shows up in an attendance list.

It starts from the systems you actually run. General AI knowledge has its place, but someone solving a particular task with a particular tool needs to practise on exactly that. So the work begins with a list of what is in use, including whatever arrived as a new feature inside a product you already had.

It reaches the people who make decisions in the work first. Breadth appears when one more function finds its first use, and that is decided by someone who knows the task rather than someone who knows the technology.

It is followed on use rather than on satisfaction. An evaluation form captures what the participants thought of the day, not what changed in the work the week after.

Training takes effect once it sits alongside a list of which systems are in use and a named owner for each of them. The two pieces of work belong together and are best done at the same time, because the list gives the training its subject and the training gives the list someone able to use it.

What the coming months are for

Three things are worth doing, and none of them waits on a standard that has not yet been published.

Count your business areas. Not how many licences you hold, but in how many parts of the business AI is actually used for something. That is the figure that differs, and most organisations do not have it.

Identify where the next area is. It is rarely where the technology is most interesting. It is where somebody has already tried on their own.

Give that function the competence before the tool. The order is not a matter of indifference, and the reverse order is how licences end up unused.

The full EU AI Act calendar, every date and what applies from when, sits on our EU AI Act page.