In 2025, 19.95% of EU companies with ten or more employees used some form of artificial intelligence, according to figures Eurostat extracted in December 2025. Among large companies, 55.03%. Among small ones, 17%. Spain sits at 20.27%, almost double the year before.

That is an adoption figure. It counts who has bought, tested or plugged something in. It does not count who works differently as a result, which is a much harder number to get and a much smaller one.

Using AI tools does not make anyone an AI-driven company. Between the two sits a body of work that is not technical.

Starting with the tool is expensive

The pattern repeats. Someone tries an assistant, automates a task, connects a model to a database. It works. And nothing changes, because the workflow underneath is still the same one, with its duplication and its blurred ownership. AI speeds up whatever it finds: give it a badly designed process and you get a faster badly designed process.

Four things are worth writing down before you choose any technology:

  • Which process gets redesigned, not which task gets automated. Automating one step inside a workflow nobody has reviewed usually just moves the bottleneck further along, where it takes longer to notice.
  • Where the information the system will use comes from. If the answer is “from several places, depending on who asks”, the problem is not the model and no model is going to fix it.
  • Which decisions it can take alone and which it cannot. A system that drafts a reply and a system that sends it are two projects with two risk levels. Better to settle that before the first incident.
  • What result you expect and on what date you check. With no date a pilot runs forever, and it ends up being the reason so many trials never become process.

Loose tools or a common capability

Every department can buy its own tools, build its own knowledge base and set its own security criteria. For six months that is the fastest route. After that the company has several vendors, three copies of the same document and nowhere to look when something goes wrong.

In the group we work the other way round, and not out of architectural taste: our six companies run different businesses and hit similar problems. Identity, permissions, traceability and approved models live in the layer they all share; on top of it, each one builds its own thing. And we apply one rule: we do not authorise a trial unless it has a review date and a named person accountable for the result. If that date arrives and there is nothing to show, it gets switched off. Switching things off is the hardest part of the method.

Measure impact, not activity

Active users, queries or documents generated tell you how a tool is being used, not whether it was worth anything. What you measure is end-to-end resolution time, rework, sales response speed and the volume absorbed without adding headcount.

And you measure it by profile, because the effect is not uniform. In Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond followed 5,179 customer-support agents: access to a generative assistant raised productivity by 14% on average, with a 34% improvement among new or less skilled workers and minimal effect among the most experienced.

The average was hiding two different stories. Which is why a pilot that goes well in one team says little about the team next door, and why the roll-out order matters.

Somebody has to sign

Once a system drafts a reply, recommends an action or runs part of a process, the questions stop being technical: who writes its instructions, who validates its sources, who handles the exception and who faces the customer when it gets something wrong. Without those answers AI is a layer with no owner, and layers with no owner decay on their own.

This cannot be handed to the IT department. It touches strategy, operations, talent and the customer relationship, and all four have owners who are not the CTO.

Next year the Eurostat figure will be higher. It will still say nothing about who gained anything from it.