You Bought the Licences and Nobody Uses Them: Why AI Stays an Experiment

In the euro area 30% of firms do not use artificial intelligence, 33% use it experimentally, 31% use it moderately and only 7% use it significantly (ECB, Adoption and investment in AI across the euro area, Occasional Paper no. 395, August 2026). That 7% is identical across every size class. The wall is not between small and large companies: it is between trying and using.

In the companies I work with the script is almost always the same, and it changes little from manufacturing to finance. The licences get bought, sometimes in exactly the right way, with a team or enterprise subscription instead of twenty personal accounts paid for at random. Then adoption never starts, because adopting is complicated, and innovation stays with the most experienced department heads or with a handful of curious people who, at a glance, are not even 5% of the workforce. Everyone hopes the half-day AI course will fix it. It does not.

How many companies actually use AI, and how many are just trying it?

The most solid data available on this question does not come from a software vendor, it comes from a central bank. The ECB added two AI modules to its SAFE survey in June and December 2025, covering around 6,000 firms in twelve euro area countries, and published the results in August 2026.

The intensity scale looks like this: 30% do not use it, 33% experimental use, 31% moderate use, 7% significant use. The source's rounding adds up to 101 and I am leaving it as it is.

On breadth, Italy ranks last of the twelve: 52% of Italian firms use AI at some level, against a euro area average of 70% and 85% in the Netherlands. The paper places Italy and Ireland in a group it calls lagging adopters.

One point sends nearly everyone who quotes these numbers off course. The ECB's 52% and the 16.4% from Istat and Eurostat are not one the other's error: they are two different questions. The ECB asks whether the firm uses AI at any level, experimentation included. Istat asks whether the firm has adopted one or more specific AI technologies (Enterprises and ICT - Year 2025, December 2025). The first measures who switched it on, the second who put it into production. Putting them in the same sentence without saying so is the fastest way to get taken apart in a meeting.

Why does AI stay stuck at the experiment stage?

Because a tool gets bought and no process gets changed. It sounds like a line from a textbook until you look at the reasons firms give themselves.

Source: ECB, Occasional Paper no. 395, August 2026.

Cost comes last. Nine per cent, behind everything else. And that fits exactly what I see in the field: the money is already out the door. The licences have been bought, often under a sensible contract. So when an AI project does not deliver, the cause cannot be that you spent too little.

The heaviest item, skills at 25%, is even sharper in Italy: among firms that considered an AI investment and then did not make it, nearly 60% point to missing skills (Eurostat, release of 11 December 2025). And here an ambiguity needs defusing, because in most people's heads "skills" means knowing how to write a prompt. That part takes two hours to learn. The skill that is actually missing is a different one: being able to look at a process, decide which piece changes, and remove the old work instead of laying the new work beside it. Which is precisely what no half-day course transfers to anyone.

The result is what I described at the start. The tool reaches everyone, the ability to use it stays with that handful of curious people who would have adopted it anyway, even if the company had bought nothing. The company paid to enable the 5% that did not need enabling.

Is being a small company a disadvantage in AI adoption?

It depends which question you ask, and the two answers point in opposite directions.

If the question is who has started, then yes, and the gap is widening. In Italy 16.4% of firms with at least 10 employees have adopted at least one AI technology, but inside that number small and mid-sized firms sit at 15.7% and large companies above 53% (Istat, Enterprises and ICT - Year 2025). The gap between small and large went from 20 points in 2023 to 37 points in 2025. I wrote about it at length in why 76% of Italian SMEs do not invest in AI.

If the question is who succeeds, then no. The ECB's 7% of significant use is the same across all size classes, and on future spending small and large firms plan similar shares, both close to 10% of total investment. Translated: once you have started, your chance of reaching a use that produces results does not get worse because you are small.

If anything, a small company has an advantage here that large ones do not, and I have watched it work: the distance between whoever decides and whoever executes is two metres. In a thousand-person company changing a process goes through a committee. In a forty-person company it is decided by the people who work inside it, on Monday morning.

What is actually missing, if it is neither the money nor the tools?

Adoption is missing, and governance is missing. Both words are worn out, so it is worth saying what they mean when I use them.

Adoption means the way of working has changed for the people who are not curious. Not for the enthusiast who found Claude on his own back in March: for the person who opens the same folder at nine in the morning and finds it organised differently, with a piece of work she no longer has to do. As long as the change only touches the people who would have gone looking for it anyway, it is not adoption, it is a tolerated company hobby.

Governance means somebody has decided, and written down, which tools are used, where the data you put into them ends up, who answers when they get it wrong, and which processes can be touched. Without it, AI does not stay out of the company: it comes in anyway, through the side doors, on employees' personal accounts. That is the phenomenon I covered in Shadow AI.

Appointing an AI Leader helps, but on its own it is not enough, and I was among the first to write about the role (the new AI roles in a small company). Change is not an appointment, it is a discipline: it is exercised with consistency, with a rhythm, with somebody who owns the change practices instead of waiting for change to happen. In Italy this is the piece missing almost everywhere. The technology is not missing and the budget is not missing: what is missing is the person, call them an innovation manager or whatever you prefer, who sets the pace and checks that the thing is still alive a week later.

Training matters, but it matters afterwards, or at best alongside. On its own it transfers knowledge to the people who already wanted it, and leaves untouched the reason the others will not use it.

How do you move from an experiment to a use that produces results?

Five things, in this order. None of them requires buying anything else, because what you needed you have already bought.

1. Pick a process, not a tool. One process, done by someone every week, one you can put a cost on. If you need a method for choosing it, the processes to automate first in a small company starts exactly there.
2. Measure first. How many hours, how many days of waiting, how many errors, today. Without that number, in three months you will not be able to say whether it went well, and nobody will be able to defend the project when someone asks what we got out of it. It is the reason four in ten AI agent projects get cancelled.
3. Remove the old work. The step almost nobody takes. If after the change the person does the new thing and keeps doing the old one too, just to be safe, you have added work rather than removed it, and the return is negative by construction.
4. Give it an owner who has time. Not the person who is best with technology: the one with the authority to change how work is done, and half a day a week to spend on it.
5. Set a rhythm, not a launch. One check-in a week for ninety days, always the same day, always with the same number on the table. That is the discipline, and it is no more complicated than that: it is just that it has to happen in week six as well, when it is no longer interesting.

Step 1 looks like the trivial one and it is the one that is almost always missing. At a company I work with, nobody could say at the outset which processes AI could touch: there was no list, there was a feeling that something could be done. We wrote them down one at a time, and then changed a few of them in the back office. On those, processing time came down by up to 40%. It did not happen because we had bought a better tool, because the tool was the same one as before: it happened because for the first time somebody knew which processes they were looking at.

Ninety days is the right measure because it is long enough to change a process and short enough that nobody forgets why they started. If you want to see what numbers a project has to hit to pay back inside that window, the AI project ROI calculator runs the maths on your own hours and costs.

Frequently asked questions

How many European companies actually use artificial intelligence?

In the euro area 30% of firms do not use AI, 33% use it experimentally, 31% use it moderately and 7% use it significantly (ECB, Occasional Paper no. 395, August 2026, SAFE survey of around 6,000 firms across 12 countries). Italy ranks last of the twelve for the share of firms using it at any level, 52% against a euro area average of 70%.

Why do AI projects in companies fail to deliver results?

Because a tool gets bought and no process gets changed. The reasons firms give themselves confirm it: missing skills 25%, data and privacy 19%, incompatibility with existing systems 19%, while cost comes last at 9% (ECB, August 2026). If cost is the smallest problem, the return is not missing because you spent too little: it is missing because nobody removed the old work the tool was meant to replace.

Is being a small company a disadvantage in AI adoption?

For starting yes, for succeeding no. In Italy 15.7% of small and mid-sized firms have adopted at least one AI technology, against more than 53% of large companies (Istat, Enterprises and ICT 2025). But among those that have started, the share of significant use is 7% across every size class (ECB, August 2026): the ceiling is common, it is not a question of size.

The number that matters is not the one about Italy

"Italy ranks last" describes a country, not your company, and it does not tell you what to do on Monday. The two numbers that concern you are different ones: cost weighs 9% and skills weigh 25%, and the 7% that makes it does not change with size.

Put together they say one thing. You already have permission to succeed. What is missing is somebody, inside the company, keeping the rhythm for ninety days.

More on costs, returns and AI investment in the ROI and investment hub.