That 20% adoption figure measures something else
Twenty out of every hundred European companies with ten or more employees used artificial intelligence in 2025, up from 13.5 a year earlier, according to the Eurostat figures released in December. It sounds like critical mass until you open the breakdown of what they use it for: analysing written language (11.8%), generating images, video or audio (9.5%), and generating written or spoken language (8.8%). The AI that has landed in the European company is there to produce paragraphs and slides faster.
Break the same data down by company size and you can see why this belongs on a board agenda rather than in a budget line. In 2025, 55.0% of large EU enterprises used AI, against 30.4% of medium ones and 17% of small ones. Nearly forty points separate a large company from a small one, and that distance widens every year the internal debate stays on how many licences to buy.
Spain sits squarely at the European average, with 20.5% of companies of ten or more employees using AI last year according to the national technology observatory, almost double the 11.4% recorded in 2024. The same source adds the number that should worry a board more than the headline: roughly half of the companies using AI do it through off-the-shelf commercial software. Buy a licence, hand it round the team, minute it as an AI strategy.
So what is that 20% actually measuring? Tool penetration, not a redesigned company.
The chat lives in one person's browser tab
A Wakefield Research survey for PagerDuty, run among 1,250 office professionals in the United States, the United Kingdom, Australia and Japan and published in June 2026, found that two-thirds had used AI tools at work believing their company policy did not allow it. Some 88% had shared work information with public generative AI services, and 31% had pasted financial information or confidential documents and strategy into them.
That kind of use has two properties worth facing directly. The value evaporates the moment the person closes the tab, because the judgement they applied is written down nowhere and will not repeat itself tomorrow with the next customer, and in the meantime the firm's knowledge walks out through a door nobody has inventoried.
There is nothing to blame in people reaching for tools that make their day easier. What deserves blame is a company reading that private initiative as its AI strategy.
"An assistant amplifies whoever already knows how to do the work. A system does the work even when the person who designed it is on holiday. Between the two sits a redesign, and no licence ships with one."
What AI-native means in a company of 30 to 300 people
"Competitive advantage no longer lies in being digital, but in being AI-native," is how the trade publication IT User summed up the 2026 technology trend reports at the end of last year, with the idea of designing products, decisions and workflows with AI from the outset. The sentence is correct, and as it stands it is useless to the board of a forty-million-euro company, because it describes a destination without naming anything you touch on Monday.
At that size, AI-native shows up as four concrete traits you can verify inside one meeting.
It runs without anyone triggering it. The system wakes up on a business event, an order arriving, a ticket opening, a contract expiring, instead of waiting for somebody to open a tab and type an instruction. That is where the 24/7 comes from, and no headcount plan can match it by hand.
It feeds on what only you have. Price and margin history by customer, the real reasons deals were lost, after-sales incidents, supplier terms, and the accumulated judgement of the people who have been in the building for fifteen years.
It lives inside a decision with an owner. Approving a discount, prioritising a route, accepting an order, escalating an incident. When the system touches no decision that somebody signs, what it produces is reading material.
It has a measure and a stopping rule. One number reviewed monthly and a threshold below which the thing is switched off, with nobody's honour to defend.
Your own knowledge is the only asset the model does not have
Models are rented. Your competitor uses the same one you do, at a comparable price and in the same week, and the capability gaps between the serious options close every few months. The advantage cannot live in the model, because the model is a market service with a published rate card.
What no company can buy on the open market is the record of why your customers buy, at what price they stop buying, what breaks during installation when the order lands on a Friday afternoon, or what the market keeps asking for that your catalogue still does not cover. That sits scattered across your ERP, your CRM, the after-sales inbox and the heads of a dozen people. That set is the defensible moat, and in most mid-market companies it remains untouched.
The practical consequence is uncomfortable for anyone hoping to solve this with a purchase order: the hard work is organising and connecting that knowledge, while the model itself is close to a commodity you can contract in an afternoon.
The six decisions a board redesigns first
"Redesigning our processes" is a phrase that fits no board agenda. Redesigning six decisions does fit, and it produces the same transformation by accumulation. The selection rule stays constant: the decision repeats many times a month, it currently depends on one specific person's judgement, and there is a historical trail of how it turned out.
1. The price on every deal. Maximum discount, terms and exceptions stop depending on who happens to pick up the phone and start coming out of your own margin history by customer, product and volume. It is the decision with the most money on it and the one that leaves the deepest trail in the ERP.
2. Which opportunity gets worked and which gets dropped. A salesperson stops prioritising on instinct once the system scores each deal against what happened the last thousand times with comparable customers.
3. What you buy and from whom. Supplier terms, promised lead times against real ones, the full cost of a missed delivery. Here your own data beats any sector benchmark bought from a third party.
4. Which incident escalates and which resolves itself. After-sales and support hold the operational knowledge of the firm and almost never hand it back to the business in a shape anyone can decide with.
5. What gets produced, manufactured or planned next week. Demand, capacity and stock are the ground where a system running every night beats a spreadsheet reviewed on Mondays by a distance.
6. What you say to the market, and to whom. This is the only one of the six where the chat tab already helps today, which is precisely why almost every company started here, mistaking the appetiser for the meal.
None of the six requires an IT department. All six require somebody from the committee to sit down and describe how the decision is made today, on what data and with which exceptions, which is exactly the work nobody wants to do and the work that separates a useful system from a good-looking demo.
The minimum data capability, which is not a data lake
The reflex at this point is to commission an eighteen-month data programme, the fastest route to reaching 2028 with no decision redesigned. What you actually need to start is considerably shorter.
The sources that feed those decisions, and only those. Usually three or four ERP tables, a couple of CRM objects and one mailbox. Governing the data behind one decision does not require governing the data of the entire company first.
An owner per source and a written definition. What "active customer" means, what "margin" means, who fixes the field when it arrives empty. A good share of stalled AI projects stall right here, under a technical name and an ownership problem.
Secure retrieval over internal sources. The model queries your documents at answer time, with permissions inherited from the user and defences against prompt injection, and none of it ends up training a public model.
Traceability of what the system saw when it decided. It looks like paperwork until the day something has to be corrected, and it is also what the AI Act expects you to demonstrate about which data feeds each model and who approved it.
A mid-sized company starts with that. The data lake, if it ever arrives, will arrive because three systems in production asked for it. The minimum-viable governance model we published in July sets out the mandates and the policy stack that hold all of this up without hiring a CDO.
What you outsource and what you cannot rent
The question that brings a board this far is almost always the same one: do we have to build a technology department for this? The short answer is no, and the long answer is that a boundary is worth drawing before the first contract gets signed.
Build and run go outside without embarrassment, ERP integration, deployment, monitoring and on-call included. So does the seniority that does not justify a full salary line, taken as a fractional role inside a mandate, which is how a company of a hundred and twenty people reaches judgement otherwise found only in firms ten times its size.
Three things cannot be rented. The business judgement that decides which decision deserves to become a system. Ownership of the data and of the decision model, which stays in the house even when a third party builds it. And the objective, because a system running 24/7 optimises exactly what you asked it for, so the small print on that brief belongs to you and to nobody else.
That split avoids both the oversized IT function and total dependence on a supplier. To see how it orders a portfolio of concrete cases, the 90-day ROI roadmap lands it, and the business case for the board builds the numbers you need to get it approved.
What the committee decides this quarter
Three agreements fit into a single meeting and change the trajectory of the following year. Pick two of the six decisions and give each a name and an owner. Write down, in two pages per decision, how it is made today, exceptions included. Then set the measure that will say, one quarter from now, whether the system stays or gets switched off.
The rest is execution, and execution can be bought. The judgement about which decisions deserve to become systems is not sold in any catalogue, and it is exactly the conversation we have with the committees that call us. You can look at how we work, or bring us the decision already sitting on your table.
