Where the field stands, and what the real obstacle is
Eurostat, 2025 reference year (the most recent figures available, updated in June 2026), as a percentage of enterprises:
| Company size | EU27 | Denmark |
|---|---|---|
| 10–49 employees | 17.0% | 37.5% |
| 50–249 employees | 30.4% | 58.3% |
| 250+ employees | 55.0% | 74.5% |
| More than 10, in total | 20.0% | 42.0% |
Two things can be read out of this. One: four out of five small companies in the EU have not even started, so you have missed nothing. The other: inside one and the same economy the gap between large enterprises and small ones is more than threefold — the same tools, the same market. The difference is therefore not about the technology.
The obstacle that gets quoted most often is not the one you would expect either. The same survey asks separately why companies have not started:
| The obstacle | EU27 | Denmark |
|---|---|---|
| The necessary expertise is missing | 7.8% | 13.3% |
| Data protection concerns | 5.8% | 7.4% |
| The legal consequences are unclear | 5.9% | 7.8% |
| The cost looks too high | 4.2% | 2.9% |
Lack of knowledge comes up nearly twice as often as money — and that holds in almost every member state that publishes the figures. Good news, because a knowledge gap is cheaper to close than a capital gap.
You do not need an AI expert. Three numbers that back it up
One: the market buys finished products, it does not build them. In the same Eurostat round, ready-to-use commercial software is the most common source of AI in all nineteen member states that publish the breakdown — between 39% and 78% of the companies using AI bought it off the shelf. Building your own is a minority route everywhere: in Spain 10.1% of all firms use ready-made AI software against 3.4% that develop their own, in Hungary 7.6% against 1.4%. You do not have to write code, you have to choose.
Two: the biggest gain arises among the weakest performers. A customer-support study covering 5,179 agents (Brynjolfsson, Li and Raymond, Quarterly Journal of Economics, 2025) found an average of 14% more issues resolved per hour alongside the staggered rollout of an AI assistant. The breakdown is far more interesting than the average: +34% for novice and low-skilled staff, and essentially zero for the experienced ones. The authors explain it by the model spreading the proven solutions of the better colleagues. In a small company that is exactly the situation: a lot of people doing a lot of different things, none of them at professional level.
Three: the decision is on your desk, not on an expert's. If the biggest obstacle is a lack of expertise, then the first step is not buying a tool but deciding which task the machine is suited to at all. There is a method for precisely that: can AI do this task? walks you through it in a minute.
The one hard decision: which single process
There is one point in the three-month plan you have to think about; everything else follows from it. The criterion is three words:
- Repetitive. It happens at least ten times a week. If it happens twice a month, you will not have enough observations at the end to see anything.
- It has a measurable output. A count, a timestamp, an error rate. If the process ends with "and then somebody takes a look", then the measurement has to be built first, not the AI.
- Text-based or decision support. By the Eurostat figures, the technologies that work on text lead the field across the EU — text mining 11.8%, image and audio generation 9.6%, generating written or spoken language 8.8% — and the most common area of use is marketing and sales (6.9%). That is no accident: this is where a rollout meets the least friction.
And more important still: one process, not five. A randomised experiment run on a global online marketplace (Fang et al., 2025) tested seven generative AI applications on the same platform, on samples ranging from thirty thousand to 13.7 million consumers. The result: the pre-sales chatbot lifted sales by 16.3%, search-query refinement by 2.9%, product-description generation by 2.1% — while the marketing push message and the ad-headline generator moved nothing at all. The same company, the same technology, the same week. If you roll out all seven at once and the aggregate number improves, you will never find out which two pulled and which three were wasted.
Month 1 — decision and baseline. You still buy nothing here
Week 1: choosing the process. By the three criteria above. Write it down on one sheet of paper step by step, the way it works today — who does what, how long it takes, where it gets stuck.
Weeks 2–3: the baseline. For two weeks you collect three numbers, all three out of the system, not from memory:
| What you measure | Where it comes from | What to watch for |
|---|---|---|
| Count | how many quotes, enquiries, invoices went through the process | it has to be a full week, not a sample |
| Turnaround time | the timestamps between arrival and output | an estimate is not usable |
| Error rate | how many went back for correction, how many became complaints | define in advance what counts as an error |
If your system gives you no timestamp, then the project starts with timestamping, not with AI. This is not a detour: without a baseline you will have nothing to compare against in the third month, and you will make the decision on gut feeling.
Week 4: two things in writing. One is a one-page rule about what may and may not be typed into the tool — the AI policy template gives you a ready-made skeleton for it, and it also stops the team from starting to use the tool out of personal accounts. The other is clarifying your role under EU regulation: anyone who uses an AI system in the course of their own professional activity is a "deployer", and a handful of obligations come with that. Which of them apply today and which do not is what where the EU AI Act stands now sums up, with dates.
Month 2 — training and a pilot on a single team
The target group for the training is not the enthusiastic technician. This is the most counter-intuitive point of the plan, and the best supported one: by the measurement, the gain arises among the beginners and is essentially zero among the experienced. Whoever has already tried it on their own will use it on their own. The 34% sits with the ones who have not.
Four sessions, each with homework and one output. The syllabus, the homework and the question of what EU law actually demands — and what it does not — are unpacked by a separate piece: the team's AI literacy plan.
The pilot runs on a single process, with a control if at all possible. It comes in three forms, in this order:
| Control | How | How strong |
|---|---|---|
| Parallel | two people, two similar sets of customers — one with AI, one without, swap after two weeks | the best; it cancels out the difference between people |
| Alternating weeks | AI in even weeks, none in odd ones | weak; seasonality gets in the way |
| Before–after | you compare only against the baseline | the weakest; a signal, not proof |
One number a week goes on the board, and that number is read out of a system. Not "it has got a lot faster".
Month 3 — the scaling decision, with a condition fixed in advance
The exit condition has to be written down before the pilot starts, not at the end. Two sentences are enough:
It goes live if the output count rises by at least 30%, or the turnaround time falls by at least 30%, without the error rate getting worse. It stops if after three months neither number has moved, or the error rate has risen.
Why "a big, visible difference", and why not statistical significance? Because the latter is out of reach in a small company. A rigorously designed experiment (METR, 2025) needed 246 observations to detect a 19% effect. If your process produces ten items a week, that is close to half a year — and if the real effect is half that size, the required sample grows fourfold, so close to two years. That is why you aim not for significance but for a difference big enough to show up in the raw numbers. If there is only a small deviation, it has to be accepted honestly: it is not measurable.
If it goes on: write down the process, not the tool. The biggest and most consistent difference between successful and unsuccessful rollouts is not the choice of tool but whether the company rearranged the way the work runs around it. The evidence for that, and its limits, is what redesigning the workflow goes through in detail.
What this plan deliberately leaves out
There is no AI strategy document in it. In a ten-person company the strategy is the sentence that says which single process we are trying it on, and when we decide about it.
There is no certification in it. There is an international management standard for AI systems (ISO/IEC 42001:2023), and there is an American risk-management framework (NIST AI RMF 1.0, voluntary, currently under revision). Both give a useful structure — govern, map, measure, manage — but getting certified is not worth it for a small company. The structure can be borrowed for free.
There is no tool list in it. Prices and packages get rearranged within six months; not one point of the plan above depends on which product you pick. The method for comparing them is a separate piece: how to choose an AI model for the task.
And there is no promise in it that it will pay off. Measuring the return is a hard job in itself, full of misunderstandings — the most common ones are taken apart by measuring AI ROI. And the most frequent pitfalls, from every stage of a rollout, are listed by the ten typical mistakes, each with one concrete countermove.
What to take away
Three months, one process, three numbers. In the first month you spend nothing, in the second you train the weakest-performing third, in the third you decide on a condition written down in advance. There is one point in the plan you have to think about — the rest is execution.
The most likely mistake is not that you pick the wrong tool. It is that you have no baseline, and six months from now you still will not be able to say whether anything got better.