Five signals that your AI ambition is stalling on your organisation

In 2025, 46 per cent of Dutch companies with 50 to 250 employees used AI. Among companies with 250 or more people it was 67 per cent. That twenty-one point gap is the reason for this piece, because it is not stable. It widens a little every quarter, and it does not widen because large companies are smarter.

It widens because a large organisation can afford to do three things at once that a company of a hundred and fifty cannot: free people up, make mistakes, and carry on when it does not work the first time. If you cannot do that, you get stuck in a phase that looks like standing still from the outside and feels like being busy from the inside.

This piece is not about why an individual AI project fails; we wrote about that separately. It is about the signals in your organisation that show up before you start a project, or while three are running without anything changing.

What it is not: money

This is the most stubborn misconception and Statistics Netherlands contradicts it directly. Of companies with 50 to 250 people that considered AI but did not adopt it, 3 per cent name cost as the reason. Lack of relevant experience inside the company is named by 11 per cent, more than three times as often.

The top of that list is not a budget problem and not a technology problem. It is a people problem, and that is exactly the kind of problem a vendor quote does not answer.

Working late alone. In most companies this size the AI knowledge sits in one person.

Signal one: all the AI in your company was bought

Look at how AI arrives with you. Of companies with 50 to 250 people that use AI, 29 per cent do so through commercial software used exactly as it comes. Fifteen per cent have something built by an external supplier. Nine per cent have had something developed by their own employees.

There is nothing wrong with buying. The signal is in the ratio. If everything you do with AI comes out of a subscription, your AI capability is exactly as large as your vendors' offering and not a millimetre larger. You cannot adapt anything to your own way of working, you cannot exploit your own data that nobody else has, and you cannot switch without starting over.

How to spot it: ask your management team who can explain why one of your AI features does what it does. If the answer is that the vendor knows, you have the signal.

The smallest intervention: pick one process where your own data makes the difference and let someone internal work with it, even if the first result is worse than the bought alternative.

Signal two: the knowledge sits in one person

Almost every company this size has one. Someone who started on their own initiative, who knows the tools, who writes the prompts everyone else asks for. Often not someone from IT.

That person is a blessing and your biggest risk at the same time. As long as everything runs through them, the company's capability does not grow. It only grows in that one person. If they leave, it is not only the knowledge that goes but the enthusiasm too, and it collapses within a month.

How to spot it: note for a week who gets named in AI conversations. Same name every time, and this is your situation.

The smallest intervention: stop letting that person answer the questions, and have them spend half an hour every two weeks explaining to three others. Not as a course, but with a real problem from that week.

Signal three: every department started on its own

Marketing uses something for copy, administration uses something for invoices, and someone in sales has a subscription on their own credit card. At companies with 50 to 250 people, AI is most often used in administrative processes, by 21 per cent, followed by marketing and sales at 17 per cent.

That in itself is not a problem, it is even a good sign: people are looking for solutions. The problem is that nobody has the overview. You do not know which company data sits in which tool, you pay three times for the same thing, and when something goes wrong you cannot reconstruct what happened.

How to spot it: ask finance for every subscription under a hundred euro a month. It is all there.

The smallest intervention: make the list and put it on the wall. Banning does not work and is not necessary. Visibility alone changes the conversation.

Notes on a wall. Visibility changes the conversation faster than a ban does.

Signal four: the conversation is about tools, not about decisions

Listen to how AI gets discussed where you work. If it is about which package, which vendor and which feature, you are having the conversation the market wrote for you.

The conversation that matters is about decisions. Which decision do we currently make on instinct that we could make on numbers. Where do customers wait longest. Which action do we perform identically a hundred times a week. Those are questions your own people can answer and a vendor cannot.

As long as the conversation is about tools, you can only choose from what is for sale. The moment it is about decisions, you know what you are looking for and the choice gets small.

How to spot it: read the last three sets of management minutes. Do they contain product names or processes.

The smallest intervention: ban brand names for the length of one meeting.

Signal five: nobody can explain why an output came out

This is the signal that surfaces last and costs most. Something is running, it produces a list or a score or a forecast, and the question of why is no longer asked. Not because the answer exists, but because everyone knows no answer is coming.

That is a business risk before it is a legal risk. The moment a customer, an employee or a regulator asks the question, you have to be able to answer it. For high-risk applications such as recruitment and selection it also becomes an obligation from 2 December 2027.

How to spot it: take a random output from last week and ask for the reasoning. Measure how long it takes.

The smallest intervention: agree that every AI output going outside, to a customer or a candidate, is seen by a human who can explain what happened. That is called human on the loop.

Looking at an output together. Someone has to be able to explain why it came out that way.

The signal that is not a signal

There is one thing almost every director mentions that we deliberately left off the list: we do not have a data scientist.

In companies this size that is rarely the real problem, and it is often an expensive solution to something else. Hiring someone who can build models but does not know your floor produces exactly the situation described in signal one, only internally. Something appears that works and that nobody else can explain.

The order that does work in practice is the reverse. Start with someone who knows the process and can be taught the technology, not with someone who knows the technology and has to be taught the process. The first takes months, the second takes years, and that difference shows up in when the company becomes self-sufficient.

The same goes for the idea that you have to get your data in order first. Data never gets in order without a concrete reason to look at it. The reason comes from an application, and the application comes from a decision you want to make better.

How to measure this yourself, without a consultant

You do not need external research for this. Five questions, an hour with your management team, and be strict about the answers.

Can anyone inside this company explain why one of our AI features does what it does, without calling the vendor. Has the same name come up in every AI conversation this past month. Do we know which AI subscriptions are running and which company data sits in them. Are our minutes about processes or about product names. Can we justify an output from last week to someone who was not there.

Count the noes. At zero or one you are fine and technology genuinely is your next step. At two or three you are stuck on your organisation and a new tool will not fix it. At four or five the honest conclusion is that you are not building yet but buying, and you are better off acknowledging that than accelerating it.

Ask those five questions again every quarter and write down the result. Not because the score is interesting, but because the direction is. A company with three noes in March and still three in September spent six months on tools and nothing on capability.

Why the gap widens every quarter

All five signals share one property: they get worse on their own. Knowledge sitting in one person concentrates further every month. A collection of loose subscriptions gets murkier every quarter. A conversation about tools produces new tools, and therefore a longer conversation about tools.

Meanwhile the other side accelerates. A company that started building its own capability last year does things this year that were impossible last year, because the people doing it now know where it goes wrong. That is the difference between 46 and 67 per cent, and it is not a difference you close by buying harder.

The good side of this story: none of the five smallest interventions above costs meaningful money. They cost attention, and they cost the discomfort of admitting the current pace is not working.

What this actually costs if you do nothing

The uncomfortable thing about these five signals is that no invoice comes with them. Nobody bills you for capability you did not build, which is why it stays where it is.

There is a price though, and it breaks into three parts. The first is paying twice: three departments with three subscriptions doing largely the same thing quickly adds up to a few thousand euro a year in a company of a hundred and fifty, and that is the smallest item. The second is the work left undone because nobody frees up time for it while a competitor does. The third and largest is that two years from now you have the same dependency, but on a vendor who by then knows you cannot leave.

That last one is the argument for tech sovereignty in miniature. It is not about building everything yourself. It is about being able to switch without starting over, and you only build that while you do not yet need it.

What to do this month

Walk through the five signals and be strict. Two or more means your problem is not in the technology.

Then pick not the most important one but the easiest. Usually that is signal three, the list of subscriptions, because it takes an afternoon and immediately produces a conversation that would not otherwise have happened. Then do signal two: get the knowledge out of that one person.

What you do not do is launch a large programme to fix all five at once. That is precisely the kind of project that runs aground, and we have written about that already.

About this page

The figures on AI use, on reasons for not adopting it and on how AI enters companies come from the Statistics Netherlands ICT survey covering 2025, and concern companies with 50 to 250 employed persons. They are listed below with their table number so you can check them yourself. The five signals are our own ordering based on what we run into in companies like this, and are not a research result. This is the state of play on 10 August 2026.