Building AI capacity without your own AI team

You do not have a data scientist, an ML engineer, or probably anyone with AI in their job title. That is not the exception we talk about in the Netherlands, it is the norm. Of companies with 50 to 250 people using AI, 29 per cent rely entirely on off-the-shelf commercial software, and 9 per cent have had something built by their own staff. The rest sits in between, with a vendor who built something without anyone internal fully understanding why it works.

This piece is not about setting up a data department. It is about what actually works at a company that will never do that, and that is a larger part of the Netherlands than the AI content on the market suggests.

The wrong question and the right one

The wrong question is: how do we get someone who can build this. That question leads to a vacancy open for months, to an external party doing it for you and leaving with the knowledge, or to nothing, because the answer to "how do we find someone like that" is usually that you do not.

The right question is: who in this company understands the process we want to improve better than anyone. That is almost always someone who already works here. You do not build capability by bringing in someone who knows the technology and not the process, but by teaching someone who knows the process a bit of the technology. That is a shorter road than it sounds, because most of what an ordinary company needs is not mathematics. It is knowing what a good outcome looks like and what a bad one does, and the person already doing the work knows that better than anyone.

Two colleagues at one screen. The person who knows the process usually needs a mandate, not a course.

What "capability" means in practice, without the word

We use the word capability a lot, so let us make it concrete once. A company has AI capability if the answer to these three questions is yes.

Can someone here explain why an AI output is what it is, without calling the vendor. Can someone here adjust the system when the situation changes, without requesting a new quote. Does someone here know what happens if the external party leaves tomorrow.

Three yeses and you have capability. Three noes and you have a subscription. Most companies sit in between, and the point of this piece is to move you one step towards three yeses without setting up a new department to do it.

The role you need already exists, it just has a different name

In almost every company of fifty to three hundred and fifty people there is someone who fits exactly the role you are looking for, and nobody calls it that. That person knows the process from the inside, is curious about how things work, and has often already started using a tool on their own initiative.

Recognisable examples: the planner who built a spreadsheet nobody else understands. The administrator already using ChatGPT for draft emails. The salesperson who figured out how to auto-fill a quote template.

That person does not need a machine learning course. They need time, a mandate, and access to someone who can explain the next step when it gets stuck. That last part is exactly where external help is useful, only not as a builder but as a coach.

Explaining an idea on a whiteboard. External help works best as a coach, not a builder.

The four roles side by side, so you know who to approach for what.

The second row is the role most companies skip. Without an internal coach, every next project repeats the first one's learning curve, which is exactly why capability stays stuck in one person instead of growing.

The order that works: small, real, checked by yourself

Three rules that recur in almost every project of this kind, regardless of sector.

Start with a process that already exists, not a new vision. The first project should solve an existing bottleneck everyone already knows, not a new possibility that still needs explaining. Known bottlenecks have a clear definition of better, which is exactly what you need to know whether it worked.

Pick something whose output you can check yourself. A document check whose answer you can verify teaches your team more than a forecast you can only judge in six months. See also our pieces on why AI projects run aground and the signals that you are stalling: both arrive at the same point, checkability first.

Run it alongside the old process, not instead of it. Two months of running double feels inefficient and is the cheapest insurance there is. The difference between the old and the new output is exactly the proof you need to move on or stop.

Why this is faster than it sounds

The most common objection to this approach is that it is too slow, that an external firm would just build it faster. That is true for the first project and stops being true from the second.

An external firm that builds everything itself delivers a working system in weeks. What it does not deliver is someone on your side who understands why it works. On the next application you start from zero again, with the same lead time and the same cost as the first time. With the approach in this piece, the first project takes longer, usually a few weeks more, because your own people watch and help decide instead of having something placed in front of them.

Those extra weeks are the investment. From the second project the roles flip: your own people already know where it goes wrong, and an external party is mainly there for the parts that stay specialised. After three or four projects the lead-time gap between the two approaches has disappeared, and the gap in what stays in-house has already become irreversible.

What external help should and should not do

External advice is not the problem. What you let happen with it is.

Do: build a first version together, with your own people at the table watching every choice being made. Do: have someone explain why an approach was chosen, not just what the output is. Do: a handover that names, on paper, who is responsible going forward.

Do not: accept a finished system nobody sat in on. Do not: sign a contract where your own data improves the vendor's model without anything coming back to you. Do not: end up with a solution only that one consultant can adjust.

The question to ask on every quote is not what it costs, but what remains when the contract ends. Does a system remain that you can steer yourself, or does a black box remain that you have to keep renting.

A handover meeting. What matters is what remains when the contract ends.

Five questions for the first conversation with an external party

Bring these regardless of who you talk to.

Who from our side sits at the table during the build, and how many hours a week does that take. What happens to our data, and do you use it for other clients too. What can we adjust ourselves once the project is done, without calling you. What exactly do you record at handover, and to whom. And what does it cost if we want to take it over ourselves after six months.

A party with a concrete answer to all five is a party used to clients who want to build capability. A party that dodges the third and fifth questions would rather sell a subscription than an outcome you actually own.

A year in four steps

Quarter one. Pick one process, one person who knows it, and one external party who coaches rather than builds. Run it alongside the old process.

Quarter two. Measure the difference, record what was learned, and decide whether it replaces the old process or stops. This is where most companies either give up too fast or carry on too long without measuring.

Quarter three. Pick a second process, with the same person or someone else following the same pattern. As we calculated earlier, this project costs significantly fewer hours than the first, because the way of working is already known.

Quarter four. Look back at what can now be done internally that could not a year ago. That overview is your first real capability inventory, and it is worth more than a strategy document nobody reads.

What this does not solve

Be honest about the limit of this approach. A system that has to meet heavy regulation, such as AI in recruitment and selection under Annex III of the AI Act, needs specialised knowledge you will not build from your own team in a quarter. For that kind of application, external expertise is not coaching but necessity.

And a company with nobody curious about how things work has a different problem than an AI problem. This approach works with the people you already have. If those people are not there, the first step is not AI but hiring policy.

Where to start

Find this week whoever in your company already started using a tool on their own initiative. Ask what they did and why. Give that person an afternoon a week and one small, checkable process.

Do that before posting a vacancy for a role you will not fill anyway.

About this page

The figure on how AI-using companies with 50 to 250 people source their AI comes from the Statistics Netherlands ICT survey over 2025. The approach in this piece is our own experience from projects at companies this size, not a research finding. This is the state of play on 13 August 2026.