AI wins the task. You still keep the person.

Put a person and an algorithm side by side on the same problem, and the algorithm usually wins. At Plainsight we have run that comparison at enough companies to say it out loud, and it is still almost never a reason to remove the person. Those two sentences sound like they contradict each other. They do not, and the gap between them is where most AI plans quietly go wrong.

I said a short version of this to a journalist from De Tijd recently. The answer came out a lot shorter than the thinking behind it, which is a normal hazard of interviews. So here is the longer version, the one I actually give clients.

A task is not a job

When someone runs the human-versus-model comparison, they are almost always testing a task. Classify this ticket. Extract these fields. Score this review. Draft this summary. Those are clean, bounded problems with a right answer, which is exactly the shape of problem a model is good at.

A job is not that. A job is a bundle of tasks, plus the judgement about which task matters this week, plus the exceptions nobody wrote down, plus being the person a colleague walks over to when something looks off. When the algorithm wins the comparison, it won one line item out of that bundle. Reading the result as "we need fewer people" is a category error, and an expensive one, because you usually discover what the person was also doing about six weeks after they leave.

Take a real example. For Schoenen Torfs we replaced the mystery shopper, the person who quietly tests service in a shop, with an analysis of the 70,000 reviews customers leave every year after a visit. One subjective visit became a daily dashboard for the shop floor and a quarterly report with concrete actions for management.

Notice what actually changed there. The question was always "how good is our service, really". The old instrument answered it with one opinion from one afternoon. The new instrument answers it every morning with everything customers said. Nobody was made redundant by that. The judgement about what to do with the answer moved closer to the people who can act on it.

What the model does not carry

Three things stay stubbornly human in nearly every project we run, and it is worth being specific about them rather than waving at "the human touch".

The critical read. Somebody has to be able to look at an answer and know it is wrong. Not verify it end to end, that would defeat the point, but have the domain instinct to catch the answer that is confidently off. This is a skill, it is unevenly distributed, and it becomes more valuable as the volume of machine-produced answers goes up, not less.

The exceptions. The model handles the ninety percent that looks like the training data. The remaining ten percent is where your margin, your reputation and your regulator live. Those cases need someone who can reason about a situation they have never seen before.

Accountability. A model cannot sit across from a customer and own an outcome. It cannot be answerable to an auditor. Someone has to hold the decision, and that someone needs enough understanding of how the answer was produced to defend it.

If you strip a team down to the point where nobody can do those three things, you have not automated the work. You have removed your ability to notice when it goes wrong.

The risk is not being replaced. It is standing still.

Here is the part that sounds like it cuts the other way, and I mean it just as seriously: if you are not moving with this now, you do at some point risk becoming redundant.

That is not the same threat as "a model will do your job". It is slower and less dramatic. The job moves, and you do not move with it. The work shifts from doing the task to framing the problem, checking the output, and deciding what to do about it, and those are different muscles from the ones that made someone good at the old version of the role.

The people who make that shift well share something, and it is not technical skill. They start from the goal rather than the task. Ask them what they do and they tell you what they are trying to achieve, not which steps they perform. Give that person a tool that removes half their steps and they get interested. Give the same tool to someone whose identity is the steps, and it lands as a threat, which is a completely reasonable way to feel and also a problem you have to help them through rather than around.

So yes, some tasks will need fewer people. The honest response to that is not reassurance, it is speed. Get people into the new version of the work early, while there is room to learn it, instead of announcing it once the decision is already made.

What this means for how you plan

The practical translation of all of this is fairly boring, which is usually a good sign.

Do not start an AI project by asking which roles it could remove. Start by asking which question your business keeps answering badly or slowly, the way Torfs kept answering "how is our service" with one opinion. Find the instrument that answers it better. Then look honestly at what that frees up, and be deliberate about where those hours go, because if you do not decide, they get absorbed and you see no return at all.

And bring the people who do the work into it early enough that they can shape it. Not as a change-management courtesy. As the only reliable way to find out what the ten percent of exceptions actually look like, because the person doing the job is the only one who knows.

We do not walk into a company to cut it back. The interesting work is almost always the opposite: the same people, pointed at the parts of the job where a human still makes the difference.

If you are somewhere in this conversation inside your own company, our AI work and the Start with AI track are both built around exactly this starting point. Or just talk to us about the question you keep answering badly.

The interview

This post grew out of a portrait of me that Emma Verplancke wrote for De Tijd in August 2026, covering Plainsight's first three years, the move to Ghent, and the step into the CEO role. The questions were good enough that a few of them stayed with me, which is the highest compliment I have for an interview.

You can read the original here: Bo Vande Sompele, CEO van Plainsight, in De Tijd (in Dutch).