Most AI implementation projects that stall do not fail on the technology. They fail because nobody inside owns the thing once it is live. A sponsor signs the budget and asks how it is going. An owner has hours in their week, the standing to change how the work is done, and their name on the result. In a company of thirty people that owner is never a job title. It has to be taken out of a real one.
Why does an AI project stall after a successful pilot?
The pattern is consistent enough that I could nearly set my watch by it. The pilot works. Everyone in the room is pleased. Two months later I ask how it is going and the answer is a version of "good, I think", delivered by someone who has not actually looked.
Nothing broke. That is what makes it hard to see. The model still returns what it returned in week one. What went missing is the person who notices when the answers start to drift, who receives the "it did something strange yesterday" message from the warehouse and decides whether that is a bug, a gap in the instructions, or a process that quietly changed in April and nobody told the system about.
One prediction gets quoted at me constantly: Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, on escalating costs, unclear business value or inadequate risk controls. Read that list of three again. Not one of them is a modelling problem. Costs escalate when nobody is watching them. Business value stays unclear when nobody has been asked to state it. Risk controls stay inadequate when nobody has been told the controls are theirs. Those are three ways of describing an empty chair.
I have written before about why AI pilots fail to make the jump into daily work, and the mechanics are usually organisational rather than technical. This essay is about the smallest and most fixable version of that: the missing name.
Who owns an AI implementation: the sponsor or the owner?
A sponsor approves. They defend the budget in the management meeting, they want a number they can repeat, and they are genuinely necessary. In most Dutch mkb companies the sponsor is the directeur, and that is the right person for that job.
An owner lives with it. They decide where the exception threshold sits. They answer "why did it do that" for the colleague who asked. They are the one who says "the process changed, so the system has to change with it" before the system has been quietly wrong for six weeks.
There is a single question that tells you which of the two you have. Ask: who decides, next month, whether we move the confidence threshold from eighty percent to ninety? If the answer is a name, you have an owner. If the answer is "well, that is something we would discuss", you have a sponsor and a committee, and the threshold will still be sitting at eighty a year from now.
I ask that question in first meetings now. It is not a trick. It is the fastest way I know to find out whether the project I am being asked to quote has anywhere to land.
Why the enterprise answer does not fit a company of thirty
There is a serious body of thinking about this now, and it is all written for somebody else. In February 2026 Writer's chief people officer published a piece arguing that your next headcount is an AI Agent Owner, sitting alongside an AI Agent Builder and an AI Champion, under a two-layer governance model. It is a good framework, honestly reasoned. It also contains a section answering the objection "we're a 100,000-person company", and a ninety-day roadmap the author describes as workable for Fortune 500 complexity.
Three new roles. A governance layer above them. A pilot business unit to start in. If you employ thirty people you do not have a business unit. You have a company. You are not going to create a role for this, and I would argue you should not want to.
So the translation has to go the other way, and it is harder and less satisfying than hiring. The enterprise question is who do we bring in. The mkb question is what do we stop doing, so that this has somewhere to live. Nobody writes the second one down, because the answer is never flattering: something on somebody's list is about to get worse, on purpose, and the sponsor has to be willing to say so out loud.
The poldermodel problem: agreement is not ownership
I want to be careful here, because what follows is an observation about a business culture I work in and admire, not a complaint about it.
Dutch companies are unusually good at getting everyone into the room. The habit has a name and a history. What people mean by the poldermodel is usually traced back to the Wassenaar Agreement of 1982, when unions, employers and government settled on a package of shorter working hours and wage restraint in exchange for more employment. The word is younger than the practice: the politician Ina Brouwer appears to have been the first to write "poldermodel" down, in an article in 1990. And almost from the beginning, critics used the verb polderen for the slow version, where every party has to be heard before anything can move.
I see the good half of this constantly. A Dutch team will tell an external consultant to his face that his idea is wrong, then build a version everyone can live with, inside one meeting, without anybody sulking about it afterwards. That is a real competitive advantage and plenty of countries do not have it.
The cost shows up at exactly one moment: when a project needs a name rather than a consensus. The meeting where everyone agrees the AI should handle inbound quote requests is easy and pleasant. The meeting where one person says "then it is mine, and here is what I am dropping to make room for it" almost never happens on its own. It has to be asked for, out loud, by somebody in the room. Usually that has to be me, because I am the only one being paid to make the moment awkward.
How do you choose an internal AI owner?
Four things I look for, roughly in this order.
The person who complains about the process most precisely. Not the loudest complainer. The most precise one. Someone who can tell you that the quote goes out wrong specifically when the customer is a reseller and the delivery address differs from the invoice address is carrying a model of the process that exists in no document anywhere in the building. Precision of complaint is the best proxy for real process knowledge I have found.
Someone who consumes the output, not someone who supplies the input. The person downstream feels every error personally, because they are the one who has to fix it before a customer sees it. The person upstream can be told their data is fine and will cheerfully believe it.
Enough standing to change the work. They need to be able to say "from Monday we do this differently" and have it stick, without convening anything. In a flat organisation that is usually about respect rather than title, which is convenient, because at thirty people the titles are mostly decorative anyway.
Somebody who can win an argument with me. This matters more than it sounds. An owner who defers to the consultant on every judgement call is a sponsor with extra steps, and the day I walk out the system has nobody. I would much rather work with the person who tells me my exception rule is nonsense because on Fridays the whole process runs differently. In my experience they are right about that roughly every time.
How much time does an AI owner actually need?
I get asked for a number and I am wary of giving one, because the honest answer depends on how much of the process the system touches. What I can describe is the shape of it.
The first four to six weeks are the heavy part, because that is when the exceptions arrive. Every exception is a decision that has never formally been made before, and the owner is making them at whatever rate the system uncovers them. After that it settles into something much smaller and much more regular: a slot to look at what the system did, what it refused to do, and what people quietly worked around.
The important word in that paragraph is slot. Put it in the calendar as a recurring appointment with a start time and an end time. Time that is not allocated goes to whatever shouts loudest, and an AI system that is slowly drifting does not shout at all. That is precisely its problem, and it is why "we will keep an eye on it" is not a plan.
There is one test that cuts through the whole conversation. If you cannot name the thing that comes off this person's plate, the hours do not exist. You have described an intention, and intentions do not survive a busy November.
What the AI Act quietly assumes about ownership
Something shifted three weeks ago that makes this less abstract. Article 4 of the EU AI Act, the AI literacy obligation, has applied since 2 February 2025, but supervision and enforcement sit with the national market surveillance authorities, and they started supervising and enforcing as of 2 August 2026. The Digital Omnibus that came into force in mid-July 2026 softened the wording rather than removing it: providers and deployers have to support the development of AI literacy among the staff working with these systems, and no specific level is mandated.
That is not a reason to panic and I am not going to sell anybody a compliance package off the back of it; if you want the practical version, we keep an AI Act checklist for mkb up to date separately. The reason it belongs in an essay about ownership is the grammar of the thing. The regulation addresses "the deployer" as though a company were a single actor with a single memory. That is a legal fiction, and a useful one, but somebody inside the building has to make it true. In a firm of thirty, an obligation with no name attached is an obligation that nobody performs. Not out of bad faith. Because every single person reasonably assumed it belonged to somebody else. The capability question underneath it is usually smaller than people fear: what is missing is rarely a skills gap. It is a name.
What if nobody has the hours?
Then the project is not ready, and I would rather say that in the first meeting than in month four.
That answer costs me work and I have made my peace with it. A company that cannot free up four hours a week for the thing it has just described as strategically important is telling me something accurate about its priorities, and the right response is not to sell it a smaller version of the same problem.
There is a middle path that works more often than the flat no: shrink the scope until the ownership fits it. If the honest capacity is two hours a week, then build something that needs two hours a week. In practice that means one process instead of three, an assistive system rather than an autonomous one, and a deliberately narrow first project. Adoption is where these things break, not the code, and the second project is where you find out whether the company actually changed. That second project only ever happens if somebody was looking after the first one.
What I own, and what I refuse to own
For clarity, because this is easier to preach than to practise.
I own the build, the documentation, the handover, and the answer to "how does this actually work". I will sit in the owner's slot with them for the first few weeks, and I will be the person they call at eleven at night in the first month, because that is what a first month is.
I do not own the judgement about whether an exception is acceptable. That is a decision about your customers, your margin and your appetite for risk, and I do not have the standing to make it. Every time an external party ends up making those calls by default, the company has traded a technology dependency for a person dependency, which is the worse of the two, because it never appears on an invoice until the day you need to replace me.
If you are working out what to ask before you sign anything with an AI partner for your mkb, put this near the top of the list: what do you expect us to own, and from when? A partner who has not thought about that answer has not thought about what happens after go-live, which is the only part of an AI implementation that has to last.
None of this is about the model, and it has not been about the model for some time. The technology arrived. The org chart did not move. The gap between those two facts is where most AI implementation money in the Dutch mkb is currently being lost, and closing it is unglamorous and nearly free. One name, some hours, and a recurring appointment that nobody is allowed to cancel.