What AI really costs, and the budget nobody plans for

In 2025 an hour of labour in the Netherlands cost 48 euro on average. That is the number that appears in no AI quote, and it is the number that decides the difference between an implementation that pays for itself and one that only looked cheap on paper.

We wrote earlier about what an implementation costs: the licence, the build, the maintenance. That piece is about the bill you get. This one is about the bill you do not get, and it is often larger than the first.

Why the licence is never the problem

An AI subscription at a mid-sized company typically costs between two hundred and two thousand euro a month, depending on what it does. That amount is on an invoice, it goes through an approval process, and everyone deciding on it knows what it costs.

The budget nobody plans for sits in the hours around it. Who learns the system, who adapts the process, who checks the output is right in the first months, who explains to colleagues why something now works differently. Those hours do not appear on an invoice. They come out of the diaries of people who were already full, which is why they only feel like a cost afterwards.

Numbers on a screen. Twenty hours of staff time cost more than the licence they support.

The sum most companies skip

Take an average rollout at a company of a hundred and fifty people: five people spending six hours a week for the first two months getting used to it, testing it and adjusting it. That is a hundred and twenty hours each, six hundred hours in total. At 48 euro an hour that is 28,800 euro.

Compare that to a licence of six hundred euro a month, or 1,200 euro over those two months. The ratio is more than twenty to one, and that ratio shows up in almost no business case, because nobody puts the hours of their own people on a separate line.

This is not a reason not to start. It is a reason to actually budget the hours, because a project where half the cost is invisible is a project nobody can later say paid back what it cost.

The four places the time really goes

Adapting the process. An AI tool placed alongside the old process instead of replacing it costs double: the old way keeps running and the new one is added on top. The time sits in the decision to actually let something go, and that decision gets postponed more often than it gets made.

Checking in the first months. Every sensible rollout has a human review an output before it goes out. That is not a superfluous step, that is human on the loop and it simply belongs. The time it takes rarely appears in a budget, but it is on somebody's calendar regardless.

Explaining it to the rest of the company. A system that works well and that nobody understands gets ignored or distrusted. The time it takes to explain three times why something now works differently is just as real as the time it takes to build it.

Adjusting when it goes wrong. Something always goes wrong in the first months, usually something small. The question is not whether that happens, the question is whether time was scheduled to fix it, or whether the project slips because of costs nobody foresaw.

Reviewing documents by hand. The checking step rarely shows up in a budget.

Why this is exactly the capability argument

This is where the sum and our own conviction meet, and we do not want to hide it behind numbers alone: this is why we do what we do.

A company that only buys a licence pays the 1,200 euro and none of the 28,800. It skips the hours, and with them it skips the learning too. After two months the company has a tool that works, and no one who understands why.

A company that does invest the hours ends up paying more in the first quarter and less in every quarter after. The next application no longer costs six hundred hours, because the people who did the first project now know where it goes wrong. That is the difference between buying a tool and building capability, and it sits literally in the hours you do or do not budget.

Why the cheap quote is often the expensive one

There is a counterintuitive rule in projects like this: the lower the licence price, the larger the hidden-hours share usually is. An off-the-shelf package that costs little rarely fits your own process exactly, so more time goes into building around it than with a pricier solution set up specifically for your situation.

That is not an argument for the most expensive option. It is a warning against comparing two quotes on the bottom-line figure alone. A two-hundred-euro-a-month licence that needs six hundred hours of adaptation is more expensive than an eight-hundred-euro one that needs a hundred, and almost nobody does that sum before signing.

So do not just ask what it costs, ask how many hours of your own people it will take to make it fit. A vendor who cannot estimate that probably also has no good picture of how many companies like yours actually keep using the system after the first month.

The item nobody counts: the work left undone

There is a fifth place where time disappears, and it appears in no sum because it has no name: the work not done while five people are busy with the implementation.

Six hours a week per person is not time that appears from nowhere. It is time that used to go to something else, and that other work does not vanish, it slips. A quote cannot slip, but the rest of your company can, and that is exactly why an implementation that stayed on budget on paper still feels afterwards like it cost more than planned.

The fix is not complicated, but is rarely taken: name in advance what temporarily gets less attention. Not as an excuse afterwards, but as part of the decision to start. An implementation where nobody said in advance what would give way ends up giving way to customer work in practice, and you notice that last but hardest.

A budget you fill in yourself

Put in your own numbers, not ours.

Fill this in before you sign a quote, not after. A vendor who will not do this sum with you is letting you skip the comparison he would rather you not make.

What it costs to ignore these costs

Three things happen if you do not budget the hours, and none of them is that the project becomes free.

The first is that the hours exist anyway, they just appear nowhere. People make them in their own time or at the expense of other work, and that other work slips. Nobody sees it happen until the quarter is over and something else was left undone.

The second is that a project with no budgeted hours also has no budgeted checking. Precisely the step that takes the most time is the first to be skipped when things get busy.

The third is the most expensive: if the project shows no measurable result after six months, the conclusion is often that AI does not work for this company. What actually happened is that half the cost was never included in the question of whether it worked.

The item that only shows up in month three: cleaning up data

There is one cost item almost never in the first quote and almost always turning up: the time to make your own data usable. A customer file with three spellings of the same name, a stock system using categories nobody can explain anymore, an archive of quotes with no consistent structure. None of that is a problem until an AI system has to read it, and then it becomes one all at once.

This time usually only appears in month two or three, when the first results disappoint and someone discovers the cause is not the model but the data underneath. Budget another twenty to thirty per cent on top of the hours you already planned if your files are older than a year or have been maintained by multiple people without fixed agreements. Ask yourself this beforehand, not once the project is already running: how consistent is the data this system will use, and who actually knows for sure.

What changes by the third application

To show that the 28,800 euro is not a fixed cost but something that shrinks: on a second application the same five people no longer spend six hours a week, usually three to four, because the way of testing and adjusting is already known. By a third application it halves again.

That is the real reason building capability pays off over a longer horizon than most business cases cover. The first application costs 28,800 euro in hours on our worked example. The third, at the same scale, might cost 7,000. A company looking only at separate projects sees three cost items. A company looking at the total sees a curve that falls, and that is exactly the difference between buying and building.

Check this against your own second project once you have done one. It is the best way to prove, to yourself and to whoever approves the budget, that the first round's investment was not a one-off cost but the purchase of something that stays.

What to do before you start

Finish the table above before you sign anything. Ask the vendor how many hours of your own people they expect in the first two months, and put that next to their licence figure.

Schedule those hours as firmly as the budget. An hour on the calendar gets done. An hour that has to fit in "somewhere" is the first to die when things get busy.

And redo the comparison after the first two months. Not to prove a point, but to know whether your next project can be faster and cheaper, because your people now know where the time goes.

About this page

The average of 48 euro in labour costs per hour comes from the Statistics Netherlands news release on 2024 labour costs, published in 2025. Calculate with your own company's labour costs, not this national average. The worked examples are a method, not a promise. This is the state of play on 12 August 2026.