What does an AI implementation cost in a mid-sized company?

Search for what an AI implementation costs and you land on pages that never answer. It depends on your situation, get in touch. That is not always evasion. More often the answer genuinely varies, and naming the wrong figure feels worse than naming none.

That is no use to you. So below there are figures: ours, where we publish them, plus the four cost items every quote is built from. With those four you can put two proposals that look nothing alike side by side after all.

The short answer

For a company of fifty to three hundred and fifty people, a first AI trajectory, from diagnosis to a working application, breaks down into three decisions with steeply rising amounts. The first costs nothing, the second a few thousand euro, and the third is the only one that is a real investment.

That escalation is not a sales device, it is the only sensible order. You do not want to commit a hundred thousand euro to a use case nobody has yet worked out the return on.

What we charge

We publish our prices on the site, so here they are without a detour.

The scan is a fixed fee, not time and materials. That is deliberate: when the supplier earns more the longer the research runs, you can no longer tell why an analysis took three weeks.

We do not price the sprint up front, and the reason is one you can check. What a first application costs depends on which opportunity comes out of the scan, and nobody knows that before the scan is done. A supplier quoting you a sprint price before knowing what you are going to build is selling you weeks, not an outcome.

A plan being presented to a management team. The cost of the decision is usually larger than the cost of the project.

The four cost items behind every quote

Quotes look different because suppliers add up different things. These are the four items. If one is missing from a proposal it has not disappeared, it is just somewhere else.

That last one is the most important and the only one nobody invoices you for. Expect tens of hours from the people who know the work best, spread across the trajectory. By definition those are your busiest people. Skip that planning and you find the project stalled halfway through because operations came first, which is the most expensive delay there is.

Licences are the only item you can look up exactly, because they are on your vendor's price list. Multiply per user per month across twelve months before calling it small.

A budget you can actually fill in

Because one of the four items is only known after the diagnosis, no off-the-shelf total is available. What is available is a budget in which three of the four items are already fixed and the fourth is a clearly marked gap.

Put those three known items in your budget and reserve a range for the fourth. That is defensible to a board or an accountant, and it is infinitely better than a round number resting on nothing.

Where the money goes inside the build

Seeing a build quote for the first time, people expect the largest share to go to the model. It almost never does. The model is usually the cheapest component, because you rent it per use from a provider.

The money goes to the work around it. Data sitting in four places and defined in four ways. A process that runs differently in practice than in the handbook. Exceptions everyone knows and nobody has written down. That work is not glamorous and it is the difference between a demo and something people actually use on Monday.

That is also where the largest cost variance between two companies with the same question comes from. If your data is in order, a first application is a matter of weeks. If it is not, that is where it starts, and the preparation cannot be skipped. A supplier promising it is unnecessary is moving the problem to the day after delivery.

Why almost nobody names a figure

There is a genuine reason underneath. The spread really is wide: the same question costs one company a fraction of what it costs another, because the data is already there or is nowhere at all. Quoting an average that is true for nobody helps nobody.

There is also a less genuine reason. As long as no figure is on the table you cannot compare, and as long as you cannot compare you are competing on trust rather than substance. Price transparency is uncomfortable for a supplier and the buyer's only defence.

The workable middle: a supplier should name the price of the diagnosis up front, and a total for the build after the diagnosis. Refuse both and they are selling hours.

The sum that matters is payback time

An amount with no return next to it is meaningless. The worked example on our site concerns a process with 1,200 hours of manual work per year. Automate a substantial share of it and the saving comes out around EUR 68,000 per year, with a payback time under three months.

That is explicitly an example, not a promise. The value is in the shape, not the number: hours times rate times the share you actually remove. That same sum immediately shows where it does not add up. A process of two hundred hours a year almost never pays back, however good the demo was.

The goal, moreover, is rarely to replace people. It is to get more out of the people you already have by taking the repetitive work off their plate. That changes the arithmetic: you are not counting redundancies, you are counting capacity freed up for work that does generate revenue.

Two colleagues working something out at a whiteboard. Hours from your own people are the cost item that never appears in a quote.

Building it yourself is not the cheap option

The reflex on receiving a quote is: we will do this ourselves. Then cost that route out properly.

Our own salary benchmark for the Dutch AI market puts median total compensation for an AI Engineer at roughly EUR 85,000 to 95,000, and considerably higher in Amsterdam. Employer costs come on top, plus the time to find someone in a market with over 12,400 open positions against some 3,800 specialised graduates a year.

That is not an argument against building it yourself. It is an argument for making the comparison honest. One senior on the payroll is a structural cost of well over a hundred thousand a year. That can be exactly the right call when you have several years of work for them. For one first application it rarely is.

The third route, a contracting firm that builds it and leaves, is usually the fastest and over time the most expensive. Not because of the rate, but because the knowledge walks out with the builder and the next question starts you at zero again.

What it costs to get it wrong

The conversation about cost is nearly always about the investment and nearly never about the return that fails to arrive. Which is where the real money is.

A widely discussed 2025 study by MIT's Project NANDA concluded that roughly 95 percent of the enterprise generative AI pilots reviewed had no measurable effect on profit and loss. That figure is not peer-reviewed and rests partly on self-reporting, so do not build policy on it. The direction is recognisable, though, and one finding from the same work is strikingly concrete: about half of budgets went to sales and marketing, while the clearest returns showed up in dull back-office automation. Money went where the energy was, not where the payoff was.

Translated to your budget: the most expensive mistake is not a quote that is too high, it is a cheap project in the wrong place. A fifteen thousand euro pilot that delivers nothing does not cost you fifteen thousand euro. It costs you that, plus the internal hours, plus a year in which nobody wants to raise AI again because the last attempt did nothing.

The context you are deciding in

Statistics Netherlands measures the gap sharply. In 2025, 29.8 percent of small and medium-sized companies, between ten and 249 employees, used AI technology. Among companies of 250 people and up it was 66.2 percent.

A gap of more than double is not about interest. It is about the capacity to work it out. Large companies have someone whose job it is. In a company of a hundred and fifty people it lands on top of someone who is already busy.

Which is exactly why a paid diagnosis of a few thousand euro is often the cheapest line in your whole budget. Not because of the report, but because it is the only way to avoid basing a six-figure investment on a hunch.

Six questions that make any quote comparable

Ask them of every party, including us.

1. What does the diagnosis cost, separately from the build, and do I keep the output if I stop there?
2. After the diagnosis, do I get a total for the build, or hourly rates?
3. How many hours from my own people, and from which roles?
4. Which licences come on top, from which vendor, per user per month?
5. Who can maintain this once you are gone, and what does it cost to get it to that point?
6. What will we measure in six months to know whether this worked?

The sixth is the sharpest. A party unwilling to put a number on it does not know either whether it will work.

Where to start

If what you need right now is a figure to put in a budget: put in the diagnosis, not the whole trajectory. The AI Readiness Scan starts at EUR 3,900, takes ten working days and delivers the opportunities ranked with the numbers run on your own figures. After that you know what the build costs, because you know what you are building.

If you would rather see where you stand with no commitment first, the AI Photo is free and takes ten minutes.