Of the thousands of firms that call themselves agentic, Gartner reckons roughly 130 offer genuine agentic functionality. The rest sell existing chatbots and process automation under a new label. If you are looking for an AI consultant, that is the first thing to know: the word on the website is not a selection criterion, because everybody uses the same word.
We are a firm that sells this ourselves. So let me start with what we do not do: we do not build language models, we have no platform of our own, and we do not promise a percentage saving we have not measured. What follows is the procurement guide I would want if I were sitting on the other side of the table, built from measurements by Statistics Netherlands, MIT, Deloitte and Gartner. None of those figures come from an agency with something to sell.
Why do Dutch small businesses walk away from AI?
Not over the price. That assumption sits underneath almost every article on this subject, and Statistics Netherlands has measured it and refuted it. Micro-enterprises of two to nine people that considered AI and decided against it were asked why. 71.6% named a lack of experience. Cost came out far lower, at 20.3%, which put it seventh of the eight reasons Statistics Netherlands asks about. Five other objections weighed heavier than price.
Read that again. Of every five owners who considered it and walked away, roughly one walked away over money. The rest walked away over something else, and the largest group simply did not know how.
That changes what you are buying. Anyone who thinks price is the problem starts comparing quotes and picks the cheapest. Anyone who sees that knowledge is the problem starts looking for someone who explains rather than sells. Those are two different searches and they produce two different suppliers.
The scale is considerable. Statistics Netherlands counts that 13.8% of micro-enterprises and 29.8% of small and medium-sized firms use at least one AI technology, against 66.2% of large enterprises. The gap between small and large is more than fourfold. And within that group of users it is nearly always text, not work that runs on its own.
That is the opportunity and the trap at once. The opportunity: almost nobody in your category has this sorted. The trap: if more than seven in ten walk away over knowledge, the market fills with providers who address that gap commercially without closing it.
What are Dutch companies actually buying when they say they use AI?
Text, almost always. Statistics Netherlands measures seven separate AI technologies, and the split within them says more than the headline number. Among micro-enterprises, text analysis sits at 9.8% and automatic generation of written language at 6.5%, while process automation stalls at 2.3%.
In other words: when an owner says they use AI, it usually means someone is having text written or summarised. That is useful, but it is not the same as taking work off your hands. With text the human stays the engine and the tool delivers a first draft. I wrote about that gap between a working first version and something you dare to ship in building a webshop with AI: the last twenty percent. With process automation the work carries on while nobody is watching, which is exactly why it weighs so much more heavily when it goes wrong.
The distinction matters in a procurement conversation, because both are sold under the same word. So ask early: does this run when nobody is at their screen? With text tools the answer is no and you need little oversight. With anything that runs on its own the answer is yes, and then the questions about failure and maintenance further down apply.
The growth is real. Among micro-enterprises, usage climbed from 6.8% in 2023 to 10.6% in 2024 and 13.8% in 2025. That is close to a doubling in two years. But it is growth in the text category, not in the category of work that carries on by itself. Whoever sorts the second is doing something 97 in 100 comparable companies have not. You see the same head start on the brand side, where agents now help decide how your brand gets described.
How many genuine AI automation providers are there?
Roughly 130 worldwide, out of the thousands that use the word. That is the estimate from Gartner, made in June 2025, and it comes with a second figure a buyer should know: more than 40% of agentic AI projects will be cancelled before the end of 2027, usually because costs mount or the return stays unclear.
Gartner calls the phenomenon agent washing: old technology under a new name. That sounds harsher than it is. A well-built rule-based automation is often exactly what you need, and it is more predictable and cheaper than a system that makes its own decisions. The problem is not that providers sell it. The problem is that they call it agentic, which stops you seeing what you are buying.
In practice that makes the label worthless as a selection criterion. Everyone puts it on their site. What you need is a question that punctures the label, and that question is simple: which steps in this process are fixed, and which do you let a language model decide?
A provider who answers that concretely knows what they are building. A provider who retreats into the possibilities of the technology is selling you a label. The answer "actually all of them fixed, apart from these two steps" is not a climbdown. It is usually the best answer you can get, because every fixed step is a step that cannot break in a way nobody anticipated.
Is a subscription to an automation tool the same as automation?
No, and that difference costs most beginning projects their money. A tool is where you tie steps together. Automation is your process running every day without anyone needing to watch it. The work between those two is where the bill sits.
You can see it in the Statistics Netherlands figure for process automation. Tooling to connect processes has been available to everyone for years and takes little effort to try. Yet the share of companies actually running it stays low. That is not because the tooling is hard, but because the first version always works and the fifth week rarely does.
What happens then is nearly always the same. A supplier renames a field. A connection expires. An order contains a character combination nobody planned for. The flow carries on but delivers the wrong result, or nothing at all. Without someone noticing, you find out when a customer calls.
That is where the ratio that returns later comes from: rolling out is a fifth of the work, keeping it running is the rest. This is not an argument against starting yourself. Building a first flow yourself is the best way to learn where your process has exceptions.
It is an argument for keeping that first flow small and choosing something you can do without. If it stands still for a week without touching your revenue, you have a learning project. Put your order processing on it straight away and you have a risk.
Are you more likely to succeed outsourcing or building it yourself?
Outsourcing, and the difference is large. The MIT study into the state of AI in business examined more than 300 publicly documented AI initiatives, with interviews at 52 organisations and a survey of 153 executives. The conclusion that matters most to an owner is there in one line: external partnerships see twice the success rate of internally built systems.
The same study produced the figure you hear everywhere: 95% of organisations are getting zero return on their investment in generative AI. That is usually quoted as proof that AI does not work. It was not measured that way. The study looked at measurable effect on the profit and loss account within roughly six months, and that is a short yardstick for a change in how people work.
The funnel underneath is more interesting. For tools built for a single task, 60% of organisations evaluated a solution, 20% reached the pilot stage and 5% reached production. So two thirds disappear between looking and trying, and three quarters between trying and using. For general-purpose language models, such as a chatbot subscription, that same funnel reads 80, 50 and 40 percent. That gap is the whole point of the report: standalone tools make it through, anything that has to land inside your own process does not.
The study offers an explanation you may recognise in your own company: budgets go to visible front-office functions while the return sits in the back office. Customer service and marketing get the money because that is where you can show a result. If you would rather start at the front, first read how to measure whether AI search engines send you customers. Without that measurement you cannot tell afterwards what it delivered. The gains sit in invoicing, order processing, stock, and everywhere someone currently retypes data by hand. That is dull to present and it is exactly where the hours are.
What goes wrong between the pilot and daily practice?
The pilot measures something other than production. Deloitte surveyed more than 3,000 executives directly involved in AI initiatives and found that only 25% had moved at least 40% of their pilots into production. On top of that, 37% apply AI superficially, without changing the process itself.
The sharpest number from that study concerns oversight. Nearly three quarters want to work with agents within two years, but only 21% have a mature way of governing those agents. That gap is where the money disappears. Software gets attached to your customer contact or your order flow before anyone has established who steps in when it goes wrong, how you see that it has gone wrong, and when you switch it off.
In practice such a system rarely breaks spectacularly. It breaks quietly. A connection expires, a session logs out, a supplier changes a field name. The system keeps running and simply delivers nothing, or delivers something that still looks like a result. If nobody is watching, you find out weeks later.
So do not only ask a provider what the system does when it works. Ask what happens when a step fails. Does somebody get an alert, or is the error skipped silently? Is the task retried later? And who looks at that on a Saturday? These are not technical questions. This is the question of whether you have a supplier or merely a delivery. The same distinction applies on the brand side: a brand is a system, not a template, and a system needs maintenance.
What does it cost once nothing new is being built?
More than the build, and that rarely appears in a quote. The clearest public example comes from Pythian, a company of around 500 people processing 15,000 database tickets a month. They automated triage and cut average resolution time by 80%. At one of their clients, an organisation with 10,000 consultants, 10% of 20,000 IT support tickets a year were handled entirely without a human.
That ten percent is the figure to remember, because it sits beside a large majority where a human is still involved. This is a company with a technical team, a large volume of similar work and years of experience. If one in ten tickets there is handled independently, that is the top of what is realistic, not the floor.
From the same account comes the rule of thumb that belongs in every quote: rolling out is 20% of the work, keeping the accuracy up in production is 80%. That is their rule of thumb, not a measurement, and I adopt it because it matches what we see ourselves. A system that is correct today is not correct in six months, because the world around it changes. Prices shift, a platform adjusts its fields, a service you lean on shuts down.
For your procurement conversation that means one concrete question: what does year two cost, once nothing new is being built? A provider who puts a number on that has thought about it. A provider who says maintenance is not much has never run one for a year. I say that as someone who has run one for a year: maintenance is not light, and that is no reason not to do it. It is a reason to budget for it.
Which questions remove the most noise from a quote?
Five. They are not technical and you need to know nothing to ask them.
Which steps are fixed and which do you let a language model decide? This punctures the label. Any answer that is concrete is a good answer. Any answer about possibilities rather than about this process is a sales pitch.
What happens when a step fails, and who sees it? This is where a delivery separates from a service. Without an alert and without a retry, you have a system that can stop quietly without anyone noticing.
What does year two cost? See above. If no number comes back, there is no plan.
Who marks the output, you or us? Article 50 of the AI Act has applied since 2 August 2026. The European Commission states that it applies from that date: anyone supplying a system that speaks directly to people must make sure the user knows it is not a human, and must mark generated output in machine-readable form. Part of that duty lands not with your supplier but with you. You have to inform people about a deepfake, and about AI-written text on matters of public interest that nobody has editorially reviewed. So ask who is the provider here and who is the deployer, and who arranges which marking. Anyone who cannot answer that has not read the regulation.
What will not work? This is the most important one. A provider who has not named a single limitation after half an hour either has not understood the process or is not telling you everything. Everything worthwhile has edges, and someone who knows the trade can name them.
Note that none of these questions is about which model is used or how advanced it is. That is deliberate. The choice of underlying model changes every six months and is rarely what a project founders on. Projects founder on unclear ownership, on missing oversight, and on a second year nobody budgeted.
Where do you start if you have never done any of this?
With the dullest process you have. Not with customer contact, because that is the most visible and at the same time the most sensitive. Start with something where someone currently retypes data from one screen into another: entering orders, matching invoices, updating stock levels, moving product data from your own system to a marketplace.
Then pick something that happens often and where you know how often. The second part matters more than the first, because without a number you cannot tell afterwards whether it delivered anything. If you do not know how many orders are retyped by hand each week, count for a week first. That count is worth more than the first three quotes you request.
After that, do not ask for a proposal but for a conversation about that one process. A provider who understands it will ask about exceptions: what happens with a return, with a partial delivery, with a customer who appears twice in the system. Exceptions are where automation breaks, and it is a good sign when someone asks about them immediately. We do that on every project we take on, whether it concerns a brand or a website.
And keep the first assignment small enough to stop. Not because you expect it to go wrong, but because the figure this article opened with works in both directions. More than seven in ten owners walk away over a lack of experience. The only way to gain that experience is to finish one small thing completely, including the second year. After that the next decision is no longer a leap in the dark but an estimate you can make yourself.
