AI implementation: usage doubled, hiring a builder did not

How many micro-businesses hire an AI builder?

In 2025, seventeen in every thousand Dutch micro-businesses had an outside party make their AI. That party built or adapted the software, and this article simply calls it a builder. That is our own calculation on figures from Statistics Netherlands (CBS), and the number has not measurably grown since 2023. AI use itself has: among firms with two to nine persons employed, it doubled from 6.8% in 2023 to 13.8% in 2025. So the number of micro-businesses having AI built or adapted did not measurably keep pace with usage. Whether that also means AI implementation is lagging is a separate question, and we come back to it further down.

In the same report, the share of users that hired an outside party fell from 23.5% to 12.4%. As a share, hiring almost halved. As a number of businesses, it did not measurably grow, from about sixteen to seventeen per thousand, because the number of users doubled. To see this, multiply two CBS tables: the share of firms using AI, times the share of users per route. That calculation is ours, rounded to whole businesses per thousand, and every per-thousand figure in this article comes from it.

Our calculation on CBS tables 2.1 and 5.1, rounded. A firm could name more than one route, so the rows do not add up to the total.

Off-the-shelf use doubled from 33 to 67 per thousand. Adapting software in-house grew less steeply (commercial software from 13 to 22), while hiring a builder and in-house development barely moved. Most of the growth in AI at micro-businesses is software that firms buy and use as it comes.

Read it with the margins in mind. For the outside builder, the CBS interval in 2023 ran from 18.6 to 29.1 percent of users, and in 2025 from 10.3 to 14.9 percent. Take the lowest and highest points of both CBS intervals, for usage and for hiring, and you get eleven to 22 per thousand for 2023 and thirteen to 22 for 2025. That is a rough range, not a CBS interval. You only get a doubling if you compare the lowest 2023 point with the highest 2025 point. The point estimates, sixteen and seventeen, show no growth, while usage did double.

Two caveats apply. First, 2023 was the first year in which CBS put these questions to micro-businesses. Second, in 2023 users named almost one and a half routes on average, against about one in 2025, while CBS itself reports some variation in questions and answer categories across the years. Part of the drop in the share hiring a builder may therefore be down to how it was measured.

This is about micro-businesses, the group CBS surveys separately here. Among firms with 10 to 249 persons employed, 29.8% used AI, more than twice as often. These figures apply to the smallest firms, not to small and medium-sized businesses as a whole.

With clients, what we mostly look at is where AI sits: in a process that runs on its own, or in a browser tab that someone has to open. CBS does not measure that difference. The figures below come closest, and at the end there are five questions to check it for your own business.

How did micro-businesses get their AI?

Almost half of all users, 48.2% in 2025, bought commercial software and used it as delivered. In 2023 that share was also around half. CBS asks firms that use AI how they got it, with five possible answers, and the other four routes have all shrunk as a share.

Share of AI-using firms with two to nine persons employed. Source: CBS, chapter 5. Figures for 2025 are provisional, those for 2023 and 2024 revised provisional.

Multiple answers were possible, so the columns add up to more than a hundred. In 2023 the total comes to about 140, in 2025 to about 105. A user named almost one and a half routes on average in 2023, and about one in 2025. On paper, two years ago users more often combined a bought package with something they built or adapted alongside it.

Part of that gap may be a measurement effect. CBS itself writes that the questions and answer categories vary somewhat over the years. One possible reading of the rest: newcomers more often take a single tool and build nothing around it. The table does not prove that, but it fits what is there.

Either way, knowing that AI use doubled does not tell you whether even one workflow in a business now runs differently. Using AI and building it in are two different figures, and reports on AI in small business often treat them as one.

What do micro-businesses use AI for?

Marketing or sales comes first: 32.7% of users apply AI there. Administrative processes or management tasks follow at 25.9%, and logistics comes last at 2.3%. Among AI-using firms in the EU with ten or more people, Eurostat sees the same top two: 34.70% use it for marketing or sales and 31.05% for business administration. That is a different group of firms with different categories, so compare the order only, not the percentages.

As a share of users, IT security fell from 18.5% to 6.5% and accounting from 20.9% to 12.9%. Per thousand micro-businesses, accounting went from about fourteen to eighteen, a change that falls within the margins of error. IT security went from about thirteen to nine, though the margins overlap, and logistics stayed at three. Marketing went from about 22 to 45 per thousand, research and innovation from about thirteen to 32.

The breakdown by technology shows where the growth is. Text analysis rose across all micro-businesses from 3.1% to 9.8%. Generating text went from 2.8% to 6.5%, but according to CBS it was roughly flat in 2025 compared with 2024. Robotic process automation, the category closest to a built-in workflow, got no further than 2.3% in 2025, against 1.7% two years earlier.

In our article on choosing an AI partner we already argued that for small firms it is mostly about text. The figures per technology confirm that, with one nuance: last year's growth came from analysing text, not from writing it. AI that creates images, video or sound is deliberately left out of this series for 2025. A firm that only uses an image tool does not appear in these figures.

What is the difference between using AI and AI implementation?

Using AI only starts when someone thinks of it. With AI implementation, the step starts with an order, a form or a new product, and the result lands in your system without anyone copying it. We call the first AI in a tab and the second AI in a process.

A tab: someone opens a chat window, pastes in a customer email, copies the answer back and tweaks it a little. Faster than doing it by hand, but it starts with a person and stops as soon as that person has something else to do. A process: a request comes in, an AI step turns it into a draft, a fixed rule checks the fields, and the result is in your CRM.

Off the shelf is not the same as a tab. Accounting software that reads receipts and posts them automatically is bought software and still a process. CBS measures how firms get their AI, not where that AI sits in the work. This distinction is ours, and the figures in this article approximate it.

In a process you also decide once which data goes where, instead of email by email in a chat window. Privacy does not get solved that way, but you do handle it in one place. That matters for anyone on the fence: of the micro-businesses that considered AI but did not start, 48.8% named privacy as a reason and 42.7% uncertainty about the legal consequences.

We use both. Much of what we do with AI is simply a tab. Work that comes back every week goes into a process, and this article is itself an example. A research agent gathers sources, the next step writes a draft, and a separate critic tries to tear that draft apart. After that, a check opens the linked source for every figure and looks for the number there. Only when that holds does Louie read the text. That check catches a number that is not in the source, but not a faulty calculation. That is why the table at the top explains how we calculated, so you can verify it yourself.

Who gets AI built, and why do others never start?

Among the somewhat larger micro-businesses, AI users hire a builder more often. Of users with two people, 6.2% did so, and among firms of five to nine people 19.2%, three times as often.

Those firms also use AI more often: 19.0% at five to nine people, 11.6% at two-person firms. Combined, that comes to about 36 per thousand firms of five to nine people having AI built, against about seven per thousand two-person firms.

One obvious explanation, which CBS does not measure: a process needs someone who knows how the work runs, where the data lives, which integration is needed and what happens when a step fails. A tab only needs an account. In a two-person firm the owner has to free up that time. With nine people there is more recurring work to build in, and a better chance that someone gets hours for it.

CBS does not ask why firms that already use AI do not have it built in. It does ask why firms that considered AI never started. In 2025, 71.6% named lack of experience, against 56.5% in 2023. Cost was named by 20.3%. In 2023 that was 30.9%, but the margins of those two years overlap, so you cannot read a falling trend into it. What does follow: money ranks far below know-how.

The study Dialogic carried out for the Dutch Ministry of Economic Affairs, based on desk research and interviews with business owners and experts, concludes (in our translation) that AI is rarely embedded in day-to-day operations in any structural way, and that despite growing use, AI in most small businesses remains limited to isolated applications. That is what we call AI in a tab. The report is qualitative and covers small and medium-sized businesses broadly, but it lines up with what the CBS routes show. If you are looking for outside help to fill that knowledge gap, how to separate real AI providers from the rest gives you a way to choose.

In which sectors do micro-businesses use AI least?

Transport and storage, hospitality and construction all sit below 5%: 3.6%, 4.2% and 4.6%. Information and communication tops the list at 50.4%. A micro-business in that sector therefore uses AI many times more often than a hospitality business of the same size. On the point estimates it is twelve times, but the CBS interval for hospitality runs from 2.1 to 8 percent, so the gap lies somewhere between six and more than twenty times.

Source: CBS, chapters 2 and 5. The last column is our calculation. CBS gives no breakdown by route for construction and hospitality.

Trade, which includes online shops, shows why share and number need to be kept apart. Of the trading firms that use AI, 19.2% hire a builder, on paper the highest share of the six sectors CBS breaks down here, although its margins overlap with those of manufacturing and rental. But because few trading firms use AI, trade comes out at about seventeen per thousand firms, around the average. In information and communication it is over fifty.

We work with brands in e-commerce, engineering, hospitality, healthcare and sport. Hospitality and construction sit far below average AI use, trade clearly below it, and healthcare around the average. Sport falls outside the sectors CBS measures here. Recurring work that suits a process exists in all of them, such as reservations at a restaurant, quotes and work orders at a heating or plumbing contractor, or returns and product copy at an online shop. In the sectors at the bottom of the table, fewer than one in twenty comparable firms use AI. CBS does not measure how many of them have it in a process.

Does AI in your business sit in a process or in a tab?

Take every place where you use AI and ask five questions about it. Look at what happens before and after the tool.

Does it start without anyone opening a tab? If not, it sits in a tab. Fine, but then the time saved depends on someone's memory and calendar.

Does the result land in the system where it belongs? Product copy you paste from a chat window into your webshop goes through a person. Product copy that is ready as a draft the moment a product is created does not.

Is there a check that is not the same model? A language model reviewing its own reasoning usually does not improve it without outside feedback, and sometimes the answer gets worse. A fixed rule works better: are all required fields filled, does the price match the price list, does the item number exist. Only then a person, when in doubt.

Does anyone know when it goes wrong? A process that stops without anyone noticing only gets spotted when a customer calls. An alert has to go to someone who can act on it.

Who maintains it? Integrations change and tools get updates. Based on 134 days of logs from our own system, we budget for a five-to-ten-minute check every day and half an hour a week. How we arrived at that is in what an agent system costs to maintain. If you do not schedule that time, within a few months you have a system that keeps repeating the same error.

If most answers are no, do not start with a big project. Pick one flow that comes back every week, with a clear starting point and a fixed place where the result needs to go. Returns, quote requests or new products in your webshop are good candidates. Finish that one flow, including the check and the alert. Then you know how much time it saves and how much attention it needs, and those two numbers tell you what to pick next.

What does Oase do with AI in processes?

For us this falls under Technology & AI: the website or webshop where the brand lives online, and the software and automation that keep it running afterwards. After a launch we work out together what still needs ongoing attention, and work that keeps coming back can be automated. That sits next to websites and Shopify webshops on our services page.

Our rule for this comes from our own maintenance log: automate the repetition, not the decision. Anything that is the same every week can become a process. Anything that involves judgement stays with a person. Whatever goes out in a brand's name, such as a blog, a newsletter or a customer email, is checked by a person first.

Building a webshop with AI is quick, but the last twenty percent decides whether it works, and that is exactly where the integrations are, the part a tab does not have. With AI blog writing we deliver articles in which every figure links to a source we have checked, and which a person reads before they go live. How AI systems describe your brand outside your own site is a different topic, covered in brand positioning for AI agents.

Most of the time we start with the question from this article: does AI sit in a process or in a tab for you, and which recurring work would you hand over first? A chat window is sometimes enough. A flow that passes through the same hands every week usually is not.

Not sure which side your first flow is on? Tell us where it gets stuck, and we will look together at where best to begin.