UWV figures show fast adoption, but the harder HR question is guidance, trust, and who remains responsible when a tool shapes work.
AI at work is no longer waiting for a board decision. In many Dutch companies it is already inside the day. A sales email is polished before it leaves. A job advert is rewritten. Someone summarises a meeting. A difficult customer reply is made more careful. None of this looks dramatic. That is exactly why it becomes a management issue. When AI touches HR, absence and employer records can become evidence of trust, timing and responsibility, not only administrative notes.
When the record carries the case
UWV's 2025 employer survey gives the signal clearly. The share of employer establishments where AI is not used fell from 60 percent in 2024 to 34 percent in 2025. Use to a reasonable, high, or very high degree doubled from 16 percent to 32 percent. The survey covered 3,551 employer establishments in autumn 2025. UWV describes the results as representative for employers in almost all sectors, with agriculture, forestry and fisheries, and employment agencies and temporary staffing outside the scope.
The gap is not adoption. The gap is guidance. UWV reports that 61 percent of employers already using AI provide guidance to workers. Among employers that do not yet use AI but expect to do so within five years, 38 percent prepare their workers. That difference is the real HR story. Not whether AI is clever. Whether the employer still knows how work is being done.
The tool entered through the side door
A small service company can have an AI problem before it has an AI project. One employee uses a free tool to make Dutch text sound more professional. Another uses AI to summarise client notes. The owner asks a system to compare two job descriptions. A manager tests it on incoming CVs because the vacancy has been open too long.
No one is being reckless on purpose. People are trying to save time. The weak point is that every person may be applying a different rule. Which data may be entered? Which output may go to a client? Is AI allowed near salary, sickness, assessment, rosters, applicants, or performance notes? Who checks the result? If the answer is informal, the company has already built a shadow work method.
CBS adds the worker side of the same picture. It reports that 43 percent of adults in paid work use AI when performing their work. At the same time, 41 percent think AI could partly replace their work, and 4 percent think it could take over their work completely. Among those who think AI could do their work partly or fully, almost half are concerned.
That combination matters. Staff use AI and fear AI at the same time. In that setting, silence from management is not neutral. It can make people hide use, overstate gains, avoid questions, or treat every new tool as a quiet assessment of their future. A calm conversation about permitted use is part of trust.
The first change is tasks, not jobs
The Dutch debate often jumps too quickly from AI use to job loss. UWV's figures are more useful than that. It reports that 49 percent of employers expect training needs to change because of AI in the next five years. Another 45 percent expect different skills, and 29 percent expect functions to change strongly. Only 10 percent expect jobs to disappear, while 15 percent expect new jobs to be added. Another 19 percent expect AI to prevent or reduce staff shortages.
Payment pressure meets privacy
That tells me to look at job content before headcount. The bookkeeper may spend less time drafting a payment reminder and more time checking whether the facts are right. The recruiter may spend less time writing a vacancy text and more time guarding fair selection. The planner may receive faster suggestions, but still has to understand sickness, availability, contract hours, safety, and the strain on the team.
This is where small employers can underestimate the issue. A tool that rewrites a staff announcement is one thing. A tool that helps rank applicants, allocate shifts, monitor performance, or support dismissal preparation is another. Under the EU AI Act, several systems used in employment and worker management can fall into high-risk categories when the relevant conditions are met. Recruitment, candidate filtering, promotion, termination, task allocation based on personal traits or behaviour, and performance monitoring are not ordinary office uses.
The AI Act also contains an AI literacy obligation. Article 4 has applied since 2 February 2025. Providers and deployers of AI systems must, to their best extent, ensure a sufficient level of AI literacy among staff and other persons using AI systems on their behalf. For a small employer, I would not translate that into a thick binder. I would translate it into role-based understanding. The person using AI for marketing text does not need the same guidance as the person handling applicants, payroll input, or Dutch sickness files.
Human review must count
The phrase human review sounds reassuring. In HR, it can become decorative if the system has already framed the outcome. If a tool has ranked candidates, labelled a worker as risky, or suggested a performance path, the manager who clicks approve may feel involved without having truly reviewed the basis. GDPR rules on solely automated decisions with legal or similarly significant effects also point to human intervention and the right to contest in the relevant situations. The management lesson is simple: the human part must have weight.
That is not only a legal point. It is a leadership point. Workers know the difference between a manager who uses a tool as support and a manager who hides behind the tool. Applicants know when a process feels opaque. Existing staff notice when planning, assessment, or workload changes without explanation. In a tight labour market, that perception has a cost.
Labour pressure makes the temptation stronger. CBS reported 378,000 open vacancies at the end of the first quarter of 2026, with 91 vacancies per 100 unemployed persons and unemployment at 4.0 percent. UWV separately reports that 58 percent of employers experience problems from labour-market tightness. It is understandable that an owner-manager looks at AI as capacity. Draft faster, search faster, summarise faster, plan faster.
Time saved is not automatically profit. A faster offer that contains an inaccurate promise can damage margin. A quicker candidate screen that cannot be explained can damage trust. A payroll note prepared by a tool still needs source documents and responsible review. A client file summarised by AI is only useful if the company can still show what was received, what was checked, and what decision was made by a person.
The rule set can stay practical
Large organisations will build committees, registers, and formal procedures. Small companies need something lighter, but not weaker in substance. The first useful step is a map of actual use. Which tools are used? For which tasks? By whom? With which data? Does the output affect a customer, an employee, an applicant, a payroll file, an invoice, a sickness file, or a management decision?
What employers should keep readable
The next line is the data boundary. Personal data, client data, payroll information, medical details, commercial secrets, and internal records should not drift into unapproved tools because a worker is trying to be helpful. This is often where good intentions create weak records. The worker wants a better text. The company later discovers that the input was the sensitive part.
Then comes the HR boundary. AI for ordinary drafting is not the same as AI for hiring, scheduling, appraisal, monitoring, promotion, or dismissal. If a company has an ondernemingsraad, major technological or organisational changes may also raise participation questions under the Dutch Works Councils Act. That depends on the size of the enterprise and the concrete decision. Even where that formal structure is absent, the governance question remains: who is allowed to let software influence a person's work position?
The Dutch government is also building the national supervision structure around the AI Act. On 20 April 2026 it placed the draft implementation law into internet consultation, running until and including 1 June 2026. The government's own explanation of high-risk AI points to data quality, risk management, human oversight, and transparency. Those words sound large, but in a small company they become very concrete. Is the data fit for the task? Who owns the risk? Who checks? Can the decision be explained?
The founder's real question
I do not think the useful response is to frighten small employers away from AI. That would miss the reality of work. AI can make scarce time more useful. It can help a junior worker draft better. It can help a founder prepare a difficult conversation more carefully. It can reduce blank-page work and make routine administration less heavy.
The danger is not the tool itself. The danger is pretending that a tool is only a tool when it has already changed behaviour, evidence, responsibility, and trust. Once AI touches HR, the question is no longer only whether the output is good. It is whether the company can explain the route from input to decision, and whether people know the limits before the next busy day begins.
For the next week, the most practical move is modest: ask the team where AI is already used, without turning the question into a trap. Separate harmless drafting from HR and financial records. Keep sensitive data outside casual experimentation, including absence records and other HR records that need careful handling. Give short guidance by role. Decide where human review must be visible. Test one productivity claim against rework, quality, margin, and cash, not against enthusiasm.
AI at work is becoming normal in the Netherlands. That does not make it harmless, and it does not make it dangerous by default. It makes it part of employer control. The company that benefits will not be the one with the most tools. It will be the one that still knows where judgment sits.
The Polder is written for readers who need the Dutch business environment translated into practical meaning. Corrections, source policy and editorial accountability are part of the publication record.
