They counted the people first

The Dutch cabinet has decided how many civil servants it will cut by 2035, but it won't say how AI will do their work until next year.

On Monday morning a state secretary in The Hague sent a letter to the Dutch parliament with a number in it, and by lunchtime the number had left the letter and gone travelling on its own. Eric van der Burg, who holds a job the Netherlands has never had before, state secretary for Effective Government, wrote that central government should work with about 25,000 fewer full-time jobs by 2035. The cabinet of Rob Jetten wants to save €900 million on its own running costs in 2027, rising to €3.5 billion a year from 2035, and 25,000 is roughly what that money looks like when you turn it back into people. Police and defence are exempt. The split by ministry comes in the spring fiscal update, the Voorjaarsnota, in about six months. And the part that made the headlines, the plans for how artificial intelligence will actually do the work, arrive in the first quarter of 2027.

So the order of events, if you read the letter rather than the posts about it, goes like this. First the budget needed closing. Then the money was turned into a headcount. Then the headcount went public. The plan for the machines that are meant to make the headcount possible comes next year.

That order is the whole story, and I think it is worth slowing down for, because the Netherlands is not the only place doing it this way, and it will not be the last.

What the letter actually says

The bare facts are less dramatic than "Netherlands replaces 25,000 civil servants with AI", which is roughly how it spread on X within hours, but they are still large. Central government employed about 125,000 full-time equivalents in 2020 and about 157,000 in 2024, according to the government's own annual report on its operations, and Bloomberg puts the figure at a record 160,000 at the end of 2025, some 41% more than in 2018. Reuters counts around 46,500 jobs created since 2018. Twenty-five thousand off the top is around 16% of the whole, though Reuters and the Telegraph rounded it to "around 20%" and "15pc" respectively, which tells you how loosely the figure is being handled even by the people reporting it.

The cuts will not fall evenly. Van der Burg wrote that most civil servants work in delivery, the agencies that actually hand out benefits, collect taxes and process permits, so most of the target lands there. Bloomberg, citing the letter, reports that some ministries, inspectorates and delivery agencies could lose 45% or even 60% of their staff. NOS, the Dutch public broadcaster, noted that agencies already under strain, the benefits agency UWV, the immigration service IND and the tax office, the Belastingdienst, are not exempted in advance.

AI is in the plan, but it is not the whole plan, and the letter is honest about that. There are 500 rules a year to scrap or simplify, half for citizens and professionals and half for businesses. The 28 permanent advisory councils and roughly 150 advisory committees are to be thinned. The 400 or so policy evaluations the government runs each year are to be cut back. Staff will share desks, one workplace for every two people. Surplus buildings get sold, IT gets standardised, departments buy together. And, in a line that got much less attention than it deserved, the cabinet wants to look at reforming the Wet open overheid, the Dutch freedom of information law, so that answering the public's questions takes fewer civil servants.

Van der Burg himself told NOS that working more efficiently alone will not be enough. "Choices are needed about what the government will do differently, less, or no longer," he said, and those choices have not been made yet. In an interview with NRC, he gave the most concrete example of what AI might do that I have seen from anyone in the cabinet: "If you can replace a number of phone lines with a chatbot, that can save time. Then you don't have to call between nine and five, you can ask your question 24 hours a day."

That is the shape of it, then. A firm number, a firm saving, a firm date, and a chatbot on a phone line as the worked example.

Why the number came first

It would be easy to read all this as a government that has seen a demo and believes the hype. I don't think that's what happened, and the NRC interview makes it fairly plain. The governing coalition of D66, VVD and CDA had budgeted €2.4 billion in savings on the civil service. In this autumn's budget, the Miljoenennota, another billion was added. Asked why, Van der Burg said the budget had to balance. "If we hadn't made this choice, you would have had to find that billion somewhere else, which also hurts."

He made a second argument too, which I find more interesting. The central government, he said, has become the biggest competitor in the Dutch labour market. A local councillor had told him that very day that her municipality couldn't find civil servants because the national government offered better terms. There is a real case that a state which grew by a quarter in four years is crowding out the councils, schools and firms around it, and that shrinking it would help them.

None of that is about AI. It is about money and the labour market and a decades-old ambition. Trouw pointed out that almost every Dutch cabinet since Ruud Lubbers in the 1980s has promised to shrink the civil service, and that it almost never happens; full-time jobs at central government went from about 109,000 in 2016 to 160,000 in 2025. When Van der Burg was asked by NOS why this time would be different, his answer was that "everyone sees that it has to be done differently."

What is new is not the ambition. What is new is that there is now a technology to name when someone asks how. Parliament has passed non-binding motions calling on the cabinet to raise government productivity by 20% through AI, and Van der Burg pointed to those. One X account that tracks AI and pay, @raisetal, put the gap in a single line under a viral post on Tuesday: the savings are budgeted for 2027, and the 20% productivity gain has no deadline. Somebody has to stand in the space between those two things, and for now it is the 160,000 people who still work there.

The country that did it first

If you want to see how this goes when the paperwork comes out, New Zealand is about eighteen months ahead.

In May, Finance Minister Nicola Willis announced that the core public service would shrink by 8,700 roles to 55,000 full-time equivalents by 2029, which she said would save NZ$2.4 billion over four years. RNZ has since obtained sixteen documents that led up to that announcement, and Phil Pennington's reporting on them reads like a preview of the Voorjaarsnota.

In April, officials were working towards a target of 60,000 by 2029. Three weeks before the Budget, the Public Service Minister, Paul Goldsmith, told them to "think about 55,000 as a starting point". Officials came back and said 55,000 could worsen the risks to performance, and advised extending the timeframe by two years to 2031. Within days, the Public Service Commissioner, Sir Brian Roche, backed the faster timetable anyway, and said he remained "open to a lower number". The 55,000 figure, he wrote, "should be seen as a minimum so that as the transformation is implemented the full impact and value of AI is captured." It would mean, in his words, "a much [more] aggressive/targeted approach to AI."

Here is the line from the coverage that I keep coming back to. The documents do not model how many jobs AI might replace, or what the rollout would cost, or which services it would improve. The AI is the reason the number is achievable, and nobody wrote down how. A retired forestry consultant in Motueka, Roger May, emailed the finance minister to say exactly this, that replacing 8,700 knowledgeable people with software "is going to take time and money," and that he wouldn't mind betting the cost eats into the $2.4 billion. He got an acknowledgement from her office.

Roche, to be fair to him, has been open about what he thinks the deeper fix is. In a letter from August 2025 he argued that New Zealand should go from 42 agencies to somewhere between 15 and 20, "underpinned by common technology and a lot more AI," and said he would front it and defend it. "If it was my business and my money this is what I would do," he wrote. It is a coherent view. It is just a view in which the machines are the premise rather than the finding.

The evidence, such as it is

The awkward thing about this week is that the same few days brought one of the starkest pieces of evidence yet that AI has not, so far, made ordinary organisations cheaper to run.

On Saturday Torsten Slok, chief economist at Apollo, published a short note titled "No Signs of AI Payoff Outside the Tech Sector". He split the S&P 500 into the seven big tech companies and the other 493. For the seven, inflation-adjusted revenue per worker and profit margins are rising. For the 493, both are flat, and productivity growth shows no AI lift at all. "Despite massive investments and widespread hype around artificial intelligence," he wrote, "the technology has delivered zero financial payoff to corporate America outside of tech." Mainstream companies, he added, are starting to treat token spending as a cost that weighs on margins rather than a source of growth.

An independent analysis of SEC filings published the same day by Tensorust went further and found a twist. For 291 companies that are not the big seven, not banks and not in the AI supply chain, the combined operating margin over the last four quarters was 10.43%. For the four quarters before that it was also 10.43%. Not a basis point of change. The part of the 493 that did get richer is the part selling chips, memory, servers and power to the big seven, with 27 suppliers taking 66% of the group's profit growth. Micron alone went from $7.6 billion of operating profit to $59.2 billion.

The honest reading of that is not "AI does nothing". Slok's own earlier note in June argued that heavily regulated, slow-moving sectors face a long runway before process changes turn into productivity, and he listed health care, banking, energy, defence and transport among them. Tensorust's author reads the flat line as proof that big, slow incumbents are bad at adopting new tools, not that the tools are useless. Goldman's Jan Hatzius still expects AI investment to pay off in economy-wide profits eventually.

But a government ministry is about as heavily regulated and slow-moving as an organisation gets, and the Dutch delivery agencies, the ones expected to give up the most people, are exactly the kind of place where Slok says the runway is longest. The plan assumes the payoff the evidence has not yet shown, in the sector least likely to show it first, and books the saving from 2027.

The UK offers a smaller, stranger data point. Hundreds of Whitehall civil servants were given a home-built AI system called Redbox to summarise documents, answer questions and transcribe meetings. The Telegraph reports that a review of its use found staff were unwilling to put disclaimers on work revealing they had used it. The tool was used. Nobody wanted to say so. That is not a story about productivity, but it is a story about how hard it is to even measure whether you have any.

The Dutch have been here before

There is a particular reason the Dutch should be careful with this, and NRC put it gently in the middle of its interview. After the 2008 financial crisis, the government cut personal service at the UWV and the Belastingdienst on the theory that people could be helped just as well online. In 2019 the then social affairs minister, Wouter Koolmees, concluded that those cuts had gone too far. The Belastingdienst, NRC notes, recently warned that it is not managing to collect all the money it is owed.

And the Belastingdienst is where the most painful Dutch lesson about automated government already lives. In the childcare benefits scandal, the toeslagenaffaire, thousands of families were wrongly accused of fraud, with risk-scoring by the tax authority's systems part of how they were singled out. Mark Rutte's third cabinet resigned over it in January 2021. That system was not a large language model, and nobody is proposing to rebuild it. But it is the institutional memory of what happens when a stretched agency leans on automated judgement and has too few people left to notice the cases going wrong.

The trade union FNV's negotiator for central government, Jasper de Jong, said it in the bluntest possible terms to Trouw. "Only the numbers are being looked at," he said. The cabinet wants to cut "rücksichtslos", recklessly, "without a vision of where that can happen. From all sorts of departments, cries for help are coming in about staff shortages." And then: "25,000 jobs sounds good for the gallery. But this is also about 25,000 people who now live in uncertainty. They carry out all the rules that politicians draw up, and now they hear from those same politicians that they are not essential."

The Dutch financial site De Belegger made the opposite, equally fair point: 25,000 FTE is not 25,000 dismissal letters. Two half-time workers count as one. A lot of it can come from not filling posts as people retire or leave, which is what most coverage on X assumed when it said "natural attrition". It also noted, quietly, that if the cuts come mainly from not replacing leavers, the people who feel it most are the young graduates who would have been hired, and that "there is no proven return yet with which the whole shrinkage can be paid for."

The other half of the plan

One part of the Dutch plan cuts against the usual pattern and deserves credit. Bloomberg reports that the cabinet wants to work with Dutch and European suppliers where it can, to avoid depending on technology from outside Europe. The tax office is already moving off Microsoft 365's cloud onto its own servers and open-source software, and the government is building its own Linux desktop. Whatever you think of the headcount, a state that wants AI to carry more of its work and also wants to own the systems that do it is thinking about the right risk.

That risk was spelled out in New Zealand too. Commentators there warned RNZ that once a government has removed the people and competition among AI vendors thins out, token and licensing prices will likely go up. A ministry that has cut 60% of its staff around a vendor's model has very little bargaining power left at renewal time. Sovereignty, in that light, is not a flag-waving extra. It is the only protection against the saving being quietly clawed back.

The Dutch parliament supplied its own comic footnote. During question time on Tuesday, VVD MP Queeny Rajkowski, who covers the digital economy, asked why the Netherlands shouldn't invite Anthropic to list in Amsterdam, "so we can levy taxes and tell them what they can and cannot do." It went round X as a meme within hours. It was a throwaway line, and the mockery was predictable, but it shows the same instinct as the Linux desktop: if the state is going to run on someone's models, it would rather have some say over them.

What to watch

The best joke about all of this came from Daragh Cassidy, writing about Ireland, and nearly four hundred people liked it: "In Ireland, we'll probably end up hiring another 25,000 civil servants to monitor AI, analyse it, regulate it, and report on how we can use it better to increase efficiency." It is funny because the Dutch and New Zealand numbers both rose for years under governments that promised to cut them.

But I think the more useful thing to say is that both countries have now done something new. For the first time, a government has written a specific number of public jobs into its budget and named AI as part of how it will get there, before writing down how. New Zealand's papers don't model it. The Dutch plan arrives in the first quarter of 2027. The money is already counted.

So there are two documents worth waiting for. The first is the Dutch AI plan early next year, and the question to ask of it is simple: does it say which tasks, in which agencies, at what cost, will be done by software, and how anyone will know whether they are being done well? The second is the Voorjaarsnota in the spring, when each ministry learns its share, and the question there is whether the 60% cuts land on the same delivery agencies that are already telling parliament they can't keep up.

If those documents fill in the gap, the Netherlands will have done something genuinely useful and shown the rest of Europe how to shrink a state with software carefully. If they don't, it will have done what governments have always done with a hard budget and a new technology, which is to promise the technology will close the gap and let the people who are left find out whether it does.

Either way, the order is already set. They counted the people first.