On 26 October, the Dutch cabinet told parliament it expects central government to need about 25,000 fewer full-time positions over the next decade. The Straits Times called it perhaps the most ambitious government plan yet to use technology to shrink the state's payroll. The Straits Times

Before a board treats this as a precedent, it is worth reading what the plan actually says. The budget-derived staffing estimate precedes the detailed AI targets and savings plans. That order matters.

Read the letter, not the headline

The figure comes from a progress letter by state secretary Eric van der Burg. Two successive cabinets booked budget reductions on the civil service, including about €900 million in 2027, rising to €3.5 billion annually from 2035. The letter calls the 25,000 FTE figure "a rough estimate" of what that budget implies. Dutch government letter

Police and defense are exempt, so the cut falls mostly on ministries, inspectorates and delivery agencies, which employ about 160,000 FTE. For many organizations the letter expects a reduction of about 20%, and in some cases far more.

AI is one lever among several. The letter also lists fewer desks, simpler procurement, standardized internal services and fewer rules and advisory bodies. AI initiatives are under way, but concrete targets and savings plans are due only in the first quarter of 2027. Government announcement

The letter is also candid about the arithmetic. Generic efficiency, which the CPB, the government's economic policy bureau, puts at 0.5% a year, covers roughly a third of the target. The rest requires choices about what government will do differently, do less, or stop doing.

Then, in an annex on lessons from earlier rounds of cuts, comes the line every board should read before citing this plan: "Do not overestimate the staffing effects of digitalization and AI." Earlier digitalization, it says, led to new tasks, more capacity and more contractors rather than the promised reductions. (Translations from the Dutch here and below are ours.)

The same annex warns that blunt instruments such as hiring freezes drove good people out and blocked new and young staff. "Years later," it says, "you are missing entire cohorts."

None of this makes the plan unreasonable. A government can legitimately decide to do less. The civil service grew about 41% between 2018 and 2025, though the annex cautions that returning to 2019 or 2020 levels would ignore the tasks and burdens added since. Either way, doing less is a decision about scope and budget. Presenting it as an AI result would hide the decision that actually produces the saving.

Parliament's committee debate is set for 1 December.

More output is not more finished work

Software offers useful firm-level evidence. Fiona Chen and James Stratton of Harvard studied 300 million work events across 718 firms on the engineering analytics platform Jellyfish, comparing firms that adopted AI tools at different times. After firms adopted AI coding agents, lines of code rose 30%, commits 20% and pull requests 23%. Chen and Stratton working paper

Completed output did not keep pace. The paper measures it through resolved issues and closed projects, which are proxies, not a full measure of product quality or business value. Those estimates were small and statistically insignificant, and the authors can rule out a gain larger than 12% for agents.

The queue moved instead. Calendar time from submitting code to merging it rose 49%, from a baseline of about seven days. That is elapsed time, not reviewer effort, but other signals agree: the share of submissions sent back for changes nearly doubled, comments per submission rose 35%, and the share of workers doing reviews rose 14%.

The study found no statistically detectable employment effect over its observation window, which is not proof that no effect will come later.

The limits matter. This is an observational working paper, its data end in March 2026, and it covers software teams on one analytics platform. Combining two different measures, the paper puts engineers' coding time at roughly 30–42%: about 42% imputed from platform data, and 30–40% in a separate survey of 100 engineers. Speed up that part, and the constraint can move to review and integration. The same bottleneck is worth testing elsewhere.

Five tests before AI savings become job cuts

Taken together, the Dutch letter and the software evidence turn the headcount question into a measurement problem. Finished output does not have to rise for AI savings to be real. Delivering the same volume and quality with fewer labor hours is a genuine efficiency gain. Higher output alone, on the other hand, does not prove a saving. I would ask five questions before any AI-linked reduction is approved.

1. Did the cost of acceptable finished work fall? Track accepted outcomes per labor hour and total cost per accepted outcome, holding quality, timeliness and case mix comparable. Count licenses, inference, integration, supervision, rework, contractors and transition costs.

2. Where did the queue move? Measure review time, backlog, tail delays, rework and escalations alongside output. A slower queue does not tell you which role to keep, but it shows where the constraint now sits.

3. Which part of the saving comes from AI? A simple before-and-after comparison cannot separate AI from process changes, workload shifts or staffing changes. Use a comparable baseline and, where feasible, a phased rollout. Separate AI efficiency from simplification and from decisions to stop doing work, as the Dutch letter does on paper.

4. Who will check and own the result afterward? For every workflow affected, name the person accountable for the outcome and the people who review work and handle exceptions. That capacity may be treated as overhead in a cost exercise; protect it.

5. Has the redesigned workflow been piloted? Run it with representative demand and the proposed staffing arrangement, while retaining a fallback. Set stop and reversal criteria in advance. A quarter can serve as an example review window, but use enough time to include representative workload, exceptions and peaks.

One more question belongs in the decision memo: what happens if the saving does not appear? Compare the cost of retaining capacity during the transition with the cost and risk of rebuilding it if the expected savings fail to appear. Attrition and redeployment may be preferable, but they are not automatically cheaper or right for the skill mix.

Entry-level hiring deserves its own line. My concern, and this is my inference rather than the annex's finding, is that prolonged entry-level hiring freezes could also weaken the future pipeline of experienced reviewers and managers.

The people doing the work can supply evidence these tests need: the exceptions they caught, the work they completed, and the operating constraints that never show up in an output dashboard.

Start with one function

Choose one function where AI-related savings are being discussed. Establish a comparable baseline, then measure cost per accepted outcome, queue time and rework as AI is phased in. Treat as savings only what survives that measurement, and decide on headcount afterward.

Before attributing a staffing reduction to AI, show which work changed, what it now costs, and whether the service still holds up.