The Dutch civil-service target needs a service baseline before AI gets credit

A target with several mechanisms

The Dutch government statement says central-government organisational savings should reach €900 million in 2027 and €3.5 billion from 2035. It translates the longer-term figure into an estimated reduction of 25,000 full-time-equivalent roles over several years. Police and defence are among the excluded areas.

The programme is broader than an AI deployment. It includes simplifying or ending rules, standardising IT, reducing advisory bodies and evaluations, changing procurement and deciding which activities government should continue. The statement says goals and savings plans for using AI to improve productivity will arrive in the first quarter of 2027. That sequence matters: the staffing estimate exists before the detailed AI contribution plan.

Do not make AI the residual explanation

Reuters reported the headline as roughly one fifth of government jobs over a decade, while also noting the future AI plan. The ratio depends on which central-government population and exemptions form the denominator. A rounded percentage is not a task map, and a savings target is not evidence that a model can absorb the work.

Four mechanisms can reduce measured FTEs: vacancies left unfilled, tasks stopped, work transferred or consolidated, and productivity gains within retained services. AI may contribute to the fourth mechanism and can also affect the design of the others, but it should not receive automatic credit for every departure. Conversely, a lower headcount can coexist with longer queues, narrower access or more contractor spending.

Establish the service baseline

Before changing roles, record a baseline for each service: demand volume, completion time, error and appeal rates, accessibility, case complexity, staff effort, external spend and public-value obligations. Separate policy work from operational delivery and supervision. The same percentage reduction can have very different effects where work is discretionary, legally required or safety-critical.

For each proposed AI use, name the task rather than the occupation. Record permitted inputs, required review, exception classes, failure handling and the skill needed to operate or challenge the system. Pilot against historical cases and a live shadow period before changing the staffing assumption. If throughput rises but appeals or unresolved cases rise too, the claimed productivity gain has not cleared the service test.

Track transitions, not only exits

A multi-year programme needs an attrition and capability ledger. It should distinguish retirement, voluntary departure, redeployment, vacancy suppression, contractor substitution and compulsory loss. It should also show which competencies are disappearing, which are being built and where workload moves. Aggregate FTE totals can hide a new bottleneck in legal review, data stewardship, procurement or citizen support.

The Dutch statement leaves important implementation choices open. It does not allocate all 25,000 FTEs to named organisations or prove that AI can produce the required savings. Detailed choices are due through a spring budget process and a multi-year approach. Treat those future documents as decision gates, not administrative follow-up.

Reconcile the ledgers quarterly

Review the staffing, workload, service and supplier ledgers together. A vacancy that disappears from payroll but returns as consultancy spend is not the same transition as a task that was safely stopped. Likewise, a faster transaction is not a net capacity gain if appeal handling or accessibility support expands elsewhere. Reconciliation keeps the programme from optimising one visible total while exporting cost and risk.

Decision rule

Do not count an unfilled role as an AI productivity gain unless the associated service volume and quality remain within a pre-agreed band and the task transfer is traceable. Publish separate series for gross departures, net FTE, contractor spend, workload, service outcomes and validated AI contribution. This adds a service-capacity test to the escalation logic in China’s AI employment plan.