🔍 Read the full analysis: How AI Is Finding A Place In Europe’s Government Services on ThorstenMeyerAI.com
TL;DR
Developments in the Netherlands, Britain and Sweden show governments using AI in staffing plans and asylum casework while setting limits on public tools. A UK health consultation also received a large volume of responses flagged as bots. The cases raise questions about accuracy, transparency and how public participation is verified.
European governments are taking sharply different approaches to AI, as the Netherlands links it to plans for a smaller civil service, Britain faces scrutiny over AI summaries used in asylum casework, and Sweden blocks public versions of ChatGPT and Claude on government computers. The cases highlight a practical divide: AI is moving into public administration, but rules on accuracy, disclosure and oversight remain uneven.
In the Netherlands, a letter sent to parliament on October 26 by State Secretary Eric van der Burg sets out central-government spending reductions that rise to €3.49 billion a year from 2035. The government describes a possible reduction of about 25,000 full-time positions as a rough estimate. AI is one of several proposed efficiency measures, alongside simpler rules, shared workspaces and standardised technology. Police and defence are exempt from the cuts.
The letter also cautions against assuming AI will produce large staffing savings. It says previous digitisation and automation did not deliver promised reductions in staff and costs, in part because they created new work and reliance on contractors. The cabinet says it will publish AI targets and savings plans in the first quarter of 2027. The job figure is tied to a wider budget decision, not an announced AI-only layoff plan.
In Britain, the Home Office uses two tools built on OpenAI models for asylum casework: one summarises interview transcripts and another searches official country information. The digital rights group Open Rights Group (ORG) says applicants are not told when AI is used and cannot review or correct the summaries. The Home Office says trained caseworkers make decisions and the tools do not decide claims. A Home Office spot check found that 15% to 22% of sampled summaries failed review, compared with a 9% inaccuracy or incompleteness rate in the earlier pilot.
Infographic 01 · The Netherlands
25,000 fewer civil servants. AI is one tool, not the reason.
The driver is a budget cut. The government plans to use AI, and warns against expecting too much from it.
The cut grows every year
Cuts to the running costs of central government, € billion per year.
Police and defence are exempt. Many organisations face about 20% fewer staff, some up to 45%.
Where the money must come from
The letter’s own split of the cut.
AI is one of several tools inside both parts, next to simpler rules, standard IT, shared desks and joint purchasing.
AI in the plan
- “Ambitious targets and savings plans” for AI in Q1 2027
- Purchasing to examine “systematically” where AI fits
- Conditions: security, data quality, sovereignty
Also in the plan
- At least 500 rules a year scrapped or simplified
- One desk per two employees
- Fewer advisory bodies (28 councils, ~150 committees)
The letter’s own warning
- Past digitisation did not deliver the promised staff cuts
- It brought new tasks and more contractors
- Cuts may affect service and continuity
“Do not overestimate the staffing effects of digitalisation and AI.”Lesson 2 in the Dutch government’s own letter, 26 Oct 2026 · translation: Thorsten Meyer
Sources: BZK, Voortgangsbrief Slagvaardige Overheid + annex, 26 Oct 2026; Rijksoverheid news release, 26 Oct 2026. 25,000 ÷ 160,016 ≈ 1 in 6 and bar scaling: Thorsten Meyer.
Powered by Thorsten Meyer AI
AI Touches Jobs and Asylum Decisions
These cases matter because administrative AI can affect both how public services are staffed and how officials handle information about people. In asylum processing, an inaccurate summary could leave relevant details out of the record a caseworker uses, even if the software does not make the final decision. The Home Office’s stated position that human staff decide claims does not answer whether applicants can identify and correct errors in summaries.
The reported scale of checking is also relevant. ORG says roughly 17,515 cases a month pass through the summarisation tool, while experts reviewed 203 summaries in a 2026 spot check—about 0.3% of the estimated total from April to July. That sample does not establish the error rate across every case, but the difference between the pilot and spot-check results makes ongoing quality control a public concern.
Other examples show that AI’s role in government is not limited to direct casework. A UK health consultation received 118,756 replies in 2023; the Department of Health and Social Care flagged 90,835, or 76%, as bots and excluded them. The source account says the consultation’s result changed when those responses were removed. That episode illustrates how automated participation can affect public input, though the available material does not establish who generated the responses or why.
Popular right now
LG C2 Series 65-Inch OLED evo 4K Smart TV (OLED65C2PUA, 2022)
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
De'Longhi Magnifica Evo Fully Automatic Espresso Machine with Manual Milk Frother
As an affiliate, we earn on qualifying purchases.
Different Rules Across Public Services
The Netherlands’ staffing estimate sits within a plan to reduce central-government running costs, not a standalone programme to replace civil servants with AI. The letter says the covered ministries, inspectorates and agencies employed 160,016 full-time equivalents at the end of 2025. It also estimates that general efficiency measures would cover only about a third of the planned reduction; the rest would require decisions about which government activities to stop.
In Britain, the two asylum tools have different functions: the Asylum Case Summarisation tool condenses interview transcripts, while the Asylum Policy Search tool helps staff find country information. According to ORG’s account of documents obtained through information requests, the tools use GPT-4o through Microsoft’s Azure cloud. The search tool reached all decision-makers in July 2025, and the summary tool reached all caseworkers in April 2026. ORG published its findings on October 12 after six months of parliamentary questions and freedom-of-information requests.
Sweden’s Government Offices blocked browser access to public versions of ChatGPT and Claude on staff computers from October 26, according to the source material. Staff are directed to approved tools instead. This is described as a restriction tied to contracts and data handling, rather than a general ban on AI. The source also says the EU AI Act classifies AI used in asylum decisions and access to public services as high-risk, with the relevant duties due to start on December 2, 2027, following changes under the digital omnibus.
„The staffing translation is a rough estimate.“
— Dutch government progress letter to parliament, sent by State Secretary Eric van der Burg
Accuracy and Oversight Remain Unsettled
The available information does not show how often AI-generated asylum summaries contain errors across all cases. The Home Office’s spot check covered 203 summaries, and the source gives no evidence that the sample represents every caseworker, period or type of interview. It is also unclear whether applicants will be told about AI use or given a way to challenge a summary before it informs their case.
ORG says neither asylum tool appears in the UK algorithmic transparency register, although the government’s AI Playbook asks departments to list such tools. The source reports that the Independent Chief Inspector of Borders and Immigration asked ORG for more information while planning next year’s inspections; the inspector did not confirm a formal probe. The source material ends mid-sentence after referring to a July ruling by an upper tribunal judge, so it supplies no further detail on that development.
For the Dutch plan, the final employment impact remains uncertain: the government calls the 25,000 figure a rough estimate and has not yet published the AI targets and savings plans. In Sweden, the available account does not specify which approved tools staff may use or what data restrictions apply to them.
Government Plans and Reviews Ahead
The Dutch cabinet is due to publish its AI targets and savings plans in the first quarter of 2027. Those plans should clarify which functions may use AI, what savings officials expect, and how those estimates relate to the wider reductions in central-government spending and jobs.
In Britain, the next reported milestone is the Independent Chief Inspector of Borders and Immigration’s inspections planned for next year. The inspector’s request for more information from ORG may inform that work, but the source does not confirm a specific investigation. Further disclosure from the Home Office about error checks, applicant notification and the tools’ transparency-register status could clarify how the system is monitored.
The EU AI Act duties described in the source are scheduled to apply from December 2, 2027. Until then, these national decisions offer an early view of the questions governments face: what tasks AI should support, how people can contest errors, and how public agencies can verify that digital input reflects genuine public participation.
Key Questions
Is the Netherlands cutting 25,000 jobs because of AI?
No. The government’s estimate of about 25,000 fewer full-time positions is linked to broad spending reductions. AI is one of several proposed efficiency measures, and the government warns against overestimating its staffing effects.
Does AI make asylum decisions in the UK?
The Home Office says trained caseworkers make every decision and that its tools do not decide claims. One tool summarises interviews and another searches country information, but concerns have been raised about summary accuracy and applicant disclosure.
How many UK asylum summaries failed the spot check?
In a 2026 check of 203 summaries, Home Office experts marked 45 as failing the first review, or 22%, and 30 as failing the second, or 15%. The sample does not establish the error rate across all summaries.
Did Sweden ban AI in government offices?
No general AI ban is described. Sweden’s Government Offices blocked browser versions of ChatGPT and Claude on staff computers from October 26 and directed staff to approved tools instead.
When do the relevant EU AI Act duties begin?
The source material says the duties for the high-risk uses discussed are due to start on December 2, 2027, following changes under the digital omnibus.
Source: ThorstenMeyerAI.com
