The Dutch government's SSC-ICT is rolling out vlam-chat, a secure generative-AI assistant using European open-source models hosted in a government data centre, to roughly 26,500 civil servants across eight ministries ahead of full production in the second half of 2026.
Details
Maturity
Pilot
Promoter
SSC-ICT (Dutch national government) — Vlam.ai Project
Period
2025–2026 (pre-launch); production rollout H2 2026
Keywords
generative AI, digital government, civil service productivity, sovereign AI, secure LLM deployment
Context
Vlam-chat ('Veilige Lokale AI Modellen' — Safe Local AI Models) is being built by SSC-ICT as part of the Netherlands' Vlam.ai project. It runs entirely inside a Dutch Government Data Center on European open-source language models rather than commercial cloud AI, and is designed to help civil servants with writing, summarising and translation, while explicitly preventing the processing of confidential or privacy-sensitive information.
Objectives
The project aims to give civil servants a secure generative-AI assistant built on sovereign infrastructure, avoiding reliance on commercial cloud AI for sensitive government work.
Activities
Eight ministries — General Affairs; Interior and Kingdom Relations; Housing and Spatial Planning; Finance; Infrastructure and Water Management; Education, Culture and Science; Social Affairs and Employment; and Health, Welfare and Sport — have committed to a 2026 launch, giving an initial user base of roughly 26,500 civil servants. The project builds on smaller precedent pilots, including the PZH-Assistant in South Holland province and the Goes municipality's citizen-facing chatbot Guus, and sits within the Dutch government's April 2025 policy expanding civil servants' authority to use generative AI. A companion tool, vlam-search, is planned to let civil servants semantically search internal government documents.
Conclusions
As of mid-2026, vlam-chat remains in pre-production risk and security review, with broad production rollout scheduled for the second half of the year. No independently measured productivity or output-quality results have yet been published — the 26,500-user figure reflects ministry commitments rather than confirmed active usage, an evidence gap worth tracking as the rollout proceeds.
Implementation
Indicative cost
Medium (€50k–€500k) — No budget figures are published in the source. Hosting on existing Dutch Government Data Center infrastructure with open-source models avoids commercial API licensing costs, but integration across eight ministries and ~26,500 users implies a substantial national IT programme — a conservative medium-cost estimate pending curator review.
Time to results
Short (< 1 year) — Builds on earlier pilots (PZH-Assistant, Guus chatbot); sits within an April 2025 government-wide generative-AI policy; full production rollout is targeted for the second half of 2026.
Staffing & skills
SSC-ICT (shared IT service centre for the Dutch national government), Eight committing ministries providing civil-servant user base, Ministry of the Interior — coordinates the overarching generative-AI policy
Conditions for success
Sovereign/local hosting (Dutch Government Data Center) using European open-source models rather than commercial cloud AI
Explicit restriction against processing confidential or privacy-sensitive information
Building on smaller pilot precedents (PZH-Assistant, Guus chatbot)
Where it fits
Governance type
whole-of-government shared-service IT deployment
Scale
national — 26,500 civil servants across 8 ministries (committed, not yet confirmed active)
Income level
high-income (Netherlands)
Data sources
Where this practice's information was retrieved from, and when.
Hong Kong's Digital Policy Office and the HKGAI research centre built HKPilot (document drafting/translation) and HKMeeting (AI minute-taking) for civil …
