Since Monday 26 October, staff at the Swedish Government Offices can no longer reach ChatGPT or Claude in the browser on their work computers. They work with the approved tools, including Microsoft Copilot, whose security was tightened after a review. The models behind both stay in the building: part of the staff keeps using them through Intric, an AI platform bought through a public tender.

The Government Offices have closed the open browser route and kept the models on a route they control by contract.

What happened in Stockholm

Swedish Radio broke the news on 23 October, and the Government Offices confirmed the details to SVT. Digital director Mikael Wettercrantz warned staff that AI services used without a contract can have "consequences for our information security". Intric, procured this summer, runs on Swedish infrastructure, with a contract covering both cloud and on-premises use.

The Tax Agency and the National Board of Institutional Care (SiS) already block similarly, and the Enforcement Authority is preparing a decision. The Swedish Transport Administration shows why: its internal audit earlier this year found widespread, risky use of around ten different AI services. "We have reviewed several examples of problematic use of open AI services," the auditors wrote. "The examples concern both the sensitivity of the information entered and how the AI-generated answers were handled."

The contract decides

On the open web service, someone signs in with a personal account: no contract, no view of where the data ends up. On the contracted platform, a data processing agreement (DPA) sets out what the provider may do with the data, where it is stored and who can access it.

So contract, data route and control decide whether AI runs safely in an organisation. And a ban without an alternative just pushes use to where nobody is looking.

What about Germany?

The state of Baden-Württemberg draws the same line with its own technology. Its administration runs an in-house AI assistant, F13, for around 50,000 employees, on the state IT provider's servers and with open-source models (Staatsanzeiger, June 2026). Its economics ministry's AI guideline (April 2026) calls F13 "fully supported" and allows ChatGPT and DeepSeek only with restrictions.

At federal level, the finance ministry has "in principle prohibited" the use of "text-generating AI on the internet for official purposes" (dpa, February 2025). Meanwhile the federal IT service centre runs KIPITZ, an AI platform for federal agencies, in its own data centres, as a closed system that keeps administrative data out of model training.

Industry does the same. Wacker Chemie keeps DeepSeek off company computers, and Siemens offers AI models only through an internal platform, so that "Siemens data stays at Siemens" (February 2025). And in June 2025 the Berlin data protection commissioner reported the DeepSeek app to Apple and Google as illegal content, because user data goes to China without adequate safeguards.

What the examples have in common

Stockholm, Baden-Württemberg and Siemens block open access and provide a vetted route at the same time. That takes neither a public tender nor your own data centre. Five things show up everywhere:

  1. An approved tool that keeps up. Stockholm gives everyone Copilot, and Intric to those who need more.
  2. Business accounts with a DPA. A data processing agreement under the GDPR is the basis, including a clear rule on whether inputs are used for training. Germany's data protection authorities (DSK) also recommend that employers provide devices and accounts for AI use. More in ChatGPT at work: data protection.
  3. A known data route. F13 and KIPITZ run in their own data centres, Intric on Swedish infrastructure. For a business, that means known country, provider and subprocessors, with processing in the EU. More in Running AI in the EU.
  4. Clear ownership. The ministry's guideline separates approved from restricted tools and names an AI team that assesses new ones.
  5. Training on real tasks. Baden-Württemberg backs F13 with workshops, e-learning and a prompt library.

If you would like to talk this through

We are happy to take about 30 minutes for a free first call and look at your current AI routes together. If you want to go deeper, we start with an AI audit for €500: your tools, contracts and data flows at a glance. The fee is credited if we go on to implementation together.

AI that moves businesses forward.