Sweden sorts its AI by contract, not by model

Since Monday, October 26, 2026, Swedish government staff can no longer open ChatGPT or Claude from their work computers. The decision does not say these tools are bad. It says they were being used without a contract.

What was decided #

The Regeringskansliet, the administration that brings together the Swedish Prime Minister's Office and the ministries, has blocked access to ChatGPT (OpenAI) and Claude (Anthropic) on its staff's computers. The news was broken by Ekot, the investigative desk of Sveriges Radio, on October 23, 2026, then confirmed by the Regeringskansliet to the public broadcaster SVT.

Staff are steered to the services the administration has approved, including Copilot, Microsoft's AI tool. According to internal information seen by Ekot, the Regeringskansliet raised Copilot's security after a review of the service. The press office declines to say more, on the grounds that these are security matters.

The criterion: the contract #

This is the heart of the decision. The Regeringskansliet did not pick one model over another. Some of its staff keep access to ChatGPT and Claude, through an AI assistant platform the administration built itself. What is cut is direct use of a service the administration has no contract with.

In other words: the same model can be barred through one door and allowed through another. What makes the difference is who signed, and who configured the tool.

Not an isolated case #

Ekot surveyed other Swedish agencies. Skatteverket (the tax agency) and Statens institutionsstyrelse (the national board of institutional care) have already blocked these services. Kronofogden (the enforcement authority) is preparing its position. The Regeringskansliet had itself already restricted DeepSeek.

The most telling finding comes from Trafikverket, the transport administration. Its internal audit found earlier this year a widespread and risky use of about ten different AI services. The auditors write: “we have seen several examples of problematic use of open AI services” (our translation). They point both to the sensitivity of the information entered and to how the generated answers were handled.

About ten services, in a single agency, found by one audit. That is the figure to remember.

What the decision does not say #

• It is not a ban on AI: the administration approves some and builds others.
• It is not a verdict on the quality of ChatGPT or Claude.
• The decision covers only the Regeringskansliet's computers, not Swedish citizens or Swedish companies.
• Ekot specifies neither the version of Copilot approved nor what the security upgrade contains.

Copilot can run Claude #

One detail blurs the line more than it seems. In Microsoft 365 Copilot, Microsoft also offers Anthropic's models, from the maker of Claude. For Word, Excel and PowerPoint, the setting “Copilot in M365 apps with Anthropic models” is on by default for tenants in the European Union, EFTA and the United Kingdom created after March 25, 2026; for older tenants, Microsoft points to each tenant's Message Center (Microsoft documentation updated on October 23, 2026). When these models are used, processing takes place outside the European Union Data Boundary. Ekot does not say how things stand in the Swedish government's Copilot. For a company, the question is therefore not only “which tool?”, but also “which model behind it, and who turned it on?”.

And in a French company? #

Most companies with 150 to 499 people are not in the Swedish government's position. They often have neither a list of approved AI services nor a view of those used without a contract. The Swedish case comes down to three simple questions:

1. Which AI services do your employees use today, and which of them do you have a contract with?
2. Among those you approved, who decided each setting, and when?
3. For each approved tool, do you know which model runs behind it, and where the data is processed?

You are not behind. You do not know where you stand. That is where it starts.

Every Monday, Ojja's feed records what vendors change in companies' tools, from their own messages: ojja.ai/fil