# Regeringskansliet blocked ChatGPT and Claude: your playbook

From 26 October 2026 the Swedish Government Offices block the ChatGPT and Claude web interfaces on staff computers, keep Microsoft Copilot after a security review, and let some staff reach the same OpenAI and Anthropic models through the Swedish-hosted Intric platform. The decision sorts AI tools by contract, not by model, and a Swedish company on Microsoft 365 and Azure can copy it in a sprint: block consumer apps, configure Copilot's three EU caveats, and run Azure OpenAI in the EU Data Zone.

- Published: 2026-10-28 · Category: Security & Compliance · Tags: Regeringskansliet, Microsoft Copilot, Enterprise Data Protection, Azure OpenAI, EU Data Zone, Shadow AI, Defender for Cloud Apps, Intric, Sweden, GDPR
- Author: Technspire AB, Stockholm (https://technspire.com)
- Canonical: https://technspire.com/en/blog/regeringskansliet-blocked-chatgpt-claude-kept-copilot-azure-playbook

On Monday 26 October 2026 the Swedish Government Offices (Regeringskansliet) cut browser access to ChatGPT and Claude on every staff computer, from the Prime Minister's Office down through the ministries. Microsoft Copilot stays. And a subset of staff keep using the same OpenAI and Anthropic models through Intric, a Swedish-hosted AI platform the Government Offices procured in July. Read those three facts together and the decision is not about which model is good. It is about which contract the data sits under. That makes it a template any Swedish company on Microsoft 365 and Azure can copy in a week, and the rest of this piece is the copy.

## What was decided, exactly

Sveriges Radio's Ekot broke the story on 23 October from the internal announcement, and SVT published the operative sentence: "Webbläsargränssnittet för AI-tjänsterna ChatGPT och Claude kommer inte längre att nås från Regeringskansliets tjänstedatorer från och med den 26 oktober." The block is on the web interface. It is a network and endpoint control, not a policy memo.

| Item | Decision |
| --- | --- |
| Blocked | ChatGPT and Claude browser interfaces on Government Offices work computers, from 26 October 2026 |
| Approved for all staff | Microsoft Copilot, after an internal security review that, per Ekot's reporting, led to tightened settings |
| Exception | OpenAI and Anthropic models remain reachable for some staff through Intric, the platform procured on 8 July 2026 |
| Stated reason | Digitalisation director Mikael Wettercrantz: AI services used without an agreement can have consequences for information security |
| Precedent | Deepseek was restricted earlier; Skatteverket and Statens institutionsstyrelse already block; Kronofogden is preparing a decision |

Ekot's reporting adds the trigger. Staff had pasted documents and information that carry secrecy requirements into public AI services, and the employer had no control over what happened on the other side. Wettercrantz's internal note, as quoted across Swedish press, cited several examples of problematic use, covering both the sensitivity of what went in and how the AI-generated answers were then handled. The second half matters as much as the first: an answer drafted in a consumer tool and pasted into a government document has no audit trail.

## The decision is about processors, not models

The same GPT and Claude models that are blocked in the browser are allowed inside Intric. The Government Offices did not conclude that OpenAI's or Anthropic's models are unsafe. It concluded that a consumer web login has no data processing agreement with the state, no sekretess guarantees, and no logging the agency controls. Intric's contract, by its own announcement, runs on Swedish infrastructure with no dependency on foreign cloud services, and the Techtidningen interview with its CEO describes a model-agnostic platform that exposes ChatGPT, Claude, Gemini and open-weight models, deployable in Swedish cloud, on-premises or air-gapped.

So there are three tiers, and the tiers are defined by who the processor is:

- **Tier 0, no agreement.** Consumer ChatGPT and Claude logins. The vendor is a controller on its own terms, the organisation has no DPA, no retention control and no audit. Blocked.
- **Tier 1, Microsoft as processor.** Copilot under the Microsoft Product Terms and Data Protection Addendum, inside the existing Microsoft 365 contract the Government Offices already holds. Approved for everyone.
- **Tier 2, your own platform.** Intric on Swedish infrastructure, with the agency governing which models, which data and which users. Approved for the staff who need frontier models beyond Copilot.

A private Swedish company does not need Intric to reproduce this. Tier 1 is already in your Microsoft 365 licence. Tier 2 is Azure OpenAI in the EU Data Zone or Sweden Central, fronted by whatever chat interface you prefer. Tier 0 is a block list. The rest of this piece walks each tier.

## Why Copilot passed the review, and the three caveats

Microsoft's enterprise data protection (EDP) page is the document the Government Offices' reviewers will have read. It states that Microsoft Copilot and Microsoft Copilot Chat are covered by the Products and Services DPA and the Product Terms with Microsoft as data processor, that prompts and responses get the same contractual commitments as Exchange mail and SharePoint files, that they are not used to train foundation models, and that the EU Data Boundary, sensitivity labels, retention policies and Purview audit all apply. For a public body whose Microsoft 365 contract already went through procurement and a security review, that is a short path to yes.

The same page carries three footnotes that your review should not skip.

### 1. Web search queries sit outside the DPA

When web grounding is on, Copilot derives a few-word search query from the prompt and sends it to Bing with user and tenant identifiers removed. Microsoft handles those queries as a controller under the Services Agreement, and its own documentation says the DPA, HIPAA and the EU Data Boundary do not apply to them. The Product Terms add commitments (no advertising profiles, no training, treated as confidential), but the legal basis is different. The admin control is the Allow web search in Copilot policy in Cloud Policy service for Microsoft 365, with a third option that disables web search in Work mode while leaving it on in Web mode and Copilot Chat. Swedish press reports that the Government Offices tightened Copilot settings after its review, without saying which. If you only change one Copilot setting after reading this, make it this one.

### 2. Anthropic models inside Copilot are off by default in the EU, for a reason

Copilot can route to Claude models with Anthropic acting as a Microsoft subprocessor. Microsoft's subprocessor page states that these models are currently excluded from the EU Data Boundary and from in-country processing commitments, and that tenants in the EU, EFTA and the UK have them disabled by default. The toggle lives under Copilot, Settings, AI providers operating as Microsoft subprocessors, and it can be scoped to Entra security groups. A separate class, Anthropic models with Data Retention, is off everywhere and moves you onto Anthropic's own commercial terms and DPA, with Anthropic retaining most inputs and outputs for up to 30 days. The Government Offices has not stated publicly which way this toggle is set. Yours should be a deliberate decision recorded in the review, not an inherited default.

### 3. Agents carry their own terms

The EDP page says it plainly: when using agents in Copilot, check the agent's privacy statement and terms of use. An approved Copilot tenant with an open agent store is a side door back to Tier 0. Govern the store in the Microsoft 365 admin center before you announce Copilot as the approved tool.

## The playbook for a Swedish company on Azure

### Step 1: Block Tier 0 at the network and the endpoint

The Government Offices blocked the browser interface, which is the right scope. Block the consumer web apps, not the API hostnames your own Tier 2 platform will call. Three controls cover a Microsoft shop:

- **Microsoft Defender for Cloud Apps.** Its cloud app catalog tags generative AI apps as a category. Mark ChatGPT, Claude and the rest as unsanctioned, and Defender for Endpoint enforces the block on managed devices. Run the discovery report first: the list of AI apps your staff already use is the business case for Steps 2 and 3.
- **Edge and browser policy via Intune.** A URLBlocklist policy for chatgpt.com and claude.ai on managed browsers catches the case where a device is off the corporate network. Add the desktop apps to your application control list; a blocked website and an allowed native app is a common gap.
- **Entra Internet Access web content filtering**, if you run Global Secure Access, gives the same block as a category rule with Conditional Access scoping, so a pilot group can keep access while the rest of the organisation is cut over.

Write the exception process into the same change. The Government Offices kept a controlled path for staff who need frontier models. If you block without one, the next thing you discover is personal phones.

### Step 2: Configure Tier 1 before you announce it

Copilot Chat with enterprise data protection is included with Microsoft 365 business and enterprise licences, so "use Copilot instead" costs nothing extra per seat for the chat use case. The paid Copilot licence adds Microsoft Graph grounding across mail, files and Teams. Either way, five settings make it a defensible approved tool:

1. Set the **Allow web search in Copilot** policy explicitly. For a sekretess-heavy organisation, the Work-mode-off option is the sensible default.
2. Decide the **AI providers operating as Microsoft subprocessors** toggle for Anthropic and record why. Leave Data Retention models off.
3. Apply **sensitivity labels** with Purview so Copilot honours them. A label-aware Copilot will refuse to summarise what the user is not permitted to read.
4. Turn on **Purview Data Security Posture Management for AI**. It shows the prompt, the response and the derived web query side by side, which is the audit trail the Government Offices lacked for consumer tools.
5. Restrict the **agent store** and the Copilot Studio publishing rights to a named group.

### Step 3: Build Tier 2 on Azure OpenAI in the EU Data Zone

Intric's selling point is control: your data, your governance, your choice of model, Swedish infrastructure. On Azure, the equivalent for a private company is a Foundry project with Azure OpenAI deployments on Data Zone Standard in the EU zone, or Standard deployments in Sweden Central if you need a single named region. Prompts and completions stay under the Microsoft DPA, are not used for training, and are processed within the zone you chose. Front it with a thin chat app on Azure Container Apps behind Entra ID, log every prompt and response to Application Insights, and you have the audit trail and the retention policy the Government Offices built its exception on.

One caveat, and it mirrors the Copilot one. Claude models on Foundry are Anthropic-hosted and outside the EU Data Zone; we covered the detail in [Claude Fable 5.1 in Foundry: cache math and the EU caveats](/en/blog/claude-fable-5-1-foundry-cache-math-eu-caveats). If your Tier 2 must stay inside the EU zone, it is an OpenAI-model platform today, and the Government Offices' route to Claude (a Swedish-hosted broker) is not something Foundry gives you.

The token cost of Tier 2 is small next to the licensing conversation it usually gets confused with. Take an illustrative 500-person organisation where every person sends 20 prompts a day, 22 working days a month, averaging 2,000 input and 500 output tokens per prompt. That is 220,000 prompts, 440 million input tokens and 110 million output tokens a month. Using the Foundry list prices from our earlier posts on [GPT-6 Luna](/en/blog/claude-haiku-5-5-foundry-0-10-tier-100k-cliff-vs-gpt-6-luna) and [GPT-6.1 Terra](/en/blog/gpt-6-1-terra-foundry-1-5-eu-zone-warning-it-ignores):

| Model, EU Data Zone | Input per 1M | Output per 1M | Monthly tokens, 500 staff | Per user per month |
| --- | --- | --- | --- | --- |
| GPT-6 Luna | $0.12 | $0.60 | $119 | $0.24 |
| GPT-6.1 Terra | $1.20 | $6.00 | $1,188 | $2.38 |

Add a Container Apps environment, Application Insights and an API Management instance to broker keys and rate limits, and the platform runs for a few hundred dollars a month on top. The point is not that Tier 2 is cheap. It is that the cost of giving staff a governed frontier model is a rounding error against the cost of the incident the Government Offices was reacting to.

### Step 4: Write the rule as a table people can follow

A block without a rule produces workarounds. The rule should map your existing information classification to the tiers, in one table, on the intranet page the block redirects to:

| Information class | Consumer AI (Tier 0) | Copilot with EDP (Tier 1) | Own Azure platform (Tier 2) |
| --- | --- | --- | --- |
| Public | Blocked on managed devices regardless | Yes | Yes |
| Internal | No | Yes, web search per policy | Yes |
| Confidential, personal data | No | Yes, Work mode, web search off, labels enforced | Yes, EU Data Zone, logging on |
| Sekretess, regulated, trade secret | No | Only if the review signed it off per label | Yes, named region, modified abuse monitoring approved |

### Step 5: Put the review on a calendar

The Intric contract is twelve months with a twelve-month option. Microsoft's EDP and subprocessor pages both carry a 180-day update cycle and the subprocessor page was revised on 23 October 2026, three days before the block took effect. The settings that passed a review in October are not guaranteed to be the settings in April. Re-read both pages, re-run the Defender for Cloud Apps discovery report and re-check the Anthropic toggle every six months, and when a new model class arrives with new terms.

## The Swedish and EU angle

**For public bodies, this is now the reference decision.** Skatteverket and Statens institutionsstyrelse blocked first, the Government Offices followed, and Kronofogden is preparing its own decision. A municipality or region that still allows consumer AI logins on managed devices will find it hard to explain why its risk assessment differs from Regeringskansliet's. The legal driver is not the AI Act, which does not ban these tools; it is the Offentlighets- och sekretesslagen and the duty to keep sekretess-classified information inside controlled processing. A consumer login has no such control.

**For private companies, the parallel is GDPR Article 28.** Personal data pasted into a service with no processor agreement is processing without the contract the regulation requires. Copilot under the Microsoft DPA and Azure OpenAI under the same DPA both satisfy that; a consumer ChatGPT account does not. Integritetsskyddsmyndigheten does not need a new rule to act on it.

**The sovereignty tier is a procurement signal.** The Government Offices chose a Swedish vendor on Swedish infrastructure for Tier 2, not a hyperscaler region. For suppliers to the public sector, expect upphandlingar to ask for the same: model choice, EU or Swedish hosting, on-premises option. For a private company, Azure's EU Data Zone and Sweden Central regions are the pragmatic answer for OpenAI models, and the Claude exclusion from the EU Data Boundary in both Copilot and Foundry is the gap to document in your DPIA rather than discover in an audit.

**The Copilot review is not a one-off.** Microsoft's own Copilot Credits change on 2 November, which we covered in [Copilot Credits go default-on: your 2 November checklist](/en/blog/copilot-credits-go-default-on-your-2-november-checklist), shows how quickly the approved tool's defaults move. An approved tool with unreviewed defaults is an approved brand, not an approved configuration.

## Conclusion

Regeringskansliet did not ban AI. It sorted AI tools by contract, blocked the tier with no agreement, approved the tier its existing Microsoft contract already covers, and bought a governed platform for the staff who need more. Block the consumer web apps with Defender for Cloud Apps and Intune. Configure Copilot's web search, Anthropic subprocessor and agent-store settings before you call it approved. Stand up Azure OpenAI in the EU Data Zone for everyone who needs a frontier model under your own logging. Publish the classification table. Put the six-month review in the calendar. The whole programme fits in a sprint, and the Government Offices just gave you the precedent to get it funded.

## Sources

- [SVT Nyheter: AI-tjänster spärras på Regeringskansliet (23 October 2026)](https://www.svt.se/nyheter/inrikes/ai-tjanster-sparras-pa-regeringskansliet)
- [Sveriges Radio Ekot: Chat GPT och Claude förbjuder Regeringskansliet](https://www.sverigesradio.se/artikel/chat-gpt-och-claude-forbjuder-regeringskansliet)
- [Computer Sweden: Regeringskansliet spärrar Chat GPT och Claude](https://computersweden.se/article/4239687/regeringskansliet-sparrar-chat-gpt-och-claude.html)
- [Techtidningen: Regeringskansliet blockerar AI-tjänster](https://techtidningen.se/regeringskansliet-blockerar-ai-tjanster/)
- [Intric: Intric wins procurement with the Swedish Government Offices for sovereign AI platform (8 July 2026)](https://intric.ai/en/news/n/intric-wins-procurement-with-the-swedish-government-offices-for-sovereign-ai-plat)
- [Techtidningen: Regeringskansliet väljer svensk AI-leverantör](https://techtidningen.se/regeringskansliet-valjer-svensk-ai-leverantor-nagot-vi-brinner-for/)
- [Microsoft Learn: Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat](https://learn.microsoft.com/en-us/microsoft-365/copilot/enterprise-data-protection)
- [Microsoft Learn: Data, privacy, and security for web search in Microsoft Copilot](https://learn.microsoft.com/en-us/microsoft-365/copilot/manage-public-web-access)
- [Microsoft Learn: Anthropic models in Microsoft Online Services](https://learn.microsoft.com/en-us/microsoft-365/copilot/connect-to-ai-subprocessor)

---

Technspire AB builds AI agents, Azure OpenAI solutions, and production web platforms for Swedish and EU enterprises. Book a call: https://calendly.com/technspire · hello@technspire.com · More articles: https://technspire.com/en/blog · Site overview for agents: https://technspire.com/llms.txt