Briefing AI under the company's control
Your AI provider and a government dispute
Whether an AI provider is cut off or changes course is beyond a European company's control, but whether its own AI can move to another model is a choice it makes itself.
- Published
- Updated
- Reading time
- 5 min read
- Editor
- David Lahuta
In short
- In February 2026, by both sides' accounts, the US Department of War and Anthropic, a developer of AI models, could not agree on limits to military use, and the department then excluded the company's products.12
- A company in Europe cannot influence such events. It decides only whether its AI depends on one model or can move between models.
- Providers also retire older models on their own schedule, and a retired model stops answering.78
- What keeps a company free is one company AI between its people and the models, with instructions, data and records kept by the company.
- Under EU law the duties that come with using AI stay with the company, whichever model it uses.9
What happened in the US dispute?
Anthropic wrote on 26 February that the US Department of War would contract only with AI companies accepting "any lawful use", and that it would keep two exceptions, mass domestic surveillance and fully autonomous weapons.1 In the department's account, filed in court, it saw a risk that the company might change its software to enforce its own policy, and on 3 March the Secretary of War designated the company a supply chain risk.2 Anthropic stated on 5 March that it would challenge this in court.3 On 26 March a court in California ordered the measures paused and on 27 August set the designation aside under one law.46 On 25 September the appeals court in Washington, reviewing the exclusion under another law, denied Anthropic's petitions, one of three judges dissenting.5
We take no side. The case matters here for what it shows about depending on one provider.
What does a company depend on when its AI runs on one model?
On more than the model. It depends on the provider's terms of use, on the provider's schedule for retiring models, and on whatever happens to the provider where it is based.
Retiring models is routine. Anthropic and OpenAI both state in their documentation that they regularly retire older models, and that a retired model no longer accepts requests.78 Anthropic gives at least 60 days' notice for publicly released models.7 Two of its models were retired on 15 June 2026, two months after the notice of 14 April.7 Each such step means testing again whatever was built on the old model.
Moving takes time. By the appeals court's account, the department set itself a limit of 180 days to remove one provider's products.5
| What your AI depends on | What can change without you | What keeps you free |
|---|---|---|
| The model | It is retired on the provider's schedule78 | A set of your own real tasks to test any model against |
| The provider's terms | The provider changes what it allows | Terms read before use, and a second provider known |
| The provider's position | A dispute, a government measure or a new owner reaches the provider | The ability to move to another model within weeks |
| Your instructions and settings | They exist only in the provider's account | A copy kept in your own systems |
| The record of what the AI did | It sits with the provider, or nowhere | A record kept by the company |
| Personal data the provider holds | The contract ends | Return or deletion agreed in the contract10 |
How does a company stay free to switch?
By putting one company AI between its people and the models. People and the company's systems work with that AI, and it calls a model in the background. Changing the model then becomes a setting and a test, not a new project.
Three things make this work.
- Instructions, data and records stay with the company. The instructions that tell the AI how to do your work, the documents it draws on and the record of every question and action live in your systems.
- Permissions stay with the company. Each user works only with his own permissions, whichever model answers.
- A second model is tried before it is needed. A small set of real tasks, run against another model from time to time, shows whether a move would work and what it would cost.
Five questions show where you stand.
- Do you know which model and which provider each use of AI in the company runs on?
- Are the instructions, documents and records of your AI kept in your own systems?
- Could your main uses run on a second model within a few weeks, and has anyone tried?
- Does the contract give you notice before a model is retired or the terms change?
- Does the contract say how your data is returned or deleted when the service ends?
Should you act now or wait?
Act now on what costs little, and wait on the rest. Listing where AI is used and on which model, keeping instructions and records in your own systems and writing notice and exit into the contract take little effort and help whatever happens to your provider.
A tested second model can wait until AI carries work you could not stop for a few weeks. Room for several models is easier to plan in from the start than to add later.
What this means for a company in Europe
In the EU the duties that come with using AI stay with the company that uses it, whichever model runs underneath. The AI Act calls such a company a deployer.9 Since 2 February 2025 it must support the AI literacy of the people who work with AI on its behalf.9 For high-risk uses, such as filtering job applications, its own duties apply from 2 December 2027, among them human oversight by trained people and keeping the system's logs for at least six months.9 Since 2 August 2025 the providers of general-purpose AI models must give the companies that build their models into AI systems information on the model's capabilities and limits.9 When you change the model, those documents change too, so a switch is also a step in your own compliance. Where an AI provider processes personal data for you, the GDPR requires the contract to say that it deletes or returns all of that data at the end of the service, as you choose.10
Next step
If you would like to see how closely your company's AI use is tied to one provider, we are glad to look at it with you in a short meeting. What you do with the result is entirely your decision.
Sources
- 1. Anthropic, Statement from Dario Amodei on our discussions with the Department of War (26 February 2026). Read on 17 November 2026
- 2. U.S. Department of Justice for the U.S. Department of War, Response in opposition to emergency motion for stay pending review, Anthropic PBC v. U.S. Department of War, D.C. Circuit No. 26-1049 (filed 19 March 2026). Read on 17 November 2026
- 3. Anthropic, Where things stand with the Department of War (5 March 2026). Read on 17 November 2026
- 4. U.S. District Court for the Northern District of California, Anthropic PBC v. U.S. Department of War, No. 26-cv-01996-RFL, preliminary injunction order (26 March 2026). Read on 17 November 2026
- 5. U.S. Court of Appeals for the District of Columbia Circuit, Anthropic PBC v. United States Department of War, No. 26-1049, opinion (decided 25 September 2026). Read on 17 November 2026
- 6. U.S. District Court for the Northern District of California, Anthropic PBC v. U.S. Department of War, No. 26-cv-01996-RFL, order of final relief (27 August 2026). Read on 17 November 2026
- 7. Anthropic, Claude Platform documentation, Model deprecations. Read on 17 November 2026
- 8. OpenAI, API documentation, Deprecations. Read on 17 November 2026
- 9. Regulation (EU) 2024/1689, the AI Act, consolidated text as at 27 July 2026 (EUR-Lex). Read on 17 November 2026
- 10. Regulation (EU) 2016/679, General Data Protection Regulation, Article 28, EUR-Lex. Read on 17 November 2026
Change log
- Published.