A few months ago the head of IT at a financial institution in Northern Europe decided his organisation should have Claude. He did what most IT leaders would do: looked at Claude Enterprise, read the security pages, and took it to the auditors.
The auditors said no.
Not because the model was unsafe. Because a US-hosted subscription with standard terms could not be contracted for in the way the rules require. No audit rights. No say over subcontractors. No exit assistance. Nothing that survives the supplier being sold or shutting down. For a regulated firm in the EU or the UK, that is not a product it can sign for.
This is the part of the AI story that rarely gets told. Everyone argues about which model is best. In a regulated company the model is the easy part.
Regulated firms in the EU and UK can use Claude, but a public subscription on standard terms rarely passes review. DORA in the EU and the FCA and PRA rules in the UK require audit rights, control over subcontractors and exit assistance that a global SaaS product cannot offer a single mid-sized firm. The deployment that passed ran Claude inside the firm's own cloud account, through Amazon Bedrock, Microsoft Azure AI Foundry or Google Vertex AI, built and operated by a partner that signs the firm's addendum.
Why did the auditors say no to Claude Enterprise?
Since 17 January 2025 every financial entity in the EU falls under DORA, the Digital Operational Resilience Act. The UK has had its own version for longer, through the FCA and PRA outsourcing rules and, since 2025, direct oversight of the big cloud providers. The details differ, the logic is the same: the firm stays accountable for every technology supplier it depends on, and the contract has to prove it.
Does DORA apply to AI assistants?
The rules are not about AI. They are about any ICT service, and an AI assistant is several: the desktop application, the platform behind it, the model, the cloud it runs on, every connector into email and documents.
For each one the firm needs to know:
- where data is processed
- who the subcontractors are
- how it gets its data back if the supplier fails
- how it leaves
For anything that supports an important function it also needs audit rights, participation in its own resilience tests, and up to twelve months of exit assistance.
Why can't a public AI subscription meet these rules?
A global SaaS product cannot offer that to a 200-person firm. It is not malice, it is economics. So the firm is left with a product it is allowed to use in theory and cannot contract for in practice.
Sweden's Government Offices reached the same conclusion on 26 October 2026 and blocked ChatGPT and Claude outright, for lack of an agreement with the providers.
What about Claude in Microsoft 365 Copilot?
Even the most trusted vendors cut corners under the pressure to ship AI fast. Microsoft 365 Copilot now offers Claude models, and Microsoft's own documentation says that when they are switched on, data is stored in Europe but processed by Anthropic in the United States until at least the end of 2026.
Microsoft keeps the option off by default for European tenants and leaves it to administrators to decide. That is the compliant product.
What did the IT lead do instead?
He went looking for a way to make Claude private, secure and contractable. He found it in a place most people do not look: the cloud he already had.
How do you run Claude inside your own cloud account?
Anthropic's models are available as a service inside Amazon Bedrock, Microsoft Azure AI Foundry and Google Vertex AI. Run that way, the model lives in the firm's own cloud account under the cloud contract it already has, which already carries the regulatory terms. The model vendor's own terms, subcontractors and US infrastructure drop out of the chain.
And Anthropic's desktop, Cowork and coding applications can be pointed at those endpoints instead of Anthropic's own service. The firm gets the same tools its people would use on the public product, with every layer now inside its own account.
Why not build it in-house?
That solved the regulatory problem on paper. It created a new one. A private deployment is a platform: identity, device management, a configuration server, gateways, connectors, monitoring, backups, updates, someone on call. As the head of a small IT team, he had no wish to design, build and administer that himself.
So he went looking for a partner who would build it inside his account, run it, and sign his firm's addendum. That is how we met.
What does a regulated buyer actually ask an AI supplier?
What followed was a full regulated procurement:
- an NDA
- a security questionnaire of more than forty questions
- the firm's own DORA appendix, drafted by its lawyers
- an independent penetration test summary
- the business continuity plan, with recovery and failover test results
- a responsibility matrix
- a second round of questions after their legal team had read our standard agreement
Here is what the second round asked. Notice what is missing.
- Whose name is the cloud account in, and who holds the root credentials?
- Who are your subcontractors, and how do we record them in our register?
- What happens when the component users depend on to start the application fails?
- Who manages backups, and can we restore without you?
- The questionnaire says 99.9% availability. The contract says 99.5%. Which is it?
- If you cease to operate, can we keep running the software beyond twelve months?
- We want the right to terminate without penalty if you are sold.
- Which models retain our prompts at the cloud provider, and can that be switched off?
Not one question about whether the model is any good. Every question is about control, continuity and the contract. They were not evaluating an AI product. They were evaluating whether they could be accountable for one.
If you run a company with customer data, an auditor, or an ISO 27001 certificate, you will recognise the list. It is not a financial-services list. It is the list every regulated or security-conscious organisation ends up with once AI moves from an experiment to a tool people use all day.
What does a compliant Claude deployment look like?
The deployment that passed had five properties, and none of them is exotic.
- Your account, your keys. Software and data run in the firm's own cloud account, in its own name, under its own keys. The supplier works through a revocable role, and every action it takes lands in the firm's own audit log.
- The model comes through the cloud contract. The model is bought through the cloud contract, never through a separate subscription.
- One partner signs the addendum. The partner that deploys and operates the software is the one that signs the firm's addendum, because it can answer for the whole chain.
- The exit is designed up front. The exit is designed before the contract starts: infrastructure code and container images copied into the firm's account, a licence that does not end with the subscription, documentation that lets another operator take over.
- Model choices are on record. Model choices are written down. Some models retain prompts at the cloud provider for abuse monitoring; the default is the ones that do not.
Public subscription vs private deployment
| Public subscription, standard terms | Private deployment in your own cloud | |
|---|---|---|
| Where it runs | US-hosted service | The firm's own cloud account, in its own name, under its own keys |
| How the model is bought | Separate subscription | Through the cloud contract the firm already has |
| Audit | No audit rights | Every supplier action lands in the firm's own audit log |
| Subcontractors | No say | The model vendor's terms, subcontractors and US infrastructure drop out of the chain |
| Exit | No exit assistance | Designed before the contract starts |
| If the supplier is sold or shuts down | Nothing survives | Code, images and a licence that does not end with the subscription stay with the firm |
This is how banks have bought core systems for decades. What is new is applying it to AI assistants, which arrived as consumer products and are still mostly sold that way. The private deployment is the point where the consumer product becomes a system the firm can own.
Is regulation holding AI back? The model was never the hard part
It is tempting to read all this as regulation holding financial services back from AI. I think the opposite. The rules do not say no AI. They say AI on terms you can stand behind.
The firms that understand that are already rolling assistants out across the business, because they stopped waiting for a public tool to become compliant and started buying AI the way they buy everything else that matters.
And it will not stay a financial-services question for long. Healthcare, energy, telecoms, critical infrastructure and anyone with a serious client are getting their own versions of the same list. The contract, the architecture and the exit are about to become everyone's problem.
Next step: if your auditors are asking the same questions, see how a private Claude deployment works or talk to us.
Related reading:
- Claude Cowork on 3P vs Microsoft Copilot Cowork
- Claude Cowork on 3P vs Claude Enterprise: what's the difference
- Claude Cowork for business: secure deployment in your own cloud
Frequently asked questions
Can a regulated firm in the EU or UK use Claude Enterprise?
It is allowed in theory, but hard to contract for in practice. A US-hosted subscription with standard terms gives no audit rights, no say over subcontractors and no exit assistance, which DORA and the UK outsourcing rules require for anything that supports an important function.
Does DORA apply to AI assistants?
Yes. DORA covers any ICT service, and an AI assistant is several: the desktop application, the platform behind it, the model, the cloud it runs on, and every connector into email and documents.
Can Claude run inside our own AWS, Azure or Google Cloud account?
Yes. Anthropic's models are available as a service inside Amazon Bedrock, Microsoft Azure AI Foundry and Google Vertex AI. Anthropic's desktop, Cowork and coding applications can be pointed at those endpoints instead of Anthropic's own service.
Is Claude in Microsoft 365 Copilot processed in the EU?
No. According to Microsoft's documentation, when Claude models are switched on in Microsoft 365 Copilot, data is stored in Europe but processed by Anthropic in the United States until at least the end of 2026. The option is off by default for European tenants.
Do Claude models on cloud platforms keep our prompts?
Some do. For some models the cloud provider retains prompts and outputs for up to 30 days for abuse detection; for others it does not. A compliant deployment writes model choices down and defaults to the models that do not retain prompts.
What does a regulated buyer ask an AI supplier?
Questions about control, continuity and the contract: who owns the cloud account and holds the root credentials, who the subcontractors are, who manages backups, what availability goes into the contract, whether the firm can keep running the software if the supplier ceases to operate, and which models retain prompts.
Is this only a financial-services issue?
No. Any company with customer data, an auditor or an ISO 27001 certificate ends up with the same list. Healthcare, energy, telecoms and critical infrastructure are getting their own versions of it.

