Claude in the EU – Residency, Contracts, and Governance
Claude can keep data in the EU, but only via AWS Bedrock or Google Vertex AI in an EU region – Anthropic's own API has no EU residency. The contract stack includes a DPA with Standard Contractual Clauses, SOC 2, ISO 27001, and ISO 42001, no training on customer data without consent, and zero data retention on request for Enterprise.
Claude has become a common choice in Swedish enterprise solutions, not least because Anthropic has positioned itself around enterprise terms and security. But when the legal team asks where the data ends up, the answer depends entirely on which path in you choose. There are three – and only two of them can keep data in the EU.
Three paths in – only two keep data in the EU
Anthropic’s own API is the technically simplest path, but it runs on US/global infrastructure and offers no EU residency. If you want processing to happen in the EU, the path runs through a hyperscaler: AWS Bedrock or Google Vertex AI in an EU region. Then the hyperscaler becomes your data processor and Anthropic the subprocessor – a chain of responsibility that should be mirrored in your contracts and your subprocessor register.
| Path | EU residency | Chain of responsibility |
|---|---|---|
| Anthropic's API (first-party) | No – US/global | Anthropic is the data processor |
| AWS Bedrock in an EU region | Yes – in the chosen region | AWS is data processor, Anthropic is subprocessor |
| Google Vertex AI in an EU region | Yes – in the chosen region | Google is data processor, Anthropic is subprocessor |
Note that the region choice is your responsibility: Bedrock and Vertex exist in many regions worldwide, and it’s the configuration – not the product name – that decides where data is processed. Also check the model lineup: not every Claude model is available in every EU region, and the matrix changes continuously. An early check that your specific models are available in your specific region saves you from discovering the gap after signing.
The contract stack: what Anthropic commits to
For buyers with compliance requirements, Anthropic’s contract package is one of the stronger starting points among the AI labs:
- A DPA with Standard Contractual Clauses (SCC) is included in the commercial terms.
- Certifications: SOC 2, ISO 27001, and ISO 42001 – the last one specifically covers AI governance.
- No training on customer data without explicit consent.
- Zero data retention (ZDR) can be requested for Enterprise – prompts and responses then aren’t stored after the run.
- At least 60 days’ notice before a model is deprecated.
Remember that this stack applies to the relationship with Anthropic. If you run via Bedrock or Vertex, the hyperscaler’s terms need to be reviewed in parallel – it’s two contract relationships, not one.
In practice, that means your subprocessor register needs to reflect both links: the hyperscaler with region and service, and Anthropic as subprocessor with model version. Also write in how you’ll be notified when the chain changes – that notice is what should trigger your reassessment.
The CLOUD Act doesn’t go away – but it can be met
Anthropic is a US company, and so are the hyperscalers. CLOUD Act exposure therefore follows along no matter which region you choose. EDPB Recommendations 01/2020 point to technical mitigation as the way forward, and in practice that means customer-controlled encryption keys: AWS KMS on Bedrock or customer-managed keys (CMEK) on Vertex AI.
The logic is simple: data at rest that the vendor can’t read, it also can’t hand over in readable form. During the inference itself, however, data is processed in plaintext – the residual risk belongs in your DPIA as a documented and mitigated item, not swept under the rug.
A concrete path choice
A financial-sector company wants to use Claude to analyze contract documents. The compliance requirement: personal data may not be processed outside the EU. The path choice lands on Bedrock in an EU region, ZDR where possible, customer keys via KMS, and a DPIA that describes the CLOUD Act exposure as a residual risk with mitigations.
Anthropic’s own API was ruled out on the residency requirement – not on model quality. That’s a typical outcome: the path choice is more often driven by where data is processed than by which model is best at the task.
What needs ongoing monitoring
Two things change more often than contracts get rewritten: retention per feature and the model lineup per EU region. Verify both in writing before signing, and redo the check at every model upgrade – a new model version can have different terms than the one you approved. Date the statements and save them in the decision record; that’s the documentation that holds up when an auditor asks.
Thinking about Claude in your own solution? At Weapp we build AI solutions where residency and contract questions are part of the architecture from the start – get in touch and we’ll think through the right path in for your environment.
Frequently asked questions
Can we use Anthropic's own API and still comply with GDPR?
Yes, a third-country transfer isn't prohibited – but it requires a transfer mechanism like Standard Contractual Clauses, your own impact assessment, and documentation. If you want data to stay in the EU, though, that isn't enough: then Bedrock or Vertex AI in an EU region is the path.
Does Anthropic train on our data?
Not without explicit consent – that's part of the commercial terms. Still, confirm in writing that the statement applies to your contract type and your features, and save the answer in the procurement record.
What is zero data retention and who can get it?
Zero data retention (ZDR) means prompts and responses aren't stored after the run. At Anthropic, it can be requested for Enterprise agreements. Check which features are covered – retention can differ between features even within the same contract.
What happens when a Claude model is deprecated?
Anthropic gives at least 60 days' notice before deprecation. Plan revalidation into that window: a new model version can have different retention and region terms than the one you approved, so the compliance check needs to be redone, not just the swap in the code.
Do customer-managed keys fully protect against the CLOUD Act?
No. Customer-controlled keys are a technical mitigation in line with EDPB Recommendations 01/2020 and protect data at rest. During the inference itself, data is processed in plaintext, so a residual risk always exists – it should be documented in your DPIA, not hidden.