Regulated Data on Claude: ZDR, HIPAA, Retention and Data Residency

1.Who is responsible for the data

Regulations such as GDPR assign duties according to role. For commercial services, Anthropic's consumer Privacy Policy does not apply where Anthropic acts as a data processor on behalf of commercial customers. In that case the commercial customer is the controller. On the Claude API, Claude Platform on AWS and Claude in Microsoft Foundry, Anthropic is the processor. On Amazon Bedrock and Google Cloud's Agent Platform, the cloud provider is the processor, and that platform's documentation applies.

For a knowledge worker this has a practical consequence. When your team wants to analyse personal data about EU residents, the question is not 'is Claude GDPR-compliant?' It is whether your organisation, as controller, has an approved arrangement for this processing. Anthropic says its Privacy Policy, Data Processing Addendum and Help Center articles explain how it handles personal data. Those are the documents your privacy or legal team would review. The sources here don't describe the Addendum's terms, so don't guess what it guarantees.

Sources123

2.What gets retained, and who sets it

On the API, Anthropic states a few default commitments. Retained data is never used for model training without express permission. Only what is technically necessary for a feature is kept. Conversation content is not retained by default, with one exception: Covered Models require 30-day retention. Retained data is purged on the shortest practical time to live.

Some data follows its own retention rules. Chat, file and project content in claude.ai follows the retention policy an organisation sets under Organization settings > Data and privacy, unless a user deletes it sooner. The Activity Feed keeps data for 6 years. Data available through the Compliance API follows its own retention model. You don't need to know how to call the Compliance API. You need to know that it exists and that admins can use it for oversight, so your organisation's retention setting may not be the only copy.

Sources2

3.Zero data retention versus HIPAA readiness

For the Claude API, Anthropic offers two arrangements. Under zero data retention (ZDR), prompts and responses are not stored at rest once the response is returned. HIPAA readiness takes a different approach. It doesn't delete data straight away. Instead it adds safeguards such as encryption, access controls and audit logging for protected health information (PHI), under a signed Business Associate Agreement (BAA). They solve different problems, and each has gaps.

Two more limits catch people out. Beta features are generally outside the BAA unless the feature eligibility table lists them. Enabling HIPAA readiness from the Console is supported, but processing PHI through the Console is not covered. Before regulated data goes into any surface, check that this exact surface is covered, not just the organisation.

Sources2

4.Where data is processed and stored

Some requirements are about location rather than retention. Claude's data residency controls have two independent settings, and it is easy to confuse them.

Admins can enforce a policy with allowed_inference_geos, which rejects requests for any geo outside the list. The response also records where inference actually ran, so compliance teams can check it.

{
  "usage": {
    "input_tokens": 25,
    "output_tokens": 150,
    "inference_geo": "us"
  }
}

Sources4

5.Confidentiality and privilege: why the plan matters

Legal work shows why consumer and commercial terms matter. Anthropic presents this as its own perspective, not legal advice. In United States v. Heppner, a defendant used a consumer AI account without their lawyers' involvement, and the court held that use was not privileged work product. The court's reasoning rested on the consumer privacy terms and on the absence of direction from a lawyer.

Later decisions went the other way when confidentiality duties were in place. Morgan v. V2X approved uploading confidential information where the provider is contractually barred from training on, retaining or disclosing it beyond what the service needs. None of these are appellate decisions, and Anthropic encourages consulting your own counsel. The general lesson applies beyond law: confidential material belongs on a plan with commercial confidentiality commitments, used as your organisation directs.

Sources5

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.Being on Claude Enterprise means the claude.ai chat interface runs under zero data retention.Why is that wrong?

   The Teams and Enterprise product interfaces are outside ZDR. The only exception is Claude Code used through Enterprise with ZDR enabled.

   Covered in Zero data retention versus HIPAA readiness

2. 2.An organisation handling PHI needs both HIPAA readiness and ZDR to be safe.Why is that wrong?

   HIPAA readiness is the arrangement for PHI on its own. It applies safeguards rather than deletion, and ZDR is not also required.

   Covered in Zero data retention versus HIPAA readiness

3. 3.Once the BAA is signed, PHI can be used in any Anthropic tool the organisation has, including Claude Code.Why is that wrong?

   Coverage depends on the surface. Claude Code, the Console for processing PHI, consumer plans and most beta features are outside HIPAA readiness.

   Covered in Zero data retention versus HIPAA readiness