GDPR, Data Residency and FedRAMP for Claude Deployments

1.GDPR roles: the customer controls, Anthropic processes

Privacy regimes such as GDPR hinge on who decides what happens to personal data. For Claude for Work (Team and Enterprise plans), Anthropic's Commercial Terms make the customer the controller of data its users submit. In practice that means the customer decides who joins the team, instructs Anthropic on how submitted data may be used, and can access and export data such as conversation history. Anthropic acts as the processor and handles the data only to provide the Claude service on the customer's behalf.

Two consequences are worth holding onto. First, commercial data is not used to train Anthropic's models unless the customer opts into the Development Partner Program. Second, the documents that define processing are the Privacy Policy, the Data Processing Addendum and Help Center articles, with the Trust Center and Privacy Center as the places to find them. Anthropic describes its approach to GDPR as holistic, assessing privacy laws worldwide alongside customer needs in the context of AI and large language models.

Be precise about the limits of the source material here: the documentation provided for this lesson does not enumerate specific GDPR articles, data-subject request procedures or EU hosting options. What it does establish is the role split and where the governing documents live, which is what an architect needs to route a GDPR review to the right contract.

Sources12

2.Data residency: inference geo and workspace geo

Many regulated workloads also care where data is processed and stored. The Claude API separates these into two independent settings. Inference geo controls where model inference runs, per request, through the inference_geo parameter or a workspace default. Workspace geo controls where data is stored at rest and where endpoint processing such as image transcoding and code execution happens.

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    inference_geo="us",
    messages=[
        {"role": "user", "content": "Summarize the key points of this document."}
    ],
)

for block in response.content:
    if block.type == "text":
        print(block.text)
# Check where inference actually ran
print(f"Inference geo: {response.usage.inference_geo}")

The response's usage object reports the geography inference actually ran in, which gives you an auditable record per call. Administrators can enforce policy instead of trusting every caller to pass the parameter:

Three operational details show up in design questions. The parameter works on Claude 4.6 and later models; sending it to Claude Opus 4.5, Sonnet 4.5, Haiku 4.5 or earlier returns a 400 error. US-only inference on 4.6 and later is priced at 1.1x the standard rate across all token categories, while global routing keeps standard pricing. And organizations that had used the legacy global-routing opt-out were migrated automatically to allowed_inference_geos ["us"] with default_inference_geo "us", so their US-only behaviour continued without code changes.

The API rejects the request, because "global" is not in the workspace's allowed list. That is the point of the workspace control: it holds even when a caller gets the parameter wrong.

Sources3

3.Compliance credentials and FedRAMP workloads

Behind the per-regulation controls sits a set of audited credentials. For its commercial products, Anthropic lists a HIPAA-ready configuration with a BAA available, ISO 27001:2022 for information security management, ISO/IEC 42001:2023 for AI management systems, and SOC 2 Type I and Type II. Copies of the compliance documentation are requested through the Trust Portal.

US federal work adds FedRAMP, a standardized approach to security assessment for federal cloud services. Anthropic's announcement states that Claude models are authorized for FedRAMP High and DoD Impact Level 2 workloads through Google Cloud's Vertex AI. FedRAMP High covers sensitive unclassified data for agencies in areas such as healthcare, law enforcement, finance and emergency services; IL2 covers defense use with non-controlled unclassified information. The documented path is to set up a Google Cloud environment with Assured Workloads configured for FedRAMP High or IL2, access Claude through the Vertex AI Model Garden, and build against the Vertex AI API endpoints. Vertex AI serves the models as managed, serverless APIs, so agencies do not provision infrastructure.

Note what the announcement does not claim: IL5, needed for more sensitive DoD workloads, is described only as future groundwork. And because the authorization runs through a cloud provider, remember the separate rule from HIPAA: Anthropic's BAA does not apply to services purchased through a third-party cloud.

Sources45

Exam traps

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

1. 1.You can move an existing workspace's stored data to a different region by changing its workspace geo.Why is that wrong?

   Workspace geo is fixed when the workspace is created, and "us" is currently the only option; a different geo means a new workspace.

   Covered in Data residency: inference geo and workspace geo

2. 2.Claude is already authorized for DoD IL5 workloads on Vertex AI.Why is that wrong?

   The announced authorizations are FedRAMP High and IL2; IL5 is described only as future groundwork.

   Covered in Compliance credentials and FedRAMP workloads

Practise it for real

Send a US-pinned request and confirm where inference ran

1. 1.Call client.messages.create with a Claude 4.6-or-later model such as claude-opus-5-5 and inference_geo="us".

   Why: inference_geo is the per-request residency control.

   You should see: A normal Messages API response.

2. 2.Print response.usage.inference_geo.

   Why: The usage object records where inference actually ran, which is your per-call evidence.

   You should see: The value "us".

3. 3.Repeat the call against Claude Haiku 4.5 with the same parameter.

   Why: The parameter is unsupported on 4.5 and earlier models.

   You should see: A 400 error.

4. 4.In the Console, set the workspace's allowed_inference_geos to ["us"], then send a request with inference_geo="global".

   Why: Workspace restrictions enforce policy regardless of what callers pass.

   You should see: The API returns an error.