What you will be able to do
- Choose between inference geo, workspace geo and workspace geo restrictions to enforce a data residency obligation, and name the evidence each one produces
- Decide which surfaces and features keep a workload inside a zero data retention arrangement, and which quietly step outside it
- Isolate a Covered Model or a regulated workload in its own workspace without breaking the organization's other retention commitments
- Keep PHI and PII out of the fields that do not get the protections that prompts and responses get
- Pick the audit record (Activity Feed, Access Transparency, transparency log) that proves a given access obligation
Key concept
Enforced control versus convention — A control evidences a compliance obligation only when the platform itself enforces or records it: it rejects a request that would break the rule, or it writes down what happened. A setting that a team's code merely passes along is a convention. One mistake or one bypass breaks it, and nothing shows that it held.
1.Data residency: inference geo, workspace geo and who enforces them
Compliance obligations usually arrive as a sentence from an auditor or a contract: data must stay in the US, no prompts may be retained, every human access must be logged. The architect's job is to turn that sentence into a control that the platform enforces and that produces evidence. Residency is the obligation most often given the wrong control, because Claude has two separate geography settings and they govern different things.
Inference geo decides where the model runs. You set it per request with the inference_geo parameter on POST /v1/messages, or as a workspace default. Its values are "global", the default, where inference may run in any available geography, and "us". Workspace geo decides where data is stored at rest and where endpoint processing such as image transcoding and code execution happens. You choose it when you create a workspace and can't change it later, and "us" is currently the only value.
Pinning inference therefore takes the workspace-level inference settings. allowed_inference_geos lists the geos a workspace may use, and the API rejects any request naming a geo outside that list with a 400 error. default_inference_geo is the fallback when a request omits the parameter, and it must be one of the allowed geos. You set both on the Console's Data residency page or in the data_residency object of a Create or Update Workspace call to the Admin API. Set allowed to ["us"] and default to "us", and every key in that workspace runs in the US whether or not a developer remembers the parameter. A developer who writes global gets an error, not a silent reroute.
This difference decides which control an auditor accepts. Hardcoding inference_geo="us" in a shared client library is a convention: a service that imports a different client, or a script that calls the API directly, gets the global default. An egress proxy that only lets traffic reach the API hostname controls which endpoint is called, not where inference runs. Only the workspace restriction makes the wrong outcome impossible. The response also carries proof of where each request actually ran:
{
"usage": {
"input_tokens": 25,
"output_tokens": 150,
"inference_geo": "us"
}
}| Setting | Governs | Scope | What breaks it |
|---|---|---|---|
| inference_geo | Where model inference runs | Per request | A call that omits it falls back to the workspace default |
| allowed_inference_geos | Which geos the workspace may use | Every key in the workspace | Nothing at request time: other geos return a 400 error |
| default_inference_geo | Fallback geo when a request omits inference_geo | Every key in the workspace | A request may override it with any other allowed geo |
| Workspace geo | Storage at rest and endpoint processing | Fixed when the workspace is created | Cannot be changed afterward, and it does not set inference location |
Two practical limits follow. First, inference_geo works only on Claude 4.6 and later models. Sending it to Claude Opus 4.5, Claude Haiku 4.5 or earlier models returns a 400 error, so a residency plan has to account for every model it still calls. Second, US-only inference on supported models costs 1.1 times the standard rate across input, output, cache writes and cache reads. Organizations that had used the legacy global routing opt-out were migrated automatically to allowed_inference_geos: ["us"] and default_inference_geo: "us".
The deployment surface sets who answers for the data in the first place. On the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Claude in Microsoft Foundry, Anthropic is the data processor. On Claude Models on Google Cloud (Google managed) and Amazon Bedrock, the cloud provider is the data processor, and residency, retention and transparency controls come from that platform's own documentation. The runtime choice itself is covered in deployment surface selection. The point here is that the processor determines whose controls you cite as evidence.
Checkpoint 1 of 9· Check yourself
An auditor requires evidence that all model inference for a workspace stays in the US, including calls from scripts that hit the API directly. Which control makes a non-US inference request impossible rather than just unlikely?
Only the workspace restriction is enforced by the platform: a request with no geo falls back to the US default, and a request naming any other geo gets a 400 error. A hardcoded parameter is a convention that a direct call bypasses. Workspace geo covers storage, not inference. An egress proxy controls which endpoint is called, not where inference runs.
“If a request specifies an inference_geo not in this list, the API returns a 400 error.”Source: platform.claude.com
Checkpoint 2 of 9· Exam question
A fintech's compliance team calls the first-party Claude Messages API from several internal services. An auditor requires proof that no team can have inference run outside the US, even by mistake. As the solution architect, which control best enforces this obligation?
Correct answer: A — Restrict the workspace's allowed inference geos to US and set US as its default, so a request naming another geo is rejected with a 400 error.
- A. Correct. A workspace's allowed inference geos and default inference geo are enforced by the API itself. With only US permitted, a request naming another geo is rejected with a 400 error. No service team can override that with a coding mistake, which gives the auditor a server-side guarantee.
- B. Incorrect. A hardcoded parameter in a shared library only covers callers that use the library. A service that calls the API directly, or an edit that slips through review, falls back to the default routing. That is a convention backed by process, so it cannot prove that inference never runs outside the US.
- C. Incorrect. Workspace geo governs where data is stored at rest and where endpoint processing happens. Where model inference runs is a separate setting, inference geo. A US workspace geo therefore does not stop a request from being served by inference outside the US.
- D. Incorrect. An egress allow-list only decides which hostname the services can reach. Every request to that hostname is still routed according to its inference geo settings, so the proxy gives no assurance about which geography runs the model and cannot serve as evidence of US-only inference.
2.Zero data retention: which surfaces it actually covers
Under a zero data retention (ZDR) arrangement, Anthropic does not store customer prompts or responses at rest once the API response has been returned. You request ZDR from the sales team, and it applies to one organization only. Each new organization needs its own enablement by the account team, and enabling it does not extend to other organizations under the same account. Most ZDR failures in practice are not configuration bugs. They happen because a workload moved to a surface the arrangement never covered.
The arrangement has edges you need to know. Here is where each common surface stands:
- Claude Messages and Token Counting APIs: covered, for eligible features.
- Claude Code with API keys from a Commercial organization: covered. If Claude Code metrics logging is turned on, productivity data such as usage statistics is exempt from ZDR and may be retained.
- Claude consumer plans (Free, Pro, Max, Ultra), including Claude Code on those plans: not covered.
- Claude Teams and Claude Enterprise product interfaces: not ZDR-eligible. The one exception is Claude Code on Claude Enterprise, which has its own separately enabled ZDR offering for qualified accounts.
- Claude Managed Agents: not covered. Sessions are stateful and transcripts persist until you delete them. Self-hosted Managed Agents are the exception.
- Third-party tools, MCP servers and other integrations: not covered. Review each service's own data handling.
Take a firm whose contract promises that no developer prompts are retained, and whose developers use Claude Code with keys from the ZDR Commercial organization. The surface is right. The open question is the metrics logging the platform team just turned on for adoption dashboards. Either turn it off, or record the productivity data in the compliance file as an explicit exception to the ZDR terms. Moving developers to consumer accounts or Teams seats would take them out of ZDR entirely, since neither is covered.
Two more conditions apply to every arrangement. Even under ZDR, Anthropic may retain data where the law requires it or where automated trust and safety systems flag it. A flagged chat or session may have its inputs and outputs kept for up to 2 years. ZDR organizations also cannot call the Claude API directly from a browser: cross-origin requests are rejected, so web applications must call it from their own back end.
Checkpoint 3 of 9· Match them up
A firm's contract promises zero data retention for its Claude usage. Match each surface to where it stands under the organization's ZDR arrangement.
Tap a term, then the definition that fits it.
ZDR depends on the surface, not only on a setting. Commercial-organization API keys are covered, with metrics logging as a documented exception. Consumer plans and Teams or Enterprise interfaces are outside the arrangement, and Anthropic-hosted Managed Agents keep transcripts until you delete them.
“If metrics logging is enabled in Claude Code, productivity data such as usage statistics is exempted from ZDR and may be retained.”Source: platform.claude.com
Checkpoint 4 of 9· Exam question
A software firm holds a ZDR arrangement for its Claude API organization, and its developers use Claude Code with pay-as-you-go API keys from that Commercial organization. The platform team has just enabled Claude Code metrics logging for adoption dashboards. The contract promises that no developer prompts are retained. What should the architect do to keep the obligation evidenced?
Correct answer: A — Keep developers on Commercial-organization API keys, and either turn metrics logging off or document productivity data as an exception to the ZDR terms.
- A. Correct. ZDR applies to Claude Code used with API keys from a Commercial organization, but productivity data from metrics logging, such as usage statistics, is carved out and may be retained. The firm must either switch that logging off or record the carve-out in its compliance documentation.
- B. Incorrect. Personal consumer accounts sit outside a Commercial organization's ZDR arrangement and follow consumer retention terms. Documenting otherwise would misstate coverage and weaken the no-retention promise rather than support it.
- C. Incorrect. Documented ZDR coverage for Claude Code excludes productivity data from metrics logging. Recording that telemetry as covered, with no exception filed, would give the compliance record an inaccurate statement.
- D. Incorrect. Claude Teams and Claude Enterprise product interfaces are not covered by the API ZDR arrangement. Moving to Teams seats would drop the coverage the firm relies on, and documenting otherwise would be inaccurate.
Sources2
3.Stepping outside ZDR: stateful features and Covered Models
A ZDR organization can still use features that store data. Eligibility is set per feature, and features that are not eligible are usually stateful: they store jobs, files or container state, so they cannot be zero-retention by design.
ZDR and HIPAA readiness differ in a way that matters here. Under HIPAA readiness, the API blocks requests that include an ineligible feature and returns a 400 error, apart from a few client-side tools noted as accepted. Under ZDR nothing is blocked, so the control has to come from you: an allow-list of features in your own code, or a review gate before anyone adopts the Batch API, the Files API or code execution.
| Feature | ZDR eligible | HIPAA eligible | What is retained |
|---|---|---|---|
| Messages API | Yes | Yes | Prompts and responses are not stored under ZDR |
| Batch processing | No | No | 29-day retention; async storage required |
| Files API | No | No | Files kept until deleted or until their configured expiration |
| Code execution tool | No | No | Container data retained up to 30 days |
| Structured outputs | Yes (qualified) | Yes | Only the JSON schema, cached for up to 24 hours since last use |
| Web fetch tool | Yes | No | Fetched content is returned in the response; publishers may keep request data |
| Claude Managed Agents | No (self-hosted excepted) | No | Session transcripts persist until you delete them |
Model choice can also break ZDR. Claude Fable 5.2, Mythos 5.2, Fable 5.1, Mythos 5.1, Fable 5 and Mythos 5 are designated Covered Models, and they require 30-day data retention. They are not available under ZDR, whatever the feature eligibility, unless Anthropic expressly authorizes it. An organization whose retention configuration does not meet the requirement gets a 400 error:
{
"type": "error",
"error": {
"type": "invalid_request_error",
"message": "In order to access this model, your organization or workspace must have data retention enabled."
}
}The way out is a workspace boundary, not an organization-wide change. A ZDR organization can enable 30-day retention for a single workspace, and the other workspaces keep zero data retention. The order of decisions is: check the model's retention requirement first, then give the trial its own dedicated workspace with retention enabled, so every other workload stays under ZDR. The Default Workspace is the wrong place for this, because it always follows the organization's setting. Routing the trial through the Batch API doesn't help either: batch jobs are kept 29 days and are not ZDR-eligible on any model.
Checkpoint 5 of 9· Check yourself
Your organization runs under ZDR, and a research team wants a two-week trial of Claude Fable 5.2 without weakening retention for the production workloads. Which setup lets the trial run and keeps everything else zero-retention?
Covered Models need 30-day retention, and you can turn retention on for a single workspace while the others keep ZDR. Retention for the whole organization weakens every workload. The Default Workspace always follows the organization's setting. Batch jobs are kept for 29 days and are not ZDR-eligible, so using the Batch API does not solve the problem.
“Workspaces without an override continue to follow the organization default.”Source: platform.claude.com
Checkpoint 6 of 9· Exam question
A contract-review team under a ZDR arrangement wants to try a newer top-tier model that is a Covered Model requiring 30-day data retention. Several other workloads in the organization must stay under ZDR. What is the correct sequence of decisions?
Correct answer: A — Check the model's retention requirement first; if it needs 30-day retention, enable it for one dedicated workspace and keep other workspaces on ZDR.
- A. Correct. Covered Models are not available under ZDR unless retention is enabled, and an organization can enable 30-day retention for a specific workspace only. Checking the requirement first and isolating it keeps the other workloads inside ZDR.
- B. Incorrect. An organization-wide change would pull every workload off zero retention. The scoped workspace setting exists so only the new workload changes.
- C. Incorrect. The Default Workspace follows the organization's setting, so the trial there would not be cleanly scoped. Fixing retention after an audit finding reverses the correct order.
- D. Incorrect. The Batch API is not ZDR-eligible because batch jobs are stored, so it does not provide coverage for any model tier. It would add retention rather than avoid it.
Sources2
4.Keeping PHI and PII in the fields that are protected
Sensitive-data controls start with knowing which fields an arrangement protects. Under HIPAA readiness, which requires a signed BAA and a HIPAA-enabled organization, PHI is protected where it is expected to appear: in prompts and responses, in attached files such as images and PDFs, and in file names or metadata tied to message content. HIPAA readiness is an alternative to ZDR, not an addition to it. Rather than requiring immediate deletion, it applies encryption, access controls and audit logging across the data's lifecycle. An organization handling PHI uses HIPAA readiness and does not also need ZDR.
Several fields sit outside that protection, and personal data placed in them leaks out of the arrangement:
- JSON schemas for structured outputs and strict tools are compiled into grammars that are cached separately, so property names, enum values, const values and patterns must never contain PHI.
- User profiles may be used by HIPAA-enabled organizations, but the user profiles API is not a HIPAA-eligible service, so profile data must not contain PHI.
- Workspace names, user details, billing data and support tickets are not expected to contain PHI under the BAA.
- Telemetry can carry content. Claude Cowork and Claude Code (local mode) can stream OpenTelemetry events to a collector your organization runs. In current Claude Desktop versions those events carry metadata only unless the
otlpContentCapturepolicy turns content capture on, while some older versions send prompt text and tool inputs by default untilotlpContentCaptureis set to[]. Even metadata-only exports include user emails, folder paths and API error messages, and an error message can repeat part of the request. Where those fields could hold PHI, filter or redact them at your collector.
In practice, redaction usually belongs where data leaves the protected path, not inside the model call. A patient ID belongs in message content and never in an enum. A monitoring pipeline gets a redaction step at the collector before its storage, because OpenTelemetry spans from the SDK go straight from your environment to your exporter and are then held under that service's retention, outside Anthropic. Where in the request flow to put content filters is a separate question, covered in guardrail placement.
Checkpoint 7 of 9· Check yourself
A HIPAA-enabled organization's triage service uses structured outputs. To stop the model from inventing identifiers, its JSON schema lists each patient's medical record number as an enum value. A compliance review flags this. Which change keeps the PHI under HIPAA protections?
JSON schemas are compiled into grammars that are cached apart from message content and don't get PHI protections, so patient data belongs only in message content. HIPAA readiness replaces ZDR rather than adding to it. The user profiles API is not HIPAA-eligible. Strict tools use the same cached grammar pipeline as structured outputs.
“Patient-specific information should appear only in message content, where it is protected under HIPAA safeguards.”Source: platform.claude.com
5.Audit evidence and tenant isolation with workspaces
Audit obligations come in two kinds, and each needs a different record. Activity by your own organization, meaning API calls, admin actions and Compliance API reads, appears as standard event types on the Compliance API Activity Feed, which keeps data for 6 years. Access to your data by Anthropic staff is recorded by Access Transparency, which emits an anthropic_access event each time a person at Anthropic views covered content under a published reason code. Automated processing such as model serving, safety classifiers and abuse detection does not generate those events. Events show the accessing department and the reason code, but not the individual employee's identity.
curl --fail-with-body -sS -G \
"https://api.anthropic.com/v1/compliance/activities" \
--data-urlencode "activity_types[]=anthropic_access" \
--data-urlencode "limit=50" \
-H "x-api-key: $ANTHROPIC_COMPLIANCE_ACCESS_KEY" \
-H "anthropic-version: 2026-10-21"The details of how it operates decide whether the evidence exists when the auditor asks:
- Access Transparency is enabled at the organization level and covers every workspace. You cannot enroll a single workspace.
- It covers the same APIs and features ZDR covers. The Files API, the Batch API, Anthropic-hosted stateful applications and claude.ai Enterprise seats are outside it.
- Events are recorded even while the Compliance API is off, but you can read them only once it is on. An Admin API key created before you turned on the Compliance API cannot read the feed.
- After an organization's transparency log is created, each event is also committed to an append-only, signed log that you can verify independently of Anthropic's serving systems.
Tenant isolation follows the same boundary that carried the earlier controls: the workspace. Workspaces separate projects, environments or teams under one organization's billing and administration. An API key scoped to a single workspace can only reach that workspace's resources. Each workspace carries its own inference geo restrictions and can carry its own retention override. So a regulated tenant gets its own workspace and scoped keys. Its residency rule is set by that workspace's allowed geos, and its retention by the workspace's setting. Each Access Transparency event carries a workspace_id that ties the access back to that tenant. The Default Workspace cannot be renamed, archived or deleted, and it follows organization-wide settings, so keep it for workloads that need no isolation. Who may hold which keys is part of authentication and authorization posture.
Checkpoint 8 of 9· Fill the gap
To give an auditor the record of every time Anthropic personnel viewed your covered content, which activity type completes this Activity Feed query?
curl --fail-with-body -sS -G \
"https://api.anthropic.com/v1/compliance/activities" \
--data-urlencode "activity_types[]= ? " \
--data-urlencode "limit=50" \
-H "x-api-key: $ANTHROPIC_COMPLIANCE_ACCESS_KEY" \
-H "anthropic-version: 2026-10-21"Access Transparency records each human view of covered content by Anthropic staff as an anthropic_access event. Your own organization's API calls and admin actions use the standard Activity Feed event types, so they don't show vendor access.
Source: platform.claude.comCheckpoint 9 of 9· Check yourself
You serve several customers from one Claude organization. One regulated customer needs US-only inference for its own traffic and a record of any Anthropic staff access to its content. Which setup gives that customer the evidence it needs?
The workspace is the isolation boundary. Scoped keys and that workspace's allowed geos enforce the residency rule. Access Transparency can only be turned on for the whole organization, and each event's workspace_id links it to the customer. The Default Workspace can't be renamed and follows the organization's settings. A shared workspace with request tags gives no isolation that the platform enforces.
“Access Transparency is enabled at the organization level and covers all workspaces. Per-workspace enrollment is not currently available.”Source: platform.claude.com
Sources
(6)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“If a request specifies an inference_geo not in this list, the API returns a 400 error.”
↩︎ Data residency: inference geo, workspace geo and who enforces them“Workspace geo is set when you create a workspace and can’t be changed afterward.”
↩︎ Data residency: inference geo, workspace geo and who enforces them“Requests with inference_geo on Claude Opus 4.5, Claude Sonnet 4.5 (deprecated), Claude Haiku 4.5, or earlier models return a 400 error.”
↩︎ Data residency: inference geo, workspace geo and who enforces them“If a request specifies an inference_geo not in this list, the API returns a 400 error.”
↩︎ Key concept“Inference geo: Controls where model inference runs, on a per-request basis.”
↩︎ Exam trap 1“Workspace geo: Controls where data is stored at rest and where endpoint processing (such as image transcoding and code execution) occurs.”
↩︎ Prediction - 2.
“On Claude Models on Google Cloud (Google managed) and Amazon Bedrock, the cloud provider is the data processor”
↩︎ Data residency: inference geo, workspace geo and who enforces them“Under a ZDR arrangement, Anthropic does not store customer prompts or responses at rest after the API response is returned.”
↩︎ Zero data retention: which surfaces it actually covers“If metrics logging is enabled in Claude Code, productivity data such as usage statistics is exempted from ZDR and may be retained.”
↩︎ Zero data retention: which surfaces it actually covers“Claude Teams and Claude Enterprise product interfaces: These interfaces are not ZDR-eligible.”
↩︎ Zero data retention: which surfaces it actually covers“if a chat or session is flagged, Anthropic may retain inputs and outputs for up to 2 years.”
↩︎ Zero data retention: which surfaces it actually covers“Organizations with a ZDR arrangement cannot call the Claude API directly from a browser”
↩︎ Zero data retention: which surfaces it actually covers“Under HIPAA readiness, the API blocks requests that include a “No” feature and returns a 400 error”
↩︎ Stepping outside ZDR: stateful features and Covered Models“Covered Models are not available under ZDR regardless of feature eligibility.”
↩︎ Stepping outside ZDR: stateful features and Covered Models“Organizations with a ZDR arrangement can make these models available in a specific workspace by enabling 30-day retention for that workspace only.”
↩︎ Stepping outside ZDR: stateful features and Covered Models“The Default Workspace always follows the organization’s setting”
↩︎ Stepping outside ZDR: stateful features and Covered Models“29-day retention; async storage required.”
↩︎ Stepping outside ZDR: stateful features and Covered Models“PHI typically appears in message content (prompts and Claude’s responses), attached files (images, PDFs), and file names or metadata associated with message content.”
↩︎ Keeping PHI and PII in the fields that are protected“If your organization handles PHI, HIPAA readiness is the arrangement to use; you do not also need ZDR.”
↩︎ Keeping PHI and PII in the fields that are protected“Do not include PHI in JSON schema definitions.”
↩︎ Keeping PHI and PII in the fields that are protected“PHI must not be included in the profile data you send.”
↩︎ Keeping PHI and PII in the fields that are protected“SDK spans go directly from your environment to your exporter.”
↩︎ Keeping PHI and PII in the fields that are protected“The Activity Feed retains data for 6 years.”
↩︎ Audit evidence and tenant isolation with workspaces“Under ZDR, the API does not block these features; using one is a choice to step outside your ZDR arrangement for that specific data”
↩︎ Exam trap 2“Organizations with a ZDR arrangement can make these models available in a specific workspace by enabling 30-day retention for that workspace only.”
↩︎ Exam trap 3“If metrics logging is enabled in Claude Code, productivity data such as usage statistics is exempted from ZDR and may be retained.”
↩︎ Exam trap 4“Under ZDR, the API does not block these features; using one is a choice to step outside your ZDR arrangement for that specific data”
↩︎ Prediction“Workspaces without an override continue to follow the organization default.”
↩︎ Checkpoint“Patient-specific information should appear only in message content, where it is protected under HIPAA safeguards.”
↩︎ Checkpoint - 3.
“If these can contain PHI, filter or redact them at your collector.”
↩︎ Keeping PHI and PII in the fields that are protected - 4.https://support.claude.com/en/articles/14477985-monitor-claude-cowork-activity-with-opentelemetryOfficial docs
“Events only carry metadata unless you turn on content capture with the otlpContentCapture device policy”
↩︎ Keeping PHI and PII in the fields that are protected - 5.
“Events represent human access, not automated processing.”
↩︎ Audit evidence and tenant isolation with workspaces“Your API calls, admin actions, and Compliance API reads are covered by standard Activity Feed event types.”
↩︎ Audit evidence and tenant isolation with workspaces“Anthropic records Access Transparency events even while the Compliance API is off, but you can read them only while it’s on.”
↩︎ Audit evidence and tenant isolation with workspaces“Create an Admin API key after you turn on the Compliance API, because a key created earlier can’t read the feed.”
↩︎ Audit evidence and tenant isolation with workspaces“The log is an append-only, signed record that you can verify independently of Anthropic’s serving systems.”
↩︎ Audit evidence and tenant isolation with workspaces“Individual employee identity is not disclosed”
↩︎ Audit evidence and tenant isolation with workspaces“Access Transparency is enabled at the organization level and covers all workspaces.”
↩︎ Exam trap 5“Access Transparency is enabled at the organization level and covers all workspaces. Per-workspace enrollment is not currently available.”
↩︎ Checkpoint - 6.
“API keys can be scoped to a single workspace. In this case, they can only access resources within that workspace.”
↩︎ Audit evidence and tenant isolation with workspaces“Every organization has a Default Workspace that cannot be renamed, archived, or deleted.”
↩︎ Audit evidence and tenant isolation with workspaces