Claude BAA Coverage and PHI Routing Across the Anthropic API, Amazon Bedrock, and Google Vertex AI
Treat each way of reaching Claude as a separate destination with its own agreement, covered services, and feature conditions. Send PHI only to destinations whose executed BAA and feature eligibility you have verified. Limit every other surface to de-identified or non-PHI data. Enforce the decision at runtime with allow, redact, and block actions, and record each decision as audit evidence for 164.312(b).
Enforcing the routing rules in the execution path
- 1
1. Resolve the destination
Determine the surface, account, and feature set the request would use, including tool calls and file attachments.
- 2
2. Detect PHI
Inspect the payload for PHI before egress. Test detection accuracy on representative data.
- 3
3. Look up the policy
Match the destination to its tier and permitted data class in the current policy version.
- 4
4. Allow, redact, or block
Allow PHI to Tier 1 destinations. Redact PHI bound for Tier 2. Block PHI to unverified or non-covered destinations.
- 5
5. Fail closed
If detection or policy evaluation is unavailable, block PHI-bound traffic rather than passing it through.
- 6
6. Record the decision
Write destination, action, policy version, and timestamp for every decision.
Confirm before any PHI is sent
- The executed agreement that covers PHI for this access path, with its version and date.
- The models and regions covered under that agreement.
- Eligibility or exclusions for tools, file handling, batch, prompt caching, web search, and retention settings.
- Data handling and retention terms that apply to prompts and outputs.
- Whether any consumer-facing Claude product is reachable from your network, and how it is blocked for PHI.
- How the vendor communicates changes to coverage, and who receives those notices.
Why BAA coverage is a per-surface routing problem
HHS states that a covered entity or business associate may use a cloud service to store or process ePHI, provided it enters into a HIPAA-compliant business associate agreement with the cloud service provider that will be creating, receiving, maintaining, or transmitting ePHI on its behalf, and otherwise complies with the HIPAA Rules. The agreement attaches to the specific provider receiving the data.
Claude can be reached through the Anthropic API, Amazon Bedrock, and Google Vertex AI. That means three distinct provider relationships: Anthropic, AWS, and Google Cloud. Each has its own agreement, its own list of covered services, and its own terms for features and data handling. A BAA signed with one provider says nothing about the others.
The Security Rule applies to ePHI that is created, received, used, or maintained by a covered entity or business associate. A prompt or file sent to a model endpoint is plausibly ePHI in transit to a business associate. The practical risk is rarely the contract itself. It is the gap between contractual eligibility and runtime behavior: a developer picks an endpoint, an SDK default, a feature, or a consumer-facing product that the agreement does not cover, and PHI follows. Routing needs to be an explicit, enforced model rather than developer discretion.
What to record for each Claude destination
Build one routing record per destination. The unit of routing is the surface plus the account plus the feature set, not the vendor name. Populate each record only from executed agreements and current vendor eligibility documents, and note the document version and date.
- Anthropic API: Record the executed Anthropic agreement, covered models, covered features, and data retention settings. Default tier until verified: non-PHI or de-identified only.
- Amazon Bedrock: Record the AWS agreement reference, the covered services and models in your region and account, and feature conditions. Default tier until verified: non-PHI or de-identified only.
- Google Vertex AI: Record the Google Cloud agreement reference, covered services and models, and feature conditions. Default tier until verified: non-PHI or de-identified only.
- Consumer-facing Claude products: Treat as outside the PHI tier unless a specific agreement says otherwise. Default action: block PHI, and block the destination where policy allows.
A decision model for PHI routing
Use three tiers. Tier 1 is PHI-eligible: the destination, account, and feature set all map to a verified BAA entry. Tier 2 accepts de-identified or non-PHI data only. Tier 3 is blocked for all Claude traffic. Any destination without a verified entry defaults to Tier 2 or Tier 3. Default-deny is a design recommendation rather than a quoted requirement, but it follows directly from the BAA prerequisite HHS describes.
Feature eligibility belongs in the policy schema. Tools, file handling, batch processing, prompt caching, web search, and data retention settings can each carry different conditions from the base model endpoint, and those conditions can change over time. Give each feature its own attribute in the routing table, set to not eligible until the vendor documentation confirms otherwise. Then a request that uses an unverified feature on an otherwise eligible endpoint is treated as a different destination.
There are tradeoffs. Strict feature gating reduces what teams can build against PHI. Redaction can remove context the model needs. PHI detection is imperfect, so accuracy has to be tested on your own data rather than assumed. Decide in advance whether the system fails closed when detection or policy evaluation is unavailable. For PHI-bound traffic, failing closed is the defensible default.
Evidence for 164.312(b) and where runtime enforcement fits
Audit controls are a distinct technical safeguard standard at 45 CFR 164.312(b), and 164.312 applies to covered entities and business associates in accordance with 164.306. Pull the verbatim text of 164.312(b) from the eCFR before mapping evidence, and re-read it at each review. As of November 2026, the eCFR source note for 164.312 lists 2003 and 2013 actions. A Federal Register document on strengthening Security Rule cybersecurity was published in January 2025, so check current status at each review.
HHS says covered entities are not required to monitor how business associates carry out privacy safeguards. That statement concerns the Privacy Rule and does not remove your own Security Rule obligations or the need for your own technical controls. Your evidence should therefore show what your environment allowed, redacted, or blocked, not only what the vendor agreed to.
For each decision, retain the destination, the action, the policy version in force, the timestamp, and the requesting identity. Store the detection outcome by category rather than the PHI itself. Keep the policy version history so a reviewer can see why a request was handled as it was.
Trussed enforces policy in the execution path, so these decisions are enforced at runtime, not reviewed after the fact. It applies allow, redact, and block by destination, including models, tools, and MCP servers, at the level of individual interactions and tool calls. The audit-ready evidence is generated by enforcement itself and mapped to HIPAA, including 164.312(b). It can be deployed self-managed in the customer cloud with configurable retention and data residency, and it runs in production with healthcare organizations.
PHI routing decision by destination
Identify destination
Surface, account, and feature set for each request.
Check the routing table
Executed BAA reference, covered features, permitted data class.
Enforce
Allow, redact, or block before the request leaves the environment.
Record
Destination, action, policy version, and timestamp as evidence.
Turn the routing table into an enforced control
Talk to an expert about enforcing PHI routing by destination in the execution path, with audit-ready evidence mapped to HIPAA.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.