In short
- To send PHI to a large language model you need a BAA with whoever processes it, either AWS for Amazon Bedrock, which is on AWS's HIPAA eligible services list, or Anthropic for its HIPAA-ready services, including its first-party API.
- Both agreements cover specific services and features, and Anthropic's coverage table lists which API features are covered, which are not and which are unavailable to HIPAA-ready organizations.
- Beyond the BAA, on Bedrock you also choose a data retention mode, decide whether to turn on invocation logging, and pick geographic or global cross-Region inference, and each choice changes where PHI is stored.
Running an LLM under a HIPAA BAA means two things: the company that processes your prompts has signed a business associate agreement with you, and you only use the services and features that agreement covers. For Claude there are two common routes. Amazon Bedrock is on AWS's HIPAA Eligible Services Reference, so the AWS BAA you accept in AWS Artifact covers it. Anthropic also provides a BAA for its HIPAA-ready services, including its first-party API, with a published list of covered and uncovered features. After the contract, three settings decide where PHI actually lands: retention, logging and region.
This guide is for engineering leads adding generative AI to a product that handles protected health information. It compares the two routes and the configuration choices on each, as the vendors document them at the time of writing. Both vendors change their pages often, so check the linked pages before you rely on any detail.
Why does an LLM provider need a BAA?
A model provider that receives PHI in prompts is creating, receiving or transmitting PHI on your behalf, which is what a business associate does. Under 45 CFR 164.314(a), the business associate contract must require the associate to comply with the HIPAA Security Rule, to make sure any subcontractors that handle electronic PHI agree to the same, and to report security incidents, including breaches of unsecured PHI. The same requirements apply between a business associate and its own subcontractor, which matters if you are a vendor serving providers: they sign a BAA with you, and you need one with the model provider underneath.
How does Amazon Bedrock fit under the AWS BAA?
AWS's HIPAA Eligible Services Reference lists Amazon Bedrock and Amazon Bedrock AgentCore among the services eligible to create, receive, process, maintain or transmit PHI. You accept the AWS BAA in AWS Artifact for an account or for a whole AWS organization. AWS's rule is that you may use any service in a HIPAA account but should process, store and transmit PHI only in eligible services.
AWS also documents how third-party models are isolated. In each Region where Bedrock runs, there is one model deployment account per model provider, owned and operated by the Bedrock service team. Model providers have no access to those accounts, and therefore no access to Bedrock logs or to customer prompts and completions. AWS's data retention page adds that Bedrock does not share your content with model providers under any current retention mode.
What does Bedrock retain, and can you control it?
Bedrock now lets you set a data retention mode per Region, at the account level or per project. The setting applies across the Messages, Chat Completions and Responses APIs, and a Region you have not configured falls back to each model's default. The modes, from least to most permissive:
- none: no request or response data is written to durable storage by AWS or shared with the model provider.
- default: the model's own retention policy applies. AWS may retain data for safety and abuse prevention, and the provider does not receive it.
- aws_review: inputs and outputs may be retained for human review by AWS, within AWS, for models whose provider requires that review as a condition of access. Content is not shared with the provider.
- provider_data_share: a legacy mode that, according to AWS, today behaves like aws_review because Bedrock does not share content with providers.
- inherit: no setting at this scope; the broader scope decides.
Each model declares the modes it accepts. AWS's example is that Claude Fable 5, 5.1 and 5.2 require aws_review, so they are unavailable until you allow it, and for those models prompts and completions are retained within AWS for up to 30 days and may be reviewed by AWS. AWS's documentation says a model that accepts none, such as Claude Opus 4.8 in its example, does not retain data whatever mode you set. If cross-Region inference is on, retained data is stored in the Region where the request was processed. You can restrict which modes anyone in your organization may set with IAM policies or service control policies.
This is a decision for your privacy officer as much as for engineering. Before you allow a mode that permits human review of prompts that contain PHI, check it against your own BAAs with customers, your notice of privacy practices and your risk analysis.
Where else can PHI end up on Bedrock?
- Invocation logging. It is off by default. Once enabled, it records full request and response bodies, up to 100 KB inline and larger bodies as separate S3 objects, in CloudWatch Logs or S3 in the same account and Region, and keeps them until you delete the configuration. Treat those destinations as PHI stores with their own encryption and access rules.
- Cross-Region inference. Geographic inference profiles route requests within a geography such as the US or EU, while global profiles route to any supported commercial Region. AWS recommends geographic profiles when you have data residency requirements. CloudTrail records the Region that processed each request.
- Tags and names. AWS recommends never putting sensitive information in tags or free-form fields such as a Name field, because they may be used in billing or diagnostic logs.
- Your own application. Prompt templates, caches, traces and analytics events often hold the same text you sent to the model. They need the same controls as the database.
AWS's general guidance for Bedrock also applies: protect credentials with IAM, use multi-factor authentication, connect over TLS, which AWS requires at version 1.2 or above, and log API activity with CloudTrail.
What does Anthropic's own BAA cover?
Anthropic's privacy center says it provides a BAA covering its HIPAA-ready services, such as its first-party API, Claude in Microsoft Foundry and Claude Enterprise plans. For the first-party API, the organization's Primary Owner signs the BAA and then contacts Anthropic or its sales team to have it turned on. The BAA covers only the single organization that accepted it, and covered models require 30-day data retention unless Anthropic expressly authorizes otherwise.
Coverage is listed feature by feature. For a HIPAA-ready API organization, the page lists:
- Covered: the Token Counting, Models, Org Management and Compliance APIs, and Computer Use.
- Messages API, partly covered: prompt caching, structured outputs, memory, web search, and the bash and text editor tools are covered. The page states that Messages API features not listed are not covered.
- Not covered and not accessible to HIPAA-ready API users: the Batch API, the Files API and Skills API (both beta), code execution and web fetch.
- External MCP servers: available, but third-party data flows are not covered.
- Version dates: some Messages API features are covered only under BAA versions accepted after April 1, 2026.
The table also covers Claude Code. Claude Code running through Bedrock, Google Cloud or Foundry, or through an LLM gateway, is listed as not covered by Anthropic's BAA. Each route needs its own agreement and its own check.
Bedrock or Anthropic directly: how do you choose?
- Contracts: if your PHI already lives on AWS under an accepted BAA, Bedrock adds a service to an existing agreement. Going direct adds a new business associate to your chain.
- Features: compare the exact features you need against each route. A feature that is covered on one route may be unlisted or unavailable on the other.
- Retention: Bedrock lets you set and lock retention modes per Region and project. Anthropic's covered models require 30-day retention under its BAA.
- Region and residency: Bedrock's geographic inference profiles keep processing within a geography. Check the equivalent terms for any other route.
- Model access: confirm the model version you need is offered on the route you choose, and note that some Bedrock models require a more permissive retention mode before you can call them.
Our AI assistants run on Claude via AWS Bedrock or Anthropic directly, with provider switching. Switching only stays safe if each route has its own list of allowed features and settings, kept in configuration and reviewed when either vendor updates its documentation. Otherwise a fallback can send PHI, unnoticed, to a feature the other route's agreement does not cover.
What does the BAA not do for you?
A signed BAA does not decide who in your product may see which patient, what goes into a prompt, or what happens to the answer. You still need your own risk analysis under the Security Rule, access controls in the application and a plan for testing the model's output. On our projects, we sign a BAA before anyone sees PHI, scope access so each clinician sees only their own patients, and run a penetration test before launch. We also built an assistant that reads a patient's records and answers with linked sources, on Claude under a BAA, inside a live telehealth platform. Our AI assistants cite their sources, which is how clinicians can check an answer without trusting the model blindly.
Onix Systems has been building software since 2000 and has 370+ engineers. If you are choosing between Bedrock and a direct Anthropic agreement, or checking an AI feature that already handles PHI, talk to one of our engineers. Our free five-day technical audit ends with a written report on your BAA chain, data flows and settings.
Sources
- AWS: HIPAA Compliance
- AWS: HIPAA Eligible Services Reference
- Amazon Bedrock User Guide: Data protection
- Amazon Bedrock User Guide: Data retention
- Amazon Bedrock User Guide: Model invocation logging
- Amazon Bedrock User Guide: Cross-Region inference
- Anthropic Privacy Center: Business Associate Agreements (BAA) for commercial customers
- 45 CFR 164.314: Organizational requirements (Cornell LII)