Is ChatGPT or Claude HIPAA compliant?

A hospital leadership team wants an assistant that answers questions from Epic and finance data. Security asks three questions: is there a BAA, does patient data leave our tenant, and can every answer be audited? The model vendor answers only the first.

Vendor coverage changes often. The summary below reflects each vendor's published pages as of November 2026; check the vendor's current list before relying on it.

Which ChatGPT products can be covered by a BAA?

OpenAI's help center describes two routes. For the API, organizations can request a BAA, and HIPAA eligibility depends on the account being set up with the retention configuration OpenAI specifies. For ChatGPT, OpenAI says only sales-managed ChatGPT Enterprise or Edu accounts are eligible, and that it does not offer a BAA for ChatGPT Business. Consumer ChatGPT plans are not covered.

OpenAI also lists healthcare-specific workspace options with their own covered features. Some features, such as certain connected-app automations, are excluded from BAA coverage; review the feature list, not only the plan name.

Which Claude products can be covered by a BAA?

Anthropic's privacy center says its first-party API is covered once the organization's primary owner signs a BAA, and Claude Enterprise is covered only after a primary owner activates HIPAA settings. Consumer plans (Free, Pro and Max) are handled separately and are not covered. Anthropic notes that covered models require its standard 30-day retention rather than zero data retention unless it authorizes otherwise, and that some surfaces, beta features and connectors fall outside the BAA. For plans the page does not mention, check with Anthropic.

What about Azure OpenAI and Microsoft Foundry?

Microsoft makes its HIPAA BAA available through the Product Terms and the Data Protection Addendum by default to customers who are covered entities or business associates, with no separate contract to sign. Coverage applies to in-scope services; Microsoft publishes that scope in its compliance offerings document on the Service Trust Portal. Running the model inside your own Azure subscription, behind private networking, keeps prompts and responses within infrastructure your security team already governs. Anthropic lists Claude in Microsoft Foundry among its HIPAA-ready options.

What does a BAA not cover?

A BAA defines how the vendor handles PHI it receives. It does not decide whether a given use is permitted, who inside your organization may ask about which patients, or what happens to an answer after it is generated. Microsoft states plainly that using its services does not on its own achieve HIPAA compliance.

• Access control: which users can reach which records through the assistant.
• Minimum necessary: whether the prompt carries more PHI than the question needs.
• Audit: a record of who asked what, which data was read and what came back.
• Retention and exports: where answers, files and logs end up.

How do hospitals deploy AI on patient data?

Three patterns are common. A business ChatGPT or Claude workspace under a BAA suits drafting and general questions that include some PHI. An API under a BAA, wrapped in your own application, suits defined tasks where you control the prompt. A model inside your own cloud tenant (Azure OpenAI, Claude in Microsoft Foundry or a privately hosted model) suits assistants that query clinical and finance data directly, because the data, the model and the audit log stay under your policies.

Where does Databasin fit?

Databasin provides the third pattern for questions about Epic, Workday and finance data. It installs from the Azure Marketplace into your own tenant, calls Azure OpenAI in your subscription or another model you approve, and answers from governed tables rather than raw exports. Agents are read-only, every run is audited, and each answer shows the SQL behind it. Databasin is HIPAA-ready and a BAA is available. Our team implements the connections, definitions and access rules with your security and analytics staff.

Example: the same question three ways

An illustrative question: "How many sepsis encounters did each unit have last quarter?"

Checks before PHI reaches a model

1. Confirm the BAA and its scopeName the product, the plan and the features covered, and keep the vendor's current list on file.
2. Choose where the model runsVendor workspace, API or your own tenant, decided by the data the assistant must reach.
3. Limit what each user can ask aboutApply your existing access rules to the data behind the assistant.
4. Keep an audit trailRecord the question, the data read and the answer for every run.

Sources