Claude is not HIPAA compliant by default: as of November 2026, Anthropic offers BAA coverage for eligible Enterprise, first-party API and Microsoft Foundry use with required configuration. Team and individual plans cannot enable HIPAA, and some features remain outside coverage. Identify the product and features staff will use, then approve that configured route before it receives protected health information.
A clinician asks to use Claude for a patient letter. An application vendor also proposes a tool that calls Claude. Security needs to review each request separately: they can send patient information through different products, endpoints and agreements even though they name the same model family.
Start the review with the application staff will open and the endpoint it will call. This guide follows Anthropic's product coverage as of November 2026. For OpenAI products and Azure OpenAI, see the ChatGPT HIPAA guide.
Anthropic's BAA covers three Claude routes
An organisation can arrange Anthropic BAA coverage for eligible Enterprise, first-party API or Microsoft Foundry use. Each route requires setup before it handles PHI. The approval needs to identify which of these services the task will use and who has completed that setup.
For Enterprise, the organisation's Primary Owner enables HIPAA under Data and privacy and accepts the BAA. A standard Enterprise workspace has no BAA coverage until the Primary Owner completes that action. Check that setting in the organisation where staff will perform the patient-data task.
For the first-party API, the Primary Owner signs the BAA and asks Anthropic to enable the covered use. The developer must then use eligible API features under that arrangement. Confirm the enablement even if the application already calls the API for other tasks.
Anthropic also lists Claude in Microsoft Foundry as eligible under its BAA with the required HIPAA configuration. Feature availability differs on that route. Record the service and features used by the application. Security can then review the request path that will handle the patient records.
Team and individual Claude plans can't be covered
Claude Team and the individual Free, Pro, Max and Ultra plans cannot enable HIPAA. A paid account on one of those plans does not provide a route for the proposed PHI task.
An Enterprise agreement covers the organisation that accepted it. It does not extend to a separate personal account or another organisation because a person belongs to both. Staff must use the account within the organisation covered by the agreement for that task.
Existing API agreements also need attention. Anthropic says an API BAA signed before 2 December 2025 does not extend to the HIPAA-ready Enterprise plan. An organisation adding Enterprise under that older agreement needs a new BAA for that access.
Staff instructions need to name the approved organisation and feature. Someone told only to "use Claude" could open a familiar personal account, outside the arrangement security reviewed. Naming the permitted route lets that person check the account before starting the patient-data task.
Some Claude features stay outside the BAA
The feature staff select matters after HIPAA has been enabled. Some available features remain outside the BAA, and others send records to a third party. Review the proposed task far enough to identify those recipients as well as the Claude service.
Enterprise chat, projects and artifacts within chat, web search and research are eligible when HIPAA is enabled. File creation and code execution have a network-access restriction. Connectors, MCP, Enterprise Search and Claude in Chrome can involve third-party data flows that Anthropic's BAA does not cover.
Claude Code in the terminal or the Desktop Code tab, and Cowork in Claude Desktop, have a separate local-mode HIPAA configuration. The Primary Owner applies it with support from Anthropic's account team. Identify the permitted terminal or Desktop surface in the approval. Cloud sessions and beta features have different coverage.
The table summarises Anthropic's November 2026 coverage. For each proposed task, match the row to the product staff will use and the configuration the Primary Owner has applied.
| Claude product or feature | Under Anthropic's BAA | What it takes |
|---|---|---|
| Enterprise chat, projects and artifacts in chat, web search, research | Eligible | Enterprise HIPAA enabled |
| Messages API | Eligible features covered | Signed BAA, enablement and required retention |
| Claude in Microsoft Foundry | Eligible | HIPAA configuration and supported features |
| Claude Code in terminal or Desktop Code tab, Cowork in Claude Desktop | Eligible local use | Primary Owner applies HIPAA configuration with Anthropic account team |
| Connectors, MCP, Enterprise Search and Claude in Chrome | Third-party data flows excluded | Separate review of enabled features and recipients |
| Skills | Excluded | Keep PHI out of uncovered use |
| Claude Code cloud sessions, Cowork in the cloud, beta features | Excluded | Use an eligible configured surface |
| Team, Pro, Max, Ultra and Free | HIPAA enablement unavailable | Use an eligible product and agreement |
Covered Claude models keep data for 30 days
Anthropic's Covered Models require 30-day retention. Zero data retention needs Anthropic's express authorisation for those models. Before a developer selects one for the application, the organisation needs to know which model and retention arrangement Anthropic has authorised for its use.
Include the retention requirement when reviewing the model and service. A team may have chosen zero data retention for another application, but that choice does not establish which models it can use under the HIPAA arrangement. Check which model and retention combination Anthropic has authorised for this organisation.
The review should also account for copies created outside the model service. An application can retain its own prompts, outputs or exported documents. Decide how those records are stored and who can access them as part of the proposed patient-data task.
Keep the model-service retention period and the application's own retention policy distinguishable. List the records held by each system and the setting or procedure used to remove them. Security and the application owner can then review each copy.
Bedrock and Google Cloud use their own BAAs
A Claude request sent through a cloud provider needs review under that provider's arrangement. Match the application endpoint and its features to the provider's BAA and eligible-service list. An approval for a different Claude route does not settle that review.
AWS lists Amazon Bedrock among its HIPAA-eligible services and states that customers must configure eligible services consistently with their responsibilities. Review the AWS agreement and the application's configuration together when choosing that route.
For a Google Cloud route, including a proposal described as Claude on Vertex AI, check the current Google Cloud covered-service list and the terms for the selected model endpoint. Check the features used by the request as well as the named model. Coverage follows the eligible service under the agreement.
Anthropic separately states that Claude Code through Bedrock, Google Cloud or Microsoft Foundry is outside Anthropic's BAA. That differs from its eligible Claude API service in Microsoft Foundry. Name the product in the approval alongside the cloud. This keeps the approved API route distinguishable from a coding-agent session governed by different terms.
A BAA leaves access and audit to you
The application still needs to enforce your organisation's patient-data permissions. A vendor agreement covers a service, but someone must decide which staff can use the application and which records their task requires. Those decisions determine the access it should receive.
Define the permitted task and the minimum information needed to complete it. A patient letter, a population summary and a finance review can require different records and different recipients. The assistant's source access should reflect those differences.
Keep enough evidence to review what the application accessed and produced. That includes the query or source records behind an answer and the destination where a person saved or shared it. An output can create another copy of PHI even when the original request used an approved service.
Name the people who will review access changes and maintain the approved configuration. The model name alone does not establish HIPAA compliance. The deployment, agreements, data flows and operating procedures need to describe the same application, including what happens after a user receives an answer.
Databasin keeps patient data in your tenant
We prepare approved Epic and finance records for questions inside your Azure environment, following the institution's data permissions. Databasin Private installs from the Azure Marketplace into your tenant, where your network, keys and policies govern the deployment.
We configure a connector for Claude, GPT or an internally approved model. Its endpoint determines where the request is handled, so the model route forms part of the deployment design. Review that route with the data sources and permissions the assistant needs.
Databasin is HIPAA-ready, with a BAA available. People can inspect the SQL accompanying an answer. Query logs and lineage let them follow the result back to the data used. Investigation agents have read-only production-source tools and limits on their runs.
The secure AI page explains how those controls apply to business questions. We build against the institution's approved sources and access scope. The deployment defines the model endpoint, who operates the assistant and the evidence available to review an answer.
Example: five Claude uses checked before rollout
These illustrative requests could reach a hospital in the same week, with each requiring a decision about its Claude product and feature.
| Use | Product | Covered by Anthropic's BAA? |
|---|---|---|
| Draft a patient letter | Enterprise chat with HIPAA enabled | Yes, within the accepted agreement |
| Summarise notes inside your own app | First-party API with signed BAA and enablement | Yes, using eligible features and required configuration |
| Read an EHR page in the browser | Claude in Chrome | Third-party data flows are outside coverage |
| Debug a script on PHI test data in a cloud session | Claude Code on the web | No |
| Analyse an export in a shared workspace | Claude Team | No, Team cannot enable HIPAA |
Approval follows the product, feature and configured data path. Review the browser integration and cloud coding session separately from the covered chat task.
Five checks before PHI reaches Claude
- Confirm the plan and the signed BAAMatch the agreement to the organisation and product that will process the records, including any Enterprise access added to an older API arrangement.
- Apply the HIPAA configurationHave the Primary Owner complete the required enablement and review the separate configuration for local Claude Code and Cowork use.
- Turn off uncovered featuresRestrict the proposed PHI task to covered features and review any connector or external recipient before allowing records through it.
- Keep Anthropic's current table on fileKeep a dated copy of the coverage table with the approval and record which product features the application will use.
- Decide where the model runsIdentify the endpoint, provider agreement and retention settings alongside the data sources, user permissions and copies created by the application.
Sources
Anthropic Privacy Center, Business Associate Agreements for Commercial Customers, 2026 · privacy.claude.com/en/articles/8114513-business-associate-agreements-baa-for-commercial-customers
Claude Help Center, HIPAA-ready Enterprise plans, 2026 · support.claude.com/en/articles/13296973-hipaa-ready-enterprise-plans
Amazon Web Services, HIPAA Eligible Services Reference, 2026 · aws.amazon.com/compliance/hipaa-eligible-services-reference
Google Cloud, HIPAA compliance on Google Cloud and Google Workspace, 2026 · cloud.google.com/security/compliance/hipaa
Plan AI on your patient data with us
We can walk through your security requirements and how an assistant over Epic and finance data would run in your tenant. A demo explores the questions people need to answer, the model route and the evidence available for review.