HIPAA-compliant AI tools for healthcare teams

By Mat Steinlin, Head of Information Security

Last updated: July 2026

Most compliance conversations about AI start with the LLM APIs: Claude, OpenAI, Gemini. Those are the right first questions. But healthcare teams use AI across a much wider surface: coding assistants in their development workflows, meeting tools in clinical and operational meetings, documentation AI for notes and records, productivity software embedded throughout the organization.

Each of these is a potential PHI exposure point. Each needs the same evaluation: is there a BAA, what does it cover, and is this tool being used in a way that requires one?

Quick reference: BAA status by tool

Coding assistants

The compliance question for coding tools is distinct from the question for clinical tools. The concern isn't patient data flowing through the interface. It's PHI in the codebase, test fixtures, database dumps, and development environments that gets swept into AI context windows.

GitHub Copilot

GitHub Copilot Enterprise includes a BAA through Microsoft's enterprise agreement, the same DPA that covers Azure OpenAI and Microsoft 365 Copilot. GitHub does not publish a standalone Copilot-specific HIPAA page. Verify coverage with your Microsoft account representative if you need written confirmation.

Individual and Team plans do not include a BAA. If your organization uses Copilot on non-enterprise plans and developers work in codebases containing PHI (in test data, logs, database seeds, or configuration files) that usage is not covered.

The practical guidance: audit which Copilot plan your engineering team is on, confirm enterprise BAA coverage, and establish policies about what data is acceptable in development environments. PHI in test fixtures is a common problem that precedes any AI tooling discussion.

Cursor

Cursor is an AI-native code editor that sends code context (including file contents, highlighted code, and repository context) to AI models for completions, chat, and other features. Cursor offers a HIPAA BAA for Enterprise customers, with two requirements: Privacy Mode must be enabled and locked organization-wide, and a BAA must be signed before any PHI is submitted.

Privacy Mode must be configured at the organization level, not left to individual developers. Third-party integrations used alongside Cursor are not automatically covered by the BAA. If your engineering team uses Cursor on a non-Enterprise plan and works in environments containing PHI, those interactions remain uncovered.

JetBrains AI Assistant

No HIPAA certification or BAA availability appears in official JetBrains documentation. Standard enterprise licensing covers the IDE, not the AI services. The same PHI-in-codebase risk applies as with any AI coding tool: if developers work in environments where PHI may appear in code, logs, or test data, those interactions are not covered.

Error tracking and observability tools

Error tracking tools are a compliance gap that most healthcare engineering teams overlook. Tools like Sentry, Datadog, and LogRocket receive PHI through default request/response logging, even when developers never consciously sent patient data to them. The exposure is in the default configuration: full request bodies are logged, and those request bodies frequently contain PHI. One practitioner described pulling social security numbers and diagnoses out of a client's Sentry within ten minutes of starting an engagement. Sentry does offer a HIPAA BAA on Business and Enterprise plans. The required action is not just signing the BAA. Audit the default logging configuration to prevent PHI from reaching these services in the first place.

Meeting and documentation tools

Clinical and operational meetings frequently involve discussions of patient cases, treatment decisions, and care coordination. Meeting AI tools (those that record, transcribe, and summarize) are directly in scope for HIPAA if PHI is discussed.

One important caveat for behavioral health and substance use disorder (SUD) use cases: 42 CFR Part 2 is a stricter, separate consent regime that governs SUD treatment records. A standard HIPAA BAA does not automatically cover Part 2; a reader could have a fully valid BAA in place and still be in violation of Part 2 if the agreement wasn't specifically drafted to incorporate it. If your use case involves behavioral health or SUD, confirm that your vendor's BAA explicitly incorporates Part 2 terms before using any meeting AI tool in those contexts. Don't assume HIPAA coverage extends to it.

Otter.ai

Otter.ai offers HIPAA compliance on their Enterprise plan only, not Business or consumer tiers. A BAA must be executed before any PHI is handled; Otter is not a covered entity by default and only becomes a business associate once a signed BAA is in place. Contact your account executive to start the BAA process.

Getting the BAA in place is only the starting point. Otter's HIPAA guidance identifies several controls that organizations are responsible for configuring: disabling public and link-based transcript sharing, enforcing 2FA and SSO, managing the Notetaker auto-join feature (which can automatically record scheduled meetings), and setting custom data retention policies. These controls are not on by default; they require admin configuration.

The key risk: employees using personal Otter.ai accounts for work meetings. If clinical staff record patient discussions through a personal or Business-tier account, those interactions fall entirely outside any organizational BAA.

Fireflies.ai

Fireflies.ai offers HIPAA compliance on their Enterprise plan only. The platform is SOC 2 Type II certified. A BAA must be in place before handling any PHI. Fireflies has been evaluated by organizations with heavy PHI workflows as part of the same AI meeting tool decision that includes Copilot and Zoom AI Companion; the BAA and admin control requirements mirror those of other enterprise meeting tools. The same governance requirements apply: ensure staff are not using personal or lower-tier Fireflies accounts for meetings where PHI is discussed.

Granola

Granola is an AI meeting notes tool. As of June 2026, Granola does not offer a HIPAA BAA. Healthcare organizations should not use Granola for meetings where PHI may be discussed until a BAA path becomes available. The risk extends beyond typed content: Granola transcribes meetings, which means verbal PHI discussed during calls (patient cases, treatment decisions, care coordination) is captured in Granola's transcripts. Healthcare teams often treat meeting transcription as lower-risk than document AI, but the PHI vector is the same.

Purpose-built clinical AI tools

A separate category of AI tools is designed specifically for healthcare workflows: AI medical scribes, clinical documentation assistants, and ambient AI tools that operate in clinical environments. These are purpose-built for regulated data from the ground up.

Notable examples:

• Freed is an AI medical scribe with a HIPAA BAA in use with hundreds of health systems. BAA covers the entire organization, not just individual users. Designed for clinician documentation; records and summarizes patient encounters.

• Nabla is an ambient AI tool for clinical documentation. A BAA is required before using Nabla with PHI; organizations cannot use the service with patient data until a BAA is executed. Supports real-time transcription and note generation in clinical settings.

• Abridge is a clinical AI platform developed with academic medical centers, with a BAA. Used by several major health systems.

• Suki is an AI clinical scribe with a HIPAA BAA and SOC 2 Type II certification. It integrates directly with most EMRs and is used for note documentation. Practitioners in clinical communities have named it alongside Abridge as a recommended approved ambient scribe for its EMR integration coverage.

• Heidi Health is a clinical AI scribe positioned for smaller practices. A HIPAA BAA is available; the specific tier requirement for BAA eligibility is not confirmed in public documentation; verify with Heidi Health directly before use with PHI.

For clinical documentation specifically, purpose-built tools with existing healthcare compliance programs are the lower-risk path compared to adapting general-purpose meeting AI for clinical use.

Voice dictation AI

Voice dictation tools occupy a distinct category from meeting AI and clinical scribes. They capture real-time speech to text as a general-purpose input method, and healthcare workers use them across a wide range of workflows: composing notes, drafting messages, filling forms.

Wispr Flow

Wispr Flow is an AI voice dictation tool with HIPAA BAA coverage available on all plans, including individual plans. Enabling HIPAA mode is self-serve: users activate it via Settings > Data and Privacy. Enterprise organizations can enforce HIPAA mode org-wide. This self-serve BAA path is notable because it contrasts with sales-gated BAA processes at other AI vendors. Practitioners have specifically called out Wispr Flow's accessibility compared to enterprise-only BAA options elsewhere.

Productivity AI

The productivity AI category is the most diffuse and in some ways the most difficult to govern. These are features embedded in tools that organizations already use extensively: email, documents, calendars, project management.

Microsoft 365 Copilot

Microsoft 365 Copilot is covered under Microsoft's enterprise HIPAA BAA for eligible M365 enterprise plans and appears explicitly in Microsoft's in-scope services table for the Commercial environment. This is the same Microsoft DPA that covers Azure OpenAI.

Coverage requires an M365 enterprise agreement with the appropriate data protection terms in place. Not all M365 plans include the data protection addendum. Verify with your Microsoft account representative that your specific plan is covered before enabling Copilot for clinical or operations staff who handle PHI.

Note: Microsoft 365 Copilot and GitHub Copilot are separate products with separate compliance postures. Having a Microsoft Enterprise Agreement does not extend Azure or M365 BAA coverage to GitHub Copilot.

Google Workspace Gemini

Coverage depends on your Workspace plan and configuration. See Gemini / Vertex AI for the full breakdown. The short version:

• Enterprise Workspace plans with appropriate data processing terms in place can be covered

• Business plans have different terms

• Consumer-tier Workspace does not have BAA coverage for AI features.

Notion AI

Notion offers a HIPAA BAA for their platform. That BAA does not cover Notion AI.

Notion AI sends content to AI providers under Notion's own agreements, not the customer's agreements. This means PHI in Notion pages that gets processed through Notion AI features (summarization, editing, Q&A) is being sent to third-party AI infrastructure outside the scope of your organization's BAA with Notion.

Organizations with Notion BAAs in place should not use Notion AI features with content containing PHI until Notion offers BAA coverage for AI features specifically.

Slack AI

Slack offers a HIPAA BAA for Enterprise Grid customers; the BAA is not available on Pro or Business plans. Slack AI is not explicitly addressed in the HIPAA compliance documentation. Verify with your Slack account representative whether AI features are included in your BAA coverage before using them in channels where PHI may be present.

The shadow AI problem

Every BAA conversation starts with the tools an organization has evaluated and contracted for. The real exposure is often in the tools being used without organizational review.

Healthcare workers use consumer AI tools routinely: ChatGPT for documentation drafts, Claude.ai for research summaries, Otter.ai personal accounts for meeting notes. These are not covered by organizational BAAs. They are happening regardless of policy.

The AMA's 2026 Physician Survey on Augmented Intelligence found over 80% of physician respondents currently use AI in a professional context, double the share reported in 2023. The same survey found physicians are significantly more concerned about patient privacy when using non-institutional AI tools (71%) than institutional ones (42%), a gap that reflects how much unsanctioned use is already happening. Research in JAMA documented 17% of healthcare workers admitting to using unapproved AI tools, most commonly for clinical documentation.

The gap isn’t typically that employees are trying to violate HIPAA… It's that the consumer tools are better or faster at the task, the approved alternatives are cumbersome, and the compliance boundary is not visible to the end user. Claude for Enterprise and Claude.ai Free look identical. ChatGPT Team and ChatGPT Plus look identical.

Practitioners who discuss this in clinical communities have identified a practical workaround: use AI for templates and frameworks, fill in patient details yourself afterward. That approach captures roughly 80% of the time savings without routing PHI through any AI tool. It's a useful bridge between the approved AI footprint your organization has and the workflows your staff are already trying to run.

For a framework for discovering and governing unauthorized AI tool usage across your organization, see Shadow AI in healthcare.

What you still need

Every BAA in this chapter covers how that vendor handles your data on their infrastructure. Your audit trail, access controls, and governance policies across all of these tools are yours to build and maintain; no vendor agreement creates them for you.

For teams building healthcare applications rather than operating within healthcare organizations, the relevant compliance surface is the LLM APIs and infrastructure under your application. Aptible LLM Gateway handles the compliance layer for that surface. For the tools your team uses day-to-day, the evaluation framework in this chapter applies.

FAQs

Is Cursor HIPAA compliant?

Cursor offers a HIPAA BAA for Enterprise customers, with Privacy Mode enabled and locked organization-wide. Non-Enterprise plans have no BAA coverage. Do not use non-Enterprise Cursor in environments where PHI may be present.

Is Otter.ai HIPAA compliant?

Otter.ai offers a HIPAA plan with a BAA. This is a paid plan, not the consumer tier. Consumer Otter.ai accounts have no BAA coverage. Verify that your organization is on the HIPAA plan before using Otter for meetings involving patient discussions.

Is Notion AI HIPAA compliant?

No. Notion offers a platform BAA but it does not cover Notion AI features. Do not use Notion AI with PHI-containing content.

Is GitHub Copilot HIPAA compliant?

GitHub Copilot Enterprise is covered under Microsoft's HIPAA BAA. Individual and Team plans are not. Verify your organization's Copilot plan before allowing it in environments where PHI may appear in code.

Can I use AI note-taking tools in clinical settings?

Purpose-built clinical AI tools like Freed, Nabla, Abridge, Suki, and Heidi Health offer HIPAA BAAs and are designed for clinical environments. General-purpose meeting tools like Otter.ai and Fireflies.ai have HIPAA enterprise plans available. Consumer-tier tools (personal Otter.ai, Granola, consumer transcription apps) should not be used in clinical settings where PHI is discussed.

Does my Microsoft Enterprise Agreement cover GitHub Copilot?

No. Microsoft 365 Copilot and GitHub Copilot are separate products with separate compliance postures. Your M365 or Azure BAA does not extend to GitHub Copilot. GitHub Copilot Enterprise requires its own BAA coverage through the Microsoft enterprise agreement specifically for Copilot. Verify with your Microsoft account representative.

What AI tools are safe to use with PHI?

Tools with HIPAA BAAs in place, used according to the terms of those agreements, with appropriate organizational controls. The BAA is necessary but not sufficient. Your audit controls, access management, and governance framework are equally required.

What's next

Once you've audited the tools your team uses, the next challenge is managing BAAs and compliance requirements across all of them. See Managing BAAs across multiple AI vendors for how to manage the full stack without building each vendor's compliance controls independently.

For a governance framework for shadow AI (the tools being used without organizational review), see Shadow AI in healthcare.

__

Many of the tools in this chapter (Notion, Slack, Otter, and others) now publish MCP servers, which means your team can connect to them directly from Claude or other AI clients. That's powerful and also a new compliance surface: when an AI agent calls an Otter MCP server in a meeting where PHI is present, or queries a Notion workspace that contains patient data, those tool calls are in scope.

Aptible MCP Gateway provides audit logging, access controls, and PHI de-identification for MCP tool calls, so your team can use these integrations without routing PHI through uncontrolled connections. Talk to an engineer to see how it fits.