HIPAA and generative AI: a practical guide

Scope of this guide

This guide is for covered entities (health plans, clearinghouses, and providers who conduct standard electronic transactions) and business associates, and focuses on the HIPAA Privacy and Security Rules as they apply to AI tools.

When an AI vendor needs a BAA

A vendor that creates, receives, maintains, or transmits PHI on behalf of a covered entity or business associate is generally a business associate and must sign a Business Associate Agreement before PHI is shared. AI vendors are no exception. Key points:

• The BAA must cover the specific product and plan. Many vendors offer BAAs only for enterprise tiers or specific services. A BAA for one product does not automatically extend to another from the same vendor.
• Consumer accounts are not covered. An employee's personal account is outside any BAA your organization signed.
• Configuration may be part of the deal. Some BAAs require specific settings (for example, particular retention options or disabled features) for coverage to apply.
• Sub-processors matter. If the AI product relies on a third-party model provider, that provider handles PHI too; the vendor's own agreements must flow down appropriately.

Minimum necessary and de-identification

The minimum necessary standard applies to most uses and disclosures of PHI (with exceptions such as disclosures for treatment). For AI, that means sending only the PHI the task needs — for example, the relevant note rather than the full chart.

Properly de-identified data is not PHI. HIPAA recognizes two methods: Safe Harbor (removing 18 specified identifiers and having no actual knowledge that the remainder could identify an individual) and Expert Determination. Removing names alone is not de-identification; dates, geographic detail, and free-text notes frequently contain identifiers.

Security Rule safeguards for AI tools

Common gaps we see

• Clinicians or billing staff using personal AI accounts to draft letters or summarize notes.
• An AI scribe or documentation tool approved clinically, without confirming the BAA covers the specific product.
• AI features enabled inside an EHR or practice-management system without reviewing what data they send and where.
• Developers testing AI integrations with real patient data on developer API tiers.
• No logging of what AI tools accessed.

Practical control checklist

• Inventory every AI tool, including AI features inside existing clinical and business systems.
• For each tool that may touch PHI, confirm a BAA covers the specific product and plan; record it.
• Prohibit PHI in consumer AI accounts in policy and training.
• Configure retention and disable features not covered by the BAA.
• Enforce SSO and MFA.
• Enable audit logging and retain logs.
• Use de-identified or synthetic data for development and testing.
• Include AI in your annual security risk analysis.
• Add AI scenarios to incident response.

FlashTower flags any vendor-hosted AI system with PHI access and no recorded BAA as Critical, and maps findings to HIPAA Security Rule references to help organize remediation. FlashTower does not determine HIPAA compliance. See FlashTower for healthcare.

AI inside systems you already use

Much AI in healthcare arrives as features in existing products: EHR summarization, ambient documentation, patient-messaging assistants, revenue-cycle automation, and call-center tools. Existing BAAs with those vendors may or may not extend to new AI features, particularly when a third-party model provider is involved. When a vendor announces an AI feature, ask what data it sends, to whom, whether the BAA covers it, and whether it can be disabled until reviewed.

Patient-facing AI

Chatbots and messaging assistants that interact with patients raise additional considerations beyond HIPAA: accuracy of clinical information, escalation to clinicians, accessibility, and state laws on AI disclosure in healthcare communications. Require human escalation paths, clear disclosure that the patient is interacting with an automated system where appropriate, and clinical review of content.

Workforce guidance that clinicians follow

Clinicians adopt AI to save documentation time, so guidance must offer a faster approved path, not only prohibitions. Publish a one-page list of approved tools for PHI, what each may be used for, and what is never permitted — for example, entering PHI into any personal account or unlisted app. Include two or three realistic examples (drafting a patient letter, summarizing a referral) showing the approved way to do each task.