Can an AI System Pass a HIPAA Audit?

Why healthcare teams keep asking this question

AI vendors market to healthcare practices constantly, and the compliance question comes up in almost every sales conversation. The problem is that most vendors answer it vaguely, pointing to a SOC 2 badge or a privacy policy and calling it good enough.

HIPAA doesn't work that way. It's a framework built around covered entities and their business associates, not software products. When an audit happens, the auditor is looking at how your organization handles protected health information, not whether your AI vendor has a clean landing page.

What a HIPAA audit actually examines

HIPAA audits review three rule sets: the Privacy Rule, the Security Rule, and the Breach Notification Rule. For an AI deployment, the Security Rule does most of the work. Auditors want to see a signed Business Associate Agreement with every vendor that touches PHI, access controls that limit who can query the system, encryption at rest and in transit, and audit logs showing who accessed what and when.

If your AI system runs through a public API like OpenAI's standard tier, you almost certainly have a problem. Those endpoints process data on shared infrastructure, and most standard API agreements don't constitute a valid BAA. A private deployment, whether self-hosted or in a dedicated cloud environment, gives you the isolation you need to satisfy the Security Rule's technical safeguards.

The model itself, whether it's Llama 3.1, a fine-tuned clinical variant, or something else, doesn't get a HIPAA certification. What gets you through the audit is the architecture around it: network segmentation, role-based access, encrypted storage, a valid BAA, and documented workforce training. Get those right and the AI component is just another system in your environment.

When the answer gets more complicated

If your AI system touches PHI in real time, such as reading patient records from Epic or transcribing intake calls, the bar gets higher. You'll need to document a full risk analysis under 45 CFR 164.308(a)(1), not just point to a vendor's compliance page. Multi-agent systems that pass PHI between components need BAAs evaluated at each handoff point, not just at the front door.

Smaller practices sometimes believe that de-identified data puts them outside HIPAA's reach. That's partially true, but de-identification under the Safe Harbor or Expert Determination methods is stricter than most teams realize. If there's any doubt about whether your data qualifies as truly de-identified, treat it as PHI and build accordingly.