The question every health system is asking
A clinician wants to paste a patient note into an AI tool to draft a summary. Compliance wants to know one thing first: is this AI HIPAA compliant?
The honest answer is that no model is "HIPAA compliant" on its own. Compliance is a property of the whole deployment — where the data goes, who can see it, and what is logged.
What HIPAA actually requires of an AI system
HIPAA cares about protected health information: where it is stored, how it moves, and whether you can prove control. For an AI system that means a few concrete things.
You need a Business Associate Agreement with anyone who touches PHI. You need access controls and an audit trail. And you need assurance the data isn't used to train someone else's model.
So "which AI is HIPAA compliant" is the wrong question. The right one is: under what deployment can this AI handle PHI safely?
The limits of a BAA with a cloud vendor
Major vendors will sign a BAA for their enterprise tiers, and that genuinely matters. But a BAA is a contract — a promise about behavior, backed by penalties, not a guarantee about architecture.
Is Claude AI HIPAA compliant? Is Gemini? On a covered enterprise plan with a signed BAA, those vendors support HIPAA workloads. The consumer apps are a different story, and that gap is where most real exposure happens.
The risk isn't usually the enterprise contract. It's the staff member pasting PHI into a free chatbot that was never in scope.
Why architecture beats assurance
A BAA tells you a vendor promised not to misuse your data. An air-gapped or on-premise deployment makes the promise unnecessary, because the PHI never leaves your infrastructure in the first place.
That is the difference between a policy and a guarantee. Open models — Llama, Mistral, and peers — now handle clinical text well enough that you no longer trade capability for control.
This is the basis for HIPAA-aligned AI for hospitals that you own: hipaa compliant generative AI running on your servers, where PHI stays inside your perimeter and every action is logged to your own audit trail.
What clinicians actually use it for
Capability is no longer the blocker, so the use cases are practical:
- Clinical Documentation Agent — drafts notes and summaries from the encounter, with the text staying inside your environment.
- Medical Coding Agent — assigns ICD-10 and CPT codes and flags claim issues before they cause denials.
- Prior Authorization Agent — assembles auth requests against payer rules and tracks status.
- Clinical Support Agent — surfaces evidence and drug-interaction checks grounded in your own protocols.
Each runs against your EHR through connectors rather than shipping a copy of patient data somewhere else.
Where to start
Pick one workflow with clear value and low risk — internal protocol search or coding support is a common first step — and run it on-premise against a single service line.
Prove the security model and the output quality on real charts before expanding. The point is to use AI on terms that survive an OCR audit, not to adopt it everywhere at once.
Checking a specific model
Evaluating a particular vendor against HIPAA? Each of the major models carries a BAA on some routes and not others — the per-model breakdowns walk through exactly which:
- Is ChatGPT HIPAA compliant? — OpenAI's BAA, ChatGPT Enterprise and Edu, and where PHI still leaks.
- Is Claude HIPAA compliant? — Anthropic's BAA, Bedrock and Vertex routes, and the audit-trail gap.
- Is Gemini HIPAA compliant? — which Google Cloud and Workspace routes carry a BAA, and which do not.