AI and HIPAA: What Every Healthcare CIO Needs to Know

It starts with the best of intentions. A physician, facing hours of documentation at the end of an already long shift, discovers that ChatGPT can draft clinical notes from conversational bullet points in seconds. A nurse coordinator uses Claude to summarise complex care plans for patient handoffs. A medical coder uses an AI tool to interpret ambiguous documentation and suggest billing codes more quickly than any reference manual allows.

In every one of these scenarios, protected health information is leaving your organisation and entering an AI system that almost certainly has no Business Associate Agreement with you, processes data on infrastructure you cannot audit, and may use that data to train future model versions. The Office for Civil Rights at HHS has been explicit: submitting PHI to an AI provider without a compliant BAA is a HIPAA violation. The enforcement clock is ticking.

The PHI Problem in AI Prompts

HIPAA defines Protected Health Information broadly: any individually identifiable health information maintained or transmitted in any form that relates to the past, present, or future physical or mental health condition of an individual. The 18 HIPAA safe harbour identifiers include names, dates, geographic identifiers, telephone numbers, and — most relevant in clinical AI contexts — "any other unique identifying number, characteristic, or code."

Clinical AI prompts almost always contain PHI. Not because clinicians are careless, but because the entire value proposition of AI in clinical workflows depends on giving the model context about a specific patient's situation. You cannot ask an AI to help draft a medication reconciliation note without providing the patient's medication list. You cannot ask it to summarise a patient's care trajectory without providing the relevant clinical history. The PHI is the context — and without the context, the AI assistance is generic and clinically useless.

This is not a problem that can be solved by telling clinicians to anonymise before pasting into AI tools. Anonymisation is complex, time-consuming, and error-prone under clinical workflow conditions. A consistent de-identification process that reliably meets HIPAA's Safe Harbor standard requires technical controls, not individual discipline.

The HIPAA BAA Gap

OpenAI, Anthropic, and Google all offer enterprise agreements that include Data Processing Agreements for their business-tier API products. These are distinct from the consumer-facing ChatGPT, Claude.ai, and Gemini interfaces, which do not include BAAs and explicitly disclaim HIPAA compliance in their terms of service.

The critical point is that most healthcare staff who are using AI tools in their workflows are using the consumer interfaces — not the enterprise API products with BAAs in place. A clinician who downloads Claude.ai or visits ChatGPT.com to help with clinical documentation is using a product that has no HIPAA coverage, regardless of whether their organisation has an enterprise agreement with the same vendor.

OCR Enforcement Risk

The HHS Office for Civil Rights has signalled clearly that AI is on its enforcement agenda. The 2024 OCR Annual Report included a dedicated section on AI and HIPAA risk for the first time. The 2025 OCR guidance on Remote Communication Technologies, which addressed telehealth security, explicitly extended to AI-assisted clinical communication tools. Healthcare organisations operating in 2026 are operating in an environment where AI-related HIPAA enforcement actions are a matter of when, not if.

The enforcement exposure is not limited to headline data breach incidents. Routine use of non-BAA-covered AI tools by clinical staff, even without any identifiable adverse outcome, constitutes a violation of the HIPAA Security Rule's technical safeguards requirements if PHI is transmitted without appropriate security controls in place. Under the HIPAA audit programme, these violations are discoverable in the normal course of compliance review.

Technical Safeguards Mapping

The HIPAA Security Rule's Technical Safeguards (45 CFR § 164.312) require specific controls for ePHI. Here's how an AI DLP solution addresses each requirement:

The Five-Step AI HIPAA Compliance Programme

1. Inventory AI Tool Usage

   Use DNS log analysis, proxy inspection, and browser extension telemetry to discover which AI tools clinical and administrative staff are currently using. This baseline is essential — you cannot govern what you don't know exists.

2. Classify Each Tool by BAA and Data Classification Status

   For every AI tool discovered: does a HIPAA BAA exist for the tier your staff are using? What data classification is permitted? Build a clear sanctioned / conditional / blocked list with the specific conditions of each.

3. Deploy AI DLP Scanning at the Traffic Layer

   Implement prompt-level PHI detection for all AI traffic — identifying and blocking prompts containing PHI destined for non-BAA-covered services, while permitting appropriately protected traffic to BAA-covered enterprise services.

4. Establish Session-Level Audit Logging

   Configure complete logging of AI interactions for all clinical staff using AI-assisted tools. This logging is the audit control that HIPAA's Technical Safeguards require, and it's the foundation for any incident investigation.

5. Update BA Agreement Register and Acceptable Use Policy

   Formalise the AI-specific additions to your Business Associate Agreement register and acceptable use policy. Document which AI tools are approved for which use cases, the conditions of approval, and the review cycle for each.