AI and HIPAA Business Associate Agreements: What Must Be in Place

Healthcare organizations adopting AI face a question that consumer AI tools were not designed to answer: who is responsible for protected health information once it reaches a model? The answer is that any vendor or model provider that receives PHI on the organization's behalf is a business associate, and a business associate agreement must exist before the first record flows, with configuration and safeguards that match it. This guide covers when a BAA is required, what it must cover, and the mistakes that create violations, drawing on FISTA Solutions' AI enablement practice. The safety context is in the AI safety in healthcare operations whitepaper and the hospital setting in ai in hospitals. This article is general guidance, not legal advice; organizations should confirm obligations with counsel and privacy officers.

Which AI vendors are business associates?

Vendor type	Touches PHI when	BAA required
Model providers via API	Prompts or documents containing PHI are sent	Yes, on a tier that offers one
AI application vendors	The application processes patient records	Yes
Transcription and documentation tools	Clinical conversations or notes are processed	Yes
Infrastructure hosting AI components	PHI is stored or processed on it	Yes, typically already in place
Analytics on de-identified data	Data meets the de-identification standard	No, if standard is met

Provider selection with BAA availability in mind is in how to choose an llm provider.

What must the BAA cover?

Permitted and required uses and disclosures of PHI; a prohibition on other uses, including training or improving models; appropriate administrative, physical, and technical safeguards; reporting of breaches and security incidents within defined timelines; flow-down of the same obligations to subcontractors that touch PHI, including any model providers behind an application vendor; support for individual rights where the vendor holds records; return or destruction of PHI at termination; and compliance with applicable rules. Vendor review practice is in ai third party risk management.

Why does the service tier and configuration matter?

Providers offer BAAs on specific tiers, often enterprise or healthcare offerings, with settings that disable training on inputs, limit retention, and support audit logging. A BAA signed for one tier does not cover usage through another, and a covered tier used with retention or training settings that contradict the BAA is a gap. Verify the tier, the settings, and the region before PHI flows, and monitor that they remain in place. Questionnaire practice is in the ai vendor security questionnaire.

What safeguards must the covered entity still apply?

The BAA governs the vendor; the covered entity's own program governs its AI components: risk analysis including AI-specific threats, minimum necessary PHI in prompts and retrieval, access controls so assistants read only what the user may, encryption of indexes and logs, activity logging with redaction, workforce training, and incident procedures. A compliant vendor behind a leaky application is still a breach. Access design is in ai access control and leakage controls in ai data leakage prevention.

How does de-identification fit?

Data de-identified to the applicable standard is not PHI, and many analytics, model development, and evaluation uses can proceed on it without a BAA. De-identification must actually meet the standard, and re-identification risk in AI contexts, where models can infer identity from combinations, should be assessed. Do not treat removal of obvious identifiers as de-identification. Privacy assessment practice is in the ai privacy impact assessment checklist.

What happens when a vendor has an incident?

The BAA's reporting obligations trigger; the covered entity assesses whether PHI was compromised, which requires knowing what PHI was sent and retained, which requires lineage and logs; notification obligations follow the assessment. AI incidents such as leakage through outputs or logs are within scope. Practice is in ai incident disclosure and trail design in how to build an ai audit trail.

What records should be kept?

The inventory of AI vendors and the PHI each receives; executed BAAs and the tier each covers; configuration evidence for training, retention, and logging settings; risk analyses; access and activity logs; workforce training records; and incident records. Record practice is in ai record keeping requirements.

What mistakes create violations?

Clinicians using consumer AI tools with patient information; BAAs assumed because the vendor's website says compliant; the wrong tier or settings in use; application vendors without flow-down to their model providers; PHI in logs the entity controls; and de-identification that does not meet the standard. Each is common and each is avoidable with inventory, contracts, and configuration checks.

What does compliant practice look like?

A telehealth provider deploying a documentation assistant inventories PHI flows, selects a model provider tier with a BAA, verifies training and retention settings, signs the BAA before any PHI is processed, applies minimum necessary and access controls in its application, encrypts and redacts logs, trains clinicians, and documents the risk analysis. Configuration is re-verified quarterly and after provider changes. The setting is in ai in telehealth.

How FISTA Solutions helps with HIPAA and AI

FISTA Solutions builds healthcare AI systems with minimum necessary data flows, permission-aware retrieval, encrypted and redacted logging, and audit trails, and helps clients select BAA-covered tiers, verify configuration, and extend risk analyses to AI. The AI enablement practice leads privacy-by-design, AI agents ship with the safeguards, and forward deployed engineers embed with client compliance and clinical informatics teams. The record behind the approach is 150+ projects with 99.9% uptime.

To put AI on PHI with the agreements and safeguards in place, message FISTA on WhatsApp, or read the AI safety in healthcare operations whitepaper for the wider safety program.