HIPAA Compliance by Vertical
HIPAA Compliance for AI Scribe and Ambient Documentation Vendors
HIPAA for AI scribe and ambient documentation vendors: encounter audio, model-provider subcontractors, retention, training data, and hospital security reviews.
Key takeaways
- PHI in scope: Audio recordings of clinical encounters, Transcripts and generated clinical notes, Patient and encounter context pulled from the EHR, Clinician edits and feedback on generated notes.
- Start with these controls: A subcontractor BAA with every speech-to-text and model provider that receives PHI; Documented retention and deletion for audio, transcripts, and generated notes; A written position on model training, permitted by the BAA or limited to de-identified data.
- The most common mistake: A model or transcription provider in production without a BAA, or with a BAA that does not cover the service used
- A signed BAA is required with every vendor that creates, receives, maintains, or transmits PHI on your behalf, and no tool is HIPAA compliant on its own.
PHI handled in this environment
- Audio recordings of clinical encounters
- Transcripts and generated clinical notes
- Patient and encounter context pulled from the EHR
- Clinician edits and feedback on generated notes
- Logs, evaluation sets, and any data used to improve models
Common compliance pain points
- Model providers are subcontractors: If your product sends encounter audio or transcripts to a speech-to-text or language-model provider, that provider is your subcontractor business associate and needs a BAA with you. Check which of the provider's services its BAA actually covers, because provider BAAs often list specific eligible services.
- Retention of audio and transcripts: Every recording and transcript is PHI. Decide how long you keep audio after the note is signed, document the decision, and make sure deletion reaches backups, logs, and vendor copies.
- Using customer data to improve models: A business associate may use PHI only as its BAA permits. Using encounter data to train or evaluate models has to be allowed by the BAA, or the data has to be de-identified under 45 CFR 164.514 first.
- Hospital security reviews: Health systems commonly run a security review before an AI scribe rollout. Expect questions about your risk analysis, subcontractors, data retention, access controls, incident response, and whether customer data trains models, and often a request for a SOC 2 report.
Priority controls
- A subcontractor BAA with every speech-to-text and model provider that receives PHI
- Documented retention and deletion for audio, transcripts, and generated notes
- A written position on model training, permitted by the BAA or limited to de-identified data
- A HIPAA Security Rule risk analysis covering the full audio-to-note pipeline
- Access controls and audit logs on every system that holds encounter data
- SOC 2 Type II if health system customers require it
Common mistakes
- A model or transcription provider in production without a BAA, or with a BAA that does not cover the service used
- Encounter audio kept indefinitely after the note is signed
- Training models on customer encounters without that use being permitted by the BAA
- Raw transcripts landing in application logs or error tracking
- No risk analysis covering the audio pipeline when a health system asks for one
Enforcement context
Since the 2013 Omnibus Rule, business associates are directly liable for Security Rule compliance and for uses and disclosures their BAAs do not permit, which includes using encounter data in ways the BAA does not allow. HHS's January 2025 Security Rule proposal (90 FR 898) would, if finalized, require business associates to verify their technical safeguards to covered entities at least once every 12 months; it is a proposal, not current law.
Frequently asked questions
What PHI do AI scribe organizations handle?
Audio recordings of clinical encounters; Transcripts and generated clinical notes; Patient and encounter context pulled from the EHR; Clinician edits and feedback on generated notes; Logs, evaluation sets, and any data used to improve models. Any of these, combined with an identifier, is protected health information when a covered entity or business associate holds it, and in electronic form it brings the Security Rule into scope.
What are the priority HIPAA controls for AI scribe?
A subcontractor BAA with every speech-to-text and model provider that receives PHI; Documented retention and deletion for audio, transcripts, and generated notes; A written position on model training, permitted by the BAA or limited to de-identified data; A HIPAA Security Rule risk analysis covering the full audio-to-note pipeline; Access controls and audit logs on every system that holds encounter data; SOC 2 Type II if health system customers require it.
What are the most common HIPAA mistakes in AI scribe?
A model or transcription provider in production without a BAA, or with a BAA that does not cover the service used; Encounter audio kept indefinitely after the note is signed; Training models on customer encounters without that use being permitted by the BAA; Raw transcripts landing in application logs or error tracking; No risk analysis covering the audio pipeline when a health system asks for one.
What does HIPAA enforcement look like for AI scribe?
Since the 2013 Omnibus Rule, business associates are directly liable for Security Rule compliance and for uses and disclosures their BAAs do not permit, which includes using encounter data in ways the BAA does not allow. HHS's January 2025 Security Rule proposal (90 FR 898) would, if finalized, require business associates to verify their technical safeguards to covered entities at least once every 12 months; it is a proposal, not current law.