Written and maintained by CASRAI Editorial Board
Last updated
Regulatory & Clinical Governance Notice: This reference explains federal health privacy regulations, cryptographic security controls, and clinical documentation workflows for administrative and health informatics evaluation. It does not constitute legal counsel or patient-specific clinical advice.
Healthcare privacy officers, clinical informatics architects, and behavioral health practice administrators evaluating a HIPAA-compliant AI medical scribe or HIPAA-compliant AI note taker must verify whether an ambient transcription and large language model (LLM) summarization platform satisfies both contractual requirements under 45 CFR §§ 164.502(e) and 164.504(e) and technical security controls under the HIPAA Security Rule (45 CFR § 164.312). Because ambient clinical documentation captures raw multi-speaker patient-clinician audio containing electronic Protected Health Information (ePHI), deploying an unvetted consumer note taker or failing to distinguish medical progress notes from protected psychotherapy notes creates immediate regulatory exposure.
What Is a HIPAA-Compliant AI Medical Scribe in Healthcare Workflows?
A HIPAA-compliant AI medical scribe is an ambient clinical intelligence system that captures patient-clinician encounters, performs medical automatic speech recognition (ASR) and speaker diarization, and synthesizes structured clinical documentation—such as a SOAP note—under a countersigned Business Associate Agreement (BAA) and enforced technical safeguards. Unlike general consumer transcription apps, a healthcare-grade AI medical scribe operates within a zero-training data boundary where patient audio and transcripts are never used to train foundational models accessible to external tenants.
In clinical and research workflows, the compliance boundary spans four distinct lifecycle stages:
- Pre-Encounter Notice and Consent: Capturing verbal or written patient consent (and complying with state-specific two-party wiretapping/recording statutes) before activating microphone capture.
- Encrypted Audio Streaming: Streaming audio over Transport Layer Security (TLS 1.2/1.3) to an ephemeral memory buffer or encrypted cloud storage volume (AES-256 at rest).
- Zero-Data-Retention (ZDR) Inference: Executing ASR transcription and LLM clinical note generation under contractual zero data retention or automated short-window audio purging immediately after clinician sign-off.
- Human-in-the-Loop EHR Attestation: Staging the draft note into the Electronic Health Record (EHR) so the licensed clinician reviews, edits, and authenticates the entry prior to permanent medical record commitment.
What Security Controls Are Required for a HIPAA-Compliant AI Medical Scribe?
Marketing claims that an AI note taker is “HIPAA compliant” have no legal standing without both a binding contract and verifiable engineering controls. Under 45 CFR Part 164, covered entities must verify six mandatory compliance pillars before transmitting clinical audio or transcripts:
| HIPAA Citation / Standard | Regulatory Requirement | AI Medical Scribe Engineering Implementation |
|---|---|---|
| 45 CFR §§ 164.502(e), 164.504(e) Business Associate Contracts |
Written contract binding the vendor and all downstream sub-processors before ePHI is created, received, maintained, or transmitted. | Executed BAA covering the AI scribe vendor and chained BAAs with underlying cloud ASR/LLM hosts (e.g., AWS Bedrock, Azure OpenAI, Google Vertex AI). |
| 45 CFR § 164.312(a)(1) Access Control |
Unique user identification, emergency access procedures, automatic logoff, and encryption/decryption. | Role-Based Access Control (RBAC), SAML 2.0 / OIDC Single Sign-On (SSO), session inactivity timeouts, and AES-256 encryption at rest. |
| 45 CFR § 164.312(b) Audit Controls |
Hardware, software, and procedural mechanisms that record and examine activity in information systems containing ePHI. | Immutable audit logs recording every audio stream creation, transcript view, LLM regeneration, EHR push, and audio deletion event with user ID and UTC timestamp. |
| 45 CFR § 164.312(c)(1) & (e)(1) Integrity & Transmission Security |
Protect ePHI from improper alteration/destruction and guard against unauthorized access during electronic transmission. | TLS 1.2+ / TLS 1.3 encryption in transit, cryptographic payload checksums, and isolated Virtual Private Cloud (VPC) inference endpoints. |
| 45 CFR § 164.502(b) Minimum Necessary & Model Training |
Prohibits unauthorized secondary uses or disclosures of ePHI outside Treatment, Payment, and Healthcare Operations (TPO). | Contractual prohibition against using customer audio, transcripts, or clinician edits to train or fine-tune shared/public foundation models without de-identification under 45 CFR § 164.514. |
| ONC HTI-1 (45 CFR § 170.315(b)(11)) Decision Support Transparency |
Certified health IT modules supplying predictive Decision Support Interventions (DSI) must expose model source attributes and governance risk mitigations. | Traceability linking generated assessment/plan statements back to source transcript segments so clinicians can detect hallucinations or omissions. |
Is Otter.ai, Fireflies.ai, or Read AI HIPAA Compliant for Clinical Notes?
Clinicians and therapists frequently search whether general-purpose corporate meeting assistants—such as Otter.ai, Fireflies.ai, or Read AI—can be used as a HIPAA-compliant AI note taker. General-purpose meeting bots differ fundamentally in architecture and default privacy posture from dedicated healthcare ambient scribes:
| Platform Category & Examples | Default / Pro / Business Tiers | Enterprise Tier BAA Availability | Operational & Clinical Workflow Risks |
|---|---|---|---|
| Otter.ai (General Meeting Transcription) |
Not HIPAA compliant. Basic, Pro, and standard Business plans do not offer a BAA and cannot legally process PHI. | Otter offers a BAA exclusively on custom Otter Enterprise plans. | Designed for corporate calendar auto-join and persistent searchable audio archives; default settings email meeting summaries to participants unless restricted by enterprise admin policy. |
| Fireflies.ai (General Meeting Assistant) |
Not HIPAA compliant on Free, Pro, or standard Business self-serve tiers without an executed BAA. | Fireflies offers a HIPAA BAA only on the Enterprise tier with HIPAA mode enabled (private storage, zero model training). | Requires disabling calendar bot auto-join on unconsented telehealth links and disabling public link-sharing of transcripts. |
| Read AI (Meeting Analytics & Summaries) |
Not HIPAA compliant on Free, Pro, or standard Enterprise tiers without a custom healthcare addendum. | Offers HIPAA BAA only via custom Enterprise security agreements with workspace retention lockdowns. | Captures video/engagement telemetry alongside audio; unsuitable for clinical encounters unless video telemetry and persistent storage are disabled under a signed BAA. |
| Dedicated Clinical AI Scribes (e.g., Abridge, DAX Copilot, Suki, Freed, Heidi) |
Built specifically for clinical encounters; BAAs offered as standard healthcare contracting (verify tier before use). | Standard executed BAA covering ASR, clinical LLM synthesis, and EHR integration pipelines. | Supports ephemeral audio streaming, medical vocabulary/ICD-10/SNOMED mapping, and structured EHR write-back without calendar bot auto-join hazards. |
As detailed in our guide to using AI with PHI in research under a HIPAA BAA, using a consumer or self-serve team tier of Otter.ai, Fireflies.ai, or Read AI to transcribe patient encounters constitutes an unauthorized disclosure of ePHI to an uncontracted vendor, even if the vendor uses TLS encryption.
HIPAA-Compliant AI Therapy Notes: Progress Notes vs. Psychotherapy Notes (45 CFR § 164.508)
Behavioral health clinicians searching for a HIPAA-compliant AI note taker for therapists or HIPAA-compliant AI therapy notes face an additional statutory requirement under the HIPAA Privacy Rule. Federal regulations draw a strict legal boundary between routine behavioral health progress notes and psychotherapy notes:
- What Qualifies as Psychotherapy Notes (45 CFR § 164.501): Notes recorded in any medium by a mental health professional documenting or analyzing the contents of conversation during a private counseling session (individual, group, joint, or family) and that are kept separate from the rest of the individual’s medical record.
- What Is Excluded from Psychotherapy Notes (Standard Progress Notes): Under 45 CFR § 164.501, the definition explicitly excludes medication prescription and monitoring, counseling session start and stop times, the modalities and frequencies of treatment furnished, results of clinical tests, and any summary of diagnosis, functional status, treatment plan, symptoms, prognosis, and progress to date.
- Heightened Authorization Rule (45 CFR § 164.508(a)(2)): Unlike routine clinical progress notes, which may be used and disclosed for standard Treatment, Payment, and Healthcare Operations (TPO), uses and disclosures of psychotherapy notes require a standalone, specific patient authorization except for narrow originator use, supervised mental health training, or legal defense by the covered entity.
Consequently, when a therapist uses an ambient AI scribe during a counseling session:
- Audio Stream Sensitivity: Even if the final output pushed to the behavioral health EHR is a standard clinical progress note (documenting modality, start/stop times, symptoms, and treatment plan), the raw audio recording and verbatim transcript capture the full intimate contents of the counseling conversation.
- Ephemeral Processing Mandate: Behavioral health practices should configure their AI scribe to operate in ephemeral audio mode—streaming audio in memory to generate the structured progress note and immediately purging both raw audio and verbatim counseling transcripts upon clinician sign-off—or store separated psychotherapy notes in a dedicated, access-restricted vault governed by explicit patient consent.
How Does a HIPAA-Compliant AI Medical Scribe Integrate with Electronic Health Records?
Enterprise AI scribes integrate with hospital and ambulatory EHRs through HL7® FHIR® R4 and SMART on FHIR APIs rather than clipboard copy-paste:
- Encounter Context Retrieval: Using SMART on FHIR launch scopes (`launch`, `patient/Patient.read`, `patient/Encounter.read`), the scribe mobile or desktop client binds the audio stream to the exact scheduled `Encounter` and `Patient` identifier, preventing wrong-chart documentation errors.
- Structured Note Staging (`DocumentReference` & `Composition`): Once the clinician reviews the generated history of present illness (HPI), physical exam, and assessment/plan, the scribe writes the draft text back to the EHR via `POST /DocumentReference` or vendor-specific ambient note staging APIs with `docStatus: “preliminary”`.
- Audit Trail and Authentication: The note remains in an unauthenticated draft state inside the EHR until the attending clinician or therapist inspects the text, resolves any AI-generated discrepancies, and applies an electronic signature (`docStatus: “final”`).
How Should Ambient AI Scribes Handle Audio Retention and Patient Consent?
To minimize breach liability under the HIPAA Breach Notification Rule (45 CFR §§ 164.400–414) and state biometric/wiretapping laws, healthcare organizations enforce a five-point governance checklist for audio retention and patient consent:
- Explicit Pre-Recording Patient Consent: Notify the patient before starting recording that an ambient AI medical scribe is transcribing the visit, explain that no audio is retained for commercial model training, and offer an immediate opt-out with zero impact on clinical care.
- Pause/Resume Controls for Sensitive Examinations: Train clinicians to pause ambient recording during sensitive disclosures or third-party family discussions when requested by the patient.
- Zero Persistent Audio Storage (or ≤7–30 Day Auto-Purge): Configure the enterprise workspace so raw audio files are deleted immediately upon note finalization (or purged automatically within a strict 24-hour to 7-day window if temporary replay is required for delayed charting).
- De-Identification Verification Before Model Fine-Tuning: If an enterprise health system permits vendor fine-tuning on specialty templates, require formal Expert Determination or Safe Harbor de-identification under 45 CFR § 164.514(b)—noting that raw voiceprints constitute biometric identifiers (`§ 164.514(b)(2)(i)(P)`) and can never be treated as de-identified audio.
Frequently Asked Questions
What is a HIPAA-compliant AI medical scribe in healthcare workflows?
A HIPAA-compliant AI medical scribe is an ambient speech-to-text and clinical LLM platform governed by an executed Business Associate Agreement (BAA) under 45 CFR § 164.504(e) and technical safeguards under 45 CFR § 164.312. It captures patient-clinician dialogue, generates structured draft notes (such as SOAP notes), prohibits foundation model training on ePHI, and stages notes for clinician review and authentication in the EHR.
How does a HIPAA-compliant AI medical scribe integrate with electronic health records?
It integrates via HL7 FHIR R4 and SMART on FHIR APIs. The application launches within the EHR session to bind `Patient` and `Encounter` context, processes audio within an encrypted zero-data-retention pipeline, and writes the structured draft back via `DocumentReference` or `Composition` resources with `preliminary` status until the clinician signs the note.
What security controls are required for a HIPAA-compliant AI medical scribe?
Required controls include an executed BAA covering the vendor and all cloud AI sub-processors, AES-256 encryption at rest, TLS 1.2/1.3 encryption in transit, SAML 2.0/OIDC SSO with automatic session logoff, immutable audit logs (45 CFR § 164.312(b)), contractual zero model training on ePHI, and ephemeral or auto-purged audio retention policies.
Is Otter.ai or Fireflies.ai HIPAA compliant for clinical notes?
Not on Free, Pro, or standard self-serve Business tiers. Both Otter.ai and Fireflies.ai offer a HIPAA Business Associate Agreement (BAA) exclusively on custom Enterprise tiers with HIPAA security controls enabled. Even on Enterprise tiers, healthcare organizations must disable automatic calendar-bot meeting joins and restrict persistent audio/transcript sharing.
How should ambient AI scribes handle audio retention and patient consent?
Clinicians must obtain informed patient consent before activating microphone capture (honoring two-party recording states), provide an opt-out option, and configure the AI scribe to purge raw audio recordings immediately after note authentication (ephemeral processing) so biometric voice data is not stored indefinitely.
