Compliance & Privacy

Is Ambient AI Transcription HIPAA-Safe? BAAs, Consent Scripts, and State Rules

A practice owner's guide to evaluating whether an ambient AI scribe is HIPAA compliant — what a BAA must cover, how to word session consent, and how state recording-consent laws apply.

Is Ambient AI Transcription HIPAA-Safe? BAAs, Consent Scripts, and State Rules

Quick answer (for featured snippet)

Ambient AI transcription is HIPAA-safe only when the vendor signs a BAA that names the transcription and inference steps specifically, encrypts audio and text in transit and at rest, and gives you a documented retention and deletion policy. Separately, you need patient consent to record every session — this is a clinical and ethical requirement, not just a state law question. Treat every session as requiring explicit consent, whether your state is one-party or two-party for recording law.

This post is part of our complete guide to HIPAA-Compliant AI for Behavioral Health Practices.


Why "ambient" changes the compliance picture

Ambient AI scribes listen through a session and produce a note afterward, instead of requiring you to dictate or type. That convenience is the whole selling point — but it also means raw audio of a clinical conversation now exists somewhere outside the room, even briefly, before it becomes text. That's a different risk profile than typing your own notes, and it's worth treating it as one.

Two questions matter, and they're easy to conflate:

  1. Is the vendor's handling of that audio HIPAA compliant? (A business associate question.)
  2. Do you have the patient's permission to record them at all? (A consent and clinical ethics question, layered under state recording law.)

Vendors that only answer the first question and skip the second are giving you half the picture.

What a real BAA covers for an ambient scribe

A signed Business Associate Agreement is necessary but not sufficient. Ambient transcription typically involves more processing steps than a text-based tool, so the BAA needs to name each one:

  • Audio capture — how the session audio reaches the vendor (browser mic capture, a companion app, a phone line) and whether it's encrypted the moment it leaves your device.
  • Transcription/inference — whether the vendor runs its own model or calls a third-party API. If it's a third party, that provider needs to be a subcontracted business associate under the same BAA chain, not a silent dependency.
  • Storage — where raw audio and the resulting transcript live, for how long, and whether audio is deleted after the note is generated (many vendors discard audio once transcription completes — ask if yours does).
  • Model training — whether your session data is used to improve the vendor's models, and whether you can opt out. This should be a stated policy, not something you infer from silence.

If a vendor's BAA is a generic template that doesn't distinguish audio from text, ask directly which of the above it covers. A vague answer is itself the answer.

Consent: separate from the legal recording question

Every state has a law governing whether you need one party's consent (yours) or all parties' consent (yours and the patient's) to record a conversation. That statute matters for legal exposure, but it is not the compliance bar for clinical practice.

Clinical documentation standards and most licensing boards expect informed consent before recording a session for any purpose — including AI transcription — independent of what your state's wiretapping law technically requires. Practically:

  • Don't rely on "my state is one-party." Get explicit consent anyway. It protects the therapeutic relationship and gives you a defensible record if a patient later disputes that a session was recorded.
  • Document the consent, not just the recording. A checkbox in your EHR or a signed intake line is worth more than an assumption that a general HIPAA notice covers it.
  • Re-confirm for telehealth. A patient joining from a shared space or a different state may have different expectations; a quick verbal confirmation at the start of a session costs a few seconds.

A short verbal consent script clinicians can adapt:

"Before we start, I want to let you know I use an AI tool to transcribe and draft notes from our session. The audio is processed securely and I review everything before it goes in your chart. You're welcome to say no — I'll take notes by hand instead, and it won't change anything about your care. Are you comfortable with that?"

Keep a written version in your intake paperwork for the same reason you keep any other informed-consent documentation: it's the artifact that shows the conversation happened, not just that a policy exists.

A concrete workflow example

A solo therapist running a mixed telehealth/in-person caseload wants to start using an ambient scribe. Before the first session:

  1. Adds one line to the intake packet: AI-assisted documentation disclosure, with an opt-out box.
  2. Confirms with the vendor, in writing, that the BAA names both the transcription step and any third-party model provider, and that raw audio is deleted after note generation.
  3. Builds the verbal script above into the first two minutes of every new-patient session.
  4. Sets up a fallback: for patients who decline, the note-taking workflow reverts to a structured typed template — no separate software, just skipping the ambient capture.

Four sessions in, one patient declines. The therapist takes handwritten notes during the session and types them into the EHR afterward, using the same note structure. Nothing about scheduling or care changes — only the source of the note.

What to check before you sign up

  • Ask for the BAA before your first paid session, not after — a vendor that stalls here is telling you something.
  • Confirm whether audio is retained after transcription, and for how long if so.
  • Build the consent line into your intake process on day one, not as a later add-on.
  • Keep a manual fallback documented so declining patients don't create a workflow gap.

Ambient transcription doesn't have to be a compliance gamble. The vendor's BAA answers whether the technology is handled safely; your consent process answers whether you have permission to use it. Both need to be in place before the first recorded session — not verified after the fact.

PsyFi Notes processes session audio under a signed BAA, deletes raw audio automatically within 30 days of transcription (or immediately when you delete the session), never uses session data for model training, and is built around the consent-first workflow described above. For the fuller compliance picture beyond ambient scribes, see HIPAA-Compliant AI for Behavioral Health Practices and Consent & Liability: Template Language for AI Intake.


PsyFi Notes is HIPAA-aligned clinical documentation software backed by a Business Associate Agreement. For questions about BAAs and consent for ambient transcription in your practice, contact our onboarding team at [email protected].

Frequently asked questions

Is ambient AI transcription HIPAA compliant?
It can be, but compliance depends on the vendor, not the technology. You need a signed BAA covering the transcription and inference steps, encryption in transit and at rest, and a documented retention/deletion policy — not just a marketing claim of 'HIPAA compliant.'
Do I need patient consent to record a therapy session for AI transcription?
Yes, always — regardless of your state's recording-consent law. Clinical ethics and most licensing boards require informed consent before recording, separate from any wiretapping statute.
What's the difference between one-party and two-party consent states for ambient scribes?
One-party states legally allow recording with only the clinician's consent; two-party (all-party) states require every participant to agree. Clinically, treat every session as if two-party consent applies and get explicit patient agreement either way.
What should a BAA for an ambient scribe explicitly cover?
The audio capture step, the transcription/inference step (including any third-party model providers), storage location, retention period, and whether session data is used to train models. If any of those steps aren't named, ask why.
Can a patient decline ambient AI recording and still get care?
Yes. Every ambient scribe workflow needs a manual fallback — typed or dictated notes — for patients who decline. Declining should never affect the care they receive.

All articles