AI medical scribe for therapists — what's actually different
AI medical scribes were built for primary care. Therapy is a different problem. A practical guide to what changes when the encounter is 50 minutes of talk, two voices, and content that should not leave the room.
The phrase AI medical scribe showed up in serious volume in the second half of 2024. Abridge raised at $5B. Nuance DAX got folded deeper into Epic. Suki, Doximity, and Heidi all advertised in primary-care journals. By early 2026 the term has settled into a stable meaning: software that records a clinical encounter, transcribes it, and drafts a structured note so the clinician does not have to write it from memory at the end of the day.
That definition was written for primary care. Therapy is a different problem.
This post is the version of the explanation I wish someone had given me before I built one.
The primary-care frame
A primary-care encounter is roughly twelve minutes. There are two people in the room some of the time, three or four people in the room some of the time (clinician, patient, MA, sometimes a family member). The vocabulary is bounded — symptoms, medications, vitals, ICD-10 codes. The output of the visit is a SOAP note that goes into Epic or Cerner via an HL7 or FHIR API. The clinical signal is dense; the time is short. A scribe's job is to compress.
Almost every AI medical scribe shipping in 2026 was designed against that frame. Abridge, Nuance DAX, Suki, Heidi, Augmedix, DeepScribe, Iris — they all do roughly the same thing: ambient mic, cloud upload, server-class transcription, frontier-model drafting, API push back into the EHR. They are remarkable products for what they were built for.
A 50-minute talk-therapy session is none of those things.
What changes when the encounter is therapy
Five things change, and each one breaks something a primary-care scribe was good at.
1. The encounter is long, and most of it is not data
Primary care wants the clinician's words and the patient's symptom complaints. Therapy wants the patient's exact words, sometimes verbatim, often the silences between them. A long pause after a question about a deceased parent is clinical signal. A primary-care scribe trained to compress will smooth it away. A therapy scribe needs to preserve the rhythm.
2. The vocabulary is narrative, not coded
There is no ICD-10 for “the client made eye contact for the first time when discussing her stepfather.” A therapy note records what was said and what was processed, not which symptoms presented at which severity. The note formats — SOAP, DAP, BIRP, EMDR, Intake, Progress — are narrative-shaped, not coded-shaped. The model has to be tuned for that.
3. The content is asymmetric in sensitivity
Primary care exposes vital signs and prescription history if it leaks. Therapy exposes a client's inner life — trauma history, suicidal ideation, custody-relevant disclosures, EMDR processing blocks. The asymmetry of harm is severe. A breach of a primary-care scribe is bad. A breach of a therapy scribe is career-ending for the clinician and life-altering for the patient.
This is the change that most cloud-based AI medical scribes have not absorbed. The same architecture that was acceptable for vital signs is not obviously acceptable for trauma narratives.
4. The EHR is different — and the API push is a liability
Primary-care EHRs (Epic, Cerner, Athena, eClinicalWorks) have institutional IT departments, signed BAAs with every vendor, and well-defined integration patterns. Therapy EHRs (SimplePractice, TherapyNotes, TheraNest) are solo-clinician tools without the same integration surface. The API push that's table-stakes for a primary-care scribe is harder to negotiate for therapy.
It is also undesirable. An AI scribe with an authenticated API handle on a therapist's EHR is, by definition, a server with a key to the patient's chart. That is a more attractive target than an AI scribe that ends at a copy-paste step.
5. The buyer is different
A primary-care scribe is bought by a hospital's IT committee after a six-month security review. A therapy scribe is bought by a solo clinician on a Tuesday after reading two essays and listening to a podcast. The compliance posture has to fit on a page. The architectural attestation has to pass an ethics consultant's read on a Saturday morning. The pricing has to be a number a private practice can absorb without a P&L meeting.
What an AI medical scribe for therapists should look like
Five characteristics fall out of those changes. None of them are revolutionary; together they describe a different product.
- Two-voice tuning. Speaker diarization that knows therapist from client across a 50-minute hour, with multi-speaker support for couples and family work. Not the four-voice ambient mode of a primary-care scribe.
- Therapy-native templates. SOAP, DAP, BIRP, EMDR, Intake, Progress — and a custom-format option for whatever your panel demands. Not a generic SOAP that papers over the differences.
- Architectural privacy, not BAA privacy. Audio, transcript, and note all stay on the device. No vendor server in the loop. The compliance posture is “we never had the data,” not “we have a contract about the data we have.” The full comparison lives at Bryl vs Mentalyc, Upheal, Heidi, Blueprint, Eleos, Twofold.
- Copy-paste, not API push. One paste into SimplePractice or TherapyNotes is the privacy guarantee, not a workflow gap. Why this matters is on the glossary page.
- A signing step that actually means something. A press-and-hold gesture, a tamper-evident fingerprint of the source transcript, a notarial frame that makes silent edits detectable. Therapy notes have to hold up under a subpoena; the signing step is what holds them.
Why this is a separate product, not a feature
The cloud-based AI medical scribes shipping today could, technically, add an “on-device mode.” They will not. The architecture of a cloud scribe is designed around the cloud — the model serving, the transcript pipeline, the EHR integrations, the team, the compliance org. Removing the cloud isn't a feature flag; it's a different company.
This is the gap Bryl was built for. One ICP, one architecture, one product. The longer version of why is here; the technical anatomy is in What on-device AI actually means for therapy notes.
If you are a licensed psychotherapist evaluating an AI medical scribe right now, the question worth asking is not which cloud scribe has the best transcription accuracy or the smoothest EHR integration. The question is whether the product was designed for the encounter you're recording. Most of them were not.