Choosing an AI scribe: a therapist's evaluation checklist
An AI scribe sits between your clinical voice and your client’s record, so the decision deserves more rigor than a feature list and a free trial. The work of choosing an AI scribe as therapists do it is not finding one that produces a readable note. Most can. The hard part is judging where the audio goes, how much editing the draft actually demands of you, and whether you can walk away with your data intact a year from now. This is a structured way to evaluate the field on the dimensions that matter clinically, plus the questions to put to vendors and a low-stakes way to test before you trust it with real sessions.
Choosing an AI scribe: start with the data path, not the demo
The first question is not “how good is the note” but “where does my client’s voice go.” Many scribes upload audio to a cloud server, transcribe it there, and run a hosted model to draft the note. That means a recording of a session, and the text derived from it, travels off your machine and sits on someone else’s infrastructure, governed by their retention policy and their breach exposure.
Ask the vendor plainly: does any audio or transcript leave the device, and if so, where is it stored, for how long, and who can access it? A signed business associate agreement is necessary if data is processed in the cloud, but treat “HIPAA compliant” as a floor, not a feature. It tells you a vendor has paperwork and safeguards; it does not change the fact that the data left the room. On-device processing changes the question itself: if the audio never travels, there is no cloud copy to subpoena, breach, or retain. That is the design premise behind the best AI scribe options for Mac that run locally.
Accuracy means edit load, not a marketing number
A transcript that sounds broadly accurate can still miss the clinically loaded part: a medication name, a risk disclosure, the difference between “I want to hurt myself” and a hypothetical. Those are the lines you cannot afford to lose, and they are exactly the ones that get garbled in crosstalk. Ignore any accuracy percentage a vendor quotes; it is hard to verify and usually measured on clean audio, not a real consulting room with two voices and a tissue box.
What you can measure is your own edit load. After a draft is generated, how much do you correct, restructure, or delete? Watch specifically for confabulation — plausible details the model invents that were never said. In a clinical note, an invented symptom or quote is worse than an omission. A good scribe under-claims and leaves gaps you fill, rather than fabricating completeness.
Workflow fit: does it match how you already write
A scribe that produces a tidy note in a format your payer rejects has solved the wrong problem. Confirm it generates the structure you actually use — SOAP, DAP, or BIRP — and that switching between them is straightforward. Then look at the editing experience, because that is where you will spend your real time:
- Can you fix the draft inline, the way you would edit your own writing?
- Does it preserve your phrasing, or force a house style you have to override every time?
- How many steps from session end to a note you would sign?
Remember the draft is yours to finish. The software produces a starting point; you remain the author of record, you review every line, and you sign it. Any tool that implies the note is finished the moment it is generated is describing a liability, not a feature.
Cost model over the long run
Pricing structure tells you how the vendor thinks about your relationship. Map the real total cost over three to five years, not the monthly number.
| Model | What it usually signals |
|---|---|
| Per-seat subscription | Ongoing dependency; cost scales with time, not value |
| Per-note or per-minute | Usage anxiety; you ration a clinical tool to save money |
| One-time / lifetime | You own the tool; no meter running on your sessions |
A recurring fee for a tool that runs entirely on your own hardware is worth questioning. If nothing leaves your Mac and there is no server cost to cover, a perpetual monthly charge is renting something you could own outright. A side-by-side comparison of the trade-offs is the fastest way to see which model fits a long practice.
Continuity and export: the exit test
Vendors fold. Apps get acquired and shut down. Before you commit years of notes, ask the question that separates a tool from a trap: if you disappear tomorrow, can I export every note and transcript in a portable format? You want plain text, PDF, or a clean structured file you can move into any EHR — not a proprietary database only their app can read. Ownership of your clinical record is non-negotiable; a scribe that holds it hostage has failed before you start.
Test it on a low-stakes session first
Do not let a new scribe near a high-acuity session on day one. Test it the way you would test any clinical instrument:
- Use it on a routine, low-risk session — a stable client, a check-in, with consent and per your jurisdiction’s recording rules.
- Write the note both ways once: let the scribe draft it, then compare against the note you would have written unaided.
- Score the gap. Count the corrections, the inventions, the omissions, and the time saved net of editing.
- Confirm the data path with a real recording, not a sales call — then export the note and reopen it elsewhere.
Recording rules and consent requirements vary by state, board, and payer, and this is not legal advice, so confirm yours with your licensing body or attorney before you record anyone. That caveat aside, a single honest test session tells you more than any feature list.
The right scribe is the one that quietly shortens the distance between a session ending and a note you are willing to put your signature on, without asking you to send your clients’ voices anywhere or surrender ownership of your own records. It does not finish your thinking for you; it gives you a draft to sharpen, and it stays out of the way while you do the part that is genuinely yours. Hold every candidate to those five questions — data path, accuracy, workflow, cost, and exit — and most of the field sorts itself out before you have signed a single note.