An AI medical scribe is software that listens to a clinical visit and writes the note for you. It transcribes the conversation, pulls out the clinically relevant parts, and drafts structured documentation, usually a SOAP note, that you review and sign. No keyboard during the visit, no chart waiting for you at 9pm.
That’s the short answer. The longer answer is worth ten minutes, because the products sold under this label differ in ways that matter clinically: what they do with your audio, which languages they understand, what they draft beyond the note, and what “review” actually means when you’re 14 patients deep on a Tuesday.
Key takeaways
- An AI scribe converts visit audio into a draft note. You stay the author; it does the typing.
- Primary-care physicians log a median of 36.2 minutes of EHR time per 30-minute visit. The note now outlasts the appointment it describes.
- Published US pricing runs roughly $39–$199 per clinician per month, and several big vendors publish nothing at all.
- The two questions that separate vendors fastest: what happens to the audio, and how honest the vendor is about what the tool can’t do.
of EHR time logged per 30-minute primary-care visit (JAMA Network Open, 2023)
to review the AI draft note after the visit ends
entry price per clinician (US, Assist annual billing)
How does an AI medical scribe actually work?
Four steps, all of them during or right after the visit.
First it listens. You start a session on your phone or laptop, get the patient’s consent, and see the patient like normal. The scribe captures the conversation ambiently. Nobody dictates anything.
Then it transcribes. Speech becomes text. The good systems handle two speakers, interruptions, and the way real consultations loop back on themselves. The genuinely hard version of this problem is a crowded room with a relative answering half the questions and a thick accent the model has never heard. More on that below.
Next it structures the transcript. A language model turns it into clinical documentation: subjective, objective, assessment, plan. This is the step where products diverge. Some stop at the note. Others draft the downstream artifacts too, and AI Medical Scribe by Patient Square also returns ICD-10 suggestions and a prescription draft with the SOAP note.
Then you review and sign. The draft appears about two minutes after the visit ends. You read it, fix what needs fixing, and sign. Nothing enters the record without you.
The honest framing: this is a very fast, very patient junior scribe with perfect recall of the last conversation, not a colleague. It drafts. You decide.
Does the four-step loop hold up outside a demo? The largest real-world read so far says yes. The Permanente Medical Group ran ambient AI scribes across 7,260 physicians and 2,576,627 patient encounters over 63 weeks (October 2023 to December 2024), and reported an estimated 15,791 hours of documentation time recovered, roughly 1,794 eight-hour workdays, in results published in NEJM Catalyst and summarized by the AMA. Eighty-four percent of those physicians said the tool improved how they communicated with patients, and 47% of patients noticed their doctor spending less time on the computer. That’s not a pilot with 30 friendly users. It’s a deployment at the scale of a large medical group, and it’s the kind of evidence worth weighing before you trust any product with your charts.
What does an AI scribe hand back after the visit?
Baseline, every credible product: a structured note. Beyond that, read the feature list carefully, because the word “scribe” covers a wide range.
The AI Medical Scribe is one module inside Practice Copilot, Patient Square’s AI platform for the whole practice. It listens during the visit and hands back a structured SOAP note, ICD-10 suggestions, and a prescription draft, ready to review and sign about two minutes after the visit. The prescription is a draft built from the plan you stated in the room; you review the drugs, doses, and instructions and sign. It doesn’t run an automated interaction or dosing check, and it never goes to a pharmacy on its own. We’re deliberate about that boundary: a draft an LLM wrote is still a draft an LLM wrote, so the clinician stays the safety check, exactly as with a script written by hand.
Things some vendors include that we currently don’t: after-visit patient summaries, referral letters, per-specialty note templates, and direct EHR write-back. If those are must-haves for your workflow today, weigh that honestly. Our pricing page lists exactly what ships now.
AI scribe vs human scribe vs dictation: what’s the difference?
| Human scribe | Dictation software | Ambient AI scribe | |
|---|---|---|---|
| Who does the work | A person, in the room or remote | You, after the visit | Software, during the visit |
| Your time per note | Review only | 5–10 min of speaking + cleanup | ~2 min review |
| Cost shape | A salary (plus hiring, training, turnover) | Low subscription | $39–$199/mo, flat per clinician |
| Captures the visit itself | Yes | No, you reconstruct it from memory | Yes |
| Patient in the room | A third person present | Just you | Just you and a phone |
| Scales to a group practice | Linearly, painfully | Per clinician | Per clinician, flat |
Dictation was the first wave: it moved typing to talking, but you still spent evenings narrating visits you’d already done once. Human scribes solve the time problem and create a staffing one. Ambient AI is the first version where the documentation happens while the medicine happens.
What happens to the visit audio?
Ask this before you ask about price. Seriously.
A visit recording is among the most sensitive artifacts in healthcare. It’s more revealing than the note, because it contains everything that didn’t make the note. Vendors handle it very differently. Some retain audio for days or weeks for model improvement or dispute resolution. Some let you opt out. Some are vague.
Our position: visit audio is processed in memory and discarded the moment the note is drafted. There is no audio archive, not for us, not for the practice, not for anyone. What survives is the note you reviewed and signed. Notes are encrypted in transit (TLS 1.2+) and at rest (AES-256), they belong to your practice, and you can export or delete any visit at any time. The full posture is on our security page.
The floor for any vendor is a signed BAA, plus encryption in transit and at rest and access you can audit. A vendor who can’t answer “where does the audio live and for how long” in one sentence has answered the question.
How much does an AI medical scribe cost in 2026?
From published pricing pages (fetched June 2026):
- Self-serve band: Freed runs $39–$119 per month across its tiers depending on billing; Commure Scribe lists $59–$89; Twofold $49–$69; Sunoh lists $149 promotional ($199 regular).
- The opaque tier: several large vendors (Suki, DeepScribe, Nabla) publish no pricing at all and route you to sales. You get a quote, not a list price.
- Patient Square’s Practice Copilot starts from $79/clinician/month on annual billing (the Assist plan); Copilot ($119) adds a bundled AI EHR and messaging, and Autopilot ($199) adds an AI receptionist and follow-ups. Month-to-month and committed-annual rates are on the pricing page. Every plan starts with a 7-day free trial.
One opinion, stated as one: we think pricing pages that say “talk to sales” for a per-clinician software subscription tell you something about how the renewal conversation will go. Transparent list pricing keeps vendors honest, including us.
Does the subscription pay for itself? The arithmetic is short. Primary-care physicians in a 2023 JAMA Network Open study logged a median of 36.2 minutes of EHR time per 30-minute visit, and an earlier time-motion study in Annals of Internal Medicine found nearly two hours of EHR and desk work for every hour of direct patient care, plus another one to two hours of after-hours “pajama time” most nights. Recovering even a fraction of that against a $79 monthly line item isn’t a close call. Still, run the numbers for your own visit volume, not ours.
- EHR & desk work67%
- Direct patient care33%
Where do AI scribes still fall short?
Anyone selling you perfection is selling. The failure modes worth knowing:
Noisy, multi-speaker rooms. A quiet consult room is the easy case. A crowded clinic with a caregiver answering half the questions, or a pediatric visit with a parent in the room, is the real test, and transcription quality varies sharply across products there.
Language is the other big one. “Multilingual” without a named list is marketing. Read any vendor’s language list literally and test it on your own patients during a trial, because accents and code-switching are where models quietly break. We capture English and more languages, with notes always returned in clean clinical English, and in low-connectivity settings capture works offline with on-device encryption and syncs later.
Drafts contain errors, too. Models mishear drug names, compress two complaints into one, occasionally write something plausible that didn’t happen. This is why the review-and-sign step is load-bearing, why we treat the clinician as the safety check on every Rx draft rather than an automated screen, and why you should distrust any vendor whose pitch implies you can skip reading the note.
And the chart is yours, legally. No scribe, silicon or human, moves the documentation responsibility off the signing clinician. A tool that writes a complete record at the pace of the visit is the difference between signing at 6pm and signing in bed, and between a defensible chart and a backlog you’ll be answering for later.
Five questions to ask before you sign with any vendor
- “What happens to the visit audio, and when is it deleted?” Accept only a one-sentence answer with a timeline. Ours: processed in memory, discarded at note draft.
- “Will you sign a BAA?” Vague answers here are disqualifying. So is anyone implying they’re “HIPAA certified,” because no such certificate exists.
- “What exactly is in the draft: note only, or codes and prescriptions too? What checks them?” Make the demo show a wrong drug pairing and watch what happens.
- “What does my actual patient mix sound like to your model?” Test with your accents, your languages, your interruptions, during a free trial, on day one, not after procurement.
- “What’s the price after year one?” Launch rates, annual commitments, and month-to-month rates differ. Get the whole ladder in writing. Ours is published.
A scribe that survives those five questions will probably survive your Tuesday clinic. Take the 7-day trial, run it on real visits, and read every draft closely for the first week. That’s the evaluation that matters, and it costs you nothing but the week.