AI medical scribe: what it takes to chart into your EHR
An ambient scribe listens, drafts the note into your template, and the clinician edits and signs it. The real work is getting that draft into your EHR safely, with consent handled and a human still deciding what gets filed.
Answered in short
5 things that decide this
- 01An ambient AI scribe records the visit, transcribes it, and drafts a structured note into your EHR's template. The clinician reads it, edits it, and signs it before it becomes part of the chart.
- 02The best available randomised trial found 41 seconds saved per note by the better of two ambient scribes tested (Lukac et al., NEJM AI, 2025), a real but modest gain, not the hours some vendors imply.
- 03Nuance DAX, Abridge and Suki do the listening and drafting well. They generally do not touch re-keyed demographics, order entry, referral letters or coding, which is where staff time keeps going.
- 04A practice needs a consent script, a signed BAA, a defined place for the audio to live and expire, a de-identification policy, and a clinician sign-off gate before any scribe goes live, packaged or custom.
- 05A custom build fits where your template, your EHR's access model or your billing codes do not match what a packaged scribe assumes. The clinical decision never moves off the clinician.
Listening and drafting is the easy 80%
You run the visit. A microphone, usually on a phone or a dedicated device, records the conversation with the patient's consent already given. The model transcribes it, pulls out the history, the exam findings and the plan, and drafts a note shaped to your template: SOAP, a dental chart note, a SOAP-like DVM record, whatever your specialty runs. You read the draft, correct anything wrong, and sign it. That's the whole job an ambient scribe was built for, and the better tools do it well.
One number here is worth knowing, and it's smaller than the marketing suggests. In a 238-physician randomised trial, the better of two ambient scribes tested saved 41 seconds per note against a control group with no scribe (Lukac et al., NEJM AI, 2025). Real, measured, and worth having across a full patient day. It's not the hour-back-per-shift some sales decks promise, and a vendor who won't cite a trial is asking you to take their word for the number instead.
The 20% around the note that no packaged scribe touches
Nuance DAX, Abridge and Suki, at the category level, all do the same core job: capture the conversation, draft the note, hand it back for sign-off. None of them, on their own, re-keys the demographics your intake form already collected. None files the referral, submits the prior authorization, or picks the CPT and ICD codes the visit implies. Your front desk still types the same address a third time, and a biller still translates the note into a claim by hand.
That gap is exactly what a scribe was never scoped to close. It's also where most of a practice's manual hours actually live once the note itself stops being the bottleneck. Writing a beautiful note into a system nobody else touches moves the work. It doesn't remove it.
- 01Demographics and intake data, re-typed instead of flowing from the form to the chart.
- 02Orders, referrals and prior-auth packets, still assembled by staff after the note is signed.
- 03CPT and ICD coding, translated from the note by a person rather than proposed alongside it.
- 04Anything the scribe transcribes that touches PHI outside your EHR's own access controls.
What a practice needs in place first
None of this is optional, and it's the same list whether you buy a packaged scribe or build one. You need a consent script your front desk and clinicians actually use, not a line buried in the intake paperwork nobody reads aloud. You need a signed BAA naming every vendor the audio and the draft note touch, including any transcription step that runs on a third party's servers.
You need a stated answer to where the audio goes, how long it's retained, and when it's deleted, because a recording of a clinical conversation is PHI the moment it exists. A de-identification policy covers anything a model sees outside the encounter itself. A template mapping puts the draft in the fields your EHR actually uses. And a clinician sign-off gate that can't be skipped under load is the whole reason a scribe stays a drafting tool and not a decision-maker.
- 01A consent script staff say out loud, not a clause in paperwork.
- 02A signed BAA naming every vendor in the chain, including transcription.
- 03A stated retention and deletion policy for the raw audio.
- 04A de-identification rule for anything a model processes outside the encounter.
- 05A template mapping into your EHR's actual fields, not a generic note dump.
- 06A sign-off gate the clinician can't skip, whatever the day's volume.
- Your templateMapped to the fields your specialty actually charts, not a generic SOAP dump
- Your EHR's access modelWrite-back through FHIR, HL7 or a vendor API, to the depth your account exposes
- Your codesCPT and ICD proposals drafted alongside the note, reviewed before submission
- Sign-off gateThe clinician reads, edits and signs. Nothing files itself
A packaged scribe is built for the average template. Where yours doesn't match, or your EHR's write-back needs work a vendor won't do for one account, that's where a custom build earns its place.
Some of the systems we have shipped
Related questions
01Do ambient scribes work for specialties outside primary care, like dental or veterinary?+
The transcription and drafting approach carries over, but the template does not. A dental chart note and a DVM SOAP record don't look like a physician's visit note, so the mapping into your PMS or PIMS is specialty-specific work either way.
02Can a scribe suggest billing codes as part of the draft?+
A handful of packaged tools offer a coding suggestion as an add-on. Ask a vendor to show it inside your own account rather than a demo built to look clean, and either way, a biller or the clinician reviews the code before it goes on a claim.
03Where does the audio recording actually go?+
That has to be a specific answer before you sign anything: which vendor's servers, which region, how long it's kept, and when it's deleted. If a vendor can't answer that plainly, treat it as a compliance flag, not a technical detail.
04Is a scribe the same thing as an AI EHR integration?+
No. A scribe drafts a note. An integration is the write-back and data flow between systems, which is a separate and often larger piece of work. Most practices need both, but they're not the same build.
05What happens if the scribe mishears something clinically important?+
That's exactly what the sign-off gate exists for. The clinician is reading and editing the draft against their own memory of the visit before it becomes part of the record, not accepting it unread.
Related
- AI, automation and custom software for healthcare →The whole chain, from the front desk to the chart to the claim.
- EHR integration for healthcare practices →Where scribe write-back fits inside a wider integration build.
- Epic integration: what a real build involves →The access-model reality behind any EHR write-back.
- The EHR does 80 percent, and people copy the rest →The manual work sitting around every EHR, scribe or not.
- HIPAA-compliant development for healthcare →The BAA, PHI and audit-trail requirements behind any clinical AI build.
