AI medical scribes: how ambient documentation works, and where it fails

An AI medical scribe listens to a consultation, turns speech into a transcript, pulls out the clinical facts and drafts the note for the clinician to review and sign. Done well, it gives doctors their attention back. Done badly, it produces confident notes that nobody can check. This guide explains how the pipeline works, where it fails, and what a safe ambient AI scribe looks like.

We write this as a team that builds a scribe and uses it in clinic. So we will be specific about the weak points as well as the strengths. If you are about to choose a scribe, read this alongside our 12-point AI scribe evaluation checklist.

What an ambient scribe actually does

Older dictation tools turned your voice into text, and you still had to structure the note yourself. Ambient documentation is different. The scribe listens to the natural conversation between doctor and patient. You do not dictate to it. It then does the structuring for you. Most ambient scribes run the same six stages:

1. Capture. A microphone records the visit, usually through the browser on a laptop or tablet.
2. Transcription. Speech recognition turns audio into text, sentence by sentence, often while the visit is still running.
3. Speaker separation. The system works out who said what: doctor, patient or companion. This matters because “I have chest pain” means something different from each speaker.
4. Fact extraction. A language model reads the transcript and pulls out clinical facts: symptoms, history, exam findings, assessments, plans, medications and orders.
5. Note assembly. The facts are arranged into a structure such as SOAP or a specialty template.
6. Review and sign-off. The clinician reads, corrects and signs. Until this step, the note is only a draft.

Some scribes stop at a text note that you copy into your EMR. Others write back into the chart. They fill structured fields, suggest ICD-10 and CPT codes, and stage prescriptions, lab and imaging orders for you to approve. The further downstream a scribe reaches, the more time it can save. It also means an error can travel further, so the review step carries more weight.

Where ambient scribes fail

Every stage above can go wrong. These are the failure modes we see discussed most often by clinicians, and the ones we design against:

The last row is the hardest to fix with technology alone. A well-formatted note invites trust. The risk is not only that the AI makes mistakes. The risk is that a busy clinician signs them. So a safe design has to make checking fast. It cannot just tell clinicians to be careful.

What good looks like: five design tests

When we look at any scribe, including our own, we ask five questions.

• Can every statement be traced? For every line in the note, can the clinician see the exact words it came from, and ideally hear the audio? If a fact has no source, it should not be in the note.
• Is review per fact, not per page? Accepting or rejecting single facts is faster and safer than proofreading a finished paragraph.
• Does a rejection flow everywhere? If you reject a plan item, it should disappear from the note, the export and the chart, not just from one screen.
• Does anything happen without a click? Orders, codes and bookings should be staged for approval, never committed on their own.
• Where does the audio go? Know which company processes the recording, where it is stored and for how long, and whether it is used for training.

How Verto approaches it

Verto is the scribe built into the Helix EMR, so the note, the orders and the codes all land on the same patient record. The design follows the tests above:

• Fact cards with tap-to-prove. Verto turns the conversation into fact cards grouped by SOAP section. Tap any fact to see the exact utterance, who said it and when. Once the visit is saved, you can replay the audio segment, if the patient consented to audio retention.
• A “needs a source” lane. Facts Verto cannot trace to something said in the visit are kept apart. They never reach the note unless a clinician confirms them.
• One decision, every surface. Rejecting a fact rebuilds the note, and exports are always rebuilt from reviewed facts. A rejected fact cannot reach a printed or downloaded note.
• Staged orders. Prescriptions, labs, imaging, referrals and follow-ups mentioned in the visit are staged for one review pass. Surgery is never booked automatically.
• Codes proposed, chart confirms. ICD-10 suggestions appear in the chart's diagnosis block with accept and dismiss buttons. Only codes you accept travel onward.
• Specialty structure. Notes follow templates for the specialty you practise. Helix covers 27 medical specialties, and Verto autofills specialty chart fields with a confidence level for each field.
• Private by design. The AI models run in Helix's own cloud, not through third-party AI APIs, and customer data is never used for training. Our security page covers data residency and audit.

Verto also handles bilingual visits, where the patient speaks Arabic and the doctor dictates in English. That topic needs its own article, so read our guide to the Arabic–English AI scribe.

Rolling out a scribe safely

Scribe projects rarely fail because the software is poor. They fail because nobody planned the first month. A simple rollout looks like this:

1. Agree on consent wording. Tell patients the visit is being recorded to help write the note. Decide whether audio is kept, and for how long. Check your facility's policy and your regulator's current guidance.
2. Start with a small group. Pick two or three clinicians with different styles, including one skeptic.
3. Set up the basics. Use a decent microphone, keep background noise down, and enroll each clinician's voice so speaker identification works.
4. Review everything at first. For the first weeks, read every fact against the transcript. You will learn the error patterns of your own rooms.
5. Measure your own time. Time a few ordinary clinic sessions before and after. Clinicians using Verto tell us they spend 80% or more less time on documentation, but your specialty, visit length and habits will decide your result.
6. Write down the rules. Record who signs, what can never be accepted without reading, and how to report an error.

The bottom line

An AI medical scribe is a drafting tool. The clinician still owns the note. The best scribes accept that and are built around fast, evidence-based review. A scribe that cannot show you where a sentence came from should not be writing your notes.