An AI scribe is software that listens to a consultation through a phone or computer microphone, turns the conversation into a transcript, and uses a language model to draft a structured clinical note for the clinician to check, edit and sign. It documents; it does not diagnose. The clinician remains responsible for every word that goes into the record.
The reason to understand the mechanics is that the draft looks finished. Its errors are small and plausible: a dropped "no" that turns "no bladder changes" into "bladder changes", a left knee that becomes a right knee, an examination finding the clinician never actually performed. Each stage of the pipeline produces its own kind of mistake, so knowing the stages tells you where to look when you review.
The consultation, stage by stage
Most AI scribes, whether they are general products such as Heidi or Nabla, enterprise systems such as Microsoft Dragon Copilot, or a feature built into practice software, follow the same sequence.
Before the patient sits down
- Template choice. The clinician picks, or the system defaults to, a note template: a general consultation, a SOAP note (subjective, objective, assessment, plan), a specialty letter. The template decides the headings the draft will use.
- Context. Tools linked to the clinical record may pull in the patient's name, age and problem list. Standalone tools start with nothing and know only what is said aloud.
- Consent. The patient is told the consultation will be transcribed by an AI tool and asked if that is all right. The scribe does not do this for you.
During the consultation
- Capture. The clinician taps start. The microphone records the conversation, either "ambiently" (the whole natural conversation) or as dictation (the clinician speaking the note after the patient leaves). Audio is usually sent to the vendor's servers, either streamed as it is recorded or uploaded when the clinician stops.
- Speech recognition. A speech-to-text model converts the audio into words. Medical vocabulary, drug names and numbers are the hard parts.
- Speaker separation. Often called diarisation, this labels which words came from the clinician and which from the patient (and any relative or interpreter). It matters because "I've had chest pain" means something different from "have you had chest pain?"
After the patient leaves
- Drafting. A large language model reads the labelled transcript and writes the note under the template's headings, removing small talk and repetition. This usually takes seconds.
- Review and edit. The clinician reads the draft against their memory of the consultation, corrects it, and signs it. This is the only step that makes the note trustworthy.
- Export. The signed note is copied or pushed into the clinical record. Deeper integrations write straight into the right fields; simpler tools rely on copy and paste.
- Deletion or retention. The audio is deleted or kept according to the vendor's settings, and the transcript is kept for a set period or deleted. More on this below, because it varies more than people expect.
What goes in and what comes out: a short example
An illustrative extract from a transcript, lightly cleaned, for a patient with back pain:
Clinician: So how long has this been going on?
Patient: About five days. I was moving boxes at the weekend.
Clinician: Any pain going down your legs?
Patient: No, just here in the low back, more on the right.
Clinician: Any numbness, or any changes with your bladder or bowels?
Patient: No, nothing like that.
Clinician: What have you been taking?
Patient: Ibuprofen, it takes the edge off.
Clinician: OK, I'm just going to press here... that's tender on the right side. Can you lift your leg for me, straight... both sides fine, no pain down the leg. Reflexes are normal and sensation is normal in both legs.
And the drafted note (illustrative), using a SOAP template:
S: Low back pain for 5 days after lifting boxes. Right-sided, no
radiation to legs. No numbness. No bladder or bowel changes.
Ibuprofen gives partial relief.
O: Tenderness right lumbar paraspinal region. Straight leg raise
negative bilaterally. Reflexes and sensation normal in lower limbs.
A: Mechanical low back pain. No red flag features.
P: [from the rest of the consultation] Continue simple analgesia,
keep active, review in 2 weeks if not improving. Safety-netting
advice given on leg numbness and bladder or bowel changes.Notice two things. First, the assessment "no red flag features" is the model summarising; the clinician must agree with it before signing, because it is a clinical conclusion. Second, the objective findings only appear because the clinician said them aloud. Had the leg raise been done in silence, the scribe would either have left it out or, worse, filled the gap with a plausible "neurological examination normal".
Where the errors come from at each stage
This is the part worth pinning next to the screen. Each example is illustrative but typical of what published reviews of AI scribes describe: omissions, errors and hallucinated content.
| Stage | Typical error | Example | How to spot it |
|---|---|---|---|
| Speech recognition | Misheard drug name or number | "Amlodipine 5 mg" becomes "10 mg" | Check every dose, date and measurement |
| Speech recognition | Dropped negation | "No bladder changes" becomes "bladder changes" | Read red-flag lines word by word |
| Speaker separation | Clinician's question recorded as patient's symptom | "Any chest pain?" becomes "reports chest pain" | Ask: did the patient say this, or did I ask it? |
| Speaker separation | Relative's words credited to patient | Daughter's worry recorded as patient's complaint | Check who reported what in the history |
| Drafting | Invented finding | "Chest clear on auscultation" when the chest wasn't examined | Check every objective finding was done |
| Drafting | Wrong side | Right knee becomes left knee | Check laterality in every line |
| Drafting | Omission | Safety-net advice given but not recorded | Check the plan matches what you said |
| Drafting | Over-confident summary | "Likely viral" when you said "could be viral" | Check the assessment says what you meant |
The invented finding is the most dangerous because it reads like good documentation. A realistic way it shows up: a clinician signs twenty notes at the end of a clinic, and weeks later a colleague reads "abdomen soft, non-tender" in a note for a patient whose abdomen was never examined. Nothing flagged it, because it looked exactly like what a thorough clinician would have written.
A 60-second review before signing
Reviewing a note properly does not mean re-reading it from the top as if it were an essay. It means checking the lines most likely to be wrong, in the same order every time, so the habit survives a busy clinic. A routine that fits in about a minute:
- Red flags and negatives. Read every "no" and "denies" line word by word. These are where a dropped word changes the meaning.
- Numbers. Doses, frequencies, durations, measurements, dates. Compare each with what you said.
- Sides and sites. Left or right, upper or lower, which tooth, which joint.
- Examination. For each finding listed, ask: did I actually do that? Delete anything you did not.
- Who said it. Check that symptoms came from the patient, not from your questions or a relative's worries.
- Assessment and plan. Make sure the wording matches your certainty ("possible", "likely", "confirmed") and that safety-net advice you gave is recorded.
An illustrative correction from that routine: the draft said "Plan: increase sertraline to 100 mg". The clinician had actually said "we might think about increasing it at the next review". Step six caught it. Signed as drafted, the note would have recorded a medication change that never happened, and the next clinician to read it might have acted on it.
When the room makes transcription harder
Scribes are tested in quiet rooms with two clear voices. Real consultations are often messier, and accuracy drops in predictable situations:
- Several voices. A parent answering for a child, a partner chipping in, an interpreter repeating everything. Speaker separation struggles with three or more voices, so check the history for who said what with extra care.
- Background noise. A crying baby, a busy corridor, a fan. Put the phone or microphone between you and the patient, not on the far side of the desk.
- Quiet or muffled speech. Masks, a patient who speaks softly, a clinician turned towards the screen. Repeat key facts back ("so that's three weeks, yes?"), which helps the transcript and the patient.
- Sensitive moments. A patient who wants to say something off the record. Pause the recording visibly and say that you have, then restart it.
In an illustrative interpreted consultation, the scribe recorded the interpreter's first-person translations as the clinician's words, so the history read "I have had headaches for a month" under the clinician's name. The fix was simple once spotted: the clinician now says "the interpreter is translating for the patient" at the start, and reads the history section first when reviewing any interpreted note.
What happens to the audio, transcript and note
A scribe creates three things, and each has its own lifespan. The audio recording, the transcript and the drafted note. Vendors handle them differently, and the defaults change:
- Heidi says it does not keep the audio.
- Nabla discards the audio and keeps transcripts for 14 days by default.
- SimplePractice, which sells a note taker for therapists, switched its default in June 2026: new users now start with de-identified transcript retention turned on. It is a useful reminder that a setting you checked at sign-up can change later.
Before any patient is recorded, get written answers to these questions from the vendor: where the audio is processed and stored; when the audio is deleted; how long the transcript is kept and whether you can shorten that; whether any of it is used to train or improve models; which vendor staff can reach it; and whether the vendor will sign the data-processing agreement your data-protection rules require. Health information gets extra protection under data-protection law, so ask your data-protection adviser if any answer is unclear. The broader list is in the patient data and AI confidentiality checklist.
Ambient scribe, dictation and a human scribe compared
| Ambient AI scribe | AI dictation | Human scribe | |
|---|---|---|---|
| What it captures | The whole conversation | Only what the clinician says afterwards | What the scribe hears and understands |
| Clinician effort during consult | Almost none, beyond saying findings aloud | None during, a minute or two after | None |
| Main risk | Misattribution and invented findings | Omitting what the clinician forgets to say | Cost and availability |
| Patient experience | Clinician looks at the patient, not the screen | Same during; clinician talks to the device after | A third person in the room |
| Best for | History-heavy consultations | Procedures, short reviews, noisy rooms | High-volume specialist clinics |
Many clinicians end up using both AI modes: ambient for new problems, dictation for quick follow-ups. Many AI scribe products offer both.
What an AI scribe does not do
- It does not diagnose. Some products add coding suggestions, letters or prompts about next steps. Anything that suggests a diagnosis or treatment is a different and more tightly regulated category of software in many places; ask the vendor what classification its features have where you practise.
- It does not see. A rash, a limp or a wince exists in the note only if someone says it aloud.
- It does not know the history unless it is linked to the record. A standalone scribe will not know the patient had the same problem last year.
- It does not keep an evidential recording in most setups, because the audio is deleted. If you need a recording of what was said, that is a separate decision.
- It does not replace consent. You still tell the patient, and you still stop recording if they ask.
How patients are told, and what they ask
A short, plain explanation at the start works better than a form. An illustrative script:
"Before we start, I use a computer tool that listens to our conversation and types up my notes, so I can look at you rather than the keyboard. I check and correct everything it writes. The recording is deleted once the note is done. Is that all right with you? If you'd rather I didn't use it, that's completely fine."
The most common patient questions are "who hears it?", "is it kept?" and "can you turn it off for this bit?". Have answers ready, and make pausing the recording easy mid-consultation. Detailed wording for posters, forms and scripts is in consent wording for AI note-taking, and the wider rollout is covered in rolling out an AI scribe without losing patient trust.
Speaking so the scribe gets it right
Small changes in how the clinician talks make a large difference to the draft. Compare two illustrative examination segments.
Before: "OK, let me just have a look... mm... and the other one... fine." The scribe has nothing to work with and may invent "examination unremarkable".
After: "Right knee: mild swelling, no redness, flexion to about 110 degrees, ligaments stable. Left knee normal for comparison." Every finding, side and number is in the transcript, and the patient hears what you found, which many like.
Three habits help most. Say the side and the number every time. Say findings as you find them, not in your head. And close with a spoken summary ("So, to sum up, the plan is...") because the plan section of the note will lean heavily on it.
Trying one fairly in your own clinic
A two-week trial tells you more than any demo. Pick one or two scribes, use them on a mix of consultation types, and score every note before editing it. An illustrative scoring sheet from one clinician's first week:
| Measure | Scribe A | Scribe B |
|---|---|---|
| Notes drafted | 42 | 38 |
| Notes needing no correction | 11 | 6 |
| Notes with a factual error (dose, side, finding) | 4 | 9 |
| Invented findings found | 1 | 3 |
| Average review time per note | 1 min 50 s | 2 min 40 s |
| Patients who declined | 2 | 1 |
Scribe B looked faster in its demo, but the error count decides it. Scoring before editing is the important discipline, because once a note is corrected you forget how wrong it was. Whether the time saved justifies the cost for a small clinic is its own calculation, set out in whether an AI scribe is worth it for a small private clinic, and prices are compared in what small clinics pay for AI scribes. Heidi, for example, has a free tier alongside its paid plans, which makes a no-cost first trial possible; check its current plan limits before you start.
Whichever tool you choose, keep counting errors after the trial ends. A scribe that was accurate in week one can drift when the vendor updates its models, and a monthly sample of ten notes read against the transcript is the cheapest way to notice.
More questions about AI scribes
Does an AI scribe work for telephone and video consultations?
Most do. For video, the scribe usually runs on the same computer and captures both sides through the system audio or the microphone and speakers. For telephone, some tools offer a dial-in or speakerphone mode. Check how the tool captures the far side of the call, because a scribe that only hears the clinician produces a one-sided note.
Can the scribe handle a consultation in another language?
Many scribes transcribe several languages and some draft the note in the clinician's language from a conversation in another. Accuracy varies by language and accent, and interpreted consultations with three voices are harder. Test with your real patient mix before relying on it, and check medication names and numbers with extra care.
Is the audio recording kept as part of the medical record?
Usually not. Many scribes delete the audio soon after the note is drafted, and some keep the transcript for a short period. The signed note is the record. If you want recordings kept, that is a separate decision with its own consent and storage rules, so check the vendor's retention settings and your own record-keeping policy.
Further reads
- What to Ask Before Buying Any AI Tool for a Medical Practice — The questions to put to any AI vendor before a clinic signs.
- A Worked AI Implementation Plan for a Small Medical Practice — A worked plan for bringing AI into a small medical practice.
- How Physiotherapists Use AI to Write Treatment Notes Faster — How the same technology works in physiotherapy sessions.
- How Small Vet Practices Use AI to Cut Time on Clinical Notes — How vet practices use scribes, with pets and owners in the room.
- Are AI Meeting Note-Takers Safe for Client Calls? — The general safety questions for any tool that records conversations.
- 10 Admin Tasks a Small Clinic Can Hand to AI This Month — Other clinic admin jobs AI can take off your team.
- Creating Home Exercise Programmes With AI for Physio Patients — A physio-led workflow for AI-assisted home exercise programmes: shorthand to patient instructions, pain rules, delivery tools and adherence messages.
- Can Therapists Use ChatGPT for Session Notes? — Personal plan, de-identified shorthand, business plan or a therapy note tool: where each lands, and why removing a name rarely removes the client.
- Should Therapists Tell Clients They Use AI? — Which AI uses therapists must tell clients about, ready-to-adapt wording for contracts and first sessions, and what to do when a client says no.
- Eight Practical AI Uses for an Independent Optician — Recalls, enquiries, plain-English prescriptions, lens reorders, frame copy, reviews, referral letters and stock: how each works in a small practice.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft Dragon Copilot product pages; Heidi pricing and product pages; Nabla and Heidi data retention statements; published reviews of ambient AI scribes in clinical documentation research.