What is an AI medical scribe? A guide for doctors

Q: Is an AI medical scribe the same as dictation software?

No. Dictation transcribes what you deliberately say into a microphone, and you still compose the note. An ambient AI medical scribe listens to the whole clinician-patient conversation and drafts a structured note from it, which you then review, edit and sign.

Q: How much time does an AI medical scribe actually save?

Less than marketing suggests. In a randomised trial at UCLA Health, note-writing time fell by about 41 seconds per note in the best-performing arm — roughly 9.5% more than the control group. Uncontrolled studies report larger savings, but they lack a control group and recruit enthusiastic early adopters.

Q: Are AI medical scribes accurate?

Modern ambient scribes report overall error rates of about 1-3%, lower than older dictation systems at roughly 7-11%. But they fail differently: fabricated content, silent omissions, misinterpretation and speaker misattribution. The clinician remains responsible for checking the draft against what actually happened.

Q: Do patients have to consent to an AI scribe?

In general yes, and you should document it. RACGP guidance states that consent to the use of an AI scribe must be sought at the start of the consultation and recorded. Recording law varies by jurisdiction, so check local requirements and your indemnity provider's position before adopting one.

An AI medical scribe is software that listens to a clinical encounter — in person or over video — and drafts the clinical note from the conversation, so the clinician reviews, edits and signs instead of typing from scratch. It is usually called an ambient scribe, because it runs in the background rather than being dictated to. Crucially, it produces a draft: the clinician remains responsible for the accuracy of the record.

This guide is written for clinicians weighing one up. It covers what an AI scribe actually is, how the pipeline works, what the 2025 evidence really shows about time saved and burnout, how accurate these tools are and how they fail, the regulatory and consent position, and a practical checklist for evaluating a product before you adopt it.

Key takeaways

An AI medical scribe drafts the note from the consultation; you review and sign it. Roughly 30% of physician practices were using one by late 2025. Randomised evidence shows modest documentation-time savings — not the dramatic figures in vendor marketing — while burnout and cognitive-load improvements look more consistent. Error rates are low overall (~1–3%) but the failure modes are new: fabrication, omission and speaker misattribution. Consent and record accuracy remain your responsibility.

What is an AI medical scribe?

An AI medical scribe is a documentation tool that captures the spoken clinical encounter and generates a structured clinical note from it — typically a SOAP note, a referral letter or a progress note — which the clinician then checks, corrects and signs. It replaces the typing, not the clinical judgement. The term "ambient" distinguishes it from dictation: you talk to the patient, not to the microphone.

The category has grown quickly. A 2025 commentary in npj Digital Medicine reports that AI scribes are "now used by approximately 30% of physician practices, with adoption growing across major health systems and electronic health records (EHR) platforms", and describes one health system where more than 7,000 physicians used AI scribes across more than 2.5 million patient encounters in 14 months (Topaz, Peltonen & Zhang, npj Digital Medicine, 2025).

The problem it targets is well documented. In the landmark "Tethered to the EHR" study of 142 family-medicine physicians, clinicians spent 355 minutes — 5.9 hours — of an 11.4-hour workday in the electronic record per day, including 1.4 hours after clinic hours (Arndt et al., Annals of Family Medicine, 2017). That after-hours tail is the "pyjama time" AI scribes are sold to eliminate. For the wider economics of that burden, see our piece on the hidden cost of clinical documentation.

How does an AI medical scribe work?

Four stages, in sequence: capture the audio, transcribe it to text, summarise that transcript into a structured clinical note using a large language model, and hand the draft to the clinician to verify and sign. Everything upstream of the clinician's signature is assistive; nothing is filed to the record automatically unless the product is configured — and regulated — to do so.

Capture. A phone, tablet or desktop microphone records the consultation, with the patient's knowledge and consent. Some tools stream in real time; others process after the encounter ends.
Transcription (ASR). Automatic speech recognition converts audio to text, usually with speaker diarisation to separate clinician from patient.
Summarisation. A large language model turns the raw transcript — which is long, repetitive and full of social conversation — into a clinical note in the requested format and template.
Review and sign. The clinician reads the draft against what actually happened, corrects errors and omissions, adds anything not spoken aloud, and signs. This step is not optional.

The summarisation step is where the real difficulty lies: a transcript is not a note. Deciding what was clinically relevant, what belongs in the Objective section rather than the Subjective, and what should be left out entirely is a judgement task, and it is the step most likely to go wrong. We break the pipeline down further in how AI medical scribes work.

How is an AI scribe different from dictation or a human scribe?

Dictation transcribes what you say into the microphone; an AI scribe listens to the whole consultation and writes the note for you. A human scribe does the same job but is a person in the room (or on a remote link), bringing judgement and visual context that an audio-only system cannot.

Dictation software converts your spoken narration to text verbatim. You still compose the note. Older speech-recognition systems carry error rates of roughly 7–11%, driven by medical jargon and accent variability (Topaz et al., 2025).
Human scribes document in real time alongside you. In randomised trials they are more than four times as likely to produce notes physicians rate as accurate, compared with physicians documenting for themselves (ibid.). They also see the room — body language, visible distress, social context an audio feed misses.
Ambient AI scribes need no narration and no second person. Reported overall error rates are lower (≈1–3%), but they introduce failure modes a human scribe does not have: hallucination, silent omission and speaker misattribution (ibid.).

The honest summary is that these are different trade-offs rather than a straight upgrade. We compare the two directly in AI scribe vs human scribe.

Do AI medical scribes actually save time?

Yes, but less than the marketing suggests. The best-designed evidence — a randomised trial — found savings measured in seconds per note, not hours per day. Uncontrolled before-and-after studies report larger gains, but they lack a control group and tend to recruit enthusiastic early adopters, so they overstate what a typical clinician should expect.

In a three-arm pragmatic randomised trial at UCLA Health, 238 physicians across 14 specialties were assigned 1:1:1 to Microsoft DAX Copilot, Nabla, or usual care over roughly two months and about 72,000 encounters. Average time spent writing each note fell from 4 min 30 s to 3 min 49 s in the Nabla arm (about 41 seconds) against 4 min 22 s to 4 min 4 s (18 seconds) in the control arm — a reduction described as 9.5% larger than control and statistically significant. The DAX arm's reduction did not reach statistical significance (UCLA Health, reporting the NEJM AI trial, November 2025).

Uncontrolled studies are rosier. A quality-improvement study of 45 clinicians across 17 specialties found documentation time fell by a median of 2.6 minutes per appointment and after-hours EHR work by 29.3%; an observational study of 119 allied-health professionals found a 33% reduction in documentation time. But the same review notes a study in which AI scribes saved only 34 seconds per note, with wide individual variability (Topaz et al., 2025).

What this means in practice

Expect a real but modest reduction in keyboard time per note, with large variation between clinicians, specialties and products. Treat any vendor claim of "two hours a day saved" as a hypothesis to test on your own encounters during a trial — not a number to build a business case on.

Do AI scribes reduce physician burnout?

The burnout signal is more consistent than the time-saving signal — which is itself interesting, because it suggests the benefit is partly about cognitive load and attention rather than minutes. Clinicians repeatedly report lower exhaustion and lower note-related mental effort even when the measured time saved is small.

A multi-site quality-improvement study of 263 ambulatory clinicians across six US health systems found the proportion experiencing burnout fell from 51.9% to 38.8% after 30 days of ambient AI scribe use (adjusted OR 0.26; 95% CI 0.13–0.54; P < .001), with note-related cognitive task load falling from 7.10 to 4.46 on a transformed 10-point scale (Olson et al., JAMA Network Open, October 2025). In a separate pilot across Mass General Brigham and Emory Healthcare involving 1,430 clinicians, burnout prevalence at Mass General Brigham fell from 50.6% to 29.4% at 42 days and from 52.6% to 30.7% at 84 days (You et al., JAMA Network Open, August 2025).

Read those numbers with the caveats the authors themselves give. Neither study had a control group; both relied on voluntary participants; survey response rates in the Mass General Brigham / Emory pilot ranged from 11.1% to 30.4%. Enthusiastic early adopters answering a survey about a tool they chose to pilot is not the same as a representative clinician. For the wider picture on what drives — and relieves — burnout, see our pillar on physician burnout.

How accurate are AI medical scribes, and how do they fail?

Modern ambient scribes report overall error rates of roughly 1–3%, better than older dictation systems. But the errors are qualitatively different, and in clinical documentation a small percentage can carry serious consequences. Reported rates also vary widely because studies define "error" differently — some count only factual inaccuracies, others include omissions and inconsistencies. For the full evidence base — three independent evaluations, a government audit of 20 pre-approved products, and why the published figures disagree so much — see our deep dive on how accurate AI medical scribes are.

The documented failure modes are worth knowing by name, because they tell you what to look for when you review a draft (Topaz et al., npj Digital Medicine, 2025):

Hallucination (fabrication) — content with no basis in the encounter, such as documenting an examination that never happened, or a diagnosis nobody mentioned.
Omission — symptoms, concerns or findings that were discussed but are missing from the note. This is the hardest to catch, because nothing on the page looks wrong.
Misinterpretation — context-dependent statements read the wrong way, producing incorrect medications, treatments or plans.
Speaker misattribution — patient statements recorded as the clinician's, or vice versa, when diarisation fails in a multi-speaker room.
Performance disparities — speech recognition underlying these tools shows systematically higher error rates for some speakers, including documented reduced accuracy for African American patients, and for non-standard accents and limited English proficiency.

There is also a subtler risk the same authors raise: if a scribe documents everything discussed, the note can bloat into noise; if it filters, it may filter on criteria you cannot inspect. Either way, reviewing a long AI-generated draft is itself cognitive work — a real cost to set against the time saved.

The practical mitigation is a disciplined review habit. Read the draft against the encounter, not for internal plausibility — a hallucinated note reads perfectly well. Check the Plan and medications first, then scan for what is missing, which is the error class a quick proofread will not surface. Our SOAP note guide sets out what a complete note should contain, and the free printable SOAP note template gives you the same structure on paper.

Is an AI medical scribe a regulated medical device?

It depends on what the product does. In the UK, the MHRA clarified in July 2026 that ambient voice technology products "intended solely for transcription, summarising of clinical conversations" are not regulated as medical devices under the current framework, while products intended to support diagnosis, treatment or prevention — or that take automated action such as placing orders without clinician review — are regulated and must meet the relevant safety and performance requirements (MHRA, 29 July 2026).

That line matters when you evaluate a product. A scribe that only drafts text sits outside device regulation in the UK; the moment it suggests a diagnosis, recommends treatment or acts on the record without a clinician in the loop, it is in a different regulatory category. Rules differ by jurisdiction, so check your own regulator and your indemnity provider before adopting. Not being a regulated device is not the same as being validated — it simply means a different set of obligations applies to you as the adopter.

Do you need patient consent to use an AI scribe?

Assume yes, and document it. The RACGP's guidance is explicit: "Patient consent to the use of an AI scribe must be sought at the beginning of the consultation and documented in the record." The same guidance states that "the output of an AI scribe must be carefully checked for accuracy as these tools can produce errors and inconsistencies" (RACGP, AI scribes and data quality).

Two further points from that guidance are easy to miss. First, recording law varies by jurisdiction and recording without consent can be a criminal matter in some places. Second, AI scribes "predominately only deliver free-text consultation notes" and rarely populate structured fields such as allergies or observations — so a record built on scribe output alone may be incomplete for medicines reconciliation, decision support and audit. Structured data entry still needs doing.

How should you evaluate an AI medical scribe before adopting one?

Run a real pilot on your own encounters, with your own specialty vocabulary and your own patient population, and measure rather than impress yourself. The questions below separate products that survive clinical use from those that demo well.

Accuracy on your patients — including accented speech, code-switching and the languages your clinic actually works in. Test it, do not take a claim on trust.
Omission behaviour — ask explicitly how the product handles information it judges irrelevant, and audit a sample of notes against the encounter for what is missing.
Note format control — can you shape the template (SOAP, problem-oriented, specialty-specific) rather than accept the vendor's house style? See SOAP vs POMR.
Review workflow — how fast is it to correct a draft? A scribe that saves typing but costs five minutes of careful reading has moved the work, not removed it.
Data handling — where is audio and text processed and stored, which sub-processors are involved, what is the retention policy, and can you get your data out?
Evidence offered — ask for controlled data on time and note quality, not testimonials. Ask what the measured savings were for clinicians in your specialty.
Governance fit — consent workflow, clinical-safety documentation, data protection assessment, and your indemnity provider's view.
Cost against measured benefit — price it against the time you actually measured in the pilot, not the vendor's figure.

Where Shifaa AI fits

Shifaa AI is an AI clinical assistant for doctors and clinics, currently in private beta, with AI medical scribe functionality among its features. The design principle is the one this article argues for throughout: the assistant drafts, the clinician decides. It is clinical decision support and documentation assistance — it does not diagnose patients, and every note is a draft for you to review, edit and sign. Processing uses disclosed third-party AI sub-processors, which is a question you should ask of any vendor, including us.

Frequently asked questions

Is an AI medical scribe the same as dictation software?