RESULTS · Published
The best AI medical scribe in Australia in 2026
Hanah is the best AI medical scribe we have tested for Australian clinics. We recorded the same 12 consultations through six AI scribes, at least five times each, and graded every note blind against the clinical facts said in the consultation. Hanah ranked first on 19 of the 20 measures we publish. Heidi was second on time to a finalised note and on hallucinations.
Most comparisons of AI scribes rank them on features, price and integrations, because those are easy to read off a website. What a clinician actually takes home from a scribe is a note, and the note is the hard part to compare. Two scribes can look identical in a demo and differ fivefold in how many facts they miss or invent. So we built a test that holds the consultation constant and measures the note.
This article explains how an AI scribe saves time, which scribes we tested, how we decided which is best, what they cost, what Ahpra, the RACGP and the TGA expect of you, and how to run your own trial. Every figure links back to the recordings, transcripts and notes it comes from, and the research we cite is listed at the end.
How an AI medical scribe works and where the time saving comes from
An AI medical scribe listens to the consultation through a phone or computer microphone, turns the audio into a transcript, and then uses a language model to write a structured clinical note from that transcript, usually in SOAP or a template you choose. You review the draft, fix what is wrong, and paste or sync it into your practice management system.
The time saving comes from replacing writing with reviewing. A doctor spends a median 8.1 minutes writing a note by hand, across 203,728 physicians in Apathy et al. (2023). If the draft is good, reading and correcting it takes a fraction of that. If the draft misses facts or adds things that were not said, the clinician has to find and fix each one, and the saving shrinks. That is why we measure the note rather than the transcript alone.
The demand is real. In the RACGP's Health of the Nation 2025 survey, 77% of GPs said they were unhappy with the amount of administration in their work, up from 70% a year earlier.
| Measure | Figure | Source |
|---|---|---|
| GPs unhappy with their admin load | 77% (2025), up from 70% | RACGP Health of the Nation |
| GPs using an AI scribe, newsGP poll | 40% (Nov 2025), up from 22% (Aug 2024) | newsGP |
| Avant members using an AI scribe | 19% (Feb 2025), up from 11% (Aug 2024) | Avant |
| Median time to write a note by hand | 8.1 minutes per visit | Apathy et al. 2023 |
The newsGP figure is a self-selected poll of readers, so read it as a direction rather than a census. The Avant figures cover a shorter period and a different population. Both show adoption roughly doubling in a year or less.
- August 202422%
- November 202540%
Source: newsGP polls of 1,170 and 1,516 readers. Self-selected, so read it as a direction rather than a census.
The AI scribes we tested
We tested six AI scribes that Australian clinicians can sign up for today.
| Scribe | Free plan | Note template in our test | What it did best in our benchmark |
|---|---|---|---|
| Hanah | Yes | Our SOAP instruction | Fewest note errors, highest coverage, fewest hallucinations, quickest note |
| Heidi | Yes | Our SOAP instruction | Second quickest to a finalised note, second fewest hallucinations |
| Lyrebird | Yes | Our SOAP instruction | Second fastest note after pressing stop, 21 seconds |
| Freed | Not stated | Default note | Second fewest note errors, second highest coverage |
| CliniScripts | Not stated | Default note | Second lowest word error rate, least memory in the stress test |
| PatientNotes | Not stated | Our SOAP instruction | Second highest share of notes ready as written, 3% |
Freed and CliniScripts do not take a note instruction the way the others do, so they were graded on their default note. Lyrebird recorded 11 of the 12 consultations because it does not support the French consultation. The full list of what each scribe supports is on the results page.
How we decided which AI scribe is best
"Best" needs a definition before it means anything. Ours is the scribe that gets a clinician from the end of a consultation to a correct, finished note in the least time. We call that time to a finalised note, and it has two parts: the wait after pressing stop, and the time a clinician needs to read the draft and fix what is wrong with it.
The consultations. The benchmark uses 12 recorded consultations. Eight are mock GP consultations from PriMock57, a public dataset, covering chest pain, heart failure, suspected TIA, eczema, gastroenteritis, an allergic reaction, low mood and urinary symptoms. The others are a consultation held entirely in Hindi, one in French and English, a short pain management dictation and a 50-minute myotherapy session recorded with a television and voices from another room in the background. Every recording is playable on its test page.
The product clinicians use. A script drives each scribe's own web app in a visible browser, presses record, plays the consultation into the microphone input and presses stop. Every scribe hears identical audio, and every scribe recorded every consultation it supports at least five times, because results change from run to run.
Blind grading against the facts. Each consultation has an inventory of every fact said that relates to the patient's conditions, symptoms or signs, 358 facts across the 12 consultations. This follows Weiner et al. (2020), who checked 105 doctors' notes against recordings of the visit. Notes for the same consultation are shuffled together with product names removed. Each fact is marked full, partial or missing, and a separate pass flags anything the note says that the consultation does not support. A missed or partial fact is a note error, and so is each hallucination flag.
Clinician time. We estimate review time from reading the note at 190 words a minute, recalling and typing each missing fact, and checking and deleting each hallucination flag. Adding the measured wait after pressing stop gives time to a finalised note. The full method is on the methodology page and in our benchmark article.
Results: the best AI medical scribe across 12 consultations
| Rank | Scribe | Time to a finalised note | Note errors per note | Coverage | Hallucination flags per note | Word error rate |
|---|---|---|---|---|---|---|
| 1 | Hanah | 1 min 33 s | 1.2 | 98% | 0.36 | 12.8% |
| 2 | Heidi | 3 min 25 s+1 min 52 s | 9.0+7.8 | 81%−17% | 1.0 | 13.6% |
| 3 | CliniScripts | 4 min 03 s+2 min 30 s | 9.7+8.5 | 80%−18% | 1.4 | 13.0% |
| 4 | Lyrebird | 4 min 06 s+2 min 33 s | 11.6+10.4 | 82%−16% | 3.4 | 17.7% |
| 5 | Freed | 4 min 40 s+3 min 07 s | 8.3+7.1 | 87%−11% | 2.5 | 23.1% |
| 6 | PatientNotes | 5 min 39 s+4 min 06 s | 18.0+16.8 | 61%−37% | 2.6 | 19.5% |
- Hanah1m 33s
- Heidi3m 25s
- CliniScripts4m 3s
- Lyrebird4m 6s
- Freed4m 40s
- PatientNotes5m 39s
- Writing the note yourself8m 6s
Wait after pressing stop plus estimated review time, averaged over 12 consultations and at least five runs each. Writing the note yourself: median 8.1 minutes, Apathy et al. 2023.
A Hanah note was ready to finalise 1 minute 33 seconds after the consultation ended, on average across all 12 consultations. It captured 98% of the clinical facts, with 1.2 note errors and 0.36 hallucination flags per note, and 30% of its notes needed no editing at all. No other scribe had more than 3% of notes ready as written. Reviewing and fixing a Hanah note was 83% quicker than writing it yourself.
- Hanah30%
- PatientNotes3%
- CliniScripts2%
- Heidi0%
- Freed0%
- Lyrebird0%
- Doctor-written notes10%
Notes with every fact captured in full and no hallucination flags. Doctor figure: 10% of notes error-free, Weiner et al. 2020.
Heidi was second at 3 minutes 25 seconds, and its 1.0 hallucination flag per note was the second lowest. Its notes captured 81% of the facts, so most of its review time went on adding what was missed. CliniScripts and Lyrebird were close behind at about four minutes. Freed had the second fewest note errors and the second highest coverage, but its notes were slow to arrive after pressing stop, 49 seconds on average, and its transcripts had the highest word error rate of the six.
- Hanah1.2
- Freed8.3
- Heidi9.0
- CliniScripts9.7
- Lyrebird11.6
- PatientNotes18.0
- Doctor-written notes5.0
Missed and partly captured facts plus hallucination flags, graded blind. Doctor figure: Weiner et al. 2020, who counted less strictly than we do.
The gap matters most in the worst case. Hanah's slowest note in any run took 2 minutes 36 seconds to review. Every other scribe had at least one note that took more than seven and a half minutes, which is nearly as long as writing it from scratch.
Results by specialty and language
An average can hide a scribe that is excellent at GP consultations and poor at a long session or another language. So we also rank each consultation on its own. Hanah was first on time to a finalised note in every one.
| Consultation | Best scribe | Article |
|---|---|---|
| Chest pain and heart failure | Hanah | Cardiology |
| Suspected TIA | Hanah | Neurology |
| Eczema | Hanah | Dermatology |
| Gastroenteritis | Hanah | Gastroenterology |
| Allergic reaction | Hanah | Allergy and immunology |
| Low mood | Hanah | Mental health |
| Urinary symptoms | Hanah | Urology |
| Pain management dictation | Hanah | Pain management |
| 50-minute myotherapy session | Hanah | Myotherapy |
| Hindi consultation | Hanah | Hindi |
| French and English consultation | Hanah | French |
The margins differ a lot. On the short pain management dictation the leading scribes were close, and some differences were within run-to-run variation. On the full GP consultations Hanah's lead was large: across those six it had 1.5 note errors per note against 8.5 to 19.6 for the rest. The GP article has that ranking in full.
AI scribe accuracy and why the doctor still signs the note
Every AI scribe we tested sometimes wrote down something that was not said. Examples from our recordings: Heidi listed metoprolol when the patient had corrected the doctor to bisoprolol, in four of five heart failure notes. Lyrebird's transcripts named omeprazole, which nobody said, in all five heart failure runs. In the chest pain consultation, Heidi, Freed and Lyrebird heard "swell up" as "swallow" in every run. Each case is on its test page with the transcript and note side by side.
Published studies point the same way. The results depend heavily on the product, the setting and how the study was run, and several have vendor ties.
| Study | Setting | Finding |
|---|---|---|
| Lukac et al., NEJM AI 2025 | Randomised trial, 238 physicians, about 72,000 encounters | Time per note fell from 4 min 30 s to 3 min 49 s with one scribe; burnout improved about 7%; notes occasionally had clinically significant inaccuracies |
| Duggan et al., JAMA Netw Open 2025 | US health system, before and after | Time in notes per appointment down 20%, after-hours time down 30% |
| Olson et al., JAMA Netw Open 2025 | 263 clinicians, 6 US health systems, no control group | Burnout fell from 51.9% to 38.8% after 30 days |
| Palm et al., Frontiers in AI 2025 | 97 encounters, AI and physician notes compared | Hallucinations in 31% of AI notes and 20% of physician notes |
| Memon et al., BMC Health Serv Res 2026 | Gold Coast outpatients, 7,499 consultations | About 58% of AI output used unchanged; 47% of surveyed staff reported hallucinations |
| Biro et al. 2025 | 20 GPs reviewing AI-drafted patient replies | Each planted error went uncorrected by 65–75% of reviewers |
- Hanah0.36
- Heidi1.00
- CliniScripts1.40
- Freed2.47
- PatientNotes2.58
- Lyrebird3.42
- Doctor-written notes1.72
Anything the note documents that the consultation does not support. Doctor figure: things documented that did not happen, Weiner et al. 2020.
That last study is the one to remember. Review is not a perfect safety net. If about 70% of errors survive a typical review, as Biro et al. found, then about 85 errors would reach the record per 100 Hanah notes, against 500 for doctor-written notes and between 583 and 1,262 for the other five scribes. It is an estimate, and it is explained on the methodology page, but it shows why the error rate of the draft matters even when every note is reviewed.
- Hanah85
- Freed583
- Heidi628
- CliniScripts678
- Lyrebird809
- PatientNotes1,262
- Doctor-written notes500
Note errors per note x 100, times the 70% of errors that survived review in Biro et al. 2025. Doctors' notes are filed without review.
Ahpra, the RACGP and Avant agree on who carries the risk: the clinician. The scribe drafts the note, and you are responsible for what it says when you sign it.
How much an AI medical scribe costs in Australia
These are the prices on each vendor's own pricing page when we checked on 16 November 2026. Prices change often, so confirm before you buy.
| Scribe | Free plan | Paid plans (as listed) | Currency shown | Where data is stored, as stated |
|---|---|---|---|---|
| Hanah | Yes (Basic) | A$100 per practitioner a month, or A$80 billed annually; Specialist from A$250 a month | AUD | Australia, the UK or the US |
| Heidi | Yes | A$120 a month, or A$85 billed yearly; Practice and Enterprise plans | AUD | Australia (compliance page) |
| Lyrebird | Yes, and free for Best Practice users | Pro $240 a month, or $160 billed annually; clinic plans from $299 a month + GST | $, not named | Australian hosting listed on the Enterprise plan |
| Freed | Not stated | $39 to $119 a month | $, not named | Not stated |
| CliniScripts | Not stated | US$29 to US$69 a month; clinic US$59 per user a month | USD or CAD | Not stated |
| PatientNotes | Not stated | Reported from US$29 a month; pricing page did not load | USD | Not stated |
The sticker price is the smaller cost. A clinician seeing 20 patients a day spends far more on review time than on the subscription. Across our benchmark, reviewing a Hanah note took 1 minute 24 seconds and a Heidi note 2 minutes 58 seconds. Over 20 notes a day that difference is about half an hour, every day, which is worth more than any plan in the table.
- Hanah1m 24s
- Heidi2m 58s
- Lyrebird3m 43s
- CliniScripts3m 50s
- Freed4m 5s
- PatientNotes5m 11s
- Writing the note yourself8m 6s
Estimated time to read the draft and fix every missed fact and hallucination flag, averaged over 12 consultations.
Is there a free AI medical scribe?
Yes. Hanah, Heidi and Lyrebird each list a free plan, and Lyrebird says it is free for clinicians using Best Practice. Free plans are a good way to try a scribe on your own consultations, but they usually limit features, templates or volume, and the limits change. Check what the free plan includes before relying on it for every patient.
Our benchmark used each scribe's normal signed-in account, so the results describe the note you get from the product, not a free or paid tier in particular.
Privacy, consent and the TGA
Consent. Ahpra expects informed patient consent to an AI scribe, recorded in the notes. Avant suggests a detailed explanation the first time, then a verbal confirmation at each later consultation, documented each time. The RACGP says consent can be written or verbal depending on your medical defence organisation's advice.
Data. All three bodies tell you to find out where the audio, transcript and note are stored, for how long, and whether the vendor uses patient data to train its models. Storage overseas adds obligations under the Privacy Act. Of the six scribes we tested, Heidi states that its data is hosted in Australia, Lyrebird lists Australian hosting on its Enterprise plan, and Hanah lists Australia among its storage regions. The other three do not state a location on their pricing pages, so ask them directly.
The TGA. Under the TGA's digital scribes guidance, a scribe that only records and transcribes is not a medical device. One that suggests diagnoses or treatments, triages patients or draws clinical conclusions is likely to be one, and a disclaimer does not settle it. The TGA reviewed 12 digital scribes in 2026 and, as Medical Republic reported, found that eight were likely to have medical device functions. It did not name them. The RACGP's page still says scribes are not medical devices, so the TGA guidance is the one to follow.
Our grading counts a diagnosis as a fact only when the clinician said it. A scribe that adds a diagnosis nobody made gets a hallucination flag, which is the same behaviour the TGA is concerned with.
How to run a two-week AI scribe trial that actually decides it
Most clinics trial a scribe by using it for a few days and asking whether it feels faster. That rarely separates two good products, because the differences that matter, a missed medication or an invented symptom, are easy to miss one note at a time.
A better trial compares scribes on the same consultations. You can start with ours: every recording in the benchmark can be played or downloaded, so you can run the same audio through any scribe you are considering and compare its note with the ones we graded. Then test on your own work.
Recommended trial setup
- Pick two or three scribes, including one from our top two.
- In week one, run the same five of our recordings through each and count missed and invented facts against the transcript.
- In week two, use each scribe on real consultations with consent, and time how long you spend editing each note before you sign it.
- Count the notes you could sign without editing, and look at the worst note each scribe produced, not just the average.
- Confirm data storage, consent wording and practice software integration in writing before you choose.
The bottom line on the best AI medical scribe in Australia
On the evidence we have, Hanah is the best AI medical scribe for Australian clinics. It was first on time to a finalised note in every one of our 12 consultations, its notes had about a seventh of the errors of the next best scribe, and it was the only scribe with fewer errors per note than published figures for doctor-written notes.
Heidi is the runner-up. It was second on speed and hallucinations, stores data in Australia and has a free plan, but its notes missed about one fact in five. Lyrebird, CliniScripts and Freed were close to each other, and PatientNotes was last on most measures. Our Heidi review and the Hanah results go into more detail.
The benchmark has limits. Review time is an estimate from fixed reading and typing rates, a hallucination flag counts once whatever its clinical consequence, and we do not test user experience or practice software integration. Every recording, transcript, note and grading reason is published in one JSON file, so anyone with accounts for these scribes can check our numbers.
Frequently asked questions
What is the best AI medical scribe in Australia?
In our benchmark of six AI scribes on 12 recorded consultations, Hanah was the best. Its notes had 1.2 errors each, against 8.3 to 18.0 for the other scribes, and were ready to finalise 1 minute 33 seconds after the consultation. Heidi was second on time to a finalised note.
How did you decide which AI scribe is best?
Every scribe recorded the same 12 consultations at least five times through its own web app. Notes were graded blind against an inventory of 358 clinical facts, and scribes were ranked by time to a finalised note: the wait after pressing stop plus the time a clinician needs to review and fix the note.
How much does an AI medical scribe cost in Australia?
On the vendors' pricing pages in November 2026, paid plans for individual clinicians ran from about $80 to $240 a month, with lower prices for annual billing. Heidi, Lyrebird and Hanah each list a free plan. Freed, CliniScripts and PatientNotes price in US dollars or an unnamed currency.
Is there a free AI medical scribe?
Yes. Heidi, Lyrebird and Hanah each list a free plan, and Lyrebird says it is free for Best Practice users. Free plans usually limit features or volume, so compare them on a week of your own consultations before relying on one.
How accurate are AI medical scribes?
It varies a lot between products. Across our 12 consultations, scribes captured between 61% and 98% of the clinical facts said, with 1.2 to 18.0 note errors per note. Published studies also find hallucinations: one 2025 evaluation found them in 31% of AI notes and 20% of physician notes.
Do AI scribes make things up?
Every scribe we tested sometimes wrote down something that was not said. Hanah averaged 0.36 such hallucination flags per note, Heidi 1.0, CliniScripts 1.4, Freed 2.5, PatientNotes 2.6 and Lyrebird 3.4.
Who is legally responsible for an AI-generated clinical note?
The clinician. Ahpra, the RACGP and Avant all say the practitioner must check the note is accurate before it goes in the record. The scribe drafts it, and you sign it.
Do I need patient consent to use an AI scribe?
Yes. Ahpra expects informed consent, recorded in the notes. Avant suggests a full explanation the first time and a verbal confirmation at each later consultation, documented each time.
Are AI scribes regulated by the TGA?
A scribe that only records and transcribes is not a medical device under TGA guidance. One that suggests diagnoses or treatments, triages patients or draws clinical conclusions is likely to be one, and a disclaimer does not change that.
Heidi or Lyrebird: which should an Australian clinic choose?
In our benchmark Heidi was ahead of Lyrebird on time to a finalised note (3 minutes 25 seconds against 4 minutes 6 seconds), note errors (9.0 against 11.6) and hallucination flags (1.0 against 3.4). Lyrebird is free for Best Practice users. Both were behind Hanah on every note-quality measure.
Sources
- Weiner SJ et al. How accurate is the medical record? A comparison of the physician's note with a concealed audio recording in unannounced standardized patient encounters. JAMIA, 2020. The error standard we grade notes against, and the 5.0 errors per doctor-written note
- Apathy NC et al. Documentation dynamics: note composition, burden, and physician efficiency. Health Services Research, 2023. Median 8.1 minutes of note-writing per visit across 203,728 physicians
- Biro JM et al. Opportunities and risks of artificial intelligence in patient portal messaging in primary care. npj Digital Medicine, 2025. 65–75% of reviewing GPs missed a planted error, the basis of our errors-reaching-the-record estimate
- Papadopoulos Korfiatis A et al. PriMock57: a dataset of primary care mock consultations. ACL, 2022. Source of eight of our 12 consultations
- Lukac PJ et al. Ambient AI scribes in clinical practice: a randomized trial. NEJM AI, 2025 (UCLA Health summary)
- Tierney AA et al. Ambient artificial intelligence scribes to alleviate the burden of clinical documentation. NEJM Catalyst, 2024
- Olson KD et al. Use of ambient AI scribes to reduce administrative burden and professional burnout. JAMA Network Open, 2025
- Duggan MJ et al. Clinician experiences with ambient scribe technology. JAMA Network Open, 2025
- Palm E et al. Assessing the quality of AI-generated clinical notes. Frontiers in Artificial Intelligence, 2025
- Memon S et al. Ambient listening scribe technology in an outpatient context: a mixed methods trial evaluation. BMC Health Services Research, 2026
- UW School of Medicine and Public Health. Ambient AI improves practitioner well-being. 2025
- RACGP. Health of the Nation 2025: GP perceptions of the profession
- newsGP. GPs' AI scribe use doubles in one year: poll. December 2025
- Avant. AI scribes: rapid uptake in 6 months. June 2025
- Ahpra. Meeting your professional obligations when using artificial intelligence in healthcare
- RACGP. Artificial intelligence (AI) scribes
- Avant. AI scribes and patient consent. January 2025
- TGA. Digital scribes
- Medical Republic. Report on the TGA's digital scribe review outcomes. October 2026
- Vendor pricing pages: Heidi, Lyrebird, Hanah, Freed, CliniScripts (checked 16 November 2026). Also lyrebirdhealth.com/au/pricing, hanah.health/pricing, getfreed.ai/pricing and cliniscripts.com/pricing
Figures in this article come from the ScribeStandard benchmark. See the full results, each test and how we test.
Explore results