RESULTS · Published
The best AI scribe for GPs
Hanah was the best AI scribe we tested on GP consultations. It was first on time to a finalised note in each of the six consultations, and it had the fewest note errors in each. Heidi was second and Lyrebird third.
The tests
The consultations are GP video appointments from PriMock57, a public dataset of mock primary care consultations. Six AI scribes recorded each one five times through their own web apps, and every note was graded blind against the clinical facts said in the consultation. We rank scribes by time to a finalised note, which is the wait after pressing stop plus the time a clinician needs to review the note and fix what is wrong.
The benchmark has eight PriMock57 consultations. This ranking covers the six that every scribe has published notes for: gastroenteritis, eczema, chest pain, allergic reaction, low mood and urinary symptoms. Hanah's notes for the heart failure and suspected TIA consultations are not yet in the published results, and those two are left out of every scribe's average. Hanah's published notes were regenerated on 11 October 2026, and its wait after pressing stop is an estimate from measured generation time.
Results
| Rank | Scribe | Time to a finalised note | Note errors per note | Coverage | Hallucination flags per note | Word error rate |
|---|---|---|---|---|---|---|
| 1 | Hanah | 1 min 50 s | 1.5 | 98% | 0.4 | 14.0% |
| 2 | Heidi | 3 min 31 s+1 min 41 s | 9.9+8.4 | 79%−19% | 1.3+0.9 | 14.6%+0.6% |
| 3 | Lyrebird | 3 min 56 s+2 min 06 s | 11.2+9.7 | 83%−15% | 3.7+3.4 | 17.5%+3.5% |
| 4 | CliniScripts | 4 min 24 s+2 min 34 s | 11.6+10.1 | 75%−23% | 1.5+1.1 | 14.8%+0.8% |
| 5 | Freed | 4 min 51 s+3 min 01 s | 8.5+7.0 | 87%−11% | 2.5+2.1 | 19.5%+5.5% |
| 6 | PatientNotes | 5 min 56 s+4 min 06 s | 19.6+18.2 | 58%−40% | 2.9+2.5 | 19.6%+5.6% |
A Hanah note was ready to finalise 1 minute 50 seconds after the consultation ended, on average across the six consultations. It captured 98% of the facts, with 1.5 note errors and 0.4 hallucination flags per note. Reviewing and fixing it took 79% less time than writing the note yourself, and 30% of Hanah notes were ready to finalise as written. Hanah's slowest run on any of the six consultations took 2 minutes 46 seconds, which was quicker than every other scribe's average.
Heidi was second at 3 minutes 31 seconds, with 9.9 note errors per note and 79% coverage, and it had the second fewest hallucination flags at 1.3 per note. Lyrebird was third at 3 minutes 56 seconds. Freed had the second fewest note errors at 8.5 and the second highest coverage at 87%, and it took 4 minutes 51 seconds to a finalised note.
Hanah had the lowest word error rate at 14.0%, with Heidi at 14.6% and CliniScripts at 14.8%.
Results by consultation
| Consultation | First | Second | Article |
|---|---|---|---|
| Gastroenteritis | Hanah, 1 min 44 s | Heidi, 3 min 19 s | Read |
| Eczema | Hanah, 2 min 01 s | Heidi, 3 min 17 s | Read |
| Chest pain | Hanah, 1 min 31 s | CliniScripts, 2 min 58 s | Read |
| Allergic reaction | Hanah, 2 min 21 s | Heidi, 4 min 44 s | Read |
| Low mood | Hanah, 1 min 37 s | Heidi, 2 min 45 s | Read |
| Urinary symptoms | Hanah, 1 min 46 s | Lyrebird, 3 min 34 s | Read |
The times are each scribe's average time to a finalised note on that consultation.
Findings
- In the chest pain consultation, the patient said they had no allergies other than aspirin. Hanah recorded this in every run, and every other scribe missed it in every run.
- The same patient said aspirin makes them swell up. Heidi, Freed and Lyrebird transcribed "swell up" as "swallow" in all five runs. Hanah and CliniScripts heard it correctly every time.
- In the gastroenteritis consultation, the doctor said antibiotics were not needed. Hanah recorded this in every run and Lyrebird in four of five. Heidi, Freed and PatientNotes missed it in every run.
- In the low mood consultation, the patient said they had no previous mental health diagnoses. Hanah recorded this in every run, and no other scribe recorded it in full in any run.
- The doctor in the urinary symptoms consultation said the infection could be treated without testing the urine. Hanah recorded this in every run, and no other scribe recorded it in any run.
- Heidi transcribed furosemide as "fruit" in all five heart failure runs. Lyrebird's transcripts named omeprazole, which nobody said, in all five runs and missed bisoprolol in every one.
Our Heidi review covers how Heidi did across all 12 consultations, with a video walkthrough.
Frequently asked questions
What is the best AI scribe for GPs?
In our test of six AI scribes on six recorded GP consultations, Hanah was the best. Its notes were ready to finalise soonest after the consultation, 1 minute 50 seconds on average, and had the fewest note errors in every consultation. Heidi was second and Lyrebird third.
How accurate are AI scribes in GP consultations?
Across six GP consultations, the scribes captured between 58% and 98% of the clinical facts that were said. Note errors, counting missed facts and unsupported claims together, ranged from 1.5 per note for Hanah to 19.6 for PatientNotes. Every note still needs a clinician's review before it is finalised.
Do AI scribes make things up in GP notes?
Every scribe we tested added some details to its notes that were not said in the consultation. We count these as hallucination flags. Hanah had the fewest at 0.4 per note, and Lyrebird the most at 3.7.
How much time does an AI scribe save a GP?
A doctor spends a median 8.1 minutes writing a note by hand (Apathy et al. 2023). Reviewing and fixing a Hanah note took 79% less time than that. The other scribes saved between 34% and 62%.
How were the AI scribes tested?
Each scribe recorded the same GP consultations from the PriMock57 dataset five times through its own web app. Every note was graded blind against an inventory of the clinical facts said in the consultation. The recordings, transcripts, notes and grading reasons are published on each test page.
Figures in this article come from the ScribeStandard benchmark. See the full results, each test and how we test.