Can AI Cut Clinical Documentation Time?
Ambient AI scribes listen to a consultation and draft the note. Randomised trials now show they return clinician time, but the measured saving is smaller than vendors claim and uneven between products. Every draft still has to be checked, and Australian regulators are clear that the clinician who signs it owns it.
On this page
Better: Note and confidence
Moderate confidenceNotes can match human quality; errors still occurIn one US trial, scribe drafts scored highly on a quality instrument when rated by an AI model, before clinician review; blinded GPs rated AI notes comparable to human notes in a simulated Australian study. At Gold Coast Health, 47% of staff survey respondents reported hallucinations.Faster: Note and confidence
Moderate confidenceMeasured savings are real but smaller than vendor claimsIn two randomised trials measured from health record logs, one product cut time on notes by 0.36 hours a day (a secondary outcome); in the other, one product cut time per note by 9.5% and another made no significant difference. Microsoft's five minutes per encounter is a self-reported vendor survey.Cheaper: Note and confidence
Low confidenceSome transcription savings; capacity gains are smallOne Australian children's health service reported transcription cost savings of AUD 16,903 but no definitive effect on overtime. A US multisite study found less than one extra visit a week, and we found no published cost-benefit analysis.Safer: Note and confidence
Governance-gatedSafe only with review, consent and data controlsRACGP, Ahpra and the Commission all place responsibility for the record on the clinician. The TGA says a scribe that suggests diagnoses or treatments is likely to be a medical device.The problem, in numbers
Clinical notes take clinician time, and AI scribes have spread quickly. In newsGP reader polls, the share of respondents using an AI scribe rose from 22% in August 2024 to 40% in November 2025.Source 1 Those are self-selected polls of readers, not a census of general practice.
The time at stake is measurable. In a randomised trial at UCLA Health, physicians spent a median of between 4 minutes 23 seconds and 5 minutes 22 seconds writing each note before the study began, depending on the group, as recorded by the health record system.Source 5
This review asks how much clinician time an ambient scribe actually returns, and what checking its draft costs. Our reviews of diagnosis and medication safety deal with clinical decisions. This one is about the record.
What improves, and why
An ambient scribe records the conversation and uses generative AI to summarise it into a draft, which is then added to the health record.Source 4 In our reading, the gain comes from replacing typing with reading and editing. Whether that is faster depends on how much the clinician has to fix.
The strongest evidence comes from two randomised trials published in NEJM AI in late 2025, both measuring time from health record logs rather than asking clinicians. At UW Health, 66 practitioners in two US states were randomised to staggered starts with Abridge over 24 weeks. Of 71,487 notes, 38% were drafted with the scribe, and measured time on notes fell by 0.36 hours a day, a secondary outcome.Source 6 On a structured quality instrument, drafts scored between 3.97 and 4.99 out of 5 across domains, but they were rated by an AI model before clinician review, not by clinicians.Source 6
A larger observational study points the same way. Across five US academic health systems and 8,581 clinicians, of whom 1,809 adopted a scribe, adoption was associated with 16.0 fewer minutes of documentation time per 8 scheduled patient hours and 0.49 more visits a week.Source 7 Record time outside work hours did not change significantly.Source 7
A scribe does not remove the note. It turns writing into checking, and the saving is whatever the checking leaves behind.
What the evidence actually says
The measured effects are real but smaller and more uneven than vendor claims. Microsoft says clinicians report saving five minutes per encounter with DAX Copilot, based on its own survey of 879 clinicians across 340 organisations; that is vendor data, and self-reported.Source 5 In the UCLA trial, 238 physicians were randomised between DAX, Nabla and usual care. The DAX group showed no significant change in measured time per note against control, while Nabla users saw a 9.5% reduction.Source 5 In absolute terms, estimated time per note fell by 41 seconds with Nabla and 23 seconds with DAX, against 18 seconds in the control group.Source 5 The authors warn that their metric missed editing done inside the scribe software, so the saving may be overstated.Source 5
Large deployments show the same pattern. At a 14-hospital Boston system, enrolling 902 outpatient clinicians in Heidi was associated with an immediate fall of 1.21 minutes in note-active time, larger among high users, but no change in same-day or after-hours note closure or in work relative value units, a productivity measure. It was observational.Source 7 In the multisite study, gains were greatest in primary care and among clinicians using the scribe in at least half their visits.Source 7
Australian evaluations are smaller, and most did not measure clinician time. At Gold Coast Health, 100 outpatient clinicians used a scribe across 7,499 consultations over 16 weeks; 84% of the 43 who answered the staff survey reported a positive effect on efficiency, a self-reported result.Source 8 At the St Vincent's Hospital Melbourne emergency department, a modelled estimate, based on word counts and assumed typing speeds rather than timing, put the saving at 7.1 hours across 248 presentations, or 4.9 hours for fast typists. That study is vendor-linked: the lead author advises the scribe's vendor and three co-authors hold roles there, though the authors state it had no input.Source 8 On quality, blinded GP raters in a simulated Australian study found AI notes comparable to human notes overall, and found hallucinations in both.Source 8
Where it falls short
Errors still occur. At Gold Coast, 47% of staff survey respondents reported hallucinations, such as undiscussed medications, allergies and measurements.Source 8 In the UCLA trial, the one reported adverse event, graded mild, was counselling left out of the plan.Source 5
Uptake is patchy. At St Vincent's, 9 of 40 consultants and 11 of 23 trainees used the tool.Source 8
Minutes do not always become capacity. Boston found no change in work relative value units, and the multisite study found less than one extra visit a week.Source 7
Vendors have not published. The RACGP says companies selling AI scribes have not published clinical validity and safety data in academic journals.Source 1
Where governance decides the outcome
Australian guidance agrees on one point: the clinician owns the note. The RACGP says GPs remain responsible for the accuracy of records and must obtain consent first, and in some jurisdictions recording a private conversation without consent is a criminal offence.Source 1 Ahpra says the practitioner is responsible for checking the accuracy and relevance of records created using generative AI.Source 2 The TGA, in guidance updated on 16 October 2026, says a scribe that only records, transcribes and translates, without adding clinical content, is not a medical device, and it lists summarising as an administrative function; one that suggests diagnoses or treatments is likely to be, and a disclaimer does not change that.Source 3 Ahpra's page, reviewed in 2024, still says scribes are usually not TGA regulated, so the newer guidance should be checked.Source 2,Source 3 The Commission warns of automation bias, asks that records be labelled when AI was involved, and says the data must be stored and processed in Australia unless the patient explicitly consents otherwise.Source 4 The privacy regulator reports advising on the overseas transfer of health data by AI tools.Source 4
How to think about adoption
Measure before and after from the record system, not from surveys, and count editing time inside the scribe as well as in the record. Check each function against the TGA's guidance, since a feature that suggests a diagnosis changes the product's status. Set a consent process that meets state surveillance laws, and confirm where audio and notes are stored. Review a sample of notes against the recording, and track error reports. In our reading, the time is worth having, but only if the sign-off is treated as part of the work rather than an afterthought.
BFCS.ai does not fabricate figures. Time savings here are labelled as measured (from health record logs), modelled or self-reported, and they should not be added together. The two randomised trials are single health systems in the US; the largest studies are observational, and four sources (the JAMA and Annals studies and two Australian studies) were read in abstract only. Microsoft's figure is a vendor survey. We found no Australian randomised trial and no published cost-benefit analysis.
Sources
- 1
Royal Australian College of General Practitioners, "Artificial intelligence (AI) scribes" fact sheet (first published August 2024; publication information says last updated October 2025, while the page header says last updated 29 June 2026). racgp.org.au
newsGP (RACGP), "GPs' AI scribe use doubles in one year: Poll," 12 December 2025: racgp.org.au
Back to textWhat we took from it
AI scribes cannot completely replace the work GPs undertake to prepare documentation; output must be carefully checked by the GP because scribes can produce errors and inconsistencies; GPs are ultimately responsible for ensuring records are accurate and up to date; "There is limited data on the clinical utility/validity and patient safety" and companies with AI scribes on the market have not published this data in academic journals; GPs must obtain consent before using an AI scribe, some medical defence organisations advise written consent; in some Australian jurisdictions recording a private conversation without consent is a criminal offence.
22% of poll respondents were using AI scribes in August 2024 (1,170 votes) and 40% in November 2025 (1,516 votes). newsGP weekly online reader poll: respondents chose to vote, and the article does not say whether all of them were GPs.
- 2
Australian Health Practitioner Regulation Agency and National Boards, "Meeting your professional obligations when using Artificial Intelligence in healthcare" (page reviewed 22 August 2024). ahpra.gov.au
Back to textWhat we took from it
Practitioners must apply human judgment to any output of AI; TGA approval does not change that responsibility; "If using an AI scribing tool, the practitioner is responsible for checking the accuracy and relevance of records created using generative AI"; a generative AI scribe will generally require input of personal data and therefore informed consent, ideally noted in the health record; there may be criminal implications if consent is not obtained before recording. The page also states that generative AI scribing tools are usually not regulated by the TGA; that statement predates the TGA guidance in footnote 3.
- 3
Therapeutic Goods Administration, "Digital scribes: When digital scribes are regulated as medical devices" (last updated 16 October 2026). tga.gov.au
Back to textWhat we took from it
Digital scribes that only record, transcribe and translate clinical conversations into written records, without adding new clinical content or suggestions not explicitly stated by the health professional, are not medical devices; a scribe is likely to be a medical device if, for example, it identifies, recommends or suggests diagnoses, treatments or investigations, prioritises or triages patients, or generates clinical conclusions; general disclaimers such as "this product is not for diagnosis" are not acceptable where medical claims are made or the product includes features consistent with a medical device; if updates add a therapeutic function, the product becomes a medical device. The TGA's separate page on its digital scribe review outcomes could not be opened (automated access was blocked), so its findings are not used here.
- 4
Australian Commission on Safety and Quality in Health Care, "AI Safety Scenario: Ambient scribe," Version 1.0, August 2025 (PDF). safetyandquality.gov.au
Office of the Australian Information Commissioner, "Annual report of the Australian Information Commissioner's activities in relation to digital health 2025-26" (signed September 2026): oaic.gov.au
Back to textWhat we took from it
Ambient scribes record and then use generative AI to summarise a consultation, which is added to the health care record; generally not regulated by the TGA as a medical device, "nor does ambient AI scribe software have rigorous supporting evidence on safety and performance"; recordings and summaries are personal and sensitive information under the Privacy Act 1988 and "must be stored and processed in Australia unless the patient has provided explicit consent" for offshore handling; clinicians are susceptible to automation bias; all outputs must be reviewed for accuracy and quality; responsibility for the quality and accuracy of scribe summaries in records remains with the clinician; label records indicating AI was involved; report adverse events and near misses.
OAIC staff met Ahpra, the Commission, the TGA, the Australian Digital Health Agency, Austrade and the RACGP to discuss the uptake of AI scribes and provided advice on matters including the transfer of health data overseas by AI tools.
- 5
Lukac, P. J., Turner, W., Vangala, S. et al., "Ambient AI Scribes in Clinical Practice: A Randomized Trial," NEJM AI 2(12), 2025, doi:10.1056/AIoa2501000 (peer-reviewed; full text read via PubMed Central). pmc.ncbi.nlm.nih.gov
Vendor data: news.microsoft.com
Back to textWhat we took from it
Pragmatic three-group randomised trial at UCLA Health, 4 November 2024 to 3 January 2025: 238 outpatient physicians in 14 specialties, 79 assigned to Microsoft DAX Copilot, 79 to Nabla and 80 to usual care; primary outcome time-in-note, an Epic Signal metric measured from the record system, comparing December with a six-month baseline; baseline median time-in-note 4 minutes 23 seconds (DAX group), 4 minutes 47 seconds (Nabla) and 5 minutes 22 seconds (control). DAX was used in 33.5% of 24,696 visits and Nabla in 29.5% of 23,653. Nabla users had a 9.5% decrease in time-in-note versus control (95% CI -17.2% to -1.8%); DAX showed no significant change (-1.7%; 95% CI -9.4% to +5.9%). Estimated declines were 41 seconds (Nabla), 23 seconds (DAX) and 18 seconds (control). The metric does not count time spent editing drafts inside the vendor platforms, so the authors say time savings may be overestimated. Survey respondents rated clinically significant inaccuracies as occurring "occasionally"; one grade 1 (mild) adverse event, in which extensive patient counselling was left out of the assessment and plan. The authors call for a robust cost-benefit analysis. Funded by UCLA.
Microsoft, "Microsoft Dragon Copilot provides the healthcare industry's first unified voice AI assistant," news release, 3 March 2025: "clinicians reporting five minutes saved per encounter," citing a Microsoft survey of 879 clinicians across 340 healthcare organisations using DAX Copilot, July 2024. Vendor survey (self-reported, not peer-reviewed).
- 6
Afshar, M., Baumann, M. R., Resnik, F. et al., "A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence to Improve Health Practitioner Well-Being," NEJM AI 2(12), 2025, doi:10.1056/AIoa2500945 (peer-reviewed; full text read via PubMed Central). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
24-week stepped-wedge, individually randomised pragmatic trial at UW Health across ambulatory clinics in two US states and eight specialties: 66 practitioners randomised to three six-week sequences using Abridge. 71,487 notes authored, 27,092 (38%) with ambient AI. Coprimary outcomes were burnout measures; among secondary outcomes, measured from EHR audit logs and normalised to an 8-hour workday, time spent on notes fell by 0.36 hours a day (95% CI -0.55 to -0.17); a 0.50 hours a day fall in work outside work was no longer significant after removing the top 3% of daily observations. Documentation quality (PDSQI-9) mean domain scores ranged from 3.97 to 4.99 out of 5, for 7,966 randomly sampled drafts scored before practitioner review by an LLM-as-a-judge method that the authors say was validated against physician ratings, not by human raters. Funded by UW Hospital and Clinics and the NIH; no funding from the AI company, which provided technical support.
- 7
Rotenstein, L. S., Holmgren, A. J., Thombley, R. et al., "Changes in Clinician Time Expenditure and Visit Quantity With Adoption of Artificial Intelligence-Powered Scribes: A Multisite Study," JAMA 335(16), 2026, pp. 1408-1417, doi:10.1001/jama.2026.2253 (peer-reviewed; abstract and disclosures only read, via PubMed). pubmed.ncbi.nlm.nih.gov
Booth, S. E., Gould, R., Cibotti, N. et al., "Impact of Unrestricted Artificial Intelligence Scribe Deployment in a Single Health System on Clinician Workload and Productivity: A Quality Improvement Study," Annals of Internal Medicine, published online 27 October 2026, doi:10.7326/ANNALS-25-05294 (peer-reviewed; abstract only read, via PubMed). pubmed.ncbi.nlm.nih.gov
Back to textWhat we took from it
Longitudinal cohort at five US academic institutions, June 2023 to August 2025: 8,581 ambulatory clinicians including 1,809 AI scribe adopters; adoption meant receiving access, by opt-in at four of five sites. Difference-in-differences: 13.4 fewer minutes of EHR time, 16.0 fewer minutes of documentation time (both per 8 scheduled patient hours) and 0.49 more weekly visits; EHR time outside work hours did not change significantly; changes were greatest for primary care, advanced practice clinicians, female clinicians and those using scribes in 50% or more of visits. Disclosures include an author whose employer had an institutional investment in Abridge.
Retrospective interrupted time series at a 14-hospital Boston health system: 902 outpatient clinicians enrolled in Heidi (605 users, 297 enrolled non-users), 1,637,227 encounters, 1 July 2024 to 31 January 2026. Enrolment was associated with an immediate reduction in note-active time of 1.21 minutes (95% CI -1.74 to -0.68; the abstract does not state the unit, such as per note or per day), larger among high users; no differences in same-day or after-hours note closure or work relative value units; observational, and AI-scribed notes could not be distinguished from others. Primary funding source: none.
- 8
Australian evaluations. pmc.ncbi.nlm.nih.gov
Kapoor, V., Skellern, C., McGrath, P. et al., "Improving Clinical Efficiency Using Artificial Intelligence Scribe for Documentation (ICE-AID Study)," Journal of Paediatrics and Child Health, online 27 May 2026 (peer-reviewed; abstract only read, via PubMed): pubmed.ncbi.nlm.nih.gov
Akhlaghi, H., Freeman, S., Sun, K. et al., "Evaluation of an AI Scribe Tool in the Emergency Department: A Single-Arm Observational Study," Emergency Medicine Australasia 38, e70272, 2026 (peer-reviewed; full text read via PubMed Central): pmc.ncbi.nlm.nih.gov
Foo, D., Tan, J., Stevens, S. et al., "A comparative analysis of AI scribes versus human documentation in simulated general practice consultations," Australian Journal of General Practice 55(4), 2026, pp. 226-232 (peer-reviewed; abstract only read, via PubMed): pubmed.ncbi.nlm.nih.gov
Back to textWhat we took from it
Memon, S., Brand, A., Taylor, B. et al., "Performance, acceptability, and impact of ambient listening scribe technology in an outpatient context: a mixed methods trial evaluation," BMC Health Services Research 26, 182, 2026 (peer-reviewed; full text read via PubMed Central): Gold Coast Hospital and Health Service, 16-week trial from July 2024, 100 outpatient clinicians, 7,499 consultations; staff survey 43 responses (43%); 84% of respondents reported a positive impact on efficiency (self-reported); 47% reported hallucinations, of whom 20% said they occurred frequently, including undiscussed medications, investigations, allergies and measurements; a text-overlap (ROUGE) analysis of 21 draft and final note pairs gave a median F1 of 0.58, which the authors report as 58% of AI text used verbatim; authors declare no competing interests.
quaternary children's hospital and health service; 131 participants; transcription cost savings of AUD 16,903 but no definitive impact on overtime claims. The abstract's letter-turnaround figure is not used because its unit could not be confirmed.
St Vincent's Hospital Melbourne emergency department, 2 June to 6 July 2025; 9 of 40 consultants and 11 of 23 trainees used the tool, for 248 presentations; modelled estimate (from word counts and assumed typing and reading speeds, not timed) of 7.1 hours saved across the 248 presentations at average typing speed and 4.9 hours for fast typists. Vendor-linked: the lead author is a medical advisor to the vendor, NexusMD.ai, and three co-authors hold roles there (chief technology officer, chief product officer and a medical director); the vendor provided the software; the authors state the vendor had no funding, input or influence.
RACGP exam cases with four patient actors, two GPs and three blinded GP raters; AI notes comparable or superior overall but not statistically significant (P = 0.071); both AI and human notes contained hallucinations.