Evidence · strength 2 of 5 · searched to 2 Nov 2026
Do AI scribes reduce doctors' paperwork?
No clear effect so far: the studies do not show that AI scribes reduce clinicians' time on notes. 3 of 6 studies found a clear gain in time on notes.
What randomised and controlled studies found when clinicians used ambient AI scribes that draft notes from the consultation: time on notes and burnout.
VerdictNo clear effect so far
No good evidence. Studies so far find no clear effect either way on the headline outcome.
ImpactNo clear effect
3 of 6 studies found a clear gain in time on notes.
Evidence strength2 of 5
We have low confidence in this estimate.
Cost per user a year?
Scribe products are priced by contract; published list prices are rare. Too varied to band.
Pilot rating under our v0.3 rules. The rules are published on our method page but not yet independently reviewed, so ratings may change as the method is checked and new studies arrive. Our method.
What the studies found
Time (headline)
No clear effect- Lukac et al. (2025): −9.5% Time-in-note per note (Epic Signal), Nabla vs control, ITT (95% CI -17.2% to -1.8%; P=0.02).
Nabla users experienced a 9.5% (95% confidence interval [CI], −17.2% to −1.8%; P=0.02) decrease in time-in-note versus the control group
- Lukac et al. (2025): −1.7% Time-in-note per note (Epic Signal), DAX vs control, ITT (95% CI -9.4% to +5.9%; P=0.66). Not statistically clear.
DAX users exhibited no significant change versus the control group (−1.7%; 95% CI, −9.4% to +5.9%; P=0.66).
- Afshar et al. (2025): −0.36 hours per day; about -21.9 minutes Time spent on notes per 8 hours of patient time (EHR audit log) (95% CI -0.55 to -0.17).
time spent on notes decreased (−0.36 hours per day; 95% CI, −0.55 to −0.17).
- Afshar et al. (2025): −0.5 hours per day Work outside work (EHR time outside scheduled hours) (95% CI -0.90 to -0.09; not significant after removing top 3% of observations (-0.19; -0.40 to 0.03)). Not statistically clear.
The reduction in WoW (−0.50 hours per day; 95% CI, −0.90 to −0.09) was sensitive to exclusion of extreme values
- Rotenstein et al. (2026): −13.4 minutes per 8 scheduled patient hours Total EHR time per 8 scheduled patient hours (difference-in-differences, adopters vs non-adopters) (95% CI 9.1 to 17.7 fewer minutes).
AI scribe adoption was associated with 13.4 (95% CI, 9.1-17.7) fewer minutes of EHR time, 16.0 (95% CI, 13.7-18.3) fewer minutes of documentation time
- Rotenstein et al. (2026): −16 minutes per 8 scheduled patient hours Documentation time per 8 scheduled patient hours (DiD) (95% CI 13.7 to 18.3 fewer minutes).
AI scribe adoption was associated with 13.4 (95% CI, 9.1-17.7) fewer minutes of EHR time, 16.0 (95% CI, 13.7-18.3) fewer minutes of documentation time
- Rotenstein et al. (2026): −10% Documentation time per 8 scheduled patient hours, relative change (difference-in-differences).
- Martini et al. (2026): −2.4 minutes per note Active EHR time in the note, AI-scribed vs the same physician's unassisted notes (adjusted). Not statistically clear.
Physicians spent less active time in AI-scribed notes (median 10.0 versus 12.5 min-2.4 minutes after adjustment)
- Shin et al. (2025): −9.4 minutes per scheduled day After-hours EHR work ('pajama time'), difference-in-differences (95% CI -41.2 to +22.4). Not statistically clear.
There was no significant change to pajama time (-9.4 minutes per scheduled day, 95% CI: -41.2 to +22.4)
- Shin et al. (2025): +0.2 minutes per note Time in notes per encounter, difference-in-differences (95% CI -6.6 to +6.9). Not statistically clear.
time in notes per encounter (+0.2 minutes per note, 95% CI: -6.6 to +6.9)
- Pearlman et al. (2025): −8.5% Mean EHR time per appointment, users versus weighted non-users (95% CI -12.8% to -3.9%; P < .001).
AI scribe use was associated with an 8.5% (95% CI, -12.8% to -3.9%; P < .001) lower mean EHR time (ie, 2.4 minutes)
- Pearlman et al. (2025): −15.9% Mean time in notes per appointment, users versus weighted non-users (95% CI -21.2% to -10.4%; P < .001).
and a 15.9% (95% CI -21.2% to -10.4%; P < .001) lower mean time in notes (ie, 1.8 minutes) with no significant differences in other outcomes
Errors
No clear effect- Etagiuri et al. (2026): +10 count of medical errors Medical errors found in clinic letters (surgeon review), AI scribe vs conventional dictation. Not statistically clear.
There were 19 medical errors in the AI scribe group compared to only nine errors in the conventional group.
- Martini et al. (2026): −0.8 points 72-hour return to the ED (95% CI -1.5 to -0.1).
72-hour ED return was not increased (-0.8 percentage point; 95% confidence interval -1.5 to -0.1).
Output
Gain- Rotenstein et al. (2026): +0.49 visits per week Weekly visits delivered (DiD) (95% CI 0.17 to 0.81).
and 0.49 (95% CI, 0.17-0.81) additional weekly visits delivered.
- Shin et al. (2025): −0.03 wRVUs per encounter Work RVUs per encounter, difference-in-differences (95% CI -0.5 to +0.44). Not statistically clear.
or work Relative Value Units (wRVUs) per encounter (-0.03, 95% CI: -0.5 to +0.44)
- Shi et al. (2026): +10.5 points Share of encounters with written patient instructions after scribe availability (difference-in-differences) (P = .018).
AI scribe availability was associated with a 10.5 percentage-point increase in patient instructions documentation frequency (P = .018)
- Shi et al. (2026): +4.32 adjusted odds ratio Odds of patient instructions when the scribe was used in the encounter (95% CI 4.14 to 4.50).
encounter-level AI use was associated with higher odds of documentation (aOR 4.32; 95% CI, 4.14-4.50)
Quality
Gain- Afshar et al. (2025): +0.93 points on 0-10 scale; 6.87 vs 5.94 ICD-10 diagnosis billing compliance score (0-10), professional coders, AI-generated vs human-authored notes (AI 95% CI 6.76 to 6.98; human 95% CI 5.81 to 6.07; P<0.001).
a mean ICD-10 compliance score of 6.87 (95% CI, 6.76 to 6.98) for ambient AI generated notes versus 5.94 (95% CI, 5.81 to 6.07)
Wellbeing
Gain- Lukac et al. (2025): +2.83 points on 10-50 scale Mini-Z 2.0 total score (10-50, higher is better), DAX vs control (DAX 95% CI +1.28 to +4.37; Nabla 95% CI +1.14 to +4.23). Self-reported.
Increases in total Mini-Z (scale 10–50; DAX 2.83 [95% CI, +1.28 to +4.37]; Nabla +2.69 [95% CI, +1.14 to +4.23])
- Lukac et al. (2025): −0.32 points on 0-4 scale Work exhaustion (Professional Fulfillment Index, 0-4, lower is better), DAX vs control (DAX 95% CI -0.55 to -0.08; Nabla 95% CI -0.46 to +0.01). Self-reported.
PFI-WE (scale 0–4; DAX 0.32 [95% CI,−0.55 to −0.08]; Nabla −0.23 [95% CI, −0.46 to +0.01])
- Lukac et al. (2025): −39.9 points on 0-400 scale Physician task load (0-400, lower is better), DAX vs control (DAX 95% CI -71.9 to -7.9; Nabla 95% CI -63.8 to +0.4). Self-reported.
reductions in PTL (scale 0–400; DAX −39.9 [95% CI, −71.9 to −7.9]; Nabla −31.7 [95% CI, −63.8 to +0.4])
- Lukac et al. (2025): +2.69 points on 10-50 scale Mini-Z 2.0 total score (10-50, higher is better), Nabla vs control (95% CI +1.14 to +4.23). Self-reported.
- Lukac et al. (2025): −0.23 points on 0-4 scale Work exhaustion (0-4, lower is better), Nabla vs control (95% CI -0.46 to +0.01). Self-reported. Not statistically clear.
- Lukac et al. (2025): −31.7 points on 0-400 scale Physician task load (0-400, lower is better), Nabla vs control (95% CI -63.8 to +0.4). Self-reported. Not statistically clear.
- Afshar et al. (2025): −0.44 points on 5-point scale Work exhaustion/interpersonal disengagement (Stanford PFI, 1-5 scale; co-primary) (95% CI -0.62 to -0.25; P<0.001). Self-reported.
Ambient AI use had a significant reduction in work exhaustion/interpersonal disengagement (−0.44 points; 95% confidence interval [CI], −0.62 to −0.25; P<0.001)
- Afshar et al. (2025): +0.14 points on 5-point scale Professional fulfilment (Stanford PFI, 1-5 scale; co-primary) (95% CI 0.004 to 0.28; P=0.04 (not significant at the pre-set 0.025)). Self-reported. Not statistically clear.
a nonsignificant increase in professional fulfillment (+0.14 points; 95% CI, 0.004 to 0.28; P=0.04) on a five-point Likert scale.
- Afshar et al. (2025): −6.72 points Task load (secondary) (robust 95% CI -10.43 to -3.00; adjusted P=0.001). Self-reported.
ambient AI was associated with reduced task load (−6.72; robust 95% CI, −10.43 to −3.00; adjusted P=0.001)
- Shin et al. (2025): +29.9 points on KLAS Net EHR Experience KLAS Net EHR Experience score, change in DAX group minus change in controls. Self-reported. Not statistically clear.
KLAS Net EHR Experience improved among intervention providers from 52.6 (70th percentile) to 75.2 (99th percentile) but dropped from 37.3 (38th percentile) to 30 (14th percentile) among control providers.
- Shin et al. (2025): −33 points Share experiencing burnout, change in DAX group (control unchanged). Self-reported. Not statistically clear.
Experiencing burnout dropped from 8 (89%) to 5 (56%) among intervention providers but remained stable at 3 (38%) in the control group.
A minus sign on time means the task took less time. Bands are provisional; see how impact is banded. Effects are not pooled yet: each band is set from one result per study, and where studies use different units we show only the direction.
How sure are we?
- Starting level3 of 5, from 2 grade A or B studies.
- Down one: old models0 of 8 studies show they used a model released since May 2025 (the 18-month window); the rest used older models or do not say which.
- Publication biasnot testable with fewer than 10 studies.
- Result2 of 5. We have low confidence in this estimate.
Behind the average
- First month of use (November) (Lukac et al. (2025)): Smaller, non-significant reductions in the first month (Nabla -5.0%, DAX 0.0%), suggesting a learning curve
- Practitioners above the high-burnout threshold (Afshar et al. (2025)): Share fell from 56.1% at baseline to 35.4% by the end of the study (descriptive)
- Primary care clinicians (Rotenstein et al. (2026)): 25.0 fewer minutes of EHR time and 26.9 fewer minutes of documentation time per 8 patient hours
- Female clinicians (Rotenstein et al. (2026)): 19.0 fewer EHR minutes, 19.9 fewer documentation minutes and 6.2 fewer after-hours minutes
- Top quartile of use (>80% of appointments) (Pearlman et al. (2025)): EHR time per appointment fell 5.9 minutes within this group (median 29.0 to 23.1)
- Male clinicians (Pearlman et al. (2025)): EHR time -5.1% (95% CI -11.2 to 1.4; P = .12), note time -8.2% (-16.3 to 0.8; P = .07); no clear reduction
- Offsite (not hospital-based) clinicians (Pearlman et al. (2025)): EHR time -2.4% (P = .54), note time -5.8% (P = .24); no change
- Surgical or procedural clinicians (Pearlman et al. (2025)): EHR time -2.5% (P = .57), note time -10.0% (P = .08); no significant change
- 10 or fewer years in practice (Pearlman et al. (2025)): EHR time -3.1% (P = .31); note time -11.1% (-17.9 to -3.8; P < .001) in Table 3 although the text says 'no changes'
- Primary care (pre-specified) (Shi et al. (2026)): +16.6 percentage points (SD 3.9; no P reported)
- Non-primary care (pre-specified) (Shi et al. (2026)): +4.9 percentage points (SD 3.6; no P reported)
How, where and why it works
Why it works
2 of 4. The AI drafts the note during the visit, so less typing after it; studies measure time in the record system.
For whom and where
2 of 4. One multi-site study found larger savings for primary care clinicians; one trial found little effect in the first month.
What it takes
2 of 4. Studies report training, consent from patients and a learning curve of weeks.
The full cost
1 of 4. Time editing drafts inside the scribe app is often not counted, so savings may be overstated.
Four separate scores from 0 to 4, never added up: how much the studies tell us about why it works, for whom, what it takes and what it really costs (the EMMIE method).
What to do next
Measure your pilot →If you are trying it yourself
- Read and correct every drafted note before signing it.
- Tell patients and record their consent.
If you are rolling it out
- Compare products head-to-head, as one trial did.
- Measure note time from audit logs and burnout with validated surveys.
- Check note quality, not just speed.
What it takes
- Patients must be told and agree before recording.
- Expect a learning curve of several weeks.
- Clinicians remain responsible for every note.
What can go wrong
- Errors or omissions in drafted notes.
- Time spent editing inside the scribe app is often not measured.
The studies
| Study | Grade | Design | People | AI model (released) | Company role |
|---|---|---|---|---|---|
| Ambient AI Scribes in Clinical Practice: A Randomized Trial Paul J. Lukac, William Turner, Sitaram Vangala, Aaron T. Chin, Joshua Khalili, Ya-Chen Tina Shih, Catherine Sarkisian, Eric M. Cheng, John N. Mafi, 2025. NEJM AI, 26 Nov 2025 (journal) | B | field RCT, deployed work | 238 outpatient physicians, 14 specialties, one academic health system (UCLA): DAX 79, Nabla 79, usual-care control 80 | Microsoft Dragon Ambient eXperience (DAX) Copilot version 2.0 (underlying LLM not stated); Nabla version 1.5 (underlying LLM not stated) | none |
| A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence to Improve Health Practitioner Well-Being Majid Afshar, Mary Ryan Baumann, Felice Resnik, et al. (22 authors; last author Joel Gordon), 2025. NEJM AI, 26 Nov 2025 (journal) | B | field RCT, deployed work | 66 health care practitioners (attending physicians and advanced practice practitioners), ambulatory clinics of one academic health system in Wisconsin and Illinois; 71,487 notes, 27,092 generated with ambient AI | Abridge (general-availability release as of 10 Oct 2024; components include bespoke LLMs) | none |
| Changes in Clinician Time Expenditure and Visit Quantity With Adoption of Artificial Intelligence-Powered Scribes: A Multisite Study Lisa S. Rotenstein, A. Jay Holmgren, Robert Thombley, et al. (31 authors; last author Rebecca G. Mishuris), 2026. JAMA 2026 (journal) | C | other quasi-experiment, deployed work | 8,581 ambulatory clinicians including 1,809 AI scribe adopters, at 5 US academic health systems | Ambience, Nuance DAX Copilot or Abridge, individually or in combination (underlying LLMs not stated) | none |
| Evaluating the Impact of Artificial Intelligence Scribes on Clinical Workflow and Documentation Quality: A Randomized Controlled Trial Abdu Etagiuri, Daniel Chaudhry, Jano van der Merwe, Michaela Nickol, Johannes M. van der Merwe, 2026. The Journal of Arthroplasty (journal) | C | RCT, deployed work | 304 patients in one surgeon's adult reconstruction practice (140 conventional documentation, 139 AI scribe analysed) | AI scribe (product not named in the abstract) | none |
| Ambient Artificial Intelligence Scribe Use and Objective Documentation Effort Among Emergency Physicians Martini, Heaton, Walker, Urumov, Rappaport, Hodgson, 2026. Annals of Emergency Medicine (journal) | C | other quasi-experiment, deployed work | 275 attending emergency physicians across 20 EDs (3 academic, 17 community); 31% of encounters AI-scribed | Ambient AI scribe (product not named in the abstract) | none |
| The Influence of Artificial Intelligence Scribes on Clinician Experience and Efficiency among Pediatric Subspecialists: A Rapid, Randomized Quality Improvement Trial Shin, Williams, Braykov, Jahan, Meller, Orenstein (surnames as on PubMed; Emory University and Children's Healthcare of Atlanta), 2025. Applied Clinical Informatics (journal) | C | field RCT, deployed work | 23 pediatric subspecialists randomised (12 DAX, 11 control); 17 completed both surveys | Microsoft/Nuance DAX Copilot | none |
| Use of an AI Scribe and Electronic Health Record Efficiency Pearlman, Wan, Shah, Laiteerapong (surnames as on PubMed; University of Chicago), 2025. JAMA Network Open 8 (journal) | C | other quasi-experiment, deployed work | 125 AI scribe users and 478 propensity-weighted non-users (ambulatory attendings and APPs) | Abridge AI scribe | none |
| Association of ambient artificial intelligence scribe availability with patient instruction documentation and readability in ambulatory care Shi, Creasman, Schmajuk, Oates, Byron, Garcia-Agundez, Yazdany (surnames as on PubMed; UCSF), 2026. Journal of the American Medical Informatics Association (journal) | C | other quasi-experiment, deployed work | 391,061 adult ambulatory internal medicine encounters at one medical centre; 399 clinicians (146 AI adopters, 253 non-adopters) | Ambience Healthcare AI scribe | none |
Grades: A, a well-run randomised trial in real work; B, a strong natural experiment in real work or a well-run trial on realistic tasks; C, a trial with clear limits or a matched comparison; D, before-and-after, surveys or case studies. How studies are graded.
Studies we found but do not count
- Comparing ambient scribes: a randomized crossover clinical trial addressing ambient scribe technologies' impact on physician burnout (Chowdhury et al. (2026), grade C). It compares two AI scribes with each other; its comparison with no scribe is before-and-after only. Shown as context.
These studies are real and often cited, but they fall outside our rules for this rating. What counts as a study.
Does the rating hold up?
- Run without company-run studiesNo change: no such studies.
- Run without working papers and preprintsNo change: no such studies.
How up to date is this evidence?
Method →Evidence searched to 2 Nov 2026. The studies do not all state model release dates. Rating computed 5 Nov 2026 under Module 15 v0.3 (pilot).
How this page was made: AI research agents searched for studies and read each paper; a second pass checked every number on this page against the paper itself, with the quote and where it appears. The rating is computed by fixed rules, not chosen. No person outside AI Adoption has reviewed it yet. This page is not yet in search engines: health, legal and finance pages wait for a named domain reviewer, and none is appointed yet. Editorial owner: AI Adoption editors. Spot a mistake? Corrections.
Disclosure: AI Adoption uses AI tools, including Claude from Anthropic, in its research and editorial work. No company pays for or has any say in a rating.