Ambient AI scribes, software that records a clinic visit and drafts the visit note, have now been tested in randomized trials. In a UCLA trial published in NEJM AI on November 26, 2025, with 238 outpatient physicians, one scribe product cut time spent writing notes by 9.5 percent compared with usual care, while the other showed no clear difference (a 1.7 percent drop that could be chance). Measures of burnout and task load moved in a favorable direction, but the authors did not run formal tests on them.
Key takeaways
- The UCLA trial randomly assigned 238 physicians in 14 specialties to one of two scribe products or to usual care between November 4, 2024, and January 3, 2025.
- Time in notes per note fell by about 41 seconds in one scribe arm, about 23 seconds in the other, and about 18 seconds in the control group.
- Only the first product's time reduction was statistically distinguishable from the control group.
- Physicians rated clinically significant inaccuracies as occasional, and omissions were the most cited error type.
- One site, English-language visits, and a two-month window limit how far the results travel.
What an ambient scribe does
An ambient scribe listens to the conversation between clinician and patient, then produces a draft note inside the electronic health record. The clinician can edit the draft before it becomes part of the chart. The technology targets note writing, a routine documentation task in outpatient care. The 2025 trials compare physicians offered a scribe with a randomized control group that kept working as usual, which is a stronger design than a before-and-after pilot.
How the UCLA trial was built
The trial, reported by Lukac and colleagues, ran at a single academic institution and was pragmatic, meaning physicians used the tools in ordinary clinics rather than in a lab. Randomization was constrained so that the three groups started with similar baseline note time, burnout scores, and clinic days per week. Only English-language visits were included.
| Group | Physicians | Share of visits where the scribe was used |
|---|---|---|
| Microsoft DAX Copilot v2.0 | 79 | 33.5 percent of 24,696 visits |
| Nabla v1.5 | 79 | 29.5 percent of 23,653 visits |
| Usual care (control) | 80 | Not applicable (24,020 visits) |
About 15 percent of physicians assigned to a scribe never used it. Because the main analysis counted everyone who was assigned a tool, whether or not they used it every day, the results describe what happens when a clinic offers a scribe, not what happens in each visit where it runs.
What changed in documentation time
The primary outcome was time-in-note, measured by the electronic record system as minutes spent writing notes divided by the number of notes, compared with each physician's prior six-month baseline. Relative to the control group, the Nabla arm showed a 9.5 percent reduction (95% confidence interval 1.8 to 17.2 percent, P=0.02). The DAX arm showed a 1.7 percent reduction, with an interval running from a 9.4 percent drop to a 5.9 percent rise, so the data are compatible with no effect. In absolute terms, average time per note fell from about 4 minutes 30 seconds to 3 minutes 49 seconds in the Nabla group, a saving of roughly 41 seconds. Controls improved by about 18 seconds without any scribe, which shows how much ordinary variation exists over a few months. Time spent on unscheduled days and outside scheduled hours did not differ clearly between groups.
Burnout, task load, and what was not tested
Secondary outcomes pointed the same direction for both products. A composite well-being score (the Mini-Z, scored 10 to 50) rose by about 2.8 points in the DAX arm and 2.7 in the Nabla arm. Work exhaustion, scored 0 to 4, fell by 0.32 points with DAX, with an interval that excluded zero, and by 0.23 with Nabla, with an interval that just reached zero. Self-reported task load fell by 39.9 points on a 0 to 400 scale with DAX, and by 31.7 points with Nabla, again with an interval touching zero. The authors did not test these secondary outcomes as hypotheses and described them as needing confirmation in larger, multicenter trials. Single-item burnout showed odds ratios below one for both products, but the intervals were wide and included no change.
Errors and the limits of the time metric
Physicians rated clinically significant inaccuracies as occasional, scoring them 2.7 for DAX and 2.8 for Nabla on a five-point scale, and bias as rare. The most cited error types were omissions, problems with structure or content, and mistakes in resolving pronouns; others included wrong speaker attribution and invented content. One mild adverse event was reported, in which extensive patient counseling did not appear in the assessment and plan or the patient instructions. Physicians rated the drafts as at least as good as their own only at a neutral level.
The authors list several limits. The electronic record's metric does not capture time spent editing a scribe's draft, so time savings may be overstated. The trial ran for two months at one institution, contract constraints limited trial length and user numbers, and the trial registration was submitted on December 9, 2024, after the study began on November 4, 2024.
A second randomized trial at UW Health
A separate NEJM AI trial from the University of Wisconsin-Madison, led by Majid Afshar, randomized 66 providers at UW Health across a 24-week stepped-wedge schedule in which each clinician gained access to an ambient scribe at week one, seven, or 13. Its primary outcomes were professional fulfillment and a burnout composite. A summary of the authors' January 30, 2026, grand rounds talk, posted by the NIH Collaboratory's Rethinking Clinical Trials program, reports lower burnout and a trend toward better fulfillment. The summary lists a burnout reduction of about 20 percent, a trend toward better professional fulfillment that was not a clearly confirmed effect, 21 fewer minutes in notes, and about 30 fewer minutes of work outside of work. It does not say over what period those minutes are counted, and secondary write-ups disagree on whether the fulfillment result was statistically significant, so confidence intervals and the primary estimates from the full paper are left out here.
As of October 2026, these two NEJM AI reports are the randomized evidence this article can verify, and no newer completed randomized trial of ambient scribes turned up in a search of trial registries and the literature. Both ran in 2024 and 2025, on specific products and at single health systems, so their numbers describe those tools and settings rather than every scribe now on the market.
Questions a clinic or patient can ask
- Is the visit audio stored, for how long, and who can access it?
- Does a clinician review and edit every draft before it enters the record?
- How does the clinic track omissions and errors after rollout?
- Were the reported benefits measured in a randomized comparison or in a vendor pilot?
- What happens to patients who prefer that a visit not be recorded?
This article gives general information about published research. It is not medical or legal advice, and it does not recommend or rank any product. Decisions about care belong with a patient and clinician, and questions about privacy duties belong with the clinic's privacy office or a qualified professional.
FAQ
Do AI scribes save doctors time?
The evidence so far shows small savings that vary by product. In the UCLA trial, one product reduced time per note by about 41 seconds against about 18 seconds for controls, and the other was not clearly different from usual care.
Does the clinician still write the note?
The scribe produces a draft, and the physicians in the UCLA trial could edit it. The trial's time metric did not capture editing time, which is one reason its savings may be overstated.
Are AI scribes accurate?
Physicians in the UCLA trial rated clinically significant inaccuracies as occasional, and omissions were the most common problem. Review of each draft by the clinician remains the safeguard.
Does one trial settle whether scribes reduce burnout?
No. The burnout measures in the UCLA trial were secondary outcomes that the authors did not test formally, and the second trial in Wisconsin had only 66 providers. Larger multicenter studies are still needed.
Related reading
- How Canada regulates and uses AI in health care
- Where bias enters health AI
- AI in nursing education
- AI in orthodontics and dental care
More in AI in health.
Last verified: October 2026
Sources
- Lukac PJ et al., Ambient AI Scribes in Clinical Practice: A Randomized Trial, NEJM AI, November 26, 2025 (author manuscript in the free archive, PMC12768499)
- Grand Rounds January 30, 2026: A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence to Improve Health Practitioner Well-Being, NIH Pragmatic Trials Collaboratory, Rethinking Clinical Trials, February 2026
- Afshar M et al., A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence to Improve Health Practitioner Well-Being, NEJM AI, 2025 (identified through the grand rounds summary; full text not retrieved)
- US federal trial registry record NCT06792890, A Randomized Controlled Trial of Ambient Artificial Intelligence Scribe Technologies, U.S. National Library of Medicine