The randomized answer
The first randomized trial of ambient AI scribes appeared in NEJM AI in November 2025. Paul Lukac and colleagues at UCLA Health assigned 238 outpatient physicians across 14 specialties, one to one to one, to Microsoft's DAX Copilot, to Nabla, or to usual care, and ran the trial across roughly 72,000 encounters between 4 November 2024 and 3 January 2025.
Nabla users cut time-in-note by 9.5% against control (95% CI, -17.2% to -1.8%; P=.02). DAX users showed no significant change (-1.7%; 95% CI, -9.4% to +5.9%; P=.66). In absolute terms the Nabla arm moved from 4 minutes 30 seconds a note to 3 minutes 49 seconds, a gain of 41 seconds, while the control arm improved by 18 seconds over the same weeks.
Two products with near-identical positioning, in one trial, at one institution, in the same months, and one of them moved the primary outcome. A procurement conversation that treats ambient scribes as a single category with a single expected benefit has already lost the distinction the only randomized evidence draws.
Two cohorts, one modest picture
The largest observational study appeared in JAMA in April 2026. Lisa Rotenstein and colleagues followed 8,581 ambulatory clinicians at five US academic health systems, 1,809 of them adopters, in a difference-in-differences design. Adoption was associated with 16.0 fewer minutes of documentation time per eight scheduled patient hours (95% CI, 13.7 to 18.3) and 13.4 fewer minutes of total record time (95% CI, 9.1 to 17.7). Time in the record outside scheduled hours held steady.
A University of Chicago cohort published in JAMA Network Open in October 2025 compared 125 scribe users with 478 weighted controls using Epic telemetry. Total record time per appointment fell 8.5%, around 2.4 minutes, and time in notes fell 15.9%, around 1.8 minutes. After-hours documentation, visit volume and appointment length each held steady.
That study also teaches something in its own two analyses. Among users alone, time to close an encounter fell by a median of 7.1 hours. Under propensity weighting against controls, the same outcome lost statistical significance. One study, one dataset, and the headline depends on which comparison gets quoted.
The authors are equally direct about what their instrument can see. Epic telemetry separates active engagement from a record left open during a face-to-face encounter poorly, which they judge likely to understate the reductions they measured. A measurement standard stated that plainly is worth more than a larger number stated without one.
Where the saved time lands
A prospective time-motion study published in JMIR Medical Informatics in March 2026 put five trained observers into 169 consultations at an academic medical center in Singapore, with nine clinicians each observed both with and without an in-house ambient scribe. Documentation time per consultation fell 15.0%, from a mean of 5.3 minutes to 4.5 (P=.04).
Consultation duration held steady, at a mean of 11.5 minutes against 10.9 (P=.42). Total cycle time per patient, counting pre-clerking and post-consultation documentation, also held steady, at 13.7 minutes against 13.2 (P=.57). What moved was attention: the share of the consultation spent in eye contact rose from 69.6% to 77.1% (P=.009).
The large cohorts agree. The JAMA study found adopters delivered 0.49 additional visits per week (95% CI, 0.17 to 0.81), about one more patient a fortnight. The Chicago cohort found no change in visit volume at all. Across three designs on two continents, after-hours documentation stayed where it was.
So the saving lands inside the consultation and in the attention a patient receives. That is a real benefit with a precise name. Booking it as recovered appointment capacity puts a revenue forecast on top of half a visit a week.
The largest number, divided
The Permanente Medical Group reported the field's biggest deployment figure in NEJM Catalyst in April 2025: more than 2.5 million ambient scribe uses in the first year, and an estimated 15,791 hours of documentation time saved, which the group described as 1,794 eight-hour workdays. This is single-group deployment data with no control comparison, and the paper presents it as such.
Divide the two published figures and the same result reads differently. Those 15,791 hours across 2.5 million uses come to roughly 23 seconds a visit. Both descriptions are accurate. One of them sizes a purchase.
A Stanford emergency department shows the same arithmetic running the other way. In JMIR AI in July 2026, scribe use across 10,344 encounters handled by 100 attending physicians was associated with a 72.6-second reduction in on-shift documentation per encounter (95% CI, 63.8 to 81.4). The authors then observe that across twenty encounters this comes to roughly 24 minutes per eight-hour shift.
Both figures come out of one regression, and the distance between them is an assumption about how many encounters a physician documents. Stated in the paper with that assumption attached, it is a fair extrapolation. Repeated in a deck without it, it becomes a number with no measurement standard behind it.
Burnout, stated at its real strength
In JAMA Network Open in October 2025, Kristine Olson and colleagues surveyed clinicians at six US health systems before and after 30 days with the same ambient platform. Among the 263 analyzed, the proportion reporting burnout fell from 51.9% to 38.8% (odds ratio, 0.26; 95% CI, 0.13 to 0.54). Note-related cognitive load and undivided attention on the patient each improved by more than two points on ten-point scales.
Two qualifications belong in the citation. The design is a pre-and-post quality improvement study built on self-report, with no control arm. Several authors were employed by or advising the vendor, which the paper discloses.
The randomized trial points the same way and stops short of settling it. Users of either scribe improved on the Mini-Z scale (DAX 2.83; Nabla 2.69, both intervals excluding zero), while the task-load and work-exhaustion intervals in the Nabla arm crossed zero. The authors call these secondary findings suggestive and ask for larger multicenter trials before anyone treats them as confirmed.
The Peterson Health Technology Institute, an independent evaluator, reached a compatible position in March 2025: early evidence on burnout and cognitive load is positive, while health system leaders report gaps in the evidence on productivity and financial performance. Well-being is the strongest claim this technology has, and it is worth stating at exactly that strength.
One revenue number, three explanations
In JAMA Network Open in January 2026, A. Jay Holmgren and colleagues at UCSF examined roughly 1.2 million ambulatory encounters across 1,565 physicians, 698 of whom adopted a scribe. Access was associated with a 5.8% rise in weekly relative value units and 2.8% more encounters per week, with no increase in claim denials. The team put the gain at about 1.81 RVUs a week, which UCSF translated to roughly $3,000 per physician per year.
The invited commentary published alongside it, by Shreya Shah and Patricia Garcia of Stanford, declines to say what the rise means. It asks whether higher RVUs reflect more services delivered, improved documentation accuracy, or fuller capture of complexity through hierarchical condition category coding. It then notes that recent downcoding policies at Aetna and Cigna may signal a coding arms race that could offset the gains.
That is a large, statistically sound, journal-published number whose business meaning is genuinely unresolved by the journal that published it. A real result and a usable result are separate things, and the gap between them is where most AI business cases quietly fail.
Four questions, then a gate
What was the comparison? A before-and-after on volunteers who chose the tool measures their enthusiasm alongside the software; the JAMA study used non-adopting colleagues, Chicago used weighted controls, and UCLA randomized. What is the denominator? Per encounter, per eight scheduled hours, per week, per adopting clinician, per licensed clinician. In the UCLA trial the scribes were used in 33.5% and 29.5% of eligible visits, so a per-use figure quoted to a buyer paying for every seat overstates the return by the whole adoption gap.
Who held the stopwatch? Record telemetry, a self-report survey and a trained observer with a clipboard measure three overlapping things, and the Chicago authors say which direction their instrument errs in. What was left out? A UC Davis pilot in JMIR Medical Informatics in April 2026 had 31 physicians generate 7,545 notes and evaluate 356 of them, finding omissions in 18%, hallucinations in 11.5%, and 5.3% of notes holding an error rated as serious or imminent risk if left uncorrected. Physicians changed a median of 9.0% of the generated words and filed 14.9% of notes unedited. Review effort is a cost line, and it sits inside every time figure above.
We score a system against the client's own data, at something close to real volume, on criteria written down before the run. Each condition answers something this literature demonstrates. Own data, because the measured effect varied by product, by specialty and by clinician group in every study here. Real volume, because the gains concentrate in clinicians who use the tool in most of their visits, and about a third of adopters in the JAMA cohort reached that level.
Criteria fixed in advance, because a technology with four legitimate success metrics offers four chances to pick the flattering one once the result is in. The honest case for ambient documentation reads: documentation time down by single to low double digits, the field's strongest evidence sitting on clinician well-being, an accuracy profile that requires a clinician to read every note, and a capacity gain near zero. In many organizations that is worth funding. It is also a case a scored gate can settle on your own clinicians, in weeks, before a multi-year licence is signed.
Why we run the gate first
We build and run production AI for organisations where a wrong answer carries real cost, and healthcare is where that bar gets tested. The literature above is the reason an engagement here opens on a measurement. A technology can be genuinely worth deploying and still return a third of the figure in the deck, and the only way to learn which applies to you is to run it against your own records.
So a build begins with a proof gate. The system runs on the client’s data, at something close to their real volume, scored on criteria written down before the run. The gate is built to be failed: an answer of "this does not clear" arrives in weeks, on fixed scope, while the budget for the full build is still uncommitted. That is the cheapest useful answer available in the whole engagement.
For ambient documentation we would score well-being and documentation time as two separate outcomes, and we would keep throughput out of the business case until a clinic has measured it in its own schedule. The evidence carries the first two comfortably. The third has been looked for by several research groups on two continents and has yet to appear.
