A trial of two AI scribes found different gains in doctors’ note-writing time

Two ambient AI scribes produced different results in a randomised UCLA Health trial. Nabla reduced physicians’ note-writing time by 9.5% relative to usual care, while Microsoft DAX’s smaller change was not statistically significant. Both tools record the appointment and produce a draft for the physician to review; the trial measured that documentation workflow, rather than assuming a faster draft means a shorter working day.[1][2]

The trial ran from November 2024 to January 2025 across 238 physicians and 14 specialties. Physicians received Microsoft DAX Copilot, Nabla or usual care. The intervention was a workflow: record the consultation, generate the draft, then have the physician review and edit it. A note that appears quickly can still require work before it is ready to enter the clinical record.[2]

One product reduced the primary time measure

Nabla physicians’ estimated writing time fell from 4 minutes 30 seconds to 3 minutes 49 seconds per note. The control group also improved, from 4 minutes 22 seconds to 4 minutes 4 seconds. The improvement occurring in both groups is why the comparison with usual care matters: the full 41-second fall cannot simply be credited to the AI.[2]

The trial also found favourable changes in some measures of work experience, but these were secondary outcomes. Physicians reported occasional clinically important inaccuracies, including omissions and errors identifying the person being described. The workflow still included clinical review, and the primary outcome was time writing notes rather than total labour around each visit.[1]

What adoption looked like across five health systems

A separate JAMA study published in April 2026 examined 8,581 clinicians, including 1,809 given access to scribes, at five US academic health systems. It used differences in changes over time between adopters and nonadopters; adoption was largely voluntary. That wider observation supports a view of ordinary deployment, while retaining the possibility that the groups differed in ways the statistical adjustments did not capture.[3]

These measures answer different questions. Saving time in one part of the record does not guarantee the same reduction in the whole record, and neither guarantees an earlier finish. Scheduling, inbox work and what clinicians do with recovered time can all influence the final working day. The figures are most useful when kept attached to the activity and denominator actually measured.

A good note has several jobs

Interviews with 24 clinicians at the University of Iowa add an explanation for the uneven experience. Some described better eye contact and more attention during appointments, while also finding drafts too detailed in one section and too thin in another. Clinicians used notes to convey their reasoning, communicate with colleagues and handle sensitive information; producing a transcript did not finish those tasks.[4]

The interviews also found less perceived benefit in settings where procedure or inpatient notes already followed tightly standardised formats. For a conversational outpatient visit, capturing the exchange can remove substantial drafting work. For a largely templated task, the remaining work may lie elsewhere. These are qualitative accounts of workflow, not another measured percentage saving.[4]

That is the opportunity the evidence makes concrete: an AI scribe can change both the effort of producing the record and the attention available during the visit. The best result will come from a draft that already fits the clinical purpose closely enough that reviewing it takes less work than creating it. Faster generation is one step in that process; the finished, useful note is the product.

Report an error or suggest a correction