AI Medical Scribes: What the Randomized Trial Shows

In short: A randomized trial at UCLA tested two AI "ambient scribe" tools that listen to a doctor's visit and draft the clinical note. Published in NEJM AI in late 2025, it found one tool cut the time doctors spent writing each note by 9.5%, about 41 seconds, while the other showed no statistically significant change. Burnout scores improved modestly, but the notes sometimes contained errors that doctors had to catch.

Key Takeaways

• The trial, by UCLA researchers and published in NEJM AI, randomized 238 physicians across 14 specialties to Nabla, Microsoft DAX or usual documentation, between November 2024 and January 2025.
• Nabla users cut time per note by 41 seconds (9.5%), from 4:30 to 3:49. The DAX difference was not statistically significant compared with the control group, according to UCLA Health.
• Both tools showed roughly a 7% improvement in validated burnout scores compared with controls, per UCLA Health.
• Notes occasionally contained clinically significant inaccuracies, such as omissions and pronoun errors, and one mild patient safety event occurred.
• It was one academic medical center over about two months, so the authors call for longer, multi-institution studies.

The Evidence

The trial. UCLA Health announced on November 26, 2025 a randomized clinical trial published in NEJM AI. It compared two AI scribe tools, Nabla and Microsoft DAX, against standard documentation practice. It enrolled 238 physicians across 14 specialties and ran from November 2024 to January 2025. Funding listed by UCLA includes its Department of Medicine, its Faculty Practice Group, the National Institutes of Health's National Institute on Aging and the UCLA Healthcare Value Analytics Consortium.

The numbers. The primary outcome was documentation efficiency. Per UCLA, Nabla users saved 41 seconds per note, a 9.5% improvement, from 4:30 to 3:49. DAX showed no statistically significant reduction compared with the control group. UCLA reports roughly 7% better burnout scores for both tools than for controls.

Why it was studied. UCLA's Dr. Paul Lukac is quoted saying that "doctors often spend two hours on paperwork for every hour of patient care."

A caution on sources. Nabla's press release about the trial is vendor marketing and does not discuss limitations. We rely on UCLA's release for the figures. Different outlets give different encounter counts for the trial, so we do not quote one.

What the Technology Does

1. Capture. With the patient's consent, the tool records the conversation during a visit.
2. Draft. An AI model turns the conversation into a structured clinical note.
3. Review. The physician reads, edits and signs the note. Accountability stays with the doctor.

In the trial, physicians described the tools as easy to learn and said they helped them engage more directly with patients, according to the NEJM AI summary reported by Nabla, a vendor, so treat that as self-reported.

What Changed

Many hospitals adopted AI scribes quickly on the strength of vendor claims and early pilots. This is a randomized comparison with a control group in routine practice, which is a firmer way to ask whether the tools save time. The answer was nuanced: a real but modest saving for one tool and no clear saving for the other.

Who Benefits

• Physicians. Less time typing notes and a modest improvement on burnout measures.
• Patients. Doctors who report engaging more directly in the visit. That is the physicians' perception and not a measured patient outcome in our sources.
• Health systems. A possible route to easing administrative load. The sources give no cost data, so we state none.

Limitations and Open Questions

• Accuracy. UCLA reports AI-generated notes occasionally included clinically significant inaccuracies, such as information omissions and pronoun errors, and one mild patient safety event occurred. Every note needs real human review.
• Modest and uneven effect. 41 seconds per note is meaningful at volume but not transformative, and results differed between tools.
• Short and local. One academic medical center and about two months. Findings may not generalize, and UCLA says longer, multi-institution studies are needed.
• Privacy and consent. Recording consultations raises data-protection questions. See our guide to protecting patient data.

This article is general information and not medical advice.

What Happens Next

Expect more trials across specialties and settings, independent audits of note accuracy, and studies of whether time savings last. Buyers should ask for results from randomized or independent evaluations, error rates, and how the tool fits a clinic's own workflow.

The Short Version

AI scribes are no longer just a promise: a randomized trial found a modest documentation-time saving for one tool, no significant saving for the other, and small burnout improvements, alongside occasional note errors. They are best treated as a draft-writing aid that needs careful clinician review. For related evidence on clinical AI, see our posts on AI sepsis early warning and AI in UK healthcare.

Related Insights

• AI Sepsis Early Warning: What the Evidence Shows
• AI Mammography Screening: What the Swedish Trial Shows
• AI in UK Healthcare: What to Check Before You Buy
• EU AI Act: What Applies Now and What Was Delayed

Sources

• UCLA Health, UCLA study finds AI scribes may reduce documentation time and improve physician well-being, November 26, 2025
• Nabla, NEJM AI trial reports efficiency gains for physicians using Nabla's ambient AI assistant, December 2025 (vendor press release)
• Lukac et al., Ambient AI Scribes in Clinical Practice: A Randomized Trial, NEJM AI, 2025 (figures as reported by UCLA Health and the sources above)