Authors: Paul J. Lukac (1Department of Pediatrics, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA; 2UCLA Health Information Technology, UCLA Health, University of California, Los Angeles, Los Angeles, CA, USA), William Turner (3Division of General Internal Medicine and Health Services Research, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA), Sitaram Vangala (3Division of General Internal Medicine and Health Services Research, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA), Aaron T. Chin (2UCLA Health Information Technology, UCLA Health, University of California, Los Angeles, Los Angeles, CA, USA; 4Department of Pediatrics, Division of Immunology, Allergy and Rheumatology, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA), Joshua Khalili (5Department of Medicine, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA), Ya-Chen Tina Shih (5Department of Medicine, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA; 6Department of Radiation Oncology, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA; 7Department of Health Policy and Management, Fielding School of Public Health, University of California, Los Angeles, Los Angeles, CA, USA; 8Program in Cancer Health Economics Research, Jonsson Comprehensive Cancer Center, University of California, Los Angeles, Los Angeles, CA, USA), Catherine Sarkisian (3Division of General Internal Medicine and Health Services Research, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA; 9Veterans Administration Greater Los Angeles Geriatrics Research Education and Clinical Center, Los Angeles, CA, USA), Eric M. Cheng (2UCLA Health Information Technology, UCLA Health, University of California, Los Angeles, Los Angeles, CA, USA; 10Department of Neurology, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA), John N. Mafi (3Division of General Internal Medicine and Health Services Research, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA, USA; 11RAND Corporation, Santa Monica, CA, USA)
Categories: Article
Source: NEJM AI
Doi: 10.1056/aioa2501000
Authors: Paul J. Lukac, William Turner, Sitaram Vangala, Aaron T. Chin, Joshua Khalili, Ya-Chen Tina Shih, Catherine Sarkisian, Eric M. Cheng, John N. Mafi
Ambient artificial intelligence (AI) scribes record patient encounters and rapidly generate visit notes, representing a promising solution to documentation burden and physician burnout. However, the scribes’ impacts have not been examined in randomized clinical trials.
In this parallel three-group pragmatic randomized clinical trial, 238 outpatient physicians, representing 14 specialties, were assigned 1:1 via covariate-constrained randomization (balancing on time-in-note, baseline burnout score, and clinic days per week) to either one of two AI scribe applications — Microsoft Dragon Ambient eXperience (DAX) Copilot or Nabla — or a usual-care control group from November 4, 2024, to January 3, 2025. The primary outcome was the change from baseline log writing time-in-note. Secondary end points measured by surveys included the Mini-Z 2.0, a four-item physician task load (PTL), and Professional Fulfillment Index — Work Exhaustion (PFI-WE) scores to evaluate aspects of burnout; work environment; stress; and targeted questions addressing safety, accuracy, and usability.
DAX was used in 33.5% of 24,696 visits; Nabla was used in 29.5% of 23,653 visits. 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, whereas DAX users exhibited no significant change versus the control group (−1.7%; 95% CI, −9.4% to +5.9%; P=0.66). 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]) and 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]), and 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]) scores suggest improvement for users of either scribe versus the control. One grade 1 (mild) adverse event was reported, while clinically significant inaccuracies were noted “occasionally” on five-point Likert questions (DAX 2.7 [95% CI, 2.4 to 3.0]; Nabla 2.8 [95% CI, 2.6 to 3.0]).
Nabla reduced time-in-note versus the control. Both DAX and Nabla resulted in potential improvements in burnout, task load, and work exhaustion, but these secondary end point findings need confirmation in larger, multicenter trials. Clinicians reported that performance was similar across the two distinct platforms, and occasional inaccuracies observed in either scribe require ongoing vigilance. (Funded by the University of California, Los Angeles, Department of Medicine and others; ClinicalTrials.gov number, NCT06792890.)
Burnout afflicts nearly half of physicians in the United States, approaching endemic levels and fueling a workforce exodus that amplifies an already critical physician shortage.^1–5^ Pernicious symptoms of exhaustion, depersonalization, and a diminished sense of personal accomplishment have far-reaching consequences. The effects manifest across the health care ecosystem, jeopardizing care access, doubling the risk of patient-safety events, incurring billions of dollars in undue costs, and endangering physician well-being.^6–9^ While causes of burnout are complex and multifactorial, electronic health records (EHRs) and related documentation burden are frequently cited culprits.^10–14^ This excessive EHR charting forces physicians to spend as much as 2 hours documenting for every hour of direct care.^15,16^ A systematic review detailing EHR characteristics associated with burnout identified “insufficient time for documentation” as a top contributor.^17^
To address this problem, human scribes, both in person and virtual, have been deployed with some success, though these options present challenges around cost and accessibility, particularly among less well-resourced specialties and settings.^18–21^ Digital scribes, which incorporate artificial intelligence (AI) to generate drafts that still require human editing, have shown variable impact, with little to no efficiency gain or burnout improvement, but generally positive user reception.^22–24^ Fully autonomous ambient AI scribes, which leverage large language models (LLMs), have been received with great enthusiasm by both industry and providers and are less expensive than human scribes, increasing scalability and the potential for widespread adoption.^25–28^ At present, over 50 ambient AI companies offer serviceable products, though few studies have rigorously evaluated their impact.^2,29–31^ Two small, nonrandomized pilot studies reported that Dragon Ambient eXperience (DAX) Copilot usage decreases documentation time, EHR time, physician task load (PTL), and aspects of burnout.^32–34^ However, a larger, nonrandomized study of DAX showed no significant changes in financial and EHR-use metrics, while, discordantly, survey respondents in the intervention group subjectively reported decreased EHR time.^35,36^ Separately, in a large pilot study of Nabla, researchers reported that Nabla usage resulted in small reductions in charting time and favorable physician reception.^37^
Despite the many questions surrounding the use of AI in health care as well as the novel challenges of generative AI (genAI), there is a dearth of randomized clinical trials evaluating its effectiveness and safety in real-world settings.^29,38–40^ This trial, evaluating two AI scribes, aimed to investigate their effects on documentation time and physician psychometrics, as well as their usability, accuracy, and safety. In addition, with many AI options available to health care providers, this study sought to compare two leading vendors with a control group to inform future investment in this nascent technology.
Enrolled physicians (N=238) were randomly assigned 1:1 to either of two intervention groups, one provisioned with Microsoft DAX Copilot (N=79) and one provisioned with Nabla (N=79), or to a contemporaneous control group (N=80). To achieve cohort balance, the study statistician performed covariate-constrained random assignment on physician baseline time-in-note (an Epic Systems, Inc. [Verona, WI] Signal metric), a single-item burnout score,^41^ and the number of self-reported clinic days per week. Further details on covariate-constrained randomization can be found in the Supplementary Appendix. A mandatory prestudy survey was sent in September 2024, and a poststudy survey was sent on January 4, 2025. The trial was implemented from November 4, 2024, to January 3, 2025.
The study protocol was developed in accordance with Standard Protocol Items: Recommendations for Interventional Trials–Artificial Intelligence (SPIRIT-AI) and registered on ClinicalTrials.gov (NCT06792890).^42^ The study results reported in this article follow the Consolidated Standards of Reporting Trials–Artificial Intelligence (CONSORT-AI).^43^ The University of California, Los Angeles (UCLA) Institutional Review Board (IRB-24–5425) determined that the study did not constitute human subjects research. Institutional urgency to rapidly test the tools during a brief contractual trial period precluded the study’s timely preregistration. The finalized trial registration was submitted on December 9, 2024, and published on January 27, 2025; the unpublished protocol manuscript can be found in the Supplementary Appendix.
Physicians were recruited via department-wide emails and nominations from department leaders. Physicians were required to hold at least one half-day of clinic per week, and those in intervention groups who used human scribes were obligated to forego this assistance during the study. Usage was restricted to English-only visits due to a lack of internal validation of translation capabilities. Figure 1 summarizes the trial recruitment.
Microsoft DAX Copilot version 2.0 and Nabla version 1.5 were used. Both products were integrated into the EHR, allowing AI scribe–produced text to populate Epic’s native notes. Participants received 1 hour of virtual training from physician informaticists and vendor representatives 1 to 2 weeks before the study’s onset. Training was standardized for length, vendor participation, and format of internally generated educational material. Two weeks after study onset, vendors delivered a second presentation highlighting advanced features and customization. Support included an internal messaging channel and listserv, as well as vendor assistance.
In compliance with California law requiring two-party consent for audio recording, all physicians were instructed to obtain verbal consent from all relevant parties and document it in their notes.
A prestudy Qualtrics survey characterized physician demographics and clinical time data. Total usage reflects all activity during the study period (November 4, 2024–January 3, 2025) to capture overall engagement.
The prespecified primary outcome was documentation time, specifically the amount of time spent writing each note, routinely measured by an Epic Signal EHR utilization metric (time-in-note). On a provider level, this metric represents the total number of minutes a physician spends writing notes in a week, divided by the total number of notes. For each provider, these results were averaged by month. For a baseline comparison, time-in-note data were collected for the 6 months preceding the study’s onset, while study data were collected from November 4, 2024, to January 3, 2025. We allowed 1 month’s lead time for participants to reach proficiency and thus only compared the second intervention month (December 2, 2024, to January 3, 2025) with the baseline for all participants, aggregated within each study group. Notably, the time-in-note metric does not account for time spent editing the AI scribe–generated draft within the vendor platforms.
Prespecified secondary outcomes included the following validated survey instruments (full descriptions in Table S1): Mini-Z 2.0 to assess burnout, work environment, work pace, and EHR stress (10–50 scale, where lower scores=worse burnout, more stressful workplace).^2,44,45^ This 10-question survey contains a single-item burnout question that has been validated against a historical benchmark, the Maslach Burnout Inventory (MBI),^44,46^ whereas the remaining questions show convergent validity with MBI subscales.^47^ Mini-Z 2.0 is often reported as a single-item burnout measure along with the totals of its two five-question subscales — Supportive Work Environment and Work Pace and EMR Stress. However, the unidimensional composite score can be used to “portray an overall ‘joy score’,”^44^ representative of the single-item burnout question and the two subscales, and has been reported in the context of digital-scribe use.^22^Four-item PTL to assess cognitive load related to stress from EHR documentation (0–400 scale, where lower scores=less cognitive load).^48,49^ PTL has been demonstrated as a possible mediator between EHR usability and MBI-defined burnout, wherein better usability is associated with a lower PTL and lower odds of burnout.^48^ As a continuous effect, it has shown direct correlation with burnout, where each 40-point (10%) decrease in PTL correlated with 33% lower odds of burnout in one national study.^49^Professional Fulfillment Index — Work Exhaustion (PFI-WE; 0–4 scale, where lower scores=less exhaustion).^50,51^ PFI-WE has demonstrated convergent validity with the MBI — Emotional Exhaustion subscale.^50^
The poststudy survey addressed other prespecified topics via five-point Likert questions, such as usability, occurrence of inaccuracies and biases, and perceived risks to safety.
Two additional Epic Signal metrics — time in the EHR on unscheduled days, which reflects the average number of minutes a physician spends in the EHR on days with no scheduled patients, and time in EHR outside scheduled hours — were collected; baseline comparisons were conducted as with time-in-note.
Sample size was constrained by exclusion criteria and a contractual limitation stipulating a maximum of 100 concurrent users of each tool. A sample size of at least 79 physicians per condition (the size of our smallest condition) provides 80% power to detect effect sizes as small as 0.50 standard deviations, assuming a two-sample t-test and a two-sided 0.025 significance level (twofold Bonferroni correction for the comparison of each tool with the control condition) for the primary outcome.
Using prestudy data from a comparable time frame, the estimated log-scale standard deviation of change for the time-in-note metric was approximately 0.31. This design provided sufficient power to detect a 15.5% relative improvement in the primary outcome. Given a baseline-year geometric mean time-in-note of approximately 4 minutes 43 seconds (s), this corresponds to an absolute difference of roughly 44 seconds. This effect size falls within the range of values reported.^32,34,37^
The primary analyses — comparing DAX versus the control and Nabla versus the control — were conducted according to the intention-to-treat principle. Time-in-note was log-transformed to address nonnormal distribution and compared with providers’ prior 6-month baseline. A linear mixed-effects model, which included a study group effect, a period effect (second month vs. first month), and the interaction of these terms, was utilized to model change in log-time-in-note, accounting for repeated measurements over time. Linear contrasts were used to evaluate the effect of each tool versus the control in the second month (primary hypothesis). Survey-derived (Mini-Z 2.0, PTL, PFI-WE) quantitative outcomes were analyzed using unadjusted linear regression models of the absolute change from baseline to postintervention. Binary outcomes (e.g., dichotomized burnout) were analyzed using logistic regression models of postintervention responses, adjusting for baseline response.
We used a multiplicity-adjusted significance level of 0.025 for primary outcome hypothesis testing. For all other outcomes, we did not perform hypothesis testing and only report estimated differences and 95% confidence intervals. Results were collected on January 15, 2025, and all analyses were performed from January 16, 2025, to May 19, 2025, using R version 4.4.2 (https://www.r-project.org/).
Post hoc exploratory analyses, examining the comparisons of Nabla versus DAX and any scribe use versus the control, and the association of scribe usage rates with outcomes, can be found in the Supplementary Appendix (Table S2, Fig. S1).
In total, 238 physicians representing 14 specialties were enrolled (Table 1). Females were overrepresented (60.5%), and nearly half of the providers (46.6%) reported being 35–44 years old. At baseline, median time-in-note was 4 minutes 23 seconds for the DAX group, 4 minutes 47 seconds for the Nabla group, and 5 minutes 22 seconds for the control group.
DAX was used at 8271 of 24,696 (33.5%) of patient visits, whereas Nabla was used at 6981 of 23,653 (29.5%) visits. Control group encounters totaled 24,020. Approximately 15% of treatment-group physicians never used their assigned scribe. Feedback left by 28 physicians who indicated low or no use in poststudy survey responses or responded to the survey despite never having completed a scribe encounter can be viewed in Table S3.
Per intention-to-treat, time-in-note for each provider reflects the average of all notes written per month, regardless of scribe use. In our mixed model analysis comparing baseline log time with December data, time-in-note declined by an estimated 18 seconds (from 4 minutes 22 seconds to 4 minutes 4 seconds) in the control group, 23 seconds (from 4 minutes 29 seconds to 4 minutes 6 seconds) in the DAX group, and 41 seconds (from 4 minutes 30 seconds to 3 minutes 49 seconds) in the Nabla group. The reduction in the Nabla group was significantly larger than in the control group (−9.5% [95% confidence interval (CI), −17.2% to −1.8%; P=0.02), while the reduction in the DAX group did not significantly differ from the control group (−1.7%; 95% CI, −9.4% to +5.9%; P=0.66) (Table 2). Although not statistically significant in a month-to-month comparison, reductions in November (Nabla=−5.0% [95% CI, −12.7% to +2.7%]; DAX=−0.0% [95% CI, −7.6% to +7.7%]) were smaller than in December.
The poststudy survey was completed by 61 of 80 (76%) control-group physicians and 65 of 79 (82%) physicians in each intervention group. Composite Mini-Z scores increased, whereas PTL and PFI-WE scores decreased for users of either scribe versus the control (Table 2).
Inaccuracies were noted “occasionally” (DAX=2.7 [95% CI, 2.4 to 3.0]; Nabla=2.8 [95% CI, 2.6 to 3.0]) and bias “rarely” (DAX=1.6 [95% CI, 1.4 to 1.8]; Nabla=1.7 [95% CI, 1.5 to 1.9]) (Table 3). Respondents who noted “inaccuracies” and “bias” most frequently described omissions (N=12); structural concerns (e.g., undesired formatting, oversimplified language, or too much detail) (N=11); and pronoun resolution errors (N=8) (Table 4).
One adverse patient safety event was reported. Five physician co-authors (P.J.L., J.N.M., C.S., E.C., and A.C.) independently deemed the event, described as “extensive patient counseling was not included in the assessment/plan or the patient instructions,” to be grade 1 (mild).^52^
Of all usability questions, physicians rated both scribes least favorably in response to, “The tool generated notes at least as good as my own,” where the average rating was “neutral” (DAX=3.4 [95% CI, 3.1 to 3.7]; Nabla=2.9 [95% CI, 2.6 to 3.2]). Users rated both scribes highly on improving physicians’ ability to engage with patients (DAX=4.2 [95% CI, 4.0 to 4.4]; Nabla=3.8 [95% CI, 3.5 to 4.1]) and patients’ receptiveness to their use (each scribe=4.4 [95% CI, 4.2–4.6]) (Table 3). Figure S2 illustrates the answer distributions for each Likert-scale question.
There were no differences in time spent in the EHR during unscheduled days or time spent in the EHR outside scheduled work hours (Table 2).
In this randomized controlled trial of LLM-powered ambient AI scribes, we observed a modest reduction in the time spent on documentation among Nabla users compared with the control. Secondary end points indicate potential improvement in burnout, cognitive task load, and work exhaustion for DAX and Nabla users. Physicians found the scribes easy to use and felt that the technology allowed them to better engage with their patients, results which may reflect additional mechanisms for reduced burnout aside from documentation time-savings. For one of the most anticipated and rapidly adopted technological innovations in U.S. health care since the Health Information Technology for Economic and Clinical Health (HITECH) Act incentivized EHRs, these empirical findings are consistent with prior observational studies showing optimistic results^32–34,37^ and reveal similar performance and reception across the two platforms. More broadly, by embedding a randomized trial within routine practice, our study provides high-quality real-world evidence and can serve as a model for scientifically sound and ethically responsible AI integration.
A central challenge to AI-scribe adoption lies in justifying investment in this costly technology; efficiency gains, such as decreased time-in-note, represent one plausible rationale. In a recent report from Peterson Health Technology Institute, health-system leaders identified increasing the “number of patient encounters per period” and “accuracy or completeness of coding for billing purposes” as two financial metrics.^31^ The former, however, contradicts a potential benefit documented in this study and others — improved clinician well-being — and risks exacerbating physician disillusionment. Prior research suggests that burnout costs U.S. health systems 7600 per physician per year (2015 dollars),^7^ translating to 36.5 million annually if applied to the 4800 physicians at UCLA. Any improvement in physicians’ experience of work and reduction in stress from work could translate to savings from decreases in burnout-associated turnover and reduced clinical hours, and future research should determine the precise budgetary impact of this nascent technology.
On a broader scale, generative AI (genAI) is increasingly seen as a potential means to ease physicians’ documentation burden, a known contributor to EHR-mediated burnout.^17,53,54^ Two prominent early use cases of genAI, AI scribes and generative pretrained transformer (GPT)–produced draft responses to patient messages (Epic Systems), have shown signs of reducing PTL and burnout, despite limited or no efficiency gains.^32–34,51^ Building on these observations, our study provides preliminary support for AI scribes’ potential benefit to physician well-being, whether accompanied by efficiency gains (Nabla group) or not (DAX group). Holistically, the Mini-Z 2.0, PTL, and PFI-WE assess both drivers of burnout (e.g., workload, work conditions) and outcome states (e.g., burnout/exhaustion), and we noted concordant potential improvement across both the DAX and Nabla groups. Nevertheless, these secondary outcomes warrant cautious and nuanced interpretation and need to be confirmed in larger trials.
When our physicians reported on inaccuracies, they highlighted multiple manifestations, ranging from omissions to pronoun errors, magnifying the ambiguity and nuance in the clinical output of LLMs and suggesting that ongoing physician oversight will be necessary to ensure documentation fidelity and to prevent unintended downstream consequences, such as clinically significant omissions. Yet there remains a wide gap between our ability to deploy genAI at scale and our ability to validate genAI at scale. As noted by Bedi et al.,^55^ this is partly attributable to a lack of standardized tasks and dimensions of evaluation; for example, with AI scribes, should we focus on accuracy and factuality or omissions and comprehensiveness? Historically, quality assessment of EHR documentation has relied on manual human review with frameworks such as the Physician Documentation Quality Instrument — 9 (PDQI-9).^56^ Given that LLMs are nondeterministic and subject to versioning, which can influence output over time, relying on frequent user feedback for quality assurance is both unrealistic and may itself exacerbate task load.^57^
One potential solution would be for LLMs to augment or replace human evaluators (i.e., LLMs as quality control agents).^58^ As noted by Croxford et al., this too comes with pitfalls, such as the rapid evolution of LLMs outpacing our ability to validate the LLM evaluators, LLMs’ inherent reliance on and sensitivity to prompts, and the challenge in replicating a physician’s nuanced clinical judgement, which is necessary to determine if generated content is meaningful in the context of a patient’s clinical course.^59^ To be sure, the medical profession must embrace AI education and promote widespread AI literacy — essential steps toward safely and effectively integrating these tools into clinical practice.
Our study highlights several strengths and methodological insights. First, a randomized clinical trial represents the ideal method to control for selection biases, which is particularly important for optional workflows, such as whether or not to use an AI scribe. Second, unlike static interventions in pharmaceutical trials, these tools rapidly evolve, even over the course of a months-long study. Thus, a short, contemporaneous study period was advantageous in minimizing bias from product changes while also avoiding the pitfalls of offering each vendor sequentially, where one product could mature more than the other.
This study has limitations. First, it was conducted among 238 physicians at a single academic institution on English-only encounters from November 4, 2024 to January 3, 2025, and thus, our findings may not be broadly applicable to different practice settings or other times of the year. Second, our participants were mostly female, which could reduce generalizability given that females represent 38.1% of all physicians nationally.^60^ However, the inclusion of multiple specialties strengthens overall generalizability. Third, our trial was tightly scheduled due to limitations imposed by contract terms, which stipulated the length and number of users, and contributed to the postcommencement trial registration. This may have limited the degree of impact, given the time it takes to gain dexterity. The brevity may have also disincentivized physicians from investing time in learning the tool and customizing it, which could lead to an underestimation of positive results. Of note, while formal interaction testing was negative, we observed a month-to-month reduction trend in time-in-note, which may suggest a learning curve and potentially larger reductions over time. Fourth, physicians had the ability to edit the AI-generated text in both the scribe platforms and in Epic’s note-writing interface, and we learned midway through the trial that Epic’s Signal metrics do not account for platform time. Hence, reported time savings for time-in-note may represent an overestimation. This underrecognized limitation affects all AI scribe studies reporting Epic’s Signal metrics. Fifth, Nabla was not directly integrated within Epic’s mobile application (Haiku), whereas DAX was. Thus, Nabla users had to launch an encounter from a desktop to prompt a push notification on their phone, potentially affecting the physicians’ usability impressions. Sixth, there is potential for nonresponse bias with the poststudy survey, which could skew survey results positively or negatively.
We observed modest improvement in our primary outcome, time-in-note, for Nabla users. Though secondary psychometric end points suggest both DAX and Nabla may attenuate burnout and enhance the physician work experience, our preliminary findings underscore the need for future long-term studies to validate these trends across multiple institutions, establish a robust cost–benefit analysis, measure downstream effects on quality of care and patient safety, and precisely identify clinicians who will benefit most from this technology.