Does AI Documentation Reduce Clinician Burnout? What the Research Says

The one randomized controlled trial: mixed results by tool

The only published RCT of ambient AI scribes randomized 238 physicians across 14 specialties to one of two commercial AI scribes (Microsoft DAX Copilot or Nabla) or usual care. Nabla users saw a statistically significant 9.5% reduction in time-in-note versus control; DAX users did not see a significant reduction versus control. Both groups showed modest improvements in secondary burnout-related measures (Mini-Z score, physician task load, work-exhaustion score), which the authors explicitly describe as needing confirmation in larger, multicenter trials. Both tools produced "occasional" clinically significant inaccuracies by physician report.

Source: Lukac et al., NEJM AI, 2025

Randomized controlled trial - the strongest study design in this literature, but still a single trial.

An uncontrolled pilot showed large effects

A 3-month quality-improvement pilot of an ambient AI scribe (DAX Copilot) at Stanford Health Care with 48 physicians found large, statistically significant reductions in task load and burnout scores on paired pre/post surveys, plus improved usability ratings.

Source: Shah et al., JAMIA, 2025

Before/after design with no control group - a real limitation, since burnout and task-load scores can shift for reasons unrelated to the tool (seasonality, concurrent workplace changes, novelty effect).

A small pilot found no significant change in burnout measures

A 4-week pilot with 10 primary care physicians at Cleveland Clinic found the only statistically significant pre/post change was a reduction in characters typed. Mini-Z burnout scores and NASA Task Load Index scores did not change significantly, though physicians reported optimism that the tool could eventually reduce workload once integration barriers were addressed.

Source: Alpert et al., Digital Health, 2025

Small sample (n=10) and short follow-up (4 weeks) - but an honest null finding worth including rather than omitting.

Efficiency gains didn't extend to every burden measure

A retrospective cohort study comparing AI scribe users to a propensity-matched control group found significantly less time in the EHR and in notes for scribe users, but no significant difference in after-hours "pajama time," appointment length, or appointment volume between groups.

Source: Pearlman et al., JAMA Network Open, 2025

A reminder that documentation-time improvements don't automatically translate into every downstream burden measure.

What the systematic reviews conclude, in aggregate

A review of 11 AI-scribe implementation studies (most published in 2024) found 9 of 10 reporting improved efficiency and 7 of 10 reporting positive effects on clinician wellness or burnout - but flagged variable adoption, performance limitations, and evaluation gaps. A separate review of 8 studies on AI and EHR-related burnout reported similarly favorable but preliminary findings, explicitly citing the absence of control groups, small sample sizes, and short follow-up periods as limits on how far the findings can generalize.

Source: Hassan et al., Applied Clinical Informatics, 2025; Sarraf & Ghasempour, Frontiers in Public Health, 2025

Does AI documentation reduce burnout?

The honest answer is: it can help with one piece of it. Documentation time and after-hours EHR charting are well-established drivers of burnout, and ambient AI scribes have been shown in a 2025 randomized controlled trial to reduce time spent writing notes and produce modest improvements in task-load and work-exhaustion scores (Lukac et al., NEJM AI, 2025). But burnout is multifactorial - documentation is one driver among several (workload, staffing, autonomy, organizational culture) - and even the strongest AI-scribe studies to date describe their burnout findings as needing confirmation in larger trials, not as settled proof.

What does the research say about ambient AI scribes specifically?

It's a mixed but generally positive early picture. A 2025 three-arm randomized trial of 238 physicians found one AI scribe (Nabla) significantly reduced time-in-note versus usual care, while a second (DAX Copilot) did not reach significance versus control; both showed modest improvements in secondary burnout measures (Lukac et al., NEJM AI, 2025). An uncontrolled pilot at Stanford Health Care (n=48) found large reductions in task load and burnout scores (Shah et al., JAMIA, 2025). A Cleveland Clinic pilot with 10 physicians found the only statistically significant change was less typing - not burnout or workload scores (Alpert et al., Digital Health, 2025). Systematic reviews summarizing this literature describe it as promising but limited by small samples, short follow-up, and a frequent absence of control groups (Hassan et al., Applied Clinical Informatics, 2025; Sarraf & Ghasempour, Frontiers in Public Health, 2025).

Is documentation burden really a cause of burnout?

Yes - this is the more established side of the research. A large KLAS Arch Collaborative analysis of physician survey data found physicians charting 6 or more hours a week after hours were roughly half as likely to report low burnout as those charting 5 hours or less (Eschenroeder et al., JAMIA, 2021). A 2026 study of over 9,600 family medicine residents found nearly a third reported 3+ hours of nightly after-hours charting ("pajama time"), which was associated with significantly higher odds of burnout (Barr et al., Academic Medicine, 2026). Multiple systematic reviews confirm documentation and clerical burden as a recurring, cross-study contributor to clinician burnout (Murad et al., Journal of General Internal Medicine, 2024; Budd, Journal of Primary Care & Community Health, 2023).

How strong is the evidence for AI scribes specifically?

Weaker than the evidence for documentation burden itself, and study designs vary widely. Only one randomized controlled trial of ambient AI scribes has been published to date (Lukac et al., NEJM AI, 2025); most other studies are single-site pilots, before/after comparisons without a control group, or retrospective cohorts. Several report null or mixed findings on burnout-specific measures even when documentation time improved (Alpert et al., Digital Health, 2025; Pearlman et al., JAMA Network Open, 2025). Accuracy is also a documented concern - one study found an early large language model generated an average of 23.6 errors per clinical note when producing structured notes from transcripts (Kernberg et al., JMIR, 2024), and a systematic review of AI transcription accuracy found error rates ranging from under 9% in controlled dictation to over 50% in multi-speaker conversation (Ng et al., BMC Medical Informatics and Decision Making, 2025). None of this means AI scribes don't work - it means the evidence base is real but still early.

Do these studies apply to nurse practitioners, or just physicians?

Mostly physicians. Nearly every ambient-AI-scribe study identified in this review - the NEJM AI randomized trial, the Stanford and Cleveland Clinic pilots, the JAMA Network Open cohort study - enrolled attending physicians or physician residents, not nurse practitioners specifically. The closest NP-relevant evidence is a 2026 systematic review of AI-enabled workflows across nursing staff, including nurse practitioners, which found workload reductions in half of the 20 included studies and improved emotional well-being or job satisfaction in about two-thirds - with outcomes depending heavily on training, trust, and workflow fit (Sarıköse et al., International Journal of Nursing Studies, 2026). That review covers AI tools broadly (documentation, clinical decision support, telehealth triage), not ambient scribes alone, so it's directionally relevant rather than a direct answer for NPs.

Are AI-generated notes accurate enough to trust?

Accuracy varies by tool and by study, and every study in this space recommends clinician review rather than blind trust. One evaluation of an early large language model generating SOAP notes from encounter transcripts found an average of 23.6 errors per note - mostly omissions - and concluded the output did not meet the standard required for clinical use without review (Kernberg et al., JMIR, 2024). A broader systematic review of AI speech-to-text accuracy in clinical settings found word error rates ranging from under 9% in controlled dictation to over 50% in multi-speaker conversation (Ng et al., BMC Medical Informatics and Decision Making, 2025). The randomized trial of two commercial AI scribes reported "occasional" clinically significant inaccuracies on both platforms tested, describing ongoing clinician vigilance as necessary (Lukac et al., NEJM AI, 2025).

What is "pajama time" and why does it matter?

"Pajama time" is the informal term researchers use for time clinicians spend working in the EHR after clinic hours - at home, in the evening, on days off. It's one of the most consistently measured predictors of burnout in this literature: a 2026 study of family medicine residents found those with 3+ hours of nightly pajama time had significantly higher odds of burnout, lower professional satisfaction, and lower in-training exam scores (Barr et al., Academic Medicine, 2026), and a large KLAS Arch Collaborative analysis found a similar association in practicing physicians (Eschenroeder et al., JAMIA, 2021). Notably, one retrospective cohort study of an AI scribe found it reduced total EHR time and note time, but did not produce a statistically significant reduction in pajama time specifically (Pearlman et al., JAMA Network Open, 2025) - a reminder that not every documentation-time improvement translates evenly across every burden measure.