Living EvidenceAtlas: Ambient AI Scribes

1. Clinician Burnout and Well-Being

This is the most thoroughly studied area. Three studies directly measured burnout-related outcomes.

The RCT (Afshar et al., NEJM AI, 2025) [1] is the strongest evidence available. In a 24-week stepped-wedge individually randomized pragmatic trial at an academic health system in Wisconsin and Illinois, 66 healthcare practitioners were randomized to begin using an ambient AI scribe (Abridge) at staggered intervals. The study used the Stanford Professional Fulfillment Index (PFI), a validated burnout instrument.

Work exhaustion and interpersonal disengagement decreased by 0.44 points on a 5-point scale (P<0.001). At baseline, 56% of practitioners scored above the high burnout threshold; by the end of the study, this dropped to 35%, a 21-percentage-point reduction. The number needed to treat was 1.68, meaning that for roughly every two practitioners given access to the scribe, one experienced a clinically meaningful reduction in burnout.

Critically, professional fulfillment (the positive dimension of well-being) increased by only 0.14 points and did not reach statistical significance after correction for multiple comparisons (P=0.04 vs. the required threshold of P<0.025) [1]. The authors note that professional fulfillment may be influenced by broader organizational factors like staffing levels, professional autonomy, and team climate that scribes do not address.

Multiple secondary measures also improved significantly: task load decreased, impact on personal relationships improved, perceived clinical efficiency increased, meaningfulness of work rose, and trust in AI grew. One secondary measure did not improve: perceived control over work schedule (P=0.382) [1].

This study had notable methodological strengths: no vendor funding (funded by the university hospital and NIH), all authors affiliated with the university rather than the vendor, preregistered on ClinicalTrials.gov, and analytic code publicly available.

The pre-post study (Olson et al., JAMA Network Open, 2025) [2] examined 263 clinicians across six academic and community-based U.S. health systems who used an ambient AI scribe (Abridge) for 30 days. Using a validated single-item burnout measure, burnout prevalence decreased from 52% to 39% among the 186 participants included in the burnout analysis, a 13-percentage-point drop, with 74% lower odds of burnout after adjustment.

Additional findings included a large reduction in cognitive task load (2.64 points on a 10-point scale), improved focused attention on patients (2.05 points on a 10-point scale), and a self-reported decrease of approximately 11 minutes per workday in after-hours documentation time [2].

This study had broader reach (six health systems, 263 participants) but a weaker design (pre-post with no control group, non-anonymous survey) and notable vendor involvement (three of eleven authors had affiliations with the ambient scribe company, including the company's chief clinical officer during the study period). The analysis was performed independently by the Yale-affiliated authors, and the study was preregistered. These design limitations and conflicts of interest don't invalidate the findings, but they do mean this study should be interpreted alongside, not as a substitute for, the independent RCT evidence.

Notably, the pre-post study also found that ambient scribe use was not associated with any significant increase in the number of additional patients clinicians could add to their schedule (P=0.58) [2], the only nonsignificant secondary outcome, and one that was not highlighted in earlier syntheses of this data.

The companion RCT (Lukac et al., NEJM AI, 2025) [3] tested two different products (Microsoft DAX Copilot and Nabla) in a three-arm parallel trial at an academic health system in California (238 outpatient physicians, 14 specialties). This trial was powered for documentation efficiency, not burnout, but its secondary burnout outcomes tell an interesting story. The DAX arm showed a work exhaustion reduction whose confidence interval excluded zero (PFI-WE −0.32, CI −0.55 to −0.08), and both products showed Mini-Z composite burnout improvements whose CIs excluded zero. However, as secondary outcomes from a trial not powered for these measures, these are suggestive signals that need confirmation, not proof of effect.

Utilization was approximately 30% for both products, less than half of what the Afshar RCT achieved (71%) [1], which may partly explain the weaker primary outcome results. The gap between studies highlights that implementation approach matters as much as the technology itself.

2. Documentation Time Savings

Time savings are real but more nuanced than headlines suggest.

Objective EHR data from the RCT [1] shows time on notes decreased by 0.36 hours per day (approximately 22 minutes) as measured through EHR audit logs, providing objective data rather than clinician self-report. This finding was robust across sensitivity analyses. Work done outside of normal hours also appeared to decrease (0.50 hours/day), but this finding was sensitive to statistical outliers and became nonsignificant when the top 3% of extreme values were excluded.

The companion RCT [3] measured time-in-note as a between-group percent change versus a usual-care control arm. Nabla showed a significant 9.5% reduction (P=0.02); DAX Copilot showed a nonsignificant 1.7% reduction (P=0.66). The raw within-arm time declines were 41 seconds (Nabla) and 18 seconds (DAX), but the control group also declined by 23 seconds over the same period, meaning DAX's incremental effect over doing nothing was negligible. Utilization was approximately 30% for both products, compared to 71% in the Afshar trial, which may partly explain the weaker results.

Self-reported data from the pre-post study [2] showed after-hours documentation time decreasing by approximately 11 minutes per workday (10.8 min/day), along with additional reductions in total documentation time and EHR queue time. These self-reported figures are directionally consistent with the Afshar RCT's objective measurement of 22 minutes per day, though the pre-post study measured a narrower slice of documentation work (after-hours only). The convergence across studies and measurement methods strengthens the overall time savings signal for this product.

An important caveat from both the RCT and the pre-post study: same-day encounter closure rates and follow-up timeliness did not significantly improve [1]. Documenting the clinical encounter is only a fraction of total EHR time, and both studies note that further productivity gains may depend on expanding AI capabilities beyond note generation to include downstream tasks like inbox management, order entry, and patient communication [1][2][6].

3. Note Quality and Documentation Accuracy

The RCT [1] evaluated nearly 8,000 AI-generated ambient scribe notes using the PDSQI-9, a validated note quality instrument scored by an LLM-as-judge system that was itself validated against physician ratings. Quality was high across most domains on a 5-point scale: accuracy (including falsification and fabrication) scored 4.44, thoroughness scored 4.57, usefulness scored 4.83, and linguistic quality dimensions (organization, comprehensibility, succinctness) ranged from 4.63 to 4.99.

The lowest-scoring domain was abstraction and synthesis at 3.97, reflecting the difficulty of integrating information from multiple sources and drawing clinical inferences. This makes intuitive sense: ambient scribes are fundamentally transcription-to-structure tools, and higher-order clinical reasoning remains harder for AI.

Stigmatizing language was detected in fewer than 1% of notes (12 out of 7,966). No quality drift was detected across 15 sequential monthly monitoring windows, and a planned software update during the study had no significant impact on note quality or utilization [1].

A separate study on AI-generated discharge summaries [4] compared LLM-generated (GPT-4) and physician-written narratives across 100 inpatient encounters. Blinded reviewers found comparable overall quality (3.7 vs. 3.8). AI outputs were rated more concise and coherent, but less comprehensive, and contained more errors per summary (2.9 vs. 1.8). Overall harmfulness scores remained low for both. This supports treating AI-generated clinical text as a draft requiring clinician review rather than a finished product.

4. Billing and Coding Compliance

The RCT [1] included a billing compliance analysis that compared ICD-10 coding alignment in over 6,000 notes (roughly half AI-generated, half human-authored), adjudicated by professional staff coders using a standardized rubric. AI-generated notes scored significantly higher: 6.87 vs. 5.94 on a 0–10 compliance scale (P<0.001).

This likely reflects the fact that ambient scribes generate more complete and structured documentation of clinical encounters, which in turn supports more accurate coding. This is a quantifiable financial benefit: improved coding accuracy can reduce claim denials and capture revenue that would otherwise be left on the table.

No study has yet measured the direct revenue impact of this improved coding, and the finding comes from a single health system. But it represents the first RCT-level evidence of a billing compliance advantage for AI-generated clinical notes.

5. Utilization and Implementation

Utilization (how often clinicians actually use the scribe once it's available) varied dramatically across studies, and this may be the most underappreciated finding in the evidence base.

The first RCT [1] achieved 71% utilization (note-weighted mean), with over 27,000 of 71,000 notes generated using AI. Seven of 66 practitioners showed low fidelity: three never initiated and four used the tool inconsistently. The study used a voluntary enrollment model where practitioners had to be willing to adopt the technology.

The companion RCT [3] achieved approximately 30% utilization, less than half of the first study, using different products. The lower utilization may partly explain the weaker time savings and burnout findings.

The pre-post study [2] reported that 95.9% of participants generated at least five notes using the scribe, but did not report ongoing utilization rates. Prior documentation methods among participants were diverse: 83% used manual typing, 85% used templates, 47% used dictation, and 16% had used human scribes.

The gap between 71% and 30% utilization is substantial and suggests that implementation approach, product experience, and organizational support matter as much as the underlying technology. Satisfaction surveys do not predict actual utilization; a health system where clinicians say they like the tool may still see low adoption in practice.

6. Patient Experience and Outcomes

This is the thinnest area of the evidence base. No study has directly measured patient outcomes or patient-reported experience with ambient AI scribes.

The pre-post study [2] found a small but statistically significant improvement in clinician-reported patient comprehension of care plans (0.44 points on a 10-point scale, P=0.005) and a modest improvement in urgent patient access (0.51 points, P=0.02). Both are clinician perceptions, not patient-reported measures.

The RCT [1] explicitly did not collect patient-reported outcomes, which the authors acknowledge as a limitation.

Patient consent rates were very high in the RCT (99.92%), suggesting patients are generally comfortable with ambient recording [1]. But comfort with the technology does not tell us whether it improves or changes the care they receive.

7. Safety Considerations

Error rates in AI-generated clinical text are slightly higher than in physician-written text. The discharge summary study [4] found 2.9 errors per AI summary vs. 1.8 per physician summary, though overall harmfulness was low. The RCT's note quality data [1] showed high accuracy (4.44/5) but the abstraction/synthesis weakness (3.97/5) suggests that errors of omission or oversimplification may be the primary risk.

The companion RCT reported the first adverse safety event in the ambient scribe RCT literature [3]: a grade-1 (mild) event in which extensive patient counseling was omitted from the AI-generated assessment and patient instructions. Clinicians in that study also rated inaccuracies as "occasional" (approximately 2.7–2.8 on a 5-point frequency scale). This is a concrete example of the omission risk: the scribe didn't fabricate information, but it failed to capture something clinically important that happened during the visit.

Automated fact-checking is emerging but has important limitations. The VeriFact system [5] demonstrated that individual claims in clinical notes can be verified against the patient's medical record using a retrieval-augmented LLM pipeline. Its best configuration achieved 93.2% agreement with clinician consensus on AI-generated claims, exceeding the 88.5% agreement clinicians achieved with each other. The system labels each claim as supported, not supported, or not addressed by the medical record. All models used are open-source and locally hosted, meaning no patient data leaves the institution.

However, the 93% headline requires context. Most claims in clinical notes are correct ("Supported"), so the high agreement rate is partly driven by class imbalance. The Matthews correlation coefficient (a metric less affected by class imbalance) was 0.31, indicating only moderate discriminative ability. Agreement on "Not Supported" and "Not Addressed" categories was much lower (31–34%), even among clinicians evaluating the same claims. In other words, the system is good at confirming that correct facts are correct, but substantially less reliable at catching errors.

A critical limitation for ambient scribe use: VeriFact can only verify whether what IS written in a note is supported by the medical record. It cannot detect errors of omission (content that should be in the note but isn't). For ambient scribes, where a key risk is the scribe failing to capture something the clinician said or did, this is a significant blind spot.

VeriFact was also tested on discharge summaries rather than ambient scribe encounter notes, a different document type generated in a different context. The authors suggest that fact-checking clinical notes could become analogous to spell-checking, but the current evidence supports this as a complementary safety layer, not a complete one.

The prediction that over 90% of clinical note text will be AI-generated in 2026 [6] raises a governance question: if notes are overwhelmingly AI-authored, do they still faithfully represent what the clinician actually observed, thought, and decided? This is an unresolved medicolegal and clinical quality concern that health systems should address proactively through attestation policies and documentation review requirements.