In a fictional case, a family doctor sees 22 people on a clinic day. At eight in the evening, after putting her children to bed, she opens the electronic health record at home and still has eleven notes to close. The following week, the service offers an artificial intelligence (AI) scribe: with the patient's consent, the phone records the conversation and, at the end of the visit, a draft note appears ready for review. In the first week she closes the notes at the office. In the second, she signs a note in which the patient who said "I don't drink" appears as someone who "now drinks".

This contrast sums up the debate about so-called ambient scribes. The first randomized trials suggest modest gains in documentation time and signs of less exhaustion, while studies on note quality show omissions and shifts in meaning that only careful reading by the doctor can catch (Lukac et al., 2025).

This text goes through what is known about the documentation burden, what trials and before-and-after studies show about burnout and time, what errors the generated note brings and how to check the draft before signing.

Key points

• In a randomized trial with 238 outpatient doctors, one of the two tools tested reduced note time by 9.5 percent compared with the control, and the other showed no significant difference (Lukac et al., 2025).
• In a pragmatic trial with 66 professionals, the scribe reduced the work exhaustion score, while the improvement in professional accomplishment did not reach the significance threshold set by the authors (Afshar et al., 2025).
• In a before-and-after study across six health systems, the proportion of clinicians with burnout fell from 51.9 to 38.8 percent in 30 days, with no control group (Olson et al., 2025).
• In an evaluation with simulated visit readings, 31 of 44 generated notes (70 percent) had at least one error, and omission was the most common type (Biro et al., 2025).
• The authors of the randomized trial conclude that occasional inaccuracies require continuous vigilance (Lukac et al., 2025).

• Why documentation weighs so much
• How the ambient scribe works
• What the randomized trials show
• Observational studies
• What errors the generated note brings
• Why and how to check the note
• Patient consent
• How NEXO helps train this reasoning
• Frequently asked questions
• Take-home messages

Why documentation weighs so much

The World Health Organization (WHO) describes burn-out, in ICD-11, as a syndrome resulting from chronic workplace stress that has not been successfully managed (WHO, 2019). The syndrome has three dimensions: feelings of energy depletion or exhaustion, increased mental distance from one's job or feelings of negativism and cynicism related to it, and reduced professional efficacy.

The WHO makes clear that burn-out is classified as an occupational phenomenon and not as a medical condition, in the chapter on factors influencing health status or contact with health services (WHO, 2019). The term refers only to the work context and should not be applied to other areas of life. This definition helps separate burnout from other forms of suffering, such as the one discussed in moral injury, the suffering that is not burnout, and from conditions that begin during medical school, covered in burnout in medical school.

How much time the record takes

One of the most cited studies on the topic used electronic health record event logs from 142 family doctors in a health system in southern Wisconsin over three years (Arndt et al., 2017). On an 11.4-hour workday, these doctors spent 355 minutes, about 5.9 hours, in the record: 269 minutes during clinic hours and 86 minutes outside them.

Bureaucratic and administrative tasks, such as documentation, orders, billing and choosing the ICD-10 code, took up 157 minutes, the equivalent of 44.2 percent of time in the record (Arndt et al., 2017). The inbox took another 85 minutes, or 23.7 percent (Arndt et al., 2017). This picture comes from a single American system, but it serves as a starting point for understanding the weight of the record.

The weight of alerts and notifications has its own effects, discussed in alert fatigue. Nighttime documentation also adds to accumulated tiredness, the topic of decision fatigue on shift.

How the ambient scribe works

The ambient scribe is a program that captures the audio of the consultation, usually through an app on the phone or computer, transcribes the conversation and uses a language model to produce a draft clinical note. The physician reviews, edits and signs it. In the studies discussed here, the tools were integrated into the electronic health record and used in outpatient consultations (Lukac et al., 2025; Afshar et al., 2025).

The idea is to give the physician's attention back to the patient. Instead of typing during the consultation, the physician talks, and the record arrives later. It is a proposal similar to letting the patient finish the opening statement, discussed in the patient's agenda, only applied to documentation.

What the generated note does not replace

Responsibility for the note remains with whoever signs it. In the Stanford study, the draft included an attestation section in which the clinician recorded that they had reviewed the note (Shah et al., 2025). The organization of reasoning, with subjective data, findings, assessment and plan, follows the same rules as a handwritten note, and it is worth reviewing how to organize clinical reasoning in the medical record before delegating the first version to a program.

What the randomized trials show

The two-scribe trial at UCLA

Lukac and colleagues randomized 238 outpatient physicians from fourteen specialties at an American academic institution to use one of two commercial tools, DAX Copilot or Nabla, or to continue without a scribe, between November 2024 and January 2025 (Lukac et al., 2025). Only consultations in English were included.

The primary outcome was the change in time in the note, comparing the second month of the study with the six previous months. With Nabla, the drop was 9.5 percent relative to the control, with statistical significance (Lukac et al., 2025). With DAX, the 1.7 percent drop was not significant (Lukac et al., 2025). In absolute terms, the mean time in the note fell about 41 seconds in the Nabla group, from 4 minutes and 30 seconds to 3 minutes and 49 seconds, against 18 seconds in the control group.

Actual use fell far short of the total number of consultations. DAX was used in 33.5 percent of 24,696 consultations and Nabla in 29.5 percent of 23,653 (Lukac et al., 2025). About 15 percent of the physicians randomized to a scribe never used it (Lukac et al., 2025).

The secondary outcomes, such as cognitive load, work exhaustion and the Mini-Z well-being score, improved modestly in both scribe groups, but the authors did not test hypotheses for them and say these findings need confirmation in larger, multicenter trials (Lukac et al., 2025). The chance of burnout measured by a single item was lower in the scribe groups, with confidence intervals that included the absence of effect.

The stepped-wedge trial in Wisconsin

Afshar and colleagues conducted a 24-week pragmatic trial, with individual stepped randomization, in UW Health outpatient clinics in two American states (Afshar et al., 2025). Sixty-six professionals took part, including physicians and advanced practice professionals, with the Abridge tool. Of the 71,487 notes in the period, 27,092 (38 percent) were generated with the scribe (Afshar et al., 2025).

In the two primary outcomes, the work exhaustion and interpersonal disengagement score fell 0.44 point with the scribe, with statistical significance (Afshar et al., 2025). Professional fulfillment rose 0.14 point, and the authors describe this increase as not significant, despite the P value of 0.04.

Time on notes fell by 0.36 hours per day (Afshar et al., 2025). Work outside scheduled hours fell by 0.50 hours per day, but this result was sensitive to the exclusion of extreme values and lost significance when the top 3 percent of observations were removed (Afshar et al., 2025). The study was open-label, with a group of early adopters who were younger, mostly white and female, and almost half in family medicine, and it did not collect patient-reported outcomes.

Observational studies

Before-and-after studies without a control group show larger changes, but they also carry more risk of bias. They answer whether those who used the tool felt better, not whether the tool caused the improvement.

Six health systems

Olson and colleagues evaluated use of the Abridge tool in six American health systems between February and October 2024 (Olson et al., 2025). Of the 451 clinicians enrolled, 272 answered the questionnaires before and after, and 263 outpatient clinicians entered the analysis.

Among the 186 who answered the burnout questions, the proportion with burnout fell from 51.9 to 38.8 percent after 30 days of use (Olson et al., 2025). Severe burnout, in a later analysis, fell from 18.4 to 12.2 percent (Olson et al., 2025). Cognitive load fell, attention focused on the patient rose from 6.51 to 8.56 on a 10-point scale, and the score for documentation outside scheduled hours, on the same 10-point scale, fell from 4.95 to 4.05 (Olson et al., 2025). Willingness to fit in urgent patients rose from 6.21 to 6.72, but the number of patients clinicians would accept adding did not change.

The authors note that the secondary outcomes are exploratory, without correction for multiple comparisons. The data came from a convenience sample, without a control group, with non-anonymous questionnaires and possible self-selection (Olson et al., 2025).

Stanford

Shah and colleagues followed 48 physicians who used DAX Copilot at Stanford between October 2023 and January 2024 (Shah et al., 2025). Task load fell from 68.98 to 44.56 points on a 0 to 100 scale, and the burnout score fell from 7.47 to 5.53 on a 0 to 10 scale (Shah et al., 2025).

The median time saved, according to the physicians themselves, was 20 minutes per half-day of appointments, but the variation was large: from 90 minutes saved to 90 minutes added (Shah et al., 2025). The small sample, made up of interested physicians, and the gift card offered to participants limit generalization.

Sutter Health

Stults and colleagues evaluated a pilot with 100 clinicians in California, with 57 pairs of questionnaires (Stults et al., 2025). Burnout went from 42.1 to 35.1 percent, without statistical significance (Stults et al., 2025). Time on notes per visit fell from 6.2 to 5.3 minutes. Satisfaction improved for 85.8 percent of primary care professionals, versus 36.4 percent in clinical subspecialties and 50.0 percent in surgical ones (Stults et al., 2025).

Comparison of the studies

Study	Design and site	Participants	Burnout or well-being	Documentation time
Lukac et al., 2025	Randomized trial, one academic institution	238 physicians	Modest improvements, without hypothesis testing	9.5 percent drop with one tool (Lukac et al., 2025); no difference with the other
Afshar et al., 2025	Stepped-wedge randomized trial, two states	66 professionals	Less exhaustion; accomplishment without significance	0.36 hours per day less on notes
Olson et al., 2025	Before and after, six systems	263 clinicians	Burnout from 51.9 to 38.8 percent (Olson et al., 2025)	Documentation outside scheduled hours score from 4.95 to 4.05, on a 10-point scale
Shah et al., 2025	Before and after, one institution	48 physicians	Burnout score from 7.47 to 5.53	Median of 20 minutes per half-day, reported
Stults et al., 2025	Before-and-after pilot, one organization	100 clinicians	Burnout from 42.1 to 35.1 percent, without significance (Stults et al., 2025)	From 6.2 to 5.3 minutes per visit
Pearlman et al., 2025	Retrospective cohort, one institution	125 users and 478 non-users	Out of scope	15.9 percent less on notes (Pearlman et al., 2025)

Sources: the studies cited in each row, read in full text.

Time in the medical record and productivity

Pearlman and colleagues compared 125 scribe users with 478 non-users balanced by characteristics, in an academic system, between July and September 2024 (Pearlman et al., 2025). In the weighted analysis, total time in the medical record per visit was 8.5 percent lower, about 2.4 minutes, and time in notes was 15.9 percent lower, about 1.8 minutes (Pearlman et al., 2025). Time outside working hours, visit duration and visit volume did not change.

At UCSF, Holmgren and colleagues analyzed 1,565 physicians, of whom 698 adopted the scribe (Holmgren et al., 2026). Adopters had an increase of 0.04 relative value unit per encounter and of 0.80 encounter per week, with no difference in claim denials (Holmgren et al., 2026). The authors themselves say it is necessary to check whether the increase reflects more services provided, more accurate code selection or inflated code selection.

What the time numbers hide

In the UCLA trial, the medical record metrics did not include editing time done inside the vendor platform, which may overestimate the savings (Lukac et al., 2025). That is, part of the review time may simply have moved elsewhere. The average gains, from seconds to a few minutes per note, are real, but they are far from solving on their own the burden described by Arndt and colleagues.

What errors the generated note brings

What the physicians reported in the trial

In the UCLA trial, physicians reported clinically relevant inaccuracies with a frequency described as occasional, with a mean score of 2.7 and 2.8 on a scale from 1 to 5 (Lukac et al., 2025). Examples included omissions, structural errors and pronoun swaps, the parents' history attributed to the patient and invented content. In one example cited by the authors, "not drinking alcohol" became "now drinking alcohol" in the note (Lukac et al., 2025).

One mild adverse event was recorded: extensive counseling of the patient was left out of the note (Lukac et al., 2025). In the usability assessment, the statement that the notes were at least as good as the physician's own received a neutral rating.

The test with simulated visits

Biro and colleagues read aloud 11 scripts of real outpatient visits and generated 44 drafts with two commercial products (Biro et al., 2025). They found 127 errors, a mean of 2.9 per note, and 31 of the 44 notes (70 percent) had at least one error (Biro et al., 2025). Omission was the most frequent error, followed by irrelevant or misplaced text, additions and wrong information, and the pattern varied between products.

The authors call attention to a practical point: omission is the hardest error to detect, because the physician needs to remember what was said, which becomes harder when the review happens after several visits (Biro et al., 2025). The study used simulation and few visits, and the authors call for independent tests.

What clinicians correct

Guo and colleagues analyzed 200 visits at UCI Health with drafts from two vendors (Guo et al., 2026). The most common edits involved test orders and referrals (40.0 percent), symptoms (30.3 percent), medications (27.3 percent) and description of diagnoses (25.9 percent) (Guo et al., 2026). Among the themes of the edits were correcting factual discrepancies and making the degree of diagnostic certainty explicit.

At the other end, the Afshar trial measured the quality of unedited notes with a language model acting as evaluator and found stigmatizing language in less than 1 percent of the notes (Afshar et al., 2025). This automatic assessment is useful for monitoring, but it does not replace the reading of whoever was in the room.

Why and how to check the note

A documentation error can become a care error. A missing medication, a history item attributed to the wrong person or an inverted negation can steer the next visit in the wrong direction, and the broader discussion of how this happens is in diagnostic error and the Improving Diagnosis report. When the error reaches the patient, the physician who signed the note suffers too, the subject of second victim: the physician after the error.

The studies suggest some points to watch during review:

• Look for what is missing, not only for what is wrong, because omission was the most common error and the hardest to notice (Biro et al., 2025).
• Check negations and attributions, such as "denies", "does not use" and family history, which appeared swapped in the UCLA trial (Lukac et al., 2025).
• Review medications, orders and diagnoses, the areas most edited by clinicians (Guo et al., 2026).
• Check the degree of certainty: a hypothesis under discussion should not appear as a confirmed diagnosis (Guo et al., 2026).
• Review right after the visit, while the memory of the conversation is still fresh (Biro et al., 2025).

Review as part of reasoning

The authors of the randomized trial themselves conclude that occasional inaccuracies require ongoing vigilance (Lukac et al., 2025). Guo and colleagues recommend solid and consistent review routines (Guo et al., 2026). Reading the note is also a moment to check whether the assessment reflects the reasoning that was done, and the broader debate on AI in this role is in use of artificial intelligence in clinical reasoning.

Patient consent

At an academic center in New York, Lawrence and colleagues studied consent for the scribe with 18 clinicians and 103 patients (Lawrence et al., 2025). Most patients, 74.8 percent, said they were comfortable with the tool (Lawrence et al., 2025). When they received only basic information, 81.6 percent consented, but the proportion fell to 55.3 percent when the AI features, data storage and involvement of companies were detailed (Lawrence et al., 2025).

The sample was small, restricted to English-speaking participants with high educational levels. Even so, the finding shows that consent changes depending on what is explained, and that explaining clearly is part of using the tool.

Back to the case

The physician in the fictional case gained some time, as the trials suggest, but the note with "now drinks" shows the other side. Changing one word changed the meaning of a history item, exactly the kind of error described in the UCLA trial (Lukac et al., 2025). The practical approach is to review every draft right after the visit, paying attention to negations, medications and what may have been left out.

How NEXO helps train this reasoning

In NEXO, the case of the day is fictional, with a real clinical presentation and up to six progressive clues, and you propose the diagnosis by looking for the code in ICD-10. Reviewing an AI-generated note requires the same exercise: comparing each recorded item with the real history, noticing the clue that was missing and checking whether the written hypothesis is the one you would stand behind, as in differential diagnosis training with clues.

The limits on use of the game are described on the NEXO Economy page. If you research teaching of clinical reasoning, the page for Researchers explains how to get in touch. Professionals who write or review cases find information in Curator and in Creators, and institutions can learn about the proposal on the partnerships page. Looking for the right ICD-10 code completes this training.

Frequently asked questions

Does the AI scribe reduce burnout?

Before-and-after studies show a drop in self-reported burnout, such as from 51.9 to 38.8 percent across six health systems, but without a control group (Olson et al., 2025). In the stepped-wedge randomized trial, exhaustion at work fell, while professional accomplishment showed no significant change (Afshar et al., 2025). The evidence points to benefit, but it still comes from short studies with interested adopters.

How much time does the scribe save?

Little per note. In the UCLA trial, the significant drop in note time was 9.5 percent with one of the tools (Lukac et al., 2025). In Pearlman's cohort, note time was about 1.8 minutes shorter per visit (Pearlman et al., 2025).

Can the generated note be signed without reading it?

The data advise against it. In simulated consultations, 70 percent of the notes had at least one error (Biro et al., 2025), and the randomized trial recorded occasional clinically relevant inaccuracies (Lukac et al., 2025).

Which error is hardest to notice?

Omission. Biro and colleagues observed that it requires the physician to remember what was said in the consultation, which becomes harder when the review is done hours later (Biro et al., 2025).

Is it necessary to ask the patient for consent?

The services studied asked for consent, and acceptance depended on how much was explained: 81.6 percent with basic information and 55.3 percent with detailed information (Lawrence et al., 2025).

Take-home messages

• Treat burnout as an occupational phenomenon, with exhaustion, distancing and reduced efficacy, as the WHO defines it (WHO, 2019).
• Expect modest time gains: in the randomized trial, the only significant reduction was 9.5 percent in time spent on the note (Lukac et al., 2025).
• Read burnout results with caution, because the largest improvements came from studies without a control group (Olson et al., 2025; Shah et al., 2025).
• Review each note right after the consultation, looking above all for what was left out (Biro et al., 2025).
• Check negations, medications, orders and the degree of diagnostic certainty, the areas most subject to error and editing (Lukac et al., 2025; Guo et al., 2026).
• Explain to the patient how the recording works, because consent changes with the level of detail (Lawrence et al., 2025).

References

1. World Health Organization. Burn-out an "occupational phenomenon": International Classification of Diseases. May 28, 2019. [Full text](https://www.who.int/news/item/28-05-2019-burn-out-an-occupational-phenomenon-international-classification-of-diseases).
2. Arndt BG, Beasley JW, Watkinson MD, Temte JL, Tuan WJ, Sinsky CA, Gilchrist VJ. Tethered to the EHR: primary care physician workload assessment using EHR event log data and time-motion observations. Ann Fam Med. 2017;15(5):419-26. [PubMed](https://pubmed.ncbi.nlm.nih.gov/28893811/).
3. Lukac PJ, Turner W, Vangala S, et al. Ambient AI scribes in clinical practice: a randomized trial. NEJM AI. 2025;2(12). [PubMed](https://pubmed.ncbi.nlm.nih.gov/41497288/).
4. Afshar M, Baumann MR, Resnik F, et al. A pragmatic randomized controlled trial of ambient artificial intelligence to improve health practitioner well-being. NEJM AI. 2025;2(12). [PubMed](https://pubmed.ncbi.nlm.nih.gov/41625485/).
5. Olson KD, Meeker D, Troup M, et al. Use of ambient AI scribes to reduce administrative burden and professional burnout. JAMA Netw Open. 2025;8(10):e2534976. [PubMed](https://pubmed.ncbi.nlm.nih.gov/41037268/).
6. Shah SJ, Devon-Sand A, Ma SP, et al. Ambient AI scribes: physician burnout and perspectives on usability and documentation burden. J Am Med Inform Assoc. 2025;32(2):375-80. [PubMed](https://pubmed.ncbi.nlm.nih.gov/39657021/).
7. Stults CD, Deng S, Martinez MC, et al. Evaluation of an ambient artificial intelligence documentation platform for clinicians. JAMA Netw Open. 2025;8(5):e258614. [PubMed](https://pubmed.ncbi.nlm.nih.gov/40314951/).
8. Pearlman K, Wan W, Shah S, Laiteerapong N. Use of an AI scribe and EHR efficiency. JAMA Netw Open. 2025;8(10):e2537000. [PubMed](https://pubmed.ncbi.nlm.nih.gov/41071550/).
9. Holmgren AJ, et al. Ambient AI scribes and physician financial productivity. JAMA Netw Open. 2026;9(1):e2553233. [PubMed](https://pubmed.ncbi.nlm.nih.gov/41511775/).
10. Biro J, Handley JL, Cobb NK, et al. Accuracy and safety of AI-enabled scribe technology: instrument validation study. J Med Internet Res. 2025;27:e64993. [PubMed](https://pubmed.ncbi.nlm.nih.gov/39869899/).
11. Guo Y, Hu D, Yang Z, et al. What do clinicians edit in ambient AI-drafted clinical documentation? J Am Med Inform Assoc. 2026;33(8):1457-65. [PubMed](https://pubmed.ncbi.nlm.nih.gov/42118974/).
12. Lawrence K, Kuram VS, Levine DL, et al. Informed consent for ambient documentation using generative AI in ambulatory care. JAMA Netw Open. 2025;8(7):e2522400. [PubMed](https://pubmed.ncbi.nlm.nih.gov/40694347/).