How an Ambient AI Scribe Works, and What It Takes Off Clinicians
An ambient AI scribe turns a recorded visit into a draft note that the clinician reviews and signs. Randomized trials found modest time savings and less work exhaustion, while diagnosis and treatment stay with the physician.
Checked against the versions in the published trials (Microsoft DAX Copilot 2.0, Nabla 1.5, Abridge's generally available release of October 10, 2024) and the Microsoft Dragon Copilot and Abridge product pages as of that date
Role scorecard
Family Medicine Physicians SOC 29-1215.00
- Core tasks AI does in production0 of 6
- Most mature evidenceIn production
- CostNot disclosed
- Benefit0.36 fewer hours a day on notes (UW Health RCT)
As of · Tasks: O*NET 31.01
| Core task (O*NET) | What AI or robots do | Extent | Evidence grade | Example case | Source |
|---|---|---|---|---|---|
| Prescribe or administer treatment, therapy, medication, vaccination, and other specialized medical care to treat or prevent illness, disease, or injury | Nothing in the deployments studied: the scribe was documentation-only. Microsoft says its product can capture orders spoken during the visit for the clinician | Not yet | Vendor claim | Microsoft Dragon Copilot product page, 2026 | 2, 3 |
| Order, perform, and interpret tests and analyze records, reports, and examination information to diagnose patients' condition | No deployed scribe makes diagnostic decisions. Abridge markets an assistant that answers clinical questions from the chart and the conversation | Not yet | Vendor claim | Abridge Assistant product page, 2026 | 4, 5 |
| Collect, record, and maintain patient information, such as medical history, reports, or examination results | Records the visit, transcribes it and drafts the visit note; the clinician reviews, edits and signs it before it enters the record | Assists | In production | Abridge at Kaiser Permanente, 40 hospitals and 600+ offices, since August 2024 | 6, 2, 7 |
| Monitor patients' conditions and progress and reevaluate treatments as necessary | No deployed case. Microsoft says clinicians can ask questions answered from transcripts, notes and medical references | Not yet | Vendor claim | Microsoft Dragon Copilot product page, 2026 | 3 |
| Explain procedures and discuss test results or prescribed treatments with patients | The conversation stays with the physician; the scribe aims to free their attention. Microsoft says it also drafts after-visit summaries | Not yet | Vendor claim | Microsoft Dragon Copilot product page, 2026 | 3, 4 |
| Refer patients to medical specialists or other practitioners when necessary | No deployed case of AI deciding a referral. Microsoft says its product drafts referral letters | Not yet | Vendor claim | Microsoft Dragon Copilot product page, 2026 | 3 |
Extent and grade follow our evidence grades: "Does it" needs a production case where AI performs the task on its own; "Assists" means AI helps a person do it; "Not yet" means only prototypes, vendor claims or nothing. Six of the 11 core tasks O*NET lists for this role are shown; the five left out (advising on prevention, directing staff, coordinating with other providers, planning health programs and training) fall outside what ambient scribes do in the sources below. Documentation counts as "Assists", not "Does it", because in every deployment we found the clinician reviews and signs each note. The trials enrolled several specialties; family medicine was the largest group at UW Health (45.5% of practitioners).2
Costs and benefits
| License price | Not disclosed: neither Microsoft's nor Abridge's product page lists a price (checked 2026-10-13)35 |
|---|---|
| Who paid in the studies | UW Health bought licenses under a software-as-a-service agreement, with no funding from Abridge (trial report)2; in a six-system study, each institution funded its own Abridge use8 |
| Time on notes | 0.36 fewer hours a day (95% CI 0.17 to 0.55), a secondary outcome (UW Health randomized trial; UW Hospital and Clinics and NIH funding)2 |
| Documentation time, at scale | "Nearly 16,000 hours" saved in a year across 2.5 million+ uses (operator-stated, Kaiser Permanente)4 |
| Billing | 1.81 more RVUs a week per adopting physician, which the authors translate to $3,044 a year at the 2025 Medicare fee schedule (UCSF cohort study; philanthropic gift)9 |
Seen at work
As of · Independent and operator studies of ambient AI scribes; results as each source states them
| Study | Journal, date | Design | Clinicians (n) | Product | Main outcome | Result as stated | Funder / vendor role | Source |
|---|---|---|---|---|---|---|---|---|
| UCLA Health, DAX armLukac et al. | NEJM AI, Nov 26, 2025 | Randomized trial | 238 | DAX Copilot 2.0 (Microsoft) | Change in time-in-note vs usual care (primary) | −1.7% (95% CI −9.4% to +5.9%), not significant | UCLA Department of Medicine and others; vendor representatives helped deliver training | 7 |
| UCLA Health, Nabla armLukac et al. | NEJM AI, Nov 26, 2025 | Randomized trial | 238 | Nabla 1.5 | Change in time-in-note vs usual care (primary) | −9.5% (95% CI −17.2% to −1.8%) | UCLA Department of Medicine and others; vendor representatives helped deliver training | 7 |
| UW HealthAfshar et al. | NEJM AI, Nov 26, 2025 | Stepped-wedge randomized trial | 66 | Abridge (Oct 10, 2024 release) | Work exhaustion and professional fulfillment, 5-point scale (coprimary) | Exhaustion −0.44 (95% CI −0.62 to −0.25); fulfillment +0.14, "nonsignificant" | UW Hospital and Clinics and NIH; no funding from Abridge | 2 |
| UCSF HealthHolmgren et al. | JAMA Network Open, Jan 9, 2026 | Cohort, difference-in-differences | 1,565 | Two commercial scribes (not named) | RVUs, visits and claim denials | +1.81 RVUs a week (95% CI 0.86 to 2.75); +0.80 visits a week; no change in denials | Philanthropic gift; funder had no role | 9 |
| ProvidenceHusa et al. | JAMA Network Open, May 29, 2026 | Interrupted time series | 1,547 | DAX Copilot (Nuance) | Minutes in notes per appointment | −0.26 min in the first month (95% CI −0.41 to −0.11); median 7.1 to 6.1 min | Providence research teams; no conflicts reported | 10 |
| Kaiser Permanente (TPMG)Tierney et al., summarized by the operator | KP Division of Research, Apr 1, 2025 | Operator one-year review | 7,260 | Abridge in the 2024 rollout (vendor changed after the 2023 pilot) | Use and documentation time | 2,576,627 encounters, Oct 2023–Dec 2024; "nearly 16,000 hours" saved | Operator's own research division | 4, 6 |
The outcomes differ from study to study (time in notes, burnout scales, billing), so the rows can't be compared with each other or ranked. Studies with vendor staff among the authors, and figures the vendors state themselves, are in the separate table below. RVU means relative value unit, the measure Medicare uses to price physician work.
As of · Vendor-involved study and vendor-stated figures, kept apart from the independent results above
| Source | Type | Setting | Figure as stated | Vendor role | Ref. |
|---|---|---|---|---|---|
| Olson et al., JAMA Network Open, Oct 2, 2025 | Vendor-involved study (pre-post, quality improvement) | 263 clinicians, 6 US health systems, 30 days of Abridge | Share reporting burnout fell from 51.9% to 38.8% (odds ratio 0.26) | Two Abridge employees among the authors; Abridge distributed the survey at 5 of 6 sites and provided pooled data | 8 |
| Abridge clinicians page | Vendor claim | CHRISTUS Health | "78% reduction in cognitive load" (vendor-stated) | Vendor's own marketing | 5 |
How it works
1. Consent, then capture
An ambient AI scribe is software that listens to a clinic visit and drafts the clinician's note from it, without anyone dictating. UW Health's trial required an Apple mobile device with Epic's Haiku app, and Epic's interfaces carried the audio to Abridge and the draft note back.2 Consent comes first. UW Health told patients about the scribe at check-in and again in the room, recorded only with consent and let patients decline without consequence.2 In the UCLA trial, California's two-party consent law meant physicians got verbal consent from everyone present and wrote it in the note.7 Kaiser Permanente says its tool requires patient consent.6 In the six-system Abridge study, patients were told recordings and transcripts would be erased after a short grace period.8
2. Speech recognition makes a transcript
Automatic speech recognition turns the audio into a transcript of who said what. UW Health describes the Abridge release it deployed (generally available as of October 10, 2024) as combining speech recognition with language generation, partly trained on de-identified clinical conversations with gold-standard transcripts and human annotations.2 Errors can start here. One UCLA physician reported the scribe hearing "not drinking alcohol" as "now drinking alcohol"; others reported wrong pronouns and mixed-up speakers.7
3. A language model drafts the note
A large language model (an AI system trained on huge amounts of text to predict and generate it) then condenses the transcript into a structured note, such as history, examination, assessment and plan. Abridge's system includes several components, some of them custom large language models.2 In the six-system study, the draft appeared in a standard office-note format on a secure portal, and clinicians could highlight any passage to see the transcript behind it or hear the source audio.8 Microsoft says its current product, Dragon Copilot, turns multi-party, multilingual conversations into specialty-specific notes and can also draft referral letters and after-visit summaries; those are vendor claims.3
4. The clinician reviews and signs
In each deployment described here, a clinician reviews the draft before it is filed. Kaiser Permanente says doctors review and edit the suggested summary for accuracy before it enters the medical record.46 That review is real work: UCLA's primary measure of time in notes did not count time spent editing drafts inside the vendors' own apps.7 UCLA physicians rated "The tool generated notes at least as good as my own" lowest of all the usability questions, averaging "neutral" for both products.7
5. The note lands in the health record
The signed text goes into the electronic health record (EHR), the system that stores each patient's chart. At UCLA both products filled Epic's own note fields.7 UW Health connected Abridge through Epic's private interfaces and the FHIR R4 standard, and stressed that the software was documentation-only, with no diagnostic or treatment decision support.2 Kaiser Permanente says the same of its tool: it doesn't make decisions or recommendations about care.4 That is why the scorecard above marks diagnosis, prescribing and referral "Not yet".
6. How errors are caught and measured
A hallucination is text a model produces that wasn't in its input, such as a symptom nobody mentioned. UCLA asked physicians how often a note held at least one clinically significant inaccuracy (a hallucination, omission or addition); both products averaged "occasionally" (2.7 and 2.8 on a five-point scale).7 Omissions were the most common complaint (12 reports), then structure and formatting (11) and pronoun errors (8); hallucinations proper drew one report per product. The trial recorded one patient-safety event, counseling missing from the plan and patient instructions, which physician co-authors graded mild.7 UW Health scored note quality with the PDSQI-9 instrument (means of 3.97 to 4.99 across domains on a five-point scale) and detected no drift in the software's performance.2 In its 2023 pilot, Kaiser Permanente scored a sample of 35 AI transcripts at an average of 48 out of 50 and found few hallucinations (operator-stated).4
7. What it takes off clinicians, by the evidence
The measured gains are real but uneven. In UCLA's randomized trial, Nabla cut time in notes by 9.5% against usual care, while DAX Copilot's 1.7% was not significant.7 In UW Health's trial, work exhaustion fell by 0.44 points and time on notes by 0.36 hours a day, and billing diagnosis codes improved without loss of note quality.2 Larger observational studies point the same way: Providence clinicians spent less time in notes per appointment and less after hours,10 and UCSF adopters billed 1.81 more RVUs a week with no rise in claim denials.9 Use stays partial, though. At Providence, about 8% of clinicians were active users over the study period.10 Patients had mixed views at Kaiser Permanente: in a survey of 118, two-thirds were comfortable with the tool and 8% reported some discomfort.4 These results describe documentation work only. This page explains how the technology works; it doesn't recommend any product for any practice or patient.
Sources
- O*NET OnLine (O*NET 31.0), 29-1215.00 Family Medicine Physicians: tasks, retrieved 2026-10-13.
- Afshar M, et al., NEJM AI (via PubMed Central), A Pragmatic Randomized Controlled Trial of Ambient Artificial Intelligence to Improve Health Practitioner Well-Being, November 26, 2025 (doi:10.1056/AIoa2500945).
- Microsoft, Microsoft Dragon Copilot (vendor page), retrieved 2026-10-13.
- Kaiser Permanente Division of Research, AI-assisted notetaking gains steady support from Kaiser Permanente physicians, April 1, 2025.
- Abridge, Abridge for clinicians (vendor page), retrieved 2026-10-13.
- Kaiser Permanente, Kaiser Permanente improves member experience with AI-enabled clinical technology, August 14, 2024.
- Lukac PJ, et al., NEJM AI (author manuscript via PubMed Central), Ambient AI Scribes in Clinical Practice: A Randomized Trial, November 26, 2025 (doi:10.1056/AIoa2501000).
- Olson KD, et al., JAMA Network Open (via PubMed Central), Use of Ambient AI Scribes to Reduce Administrative Burden and Professional Burnout, October 2, 2025 (doi:10.1001/jamanetworkopen.2025.34976).
- Holmgren AJ, et al., JAMA Network Open (via PubMed Central), Ambient Artificial Intelligence Scribes and Physician Financial Productivity, January 9, 2026 (doi:10.1001/jamanetworkopen.2025.53233).
- Husa RA, et al., JAMA Network Open (via PubMed Central), Ambient Artificial Intelligence Use and Clinician Documentation Burden, Productivity, and Efficiency, May 29, 2026 (doi:10.1001/jamanetworkopen.2026.15762).