Ambient AI Scribes in US Outpatient Care: Wide Adoption, Modest Measured Time Savings

The use case

Ambient scribing tools capture the spoken conversation in a consultation, turn it into a transcript, and use a large language model to produce a draft note. The clinician then reviews, edits and signs the note¹². In a typical workflow, the clinician picks the encounter from the EHR schedule, gets the patient's verbal consent, records the visit, and gets a draft that can be checked against the transcript or audio⁵. The goal is to cut documentation time, especially after-hours EHR work, which is linked to burnout⁴. In the deployments reviewed here, the tool drafts notes and does not make care decisions³.

Where it is running

The largest documented deployment is at The Permanente Medical Group (Kaiser Permanente, Northern California). Between 16 October 2023 and 28 December 2024, 7,260 physicians used ambient scribes in 2,576,627 encounters (measured usage data)³. The group changed vendor in 2024. One reason was that the new product was integrated directly into its EHR. The tool was also extended across all eight Kaiser Permanente regions³.

A national study of US hospitals using Epic found that by June 2025, 1,744 of 2,784 (62.6%) belonged to systems that had implemented or were implementing an ambient tool. Three products (DAX Copilot, Abridge and ThinkAndor) made up more than 80% of implementations⁸. Adjusted adoption was 70.2% in non-profit hospitals and 28.8% in for-profit hospitals⁸. Other named deployments in peer-reviewed studies are:

• UCLA Health (Nabla and DAX)¹
• UW Health (Abridge)²
• UCSF Health, which offered access to 2,018 physicians starting August 2023⁶
• Mass General Brigham, Emory Healthcare, Yale New Haven Health and UC Davis Health⁴

In February 2026 Epic announced a built-in AI Charting feature, which several health systems were piloting¹⁵.

Adoption by organisations overstates how much individual clinicians use the tools. At UCSF, 44.6% of physicians in the analysed sample were adopters, and only 15.2% of encounters involved an adopter⁶. In a five-centre study, only about 32% of adopters used the tool in half or more of their visits⁴. At Kaiser Permanente, the top third of users accounted for 89% of activations³.

In England, NHS England has published national guidance and set up a self-certified supplier registry¹². In June 2025 it told staff to stop using ambient tools that were not registered for medical use¹³. Published evidence on outcomes in NHS settings remains thin.

What the evidence shows

There are two randomised trials so far:

• UCLA: 238 outpatient physicians across 14 specialties were randomised to DAX, Nabla or usual care between November 2024 and January 2025. Nabla cut time-in-note by 9.5% compared with the control group. DAX showed no significant change (−1.7%)¹. Both groups reported better burnout-related scores, which the authors say need confirming in larger multicentre trials¹.
• UW Health: a 24-week stepped-wedge trial of 66 practitioners. Work exhaustion fell by 0.44 points on a five-point scale. Professional fulfilment did not improve significantly. Time on notes fell by 0.36 hours per day (measured)².

The largest controlled observational study covered 8,581 clinicians, 1,809 of them adopters, at five academic centres between June 2023 and August 2025. Adoption was linked to 16.0 fewer minutes of documentation time and 13.4 fewer minutes of total EHR time per eight scheduled patient hours. Those are relative reductions of 10.0% and 3.0%. There was no significant change in after-hours EHR work (measured, EHR metadata)⁴. Gains were larger in primary care and among clinicians who used the tool in at least half of their visits⁴.

Most burnout evidence is self-reported. A pre–post survey of 263 clinicians in six health systems using Abridge during 2024 found burnout fell from 51.9% to 38.8% after 30 days⁵. The Peterson Health Technology Institute, an independent body that drew on eight health systems, found solid evidence of lower cognitive load and burnout. It found mixed results on after-hours work¹¹.

Kaiser Permanente surveyed 118 patients. Of these, 56% said the scribe improved the quality of the visit, none said it made the visit worse, and 8% reported some discomfort with AI being used (self-reported)³. On note quality, UW Health rated AI-assisted notes between 3.97 and 4.99 out of 5 across quality domains². UCLA clinicians reported clinically significant inaccuracies "occasionally"¹.

On finance, UCSF found that adoption was linked to 1.81 more relative value units (RVUs, the unit Medicare uses to price physician work) per week (5.8%) and 0.80 more encounters per week, with no change in claim denials. That comes to about $3,044 per physician per year at 2025 Medicare rates (measured, observational)⁶. The five-centre study estimated $167 per clinician per month in additional visit-billing revenue⁴.

What it takes: cost, integration, governance

Subscription fees are reported at $200–$600 per clinician per month. This figure comes from a 2026 JAMA Network Open commentary citing press reporting from May 2025, and it has not been audited⁷.

EHR integration is a deciding factor. Kaiser Permanente switched vendor partly to get direct EHR integration³. NHS England calls integration with the patient record "an essential requirement" and warns that the cost of custom interfaces "can be substantial"¹².

In England, organisations adopting these tools are expected to:

• appoint a Clinical Safety Officer
• complete DCB0160 safety documentation and a data protection impact assessment
• monitor performance independently of the supplier
• keep pilots to a maximum of four months
• accept that liability may stay with the NHS organisation¹²

The same guidance flags weaker performance with some accents, with speakers using English as a second language, and with speech impairments¹².

Consent needs its own workflow. In the six-system study, clinicians obtained verbal consent and recordings were deleted after a short grace period⁵.

Uptake also depends on support. The five-centre authors call for "robust training and support" because so few adopters used the tool heavily⁴. Kaiser Permanente relies on physician champions and reports lower uptake in inpatient and template-heavy specialties³.

Where it falls short

The challenge question was what would make this use case look less successful than it is usually presented. Testing the main claims against the strongest contrary evidence gives the following picture.

Time saving. Headline totals hide small per-person effects. Kaiser Permanente's estimate of 15,791 hours saved across 7,260 physicians³ averages about 2.2 hours per physician over a year. The multisite study puts documentation savings at 16 minutes per eight scheduled patient hours, with no significant change in after-hours work⁴. In the UCLA trial, one of the two products showed no significant effect¹. UW Health's reduction in after-hours work stopped being significant once the top 3% of daily observations were removed². Kaiser Permanente saw a small rise in in-basket time³. The multisite authors suggest some of the time saved on notes is taken up by other EHR work⁴.

Burnout. The most-cited figure comes from a survey with no control group. It was not anonymous, the vendor distributed it at five of the six sites, and only 60.3% of enrolled clinicians completed it⁵. The UW trial found no significant gain in professional fulfilment². None of the sources reviewed shows burnout effects lasting beyond six months.

Accuracy. In simulated encounters with clean audio, tested between December 2024 and February 2025, five platforms omitted or got wrong an average of 26.3% of key clinical elements in unedited notes. Each case averaged 3.0 errors judged capable of moderate-to-severe harm. The worst performer was a free tool⁹. In simulated interpreter-mediated visits tested in August 2025, two vendors carried 55% and 60% of scripted interpreter errors into the notes¹⁰. Both studies used simulations without clinician editing. The authors point out that there is no evidence yet on whether clinician review reliably catches these errors in practice⁹.

Cost. Measured revenue gains are $167 per month in the five-centre study⁴ and about $254 per month at UCSF ($3,044 divided by 12)⁶. Both are at or below the reported licence range of $200–$600⁷. The Peterson Health Technology Institute found no significant return on investment so far, and health systems told it they had not seen more patients per session¹¹. The JAMA Network Open commentary warns that payer downcoding policies could offset coding gains, and that pushing for higher patient throughput could wear away the wellbeing benefits⁷. Adoption is concentrated in financially stronger, non-profit hospitals⁸.

Consent and privacy. A proposed class action filed on 26 November 2025 alleges that Sharp HealthCare recorded visits using Abridge without patients' consent and added boilerplate consent statements to their records. The patient bringing the suit says he was told the vendor keeps recordings for 30 days¹⁴. These allegations have not been proven.

Stage: In wide use — why

The working hypothesis was "In early use." The evidence supports wider use in US ambulatory care:

• Most Epic hospitals belonged to adopting systems by mid-2025⁸.
• Kaiser Permanente has deployed the tools across the enterprise³.
• At least eight named academic systems have published outcomes¹²⁴⁶.
• Two randomised trials have been completed¹².

This rating has limits. Within adopting organisations, a minority of clinicians use the tools regularly⁴⁶. The evidence covers outpatient care almost exclusively. Outcome data from outside the US remain limited.

What to watch next

• Sharp HealthCare lawsuit. A ruling on class certification in San Diego Superior Court would show how California recording laws apply to ambient tools¹⁴.
• Epic AI Charting. The move from pilots (announced February 2026) to general release, and whether health systems then drop third-party scribe contracts, will reshape the market and its pricing¹⁵.
• NHS England. Risk-assessment templates and a common evaluation approach were promised for early 2026; whether and when they appear is worth tracking¹².
• Research and payment. Signs to look for are follow-up beyond roughly five months in multisite data⁴, the larger multicentre trials the UCLA authors called for¹, and any payer downcoding policies aimed at notes drafted with AI help⁷.

Questions to ask before adopting

• What share of our clinicians do we expect to use the tool in at least half their visits, and how will we support them to get there?
• Will the tool write directly into our EHR, and what will the interfaces cost?
• How will we audit note accuracy ourselves, independently of the vendor, including for interpreter-mediated and accented speech?
• What consent workflow will we use, how will we record consent, and how long are audio and transcripts kept?
• Which baseline measures will we track (documentation time, after-hours work, burnout, revenue), and against what control group?