AI Workflows  /  Field note

How Health Systems Are Measuring Ambient AI Scribe ROI: What Kaiser, Sutter and Mass General Found

Ambient AI scribes are the most widely deployed generative AI tool in medicine, and named health systems are now publishing hard numbers on time, burnout and revenue. Here is how leaders define the return, where it shows up, and where the evidence still has holes.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published June 28, 2026

Health systems measure ambient AI scribe ROI across three buckets: clinician time and documentation burden (note time, after-hours "pajama time," total EHR time); clinician well-being and retention (burnout and satisfaction); and direct financial productivity (work relative value units, visit volume, and evaluation-and-management coding accuracy). Published results from named systems including The Permanente Medical Group, Sutter Health, Mass General Brigham, Emory Healthcare, and St. Luke's Health System show documentation-burden and burnout reductions commonly in the 20 percent to 40 percent range, with early single-site evidence of roughly a 5.8 percent gain in weekly RVUs. This is an analysis of reported outcomes, not medical, legal, or financial advice, and AI scribes draft notes for clinician review rather than replacing clinical judgment.

The question every CMIO is now asking

A physician finishes her last visit of the day, and for the first time in years she does not open her laptop again at 9 p.m. That small change, repeated across thousands of clinicians, is what hospital executives are now trying to price. Ambient AI scribes, the tools that listen to a clinical visit and draft a note for the clinician to review, have become the most widely implemented generative AI application in health care. Their spread has changed the conversation in the C-suite. The early question was whether the technology worked at all. The current one, posed directly in a 2025 JAMA Network Open editorial by Shreya J. Shah and Patricia Garcia titled "Ambient AI Scribes, What Is the Return on Investment?", is whether the math holds up. Most health systems pay subscription fees of roughly $200 to $600 per clinician per month, a recurring cost that has to be justified as operating margins tighten.

The return is slippery because it shows up in three different places at once, and no single number captures it. Leaders at named systems have settled on three measurement buckets: clinician time and documentation burden, clinician well-being and retention, and direct financial productivity. Each has its own data, its own caveats, and its own published evidence from real health systems. Treat any one bucket in isolation and the case looks thin. Stack all three and the picture sharpens.

Bucket one: time and documentation burden

Time is the most consistent and best-documented return, and it is where the strongest evidence lives. The Permanente Medical Group (TPMG), the physician group for Kaiser Permanente in Northern California, ran one of the largest evaluations to date. In an analysis published in NEJM Catalyst (CAT.25.0040) and covered by the American Medical Association, TPMG tracked AI scribe use over a 63-week period from October 2023 through December 2024. A total of 7,260 Permanente physicians used the technology across 2,576,627 patient encounters. The headline figure is striking: an estimated 15,791 hours of documentation time saved for users compared with non-users, the equivalent of about 1,974 eight-hour workdays in a single year. Physicians using the scribes saw statistically significant reductions in note-taking time, time spent per appointment, and "pajama time," the after-hours work done outside 7 a.m. to 5:30 p.m. that clinicians and researchers repeatedly link to burnout.

The Permanente data surfaced something else that matters for ROI modeling. The top third of users accounted for 89 percent of all AI scribe activations, and heavy users saved more than double the time per note compared with lower-frequency users. That dose-response pattern, described by TPMG co-authors Kristine Lee, MD, and Vincent Liu, MD, tells operations leaders the return scales with adoption, not just deployment. Buying the licenses is not the win. Getting clinicians to reach for the tool at every visit is.

Other systems point the same way. Sutter Health, which uses Abridge, published a JAMA Network Open study (jamanetworkopen.2025.8614) led by Cheryl Stults, PhD, with chief medical information officer Veena Jones, MD, as senior author. Across 100 clinicians spanning specialties in Northern California, mean time in notes per appointment fell from 6.2 to 5.3 minutes (p<0.001), alongside less mental demand and less of the rushed feeling clinicians describe. The Cleveland Clinic, deploying Ambience, reported that clinicians cut the average time spent writing and reviewing notes by about 14 minutes per day. Cooper University Health Care in Camden, New Jersey, using Microsoft Dragon Copilot, saved clinicians about 4.15 minutes of documentation time per patient, which it said adds up to roughly an hour or more reclaimed daily. Intermountain Health in Salt Lake City reported a 27 percent reduction in time in notes per appointment among clinicians who used Dragon Copilot for 10 or more encounters. Different vendors, different cities, one direction.

Bucket two: burnout, well-being and retention

The second bucket is harder to put a dollar figure on, yet it may carry the largest financial weight. A Mass General Brigham led study, published in JAMA Network Open (jamanetworkopen.2025.28056) and drawing on surveys of more than 1,400 physicians and advanced practice providers at MGB and Atlanta's Emory Healthcare, found that ambient documentation use was associated with a 21.2 percent absolute reduction in burnout prevalence at 84 days at Mass General Brigham, while Emory Healthcare saw a 30.7 percent absolute increase in documentation-related well-being at 60 days. Rebecca Mishuris, MD, MPH, chief medical information officer at Mass General Brigham and a study co-senior author, said physicians told the team they had "their nights and weekends back," adding that "there is literally no other intervention in our field that impacts burnout to this extent." That is a remarkable sentence from someone whose job is to be skeptical of vendor promises.

The MGB authors were candid about the limits. The findings came from pilot users with modest survey response rates, about 22 percent at 84 days at MGB and 11 percent at 60 days at Emory, so they likely reflect the experience of the more enthusiastic adopters. The team also disclosed that Mass General Brigham holds an institutional investment in Abridge, exactly the kind of conflict a careful reader should weigh. The program scaled from an 18-physician proof-of-concept in July 2023 to more than 3,000 routine users by April 2025, and the system has said it will keep tracking whether burnout gains hold, plateau, or reverse over time. Honest measurement does not stop at the press release.

Sutter's study showed the same direction with more caution on significance. Burnout fell from 42.1 percent to 35.1 percent, but the change carried a p-value of 0.12, meaning it did not reach statistical significance in that 100-clinician sample. That candor is itself instructive. Well-being signals are real and consistent across systems, yet sample sizes and survey response rates often leave individual results short of statistical proof. The JAMA editorial connects the well-being story to the balance sheet, citing evidence that burnout doubles or triples the likelihood of physician turnover, with each departing physician costing an estimated $500,000 to $1 million in recruitment, onboarding, and lost productivity. Run that math and even modest retention gains can dwarf the subscription line.

Bucket three: financial productivity and coding

The newest and most scrutinized bucket is direct revenue. A University of California, San Francisco study by Holmgren and colleagues, published in JAMA Network Open (jamanetworkopen.2025.43524) and analyzing nearly 1.2 million ambulatory encounters across roughly 1,565 physicians over two years, found that ambient AI scribe access was associated with a 5.8 percent increase in weekly work RVUs, about 1.81 additional RVUs per week, and a 2.8 percent increase in encounters, with no increase in claim denials. The accompanying editorial translated that weekly RVU gain to roughly $3,000 to $3,044 per physician per year at 2025 Medicare rates. Small in isolation, but potentially enough to offset subscription fees once it is multiplied across hundreds or thousands of clinicians. The RVU gain also grew over time, consistent with a learning curve, which suggests the early numbers understate the ceiling.

Vendor-commissioned but independently conducted KLAS Research validations have pushed the financial story further by isolating coding accuracy. A KLAS ROI validation of St. Luke's Health System's enterprise deployment of Ambience Healthcare reported per-clinician annual revenue generated through enhanced Hierarchical Condition Category capture and greater evaluation-and-management coding accuracy, framing the gains as offsetting the cost of the technology and easing downstream burden on revenue cycle teams. A separate KLAS validation of Suki across FMOL Health, McLeod Health, and Rush University System for Health found a 35 percent to 65 percent reduction in after-hours documentation and an average incremental revenue of $1,223 per provider per month. The mechanism was a coding shift. At McLeod Health, level 3 codes for established patients dropped 18.2 percent while level 4 rose 7.3 percent and level 5 rose 5 percent, which McLeod CMIO Bryon Frost, MD, attributed to "more complete" notes. McLeod reported a net gain that started at $1,004 per provider per month after subscription cost and later climbed to $2,629. FMOL Health reported a 21 percent decrease in time in notes, a 43 percent decrease in notes left open more than seven days, and incremental revenue of $862 per user per month.

Coding-based ROI carries the heaviest asterisk of the three. The JAMA editorial cautions that higher RVUs could reflect more services, better documentation accuracy, or fuller capture of complexity, and warns that downcoding policies from payers such as Aetna and Cigna could swallow the coding gains. Systems measuring this bucket should separate honest documentation of work actually performed from anything that smells like a coding arms race, and they should watch denial rates as a guardrail. Revenue that arrives through a back door tends to leave the same way.

What good measurement looks like

The systems that report credibly share a few habits. They define a pre-period baseline and measure the same clinicians before and after, as TPMG, Sutter, and MGB did. They report response rates and statistical significance honestly instead of cherry-picking the flattering number. They separate adoption from deployment, since the Permanente dose-response shows the return concentrates among heavy users. And they disclose financial conflicts, as MGB did regarding its Abridge stake. The JAMA multisite picture is a useful reality check: across five academic medical centers using Ambience, DAX Copilot, and Abridge with Epic, AI scribes modestly cut total EHR time by 13.4 minutes and documentation time by 16.0 minutes, with 0.49 more visits per week. Real, measured gains, not a revolution. For CMIOs and clinical operations leaders, the lesson is that the return here is genuine and lives in several places at once, but it has to be measured deliberately, with the rigor a system would bring to any capital investment, and with clear eyes about pilot bias and shifting payer rules.

Browse all AI Workflows

ROI figures reflect specific institutional studies; results vary by practice size, specialty, and workflow. Informational analysis, not medical, legal, or financial advice.