AI Scribes Are Saving Doctors Time. Who Gets to Keep the Hours?

Kaiser Permanente Northern California put an ambient AI scribe in front of 7,260 physicians, logged 2.58 million scribed encounters in 63 weeks, and counted more than 15,700 physician hours saved. Where those hours went is the thinner part of the record, and what exists points toward the schedule, the coding level, and the payer's audit software more often than toward the physician's evening.

An hour a day is roughly 250 hours a year. For an employed physician it is worth more than the marginal RVU, because it can become a higher savings rate now and a longer career later, and because the employer, the payer, and the physician each have a claim on it.

What the studies show

• Kaiser Permanente Northern California (NEJM Catalyst, 2025). Design: 7,260 physicians, 2.58 million encounters, 63 weeks. Time and well-being: More than 15,700 physician hours saved; high users spent less time in notes and about a minute less after-hours "pajama time" per appointment. Volume and revenue: Not reported.
• Five health systems: MGB, Emory, UCSF, Yale, UC Davis (JAMA, 2026). Design: 8,581 clinicians, 1,809 adopters. Time and well-being: 13.4 fewer minutes of EHR time and 16 fewer minutes of documentation per 8 scheduled hours; no significant change in after-hours EHR time. Volume and revenue: +0.49 visits per week; about $167 per clinician per month in marginal E/M revenue.
• UW Health (NEJM AI, 2025). Design: Stepped-wedge randomized trial, 66 practitioners, 24 weeks. Time and well-being: Work exhaustion down; documentation time down about 22 minutes a day. Volume and revenue: Not reported.
• UCLA (NEJM AI, 2025). Design: Randomized, 238 physicians, two vendors vs control. Time and well-being: One vendor cut time-in-note 9.5%; the other, not significantly; burnout and task load improved with both. Volume and revenue: Not reported; tools used in only about a third of visits.
• Mass General Brigham and Emory (JAMA Network Open, 2025). Design: 1,430 clinicians, pre/post. Time and well-being: Burnout down 21 percentage points at MGB; documentation-related well-being up 31 points at Emory. Volume and revenue: Not reported; low survey response.
• UCSF (JAMA Network Open, 2026). Design: 1,565 physicians, 698 adopters, difference-in-differences. Time and well-being: Not reported. Volume and revenue: +5.8% RVUs, +2.8% encounters per week, about $3,044 a year per physician; authors could not separate more services from higher coding.
• Providence (JAMA Network Open, 2026). Design: 1,547 users, 16,149 observation-months. Time and well-being: Sustained decline in after-hours work. Volume and revenue: Small significant rise in RVUs with no increase in daily appointments.
• Atrium Health (NEJM AI, 2024). Design: 112 users vs 103 controls, 180 days. Time and well-being: High users cut documentation hours about 7%. Volume and revenue: No significant change in RVUs, visits, or after-hours time.

The time savings are real and consistent, in the range of a quarter to half an hour a day for regular users. And in the studies that measured it, the gain showed up as RVUs and coding level at least as often as it showed up as time at home. The largest study found no change in after-hours work at all.

The hour has three possible owners

An hour freed by a scribe can go to you. That is the after-hours reduction, the "pajama time" cut, the exhaustion score that moves. Kaiser and Providence found it. The five-system study did not, and the most likely reason is the second owner.

It can go to your employer. The employer bought the scribe, at prices trade press puts at several hundred dollars per clinician per month, and it owns the schedule template. If the template adds a slot when notes get faster, the hour becomes a visit. The five-system study's extra half-visit a week, UCSF's 2.8% more encounters, and the finding that scribe users generate more RVUs without more appointments (Providence) are all the same thing seen from different angles: the employer booking the productivity gain. The Peterson Health Technology Institute's 2025 review of eight systems found most organizations reporting no increase in encounters early on and called the financial impact unclear, and by 2026 its director told reporters the consensus among insurers and providers was that scribes are raising coding intensity. A study of six adopting systems found high-level established-visit codes up 7 to 12 percentage points from 2018 to 2024.

It can go to the payer. Higher coding intensity is a documented effect, and payers have noticed. PwC's 2027 medical cost trend survey found 70% of health plans naming AI-enabled provider revenue capture as a top-three cost inflator. Cigna's reimbursement policy effective October 2025 automatically adjusts level 4 and 5 office and consultation codes down one level for a subset of providers. The policy doesn't mention scribes. They don't need to. A cleaner, fuller, more complex note is a better audit target, and downcoding by algorithm is cheaper than manual review. The scribe documents more complexity, and the payer's model downcodes it.

Follow the money

Whoever paid for the scribe, sets the schedule template, and books the revenue when RVUs rise owns the hour, and the software's quality has nothing to do with it. For 82% of U.S. physicians that is the employer on all three counts. Adoption has moved fast on the employer side: 63% of Epic hospitals had adopted ambient AI by mid-2025, and by early 2026 about 29% of physicians reported using an AI scribe, up from 20% the year before. That is a tool bought by the entity, deployed on the entity's schedule, generating revenue on the entity's ledger, and used by a physician who is paid per RVU. The physician puts in the hours, the entity books the RVUs, and compensation rises a little less than the RVUs do. Systems bought the tools partly because burnout is expensive and partly because the tools pay for themselves through coding and throughput.

Why keeping the hour matters more than it looks

For an employed physician the value of the hour is two things classical physician finance treats as fixed and AI is turning into variables.

The first is savings rate. Time reclaimed from documentation that is spent on nothing in particular still lowers the burnout that drives physicians to cut hours, take unpaid leave, or buy back their evenings with expensive conveniences. Time reclaimed and kept, rather than filled with another slot, is money that stays in your plan rather than the entity's.

The second is career length, and this is the one that changes the arithmetic. A 2026 Permanente Journal study of physicians who left clinical practice found the mean age at departure was 48.1, down from 57.1 in an earlier cohort; that is a figure for leavers, not a forecast for any individual career. Doximity's 2026 survey found 46% of physicians considering early retirement, up from 34% the prior year, which is an intention rather than an outcome. Classical physician finance assumes a thirty-year career because the compounding math needs one. Take the early exit as a stress case and a career that ends at 48 gives the plan fifteen years. The scribe is the first tool in a generation that credibly adds years back, and it only does so if the reclaimed hour lowers the load rather than raising the quota.

Offload generates capital only if you keep the dividend rather than ceding it as more volume, and the employer decides that before the go-live, when it sets the template.

What to watch in your own practice

Count your scheduled slots per session before and after any scribe rollout, and ask your employer for your after-hours EHR minutes over the same months; the EHR vendor reports that number to them already.

If you can't answer those, that is itself information about who is measuring and who isn't.

Where the book picks up

The Income Variable treats administrative offload as the one lever every specialty gets, the least speculative response to AI in the whole book, and the one most likely to be given away by accident. Chapter 7 works the two channels (savings rate now, career length later) and why they must be kept distinct. Chapter 17 is about bringing your own numbers to a contract negotiation, including the terms that bank the scribe dividend rather than handing it back as volume.

The mechanisms that decide who captures the dollar across all five income inputs are in How AI Could Reduce Physician Pay Without Replacing Physicians, and the fee-schedule side of the same story is in Why Physician Reimbursement Can Fall Even When Productivity Rises. Join the launch list for the book and the free tools.

‍