A physician finishes today’s note noticeably faster than they used to. The AI scribe listened, drafted, and closed most of the gap between seeing the patient and finishing the paperwork. At the end of the day? That’s less time hunched over a keyboard once the last patient leaves.
But ask, three months in, whether the practice is seeing more patients or capturing more revenue because of it. What comes back is a guess wearing the clothes of an answer. The tool did exactly what it promised. Where did the time it freed up actually go? That’s the question almost nobody in the practice has had a reason to ask yet.
The relief was real, the revenue was a guess
Adoption isn’t the question anymore. Nearly everyone has some form of AI running in the practice by now. Whether it paid off is a different matter, and the honest answer, per a May 2026 MGMA Stat poll of 257 practice leaders, is that fewer than half, 46 percent, will say it’s made providers more productive.
Real physician voices, twice a week
Free, and one click to unsubscribe.
That number alone doesn’t tell you much, because it’s actually hiding two different questions with two very different answers. Ask how the workday feels, and the evidence is strong and consistent. A JAMA Network Open study from October 2025, following 263 clinicians across 6 health systems, measured burnout falling from 51.9 percent to 38.8 percent within 30 days of adopting an ambient AI scribe. That’s a measured drop, not a mood survey.
Ask what the workday produces, and the evidence gets thin fast. An NEJM AI editorial from researchers at UC San Diego and Kaiser Permanente said it plainly: AI scribes are not productivity tools yet. Their own review found savings of roughly 7 to 22 minutes per provider a day, and even flagged that number as possibly generous, since some of it comes from EHR signal data rather than someone actually watching what happens in the room.
The tool improved one thing that was measured carefully. Nobody measured the other thing at all.
Nobody rebuilt the workflow around the time saved
A January 2026 cohort study out of UCSF, published in JAMA Network Open, went straight at the financial question, comparing revenue, patient volume, and claim denials between physicians who’d adopted an AI scribe and those who hadn’t, inside the same health system. A separate 150-day study of primary care clinicians found something similar: documentation time savings crept up slowly, and billed wRVUs barely moved.
It’s tempting to read that as early days, give it a few more months and the gains will show up. But the study ran 150 days, and the improvement it did find, five months in, was still a rounding error against payroll. Time alone wasn’t the missing ingredient.
Picture what that actually looks like on the ground. A physician finishes a note in four minutes instead of seven. Multiply that by twenty patients a day, and the arithmetic feels like it should add up to something: an extra slot before lunch, or a few more minutes with the sickest patient of the afternoon. Usually it doesn’t. The three minutes just get absorbed into whatever the day was already going to do to itself.
That’s worth sitting with, because the natural instinct is to treat a good tool as its own guarantee. It isn’t. The scribe does its job well: faster notes, fewer hours lost to documentation after clinic ends, and that’s a genuine win, no argument there. But a tool that saves time doesn’t automatically decide what happens to that time. Something else has to.
A separate MGMA Stat poll gets at exactly why that something else so rarely happens. Sixty-eight percent of medical groups haven’t redesigned a single role or adjusted staffing to account for the AI tools already running in their practice.
Most practices have someone with the authority to change a schedule, and someone else with the visibility to notice where the minutes went. Rarely the same person. What that costs is the easier half of this to see.
Burnout relief isn’t nothing
None of what’s above is a case against the tool itself, and it’s worth pausing to say so plainly before going further. A practice that only ever gets the experiential win, like calmer clinicians or fewer evenings lost to notes, has still bought something real, and for a field this short on people, that alone might justify the cost.
But most practices didn’t buy the tool for that reason, or not only for that reason. The pitch usually came with a number attached, like more visits, faster throughput, or a payback period. Maybe that’s an unfair thing to expect from a note-taking tool. It’s also exactly what most of them were sold as.
The redesign nobody scheduled
The minutes a scribe saves per note don’t wait patiently in some account, ready to be spent on purpose. Left alone, they get absorbed into whatever’s already filling the day.
A note wraps up five minutes early, so there’s time for one more portal message before the next room. A handoff between patients goes slightly less rushed. The schedule itself never changes shape. None of it gets logged, and none of it turns into an added slot or a dollar figure anyone could point to.
This is the norm, not a story about a few practices falling behind. That 68 percent figure describes most of the field, everyone who’s adopted the tool and stopped there.
Here’s a rough test worth running before deciding which group your own practice falls into. Pull up last month’s schedule next to one from before the tool went live, and control for the obvious: no new hire, no seasonal swing. If the number of patient slots per day still hasn’t moved, and nobody can point to a specific task that used to eat physician time and doesn’t anymore, the time is being absorbed, not captured.
Run the comparison and see which one you’re looking at. It might not be the answer you expected going in.
Somebody has to catch the time
A better AI tool was never the fix, and neither was more training on the one already running. Saved minutes and captured capacity are two different things, and turning one into the other doesn’t happen on its own just because the software works well. It also doesn’t happen from a better report or a smarter dashboard showing where the time went. A report that gets read is still just information sitting there until someone decides what to do with it.
No practice runs its staff schedule by installing software and walking away. Someone is specifically responsible for looking at the calendar every week and adjusting it, because a schedule left alone drifts into chaos. The minutes an AI tool frees up need exactly that same kind of ongoing attention, and right now almost nobody’s job includes it.
Picture the version that works. A physician’s notes get faster. Someone whose actual job is watching for this notices, within the first month, that Tuesday afternoons now run twenty minutes ahead of schedule most weeks, and adds a same-day sick visit slot instead of letting the day just end early. Nothing complicated. Just a person doing that specific work, every week, the same way someone already does it for the schedule itself.
That job exists in every practice that’s actually solved this. In most practices, it doesn’t exist yet.
This essay is cited in the KevinMD record on the electronic health record.
Karan Kanwar is the CEO of Central AI and Wing Assistant. Central AI provides an AI operating system for appointment-driven medical practices, handling calls, texts, web chat, and messages around the clock, qualifying patients and booking appointments directly into a practice’s existing systems. It is HIPAA compliant and designed to complement the tools practices already use.
Wing Assistant pairs that technology with skilled virtual assistants who take on the administrative work that pulls teams away from patient care. Kanwar writes on where automation and human support meet in health care operations, and how practice leaders can adopt both effectively. He shares updates on LinkedIn.
