The AI tool physicians actually love

Robert Wachter, M.D., chair of the Department of Medicine at the University of California, San Francisco (UCSF), has been thinking about technology and medicine for a long time. He coined the term "pajama time" to describe the hours physicians spend documenting after clinic, and he wrote about the promises and failures of the electronic health record (EHR) in his 2015 New York Times bestseller, "The Digital Doctor."

His newest book, “A Giant Leap: How AI is Transforming Healthcare and What That Means for Our Future,” published earlier this year, takes on artificial intelligence (AI) in health care — and he came out of writing it, he said, pretty optimistic.

Medical Economics sat down with Wachter to talk through what AI scribes are actually delivering, where they fall short and why he thinks their biggest contribution may not be the time they save at all.

Related content: Take note: The AI scribe era is here

Can AI scribes actually fix the pajama time problem?

Wachter was careful not to oversell the documentation benefit. AI scribes address the part of "pajama time" that comes from visit documentation, he said, but the inbox, and the volume of patient messages and portal communication that now fill evening hours, is a separate problem that will need separate solutions. The more important shift, in his view, is what scribes do for the visit itself.

The EHR changed the encounter in ways that went beyond documentation. It created an audience — Medicare, insurers, malpractice attorneys, quality measure administrators — all of whom suddenly had visibility into the clinical note and used that visibility to demand more from it. The note bloated. Physicians became, as Wachter put it, “pretty expensive grumpy data entry clerks." AI scribes have the potential to lift much of that burden and give physicians something back: eye contact, presence, the ability to actually listen.

"It's the first technology we've seen that says to physicians, we can give you back some of not only your time, but of the humanity of the doctor-patient visit," he said.

The time savings themselves have been more modest than early projections suggested. Studies, including a recent paper out of UCSF, published in JAMA Network Open by lead author AJ Holmgren, Ph.D., M.H.I., examining time savings and return on investment (ROI) specifically for AI scribes, suggest the tool roughly pays for itself in throughput terms — but doesn't deliver the five-to-ten minutes per visit that some early ROI calculations assumed. Much of the time saved from documentation gets repurposed into genuine patient interaction rather than additional appointments. Wachter said that's probably the right trade-off, even if it complicates the financial case.

Why is documentation the right first use case for AI in medicine?

Wachter framed AI scribes as deliberate "training wheels" — a first use case chosen not because documentation is the most clinically important problem AI can solve, but because it sits in the right part of a two-by-two matrix: relatively easy to do well, and unlikely to harm anyone if something goes wrong.

He traced the comparison to AI's first wave in health care, in the late 1970s and early 1980s, when the field's founders chose diagnosis as their entry point. The technology wasn't ready, the infrastructure wasn't there, and starting with the hardest and highest-stakes problem meant that when things went wrong, as they inevitably did, it damaged trust in AI broadly.

"You don't start with the hardest problem," Wachter said. "You start with easier, more tractable problems. You gain experience and you gain trust."

Scribes fit the bill. UCSF already had a model for non-physician documentation through human scribes, mostly pre-medical students, who had been doing the job for about a decade. The AI version had a clear predecessor, a clear problem to solve and a clear ceiling on how badly a mistake could hurt someone. The note might have an error. That's not nothing, but it's recoverable.

At UCSF, about 70% of the department's roughly 3,000 to 4,000 physicians now use an AI scribe, and about 90% of those say they love it.

"We've almost gotten to the point where, if we turned it off, we might lose a fair number of doctors," Wachter said.

What are the real drawbacks of AI scribes?

Wachter was candid about the limits. The tools don't yet adapt well enough to subspecialty documentation styles. Physical exam narration, something physicians have never had to do before, is awkward and remains a friction point. Some physicians tried the tools, found the error rate uncomfortable and stepped back.

He also raised a subtler concern: something may be lost when a physician stops writing the note. The act of typing is cognitively active in a way that reviewing a draft is not. Physicians sometimes caught things during documentation — a forgotten question, a connection between symptoms — that they might not catch while reading over an AI-generated summary.

"I think there's a little lost when I'm not typing my own note," he said, acknowledging it's hard to quantify.

The "human in the loop" framing, which is widely used to reassure skeptics about AI safety, also deserves more scrutiny than it usually gets. If a physician has reviewed 49 perfect notes in a row, the attention brought to note number 50 is not the same. Over-trust is a real risk, and health systems haven't yet developed robust systems to counteract it — though Wachter said conversations are beginning about approaches like embedding intentional errors periodically to keep reviewers alert, similar to phishing tests in cybersecurity.

Are AI scribes worth it for physician practices?

The financial ROI for AI scribes today rests less on throughput than on physician experience. A recently published paper in JAMA Internal Medicine, with lead author Lisa Rotenstein, M.D., MBA, M.Sc., examines how to think about ROI for tools like scribes, and Wachter said the honest answer is that the return shows up in recruitment, retention and "joy in practice," all of which carry real but harder-to-quantify economic value.

Replacing a primary care physician is not cheap. A tool that makes physicians meaningfully happier and easier to retain changes the math considerably.

But Wachter's bigger argument is strategic. AI scribes are "singles," he said. The home run hitter is still on-deck, with robust clinical decision support that helps with diagnosis, test selection and treatment. Getting there requires physician trust, and AI scribes are building it.

"The biggest benefit of scribes, when that story is finally written, will be winning hearts and minds and creating receptivity for bigger, broader and ultimately more impactful tools," he said.

Listen to our full deep-dive episode of Off the Chart: A Business of Medicine Podcast, on the state of ambient AI scribing technology: