The EHR Was Just the Scaffolding: Bob Wachter on Why AI in Healthcare Is Different This Time

Bob Wachter has seen this movie before. As Chair of Medicine at UCSF, he watched the electronic health record arrive with enormous promise and deliver enormous frustration. Physicians became expensive data-entry clerks. Patients noticed their doctors staring at screens instead of at them. His 2015 book, The Digital Doctor, captured what went wrong.

So when he tells you AI in healthcare is going to be different, it’s worth paying attention—because he’s the guy who called out the last failure.

The Scaffolding Argument

Wachter’s reframe of the EHR era is the most useful idea in the conversation. The $30 billion HITECH investment wasn’t a failure, he argues. It was infrastructure. “The mistake I made was thinking the electronic health record was the end game, and not recognizing it was simply the scaffolding.”

The data is digital now. The workflows exist. The burning platform is there—nobody is happy with the status quo. What was missing was a technology capable of doing something intelligent with all that digitized information. On November 30, 2022, that technology showed up.

Watson’s Lesson, the Scribe’s Proof

IBM Watson won Jeopardy! and promptly lost $3 billion trying to crack healthcare. “All polish and no shoe,” Wachter says. They started on the hardest problems, marketed ahead of their evidence, and got demolished by clinician skepticism.

Ambient scribes took the opposite approach. The problem was already funded—health systems were paying $30-40/hour for pre-med students to type. The workflow was contained. Failure wouldn’t kill anyone. And within two years, “what turned out to be just a remarkable innovation is now like total commoditized thing.”

That arc—from Watson’s hubris to the scribe’s quiet success—is the playbook. Start boring. Win trust. Then go bigger.

The Uncomfortable Middle

A particularly striking metaphor was Wachter’s description of where we are right now as “the uncomfortable middle”: AI that’s good enough to be useful, not reliable enough to act alone.

“Humans stink at being vigilant when they’re overseeing the result of a generally trusted technology,” he told us. “They turn their brains off. We de-skill pretty quickly.”

This is the human-in-the-loop problem that the governance frameworks haven’t caught up with. We tell patients a doctor always reviews the AI output. But the evidence on human oversight of automated systems is sobering. We develop automation bias. We stop actually checking. The oversight may be providing false comfort.

The Expert-Novice Gap

Wachter uses AI daily in clinical practice—pulling out his phone to run a case by GPT or a physician-specific tool called Open Evidence. He calls it the “curbside consult model,” and he’s convinced it makes him a better doctor.

But he’s clear-eyed about who benefits. When he prompts an AI, he brings decades of pattern recognition. He knows which 7 facts out of 100 to include. He knows which answers to trust and which to discard. A patient has none of that scaffolding.

“A lay person has absolutely no ability to know how to correctly prompt the tool and to know how to analyze the results,” he said. The leap from “helps experts” to “helps patients” requires fundamentally different tool design—more guided conversation, less open prompt box.

The Bottom Line

Wachter’s call to action is simple: jump in. Try the tools. Form your own judgment.

“I think the risks of going too slow here are far greater than the risks of going too fast,” the Mayo Clinic CEO told him. Wachter agrees—not because AI is ready for everything, but because standing still means accepting a system that already harms nearly a million Americans a year.

The EHR was the scaffolding. The question now is what we build on it.

Listen to the full conversation with Bob Wachter on Practical AI in Healthcare.