A medical chatbot nobody’s hospital approved quietly became the most-used AI tool in American medicine, and investors just priced it like a giant.
More than half of America’s doctors now lean on OpenEvidence, a free AI app built for physicians, to crack tough diagnoses. The reason it won where ChatGPT lagged: it trained only on peer-reviewed medical journals, not the open web. Last month doctors fired off 30 million questions, and the startup hit a $12 billion valuation.
Picture a packed ER at 3 a.m. A patient walks in with symptoms that don’t add up. Before the attending physician can even finish thinking out loud, two med students and a resident already have their phones out, typing the case into an app. That scene played out at Mount Sinai last summer, and the doctor watching it happen runs clinical innovation for the whole system. He had no idea how big this already was.
What Happened
OpenEvidence has become a genuine viral hit in a profession that usually treats new tech with suspicion. More than half of U.S. physicians are regular users. Last month they ran 30 million questions and consultations through it, nearly double the volume from just six months earlier.
A separate survey of 1,000 doctors found 45% using the app, roughly triple the share using ChatGPT, according to Offcall, a career service for physicians. That kind of traction sent the valuation flying: $12 billion in January, up from $3.5 billion the previous July. That is a 3.4x markup in six months.
The Backstory
Here’s the part most people missed. At Mount Sinai, a third of the system’s 9,000 doctors were already daily users before leadership even knew. Executives only found out in a meeting with the startup. “That was an ‘aha’ moment for our leadership,” said Dr. Nicholas Gavin, the system’s chief clinical innovation officer.
The doctor’s office has been a graveyard for “computer-assisted decision making” for decades. The first AI wave that actually stuck was boring but useful: scribe software that transcribes patient visits to fight burnout. OpenEvidence is riding the second wave, AI that helps with the actual diagnosis at the bedside.
The Plan
Founded in 2022, OpenEvidence made one contrarian bet: a smaller model trained on highly specialized, high-quality data could beat the giant general-purpose models in a field as information-rich as medicine. While OpenAI and Google trained on the messy open internet, OpenEvidence started with the government’s National Library of Medicine, then locked in content licensing deals with The New England Journal of Medicine, JAMA, and other peer-reviewed publishers.
The product is free to any government-verified U.S. physician. Ask it a question or feed it patient symptoms, and it returns likely diagnoses plus a “most important not to miss” list, each linked to the research behind it. A Mayo Clinic study found its answers generally accurate and evidence-based. It’s the same vertical-specialist logic that separates focused players from the general-purpose giants, a dynamic we break down in our OpenAI business model analysis.
The Business Model Angle
This is the part founders should tattoo somewhere. OpenEvidence ran three plays at once, and all three are repeatable.
Bottom-up, not top-down. It skipped the hospital IT department entirely. Doctors adopted it on their own, a pattern called “shadow AI.” By the time institutions noticed, the product was already load-bearing. You don’t sell to the gatekeeper. You make the gatekeeper irrelevant, then negotiate from strength.
Free, paid for by someone else. “We treated physicians like consumers,” said founder and CEO Daniel Nadler. During the five seconds users wait for an answer, they see ads, many from drug companies. Ads show on only 5% of questions, so the experience stays clean while pharma foots the bill. Classic consumer-internet economics, dropped into a B2B-feeling profession.
Proprietary data as the moat. Anyone can wrap a chatbot UI. Almost nobody can license the full NEJM and JAMA catalog. The licensing deals are the wall around the castle.
The Risk
Now the cold water. The research on AI in medicine is, in the experts’ word, “decidedly mixed.” AI has aced licensing exams and out-diagnosed humans on certain cases, but it has also botched research summaries and given flat-out wrong diagnostic answers.
Dr. Eric Topol of Scripps Research co-authored a paper bluntly titled “The Illusion of Readiness in Health A.I.,” which found significant competency gaps. “The potential for A.I. is great, but we’re not there yet,” he said. “It hasn’t really been tested and demonstrated in the messy, real world of medicine.”
The competition is also closing fast. UpToDate added a chatbot, Doximity bought a literature-mining startup, Abridge is bolting on decision support, and OpenAI just launched ChatGPT for Clinicians. And the bottom-up growth that made OpenEvidence cheap to scale is the same thing that spooks risk officers. Mount Sinai will now link to the app from patient records, but it is not handing over patient data. As Dr. Gavin put it, “we’re not going to just throw a patient’s data over the wall to a private company.”
Quick Questions
Is OpenEvidence actually free for doctors?
Yes. It’s a free, downloadable app for any government-verified U.S. physician. The money comes from ads, mostly pharma, shown during the brief wait for an answer.
How is it different from just using ChatGPT?
It trained only on peer-reviewed medical sources like NEJM and JAMA, not the open web, and every answer links back to the research. That focus is why nearly triple the doctors use it over ChatGPT.
Can it replace my doctor?
No, and that’s the company’s own line. “It’s not an oracle, it’s a tool,” says the CEO. It helps doctors think, but experts agree it isn’t ready to fly solo in real-world medicine.
Why is a free app worth $12 billion?
Distribution. More than half of U.S. physicians use it, 30 million questions a month flow through it, and that captive professional audience is extremely valuable to advertisers.
The Bottom Line
OpenEvidence didn’t win by building the smartest AI. It won by being the most focused, the easiest to start using, and the cheapest to adopt (free), then letting doctors pull it inside the system on their own. For founders, the lesson is sharp: in a market guarded by gatekeepers, the fastest way in is to make your product so useful that the end users smuggle it through the door for you. Just don’t confuse viral adoption with proven reliability, because in life-or-death markets, the second one is the only one that lasts.
