How OpenEvidence Became the Clinical OS for 40% of U.S. Doctors

Daniel Nadler opens with a qualifier. Across AI, adoption cycles that used to take half a decade or a decade are compressing into a year or two, in coding as in other knowledge work. The same compression hit OpenEvidence. In about 18 months, he says, it became the operating system for clinical knowledge in the United States. It is used something like 20 times more than the next platform of any kind in its specific segment: high-stakes clinical decision support for doctors. On a typical day, about 40 percent of U.S. doctors open it.

That segment is not paperwork and not scribing. Those are part of a doctor’s workflow, but the consequences differ. If you get a note wrong, you can go back and do it again. With a patient, you have one shot. Clinical decision-making is, in his view, unquestionably the highest-stakes area of medicine, and OpenEvidence is probably the only company working at the tip of that spear. Most firms have selected themselves out of the problem because, through an AI lens, they view it as too ambitious.

Nadler did not set out to build a healthcare company at all. Sequoia’s line, which he loves, is that OpenEvidence is a consumer internet company masquerading as a healthcare company. He wanted to treat knowledge workers as consumers, something he says had never been done. His first company was Kensho, a name he took from Japanese Zen. The rest of the episode splits between how medical AI actually breaks through and where a founder finds fuel — and neither answer sounds like a pitch-deck default.

At foundation, Nadler says, OpenEvidence is a search problem — a very semantic one. Flights to Barcelona collapse into a couple of keywords. A clinical question usually takes many sentences. His running example is a 44-year-old woman with moderate-to-severe psoriasis. A dermatologist would otherwise reach for a cream from a television ad, except the patient also has multiple sclerosis. The doctor is not a neurologist and does not want to refer in circles, because if everyone keeps referring, medicine never happens.

From the ether, the dermatologist may have heard that the new psoriasis biologics — IL-17 inhibitors and IL-23 inhibitors — might interact with the neurological side of a patient’s condition. That is about all they know. They did not learn it in medical school: IL-23s were FDA-approved in 2019, and the physician may have graduated in 2005. Nadler’s theme is that the golden age of biotechnology is the dark age of physician burnout, because it is impossible to keep up with the new drugs and mechanisms of action. The question is not academic. IL-17 inhibitors make MS worse. IL-23 inhibitors are safe and well tolerated. Ten years ago the likely paths were a referral loop or a 50-50 guess that might aggravate the disease.

Medical error is often repeated as the third leading cause of death in the United States after heart disease and cancer. Nadler thinks that statistic understates the damage, because it only counts death. The MS patient will not die from the wrong inhibitor; she will relapse. For every death from medical error, he estimates a factor of 10 to 100 people had a comorbidity or condition that got worse. The whole clinical string is the query. You cannot type “IL-17” into a traditional search box, and the physician does not have time to read book chapters.

What you need, he says, is a semantic understanding of the query the way another human physician would understand it. After that, the rest is fairly deterministic. From published biomedical literature you find the exact snippets in a phase 3 randomized controlled trial in the *New England Journal of Medicine* that tested each class. The answer may not even be in the abstract; it may sit in the methodology or the study population. Connecting a paragraph of clinical meaning to those snippets, across a surface of 35 million biomedical publications, is where the juice is.

The first way OpenEvidence deals with medical ambiguity, Nadler says, is to ensure the users are physicians, not patients. The team has talked about opening the product to patients since inception and so far has not. Builders have an enormous luxury when the MD is attached to the user’s name. Doctors need to protect that credential, so they use the product as a tool the way a Wall Street trader uses a Bloomberg terminal. If the terminal produced a bond quote off by an order of magnitude, the trader would frown and correct it, because the trade is theirs. Where evidence conflicts, OpenEvidence indicates the conflict. It does not pretend to present a sealed answer.

Citations were a first-class citizen six to nine months before ChatGPT started doing the same, and Nadler credits that for early adoption: people could interrogate and audit the source. The company never framed itself as an answer engine. It framed itself as a search engine, on a long continuum with Google — OpenEvidence is a Google portfolio company, and he views technology as progression rather than a net-new species. The social contract is that a paragraph-long case comes back as the three to five canonical phase 3 trials or guidelines that are responsive, including conflicting trials in *NEJM* and *JAMA*. Going out to read those papers is, he says, almost default behavior. The product is one of the largest sources of referral traffic to the *New England Journal of Medicine* after Google.

He wanted to build a consumer internet company for knowledge workers, not a healthcare company. People have a reductive picture of a consumer as a 14-year-old on TikTok. Traders, lawyers, and doctors are consumers too, and no one had treated doctors that way. They were appendages of health systems, not even gatekeepers of the technology they used — a realization he thinks begins to explain why American healthcare is a rare bipartisan consensus of dysfunction. OpenEvidence let them go to the App Store, download a free app, and start. It sounds stupidly simple. It was profound because it had never been said that way before, like the sentence in a stuck relationship that finally hits different.

Twenty years ago this would not have been possible. Today virtually every American doctor walks around with a computer they own, an iPhone or an Android. Senior leadership at UCSF, Massachusetts General, Mayo Clinic, Cleveland Clinic, NewYork-Presbyterian, Mount Sinai, and Cedars-Sinai — chief medical officers, chief physicians, CEOs in many cases — are personally avid users. Physicians also know the corpus. This is not training on tweets. It is trained on *NEJM*, *JAMA*, and other gold standards, with those partnerships in view. Gold in, gold out; garbage in, garbage out. The stack is a Bloomberg terminal for doctors: the right data, the right social contract, the right audience.

Elad Gil, who years ago founded a digital health company that delivered genetic information with a physician in the loop, describes a journalistic attitude in medicine toward what patients should see. Some of it was real worry about how people would act on the information. A lot of it, he thought, was wanting to remain the gatekeeper, or simply not wanting the questions. Nadler says he has been on both sides. As a patient he is sympathetic: medicine is not perfect, or everyone would live to 80 or 90, so patients should have a role and agency. What OpenEvidence has done is encourage physicians to generate patient handouts, a widely used secondary feature alongside prior-authorization letters and insurance appeal letters.

As a builder he is also sympathetic to the other fear. It took a graduate-level statistics course at Harvard before he really understood clinical trials. A patient who hears a trial mentioned on CNN or Fox News, then tries to read the *New England Journal of Medicine* through fear or hope, is not necessarily going to make a constructive decision. There is no good answer, he says. You want to give patients the consensus that is clear, and the tools to make sure their physician is not missing something. You can also imagine the failed case: a family member arrives with printouts asking why mother is not on this drug.

The real answer might be technical. Mother has another comorbidity, and in the presence of that comorbidity the p-value on the drug’s efficacy is not statistically robust. The family’s first reaction is “what is a p-value,” and they will not stop there; they will get really upset. A physician who has, by definition, taken at least one graduate statistics course is then trying to explain p-values to a civilian in the exam room. Nadler does not think that is a constructive outcome. The working balance is to let doctors produce handouts especially where guideline-based medicine is concerned — the content that is already consensus — rather than to dump the primary literature on frightened families.

Nadler does not try to forecast the far future of clinic visits; he thinks we are inside something like an event horizon. On doctors in the loop he reaches for aviation. Planes have been able to land themselves for a long time. There is no serious debate that the technology is there, and there is no mass movement of passengers to get pilots out of cockpits. Humans are anthropologically tribal. We do not abstract trust well; we personify it. Remove the two seats and nobody wants that plane. Leave the seats empty and still nobody. Put mannequins in as visual surrogates and, in a poll, you might see the first uptick in willingness — still a minority. Doctors, he says, have been in the loop since the ancient Greeks, if not the Egyptians, and will be for the rest of our lifetimes.

Medical education is another matter. In 1950, the doubling time of medical knowledge measured by citations was every 50 years. Estimates in the *British Medical Journal* and *Nature* put it today at every 73 days. Nadler finds that methodology aggressive because it counts every publication. Internally the team tied one hand behind its back: look only at the top quartile by citations, and pretend physicians never need the bottom three-quarters. That still doubles every five years. Go more conservative still — a doctor reads only the top 10 percent in their own specialty, no interdisciplinary medicine at all — and you enter the realm of “doable,” except that physician would spend about nine hours a day reading and would never see patients.

So the old frame, medical school as a defined period and continuing education as a wink-wink extra, is going to invert. Continuing education becomes the majority, and that is already happening. World-class physicians tell him openly that 90 to 95 percent of what they practice they learned after graduating, and in most cases after residencies and fellowships. A 70-year-old told him the majority of what he practices today he learned in the last two years. Residencies are changing at Mayo, Cleveland Clinic, and UCSF: evidence-based medicine rather than only guideline-based, the “curbside consult” (turn around and ask another doctor), distributed hive-mind decision-making. All of it sounds obvious. In 1917 the tools were gauze and scissors. IL-17 and IL-23 inhibitors for psoriasis with a neurological comorbidity are, on a historical clock, the last five seconds.

The catch is money. An interdisciplinary huddle means three or four expensive specialists on one complex patient, and the country is not minting oncologists faster. AI is what fills the gap at the edges. OpenEvidence is used by physicians in every state, county, and zip code, including rural Alaska and southwestern Georgia. Awesome things with a subscription get liked, Nadler says; awesome things that are free get fan mail. An oncologist wrote from rural Georgia: he is one of two oncologists in a 50-mile radius, serving a 75 percent African-American community with a median household income of $43,000, and he uses OpenEvidence as his curbside consult — his panel of other minds. That, Nadler thinks, is the silver lining of medical AI in healthcare deserts.

Asked what consumers might do for preventative health, Nadler warns that the answer will not be popular. He is obsessed with Japan — he named Kensho from it, has been a dozen times, was there two months before the taping. There is no net-new list. Outside the big cities, in places like Kyoto or Hakone, people of all ages are walking. Seventy- and eighty-year-olds do 10,000 to 15,000 steps a day. They do not fetishize retirement: in the traditional culture he describes, a good life is inextricable from a life with purpose, and idleness and fulfillment are incompatible. People work past 65 into their 70s and 80s, which is when mortality risk actually rises. They eat until 70 or 80 percent full. Every doctor already believed that past a certain body weight, risk goes up; 10 or 15 years ago few wanted to say it aloud. He thinks the pendulum is swinging toward a more open conversation, including neurodegeneration: no serious neurologist disputes that continuing to use your brain is a mitigant, down to writing occasionally with the non-dominant hand to form new pathways.

For founders listening from other fields, he says human psychology is not as specific as medicine. Address people as people and as consumers, speak in a way that hits different, and you can break the mold that rate-limited adoption in a skeptical industry. He does not believe “if you build it, they will come,” and he does not think that is the Apple story either. Apple, on his reading, is Nietzschean: extraordinary will to power, reality as malleable, ideas and rational thought as second-order after projections of the will. That is not a fairy tale you tell Y Combinator kids. The Enlightenment’s irony, he says, is that the notion that reason is supreme was not arrived at through reason, but through faith. MBA interview questions about which coffee shop birthed the idea are Cartesian derivatives. The more useful question is where you can find motivation that is almost compulsive.

There is no single answer. Some people spend a life proving wrong a teacher who has been dead for 30 years; he has met 75-year-olds still at it. Others are born with enormous aggression and find a constructive channel. Nadler says he was born with an unbelievable amount of aggression and, through intellect and luck, found a more useful outlet. The idea for OpenEvidence is, in his mouth, the most obvious in the world — no more creative than “let’s go to the moon.” He does not hunt for more motivation. He resists the contemporary cult of psychoanalysis. In the analysis and description of something, you kill it. He resists exploring trauma and the origins of his aggression. Most propulsion systems originate from trauma; get too close, in the Sequoia style of talking through early childhood, and you can delete the engine that was already a deep well.

That lens walks into recruiting. He learned at his first company that the correlation between being smart and output is only about 0.65. He looks for people who are exceptionally intelligent and who have a propulsion system they may not understand. Not knowing why they are driven is better, not worse. He does not like management and does not want to practice it. Business-school lore about motivating people, inspiring them, and giving constructive criticism is a real body of knowledge, and you can do it better or worse. What he seeks are the people for whom all of that is redundant, because they are already on their own warpath. The best he can do is get out of their way.