OpenEvidence AI Goes Live in Northwestern Medicine's Epic EHR

What happened

Northwestern Medicine has made OpenEvidence, the AI medical search and decision-support tool, available inside its Epic electronic health record across the whole health system. The two organizations announced the deal on October 12, 2026, and called the rollout the first phase of a wider collaboration on using AI safely at the bedside.12 Since October 8, Northwestern clinicians have been able to open OpenEvidence from inside Epic, and their existing hospital credentials sign them in automatically.15

The bigger change is who gets access. For the first time at Northwestern, nurses, pharmacists and other clinical staff can use the tool, not just physicians. The health system says the care team can use it to check drug dosing and interactions, look up current guidelines and get cited answers drawn from peer-reviewed sources, including the New England Journal of Medicine and the JAMA Network.15 Northwestern runs 11 hospitals and more than 200 diagnostic and outpatient sites, so the rollout covers a large workforce.4

Northwestern's chief medical officer, James Adams, said the health system judges new technology by two tests: does it meet a real clinical need, and can it scale across the system. He described the integration as giving clinicians one trusted source of evidence at the point of care.2 Vivien Ho, OpenEvidence's senior vice president of strategic partnerships, said that opening the tool to the full care team shows where modern hospitals are heading.4

From bookmark to workflow

The coverage agrees on the facts, which is expected because most outlets worked from the same press release.356 The analysis is in the framing. Healthcare IT News placed Northwestern in a growing group of Epic customers with OpenEvidence integrations, naming Sutter Health in California and Mount Sinai in New York.3 Fierce Healthcare's list of enterprise deals goes further, adding Cedars-Sinai, Penn Medicine, NewYork-Presbyterian and AdventHealth.10

The pattern matters more than any single announcement. OpenEvidence grew as a free, ad-supported app that individual doctors found themselves, mostly on their phones.2426 Getting embedded in the EHR moves it from a personal habit to tooling the institution has approved, with single sign-on, a defined user base and, in principle, oversight by the health system. In practice, hospitals are taking an AI behavior that clinicians had already adopted on their own and bringing it under institutional control.

The Northwestern deal arrived during a run of OpenEvidence partnership announcements. The company's announcements page lists recent deals with Memorial Sloan Kettering, Penn Medicine, Boston Children's Hospital and the American Academy of Pediatrics.9 One day after the Northwestern news, Johns Hopkins Medicine announced a partnership focused on medical education and on building AI that combines evidence with a patient's own history.10 Sara Murray, Hopkins' chief medical informatics and AI officer, said the goal is to move past standalone medical information tools toward AI built into how care is delivered.10 Northwestern's integration, which launches the tool from Epic rather than reading the patient chart, looks like an early step on the same path.

The money and the usage figures

The enterprise push is backed by a lot of capital, though the exact numbers vary by source. CNBC reported in January that OpenEvidence raised $250 million at a $12 billion valuation, with Thrive Capital and DST leading.21 Business Insider reported in late September that a further $250 million round, from hospital systems and Andreessen Horowitz, valued the company at $15 billion.23 Axios and Fierce repeated that figure, and Fierce named Byers Capital as co-lead.2810 Some coverage noted that the company has not officially confirmed the round's details.30

Usage claims vary even more. The company told CNBC in January that 40% of U.S. physicians use the product.21 The Northwestern press release says most U.S. physicians use it daily on average.2 One industry analysis puts reach at about 65% of U.S. physicians as of April 2026.35 Fierce reported that, as of September, 1.12 million license-verified U.S. clinicians used the platform, a group that includes nurses, nurse practitioners and physician assistants.10 These figures measure different things in different ways. Whichever is closest, the direction is clear: adoption is wide, and hospitals are reacting to it more than driving it.

The business model is worth keeping in mind as health systems bring the tool in-house. Several analyses say most of the company's revenue comes from pharmaceutical advertising shown to clinicians.2438 Annualized revenue was reported at about $300 million by mid-2026.2529 None of the Northwestern coverage explains how ads work, or whether they appear at all, in the Epic-embedded version. Other health systems will likely want an answer before they sign similar deals.

The evidence gap

The gap between how widely the tool is used and how well it has been studied is wide. A systematic review published in npj Digital Medicine found that OpenEvidence generally produced relevant, evidence-supported answers and fabricated citations less often than general-purpose chatbots. It also found that performance dropped in complex cases and that the tool more often confirmed doctors' existing plans than changed them.40 An earlier version of that review counted only 11 studies, which it called thin given how much the tool is used.37

The independent results vary widely. A small primary-care study rated OpenEvidence highly for clarity and relevance but low for its effect on decisions, at 1.95 out of 4.33 A December 2025 preprint tested it on difficult subspecialty board-style questions and found peak accuracy of 34% in standard mode and 41% in Deep Consult mode, in line with general chatbots on the same dataset.31 A structural heart disease comparison found OpenEvidence fully accurate on 26.7% of questions, against 66.7% for ChatGPT-4o.37 On the other side, an osteoarticular-infection benchmark put it at 94.4%.37 A company-supported evaluation using real questions from the platform found it ahead of frontier models, which shows how much test design shapes the result.35

The main lesson from this research is that accurate citations do not guarantee accurate answers. Reviewers documented cases where the tool cited a real paper but summarized it incorrectly.3437 NBC News found that many doctors click through to the cited studies only when an answer surprises them.36 That habit is the weak spot when the tool reaches more users. Northwestern specifically highlights dosing and drug-interaction checks for pharmacists and nurses, so how carefully those users verify answers will matter. One independent review said in August that OpenEvidence did not offer dedicated drug-dosing tools, which is in tension with the dosing checks Northwestern now promotes.32 Readers should treat that point as unresolved, not settled.

Why the FDA is not reviewing this

OpenEvidence can reach clinicians at enterprise scale without FDA premarket review because of how the agency defines clinical decision support. Under the 21st Century Cures Act, software aimed at clinicians is not a medical device if it meets four conditions. It must not analyze medical images or diagnostic signals. It must display or analyze medical information, such as studies or guidelines. It must support a clinician's recommendations rather than make decisions. And it must let the clinician independently review the basis for its output.12 A tool that brings up guideline text and literature for a clinician to weigh usually falls outside device regulation.12

The FDA loosened this further in January 2026. It revised its guidance to use enforcement discretion for software that gives a single recommendation when only one option is clinically appropriate.1114 The guidance was issued on January 6 and reissued on January 29.1719 Critics note that it says almost nothing about generative AI, including clinician-facing medical chatbots built on large language models.1113 Lawyers have advised developers of tools built on these models to make sure clinicians can actually trace where a recommendation came from.14 OpenEvidence's emphasis on citations fits that requirement well, though the summarization errors researchers have found test it.

The radiology contrast

The imaging boundary shows how the regulatory split works. Software that reads pixels is a regulated device without exception. The American College of Radiology repeated after the 2026 update that image-analysis tools remain under FDA oversight.13 Of the 1,524 AI-enabled devices the FDA had authorized by the end of March 2026, about 76% were radiology products.13 As a result, the most heavily used clinical AI in American hospitals, a literature-grounded answer engine, has a lighter regulatory path than a narrow chest X-ray detection algorithm.

That may change. In August the FDA published a discussion paper on regulating generative AI devices, sketching a risk map and more postmarket monitoring. Comments are due October 19, and the paper binds no one yet.13 For now, responsibility for oversight sits with hospitals. Northwestern's statement that this is the "first phase" of a safety-focused collaboration will be judged by whether it later publishes how the tool performs in practice. In this market, that kind of evidence is scarce.