The most valuable clinical AI company left Europe. Nobody made it leave.

One framing correction before anything else, because precision is the point of this blog: OpenEvidence was not kicked out of Europe. No regulator ordered it off the market, no enforcement action names it, and no ban exists. In late April 2026 the company terminated access in the European Union and the United Kingdom on its own, citing "mounting regulatory uncertainty regarding the treatment of AI systems in the European Union and the United Kingdom, including, among other rules, the EU Artificial Intelligence Act." A voluntary exit under regulatory uncertainty and an expulsion are different events with different lessons, and the real one is more interesting.

What left, and how big it was

OpenEvidence is the clinical AI search and decision-support platform that has become, by the numbers it reports, the default AI tool of American medicine: used by roughly 40% of US physicians across more than 10,000 hospitals and medical centers, handling millions of clinical consultations a month, with content partnerships including the New England Journal of Medicine and the American Medical Association. In January 2026 it raised a $250 million Series D at a $12 billion valuation, which made it the most valuable healthcare AI startup on record. The figures on usage are company-reported; the funding round is documented.

This spring, EU and UK clinicians who had adopted it received notice that access was ending. Reports place the termination in the last days of April 2026. As of this writing the platform remains unavailable in both markets, its own privacy policy warns EU users that EU safeguards do not govern the service, and no return timeline has been announced. The exit included the UK even though the UK is not subject to the EU AI Act, which tells you the company was pricing the whole regulatory neighborhood, not one statute.

The uncertainty it named was real, and then it moved

Read the company's stated reason against the calendar and the irony is sharp. In April 2026, when OpenEvidence pulled access, the EU AI Act's high-risk timeline was genuinely unsettled: the Commission had proposed making key compliance dates contingent on readiness milestones, meaning a company could not put the date its obligations began on a calendar. Weeks later, the May 7 political agreement on the Digital Omnibus replaced that mechanism with fixed dates, and the July regulation made them law: December 2, 2027 for standalone high-risk systems, August 2, 2028 for product-embedded ones. We covered the chronology in Three dates moved in June.

So the specific uncertainty OpenEvidence cited was largely resolved, in the direction of more time, within about ten weeks of its exit. Whether that changes the company's calculus is unknown; nothing public says so. But the sequence is a case study in what regulatory uncertainty actually costs. The obligations themselves did not push OpenEvidence out; by the time they bind, in 2027 and 2028, the company could have prepared twice over. What it declined to carry was the interim: operating a clinical tool in a market where the classification, the timeline, and the enforcement posture were all in motion at once. Forty percent of US physicians is a business worth protecting from that variance. Two overseas markets, apparently, were not.

The asymmetry that should bother you

Here is the part that matters for anyone responsible for a regulated product, and it is not about OpenEvidence's compliance judgment.

OpenEvidence is a clinical decision-support tool. It is built for clinicians, grounded in medical literature, and positioned inside the category of AI that the EU framework treats as high-risk. It is, in other words, exactly the kind of system the rules were written to reach, run by a company with the resources to comply, and it responded to the rules by leaving.

Meanwhile, the general-purpose engines never left. ChatGPT, Gemini, Copilot, Perplexity, and Google's AI-augmented search kept answering medication questions from EU and UK patients and clinicians through April, through the Omnibus, and through today. They are not classified as clinical decision support. They carry no medical-device obligations for those answers. The peer-reviewed evidence on how well general assistants answer drug-information questions is not encouraging; we walked through the AJHP study's numbers, including the 19% full-accuracy finding, in Citation share is not accuracy.

The net effect of the spring, in two European markets: the purpose-built, literature-grounded clinical tool is gone, and the general-purpose engines that answer the same questions with documented incompleteness remain, untouched, because the regulatory perimeter was drawn around tools that claim a clinical purpose rather than around answers that have a clinical effect. A physician in Manchester who lost OpenEvidence in April did not stop having questions. Some fraction of those questions is now going to whatever answer box is still available.

What this means if you make the products being asked about

For a pharmaceutical or device company, the OpenEvidence exit changes the composition of the answer surface in the EU and UK, and not in a comforting direction. Clinician queries that were being answered by a system grounded in NEJM content are now, at some rate no one measures, being answered by consumer engines whose drug-information answers are, on the published evidence, mostly incomplete. That shift happened without any change to any engine and without any input from any manufacturer. A market exit by a third party you have no relationship with just altered what gets said about your product to prescribers in two major markets.

There is no lever to pull on any of this. OpenEvidence's return is its own decision; the engines' availability is theirs; the regulatory perimeter belongs to legislators. What a manufacturer can do is see the surface move: capture what the remaining engines in those markets say about your products, score it against the label, date it, and re-measure, so that when the composition of the answer surface shifts again, and it will, the change shows up in your record rather than in your surprises. The companies that monitor this surface will be the ones who can say what changed in May. Everyone else will be told, eventually, by someone else.

Related: the full June record of regulatory and engine movement is in the June 2026 roundup.