The Quietest AI Win in Healthcare Is Matching Patients to Trials

After a few weeks of AI stories that made me wince, I want to tell you about one that made me smile. No chatbot doctor, no moonshot, no press release promising to reinvent medicine. Just a genuinely hard, genuinely boring data problem, solved in a way that helps real cancer patients. This is the kind of AI win I wish got more attention, precisely because it is so undramatic. It works, it is measurable, and it keeps a human firmly in charge.

The 5 percent problem

Start with a number that should bother you. Only about 5 percent of eligible cancer patients nationwide actually make it into a clinical trial. Not 5 percent of all patients, 5 percent of the ones who qualify. That is a staggering amount of missed opportunity, both for the patients who could benefit and for the research that needs them. And a big part of the reason is embarrassingly mundane: figuring out whether one patient fits one trial can mean checking more than 1,700 eligibility criteria, which takes up to an hour of a busy clinician time. Multiply that across a cancer center and most eligible patients simply never get screened.

Why matching is a data problem, not a magic problem

This is what makes it such a good fit for AI. Checking 1,700 structured criteria against a chart is exactly the kind of tedious, rule-bound work software is built for and humans are bad at sustaining. Advocate Health partnered with the startup Lind to embed an AI screener right in the electronic health record, where it reads the data that is already there and does the matching automatically. The patient information does not have to move, the clinician does not have to re-enter anything, and the hour of screening collapses into minutes. It is not magic. It is a well-scoped data problem finally getting the right tool.

Human-plus-AI, embedded in the EHR

Two design choices are doing the heavy lifting here, and both are worth stealing. First, the AI lives inside the EHR instead of in a separate portal that nobody remembers to open. Tools that live where the work already happens get used. Tools that require a detour do not. Second, it is explicitly a human-plus-AI model: the software screens and shortlists, but a clinician reviews and confirms before any patient is enrolled. Nobody gets auto-signed into a cancer trial by a model. The AI does the tedious part, the human keeps the judgment, and that division is the whole reason this is safe to run at scale.

The results, and what they do not yet prove

The early numbers are strong. In a retrospective study the tool hit more than 94 percent agreement with clinical reviewers across those 1,700-plus criteria, and it could nearly double screening capacity. The scale is real too: in 2025 Advocate screened nearly a million patients across more than 130 cancer clinics and enrolled over 14,000, running more than 1,500 trials at 50-plus sites. Now the honest caveat, because I promised you one. Agreement with reviewers on old cases is encouraging, but it is not the same as proving better prospective enrollment and outcomes. Doubling screening only helps if trial slots exist to fill, and it genuinely matters who gets screened. Watch the enrollment and equity data before you call it a home run.

Why this is the template

If you want a checklist for clinical AI that actually lands, this is it. Pick a real, structured data problem that is tedious for people and natural for software. Embed the tool where the work already happens, not in yet another login. Keep a human on the decision so the AI assists rather than decides. And measure yourself against expert reviewers with a number you can defend, like 94 percent agreement, not a demo that felt impressive. None of it is flashy. That is the point. The unglamorous stuff is where AI is quietly earning trust in medicine, one matched patient at a time.