AI-driven site feasibility can improve how sponsors shortlist clinical trial sites, but it cannot replace a conversation with the people who will actually run the study. Predictive models are strong at pattern recognition across historical performance data. They are weak at detecting what changed last month.
Site selection has always been one of the most consequential and least scientific decisions in clinical operations. A poor site list quietly drives enrollment delays, rescue sites and budget overruns. So the appeal of prediction is obvious.
What AI does well in feasibility
Machine learning models trained on historical enrollment, startup timelines, protocol deviation rates and real-world patient data can surface sites that a manual process would miss. They can also flag sites with a pattern of over-promising on feasibility questionnaires and under-delivering on enrollment. That alone is worth a lot.
AI is also useful for mapping where eligible patients actually are. Combining claims data, EHR-derived populations and epidemiology can show whether a region has the patient density a protocol requires, rather than relying on a site's optimistic estimate.
What can AI not see in site selection?
Historical data describes the past. A site's star coordinator may have left. A competing trial in the same indication may have just opened down the hall. The principal investigator may have enthusiasm for your mechanism of action that no dataset captures. Or the site may be drowning in startup paperwork for three other studies.
Predictive models also inherit the biases of their training data. Sites that have historically received more industry trials will look better on paper. Community sites and sites serving underrepresented populations often have thinner data histories, so models may undervalue them. That matters for enrollment diversity and for FDA's growing expectations on diversity action plans.
A blended model works best
The most effective approach I have seen uses AI to widen and rank the candidate pool, then applies targeted human validation to the top tier. That validation includes a direct conversation with the site's study coordinator, not just the PI. Ask about current workload, competing studies, staff turnover and realistic screen-failure expectations for your specific criteria.
Sponsors should also ask their CRO how its feasibility model works. Which data sources feed it? How recent is the data? How has the model performed on past studies in your therapeutic area? A CRO that cannot answer is offering a ranking, not a prediction.
Measure the model, not just the sites
If you use predictive feasibility, track it. Compare predicted enrollment against actual enrollment at study close. Over two or three studies you will learn whether the tool adds real signal for your programs. That evidence also strengthens your hand in future CRO negotiations.
AI makes feasibility faster and broader. Judgment still makes it right.
