Automated trial matching has been a persistent target for clinical AI, with an obvious value proposition: trials struggle to enroll, patients who would qualify are never approached, and the matching work is tedious. Progress has been real and lopsided.

Parsing eligibility criteria into structured logic is largely solved. Current systems handle the notoriously baroque phrasing of oncology protocols with high fidelity. Determining whether a specific patient satisfies that logic is where systems fail, because the required facts — performance status, prior lines of therapy, specific mutation results — are frequently in unstructured text, out of date, or absent.

Reported precision in deployed matching systems is respectable. Reported recall is poor, and the missed patients are disproportionately those with care fragmented across institutions, which is a familiar equity pattern.

The centers reporting the best results have not deployed better models. They have invested in structured capture of the handful of variables that dominate eligibility, which is a data governance project that happens to enable a matching product.