How AI Trial Matching Actually Works (and What It Can’t Do)

"AI-powered matching" appears in nearly every recruitment product pitch. Underneath the buzzword is something genuinely useful — and more bounded than the marketing implies.

What matching systems do

At core, trial matching compares structured facts about a person — age, sex, diagnoses, medications, labs, location — against a protocol's inclusion and exclusion criteria, and produces a ranked candidate list. Modern systems add natural-language processing to extract criteria from protocol documents and to read unstructured notes, plus scoring models that weight partial matches.

The value is speed and recall. A coordinator can screen a 5,000-person database against a new protocol in seconds instead of days, and the algorithm won't forget that a 2023 screen-failure is a perfect fit for this year's study. Sites using systematic matching consistently surface candidates that manual review misses — not because the algorithm is smarter than the coordinator, but because it actually looks at every record.

What they can't do

Matching output is a shortlist, not an eligibility determination. Databases are incomplete and stale; the algorithm doesn't know about the diagnosis from last month or the medication change last week. Every match still needs human pre-screening.

Garbage in, garbage out. Matching against free-text notes and outdated records produces noise. The sites that get the most from matching are the ones with disciplined, structured data capture.

It doesn't create candidates. Matching mines the database you have. If your registry is thin, the best algorithm in the world returns a short list.

Evaluating vendors

Ask three questions: What data does the match actually use? How does it handle criteria that aren't in structured fields? And can coordinators see why someone matched? Explainable scores ("age ✓, diagnosis ✓, A1c unknown") keep humans usefully in the loop. Black-box percentages breed either blind trust or blind distrust — both expensive.