A hospital already holds the information needed to find eligible patients: diagnoses, test results, medications, and visit history, recorded in electronic health records. The problem is that this information was recorded for care, not for research, so it sits in free text, inconsistent formats, and abbreviations that differ between clinics. A coordinator searching for eligible patients by hand can only look at a fraction of the records a hospital holds.
AI changes the scale of that search. It reads the records and flags the people whose recorded history appears to satisfy the eligibility criteria, producing a shortlist rather than a final answer. The shortlist is a set of candidates worth a closer look, not a set of confirmed participants. A human coordinator still reviews each flagged record, because the criteria involve clinical judgment that a record alone does not settle.
The value is in the direction of the search. Instead of starting from a clinic's full patient list and narrowing it by hand, the coordinator starts from a small group already filtered by the criteria. The same effort now covers far more records, which is what shortens the phase: not faster reading of each record, but fewer records that need to be read at all.