Protocol prose in. A ranked, cited patient list out.

Most matching tools run a search when a study opens and never run it again. Patients become eligible after that search — a biopsy comes back, a therapy fails, a lab crosses a threshold — and nobody notices, because nothing is looking.

Accrual's models re-screen as records change and as protocols open, close and amend. A patient who becomes eligible on a Tuesday surfaces on Tuesday, not at the next feasibility review. Running an AI read continuously is only affordable because it needs no human in the loop until there is something worth showing.

Structured fields carry diagnosis codes and medication lists. They do not carry ECOG status, lesion measurability, prior-therapy sequence, or the sentence in a pathology report that decides the whole question. This is why two decades of rule-based matching largely failed.

Accrual applies language models to clinical notes, pathology and imaging reports, discharge summaries and lab trajectories alongside the structured record, because criteria are written against all of it. The model extracts the criterion, finds the evidence, and cites where it found it.

Research portals fail because they ask a busy clinician to go somewhere else. The one moment a trial conversation can actually happen is when the patient is in front of them.

Accrual surfaces the AI's candidates inside the workflow the clinician already uses, with the cited evidence for each criterion in view and a one-click path to the coordinator. The clinician is confirming a shortlist, not running a search.

Site feasibility questionnaires ask a site to estimate its own population from memory. The answers are optimistic, unfalsifiable, and the reason a third of sites enrol nobody.

Because the AI has already read the population to find matches, the sponsor-side answer falls out of the same computation: how many candidates exist, how fast they can realistically be worked through, and what competing studies are drawing on the same pool.