Finding patients for an oncology trial is the hardest part of running one. A CRO feasibility lead and a principal investigator are screening a population against criteria that are deliberately tight: line of therapy, biomarker and mutation status, ECOG performance, prior treatments, washout periods, and whether the disease has actually progressed. Almost none of that lives in a clean field. It lives in unstructured progress notes and pathology reports, where a single ambiguous sentence about progression can move a patient on or off the list. The work is slow, the eligible population is small, and a missed match is a patient who never gets the therapy and an enrollment target that slips.
So the pull toward an AI agent is obvious, and reading messy clinical text is exactly what a model is good at. The question is not whether it can screen faster than a coordinator. It plainly can. The question is which of the investigator's actions an agent may perform, and which it may only propose, because in cohort identification one specific action carries the whole risk.
What an agent should do: screen and propose with evidence
Most of pre-screening is comparison and citation, the kind of high-volume reading a person is genuinely tired of. Given a protocol and a population, an agent can:
- Read line of therapy, biomarker and mutation status, ECOG, prior treatments, washout, and disease progression out of unstructured notes and path reports.
- Match each patient against the inclusion and exclusion criteria, item by item, and lay out which criteria are met and which are not.
- Cite the supporting evidence for every match, the sentence in the note, the result in the path report, so a human can check the agent's reading against the source.
- Separate the clean candidates from the borderline ones, surfacing the ambiguous cases instead of quietly resolving them.
This is throughput, and none of it requires the agent to decide anything that advances a patient. The agent screens the population and proposes candidates with the evidence attached. It does not conclude that a patient is eligible.
What stays human: the eligibility determination
One action in cohort identification is not clerical: attesting that a patient meets the inclusion and exclusion criteria. That is an eligibility determination, and it advances a human being toward enrollment in an interventional trial.
Get it wrong in the permissive direction and the consequence is not an internal note. A wrong inclusion is a protocol deviation, an ICH-GCP finding, and a data-integrity and patient-safety risk. It can put a patient on a therapy the protocol was written to keep them off, and it contaminates the dataset the trial exists to produce. ICH-GCP treats inclusion and exclusion eligibility as an investigator judgement for exactly this reason. It is not a determination a model that cannot be held accountable, and cannot explain itself to a monitor in the monitor's terms, should be the actor of record for.
The line is clean. The agent owns the screening. A named investigator owns the determination. The system has to make that line structural, not a matter of good intentions, which is the job of an AI agent control plane.
The control that makes the split structural
"An investigator reviews the agent's list" is not a control. Anyone can glance at a proposed match and mark a patient eligible; the click proves nothing about who actually judged it, or whether the same automated actor both screened the patient and confirmed the inclusion.
The control has to guarantee, structurally, that the agent which screened a patient cannot be the actor that marks that patient eligible, and that the investigator who does make the determination leaves a signature carrying its meaning. This is the maker-checker principle, the four-eye separation good clinical practice expects, enforced here against a machine. The agent is barred in code from advancing a patient to screening on its own. Marking a patient eligible is a high-risk skill, and a high-risk skill forces an approval gate by construction.
| Step | Actor | Control |
|---|---|---|
| Reading criteria out of unstructured notes | Agent | Deny-by-default, versioned skill grant |
| Matching against inclusion and exclusion | Agent | Recorded, reversible, not a final call |
| Proposing candidates with evidence cited | Agent | Surfaces matches; cannot determine eligibility |
| Eligibility determination | Investigator | Approval gate; requester cannot self-approve |
| Advancing a patient to screening | Investigator | Signed, hash-chained audit entry |
The structural part is what separates this from a policy memo. It is not enough to say the agent should stop at a proposal. The same agent must provably be unable to act as both the maker of the screen and the checker that marks the patient eligible on a single run. When it tries, the attempt is refused, and the refusal lands in the log, which is frequently the exact evidence a trial monitor wants to see.
A screen the way it actually runs
The agent screens an oncology population against a second-line trial protocol, reading line of therapy, biomarker and mutation status, ECOG, prior treatments, washout, and disease progression out of the unstructured notes and path reports. It returns two proposals. One is a fully-matched candidate, every inclusion met and no exclusion triggered, each criterion backed by a cited line in the chart. The other is a borderline case, turning on an ambiguous progression note that could read either way, with the matched and unmatched criteria laid out beside the evidence.
The principal investigator signs the eligibility determination on the matched candidate, and sends the borderline one back for chart review rather than guessing at the ambiguous note. The agent did the reading. The named investigator owns the determination the agent is designed against making alone, and the record carries both, including the case that was sent back.
Why the record is the point
A cohort decision is, in the end, a chain of decisions a sponsor will be asked to defend during inspection. Who determined this patient eligible? On what date, under which version of the protocol? On what cited evidence, and which criteria were judged met? Was a patient advanced without a named human behind the call?
A control plane produces this evidence as a by-product of doing the work. Every model call, every grant state, every gate, and every signed determination lands in an append-only, hash-chained, cryptographically signed ledger. Change one entry and the chain visibly breaks. Each case exports as an Ed25519-signed evidence pack that a trial monitor verifies offline, against a published spec, with no access to your systems. That is the form of proof an auditor trusts, and the audit-trail and signature-meaning obligations it produces evidence for are the same ones we work through in Part 11 and AI agents.
The honest version of the pitch
AI agents will not replace the investigator, and any vendor implying otherwise is selling the part of the job that carries the ICH-GCP and patient-safety liability. What agents replace is the grind, the hours spent reading notes and path reports and matching criteria before a named investigator ever has to determine anything.
That trade is worth taking only if the line between machine throughput and human judgement is enforced and recorded, not promised. Get it right and the agent comes out of the pilot it has been stuck in into a workflow you can put in front of a monitor, because every eligibility determination has a name on it, and a record that holds. The adjacent trial work runs on the same logic, see how it applies to clinical trial data with AI agents.
See how it works, or book a demo to watch an agent get blocked from approving its own work, live.