Agentic AI in Clinical Trials: Recruitment and Monitoring

More than half of aborted clinical trials fail because too few patients were accrued, and 80% of all trials miss their recruitment timeline entirely.

That's according to a 2025 survey of LLM-assisted trial recruitment from researchers at the University of Tübingen and Boehringer Ingelheim.

The bottleneck was never a lack of eligible patients. It's that finding them means a coordinator reading through charts one at a time, checking each one against criteria that have gotten steadily more complex for two decades running.

This guide covers agentic AI in clinical trials specifically, part of the broader picture in Tecla's Agentic AI in Healthcare guide: how patient matching actually works, where protocol monitoring fits, and what the FDA currently expects.

What Is Agentic AI in Clinical Trials?

Agentic AI for clinical trials is a system that reads a patient's medical record against a trial's eligibility criteria and reasons through the match, rather than just filtering on a fixed set of database fields.

A trial matching database filters on structured fields: age, diagnosis code, lab value ranges.

Agentic AI reads the parts of a chart a keyword filter can't reach, clinical notes, treatment history, physician impressions, and reasons about whether a patient actually fits.

That distinction matters because eligibility criteria have gotten genuinely harder to satisfy on paper.

Research on trial matching found the median number of criteria per trial rose from 31 in the early 2000s to 49 by 2020, with oncology trials routinely exceeding 60.

A structured filter can't tell you whether a patient's prior treatment counts as a disqualifying prior therapy or a different drug class entirely. An agentic system reads the note, reasons through the distinction, and explains which way it landed.

From Manual Chart Review to Agentic Matching

Patient recruitment automated the way most research workflows did: structured queries first. A registry search filters by diagnosis code and basic demographics, cutting a large population down to a workable list.

That filter still leaves the hard part undone. Someone has to open each remaining chart and check it against every inclusion and exclusion criterion, a process that can take hours per patient in complex trials.

Agentic AI reasons through that same chart instead of just filtering on codes. It reads clinical notes, cross-references treatment history against exclusion criteria, and produces a rationale for each match, not just a yes or no.

One published system, TrialGPT, reported a 42% reduction in screening time when coordinators reviewed its ranked matches instead of screening charts manually.

Most production deployments still keep the registry filter as a first pass. Agentic reasoning runs on the shortlist that filter produces, with a coordinator confirming every match before a patient is approached.

The Recruitment and Monitoring Workflow, Step by Step

The workflow below covers patient matching specifically. The sections after it cover protocol monitoring, the FDA's current framework, and where a coordinator's judgment still has to sit.

Patient Matching Against Eligibility Criteria

Eligibility criteria have a dual nature: universal fields like age and consent capacity, and trial-specific clinical thresholds that vary by protocol. Conventional matching handles the first well and the second poorly.

Agentic matching handles both because it reads the actual clinical narrative rather than a fixed set of fields.

A patient excluded on paper for "prior systemic therapy" might turn out, on reading the note, to have received a topical treatment that doesn't meet the exclusion at all.

Protocol Deviation Detection and Risk-Based Monitoring

Risk-based monitoring shifts oversight away from checking every data point on-site toward flagging the sites and patterns most likely to affect safety or data quality.

Tools like Medidata Detect and CluePoints apply statistical and machine learning models to this exact problem.

A deviation flagged without context is just noise added to a coordinator's queue.

The same reasoning that makes matching useful applies here: correlating a flagged deviation against that site's history and the specific protocol requirement it touches, not just surfacing a raw statistical outlier.

The FDA's Risk-Based Framework for AI in Trials

The FDA's January 2025 draft guidance ties its expectations to context of use: how much a given AI application actually influences a regulatory decision, not the technology itself.

In January 2026, the FDA and EMA jointly published ten guiding principles for good AI practice in drug development, aligning the two agencies' expectations for sponsors running trials across both jurisdictions.

Neither document treats patient matching or monitoring as exempt. A sponsor still has to document what the model does, how it was validated, and what happens when it's wrong.

The Research Coordinator's Role

The coordinator's job shifts from reading every chart in a population to reviewing the shortlist an agent already built, confirming the reasoning.

What can't be delegated is obtaining informed consent and managing the relationship with the patient and the site.

Implementation: Guardrails Specific to Clinical Trials

A matching or monitoring system that can't show its work doesn't just create an internal problem. It creates a finding waiting to happen the next time a regulator asks how a decision was made.

Rolling This Out: What to Expect

A single protocol is the right place to start, not a portfolio of active trials at once. Run the agent's matches alongside manual screening long enough to see where the two processes actually disagree, not just where they overlap.

Document validation before the model touches a live trial: what data it was trained or tuned on, how matches were checked against confirmed enrollments, and who signed off. That record is exactly what an inspection will ask for later.

Expect the eligibility criteria themselves to need review, not just the matching tool. Criteria written for a human reviewer to interpret loosely sometimes need to be made explicit before a system can reason about them consistently.

The Team Behind Production Agentic AI

The differentiator in agentic clinical research isn't the underlying language model. It's the validation record and the audit trail sitting around every match and every flag.

Tecla's Agentic AI services design, build, and operate this workflow directly, the same matching, monitoring, and documentation systems above, running in your stack with the evals and guardrails production requires.

Or bring the expertise in-house: AI engineers who've worked on live clinical research systems, past the demo stage.

Tecla runs a network of senior engineers across the US and Latin America, built over more than a decade, with a top 3% acceptance rate and first candidates in 3 to 5 business days.

FAQ