AI can now screen health records for clinical trial eligibility faster and more thoroughly than study staff, and a 2025 randomized trial in JAMA showed it nearly doubled enrollment. But finding eligible patients is only half of recruitment. Most eligible patients still never enroll, because consent depends on conversations, trust and logistics that software does not handle.
At Mass General Brigham, researchers ran an unusual experiment: a randomized trial of how to recruit for a randomized trial. They took 4,476 patients who might qualify for COPILOT-HF, a heart failure study, and split them in two. Half were screened the usual way, by study staff reading charts. The other half were screened first by RECTIFIER, a generative AI tool that read their electronic health records, including clinical notes, and flagged who looked eligible. Staff then did a quick check of the flagged charts.
The results, published in JAMA in February 2025, were striking. The AI arm identified 458 eligible patients against 284 in the manual arm. It enrolled 35 people against 19. The enrollment rate in the AI arm was nearly double.
Now look at those numbers from the other direction. Of 458 patients the AI correctly found, 35 ended up in the trial. More than nine in ten eligible people did not join.
That is the part of the story that matters most, and it is the thesis of this piece. Recruitment AI solves a real and expensive problem, which is that eligibility is buried in free-text notes nobody has time to read. But the moment it solves that problem, the bottleneck moves to a place software cannot reach: a person deciding whether to give their body and time to research. Hospitals that buy screening AI without investing in the human half will find more patients and still miss their enrollment targets.
How does AI find patients for clinical trials?
Trial eligibility criteria are written in dense clinical language: ejection fraction below a threshold, no hospitalization in the past 30 days, not on a particular drug, no history of a specific condition. Some of that information sits in structured fields in the health record, such as lab values, diagnosis codes and prescriptions. A great deal of it does not. It sits in discharge summaries, clinic letters, imaging reports and progress notes.
Older recruitment software queried the structured data. It was fast and missed a lot. Codes are incomplete, lab values lack context, and the reason a patient stopped a medication is almost never in a structured field. Human screeners read the notes, which is accurate and painfully slow.
Large language models close that gap. They can read the notes, compare them against each criterion and explain their reasoning. The National Library of Medicine’s TrialGPT, described in Nature Communications in November 2024, is a public example. In manual evaluation of more than 1,000 patient-criterion pairs, its criterion-level judgments were 87.3 percent accurate, close to expert performance, and in a user study it cut screening time by 42.6 percent.
RECTIFIER takes the same approach inside a single health system, retrieving relevant passages from the record and asking the model whether each criterion is met. The randomized JAMA study matters because it measured the outcome that counts, enrollment, rather than accuracy against a labeled dataset.
Why doesn’t better screening fix recruitment?
Because screening is the funnel’s top, and most loss happens below it.
Once an eligible patient is identified, someone must reach them. Their clinician may need to agree. A coordinator must explain the study, often more than once. The patient must weigh extra visits, blood draws, travel, time off work and the possibility of being randomized to placebo. They may want to talk to family. They may simply not answer the phone.
None of those steps gets faster because the chart review did. Some get harder. An AI that surfaces twice as many eligible patients doubles the workload for coordinators who make the calls, unless staffing grows too. A trial team that was screening-constrained becomes consent-constrained almost overnight.
This echoes a point we made about trial diversity being a staffing problem dressed up as a recruitment problem. The scarce resource in recruitment is rarely the list of names. It is trusted people with time to talk to the people on it.
Where does recruitment AI break?
Three failure points recur.
The record is wrong or incomplete. A model can only read what was written. Patients who get care across several systems, or whose notes are thin, look ineligible or invisible. That falls hardest on people who move between providers, including many of the populations trials most need to include. Our piece on why interoperability is still unsolved explains why that fragmentation persists.
The criteria are ambiguous. Many eligibility criteria require judgment: “clinically significant” liver disease, “stable” dose, “unable to comply.” Language models will make a call, and the call may differ from what the investigator intended. That is why the Mass General Brigham workflow kept a human review of every AI-flagged chart. The model narrowed the list. People still made the decision.
The patient was never asked. Screening records for research raises its own governance questions. In most health systems, reviewing records to identify potential participants is permitted under research approvals and privacy rules, but patients are often unaware it happens. Being contacted out of the blue about a trial for a condition you have not discussed with your own doctor can feel intrusive. Recruitment AI that ignores this will generate complaints faster than enrollments.
A worked example: a site that doubles its eligible list
Here is a hypothetical site running a cardiology trial, with illustrative numbers.
Before AI, one coordinator spends most of the week reading charts and finds about ten eligible patients a month. She contacts all of them. Four agree to a screening visit, and two enroll. The trial needs 40 patients from this site, so at two a month it will take 20 months.
The site adopts an AI screening tool. It now finds 25 eligible patients a month. The coordinator no longer reads charts; she reviews the AI’s flags, which takes a fraction of the time. But she can realistically hold meaningful conversations with perhaps 15 patients a month. Ten eligible patients a month go uncontacted. Of the 15 she reaches, six agree to a screening visit and three enroll.
Enrollment rose from two to three a month, a 50 percent gain. Real, and in line with the direction of the JAMA trial. But the site is now leaving more eligible patients untouched than it used to find in total. The binding constraint moved from reading to talking. If the site adds a half-time coordinator or a nurse who can run consent conversations, enrollment could rise further. If it does not, the AI’s extra capacity is wasted.
The lesson generalizes. Recruitment AI should be bought as part of a staffing plan, not instead of one.
The strongest case for recruitment AI
The case for these tools is strong, and it deserves to be put plainly.
Chart screening is one of the most tedious and least valued jobs in research. Coordinators burn out on it. Every hour a skilled coordinator spends reading notes is an hour not spent with patients. If AI takes over the reading, the human work that remains is the part that actually needs a human.
There is also an equity argument. Manual screening is inconsistent. Screeners, under time pressure, tend to look first at the patients who are easiest to assess, often those with complete records and regular clinic visits. A tool that reads every chart the same way can surface patients a busy human would have skipped. The JAMA study reported no significant differences in eligibility or enrollment by race, gender or ethnicity between arms, which is encouraging, though a single trial cannot settle the question.
And the enrollment gain in that trial was real. Nearly doubling the enrollment rate could substantially shorten a trial’s recruitment period, which means answers reach patients sooner.
All of that holds. None of it contradicts the point here. The strongest evidence for recruitment AI also shows, in its own numbers, that most eligible people still do not enroll.
What this means for hospitals, sponsors and patients
For hospitals and research sites, the decision to buy screening AI should come with a question: who will contact the extra patients it finds? Budget coordinator and research nurse time alongside the software. Measure contacted, consented and enrolled patients, not just flagged ones. Keep a human review step for ambiguous criteria, as the JAMA study did.
For trial sponsors, the implication is to write eligibility criteria that machines and humans can apply consistently. Vague criteria cost time with or without AI. Sponsors should also fund site staffing in proportion to the screening yield their tools promise. A sponsor that pays for AI screening but not for the conversations downstream is paying for a longer list.
For health system leaders, transparency matters. Tell patients that records may be reviewed to identify research opportunities, and route first contact through, or with the knowledge of, their own clinicians where possible. Trust built there is worth more than any matching algorithm.
For patients, the practical point is that being contacted about a trial is an invitation, not an obligation. It is reasonable to ask how you were identified, what participation involves and whether your own clinician knows. Discuss any decision to join a trial with a clinician you trust.
Frequently asked questions
How accurate is AI at matching patients to clinical trials?
Accuracy is now close to expert level on individual criteria. The NIH’s TrialGPT reached 87.3 percent accuracy on criterion-level judgments in manual evaluation, and a JAMA randomized trial found an AI screening tool identified more eligible heart failure patients than manual review. Accuracy falls when records are incomplete or criteria require clinical judgment, which is why human review remains standard practice.
Does AI increase clinical trial enrollment?
The best evidence so far says yes. In a randomized study published in JAMA in 2025, AI-assisted screening at Mass General Brigham nearly doubled the enrollment rate for a heart failure trial, 35 patients against 19. The gain came mainly from finding more eligible patients. Most eligible patients in both arms still did not enroll.
Is it legal to use AI to screen patient records for trials?
In many health systems, reviewing records to identify potential research participants is permitted under research ethics approvals and privacy regulations, with safeguards. The rules differ by country and institution. Using an AI tool typically requires the same approvals as manual screening, plus review of where data is processed and stored. Hospitals should confirm requirements with their ethics board and privacy office.
Can AI replace clinical trial coordinators?
No. AI can take over much of the chart review that consumes coordinator time, but the consent conversation, answering patient questions, scheduling and building trust remain human work. The evidence suggests that AI shifts coordinators’ time from reading records to talking with patients. Sites that cut coordinator staff after adopting AI are likely to see eligible patients go uncontacted.
The takeaway
Recruitment AI is one of the clearest wins for machine learning in clinical research, and the randomized evidence behind it is better than for most health AI. But its success exposes the next bottleneck rather than removing the problem. Eligibility lives in notes, and AI can now read them. Consent lives with people, and people still need other people to explain, reassure and follow up.
The sites that fill their trials will be the ones that spend the time AI saves on conversations, not on cutting staff.
If your site has adopted AI screening, did your bottleneck move to consent? Share what changed in the comments.
