AI Can Find Trial Candidates. Can Your Site Convert the Referral?

The Evidence Is Promising, but Matching Is Only the First Step

There is credible evidence that AI can reduce portions of the manual matching workload. A Nature Communications study of TrialGPT reported that its patient-to-trial matching framework reduced screening time by 42.6% in a user study. The system was designed to retrieve potentially relevant trials, assess individual eligibility criteria, and rank the resulting matches for human review. ([pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/39557832/))

That is operationally meaningful. Coordinators and investigators frequently spend substantial time interpreting long eligibility sections, reviewing fragmented records, and determining which questions still need answers. A system that narrows the field can allow qualified staff to focus their attention where clinical judgment is most valuable.

However, matching performance should not be confused with enrollment performance. A separate 2026 study evaluated an LLM-based matching platform in gastrointestinal surgical oncology clinics. The system reviewed 514 patients and produced 34 trial matches across 32 patients. Nine matches resulted in enrollment, while 25 did not. Documented reasons included ineligibility, participant decline, provider discretion, and cases with no recorded explanation. ([pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/41318261/))

This was one specialized setting and should not be generalized to every trial. Still, it illustrates the central site-level issue: identifying a plausible match does not resolve the clinical, logistical, or human reasons that determine whether enrollment occurs.

Build a Referral Acceptance Workflow, Not Just an Inbox

A common implementation mistake is treating technology-generated referrals like ordinary website leads. AI-assisted referrals may look more qualified, but they can still contain outdated diagnoses, incomplete medication histories, missing laboratory results, or interpretations that require confirmation.

Before referrals begin arriving, the sponsor, recruitment provider, CRO, and site should agree on the minimum information needed for a useful review. That does not mean transferring an entire medical record. It means identifying the specific fields required to make the next decision while protecting privacy and limiting unnecessary data handling.

The workflow should also distinguish acknowledgment from qualification. A site may be able to confirm receipt and contact a candidate quickly, but final prescreening can require records, investigator review, washout calculations, or confirmation of prior treatment. Combining these activities into one response-time target encourages superficial review or misleading status reporting.

• Define who owns each referral from initial transfer through final disposition.
• Specify the minimum data package and approved method of transmission.
• Set separate targets for acknowledgment, first contact attempt, preliminary prescreening, and clinical review.
• Create escalation rules for urgent safety information, missing records, repeated contact failures, and potentially eligible candidates nearing a protocol-defined window.
• Establish a closed-loop process so the referring system receives accurate, standardized outcomes.

The Human Handoff Still Determines Whether Access Is Real

Recruitment platforms can expand reach, but participants experience the trial through conversations with people. They may have questions about visit frequency, investigational treatment, randomization, reimbursement, time away from work, caregiving, or whether participation affects their usual medical care. A technically accurate match does not answer those concerns.

The first site contact should therefore be treated as a research interaction, not a sales call. Staff need enough protocol knowledge to explain the study at an appropriate level without overpromising eligibility or benefit. They also need a clear route for questions that require investigator input.

Communication design becomes especially important when recruitment reaches multilingual communities. A bilingual advertisement is not enough if the callback, prescreening, consent preparation, scheduling, and follow-up processes cannot consistently support the participant’s preferred language. Sites serving communities in South Florida or Puerto Rico should test the whole pathway, including translated materials, staffing coverage, voicemail, text messaging, record retrieval, and escalation to clinical personnel.

Technology may identify people who were previously difficult to reach. The site’s operating model determines whether those people encounter a practical opportunity to participate.

Sponsors Should Test Site Capacity Before Increasing Referral Volume

More referrals are helpful only when the site has the capacity to process them with reasonable speed and quality. A recruitment plan should account for coordinator time, investigator availability, expected contact rates, record-review requirements, screen-failure risk, and the number of competing studies handled by the same team.

ICH E6(R3) states that investigators should be able to demonstrate potential to recruit the proposed number of eligible participants and should have sufficient qualified staff, time, and facilities. It also makes clear that investigators retain ultimate responsibility and appropriate oversight when trial-related activities are delegated to other persons or service providers. ([database.ich.org](https://database.ich.org/sites/default/files/ICH_E6%28R3%29_Step4_FinalGuideline_2025_0106.pdf))

This makes feasibility more than an estimate of available patients. For technology-enabled recruitment, feasibility should include throughput. A site may have access to a large population but lack the staff capacity to make timely calls, obtain records, complete investigator reviews, and schedule screening visits.

A practical capacity review should ask how many new referrals the site can safely process per day, who provides coverage during coordinator absences, how quickly the investigator can resolve uncertain criteria, and what happens when a campaign generates an unexpected surge. If those answers are unclear, increasing advertising spend may increase waste rather than enrollment.

Measure the Funnel Where It Actually Breaks

Top-line referral counts rarely explain recruitment performance. Sponsors and sites need a shared disposition structure that shows where candidates are being lost and how long each stage takes.

The categories should be specific enough to support action. “Not eligible” is less useful than distinguishing diagnosis not confirmed, prohibited medication, laboratory criterion, prior treatment, timing window, comorbidity, or missing documentation. “Not interested” should not become a catch-all for someone who could not travel, could not reach the site, misunderstood the study, or never received a response.

Sites should also avoid creating so many disposition categories that staff cannot apply them consistently. A short controlled list, paired with optional notes and periodic review, is usually more useful than a complex taxonomy no one trusts.

• Referral-to-first-attempt time
• Percentage reached and number of attempts required
• Percentage completing preliminary prescreening
• Time waiting for records or outside clinical information
• Percentage requiring investigator review
• Prescreen-to-screen conversion
• Screen-failure reasons by criterion
• Participant-decline reasons
• Time from referral receipt to final disposition
• Percentage of referrals closed without a documented reason

Quality and Privacy Must Be Designed Into the Workflow

AI does not remove the investigator’s responsibility for trial conduct, and a vendor’s eligibility recommendation should not become an undocumented clinical decision. Sites need to know which information the system used, what remains unverified, and where human review is required.

The workflow should leave a clear record of referral receipt, contact attempts, source information reviewed, unresolved eligibility questions, investigator input when applicable, and final disposition. Access should be role-based, and staff should not copy sensitive information into unapproved messaging, spreadsheets, or personal devices simply because the primary platform is inconvenient.

For trials using decentralized elements, FDA recommends documenting operational responsibilities and accounting for data origins, flows, technologies, and service providers. FDA also emphasizes that decentralized activities do not change the underlying responsibilities of sponsors and investigators. ([fda.gov](https://www.fda.gov/media/167696/download?attachment=&utm_source=openai))

These expectations are relevant even when the technology is used only for recruitment or prescreening. The safest implementation is one where responsibilities, data boundaries, escalation routes, and oversight are established before the first candidate enters the system.

The Operational Test for AI Recruitment

The current wave of AI recruitment announcements is worth watching. Better matching, more targeted outreach, and faster review could reduce repetitive work and help sites find candidates who would otherwise remain unaware of research opportunities.

But the strongest technology will still encounter the realities of clinical research: incomplete records, nuanced criteria, participant preferences, investigator judgment, scheduling limitations, and the need to build trust. The practical benchmark is not how many names an algorithm produces. It is whether the complete system helps qualified candidates move through an understandable, timely, well-documented process.

For sponsors and CROs, that means evaluating the receiving workflow as carefully as the matching model. For sites, it means treating referral operations as a controlled clinical process with defined ownership and measurable capacity. That is how AI-assisted recruitment becomes more than another source of leads.