Picture a site coordinator at a large academic medical center, three weeks past the study initiation visit, staring at a list of 847 potential patients that the sponsor’s new AI platform just dropped into her inbox. The list was generated by Abridge, the AI company that Eli Lilly’s venture arm recently invested in specifically to help Lilly identify patients who may qualify for specific clinical trials. The AI listened to clinic conversations, flagged clinical language that matched inclusion criteria, and surfaced the names. All of which sounds like the recruitment problem is finally solved. But the coordinator knows something the algorithm does not: her site has two open slots in the Thursday afternoon screening clinic, a pharmacy team that needs 72 hours’ notice for IP preparation, and an IRB amendment still pending because the sponsor changed the eligibility window two months ago. Eight hundred and forty-seven names mean nothing if the infrastructure to convert them is not ready.
That gap between patient identification and patient enrolled is where most recruitment failures actually live, and it is the gap that Lilly’s investment in Abridge puts into sharp relief.
The Identification-to-Enrollment Gap
Abridge built its platform on a specific capability: AI that records and transcribes patient-physician conversations during appointments, then surfaces clinical signals that match protocol criteria. The company has been deploying this technology across health systems for documentation and EHR integration since its founding in 2018. UI Health, the academic health enterprise at the University of Illinois Chicago, deployed Abridge across inpatient, outpatient, and emergency settings and observed measurable reductions in administrative burden during a pilot phase. Lilly’s move is a logical extension: if the platform can extract clinical language from a conversation and route it to a chart note, it can also route it to a protocol eligibility checklist. That pivot from documentation tool to recruitment intelligence tool is genuinely useful. The question is what happens after the flag fires.
Industry data on this point is not encouraging. Approximately 80% of clinical trials experience delays or closures due to recruitment problems, and 11% of research sites fail to enroll a single patient across a study’s entire lifetime. Those numbers are not a failure of patient identification. Sponsors have been generating pre-screened patient lists, running social media targeting, and partnering with patient advocacy groups for years. The failure is operational conversion: the distance between “this patient matches the criteria” and “this patient has signed the ICF, cleared the screen, and been randomized.” That distance is measured in coordinator hours, IRB status, scheduling slots, and site readiness, none of which an AI flag addresses.
The financial weight behind that gap is significant. Enrollment delays cost sponsors in the range of $50,000 per day in delayed revenue and operational burn. When a sponsor invests in upstream identification technology without investing equally in the downstream site infrastructure to act on those referrals, the delay cost does not decrease. It shifts. The site absorbs coordinator time chasing unqualified leads from an algorithm that does not know about the pending amendment, the PI’s travel schedule, or the fact that the central lab courier only comes on Tuesdays.
Where the Handoff Actually Breaks
Sites I work with across our network describe a consistent pattern with sponsor-generated recruitment leads: the list arrives, the coordinator spends two to three days cross-referencing it against current eligibility criteria (often from a protocol version the sponsor has already amended), and somewhere between 40% and 60% of the names are disqualified on the first pass for reasons that had nothing to do with the AI’s clinical signal matching. Wrong age bracket because the amendment changed the window. Excluded medication the algorithm did not know to flag. Existing diagnosis that the transcription captured incompletely. The screen failure rate problem does not begin at the screening visit; it begins at the referral stage, and AI-generated referral lists can actually inflate pre-screen burden if the eligibility logic feeding the algorithm lags behind the live protocol.
The IRB documentation layer compounds this. Under FDA’s guidance on electronic informed consent (21 CFR parts 11, 50, and 56), any change to how patients are identified or approached as part of a recruitment strategy can trigger an IRB review of recruitment materials. If Abridge’s AI-generated outreach or referral workflow touches the patient before consent, the site’s IRB submission needs to reflect that mechanism. Most sites do not have a template for “AI-flagged patient referred by ambient clinical documentation system.” That is a new consent pathway description that has to be written, reviewed, and approved before the first flagged patient can be approached. Sites that discover this after activation has been granted are looking at an amendment cycle that, at most academic IRBs, runs four to eight weeks on a non-expedited track.
Abridge’s funding trajectory suggests Lilly is not the only organization betting on this capability. The company raised a $150 million Series C in February 2024 at an $850 million valuation, before Lilly’s venture investment added additional strategic weight. That capital concentration signals competitive pressure: Roche announced in March 2026 the expansion of its AI infrastructure, deploying 2,176 new NVIDIA Blackwell GPUs as part of a large-scale AI factory build. Sponsors across the top tier are moving into AI-assisted trial operations simultaneously, which means sites will face multiple, non-interoperable recruitment intelligence platforms running concurrently across their open protocols. A coordinator managing five active studies in 2027 may be receiving AI-generated referral lists from five different sponsor systems, each with its own data format, eligibility logic version, and escalation workflow.
What Changes Monday Morning
For sites, the immediate operational move is to treat any sponsor-deployed AI recruitment tool as a new recruitment method requiring IRB review of your consent and approach language, full stop. Do not wait for the sponsor to flag this. Pull your current IRB-approved recruitment materials and map every patient touchpoint that the new tool affects. If Abridge or any comparable platform is generating flags that result in a coordinator calling a patient, that call is a recruitment contact and it needs to be covered by your approved recruitment script. Raise the question with your IRB coordinator before the sponsor asks whether you are enrolling yet.
For sponsors pushing these tools to sites, the operational ask is direct: do not hand a site an AI referral list without simultaneously handing them an amended feasibility reconciliation that accounts for the site’s actual screening capacity, current IRB status, and protocol version alignment. Sites I work with have been burned by recruitment tools that generate volume the site cannot operationally process, which means the tool creates query noise rather than enrollment momentum. Build the IRB submission language for the AI recruitment workflow into your standard site activation package. If your central IRB is handling the study, get the recruitment method language approved at the central level before sites are asked to activate under the new workflow.
Lilly’s investment in Abridge is a signal worth paying attention to, but not because AI-flagged patient identification is new. The signal is that a top-five sponsor has decided that ambient clinical documentation is now a legitimate upstream channel for trial recruitment, and that decision will migrate into contract language, feasibility questionnaires, and site selection criteria faster than most sites are currently prepared for.
The sites that get ahead of this will be the ones that have already answered the IRB question, mapped their screening capacity honestly, and told their sponsor counterparts exactly how many AI-generated referrals they can convert per month given their current coordinator headcount. That number, not the size of the patient list, is what determines whether the technology delivers on its promise.
References
- Endpoints News — “Eli Lilly invests in startup Abridge for clinical trial recruitment”
- Abridge — Series C $150 Million Funding Announcement, February 2024
- Abridge — “UI Health Deploys Abridge Across Care Settings”
- Antidote.me — “25 Useful Clinical Trial Recruitment Statistics for Better Results”
- Endpoints News — “The $50K-a-Day Problem in Clinical Trial Enrollment”
- FDA — “Use of Electronic Informed Consent in Clinical Investigations: Questions and Answers”
- Roche — AI Infrastructure Expansion Announcement, March 2026
