A CRC I spoke with last month described opening a new protocol for the first time — Phase 2 oncology, complex eligibility criteria, six stratification factors — and finding zero reference to how the screening population had been defined. No patient registry data. No historical screen failure analysis. No documented rationale for the inclusion thresholds. Just criteria, pulled apparently from a prior protocol in an adjacent indication, restated with minor edits. She spent the next three weeks reconciling eligibility language against the site’s actual patient population before the PSV. The sponsor’s AI-assisted design tool had flagged zero issues.
That disconnect — between what a protocol says and what the underlying patient data actually supports — is now formally quantified. Phesi’s 2026 analysis found that fewer than one in three trial protocols are connected to documented patient data and outcomes. The company, founded in 2007 by Dr. Gen Li, has built its platform specifically around linking protocol design decisions to real clinical data. Their conclusion is direct: the industry is not solving flawed decision-making in clinical development. It is scaling it.
The Protocol Is the Blueprint — But Whose Blueprint?
Here is the operational reality that statistic creates at the site level. When a protocol’s eligibility criteria, dosing windows, visit schedules, and assessment frequencies are designed without grounding in documented patient outcomes, the site inherits every embedded assumption. Those assumptions then drive the SIV, the enrollment plan, the screen failure projection, and the budget. Sites I work with across our network routinely absorb the downstream cost of upstream guesswork — and the word “guesswork” is not hyperbole when fewer than a third of protocols can demonstrate their design choices trace back to actual patient data.
Screen failure rates are where this hits first and hardest. When eligibility criteria are drafted without reference to the real-world patient population at the sites being recruited, screen failure rates run high — sometimes catastrophically. A site projecting a 1:3 screen-to-enroll ratio based on sponsor guidance, then experiencing 1:7 in practice, is not just missing enrollment targets. That site is absorbing uncompensated hours: coordinator time, physician screening assessments, lab draws, and the administrative burden of documenting each screen failure against protocol-specific criteria. The WCG 2024 site startup analysis found median trial activation at academic medical centers running 8.12 months from site identification to study start. A protocol that was never validated against patient data before that clock started running is adding friction to every week of it.
Now add AI-assisted protocol authoring to that dynamic. Bristol-Myers Squibb announced partnerships with both Faro Health and Evinova in late 2025 and early 2026 to automate protocol design using structured AI tools. The pitch is faster first drafts, reduced cycle time, greater internal consistency. Those benefits are real. But Faro Health’s platform generates initial draft sections from templates and converts narrative protocol elements into structured assets — and if the template was built on protocols that were themselves disconnected from patient data, the AI is not introducing a new problem. It is reproducing the original problem at production speed, with the added complication that the output looks clean and internally consistent, which creates confidence it has not earned.
Phesi’s finding maps directly onto ICH E6(R3), which in its updated framework places explicit weight on risk proportionality and evidence-based protocol design. The principle is not new. What is new is the gap between what the guidance asks for and what the data shows is actually happening. Fewer than one in three protocols demonstrably connected to patient outcomes means that the majority of trial designs entering site qualification, IRB submission, and study startup are operating on assumptions that have never been stress-tested against reality — and the FDA’s December 2024 draft guidance on protocol deviations makes clear that the agency expects sponsors to have defined, documented rationale behind the criteria that generate those deviations in the first place.
Where the Breakdown Lives in Your Workflow
The failure mode Phesi is describing has a specific address in the trial lifecycle: the gap between protocol finalization and site feasibility. Sponsors finalize a protocol, then send feasibility questionnaires to sites asking how many eligible patients they can identify. Sites answer based on their clinical judgment and informal chart reviews. Nobody goes back and reconciles the sponsor’s eligibility criteria against population-level data from the actual sites being selected. The protocol’s assumptions about patient availability, co-morbidity rates, and washout compliance are never formally tested before the contract is signed.
That is a contract scope problem as much as a data problem. CRO agreements routinely define protocol translation and site training deliverables, but they rarely include a pre-activation feasibility reconciliation step that requires documented patient data to validate the eligibility framework before first-patient-in. Across our network, the sites that avoid the worst screen failure spirals are the ones where either the sponsor or the CRO ran a structured retrospective chart review — using real patient records from the specific therapeutic area at that specific site — before the budget was set. That step takes two to four weeks. Missing it costs months.
The FDA warning letter issued April 29, 2026 to investigator Sourav K. Mishra, M.D. at the All India Institute of Medical Sciences for objectionable conditions found during a March 2025 inspection is a useful illustration of what happens at the inspection end of the pipeline when documentation gaps are present at the design end. Inspectors are not just reviewing source documents and deviation logs — they are tracing decisions back to their documented rationale. A protocol whose design cannot be traced to patient data is a protocol whose site-level deviations will be difficult to contextualize when a BIMO inspector arrives and starts asking why screen failures ran at twice the projected rate.
What Changes Monday Morning
For sites: before the next SIV, ask the sponsor directly for the patient data source behind the eligibility criteria. Not the protocol synopsis — the underlying analysis. If that data does not exist or cannot be shared, document your own feasibility data independently: pull a retrospective review of charts from the prior 18 months in the relevant indication and run your own eligibility simulation. That exercise takes coordinator hours, but it is the only honest basis for an enrollment commitment. It also becomes your evidence if a contract amendment becomes necessary later because the protocol assumptions were wrong.
For sponsors and CROs: every protocol that enters the AI-assisted drafting pipeline should require a documented patient data linkage before finalization — not as a compliance checkbox, but as a contract deliverable. Define it in the CRO scope of work: a pre-activation feasibility reconciliation, completed before IRB submission, that traces the primary eligibility criteria to documented outcomes data from the target patient population at the selected sites. The AI tools you are deploying — Faro Health, Evinova, and the others — are as good as the data they are trained on and the validation gates you build around them. Without that gate, the efficiency gain in protocol drafting is real, and so is the efficiency loss at every site that inherits the result.
Fewer than one in three protocols linked to patient data means that two in three trials begin site activation with eligibility assumptions that have never been formally tested. The sites bearing the operational cost of that gap are also the sites you are depending on to close enrollment. That arithmetic has a way of resolving itself — slowly, expensively, and usually on the back half of a timeline that was already under pressure.
References
- FierceBiotech — “Clinical trial flaws ‘being scaled, not solved’ by AI: report”
- Phesi — “Our Team” (Dr. Gen Li, Founder, 2007)
- WCG Clinical — “Decoding the Top Site Challenges of 2024: Study Start-Up” (median AMC activation 8.12 months)
- Intuition Labs — “AI Clinical Trial Protocol Design: BMS, Faro Health, Evinova partnerships (2025–2026)”
- FDA — “Protocol Deviations for Clinical Investigations of Drugs, Biological Products, and Devices” (Draft Guidance, December 2024)
- FDA — Warning Letter to Sourav K. Mishra, M.D., All India Institute of Medical Sciences (April 29, 2026)
