AI for clinical trial protocol feasibility research works in two distinct ways: as infrastructure that quantifies accrual, duration, and early-termination risk before enrollment begins, and as extraction systems that convert dense protocol text into structured data for operational planning [1][2]. The most credible deployments treat model output as an input to human evaluation under alternative assumptions, not as a replacement for feasibility review [4]. Evidence from a 2026 extraction study shows that retrieval-augmented generation (RAG) pipelines achieve 89.0% accuracy against expert-supported annotations, far above the 62.6% for standalone large language models, while simulated coordinator tasks were completed at least 40% faster [2][5]. Those gains are real but bounded: the authors emphasize that expert oversight remains essential and that results may not generalize to adaptive designs or specialized therapeutic areas without tailored schemas [5]. A feasibility team should therefore focus on three decision points—whether to extract, whether to predict, and how to govern both.

What the Named Workflow Actually Involves

Protocol feasibility research typically begins with a structured review of the proposed eligibility criteria, schedule of events, endpoint definitions, and site and country assumptions. In a manual workflow, study teams read the protocol document, annotate the operational elements that affect recruitment and data collection, then compare those elements against historical site performance and disease epidemiology [1]. The AI literature describes this same workflow being split into two tasks: information extraction from the protocol text, and predictive modeling of operational outcomes such as accrual rate, duration, and probability of early termination [1][2].

A 2026 Journal of Biomedical Informatics study formalized the extraction component for clinical research coordinators. Investigators sampled interventional drug studies from ClinicalTrials.gov with identifiable protocol documents, drawing 9 trials from oncology, 7 from cardiovascular, and 7 from other therapeutic areas conducted in Canada or the United States [2]. They built semi-structured data models covering six categories, then evaluated a clinical-trial-specific RAG pipeline against publicly available LLMs with fine-tuned prompts. The RAG system's best configurations, involving Claude Sonnet 4.5 and Claude Opus 4.5, reached 89.0% average accuracy, while the standalone LLM approach reached 62.6% [2][5].

Predictive Feasibility and the Limits of Retrospective Models

The eClinicalMedicine state-of-the-evidence review identifies early risk prediction as a legitimate application domain, covering models that forecast accrual, duration, and probability of early termination from trial features, disease epidemiology, and historical site performance [1][4]. However, the same source raises a governance concern that is often omitted in vendor materials: retrospectively validated models may fail to alter operational decisions prospectively, and discrimination metrics alone do not demonstrate trial-level impact [4]. The practical recommendation is to use these tools primarily for design planning and site prioritization, while evaluating decision consequences such as time saved, screen-failure burden, staffing needs, and probability of missing event targets [4].

A commercial illustration of the predictive approach appears in Medidata's Study Feasibility product documentation. The vendor states that 43% of study sites fail to enroll a single patient, and that its platform uses AI and real-time site- and country-level performance metrics to predict high-enrolling and non-enrolling sites and forecast enrollment [6]. Medidata attributes its capability to "the industry's largest operational dataset" and claims support for scenario simulation across site and country footprint, enrollment forecasting, and diversity planning [6]. Those are vendor claims tied to a proprietary dataset, not independent validation results, and the public sources do not provide comparative accuracy figures against alternative feasibility methods [6].

Evaluation Checklist for an AI Feasibility Pilot

Select a bounded protocol sample with expert-supported reference annotations before measuring system accuracy; the JBI benchmark is a relative comparison, not a universal performance guarantee [2][5]. Document in writing whether the underlying model is locked or recalibrated during the feasibility phase, and assign a named trial role with authority to pause or retire the tool [4]. Test extraction on protocols with heterogeneous formatting and at least one complex or adaptive design element, since published results may not generalize to those settings [5]. Pair predictive feasibility output with decision consequences—screen-failure burden, staffing needs, and probability of missing event targets—rather than relying on discrimination metrics alone [4].

Governance and Technical Decision Points

A recurring theme in the primary literature is the distinction between AI-as-intervention and AI-for-trial-operations. The eClinicalMedicine review states that whether a system constitutes the intervention under evaluation or trial infrastructure depends on how its output is used, rather than on the underlying technology alone [1][4]. For feasibility research, the practical implication is that many AI tools sit in an operational gray zone: they influence protocol and site decisions but are not themselves the treatment being studied. That does not exempt them from change control, but it does change which reporting standards apply [1][4].

Model versioning and update rules are specific technical decision points. The SPIRIT-AI extension requires that when AI is part of the intervention, the protocol must state which algorithm version will be used, whether it differs from versions in prior justification studies, and what changed between versions [3]. For operational feasibility tools, the equivalent governance question is whether the model is locked, periodically recalibrated, or adaptively updated, and who has authority to pause, modify, or retire it [4]. The literature recommends encoding update cadence, recalibration data needs, approval pathways, shadow-mode requirements, rollback triggers, and freeze periods around interim analyses and database lock [4]. None of these parameters can be inferred safely from a vendor fact sheet.

A Concrete Evaluation Example

Consider a suggested workflow for a sponsor evaluating AI-assisted feasibility research ahead of a phase 2 oncology protocol. The team could select a small set of recent protocols with complete documents and ask two systems to extract the same structured elements: inclusion criteria, exclusion criteria, schedule of events, endpoint definitions, and site selection assumptions. Using the RAG benchmark as a reference point, the team would expect around 89.0% accuracy for a well-configured retrieval-augmented system, but would need to create its own expert-supported reference annotations because the published benchmark was built on disease-specific samples and may not transfer to new therapeutic areas [2][5].

The same evaluation should test what the extraction system cannot do. The arXiv version of the JBI study lists limitations that are operationally relevant: establishing complete ground truth is difficult even for expert reviewers, results may not generalize to complex adaptive designs, and the modest dataset cannot capture the full spectrum of real-world variability in protocol formatting and document quality [5]. A feasibility team should therefore design a shadow-mode test where the AI output is reviewed by a clinical research coordinator blinded to the system's confidence level, and where disagreements are logged as failure modes rather than silently corrected [4][5]. This matches the emerging regulatory interest in comparative evaluation; the FDA's April 2026 request for information on AI-enabled optimization of early-phase trials explicitly asks how performance should be evaluated across patient populations, sites, therapeutic areas, and how model drift should be detected and mitigated [7].

Where the Field Is and Is Not Ready

The eClinicalMedicine conclusion is candid about maturity levels. The field is best supported in computable eligibility, prescreening, adjudication support, and certain forms of risk-based monitoring, while it is less developed where causal interpretation, estimands, lifecycle governance, and real-world implementation under institutional constraints are most demanding [1][4]. For feasibility research specifically, extraction and basic accrual prediction have clearer evidentiary footing than claims about adaptive protocol optimization or autonomous decision-making. The next phase, according to the review, should emphasize prospective studies that evaluate the impact of decisions, interoperable protocol logic, explicit change-control policies, and trial-specific oversight for retrieval-enabled and agentic systems [1][4].

Regulatory signals reinforce this caution. The FDA request for information asks what evidence should demonstrate validity and reliability in clinical trial contexts, what metrics can assess transparency and explainability, and how privacy protections and fairness across demographic subgroups should be evaluated [7]. A separate NEJM AI study on ambient AI in routine care shows how far a governance playbook can go when documentation sustainment is treated as a first-class concern: an internal LLM coder achieved a Pearson correlation of 0.97 with professional coders, but a note template change initially reduced ICD-10 documentation accuracy from 79% to 35% before redesign and retraining restored baseline [8]. That result is a reminder that feasibility workflows are embedded in larger institutional systems where small configuration changes can produce large downstream effects [8].