Artificial intelligence and machine learning tools are now embedded in specific, identifiable parts of clinical trial operations — not as a general transformation of the field, but as point solutions applied to particular operational bottlenecks: finding eligible patients, choosing sites, designing and simulating protocols, and detecting safety signals in accumulating data. This guide sets out where AI genuinely operates in current trial workflows, what the regulatory posture toward it is, and where the real limits of current tools are. It intentionally avoids the speculative “AI will transform drug development” framing common elsewhere — the goal here is an accurate, operational picture for research administrators and clinical operations staff, not a forecast.
Why this is distinct from AI-in-manuscripts and AI-in-peer-review
CASRAI already covers AI as it intersects with authorship and publishing — see Can AI Be Listed as an Author? and AI Training Data Provenance, Copyright, and TDM Exceptions for Research. Those pages concern AI’s role in producing or reviewing the written outputs of research. This page is about a different question entirely: how AI/ML tools are used operationally to run a clinical trial — before a single manuscript exists. The two topics share a technology but not an audience, a regulatory framework, or an operational context, and treating them as one topic would obscure both.
Patient recruitment and trial matching
Matching individual patients to eligible trials, and screening candidate patients against a specific protocol’s inclusion/exclusion criteria, is one of the most mature applications of AI in trial operations, because it is fundamentally a text- and structured-data-matching problem that machine learning is well suited to. Two real, verifiable examples illustrate the current state:
- NIH-developed trial-matching algorithms. The National Institutes of Health has developed and published on AI systems (including a large-language-model-based tool, TrialGPT, from the NIH Clinical Center) that match candidate volunteers against trial eligibility criteria and generate explanations of why a given patient does or does not meet specific criteria, intended to speed up the labor-intensive manual chart-review step that recruitment coordinators otherwise perform.
- Site-level prescreening tools. Published research (e.g. in JCO Clinical Cancer Informatics) describes real-time, common-data-model-driven AI systems for semiautomated patient prescreening in oncology trials, designed to flag potentially eligible patients from electronic health record data for a human coordinator to review — the AI narrows the pool, it does not make enrollment decisions.
In every credible implementation, these tools sit upstream of, not in place of, the human eligibility determination and informed-consent process. See CASRAI’s Clinical Trial Patient Recruitment guide for the broader recruitment lifecycle this fits into, and Informed Consent for why the human step is not optional regardless of how a candidate was identified.
Site selection and feasibility
Site selection has historically relied on sponsor/CRO relationship history and self-reported feasibility questionnaires, both of which are known to correlate poorly with actual enrollment performance. Predictive-analytics approaches now draw on historical enrollment data, investigator publication and prior-trial-performance records, and regional demographic/epidemiological data to rank candidate sites and investigators by likely enrollment yield and, in some published frameworks, participant demographic representativeness. This is an operational forecasting application — it estimates which sites are statistically likely to perform well — not a replacement for the qualification, training, and monitoring obligations sponsors and CROs already hold under Good Clinical Practice. See Clinical Trial Management System for where site-selection data typically lives operationally.
Protocol design and trial simulation
Simulation-based methods — using statistical and machine-learning models to project how a candidate protocol (a given sample size, endpoint, randomization scheme, or adaptive design rule) would likely perform before the trial is run — are an established, pre-AI-era discipline in biostatistics that AI/ML methods are now extending, particularly for more complex adaptive and Bayesian designs. CASRAI’s dictionary defines this discipline directly: see Simulation of Clinical Trials for the operational definition and how it differs from a live interim analysis. The FDA’s own 2026 draft guidance on Bayesian methods in drug and biologic clinical trials reflects how much of this territory remains a statistical-methodology question with AI as one implementation route among several, not a wholesale replacement of trial biostatistics.
Adverse-event and safety signal detection
Pharmacovigilance is the application area with the longest track record of algorithmic (not just AI-branded) signal detection: disproportionality analysis methods have been applied to spontaneous adverse-event reporting databases like the FDA Adverse Event Reporting System (FAERS) for years. What has changed is the addition of machine-learning and natural-language-processing methods layered on top of that established statistical base — used to mine unstructured data (clinical notes, patient-reported outcomes, literature) alongside structured case reports, and to flag candidate safety signals for human pharmacovigilance reviewers to evaluate. The output in every credible implementation is a prioritized signal for expert review, not an automated causality determination. See CASRAI’s Pharmacovigilance in Clinical Research guide for the full signal-management workflow this fits into, and the Adverse Event (AE) dictionary term and Risk-Based Monitoring (RBM) for the related monitoring discipline these tools increasingly support.
Data monitoring and risk-based monitoring
Risk-based monitoring already uses statistical thresholds and centralized data review to target on-site monitoring effort at sites showing anomalies, rather than the older 100%-source-data-verification model. AI/ML methods extend this by flagging more subtle multivariate anomalies across accumulating trial data (data entry patterns, protocol deviation clustering, site-level outliers) than simple threshold rules can catch on their own. This remains an extension of an established methodology (see Risk-Based Monitoring (RBM)), not a distinct new discipline, and centralized statistical monitoring under ICH E6(R2)/(R3) already anticipates this kind of tool.
The regulatory posture: FDA’s 2025 draft guidance
In January 2025, the FDA released a draft guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products, aimed specifically at AI models used to produce information or data intended to support a regulatory submission’s safety, effectiveness, or quality claims. It was informed by a December 2022 expert workshop, over 800 public comments on a 2023 discussion paper, and CDER’s own experience reviewing more than 500 submissions containing an AI component between 2016 and 2023 — a concrete, verifiable indicator of how established AI use already is in submitted trial data, well ahead of most public discussion of the topic. The guidance’s central mechanism is a seven-step, risk-based credibility assessment framework: sponsors must define the specific regulatory question the AI model addresses and its context of use, assess the model’s risk, execute a credibility assessment plan proportionate to that risk, and document the results. FDA and the European Medicines Agency have also published a set of ten shared principles for AI use in drug development, reflecting an active push toward transatlantic regulatory alignment rather than each agency working from a separate framework. As of this writing the FDA guidance remains in draft form — treat its specific requirements as directional, not yet final, and check FDA’s own guidance documents page for the current status before relying on it for a submission strategy.
What AI in clinical trials is not, currently
To keep this page accurate rather than promotional: current, credible AI applications in trial operations are narrow, assistive, and sit alongside existing regulated processes — they do not currently replace investigator judgment on eligibility, IRB/ethics committee review, human causality assessment for adverse events, or the statistical rigor requirements a protocol is held to under ICH E6 Good Clinical Practice. Claims of AI autonomously running trial arms, making enrollment decisions, or independently adjudicating safety signals are not reflected in current regulatory guidance or in the published applications above, and should be treated skeptically wherever encountered.
Frequently asked questions
Is AI currently used to decide clinical trial eligibility?
No. Published and regulator-facing tools use AI to identify and prioritize candidate patients for a human coordinator or investigator to review against protocol criteria; the eligibility determination itself remains a human clinical judgment, documented through the standard informed-consent and screening process.
Does the FDA require sponsors to disclose AI use in a submission?
FDA’s January 2025 draft guidance recommends sponsors define the AI model’s context of use and document a proportionate credibility assessment when AI-generated information supports a regulatory claim about safety, effectiveness, or quality. It is guidance, not yet a final binding rule as of this writing — confirm current status directly against FDA’s guidance documents database before treating any specific requirement as settled.
How is AI-based adverse-event detection different from existing pharmacovigilance signal detection?
Statistical disproportionality analysis on structured adverse-event databases (like FAERS) predates AI branding by years. What AI/ML methods add is the ability to mine unstructured data sources (clinical notes, literature, patient-reported text) alongside that structured base, and to surface more complex multivariate patterns — the output is still a candidate signal routed to a human reviewer, not an automated determination.
Can AI replace risk-based monitoring statisticians?
No credible current implementation does this. AI/ML anomaly-detection methods extend the same centralized-review model risk-based monitoring already uses under ICH E6(R2)/(R3) — they widen what a statistical monitoring plan can catch, they don’t remove the human review layer the methodology depends on.
