Artificial intelligence in clinical trials-state of the evidence, gaps, and next steps

Artificial intelligence (AI) affects clinical trials in two distinct but overlapping ways: as the intervention under evaluation and as infrastructure supporting trial design, recruitment, monitoring, endpoint assessment, analysis, and reporting. In this manuscript, we define AI-as-intervention as AI whose output is itself part of the assigned clinical intervention being evaluated for its effect on participant care or outcomes, and AI-for-trial-operations as AI used to support trial design, conduct, or analysis without itself constituting the treatment under study. This is an important distinction because AI-as-intervention generally requires prospective clinical evaluation with prespecified estimands, prospectively governed model behaviour, and protocol-level oversight, whereas AI-for-trial-operations is often judged by workflow accuracy, impact of the decision, transportability, and safety under real-world constraints, although some uses, such as endpoint support or inferential modelling, may also require similarly explicit change control and oversight. Using this distinction as an analytical framework rather than as a division into separate parts, we examine how AI can improve each phase of the clinical-trial lifecycle, what evidence currently supports these applications, what limitations constrain their validity and transportability, and what methodological and governance safeguards are required. We synthesise evidence across safety, efficacy, operational risk prediction, network medicine, digital health technologies, retrieval-augmented generation, and agentic workflows, while aligning the discussion to ICH E6(R3), ICH E9(R1), and emerging structured-protocol standards.

AAA was partly supported by the Institute of Precision Medicine (17UNPG33840017) from the AHA, the RICBAC Foundation, and NIH grants R01HL173935-01, 1 R01 HL161008-01.