Evidence-to-decision: From exposome data to evidence to action through agentic AI

Public health decisions are uniquely difficult, weighing population benefits against harms, equity, resource constraints, and feasibility, often under deep uncertainty. The GRADE Evidence-to- Decision (EtD) framework, rooted in evidence-based medicine, offers a transparent route from evidence to action through twelve explicit criteria, and has recently been adapted for environmental and occupational health. A Human Exposome Project would generate evidence of a volume and complexity that breaks the manual assumptions on which EtD was built. I argue that agentic artificial intelligence, autonomous agents orchestrating multi-step scientific workflows, can operationalize each EtD criterion and make exposome-scale decision-making feasible, but only if it inherits the rigor of the evidence-based disciplines it is asked to accelerate. Six families of agents (evidence extraction, risk-of-bias assessment, uncertainty quantification, causality reasoning, cost-outcome analysis, and post-deployment validation) map cleanly onto the EtD criteria. Five governance requirements (traceability, versioning, context-of-use benchmarking, honest uncertainty, and human accountability) separate an evidence engine from a confident hallucination machine. The exposome demands nothing less. Plain language summaryPublic health authorities must decide which environmental exposures to act on, often with imperfect evidence and limited resources. A widely used framework called GRADE Evidence-to-Decision gives them a structured way to combine evidence with cost, equity, and feasibility considerations. If a Human Exposome Project succeeds in mapping every environmental influence on health across the lifespan, no human team can read and weigh the resulting evidence by hand. This article argues that artificial intelligence systems of coordinated specialist agents could carry that load, with one type of agent for each kind of decision input. To be trustworthy, those systems must follow the same rules as the science they support: every claim traceable to its source, uncertainty stated honestly, biases rec­ognized, and humans accountable for the final call. The article describes how such a system could be built, what to test before deployment, and which failure modes to avoid.