A panel discussion at Swiss Biotech Day explored what it takes for agentic AI to operate reliably in clinical trials. The core argument is that the bottleneck is no longer AI capability but trust — specifically, the data infrastructure, evaluation rigor, and explainability required in regulated environments. Vector-based RAG architectures fall short because clinical operations require reasoning across interconnected entities (sites, investigators, protocols, outcomes). Graph-native reasoning with temporal dynamics is presented as the more credible architectural approach. The post also covers evaluation metrics (Precision@K, groundedness scoring, human annotation) and references PSI CRO's SYNETIC system as a production example of trusted clinical AI that reduced site selection from weeks to minutes.
