Data quality and practical limits of AI in drug discovery

Multiple sources emphasize that workflow design, not technology, is the real bottleneck. Generative AI costs and hype cycles obscure practical limits. Concrete examples: 66.9% smoking data recovery, 85% IBD flare prediction, and 18-month pancreatic cancer lead time when data is right. New articles highlight wet-lab bottlenecks (Twist), dead time elimination (PhaseV), and infrastructure deals (Sage/Causaly). New: GSK-Relation $110M deal, Relation's automated data manufacturing, and clinical trial AI agent efficiency gains (data mapping from 8-12 weeks to 2-3 weeks) with 'glass box governance' framework.