As pharmaceutical R&D embraces multi-agent AI systems, the bottleneck is shifting. The challenge is no longer whether AI can reason over scientific literature — it's whether the data feeding these systems is fit for purpose. Agentic workflows that span internal datasets, external knowledge bases, and third-party sources introduce compounding risks: lost provenance, conflicting ontologies, and results that can't be reproduced. As these systems grow in complexity and tool access, their performance becomes harder to predict and harder to trust. This talk will explore what a reliable data foundation for agentic research workflows looks like, and how the relationship between data infrastructure and AI capability is evolving.