Every large pharma organization is somewhere on the same arc right now: a promising AI pilot, real enthusiasm from the data science team, and then a slow stall when someone asks how it gets into production. The model works. The data works, more or less. But the surrounding environment does not, and no one quite planned for that. This paper argues that the problem is not the AI. It is the operational harness that should have been built around it.
A landmark architectural analysis of Anthropic’s Claude Code found that only 1.6 percent of a production-grade AI system is actual decision logic. The remaining 98.4 percent is the infrastructure that makes that intelligence reliable, auditable, and safe to run at scale: permission structures, context management, tool routing, audit, and recovery logic. In pharmaceutical manufacturing, that ratio is even more demanding. The 98 percent has to enforce 21 CFR Part 11 and EU GMP Annex 11 compliance, maintain unbroken audit trails, manage instrument connectivity and sample lineage, and govern the boundary between probabilistic AI reasoning and deterministic GxP execution. An AI agent that cannot operate within those constraints is not a tool for pharmaceutical operations. It is a liability.
Drawing on Anthropic’s architectural research, Gartner’s projections on agentic AI cancellation rates, and the ISPE GAMP framework for AI validation, Vasu Rangadass, Ph.D., President and CEO of L7 Informatics, makes the case that the organizations that will win pharma’s agentic era are not the ones with the most sophisticated models. They are the ones who have the most capable operational harness.
To access this resource, please make sure to accept cookies.