This blog post outlines a framework for designing a decision-grade pilot for evaluating agentic AI systems. It emphasizes the importance of starting with a clear decision, defining autonomy levels, evaluating workflow instead of just responses, and establishing thresholds for success. The author argues that pilot evaluations should be grounded in real tasks and should measure total human effort and cost per successful task, rather than just time saved. The aim is to produce robust evidence that informs whether and how to deploy agentic AI in enterprise settings.