A framework for understanding AI tool adoption in engineering organizations, broken into five stages: inline completion, single-file agentic editing, multi-file planning, concurrent agents, and fully automated workflows. Based on telemetry from ~3 million developers, the post identifies what metrics matter at each stage (feature depth, workflow penetration, practice retention), what causes adoption stalls (usage caps, approval friction, policy constraints), and how to instrument rollouts using an AI Adoption Score. The post also argues for centralized governance with decentralized model selection, noting that heavy users average 7+ providers.
Table of contents
Five stagesGovernance from the top, selection from the bottomWhat to measureWhat can halt your climb rateNone of this works without observabilityWhere this lands271 Impressions