The four that matter most are: (1) the share of AI-generated work that reaches production without rework; (2) how often engineers override or redo agent decisions; (3) compliance and security findings attributable to AI-assisted changes; and (4) idea-to-production lead time for features initiated by AI-assisted discovery or BA work. These are leading and lagging indicators of whether AI adoption is improving the delivery system or just making it faster to generate unshippable work.