Generative AI has entered software engineering faster than the evidence base could mature. This talk asks where it genuinely delivers — and how its value can be measured beyond adoption rates, across efficiency, quality, autonomy and governance. The argument turns concrete in a real customer project: a twenty-year-old enterprise system, half a million lines, almost no automated tests, where AI agents take over the software development lifecycle and AI-generated tests become the oracle that surfaces defects humans would miss. The same project lays bare the obstacles — capability ceilings, security and data exposure, and unresolved legal questions of ownership and liability. The outlook rests on measurement rather than assumption: adopt by task, tests are the oracle, humans own accountability, and exposure scales with autonomy.