End of Volume Review & Prompt Release Capstone

The head of support operations wants automatic ticket routing live in six weeks. The company in this volume's running example (illustrative, like all its business figures) receives about 20,000 tickets a day. Today a team of agents reads each one, picks a queue and flags the ones that need a senior person, and the backlog grows every Monday. Eleven chapters have produced every piece the router needs: a prompt, a schema, a decoder that cannot break the schema, an evaluation that can block a bad release, and a registry that can roll one back. What they have not produced is the decision: which model, which prompt, at what cost, on what evidence, rolled out how.