Miller's chapter on battle-tested AI workflows identifies the pattern that separates AI deployments that fail in production from the ones that work reliably. The failure almost always happens in the same place: a technically impressive AI model without a well-designed workflow around it. A workflow defines how the AI fits into a real process — who inputs what, how errors are caught, what humans do with the AI output, how edge cases are handled, and how the system is monitored over time. An AI model without a good workflow is a powerful engine without a chassis. It runs impressively in demos and fails in practice when the inputs are messier than the test data, the users are less technical than expected, and the edge cases are more common than anticipated. Build the workflow first. Integrate the model into it.