Most AI pilots fail because they optimize for demos rather than operational conditions. A useful pilot starts from a narrow workflow, clear ownership, and explicit failure boundaries.

Define quality criteria before building: factuality, latency, cost, refusal behavior, source coverage, and escalation behavior. These criteria become the release gate.

A production-shaped pilot ends with a go/no-go decision and a known operating owner. If the metrics pass, the architecture and handoff model should already exist.