A practical guide to designing an AI agent pilot that produces usable evidence: clear scope, baseline metrics, human fallback, stop rules, and a real buy-or-kill decision at the end.