One workflow, one month, success gates written on day one. The pilot ends with a verification pack and an objective answer to "did it work?" — not a slide deck and a feeling. No customer data, credentials, or production access are needed to begin.
Name the workflow, the economic buyer, the technical owner, and the security reviewer. Fix the data boundary and the excluded uses. Most importantly: agree the acceptance tests now — pass = pilot worked is defined before any instrumentation exists. The pilot brief builder produces this one-pager on a call.
The policy gate goes in front of the workflow's tool calls; rules are tuned to the buyer's decision rights and autonomy caps. Nothing about the agent itself has to change — the gate governs what it may do, not how it thinks.
Every action — accepted, escalated, or blocked — seals into the chain. The buyer watches their own ledger build on their own workflow. Mid-pilot surprises surface here, while they are cheap.
Day-30 review delivers the verification pack: chain audit, control mapping, a live tamper demonstration, and deterministic replay of contested decisions. Each week-1 acceptance test gets a pass or a fail in front of the people who wrote it.
Production rollout proceeds on evidence: the same gate, the same rules, more workflows. The pilot's receipt chain remains verifiable forever — it is the first month of the production record, not a throwaway.
The pilot ends, and that is a success of the method: a bounded month produced a defensible "no" with receipts showing exactly where. No sunk-cost expansion, no argument about what happened.