What actually fits in a two-week pilot

One process, one number

A pilot that touches three processes proves nothing about any of them. We pick the one process whose number you already watch weekly.

Four rules

1. Real data or no pilot. Sample data proves the demo, not the agent.
2. A human approval on anything outbound. From day one, not as a later hardening step.
3. Read-only until the numbers agree. The agent proposes; the operator commits.
4. A written exit. If week two misses the baseline, we say so first.

Why fourteen days

Our revenue sits on the far side of the result. A slow pilot costs us, not you — which is exactly the incentive you want us to have.