Lesson 7: Co-Pilot approvals & Auto-Pilot graduation
This is where the daily habit and the long-term payoff meet: how to work the approval queue well, and how a task stream earns the right to run on its own.
Reading a recommendation card
Every card shows you three things:
- The action — exactly what the bee wants to do ("raise Deluxe by NPR 1,200 for tomorrow").
- The numbers — the impact, so you can sanity-check it at a glance.
- The reasoning — why, in plain language ("88% occupancy and a festival in Janakpur").
You can edit the values in place — change 1,200 to 800 — before approving. You are never forced into an all-or-nothing yes/no.
A live recommendation card — the bee's proposed rate, its rationale, and your Approve / Edit / Reject controls.
Edit, don't silently reject
When a recommendation is close but not right, edit it. An edit teaches the AI your exact preference (Lesson 6) and moves a stream toward graduation. A silent reject only says "no" — it withholds the very signal that would stop the AI making the same suggestion again. Reserve rejection for recommendations that are simply wrong, and add a one-line reason when you do.
How a stream graduates to Auto-Pilot
Auto-Pilot is granted per stream (pricing, review drafts, dispatch) and only when the evidence is overwhelming. A stream must clear all of these:
- ≥ 90% agreement between what the AI proposed and what you actually approved,
- across at least 100 real decisions,
- sustained over at least 60 days,
- with zero guardrail breaches in that window.
Miss any one and the stream stays in Co-Pilot. A single guardrail breach demotes it instantly and notifies the owner. Even once qualified, activation is owner-only and invite-gated — managers and staff cannot switch it on.
You are never locked in
Auto-Pilot is always reversible. You can pull any stream back to Co-Pilot whenever you want to resume hands-on oversight, and the Kill Switch halts everything at once. Graduation is a convenience the AI earns — never a loss of your control.