Lesson 6: Confirmed Policies — teaching the AI your preferences
Every time you approve, edit, or reject a recommendation, you are teaching the system. When a preference shows up consistently, it becomes a Confirmed Policy — a remembered rule the bees apply from then on, so you stop having to correct the same thing.
How the AI learns
- Approving a card as-is is a quiet "yes, more like this."
- Editing a card is the richest signal: it shows the AI your exact preferred outcome, not just a thumbs-down. (This is why Lesson 7 tells you to edit rather than silently reject.)
- Rejecting with a short reason tells it what to avoid and why.
Repeat a pattern — always trimming the rate hike, always routing a certain complaint a certain way — and the learning loop proposes a Confirmed Policy capturing it.
Phrasing something so it sticks
If you already know a rule you want Exubee to follow, say it in plain, absolute terms and give the reason. Note that only rules expressed as a price floor/ceiling, a relative price adjustment, a daily-change cap, or a dispatch priority minimum are enforced at the write layer — the rest become Advisory:
- "Never price Standard rooms above 9,000 — we lose our regulars."
- "Always inspect a VIP room twice before it's marked sellable."
- "For corporate bookings, don't send marketing SMS — their travel desks handle it."
Words like always, never, and for [this group] signal a hard boundary rather than a one-time request. Compare:
One-time: "Don't raise Standard above 9,000 this weekend."
Policy: "As a rule, never raise Standard above 9,000."
Reviewing and retiring policies
Confirmed policies live on your Policies page. You can read every active policy, confirm or veto an inferred one, and retire any of them when your hotel changes — a new season, a renovation, a new room class. There is no editing: retire the old policy and write a new one, so the record of what you decided and when stays honest.
Each card is badged. Enforced means the policy carries a machine-checkable rule and the system refuses a write that breaks it. Advisory means it is guidance the AI follows when recommending, with nothing enforcing it. A third badge, Needs attention, means the policy carries a rule the system recognises but cannot evaluate — so it fails closed: every write in that action family is refused until you retire the policy or restate it. Both kinds matter, and both go stale: an out-of-date "never below 2,000" floor can quietly cost you fill on a dead week — and if it is Enforced, it will cost you that fill whether or not anyone is watching.