Co-Pilot Approvals

StaYantra operates on a Co-Pilot First philosophy. All suggestions from Exubee, your AI teammate—whether adjusting a room rate, scheduling staff, or responding to reviews—appear as Recommendation Cards for human approval.

1. Recommendation Cards

When Exubee generates a recommendation, it appears on your dashboard containing:

• The Action: What Exubee wants to do (e.g., "Raise Deluxe Room rate by 1,200 for tomorrow night" — all amounts are in your hotel's base currency).
• The Rationale: Why Exubee is suggesting this (e.g., "Occupancy is at 88% and there is a festival weekend coming up in your city").
• Approval Actions:
  • Approve: Executes the action immediately.
  • Edit: Lets you change the details (e.g., change the rate hike to 800 instead of 1,200) before executing.
  • Reject: Discards the action. You will be asked for a brief reason to help train Exubee.

The Suggestions Queue — each card shows the proposed action, the reasoning, and Approve / Edit / Reject controls.

---

2. What is a "Confirmed Policy"?

When you approve a recommendation or edit it with a specific reason, the system's learning loop analyzes your behavior:

• If you repeatedly make the same edits or rejections (e.g., "never raise rates above 9,000 for standard rooms"), Exubee will formulate a Confirmed Policy.
• These policies are displayed on your Policies page. Exubee distils new policy suggestions from your decisions once a week (you can also run it any time from the Policies page), and nothing binds until an owner or manager confirms it.
• A confirmed policy that carries a machine-checkable rule — a floor, a ceiling, a required adjustment, a daily movement cap, a minimum dispatch priority — is enforced: the write itself is refused, on every path, and Exubee is told which policy and what bound to work inside. Its card is badged Enforced. So a confirmed "never above 9,000" for a room type genuinely stops a 9,500 rate being written, whether Exubee proposed it, a job ran it, or a manager typed it.
• A confirmed policy written as free text is Advisory: Exubee follows it when it makes a recommendation, but nothing at the write layer checks it. Its card is badged Advisory.
• A confirmed policy whose rule the system recognises but cannot evaluate — a floor with no price in it, a percentage that is not a number — is badged Needs attention, and it fails closed: every write in that action family is refused until you retire the policy or restate it. The refusal names the policy but carries no bound, because there is none to work inside.
• You can review, confirm, veto, or retire a policy at any time. There is no editing: a policy's wording is the evidence for it, so changing your mind means retiring the old one and writing a new one.

A policy moves through real states as it's learned, not just "on" or "off":

• Inferred — Exubee noticed a pattern in your edits or rejections but hasn't locked it in yet.
• Confirmed — the pattern now binds: when it is Enforced, the write tools refuse anything outside it; when it is Advisory, Exubee follows it in every recommendation but nothing at the write layer checks it.
• Vetoed — you reviewed an inferred pattern and rejected it, so Exubee stops proposing it as a policy.
• Manual — a rule you wrote yourself, rather than one Exubee inferred from your behavior.
• Superseded — replaced by a newer policy that covers the same situation.

Exubee also flags two situations for a manager to decide on, rather than acting silently: when a newly inferred pattern contradicts an existing confirmed or manual policy, and when a confirmed policy hasn't been used in a while (stale). Both show up as a decision for you to make, not an automatic change.

---

3. Graduation to Auto-Pilot

The ultimate goal is to save you time. Once you are comfortable with Exubee's suggestions, you can let it run on Auto-Pilot for specific streams (like dynamic pricing or review drafts).

To graduate a stream to Auto-Pilot, all four gates must be met:

1. High Agreement Rate: At least 90% agreement between what Exubee proposed and what you approved, measured over at least 100 real decisions sustained across 60 or more days.
2. No Breaches: Zero guardrail breaches in that window.
3. Non-Negative Measured Outcome: The stream's measured outcome/revenue-lift score for the window must be zero or positive. If there isn't enough outcome data to measure, the stream is not eligible — missing data does not count as a pass.
4. Owner Toggle: Even once qualified, only the Owner can switch a stream to Auto-Pilot — managers and staff cannot. Activation is invite-gated.

[!IMPORTANT] Graduating a stream to Auto-Pilot isn't a self-serve dashboard toggle alone — it also requires a one-time $150 USD activation fee and a signed legal consent (the Concept Form) before the stream goes live.

Demotion back to Co-Pilot is asymmetric — it's quick to trigger and deliberately harder to avoid than graduation is to earn. Any one of these demotes a stream immediately:

• Any single guardrail breach — instant demotion, no grace period.
• 14 consecutive days of negative measured outcome.
• A sustained intervention-rate spike — a rise of 30% or more that continues for 7 more days after you've been warned about it.

[!NOTE] Even on Auto-Pilot, all actions are written to the append-only ai_actions_ledger and can be undone using the Kill Switch or manual ledger reversion.

The Auto-Pilot Safety Switch is always visible on the main dashboard — one click pauses all autonomous actions.

---

4. The Autonomy Report Card

Every month, Stayantra generates an Autonomy Report Card for your hotel — a plain-language scorecard that turns "the AI runs your hotel" from a claim into a checkable number. It currently covers the pricing, marketing, and staff streams; coverage will expand to more streams over time.

For each stream, the report shows:

• Decisions made — how many recommendations Exubee generated in the period.
• Approved / edited / rejected — how you responded to each one.
• Measured revenue lift — the estimated impact of Exubee's decisions, shown with honest uncertainty bands rather than a single confident number.
• Guardrail breaches — any hard limits Exubee hit or crossed.
• Intervention rate — how often you had to step in and change or reject a recommendation.

This is the same evidence used to decide when a stream is ready to graduate to Auto-Pilot, so the Report Card doubles as a running preview of how close each stream is to qualifying.