When do you let the AI hit send?

You flip the switch when the drafts stop surprising you — and not everywhere at once. In practice that looks like a week or two of reading every reply before it goes out, a run of drafts you sent without touching, and a short list of message types you keep for yourself no matter how good the drafts get. StayPilot makes the moment concrete: after twenty drafts approved without an edit, it suggests full autopilot, and you decide. Here is what to look at while you approve, which conversations to graduate first, and what should never send on its own.

Draft mode is a test you are running, not a cage

New accounts start draft-only: the AI writes, you press send. The point is not to slow the software down; it is to hand you evidence. Every draft answers, on your own property, the two questions hosts actually have about AI replies: does it know my details, and does it know its limits. So read drafts actively. Three checks per draft, and they take seconds once they are habit: facts — is the code, the time, the policy right, and did it come from your listing and house manual rather than a guess? boundaries — when the message was a complaint or a refund ask, did the draft answer it, or did it correctly hold the conversation for you? voice — would a guest reading this believe the host wrote it? A draft can pass the first check and fail the second, and the second is the one that matters.

Edit the source, not the sentence

When a draft is wrong, the fix that counts is upstream. If it offered the old gate code, your knowledge base holds the old gate code; correct it there and the mistake is gone for every future guest. A draft you fixed by hand is a draft that will be wrong again next month. This is why the streak is the real signal: every unedited approval is a data point saying the knowledge base is right, and every edit is a bug report telling you exactly what to update. Two weeks of edits that keep finding new gaps means keep approving. Two weeks where your only edits are phrasing preferences means you are done.

Graduate by question, not by calendar

Guest messages do not carry equal risk. The wifi password, the parking rules, check-in time, "is the pool heated" — factual, answerable straight from the listing, and low-stakes if phrased imperfectly. These earn autopilot first, and they are most of the volume: in a beach market like Destin, the same handful of gate-code and parking-pass questions arrives every Saturday. Next come the judgment-flavored logistics — early check-in, late checkout, an extra guest — where the draft applies your stated policy and you mostly confirm it. And some categories never graduate: refunds, complaints, damage, anything touching safety or a dispute always routes to you with a draft and the context attached. That is policy, not a setting you could forget to check.

What the number actually means

Twenty unedited sends is a streak, not a timer, and the difference matters. A slow month with three bookings teaches you less than one changeover Saturday; counting approvals instead of days means the suggestion arrives when the evidence exists, whatever your calendar looks like. And autopilot is enabled per conversation, not as one global switch. The confidence bar keeps working after you flip it: a reply the AI cannot ground in your property's details is held as a draft even on full autopilot. So autopilot in practice means the routine majority sends itself in under a minute — including at 2am — while the hard minority still waits for you.

Flipping back is one move

Trust in software should never be a one-way door. You can drop any conversation — or the whole account — back to draft-only instantly, and jump into any thread yourself at any point. A guest who wants a person replies "human" and gets one. Every sent message is logged with what it said and which of your details it used, so a reply that reads wrong is traceable in one look and correctable at the source. The switch is cheap to flip in both directions, which is exactly what makes flipping it forward reasonable.

If you are weighing this for your own listing, skip the faith part: paste your listing into the demo and read the drafts it would write for the questions your guests actually ask. The question was never "do you trust AI?" It is whether these drafts, on this property, keep being right — and draft mode exists so you find that out with your hand on the send button.