Guide

Seven mistakes hosts make with AI guest messaging

The seven ways AI guest messaging goes wrong in practice, from switching on auto-send on day one to letting the AI talk about refunds, and the fix for each.

Updated 8 October 2026 · 5 min read
by the Tokay team, hosts in Bangkok

Key points

  • The tool says it can send on its own, so you let it. " four times in one afternoon.
  • The opposite mistake, and just as common. The host is nervous, keeps everything in draft mode for six months, and ends up with the same inbox work as before plus a subscription.
  • " These are not rules, they are moods, and an AI given a mood will interpret it differently for every guest.
On this page
  1. 1. Switching on auto-send on day one
  2. 2. Never switching on auto-send at all
  3. 3. Vague rules
  4. 4. Letting the AI near money
  5. 5. Two things sending messages
  6. 6. Writing the knowledge base once and forgetting it
  7. 7. Sharing too much, too early
  8. The pattern behind all seven
  9. How Tokay avoids these by default
  10. Questions hosts ask

We have run AI replies on our own apartments in Bangkok since before we built Tokay, and we made most of these mistakes ourselves. None of them are about the software being bad. They are about how hosts set it up, trust it too early or too late, and forget that the inbox is where reviews are won. Here are the seven that come up most, with the fix for each.

1. Switching on auto-send on day one

The tool says it can send on its own, so you let it. Within a week it has told a guest the pool is open when the building closed it for cleaning, and greeted the same guest "Hi Tom!" four times in one afternoon.

The AI is not the problem; the knowledge base is two days old and has holes. The fix is a supervised period. For two weeks, every reply comes to you as a draft. You approve, edit or deny. Each gap you find becomes a fact. Only then do you switch on auto-send, and only for one kind of question at a time: wifi first, check-in times next, directions after that.

2. Never switching on auto-send at all

The opposite mistake, and just as common. The host is nervous, keeps everything in draft mode for six months, and ends up with the same inbox work as before plus a subscription. Drafting saves typing. It does not save the night shift, the 6 AM "we landed, can we come now?" from Mei, or the Chinese-language question you would have put off until lunch.

The fix is the same ladder, run forward. After two weeks, look at the wifi answers. If every one was right, let those send. Keep going one category at a time. Keep refunds, complaints and anything upset on the "always ask me" list forever.

3. Vague rules

"Late checkout: be flexible." "Extra guests: use judgement." These are not rules, they are moods, and an AI given a mood will interpret it differently for every guest. Priya gets a free 3 PM checkout; Lukas, asking the same thing the next day, is told it is not possible.

Write rules as sentences a new assistant could follow: "Late checkout until 1 PM is free if nobody arrives that day. After that, 500 baht, until 4 PM at most. Anything later, ask me." If you cannot write it as a sentence, it belongs on the ask-me list.

4. Letting the AI near money

Refunds, discounts, compensation for a broken aircon, a "sorry about the noise, we will take 20% off". These should never be decided by software, and a good tool will refuse. But hosts sometimes write rules like "offer a small discount if the guest is unhappy", which hands the AI a blank cheque.

The fix: refund and discount conversations always come to you with a draft, and the draft never names a number unless you have set it. Airbnb and Booking.com have their own refund processes, and a casual promise from a bot can turn into a dispute you then have to honour. The page at /resources/refund-and-discount-requests has wording that works.

5. Two things sending messages

Your PMS fires a scheduled "Welcome, here is your door code" at 3 PM. The AI, seeing a new guest conversation, also says hello and sends the door code. The guest now has two welcomes, slightly different, and wonders which one is right.

Choose one sender per message type. If the AI handles conversations, turn off the templates that overlap, or keep templates for scheduled notices only (checkout reminder) and let the AI handle replies. Read a few conversations end to end after you change anything.

6. Writing the knowledge base once and forgetting it

The restaurant you recommend closed in March. The building moved the gym to the 7th floor. Your cleaner now needs 4 hours, not 3, so early check-in is harder. The AI keeps saying what it was told in January.

Set a monthly habit: read ten random conversations, fix anything stale. When something changes in the building, change it once at the building level so every unit updates. A tool with a three-level knowledge base (general, building, apartment) makes this a one-line edit; a tool with flat per-listing notes means editing every unit.

7. Sharing too much, too early

Before a booking is confirmed, a guest asks "which floor is it?" or "what is the exact address?". An AI set up to be helpful answers. Now someone who has not booked knows your floor, your door code pattern or your building's lax security. On platforms it can also break house rules about pre-booking contact.

Write the rule: no exact floor, no door code, no key card details before the booking is confirmed. After confirmation, send them at the right time (most hosts send access details the day before arrival). A good tool enforces this on its own, but write it down anyway.

The pattern behind all seven

Every one of these is a trust problem: trusting too soon, too late, or in the wrong place. The fix is the same shape each time. Write things down as sentences. Supervise until the answers are boring. Widen auto-send one category at a time. Keep money, complaints and anything unusual on the human list for good.

How Tokay avoids these by default

Tokay is built around that ladder. For the first two weeks every reply comes to the host to approve, edit or deny. After that the host switches on what Pam may send on her own, one kind of question at a time. Pam will not promise discounts, exceptions or refunds the host did not agree to, will not give an exact floor before booking, and asks the host when a fact is missing rather than guessing. Knowledge lives at three levels so a building change is one edit. On our own apartments this leaves about two taps a day of guest work.

Questions hosts ask

How do I know when it is safe to switch on auto-send for a category? When two weeks of drafts in that category needed no edits. Wifi and check-in times usually get there first. Early check-in and late checkout take longer because they depend on the calendar and your rule being precise.

Should the AI apologise to an unhappy guest? It should acknowledge the message and tell the guest a person is looking at it, then hand the conversation to you. The apology that could carry a refund or affect a review should be yours.

Can I just use the AI as a translator and write my own replies? You can, and some hosts start that way. It fixes the language gap but not the time gap. Most hosts who try it move to drafts within a few weeks and to supervised auto-send after that.

See it on your own listings

See how Tokay would answer this for your listings.

Twenty minutes, your apartments, your rules. We show you how Tokay answers, where it asks you, and how the cleaner finds out.

Book a demo

Any chance we could check out a bit later tomorrow? Our flight is at 11 PM.

Good news, nobody arrives tomorrow so you can stay until 1 PM at no extra charge. Please leave the key card on the table as usual.

Checked tomorrow's calendarSam updated the cleaning brief