Key points
- Not sharing the door code or exact floor with someone who has not booked yet.
- Not saying "yes" to early check-in when the calendar shows a checkout that morning.
- Not quoting a price for a late checkout unless you have set one.
On this page
Most hosts first meet AI guest messaging as a button that says "generate reply". Press it, text appears, you send it. That is the smallest version. The version that lets you sleep is a pipeline with several steps, and the steps that matter most are the ones that decide not to send. This page walks through the whole thing in plain terms, then lists what a well built system refuses to do.
Step 1: the message arrives
A guest writes on Airbnb, Booking.com or Vrbo. The platform notifies the messaging tool within seconds. Nothing changes on the guest's side; they are still in the Airbnb app.
Good systems wait a short while before acting, because guests often send three messages in a row: "Hi", "we arrive tomorrow", "what time is check-in?". Answering each one separately looks robotic. Treating the burst as one question looks human.
Step 2: understand what is being asked
The AI reads the message and the conversation so far and works out the type of question. Check-in time. Wifi. Directions. A complaint. A request for a discount. An emergency.
It also detects the language. A guest named Priya may write in English, Lukas in German, Mei in Chinese. The reply should come back in the same language without anyone switching anything on.
This classification step is where the first safety decision happens. Some categories are never answered automatically: refunds, discounts, complaints about safety, anything abusive, anything the system cannot classify with confidence.
Step 3: find the facts
The AI looks up what it needs from the knowledge base. A knowledge base is just the facts you have written down, organised so the system can find the right one for the right apartment.
The organisation matters more than people expect. The pool hours are the same for every unit in the Sukhumvit 24 tower, but the wifi password is per apartment, and the "we are a pet-friendly host" line is per host. A three-level structure (general, building, apartment) lets you write the pool hours once and have them apply to every unit in that building. Tools without this make you repeat yourself and drift out of date.
If the fact is not there, the system has two choices: guess, or stop. A system that guesses will tell Tom the hot water switch is in the kitchen when it is behind the bathroom door. A system that stops will ask you, store your answer, and never ask again.
Step 4: draft the reply
With the question type, the language and the facts, the AI writes a reply. The better ones follow rules you set: tone, length, no greeting repeated five times in one conversation, no promise the host did not make.
A few things a draft should get right that are easy to get wrong:
- Not sharing the door code or exact floor with someone who has not booked yet.
- Not saying "yes" to early check-in when the calendar shows a checkout that morning.
- Not quoting a price for a late checkout unless you have set one.
Step 5: check the draft
Serious systems run a second check before anything goes out. This can be a second model reading the draft and asking: does every claim match the knowledge base? Does it promise anything? Is the tone right for an upset guest? Hospitable, for example, describes a second-model check on its autonomous replies (as of October 2026). Tokay does the same with a separate verification pass.
If the check fails, the draft goes to the host instead of the guest.
Step 6: send, or ask the host
Now the gate. The system has a list of question types the host has allowed it to send on its own. If this question is on the list and the checks passed, it sends. If not, it puts the draft in front of the host with approve, edit and deny buttons.
Hosts should be able to widen this list over time, one kind of question at a time. Wifi first. Then check-in times. Then directions. Refund requests, never.
Step 7: learn, without drifting
When you edit a draft, the system should notice. When you answer a question it asked, the fact should go into the knowledge base. What it should not do is silently rewrite your rules because one guest got an exception.
What AI guest messaging should never do
Make things up. If a fact is not on file, the only correct behaviour is to ask the host. "I do not have that information, let me check" is a fine reply to a guest. An invented answer is not.
Promise money. No discounts, refunds, waived fees or free nights unless the host set the rule in advance. Airbnb and Booking.com have their own refund processes, and a casual "sure, we can refund that" from an AI can turn into a dispute.
Act outside the host's rules. A late checkout policy of "yes if no same-day arrival, 500 baht after 1 PM" is a rule the AI can apply. "Use your judgement" is not a rule, and a system that improvises will be wrong in ways you cannot predict.
Do anything that cannot be undone. Changing a price, cancelling a booking, sending a message to a cleaner that commits them to a time. These should be reversible or should wait for a human.
A system that is honest about these limits will sometimes answer slower than one that is not. That is the trade. On the questions that make up most of the inbox, the speed difference is nothing; on the questions that can cost you a review or a refund, the pause is the product.
How Tokay runs this pipeline
Tokay's front desk agent, Pam, follows these steps for every Airbnb, Booking.com and Vrbo message. For the first two weeks every reply comes to the host first. After that the host switches on auto-send one kind of question at a time; anything unusual, upset, sensitive or uncertain still comes to the host. If a fact is missing, Pam asks the host rather than guessing. On our own apartments this leaves about two taps a day of guest work.
Questions hosts ask
How fast does an AI reply compared with a human? Routine questions are usually answered within a minute or two, including the short wait for message bursts. Questions that go to the host for approval take as long as the host takes, which is still faster than the host noticing the message in the first place.
What happens if the guest writes in a language the AI does not know well? A good system translates before classifying, so a Korean "is there a hair dryer" is matched to the same fact as the English version. If confidence is low, it should go to the host rather than send a shaky reply.
Can the AI tell when a guest is angry? Yes, within limits. Tone detection is good enough to route complaints and upset messages to a human. It is not good enough to trust with an apology that could affect a refund, so those drafts should always be approved by the host.