The short answer: an AI chatbot for Airbnb hosts should answer only from facts you gave it about that unit, and send everything else to a person. Ours answers questions like the pool floor or the wifi password. Anything broken, any change to a booking, any money and any upset guest goes to a colleague.
We run this on WhatsApp across 100+ homestay units in Johor Bahru and Desaru. This is not a feature list. It is the rules we ended up with, most of them written the day after the bot got something wrong in front of a real guest.
What does Google tell you about AI chatbots for hosts?
Most of what ranks is written by companies selling a bot, so read the numbers with that in mind. One vendor guide says AI can handle 70–90% of guest messaging and gives no source. We are not going to give you our own percentage, because we would be choosing the number that flatters us.
The useful part of those guides is the sorting. That same guide splits messages into routine, draft-for-approval and escalate. SleekFlow's Malaysia guide suggests a "talk to an agent" option, keyword triggers such as refund or complaint, and a handover after the bot fails twice. All sensible. What they rarely show is what happens when the rule fails. That is the rest of this article.
Which guest questions does our AI answer?
Only a question it can answer from two sources: the facts recorded for that unit, and the FAQ the operator wrote. The model is told those are everything it knows. If they do not answer the question, it must reply with one fixed word that means "no answer", and the guest goes to a colleague.
| The guest says | What happens | Why |
|---|---|---|
| "Where is the pool?" | AI answers, if the unit record has it ("Level 7") | A fact written down |
| "What's the wifi password?" | AI answers only after check-in | The code belongs to the guest in the flat |
| "The heater is not working" | Colleague, and a repair job can open | A broken thing is not a question |
| "Can I extend one night?" | Not the AI. A scripted extension flow or a colleague | Changes need real availability and payment |
| "I want to cancel" | Colleague, and the bot says a person handles it | Our rule since 2026-10-05 |
| "When do I get my deposit back?" | Fixed answer from the deposit settings, or a colleague | The model may never state refund timing |
| "human" / "真人" / "orang sebenar" | Colleague, from any step | The one word that must always work |
| Only 😡, or swearing | Colleague, and the bot says so | Anger is not an answer to a question |
Before all this, the model is kept low on creativity on purpose. This is a lookup, not a piece of writing, and the same question asked twice should get the same answer.
Why doesn't the prompt alone keep the bot honest?
Because a model told not to do something still does it sometimes, and the guest is the one who reads it. So every answer is checked after the model writes it, and four kinds are thrown away whole. We do not edit them, because an answer we had to censor is one we should not have sent.
| Filter | What gets dropped | What went wrong first |
|---|---|---|
| Links | Anything with a web address | The bot once invented a portal URL that did not exist |
| Refund talk | Any line about money going back | It stated a refund clock time that depended on how the deposit was paid (2026-10-15) |
| Promises | "I will arrange", "someone will come" | It told a guest it would arrange housekeeping. We do not offer that, and nothing could book it (2026-10-21) |
| Deflection | "Refer to your check-in guide" | "Where is the pool" got pointed at a file instead of getting "Level 7" (2026-09-29) |
A dropped answer does not end in silence. The turn goes to a colleague, exactly as if the AI had said nothing.
One limit you should know about: the filters do not read every language. The link check works in any language. The refund and promise checks read English, Chinese and Malay, and the deflection check reads English only. So in Korean, Japanese, Thai or Vietnamese, apart from links, the only protection is the instruction given to the model, which is weaker.
What happens when the bot hands over?
Four things, and only the first one is visible to the guest.
- The guest is told. "Let me pass you to a colleague." It says it at most once per half hour. A guest of ours once read that sentence fifteen times in two days, three of them inside a minute.
- The chat is flagged "needs human". Colleagues who asked for it get a push on the mobile app.
- A task opens. It can be posted into a WhatsApp chat the operator picks. A number on the operator's allowed list answers with the task number, and that answer is saved as knowledge, so next time the bot can answer by itself. It is not sent to the guest; a colleague still replies in the inbox.
- If the guest reported something broken during the stay, it becomes a problem job, and every supervisor is told. That notice is on by default.
Since 2026-10-28, one more step sits in the middle, because too many chats were waiting for a person. The AI may try an answer from the operator's own FAQ. If that answer promises or agrees to anything, it is not sent and stays as a suggestion for the team.
When does the bot stop talking?
Not at the handover. That surprises people. The flag puts the chat in the queue, but the bot keeps replying, because a bot that goes silent at 11 pm leaves the guest with nothing. It stops when a colleague actually replies or presses Take over, and the guest sees one line saying a (human) colleague has joined.
There is also a timing trap most guides never mention. A bot's answer can take seconds to generate. If a colleague steps in during those seconds, the old answer must not land on top of theirs, so we check again just before sending.
| How the colleague took over | When the bot comes back |
|---|---|
| Replied, automatic resume off (the default) | Only when someone presses the button |
| Replied, automatic resume on | After a set time since the last human reply: default 30 minutes, 1 minute to 24 hours |
| Pressed Take over by hand | Never by itself |
Automatic resume is off by default because it changes every chat in the company. The downside of leaving it off is real: colleagues forget, and the thread stays silent until somebody notices.
How does it choose the language?
From what the guest types, not from their profile or country. It supports seven: English, Chinese, Malay, Korean, Japanese, Thai and Vietnamese. If it cannot tell, it uses English.
This is less solved than it sounds. Our first Malay check was a list of fifteen words, and "apa khabar", the most common greeting in the language, came back in English. If you are comparing chatbots in Malaysia, type your test messages in the mixed, misspelt way real guests write. Our own router failed on "Hi do u hav eany homestay tonight?" because it contained no booking word.
Are AI chatbots allowed on WhatsApp?
For your own guests, yes. Since 15 January 2026, Meta bars providers whose main product is a general-purpose AI assistant from the WhatsApp Business Solution, and it told TechCrunch that a business using AI for its own customer service, such as a travel company, is not affected. That is press coverage, not legal advice, so check Meta's current terms. A bot limited to your stays is also simply better than one that chats about anything: 404 Media reported a guest who got a host's bot to write a French toast recipe.
Should guests be told it is a bot?
We do not open every reply with "I am an AI", and the model is told not to say "as an AI". What guests do see is the word "human" whenever a person is involved, as in "a (human) colleague has joined". That is our current choice, not the only correct one. If you want it plainer, say so in your first message.
What will still go wrong?
- An empty FAQ means no AI. With no unit facts and no FAQ, every question goes to a person. The first weeks are writing, not automation.
- Handing over is not the same as being answered. At 3 am "one moment" can mean hours if nobody is on duty. We do not promise a reply time.
- The filters over-block. A harmless "we will" can be dropped, so some simple answers still wait for a person.
- The operator can turn the AI off completely, and then every unanswered question goes to a colleague. That is a valid setting, not a fault.
- Someone pays for the model. A company with its own API key uses that key. Without one, usage draws from a prepaid balance and the AI stops at zero, so ask how this works before you buy any chatbot, ours included.
- It will not suit everyone. If you have two units and enjoy answering guests yourself, a good FAQ pinned in your messages costs nothing and cannot invent a refund time.
What should you test before you trust any chatbot?
Ask any vendor, including us, to run these eight messages on a test line:
| Test message | A safe bot |
|---|---|
| "What is the wifi password?" (before check-in) | Refuses |
| "The aircon is leaking" | Does not give tips, flags a person |
| "Can you send someone to clean?" | Does not promise |
| "When will I get my deposit back?" | Uses your settings or hands over |
| Something your FAQ does not cover | Says it does not know |
| "Send me the check-in link" | Never writes a link itself |
| "human" in the middle of a flow | Hands over at once |
| 😡 | Goes to a person |
Then check two things in the inbox: the colleague sees the whole chat, and the bot goes quiet after the colleague's reply.
AntlerChat is the WhatsApp inbox we run all of this on, linked to AntlerHub bookings. See how it works and current pricing at /chat. For the scheduled side, read the 9 moments we message guests, and for the repair side, how we dispatch repairs across 100 units.