Introduction
A self check-in property has no front desk. So when it's 10 p.m. and a guest can't figure out the washing machine, there's nobody to ask. Either the operator's phone rings, or the guest gives up. Neither outcome is a good one.
This article walks through the AI assistant we built into the guest-only page: a 24/7, four-language companion for unstaffed operation that answers everything from how to run the appliances to where to eat nearby.
Deciding where it lives, first
Not on the public website
When people picture an AI chat, they usually picture a bubble in the bottom-right corner of the public website, open to anyone. We did the opposite: the assistant lives only inside the guest-only page.
Anyone can use it, without limit. API costs pile up on questions that have nothing to do with the property, and it becomes an easy target for abuse.
Only guests with a booking can use it. Questions naturally stay within the scope of the stay.
Once we decided it was for staying guests only, the scope of what it needs to answer became clear. It can focus on "our appliances, our facilities, our neighborhood," and politely decline everything else.
Controlling when it becomes available
We also made the assistant available only from noon, three days before check-in. If it were usable the moment a booking was made, we'd get a lot of questions that aren't relevant yet.
The important part is that this restriction isn't handled by the screen alone. Hiding a button doesn't help if someone reaches the endpoint directly. The server that actually does the work runs the same check, and simply doesn't respond outside the window.
Don't hand it straight to the AI
Three narrowing steps come first
Opening the chat doesn't drop you into a free-text box. You get choices first.
Appliances / check-in / facilities / house rules / the area / emergencies
Choose from the list of appliances in the property — washing machine, dishwasher, oven, and so on
"Basic operation," "How to clean it," and other prepared answers appear instantly
If none of that solved it, type a question. This is the first point where the AI actually runs
The key point: the first three steps never call the AI at all. The answers come straight from the database, exactly as they were written.
That means the most common questions are resolved at zero cost, with essentially zero wait, and with no chance of a wrong answer. The AI is only needed for the questions we didn't anticipate.
Adding a deliberate pause
Here's a fun detail: even the answers we can return instantly from the database get an intentional pause of about 0.7 seconds.
An answer that snaps back instantly feels mechanical and cold. A brief moment of "thinking" reads more like a conversation. Faster isn't always better — this is a small example of that.
Answering with evidence
Only what the documents actually say
The scariest thing about AI is a confident, plausible lie. "Yes, this washing machine has a dryer function" is a real problem if it doesn't.
So we made the appliance manuals and facility information searchable in advance, and built the assistant to find the passages relevant to the question first, then compose an answer using only that material. This is what's known as RAG, or retrieval-augmented generation.
If the search turns up nothing, we don't call the AI at all — we reply that we can't answer that. Being able to say "I don't know" is what earns trust in the long run. The mechanics are covered in "Turning Appliance Manuals into Knowledge."
Label the source, never blend them
When suggesting places nearby, we present two clearly separated sources: places the property has personally checked and recommends, and results from a map service.
Never blend 'our picks' with 'search results'
Listing both together makes it look like the property endorses all of them. We label the search results explicitly as places the property has not verified or endorsed, which heads off the misunderstanding.
Being careful with on-screen labels
If an appliance's control panel is labeled in Japanese, answering an English speaker with a pure translation — "select white rice" — leaves them staring at a panel with no matching text.
So the AI's instructions include a rule: keep the Japanese exactly as it appears on the panel, and add the translation in parentheses. The test is whether an overseas guest can actually find the button.
Keeping quality equal across all four languages
Japanese, English, Traditional Chinese, and Korean — all of them have complete coverage: interface text, cautions, and even the wording used to decline a question.
Refusals and emergency guidance in particular are never generated by the AI; they use fixed text written by a human in advance. Anything safety-related shouldn't be composed on the fly. That decision is explained in detail in "The Art of What AI Must Not Say."
Wrapping up
Three principles anchored the design of this AI assistant.
- Make it guest-only — keep it off the public site and narrow the scope of what it must answer
- Put three steps before the AI — solve common questions with fixed answers, cutting both cost and errors
- Ground it in documents — use only what the search finds, and answer honestly when it finds nothing
Turning Appliance Manuals into Knowledge — Building the RAG Database
How manual PDFs become searchable knowledge.
The Art of What AI Must Not Say — Guardrails and Prompt Design
Emergency routing and control over off-limits topics.
Storing No Conversation Logs, and Keeping Costs Down
How we operate the assistant with privacy intact.