Your operations are the AI's real training data

The FAQ is the easy part of a hotel AI project, and it is also the part your guests could already find.

Most hotels start by collecting FAQs. Someone exports the website's questions page, someone else adds the ten things reception gets asked daily, and that becomes "the knowledge".

What makes an assistant useful is something nobody has ever written down: how your hotel actually runs.

Examples from real properties

Housekeeping requests at one property go to a single manager's WhatsApp, Sunday to Thursday, 07:00 to 15:30. Outside those hours they go to reception. A guest asking for extra towels at nine in the evening is not an unusual request. But sending it to the manager's phone means it is read the following morning.

A chain answers all pre-arrival questions from one central number. After check-in, the guest gets their specific hotel's direct number for anything on site. Centralised before arrival, local during the stay. An assistant that does not know this either sends a pre-arrival enquiry to a hotel that has not seen the booking, or sends an in-stay request to a central team who cannot walk upstairs.

A "how is everything?" message goes out two to three hours after check-in, triggered by the check-in status in the property system rather than by a clock. That detail matters: business guests who arrive late at night are frequently checked in by hand, sometimes hours after they actually arrived. They get a different message, because "how is everything?" at 02:00 is not hospitality.

A checkout reminder goes out only at weekends, when checkout is busiest and the car park is the real problem. On a quiet Tuesday it is noise.

Vouchers: the assistant asks where the voucher was bought and which package it covers, so reception has the full picture before they reply. Without those two questions, the hand-off produces a conversation that takes four messages to get to the point.

None of this is on your website. All of it decides whether the guest is served well.

Why FAQs are the wrong starting point

An FAQ answers "what is true?". Operations answer "what happens next, and who does it?".

Almost every enquiry that matters is the second kind. "Can I have a late checkout?" is not a fact question; it is a question about who may approve it, at what time of day, at which of your properties, and whether today is a departure-heavy Sunday. Load only the FAQ and the assistant will answer confidently and be wrong in a way that costs you.

The questions to sit down and answer

Block out ninety minutes with the people who actually do the work, and write down:

Who handles what, and when. By team, by hour, by day. Include the out-of-hours fallback for every single route, because that is where the gaps live. If housekeeping finishes at 15:30, something must catch the 21:00 request.

Which system events should trigger a message. Check-in status, checkout, booking confirmation, a payment taken. Triggering on an event rather than a time is what separates a message that lands well from one that arrives at the wrong moment.

Which guests need different treatment. Business travellers, groups, travel agents, OTA bookers, returning guests on a rate you negotiated in 2019. Each of these breaks at least one of your standard rules.

What must never be handled automatically. Complaints, refunds, invoices, anything touching a card. Write this list before you write anything else.

Seasonal and holiday variation. Pool hours change, a restaurant closes midweek in February, half of the above stops applying on a bank holiday. If your assistant cannot be told about that, you will be correcting it every spring.

Start with people, then switch on AI

The hotels that launch best run about two weeks of staff answering by hand inside the same inbox the AI will use.

This sounds like a delay. It is the opposite. In those two weeks you find out what guests actually ask, in the words they actually use, which are rarely the words in your FAQ. You find out which questions your own team disagrees about. You find out where hand-offs really go, as opposed to where the org chart says they go.

Launch day then feels like a colleague stepping in, rather than a stranger being handed the phone.

Two weeks is also long enough to cover a full weekly cycle twice, including whichever day is your awkward one.

Keep it sharp after launch

After launch, staff mark weak answers with a thumbs down and correct them. The correction becomes part of the knowledge, so the same mistake does not come back.

Make that somebody's job rather than everybody's. Fifteen minutes a week, one named person, reviewing the flagged answers and the hand-offs that should not have been hand-offs. Properties that do this have a noticeably better assistant three months in. Properties that treat launch as the finish line have the assistant they launched with, slowly going out of date.

Checklist

1. Who handles what, and when: by team, by hour, by day, including out of hours.
2. Which system events should trigger a message, and which guests are exceptions.
3. Which guests need different treatment: business, groups, agents.
4. What must never be automated.
5. Two weeks of manual replies in the real inbox before you switch on auto-reply.
6. A named person, fifteen minutes a week, reviewing flagged answers.

Takeaway

An off-the-shelf bot answers questions. A custom one works the way your hotel already works. The difference is not the software. It is whether anybody wrote your operations down.

Two documents sit alongside that one: the point at which the assistant hands a guest to a person, and the tone and house rules it answers in.