First level: a draft, not a send button
The simplest way to avoid a public mistake is not to let the assistant send anything on its own. The assistant prepares a reply using data from the hotel PMS, and the team reads and confirms it with one click. You still save time because reading and approving takes much less effort than finding the information and writing from scratch.
Second level: boundaries by category, with an objective
Automation is not switched on all at once. It is decided by enquiry category. General information, quotes and groups each receive a chosen level of autonomy and, above all, an objective for the assistant to follow. One real example is: “No student groups, but reply to agencies.” Anyone outside the boundary receives a polite, reasoned refusal rather than silence.
Third level: everything remains recorded
Every conversation remains in the console and can be read in full: what the guest asked, what the assistant replied and who approved it. If a reply does not convince you, you can see it, correct it and turn that correction into the rule for next time. Control is not a switch. It means always being able to reconstruct what happened.
The question is not whether AI will ever make a mistake. It is what happens immediately afterwards, and who notices.
Where automation makes sense from the start
The risk is low and the value immediate for stable questions such as opening times, parking, pets and wifi. In the room, the concierge replies automatically at any hour and only hands over to staff when a person is needed. At 23:00, no guest wants to hear “we will ask reception tomorrow”. Where a decision is involved (special rates, exceptions or complaints), confirmation remains human until the results convince you.
How to start in practice
Start cautiously: everything remains in draft, and the team reads and approves it. After a few weeks, you can see which categories the assistant handles well and widen its autonomy one category at a time. At active properties, 40-70% of enquiries are generally handled autonomously. The range depends on how much you decide to delegate, not on a technical limit.