How do you know the answer was right, in a language nobody on the desk reads?

The problem is not translation, it is verification

Modern models answer fluently in all three languages this platform serves. Fluency is not the risk. The risk is that a confident, well-written wrong answer in a language the duty manager cannot read looks exactly like a correct one, and will not be caught at handover, in a spot check, or in a complaint until the guest arrives expecting something nobody agreed to.

What makes it checkable, and it is not a bigger model

Three things. The answer should cite the published source it came from, so a manager can check the source even without the language. Conversations should be readable in a log with the question and answer side by side. And the assistant should be scoped so tightly to published facts that there is little room for a wrong answer in any language - a restriction is worth more here than a capability.

Restrict the commitments, not the languages

The instinct is to disable the languages nobody reads. That is the wrong cut: it loses exactly the guests the hotel struggles to serve. The right cut is by subject. Rates, upgrades, late checkout, cancellation outcomes and anything a guest could call a promise go to a person in every language. Published facts - breakfast times, parking, how to reach the property - are safe in all of them.

Spot-check it the way you would a new night porter

Take ten real conversations a week in the languages nobody reads and have them read by someone who does - a colleague, a partner agency, anyone. Ten a week is enough to catch a systematic problem and cheap enough to actually happen. A hotel that never does this is not running multilingual support; it is hoping.