When a guest chat should move from an AI assistant to a person: five triggers and a handover note

A guest chat should move to a person on five triggers: money, health or safety, legal or identity requests, anger, and any fact the assistant cannot confirm. The handover, the moment a named person takes over from the assistant, needs a short written note so the guest never repeats themselves.

At 23:40 a guest writes into the chat: the room is cold, the minibar charge is wrong, and they want to speak to someone. The assistant answers all three politely and quickly. Only one of those three should have been answered by the assistant at all. The room temperature is a fact it can check; the charge is a financial record it must not touch; the request for a person is a signal it must respect. This article gives you the five triggers for an AI chat handover to hotel staff, the role that should receive each one, and a six-line note format so the guest never has to explain the problem twice.

Why does a handover rule come before the assistant?#

Because the rule is hotel policy, not a feature of the tool: it decides which guest messages a person must own, and any assistant can be held to it.

An assistant with no stop rule defaults to answering. A confident wrong answer costs more than a slow right one, because the guest acts on it. If the assistant invents a refund amount or a checkout time, the hotel has to walk back something the guest already relied on.

The rule fits on one page: trigger, who receives it, what the guest is told, how fast a reply comes. Write it before you choose the tool. It survives a change of assistant, a change of channel and a change of staff, and it is the document you hand a new supplier. For more on which actions should require a person in the first place, see which hotel AI actions need a person's approval.

Name roles, not people: front office supervisor on shift, duty manager, reservations. A rule that names an individual fails at 03:00 and on a public holiday. Set the default in one direction: when in doubt, hand over. An assistant that escalates too often is mildly annoying; one that never escalates is a liability.

The rule has to be visible twice, in the assistant's configuration and in the inbox the people actually read. A rule nobody can check is not a rule.

When should a chat about money go to a person?#

As soon as the guest asks for a decision rather than a fact: a refund, a discount, a waiver, a compensation, or a correction to the bill.

Facts the assistant may give: the published rate, the cancellation policy exactly as written, what is currently on the bill, how a payment was taken. Decisions a person must make: refunds, waiving a charge, a free upgrade, splitting a bill between two guests, extending a stay at a different rate, a group price.

The difference matters because a stated number is a commitment. The assistant may quote the policy; it may not grant an exception to the policy. If the assistant says "we will refund the minibar charge", the hotel has promised money, and nobody with authority agreed to it.

Bill corrections are financial records. A correction is made by voiding the original line and posting a reversing entry, not by deleting it, and that needs someone with the authority to do it. A request for an invoice or an electronic document belongs with whoever runs billing, because the document must match the records and the hotel's tax position.

Always say what happens next. "I have passed this to the front office supervisor, who will reply before 10:00" is a complete answer; silence is not.

Which guest messages must never be answered by an assistant?#

Anything involving health, safety or security: there the assistant's only job is to say a person is coming and to reach that person now.

Typical messages: a guest reports an injury, an allergic reaction, a fire alarm, a stranger at the door, a lost key card, a theft, harassment, or that they feel unwell. The assistant gives no medical, legal or security advice, not even cautious advice. Routing is the whole answer.

The line runs between information and judgement. "The pool is open until 22:00" is information. "Is the pool safe for my child" is judgement, and judgement belongs to a person who can see the situation.

Safety triggers bypass the queue. They reach a person on duty, not an inbox that is read in the morning. Keep the assistant's message short and specific: what it has done, who is coming, roughly when. No apology paragraph in front of it.

Log the time the message arrived and the time a person picked it up. That log is the only evidence the rule works when it matters.

When does a legal, identity or data request leave the assistant?#

Immediately, and to a named role: the assistant may explain what the hotel does with data, but it may not act on a request or confirm anything about a person.

In Türkiye every stay involves a payment, a fiscal document (an invoice or e-document), a guest identity notification (KBS) and personal data (KVKK, the personal data protection law). That is why these requests are never routine chat. Requests that must move: a copy of the data held about a guest, a deletion request, anything about a booking that may not belong to the person asking, a request from a third party, a request from a public authority.

The assistant never confirms whether a named person is staying, in which room, or until when. The same answer applies on the phone, in chat and at the desk. Identity documents are not chat material: the assistant should not ask for a scanned ID and should never repeat one back into the conversation.

The receiver is whoever owns data requests, not whoever is nearest the screen. The reply is written once, checked, and kept with the case. Where the detail of the law stops, your own case begins: confirm it with a legal adviser before you write the rule, and write down what they told you. For the contract side of this question, see AI clauses in hotel software contracts: data use and wrong answers.

How do you spot anger, and what happens when the assistant cannot confirm a fact?#

Anger is caught by signals you can count, not by guessing at feelings; an unconfirmed fact is anything the assistant cannot point to in the hotel's own records.

Countable anger signals: the same question asked a third time, a threat to post a review, an explicit request to speak to a person, a message that names a lawyer, a very long message, or one written entirely in capitals. Sentiment scoring is unreliable across languages and cultures. A short, polite message can be the angriest one in the inbox. Use signals you can count, test and explain to staff.

Hand over early rather than apologizing repeatedly. A third apology in a row reads as a wall, and the guest writes a fourth message.

Unconfirmed facts: a rate for a date the system does not hold, availability nobody has checked, a service the hotel does not run itself, a promise made by someone else, a policy the hotel has never written down. The rule is short: if it is not in the source, it is not an answer. The assistant says what it can check and passes the chat on. Group bookings, special arrangements and anything starting with "can you arrange" belong to a person by default, because each one ends in a commitment.

What must a handover note contain?#

Six things in a fixed order, so the person taking over can act without reopening the conversation.

Who: guest name, room or booking reference, language, and the channel the chat is on. What: the guest's own words, quoted once. "Guest unhappy" is not a note; "guest says the minibar was charged twice on 12 August" is. What the assistant already said: the exact commitments, including any number it quoted, so the person does not contradict it.

What was promised: who would reply, by when, and on which channel. Who owns it now: a named role, plus a second name for the next shift, so the note never sits on a desk that has gone home. What happens next: the action, its deadline, and what to do if the guest writes again before then.

Keep sensitive data out. No ID numbers, no card numbers, no health detail beyond what the action needs, and no pasting the whole transcript when four lines will do. For the messaging side, see messaging guests during the stay: consent and clear rules.

How do you measure chats that should have moved?#

Count the misses, not the answers: chats where a trigger was present and no person took over, and chats where the guest had to ask twice.

Keep three numbers weekly: chats with a trigger present, chats handed over, and chats handed over only after the guest repeated the request. Read a sample of chats that stayed with the assistant and check them against the five triggers by hand. The assistant cannot grade its own misses.

Track time to first human reply, not time to first reply. The second number is always good and tells you nothing. Watch for the two failure shapes: escalation too often, where a guest is bounced to a person over a pool towel, and escalation too late, where three messages arrive before a person sees one.

Change one thing at a time and leave it in place for two weeks before judging it, otherwise you cannot tell which change moved the number. The rule is a stop rule as much as a routing rule. If the misses do not fall, the assistant's scope is wrong, not the wording of the prompt. On running a bounded trial before a full rollout, see running an AI pilot in a hotel with a clear stop rule.

What to do next#

Write the five triggers on one page this week, name the role that receives each one, and keep the handover note to six lines. Then count the misses for a month before you change anything else. The rule is the part you own; the assistant is the part you can replace.

At AI Otel, the PMS assistant is built to answer questions about a hotel's own property in plain language, carry out routine actions, and wait for a person's confirmation on anything sensitive. You can see the hotel operating system and its PMS assistant, or request a briefing.