A voice guide for AI-drafted hotel messages: tone, words to keep and words to avoid

A brand voice guide for hotel AI messages is a one-page sheet of decisions a hotel pastes into whatever AI assistant it already uses: the tone for each situation, the facts a draft may not change, the promises it may not make, and the check a person runs before sending. The assistant drafts; a person sends.

A guest writes at 23:40 that the room is too noisy to sleep. Someone on the night shift asks an AI assistant for a reply, gets three warm paragraphs in eight seconds, and sends them. The paragraphs are polite, friendly and wrong: they promise a room move the hotel cannot make tonight. After reading this guide you will be able to build the one-page sheet that prevents exactly that: tone by situation, fixed facts, forbidden promises, and the three questions a person asks before sending.

Why do AI drafts drift without a style sheet?#

Because the assistant fills every gap you leave with the average of what it has read: generic hospitality warmth, hedged promises and adjectives that commit to nothing. A style sheet closes those gaps with decisions.

An assistant writes from the instruction it is given, not from your hotel. Asked to reply warmly, it produces the warmest average reply, which is not your hotel's voice. Drift is rarely one bad sentence. It is small moves: a refusal softened into a maybe, an extra "we will do our best", a promise sliding from possible to agreed. Three people on three shifts produce three voices; the sheet is the same voice on all three, including the night shift.

Why adjectives fail and rules work#

Voice is a set of decisions, not a list of adjectives. "Friendly but professional" decides nothing; "open with the guest's name and the fact, never with an apology" decides something. The same logic applies across the whole range of AI tools for hotel teams: instructions that name the choice outperform instructions that name a mood.

The sheet is plain text, one page, pasted into whatever assistant the hotel already uses. It is context, not a guarantee: the assistant can still ignore it, which is why the human check stays. Keep it where replies are written. A sheet nobody opens at 23:40 is decoration.

Tone by situation: welcome, delay, complaint, refusal#

Tone is not one setting. A welcome, a delay, a complaint and a refusal need four different shapes, and the sheet should name each one and show it.

Welcome and pre-arrival messages should be short and concrete: name, dates, what happens next, one offer. No adjectives about the hotel and no paragraph about the view. A delay message, whether the room is not ready, the transfer is late or maintenance is running, should lead with the fact and the new time, give the reason in one clause, then say what the hotel is doing. Never lead with an apology that hides the time.

A complaint reply should acknowledge the specific thing rather than "your feedback is valuable", state what happens next and by when, and not argue about the facts in the first reply. A refusal, on late checkout, an upgrade or a refund, should say no in the first sentence, give the reason once, then offer the nearest alternative the hotel can actually deliver. Post-stay messages carry one question, no pile of survey links, no re-selling. This matters most in messaging guests during the stay, where one line arrives on a phone screen and nothing else is read.

Decide once, for all four: sentence length, whether the guest is "siz" or "you", first name or surname, and whether emoji or exclamation marks are allowed at all.

The four shapes on one page#

What fixed facts must a draft never change?#

Every fact a guest could hold the hotel to: what was sold, what it costs, what is included, when it starts and ends, and where it is. The draft quotes these; it never rewrites them.

Room and rate names must appear exactly as they appear in the reservation system, including the board basis and what the price includes. Cancellation and payment terms come as written in the booking, not as the assistant remembers them. Check-in and check-out times, address, directions, parking, accessibility, and the pet and child policy all belong on the same list, along with the names and titles of staff and which department owns the reply.

On legal duties, keep one approved sentence and nothing beyond it: "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 glosses matter: KBS is the guest identity notification, KVKK is the personal data protection law. Anything more detailed than that sentence is a question for a legal adviser, not a draft.

Copy, never compose#

The rule for all facts: numbers, times and dates are copied, never generated. If the fact is not in the sheet or the system, the draft says a person will confirm it.

Which promises must a draft never make?#

Money, time, space and outcomes. A draft may not offer compensation, a refund, an upgrade, a discount, a particular room, a particular view, or a time the hotel cannot keep.

No compensation, refund, waiver of a cancellation fee or discount, not even as a gesture to be discussed. That is a manager's decision, in writing, after the facts are known. No room number, floor, view, quiet side or adjoining room unless it is already assigned and recorded in the system. No time the hotel cannot keep: "your room will be ready at 14:00" is a promise; "we will message you the moment it is ready" is a process.

A draft makes no admission of liability and offers no explanation of what caused a problem before anyone has checked it. It makes no statement about the law, insurance, safety or medical matters beyond the approved sentence. And it carries no invented memory: "as you requested" or "I checked with housekeeping" appears only if someone actually did. The one promise the assistant may write is "I will check and come back to you by 10:00", because a person owns that promise and can keep it.

How does a person check a draft before it is sent?#

Three questions, in this order: is every fact true and taken from the system, will the hotel keep every promise in it, and would we defend this sentence if the guest forwarded it or posted it?

True means checking the name spelling, dates, times, rate and inclusions against the reservation, not against the draft. Keepable means underlining every future-tense verb and asking who will do it and when; if there is no name, there is no promise. Defensible means reading it once as the guest's lawyer, then as the guest's friend, and rewriting it if it reads badly in either role. This three-question check is the same discipline that governs human approval for hotel AI everywhere else in the operation: the machine proposes, a person decides.

Read it aloud. Sentences that are hard to say are hard to read, and the guest is reading. Language matters too: if nobody on shift reads the language, the reply gets shorter and promises less. Log the edit: when a person changes a draft, note what changed, because the pattern of edits is the raw material for next month's update. The assistant drafts and a person sends, and that rule does not relax as drafts get better.

Three questions, in order#

What does a style sheet look like in Turkish and English?#

The same decisions, written twice, not translated once. Each language sheet carries the same facts, the same promise list and the same three checks, with the tone examples written natively.

Keep one bilingual facts table, covering room names, times and terms, and two rule sheets, one per language. The Turkish sheet uses formal "siz", geniş zaman for what is always true, short sentences, and no stock phrases such as "değerli misafirimiz" unless the hotel genuinely uses them. The English sheet uses US spelling, plain words, no "Dear valued guest" opener and no exclamation marks. Show a before and after for each situation instead of describing the tone with adjectives, and label every example as an example: the sheet is not a set of approved replies to be sent unchanged. The pair below is illustrative, not a template to copy. For the wider question of which languages your hotel should serve, see our guide to multilingual hotel service.

For example: a delay and a refusal, in both languages#

How often should the sheet be updated?#

Once a month, from evidence: the drafts people edited before sending, the replies that produced a second question, and the offers and terms that changed.

Collect the edits, not the opinions. A draft that five people change the same way is a missing rule, and the rule writes itself. Retire words as well as add them: keep a short list of phrases the hotel has decided against, each with one line on why. Refresh the facts first: rates, times, seasonal offers, closures, renovation dates. A stale fact table is worse than no table, because it is wrong with authority.

Date and version the sheet, and keep the previous version, so a bad month can be traced to the rule that caused it. Name one owner in the sheet, with a deputy who covers the night shift. Test it: give the assistant a situation the sheet has never seen and read the draft. If it drifts, a rule is missing, not a person. And do not let it grow past one page. A sheet nobody reads is worse than no sheet, because it looks like a policy.

What to do next#

Write the first version this week from material you already have. Take the ten replies your team edited most last month, pull the rule out of each edit, add the facts table and the promise list, add the three questions, date it, and paste it into the assistant you already use. The pieces fit in one order: the sheet sets the tone by situation, the facts table stops the draft inventing, the promise list stops it committing, the three questions catch what is left, and the monthly update keeps all four honest.