AI for a small hotel or pension: five jobs for a team of three

AI for small hotels means using one general AI assistant for drafting and checking: guest replies in the guest's language, listing text comparisons, review replies, a monthly numbers read and a weekly price note. A person reads and sends everything. No chatbot, no booking agent, no integration, and no guest identity or card data in the tool.

The owner of a 12-room pension answers the phone, writes the listing text, closes the day at 11 pm and replies to reviews on Sunday. Advice written for a hotel with a revenue manager does not survive that week. This guide picks five jobs that fit a few hours a week, gives you the rules page to run them under, and names the tools to skip for now.

Why does the usual AI advice not fit a team of three?#

Because most AI advice assumes departments: one person owns the tool, another checks the output, a third keeps the data behind it clean. In a small pension all three are the same person, and that person is also at the desk.

The jobs are not split. The owner-manager is reservations, front desk, marketing, purchasing and the monthly accounts, so any tool that needs an owner has no owner. Advice written for a hotel with a revenue manager assumes a rate system, a channel manager and clean historical data; a pension often has a notebook, a spreadsheet and two booking channels. The realistic budget is a few hours a week, and most of it goes to checking, not to setup, so a tool that needs a day of configuration will not be configured.

Guests already arrive with the habit. TÜİK's "Yapay Zekâ İstatistikleri, 2026" (opens in a new tab) reports that 37.6% of internet users aged 16-74 said they used generative AI in 2026, up from 19.2% in 2025 (bulletin no. 62066, 2 October 2026). Your guests are not surprised you use it; they may be surprised when it answers them badly.

The usual failure is not a wrong answer. It is a subscription nobody opens after three weeks, paid for until someone notices. So start where a mistake is cheap and reversible: text you were going to rewrite anyway, not a promise made to a guest. If you later want the department-by-department version, see where to start with AI in a hotel: back office before guests.

What is the smallest AI setup that works?#

One general AI assistant you already have access to, and one printed page of rules. That is the whole stack, and it is enough for the five jobs in this article.

Pick one assistant and stay with it for a month. Comparing tools is the most pleasant way to lose the month and learn nothing about your own work. Write the rules page before the first use, not after the first mistake; it takes twenty minutes and it is the only part of this that has to be exact.

The data rule is absolute: never paste guest identity documents, passport or ID numbers, card numbers, KBS notification details or health information. Ask the question with the person removed, and the answer is almost always the same. One rule carries the rest: a person reads every output before it leaves the building. Drafting is delegated; sending is not.

Keep a plain log: date, job, what you changed in the draft. After four weeks it tells you which jobs are worth keeping. And talk to the team before the tool. The fears are usually about being replaced or being blamed for a bad answer, and both are answered by the rules page, not by a demo, as covered in introducing AI to a hotel team: the first meeting, the fears and the rules.

The one-page rules sheet#

Job 1: how do you answer guest messages in the guest's language?#

You draft with the assistant, then read, correct and send yourself. The tool writes the first version; you own every fact in it.

Give the tool what it cannot know: room type, dates, what the rate includes, cancellation terms, check-in and check-out times, parking, pets, breakfast. Do not give it the guest. Replace the name with "the guest", drop the booking number, and describe the question in one line.

Ask for two versions, one short and one fuller, and choose. Two versions expose the assumptions the tool made. Check numbers, dates and names by hand every single time, because this is where drafting errors turn into money. Keep the property's voice: a pension's reply should sound like the person at the desk, not like a chain's template.

Write complaints, refunds and anything touching money or safety yourself. The tool may cool the wording; it may not decide the outcome. Answering in the guest's own language is the highest-value use for a small property, because the alternative is a machine translation nobody checked.

Job 2: do your listings say the same thing?#

Rarely. The same room is described on your own site, on each booking channel and in the emails you send, and each text was written at a different time by a different version of you.

The tool cannot see your listings. You paste two descriptions side by side and ask it to list what contradicts. That is the whole method, and it needs no integration. Compare named fields, not general impressions: what the rate includes, breakfast, check-in and check-out times, cancellation, extra bed, pet policy, parking, taxes.

Fix the master description first, then the channels. If you edit each channel separately, the contradiction returns next season. Do this before you open a new channel, not after, because a new channel copies whatever text you give it; see selling a small hotel online: which channels to open first. Run it again after a renovation, a rate change or a new policy, because those are the moments when one text moves and the others do not. Keep the master description in one file you can paste from, so the next comparison starts from a known text.

Job 3: who signs the review replies?#

You do. The assistant drafts, you cut the draft in half and sign it.

Write for the next guest reading the review, not for the reviewer. The reply is a listing text that happens to be in public. Never argue with a fact you cannot check: if the guest says the room was cold, the useful reply says what changed, not that it was warm.

Name the specific thing and the specific change. "We replaced the boiler in January" beats "we take your feedback seriously". Do not let a tool invent compensation, discounts or promises; those are yours to make, and a drafted apology that offers a refund you never agreed to is worse than a slow reply. Keep it short, because three sentences is usually enough and length is where drafted replies go wrong.

Reply in the language of the review, then read the Turkish or English version yourself before it goes out. Log recurring complaints: three mentions of the same thing is a maintenance job, not a communication job.

Jobs 4 and 5: what can AI do with your monthly sheet and your weekly prices?#

It can read the numbers you paste and explain what moved. It cannot know your market, and it should not set your prices.

Paste totals, not people: revenue by source, costs by heading, rooms sold, occupancy, average rate. No guest names, no booking numbers. Ask for the three biggest changes against last month and against the same month last year, and for the questions you should ask. Questions are the useful output. Ask it to explain a ratio you do not understand; that is teaching, it costs nothing, and it makes next month's sheet better.

Do not ask it to forecast. It has no demand data for your street, your town or next week's weather. The weekly price note is the fifth job: write next week's rate and one sentence of reason before the week starts, then compare at the end of it. The value is the habit of writing the reason, not the number the tool suggests. After a season you have your own record of what your market does. Keep the sheet in one file with one layout, so month-to-month comparison stays possible without reformatting; the monthly income and expense sheet for a small hotel or pension gives you that structure.

What should a small property skip for now?#

Website chatbots, AI booking agents and paid AI search services. Each one needs a system behind it that a pension does not have yet, and each one fails quietly.

A website chatbot needs a maintained answer base and someone to read the transcripts; without that it answers confidently and wrongly, in public. An AI agent that books a room needs live inventory, limits the property sets and a record of what it did: that is a system, not a subscription. Paid "AI search visibility" services sell a promise about a channel you cannot measure at 12 rooms, so you cannot tell whether it worked. AI pricing engines need a rate system and demand data; the weekly price note does the same job by hand, with a reason attached.

Anything that asks for your channel credentials before it explains what it does is a no, whatever the price. Look again when two people share the desk, when your channel mix has been stable for a season, and when a tool can read your actual system instead of a paste. One caution about market reading: when you read Ministry of Culture and Tourism accommodation figures, remember they count certified facilities, and a small pension may not appear in that count, so do not read those numbers as your own market.

What to do next#

Pick one of the five jobs, write the rules page, and run that single job for four weeks before adding a second. All five work the same way: the tool drafts or compares, a named person checks and sends, and the log tells you what to keep. That pattern is what a team of three can supervise, and it is why the tools that need a system behind them can wait.