Where to start with AI in a hotel: back office before guests

Start with back-office work a person already checks: preparing bill corrections, drafting replies for review, summarizing the day, checking registration details before the identity notification. Score each candidate on reversibility, data quality and who answers for the result. Record your own baseline first. Guest-facing work comes later, because a message already sent cannot be undone.

A general manager wants the AI to answer guests on WhatsApp first, because that is where the complaints arrive. The front office clerk wants it to stop retyping the same passport details into three screens. Only one of those two requests can fail safely. This article gives you a way to decide which task goes first: a scoring method, a shortlist of safe candidates, the Türkiye-specific tasks that reward early attention, and a baseline sheet to fill in before any tool is involved. For a wider view of the technology's uses and limits, see our guide to AI in hotels.

Why does the first AI use decide whether staff trust the next one?#

Because the first task teaches the team what an AI mistake looks like and who has to fix it. If that lesson is cheap and private, the second attempt gets a fair hearing; if it is public, the second attempt never happens.

Staff do not form an opinion of AI from a demo. They form it from the first wrong answer they have to clean up after. A back-office error is caught on a screen by a colleague; a guest-facing error is caught by the guest, in public, and often after it has been screenshotted.

The first use also sets the review habit. If the person checking the output is named and expected to change things, review survives. If the check is nominal, it disappears within weeks. And the first use teaches the hotel what its own data looks like: which fields are always wrong, which records are missing, which team keeps its work in a notebook.

A failed first attempt raises the cost of the second proposal, because the next conversation starts with the failure rather than the problem. Choose the first task for what it teaches, not for how impressive it sounds in a management meeting.

How do you score an AI candidate before you try it?#

Put every candidate through three tests: reversibility, data quality and accountability. A task that fails any one of them is not a first task, however much time it seems to save.

Reversibility asks what happens if the output is wrong. A draft on a screen is corrected in seconds; a message already sent to a guest, a payment taken or a document filed cannot be taken back. Data quality asks where the input lives today. If the answer is a spreadsheet, a WhatsApp group and one person's memory, the AI will reproduce the mess faster, not fix it. Accountability asks whose name goes on the result. If nobody is willing to sign it, the task is not ready, because an AI cannot carry a legal or commercial duty.

Two further tests help rank the survivors. Checkability: can a competent colleague tell in under a minute whether the output is right? If checking takes longer than doing, the task gains nothing. Frequency: a task done twice a week teaches you little; a task done forty times a day gives you enough cases to see the pattern of errors. Rule-heaviness matters too: tasks governed by written rules, such as a rate sheet, a bill correction or a registration form, suit early use better than tasks governed by taste.

Write the score down before the trial and keep it. A scored decision is the only defense against buying on enthusiasm. If you are new to the vocabulary, our glossary of AI agents, assistants and chatbots keeps the terms straight.

The three tests in one table#

Which back-office tasks are the safest first candidates?#

Tasks where a person already checks the work today, and where the check is quick and rule-based. You are not adding a review step; you are putting a faster first draft in front of a review step that already exists.

Preparing a bill correction for review is one. The correction is proposed as a void and a reversal, a person approves it, and no financial record is ever deleted. Drafting replies to guest emails, agency messages and review comments is another: the draft saves the blank page, the person keeps the voice and the decision.

Summarizing the day from the hotel's own numbers works the same way: arrivals, departures, open balances, rooms not yet cleaned, charges that look unusual, so the morning meeting starts from a page instead of a search. Translating a standard text, such as a house rule or a cancellation explanation, into another language for a person to check before it is used. Turning a supplier invoice, a rate sheet or a group rooming list into structured lines for a person to confirm against the original. Comparing the day's charges with the night audit close and listing the differences for a person to resolve, rather than resolving them.

In every one of these, the output is a proposal. The sentence that leaves the hotel is written or approved by a person. Before staff touch any AI tool, make sure a written policy says what they may paste into it; our guide to hotel AI policy covers that ground.

Which Türkiye-specific tasks should come first?#

The checks that sit in front of a legal step: registration data before the identity notification, e-document review before sending, and personal data handling that stays accurate, limited and time-bound. Each one ends with a person, because the duty belongs to the hotel, not to the software.

Identity records come first. Under Article 2 of the Kimlik Bildirme Kanunu (Identity Notification Law No. 1774, Resmî Gazete 11 July 1973), the responsible operator of a hotel, motel, pension, holiday village or any other lodging place must keep identity and arrival and departure records of everyone accommodated, domestic or foreign, paid or free, day by day and ready for inspection, as the consolidated text on mevzuat.gov.tr (opens in a new tab) sets out. A data check that flags a missing or mismatched field before the notification is prepared is a good early task; the notification itself is prepared and sent by a person from the check-in flow.

E-documents follow the same logic. The Revenue Administration's Tax Procedure Law General Communiqué No. 509 describes e-Fatura as letting an invoice be issued, delivered, kept and presented electronically instead of on paper, and states that neither e-Fatura nor e-Arşiv Fatura is a new document type: each has the same legal nature as a paper invoice, according to the GİB communiqué (Sıra No: 509) (opens in a new tab). That is exactly why a person reviews an e-invoice before it is sent: the document carries the same weight as paper.

Personal data is the third candidate. Article 4 of Law No. 6698 requires data to be processed lawfully and fairly, kept accurate and up to date where necessary, processed for specified, explicit and legitimate purposes, limited and proportionate to them, and kept only as long as the law or the purpose requires. Article 7 says that once the reasons for processing no longer exist, the controller deletes, destroys or anonymizes the data on its own initiative or on the person's request, as the consolidated law on mevzuat.gov.tr (opens in a new tab) states. A task that finds stale guest records for a person to review serves both articles.

Regulators point the same direction. The Personal Data Protection Authority's February 2026 document on agentic AI treats meaningful human involvement as important, says that defining when and at which stages human oversight enters can reduce the risk of unforeseen results, and describes as good practice defining and testing what data a system may access, extra human review for potentially high-risk decisions, and monitoring behavior once the system is in use.

Multilingual drafts fit the pattern: a person checks the Turkish and the foreign-language version against each other before anything reaches a guest. Keep the legal question with the professionals: confirm your own scope, dates and duties with your accountant or legal adviser, and do not let a tool decide what the law requires. For the full picture of the legal steps around a stay, see our article on hotel compliance in Türkiye.

Why do guest-facing AI tasks come later?#

Because the failure is public and permanent. A guest-facing reply, price or promise cannot be recalled, and the guest, not the hotel, decides when the conversation ends.

A message already sent to a guest cannot be undone; every other kind of AI output can be corrected, re-run or deleted before it leaves. A wrong price quoted to a guest becomes either a commitment the hotel honors or an argument it has with a customer, and both cost more than the time saved. Guest-facing work also sits on top of money and legal documents: a booking, a bill, an invoice, an identity record. The back office has to be clean before anything speaks to a guest.

The guest experience is the product, so an error there is a brand error, not an internal one. Guest-facing assistants need limits the hotel sets in advance: what they may offer, what they may confirm, what they must hand to a person, and how fast that handover happens. That is not an argument against guest-facing AI; it is an argument for earning it, by first proving the review habit on work nobody outside sees.

What baseline should a hotel record before it starts?#

Time taken and corrections needed, on one task, one person, one week, written down before the tool arrives. Without that page, nobody can say afterwards whether the work improved.

Pick one task and one person, not a department; a baseline spread across five people measures the team, not the task. Count the instances: how many times in a week does this task actually happen, and how long does one instance take from start to finished, checked output. Count the corrections: how often does the first version have to be changed, and what kind of change it is, a wrong number, a missing field, a wrong tone, a missing document.

Record who checks and how long the check takes. The check is part of the work, and often the part that decides whether the tool helps at all. Write down the error types you already see without any AI; those are the errors you will be watching for later, and they existed before the tool.

Re-measure with the same method after a fixed period, and set a stop rule in advance: if corrections rise above the level you agreed, the trial ends. The baseline is the hotel's own measurement. This article gives no expected saving, because the only honest number is the one your own front office writes down.

What is a sensible order of adoption?#

Move from the most reversible and most checked work to the least, and only move on when the previous stage's corrections have settled. Five stages are enough for most hotels.

Stage one is internal drafts nobody outside the hotel sees, such as a summary of the day or a first pass at a document. Stage two is internal checks that feed a person's decision, such as flagging a missing registration field or listing the differences at night audit close. Stage three is guest-facing drafts that a person edits and sends, such as a reply to a question about parking or a late check-out. Stage four is guest-facing replies sent within narrow, written limits, with an immediate handover to a person for anything about money, complaints or special requests. Stage five is anything that touches a payment, a fiscal document or an identity record, only with a named approver and a written rule for what the tool may never do alone.

At every stage, keep the same three tests from earlier: reversibility, data quality, accountability. A stage that stops passing them is a stage to step back from.

What to do next#

The next step is one page, written this week: choose a single back-office task a person already checks, name the person who signs the result, and record the baseline of time taken and corrections needed before any tool is involved. The order of this article is the argument: reversibility, data quality and accountability decide the candidate; the Türkiye-specific checks in front of the identity notification and the e-document come early because a person already stands there; guest-facing work waits until the review habit is proven; and the hotel's own baseline, not a vendor's promise, is what tells you whether any of it worked.