AI in hotels: a practical guide to uses, limits and safeguards

AI in hotel software either answers (questions, drafts, summaries) or acts (changes a record, a rate, a booking). Judge any use with three tests: whose data it reads, who decides, and whether the action can be undone. The uses a hotel can evaluate today all end with a person.

A vendor says the system can answer your guests, price your rooms and run your front desk. The real question is narrower and harder: which of those actions would you let a machine take at 2 a.m. with nobody watching? This guide separates what AI in hotels actually does today from what it is sold as, gives you three tests to sort any use, lists the uses that hold up, and tells you what to demand from any supplier, including us.

What does AI actually do in hotel software today?#

It does one of two things: it produces language, or it takes an action. Almost every honest conversation about AI in a hotel starts by separating the two.

Answering means a model reads text and produces text: a summary of last night, a draft reply to a guest, an answer to "how many rooms are out of order tonight". Nothing in the hotel's records changes. Acting means the same model calls a function that changes something: assigns a room, sends a message, applies a rate, posts a charge to a bill. The output is now a fact about the hotel, not an opinion about it.

The connection layer has a name. The Model Context Protocol (MCP) is an open protocol that gives AI applications a standard way to reach outside data and tools: an AI application (the host) connects through a connector (the client) to a service (the server), and a server can offer resources (context and data), prompts (templated messages and workflows) and tools (functions the model can run), as set out in the protocol's specification (Model Context Protocol project, July 2026).

Türkiye's Personal Data Protection Authority, in its February 2026 document on agentic AI, notes that there is no single agreed definition and describes the field generally as AI systems made up of AI agents acting and interacting with different degrees of independence to reach set goals. The same document gives a useful shape for the whole category: an agent asked to organize a business trip plans it from calendar, email and travel policy, presents the plan for approval, and books only after approval is given.

Why the split matters commercially: a wrong answer costs a correction, a wrong action costs money, a guest or a legal record. The two deserve different questions and different budgets.

How do you judge an AI use in a hotel? Three tests#

Ask three questions in this order: whose data does it read, who makes the decision, and can the action be undone. A use that fails any of them is not ready, however good the demonstration looked.

Whose data does it read: the hotel's own records, a guest's personal data, or public data. The answer tells you who is responsible if it is wrong and what the retention rules are. Who decides: the system alone, the system with a person confirming, or the system proposing and a person deciding. Get the actual list of actions in each group, not a promise that "sensitive things are protected". Can it be undone: a draft can be deleted, a rate change can be reversed until a guest books, a message already sent to a guest cannot be taken back.

Run the tests on the smallest version of the use, not the brochure version. "It can answer questions about occupancy" and "it can change a rate" are different products with different risk. Write the three answers down: a one-page note per use is enough, and it becomes the checklist you hand back to the vendor when something changes. Re-run the tests when the scope grows, because a tool that was read-only in March is a different tool when it gains permission to write in June.

Which AI uses hold up in a hotel today?#

The ones that end with a person: questions answered from the hotel's own numbers, drafts someone sends, briefings someone reads, alerts someone acts on, and checks on paperwork someone signs.

Questions about the hotel's own numbers#

For example: "which reservations arrive after 23:00 tomorrow", "which rooms have an open maintenance ticket", "what did restaurant sales look like last Tuesday". The system reads and answers; nobody's record changes. This is the cheapest place to learn what a system does when it is wrong, because the answer can be checked against records you already keep.

Drafts a person sends#

A model writes a first version in the guest's language; the front desk edits and sends. The value is time, not autonomy, and the handover rule matters more than the draft quality. The same logic covers a morning briefing written from the hotel's own numbers and read by the manager before the stand-up: the test is whether every line can be traced back to a record the hotel already keeps. An alert when a number moves outside its usual range belongs here too. That is a threshold, not a prediction: the hotel sets the band, the system says the number left it, a person decides why. For more on the handover point, see when to hand an AI reply over to staff.

The Turkish paperwork: identity records, e-documents, personal data#

Guest identity records. Under Article 2 of the Identity Notification Law (No. 1774), the responsible operators of hotels, motels, pensions, holiday villages and every other kind of lodging place must keep identity and arrival and departure records of everyone they accommodate, day by day in the prescribed form, and keep them ready for inspection by the general law enforcement at any time. A system can check that a record is complete before the guest leaves the desk; the notification itself stays in the check-in flow, with a person.

Electronic invoices. The Revenue Administration's 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. A system can flag a missing tax number or a mismatched amount; a licensed integrator issues the document and a person reviews it before it goes. Confirm dates, scope and your own case with an accountant or legal adviser.

Personal data housekeeping. Law No. 6698 Article 4 requires data to be processed lawfully and fairly, kept accurate, used for specified and legitimate purposes, limited to them and kept only as long as the law or the purpose needs; Article 7 says that when the reasons for processing end, the controller deletes, destroys or anonymizes the data. A system can help find and list what should be exported or deleted; the decision stays with the hotel as data controller.

For a fuller argument on sequencing, see where to start with AI in a hotel: back office before guests.

Can AI assistants book hotel rooms yet?#

Not in a way a hotel in Türkiye can plan around today. What exists is announcements, a shared technical protocol and a set of preparations a hotel can make now.

The protocol is real and open. MCP was contributed to the Agentic AI Foundation, formed by the Linux Foundation on 9 December 2025 so that it stays open and neutral; the Linux Foundation said more than 10,000 MCP servers had been published by then. On the demand side, the record so far is three data points. In October 2025, an AI assistant provider announced a plan for third-party apps inside its assistant, with two large travel booking sites first, built on MCP (Skift, 6 October 2025). Announced a plan, not launched a channel. In March 2026, The Information reported that the same provider was scaling back direct in-chat checkout; Reuters relayed the report and the company did not comment (Reuters, 6 March 2026). In August 2026, a search and advertising company began rolling out hotel booking in its AI Mode for US travelers, with ten named partners; the traveler chooses and pays with its payment service, and AI Mode will not book on its own (Skift, 27 August 2026). That it is not available in Türkiye is an inference from "US only".

What a hotel can prepare, regardless of who wins: one clean room inventory that feeds every channel, a written limit on what an assistant may hold or book, a clear meaning for a hold and its expiry, and a way to see afterwards what an assistant did. None of this is a channel a hotel can budget for yet. Treat it as a design constraint on your inventory and your records, not as a sales line. The announcements are traced in more detail in what has been said about AI assistants booking hotels.

Which phrases should make you ask questions in a sales pitch?#

Five phrases carry most of the risk: "runs the hotel", "sets your prices", "knows your guests", "learns from your hotel" and "works without people". Each is a claim about decisions, and each has one question behind it.

"It runs the hotel" deserves the hardest question. The OWASP Top 10 for LLM Applications (2025 edition) names Excessive Agency as the weakness that lets damaging actions follow unexpected, ambiguous or manipulated model output, traces it to too much functionality, too many permissions or too much independence, and recommends limiting tools and permissions to the minimum necessary, with human approval for high-impact actions. "It sets your prices": ask whether a suggestion arrives with its reasons and who approves it, because a rate change is reversible only until a guest books at the new rate. "It knows your guests": ask whose data it reads, where it is stored, how long it is kept and how it is deleted; Law No. 6698, Articles 4 and 7, set purpose limitation, retention and deletion. "It learns from your hotel": ask for the measurement, not the adjective; a trained model is a testable claim, a demonstration is not. "Works without people": ask what happens when it is wrong at 2 a.m., who notices, and what the guest is told.

Attribute nothing to a named company. These phrases circulate widely; the point is the question you ask, not who said it.

What safeguards make an AI use acceptable?#

Four: approval tiers, a record of what actually ran, an undo, and a test before anyone relies on it. Without all four, you are trusting a demonstration.

Approval tiers. Sort every action into three groups: runs directly, needs one person's confirmation, needs two. The tiers exist so that confirmations stay meaningful; if everything needs a click, people click. The MCP specification itself says the same thing for tools: for trust, safety and security there should always be a human in the loop able to deny a tool invocation; applications should show which tools are exposed, signal when a tool is used and ask the user to confirm; servers should validate inputs, apply access controls and rate-limit calls; clients should log tool usage for audit (Model Context Protocol project, July 2026). Türkiye's data protection authority recommends the same shape for agentic AI: meaningful human oversight, with conditions and stages of oversight defined in advance, testing during development of what data the system may access and what decisions it may take, extra human review for high-risk decisions at deployment, and monitoring of behaviour once in use (Kişisel Verileri Koruma Kurumu, February 2026).

A record. Every change should be logged, past entries should be append-only, and an alteration should be visible in the chain. Ask whether records can be edited, and by whom. An undo. Every change should be reversible with one exception: a message already sent to a guest. Plan that exception explicitly, because it is the one that reaches the outside world. A test before go-live: a fixed set of cases, including deliberately awkward ones, run before staff or guests depend on the system, with the results written down and repeated after any change of scope. How to build that test is covered in how to test a hotel AI assistant before staff or guests use it, and which actions need a person in human approval for hotel AI.

Where should a hotel start, and where does AI Otel stand?#

Start in the back office, with a question your team already asks every morning and a record that already exists. Guest-facing automation comes after the approvals and the records are in place.

The order of work: read-only questions first, then drafts a person sends, then alerts, then routine actions inside approval tiers. Each step teaches you what the next one needs. Before any of it: one room inventory that feeds every channel, guest records that are complete and clean, and a written list of what a machine may do alone. These are unglamorous and they decide whether anything later works. Choose the first use by the three tests, not by the demonstration: a question with a right answer, on data you already keep, with a person checking the result, is the cheapest way to learn what the system does when it is wrong.

AI Otel's PMS assistant is built: it answers questions about a hotel's own property in plain language, carries out routine actions, and a person confirms anything sensitive. It is built and working ahead of launch, not in hotels' hands, and it is not a trained model; we have no trained model yet. The same discipline applies to us as to anyone else: ask for the action list, the tiers, the record and the undo. If a supplier cannot produce them, the demonstration is all there is.

What to do next#

Pick one question your team asks every morning, answer it from your own records with a person checking the result, and keep a note of what the system did. Then take the next question. Approval tiers, records, undo and testing exist to make that first step safe enough to repeat, and they are the same four things you should demand from any supplier, including us. If you want to see how we apply them, request a briefing.