AI booking readiness checklist: what hotels can prepare now

AI booking readiness means your hotel can answer four questions in software, not in a phone call: what is free, what it costs, which rules apply, and who may confirm. It rests on one room inventory, machine-readable rate and policy rules, written permission levels for outside software, and a log of every action.

A guest asks an assistant to find a sea-view room for two nights. The assistant can see the room and the price, but not whether the deposit comes back if the trip changes. That gap, not the AI, is what ends the conversation. This article turns AI booking readiness into work you can do now, with whatever software you run: check your inventory, turn prose rules into conditions software can apply, write down the five permission verbs, set limits, and keep a log. None of it promises bookings; all of it removes the gaps that stop them.

What does an AI assistant need before it can book a room?#

Four things, all held in software: current availability, a price, the rules that apply to that price, and a path to confirm or hold the room.

Each of the four must be answerable by a machine in seconds. If any one of them needs a call to the front desk, the request stops. An assistant cannot phone anyone, so every answer it needs must already live in a system.

The usual route is a service that offers tools, meaning functions an AI model can run. The Model Context Protocol, an open standard for connecting AI applications to outside data and tools, describes exactly this: an AI application connects through a connector to a service that can offer data, prompts and tools (Model Context Protocol specification, 28 July 2026 (opens in a new tab)). We explain what this means for hotels in plain words in What MCP means for hotels: AI assistant access in plain words.

The protocol places consent on the application, not on the hotel: users must understand and agree to data access and actions, and the application must obtain explicit consent before invoking a tool (Model Context Protocol specification, Security and Trust & Safety, 28 July 2026 (opens in a new tab)). Your job is to make the answers clean enough that consent means something.

A hold is the cheapest safety valve. The room leaves sellable inventory for a fixed time and returns automatically if nothing is confirmed, so a slow traveler never blocks a room for a week.

For context, not as a promise: in October 2025 an AI company announced a plan to let third-party apps work inside its assistant, with two large booking sites first, built on MCP (Skift, 6 October 2025). An announced plan is not availability in Türkiye. In August 2026 a search company began rolling out hotel booking inside its AI mode for travelers in the United States, with ten named partners; the traveler chooses and pays, and the assistant does not book on its own (Skift, 27 August 2026). The pattern to note is that the person stays in the deciding seat.

Why does one room inventory come first?#

Because an assistant, an OTA and your front desk that each keep their own copy of availability will eventually sell the same room twice.

The mechanism is simple: one count of sellable rooms, decremented by every booking wherever it comes from. Two counts mean two truths. Holds must live in that same inventory, because a hold that exists only inside the assistant's session is invisible to the front desk and to every channel.

Room-type mapping is where this breaks in practice. A room sold as a standard double in one channel and a superior in another is one physical room with two identities. Rate plans are separate objects from room types: an assistant needs both, correctly paired, or it will quote a price that belongs to a different product.

A practical check: pick a date about 30 days out, read the sellable count in every channel at the same hour, and compare. Differences are mapping errors, not demand. Name one person who owns the mapping and approves every change to it; shared ownership is how a mapping drifts. See Channel manager mapping: room types, rate plans and testing for the full method.

Which rate and policy rules must software apply without guessing?#

Every rule an assistant has to apply must exist as a condition with a value, not as a sentence in a description.

The rules that decide money: the cancellation window and what is charged after it, the deposit amount and whether it is refundable, child age bands and extra-bed charges, minimum stay and closed-to-arrival days, and whether breakfast, tax and fees sit inside the quoted price. For example, "free cancellation until 7 days before arrival, then one night charged" is a rule software can apply. "Flexible conditions" is not.

Public descriptions are a separate problem and a separate job. Schema.org's Hotel type carries an amenity property that does not say whether a feature is included in the price or costs extra (Schema.org, Hotel type, 16 September 2026 (opens in a new tab)), which is exactly the ambiguity an assistant cannot resolve on its own. Keeping those facts consistent is its own task, covered in How AI assistants describe your hotel: keeping facts consistent.

Test the rule, not the text: give it to a colleague who has not seen the rate and ask for the total in three scenarios. If two people get two answers, software will too. Policy text and policy logic must agree; if the website says one thing and the system applies another, the guest will quote the website, and the guest will be right. State the currency and whether tax and fees are included, once, in the same place the price comes from.

What should outside software be allowed to do?#

Treat permission as five separate verbs, granted one at a time: search, quote, hold, book and cancel.

Search reads availability. Quote returns a price with its rules. Hold removes the room from inventory for a fixed time. Book creates the reservation. Cancel ends it. They carry very different risk and should never arrive as one switch.

Least privilege is the working principle: the protocol's authorization section says clients should request only the permissions a task needs, and that authorization itself is optional in the protocol (Model Context Protocol specification, Authorization, 28 July 2026 (opens in a new tab)). The tools section asks for a person who can deny a call, clear signals when a tool is used, confirmation before operations, input validation and rate limits on the server side, and logging of tool usage for audit on the client side (Model Context Protocol specification, Tools, 28 July 2026 (opens in a new tab)).

Cancel is the verb hotels forget to define, and it is the one with the most money attached. Grant per partner, with a review date in the calendar; a permission nobody remembers granting is a permission nobody revokes. A quote that cannot be held is a price the guest may never get, so quote and hold usually belong together in a first test. The wider safeguards around AI in hotels are covered in AI in hotels: a practical guide to uses, limits and safeguards.

Which limits should the hotel set?#

Limits are the hotel's own numbers, written down and enforced by software, not left to the assistant's judgment.

Start with a narrow surface: expose the simplest flexible rate plan only, and keep prepaid, non-refundable and package rates out of the first test. Set hold length deliberately: long enough for a traveler to decide, short enough that expired holds do not lock inventory, and expiry should return the room without anyone touching it.

Cap rooms per request and per day, so one misread instruction cannot take a floor out of sale. Set a value ceiling above which a person confirms, and decide whether that person is on reception, reservations or a duty manager. List blackout dates and closed-to-arrival days explicitly, in the same system the price comes from.

One caution on wording: these are commercial limits the hotel chooses. They are not the same mechanism as an automatic decision threshold inside a software product, which may be fixed in code and not hotel-editable. Write your limits on one page with an owner and a review date.

How do you mark and audit bookings that came through an assistant?#

Tag the reservation with its source at creation and keep an append-only record of every action, so a guest's claim can be checked against what actually happened.

A source or channel code on the reservation lets revenue reports separate these bookings without changing the rate or the room block. Decide the guest of record: whose name, contact and consent sit on the booking. The assistant acts for a person, and the personal data duty stays with the hotel.

Decide who sends the confirmation and in which language, so two confirmations never reach one guest with different terms. Log the action, not just the outcome: which tool ran, on whose authority, with what input and output, and when. The protocol asks client applications to log tool usage for audit purposes (Model Context Protocol specification, Tools, 28 July 2026 (opens in a new tab)). Keep the record append-only: a correction is a new entry, never an edit to the old one.

When a guest says the assistant promised something, the log is the only evidence either side has. Read it before you answer.

What should you ask your software provider?#

Ask for the five verbs, the limits, the log and a test property, and ask for each answer in writing.

Which of search, quote, hold, book and cancel can outside software perform today, and how is each granted and revoked? Where do the limits live, and can you change them without waiting for a software release? Can you see and export a log of every action, including holds that expired? Which room types and rate plans are exposed, and can you exclude one without changing the others?

Is there a test property where you can try this without touching live inventory? A fictional property is enough for a first run. How is such a booking tagged in reports, and what happens to a hold when it expires? For reference, AI Otel Core is built and runs in a public demo environment: an AI assistant can find, hold and book a room on our demo property within limits the hotel sets, and the demo hotel is fictional.

What to do next#

Pick one date about 30 days out and run the two checks in this article against it: compare sellable counts across every channel at the same hour, then hand one rate plan's cancellation and deposit rules to a colleague and see whether the totals match. Fix what fails, write the five-verb permission sheet and the limit page, and only then ask a provider for access. Inventory, rules, permissions, limits and the log are one chain: each link is useless without the one before it. If you want to see the demo environment described above, request a briefing at https://ai-otel.com/contact.