AI for Hotels: The Workflows That Pay Between Enquiry and Check-In

AI for hotels means putting an AI agent, not a general chatbot, on the workflows between a guest’s enquiry and their arrival: pricing a direct request, chasing a quiet quote, answering a pre-arrival question from your own documents, and drafting a review reply. The revenue leak in a small hotel lives in the inbox.

Direct Answer

An AI agent for a hotel reads one data source and drafts one output inside a single workflow; a person approves it before a guest sees it. The four workflows worth it are a priced reply to a direct enquiry, a follow-up on a quote gone quiet, a pre-arrival answer from the hotel’s own documents, and a drafted review reply. Guest-facing chat and in-stay automation come later, if at all.

Key Takeaways

• Direct enquiries are the highest-value workflow. A priced reply in minutes beats a perfect reply the next morning.
• Follow-up on quotes is where most small hotels lose direct bookings to an OTA, because nobody chases.
• Pre-arrival questions can be answered from your own documents with the source shown.
• An AI agent for a hotel is narrow by design: one workflow, real tool access, a person approving the output — not an autonomous assistant that books on a guest’s behalf.
• Booking changes, refunds and complaints stay with a person. The model prepares, it does not decide.

Is this an AI agent, or a chatbot with better prompts?

Neither word is doing much work on its own anymore. A chatbot answers a question from text. An agent is given access to tools, decides what to do with them inside a boundary, and hands off whatever sits outside that boundary. What Are AI Agents? and AI Agent vs Chatbot: The Difference Is Write Access, Not Intelligence make that case in general; a hotel front office is a clean example of it.

Each of the four workflows below is a narrow agent, not a chatbot and not a general assistant. It reads one data source — availability, an arrival document, a review — acts within one boundary — draft a priced reply, flag a quote as unanswered, draft a review response — and stops at the edge of that boundary every time. It does not “do hotel AI.” It does one workflow, with tool access, and a person approves the output before a guest sees it.

That is a narrower claim than the one vendors are selling. One 2026 hospitality AI guide describes agents already “handling email intake, checking live inventory across multiple properties, negotiating alternative dates… without a human touching the thread” (Ampcome). For a three-person front office, that is the wrong ambition before it is the right one. An agent that negotiates dates with a guest unsupervised also quotes the wrong price unsupervised. The version worth building first is the bounded one: narrow task, real tool access, a person in the approval loop until the system has earned the right to lose one.

Which hotel workflows should use AI first?

The first two rows are where the money is. A direct enquiry is a guest who chose to skip the platform. If the reply takes until the morning shift, a share of them will have booked through the platform by then, and the hotel pays commission on a guest it had already won. A model that drafts the priced reply from the same availability the desk would check makes the reply time independent of who is on shift.

Quote follow-up is the workflow nobody owns. The desk sends a quote, the guest goes quiet, and the thread dies because chasing is not anyone’s job at 4pm on a Saturday. A model that watches for silence and prepares the second message, for a person to approve, changes the follow-up rate from “whenever someone remembers” to “always, within two days”.

How does AI answer pre-arrival questions safely?

The safe version is retrieval over the hotel’s own documents: the arrival guide, the parking policy, the transfer options, the breakfast times, the terms. The model finds the relevant passage, answers in the hotel’s tone, and shows where the answer came from. When it cannot find a source, it says so and routes the question to the desk.

That last behaviour is the design decision that matters. A model that guesses a check-in time has created a complaint. A model that says “I could not find that, a colleague will reply” has created a ticket. IBM’s overview of chatbots describes the general mechanism; the hotel-specific rule is that anything touching a booking, a payment or a special request goes to a person, every time, until the system has run for a full season without an error the desk had to apologise for.

How to start with AI in a hotel

1. Count the direct enquiries for one month. Email, website form, phone messages transcribed. Note the hour each arrived and the hour it was answered.
2. Write down how a priced reply is built today. Which screen the desk checks, which rate rules apply, which upsells are offered. This becomes the model’s instructions.
3. Put the model on drafting only. It reads the enquiry, checks availability through the PMS if it has an API, drafts the reply. A person approves from a phone. Developers wire this with the provider’s tool-calling features; OpenAI’s function calling documentation shows the shape.
4. Add quote follow-up. Same mechanism: silence for two days triggers a drafted second message.
5. Load the arrival documents for retrieval. Start with the ten questions the desk answers most.
6. Draft review replies from the record. A reply that mentions the guest’s actual stay reads differently from a template. Google’s review management guidance explains the mechanics of responding on the profile.
7. Measure the same numbers after a month. Time to first reply, follow-up rate, desk interruptions, reviews answered.

What AI for hotels is not

It is not a website chatbot as a first step. Chat on the site captures a small share of intent and cannot quote a real price without the same integration the email workflow needs. Build the email workflow first, then decide whether chat is worth adding.

It is not in-stay automation. Room controls, digital keys and upsell screens are hardware and platform decisions with a different cost structure. This article is about the workflows a three-person front office runs from an inbox.

It is not revenue management or dynamic pricing either. Rate optimisation, demand forecasting and channel mix are a different discipline with their own tooling, usually owned by whoever runs distribution, not the front office. If “AI for hotels” brought you here looking for that, the four workflows above are the ones the front desk owns; pricing AI is worth a separate piece, not a paragraph bolted onto this one.

And it is not a replacement for the desk. The desk approves every reply that goes to a guest. The model removes the waiting, not the judgment.

Summary

For a hotel, AI pays first on direct enquiries, quote follow-up, pre-arrival questions and review replies, because those are the workflows where a guest is waiting on a person who is busy with another guest. The useful version is a narrow agent per workflow — tool access, one boundary, a human approval gate — not one general assistant that negotiates with guests unsupervised. The model reads and drafts, staff approve, and the baseline that proves it is time to first reply and follow-up rate. Booking changes, payments and complaints stay human. The same argument for restaurants is in AI for restaurants: the workflows worth automating, and the wider library is under AI for business. To map your own front-office workflow with me, start with workflow discovery.