In almost any business, selling has always meant the same thing: bringing a person to your shop window and persuading them to buy. First it was the store, then the website, then the app. In tourism the same thing happens, with one important difference: the one arriving at your shop window may no longer be a person. It is an artificial intelligence assistant that searches, compares, books and pays on behalf of a customer who never opened your page. The customer tells it what they want — “a hotel near the park, under $150, for two nights” — and the agent does the rest. Understanding that the sale no longer starts on your website but in a conversation with a machine is what separates a business that will keep selling from one that, without realising it, will stop appearing.
From the assistant that recommends to the assistant that buys: what changed in 2025
The first difference is the jump from recommending to executing. Until recently, an AI assistant helped you decide: it searched for options, compared prices and handed over links. The customer still bought on the seller’s website. That changed in 2025, when assistants started completing the purchase inside the conversation. On 29 September 2025, OpenAI launched Instant Checkout in ChatGPT together with the Agentic Commerce Protocol (ACP), an open standard developed with Stripe, starting with Etsy sellers in the United States. Google announced its Agent Payments Protocol (AP2) on 16 September 2025, and in January 2026 presented the Universal Commerce Protocol (UCP) together with Shopify. Visa announced its Intelligent Commerce programme on 30 April 2025, and Mastercard launched Agent Pay on 29 April 2025. It is not one product or one company: it is the whole payments chain building, at the same time, the way for an agent to buy for you. And in tourism there are already concrete examples: hotel booking inside Google’s AI Mode has been live in the United States since 27 August 2026, and the Booking.com and Expedia apps inside ChatGPT launched in October 2025.
Three layers to understand: discovery, purchase and payment
The second thing worth understanding is that agentic commerce is not one protocol, it is three layers. The first is discovery: the agent does not see your web page, it reads your data. The Model Context Protocol (MCP), the open standard Anthropic published in November 2024, is the most common way for an agent to query your catalogue, availability, prices and policies in real time. If that data is incomplete, out of date or in different formats, the agent recommends someone else. The second layer is the purchase: ACP covers payment and completing the purchase, and UCP covers the whole journey — catalogue search, cart, payment and orders — with a lodging variant (UCP for Lodging) that is still a draft. The third layer is payment: AP2 sells nothing, it proves with cryptographically signed mandates exactly what the user authorised; Visa and Mastercard issue tokenised credentials tied to one specific agent, with spending limits set by the customer. None of these layers is settled: in March 2026 OpenAI scaled back in-chat payment and returned completing the purchase to the seller’s website, focusing on discovery. Anyone betting everything on a single protocol is betting on a standard that is still being written.
In tourism the risk is not technical, it is about the channel: direct or through the OTA
The third point is the one that hurts most in tourism: if your inventory does not reach the agent through your own channel, it will reach it through another one. Today most of the hotel supply agents see comes from online travel agencies (OTAs), not from each hotel separately. Without a direct connection, the demand the agent creates settles by default on OTA inventory, on its terms and with its commission. Google built UCP for Lodging precisely to connect hotels with its AI Mode, and there are already MCP servers connecting a hotel’s reservation and management systems directly to AI platforms, without replacing the infrastructure. For a park or an attraction the logic is the same: the session, ticket or photo the agent cannot read as structured data — real availability, final price, cancellation policy — is a sale the agent will make somewhere else. The question is not whether your website is pretty, but whether your data can be bought without going through your website.
Who answers when the agent gets it wrong: limits, cancellations and trust
The fourth cost is trust, and it is what holds all of this back most. Most agent purchases still need a person to approve the final payment, and that is an advantage, not a flaw: it is the control that stops an agent booking the wrong date, the wrong room or a price that changed an hour ago. That is why the protocols insist on the same thing from different angles: mandates recording what the user authorised, credentials that only work for one agent and one merchant, and spending limits set by the customer. For a tourism business, that translates into three operating rules: define what an agent may buy without human intervention and what it may not — a ticket yes, a group package perhaps not; put in writing what happens if the agent books wrongly — who cancels, who refunds and how quickly; and verify that the agent asking for your data is who it says it is, because a fake agent stealing availability or prices is a new problem with old tools. Trust is not won with a protocol, it is won with clear rules before the first sale.
How to prepare: five questions before opening the door to agents
The question is not whether agentic commerce will arrive, but whether your business will be ready when it does. Check five things: whether your data is structured and in real time — availability, final price including taxes, cancellation and refund policies — in a format an agent can read, and not only on a pretty page; whether all your sales draw from the same inventory, so the agent does not sell a session that already sold through another channel; whether you have defined the limits of what an agent may buy alone and what needs human approval, with a record of that authorisation kept; whether your cancellation and refund policy is written so an agent can understand and apply it, without having to call anyone; and whether you already measure the traffic arriving from AI assistants — how many arrive, how many ask about availability and how many buy — instead of mixing it in with ordinary web traffic. If any answer is no, the next step is not to sign up for a protocol, but to tidy up your inventory data. The agent will not wait for your business to be ready: it will sell for whoever is.
Sources
- Agentic Commerce Protocol (ACP) Explained — Swell
- Agentic Commerce in 2026: Protocols, Companies, Market Map — Massive
- Agentic Payments Explained: How Visa, Mastercard, Google AP2 and ACP Let AI Agents Pay — AIfintent
- What Is Agentic Hotel Booking and How Should Hotels Prepare — Revfine
- Agentic AI in Travel Commerce 2026: Agents, MCPs, Workflows — AltexSoft
- A Retailer’s Guide To AI Shopping Protocols: ACP, UCP, And MCP Explained — Ecommerce Fastlane

