AI Agents
Getting your hotel bookable by an AI agent: the agent-ready distribution stack
Google now books hotels inside AI Mode, ChatGPT plans trips with live pricing, and the launch partner list was OTAs and the big brands, not the independents. The question has flipped from 'should we build an agent' to 'can an agent buy from us.' Here's what agent-ready distribution requires: real-time ARI, machine-readable rates, delegated payment, and where a unified guest profile is the difference between getting booked and getting bypassed.
For a decade, the hotel distribution question was “which channels do we sell through?”. OTAs, metasearch, GDS, the direct booking engine you spent years optimizing. In 2026 a different question became the more important one: can a machine buy from you at all? When Google turned on agentic hotel booking inside AI Mode in late August, the launch partners were Expedia, Booking.com, Priceline, Trip.com and the big brands. Marriott, Hilton, IHG, Choice, Wyndham. Independent hotels were not on the list. That’s the whole story in one detail: the agents are here, they’re transacting, and the properties whose inventory they can’t cleanly read are being routed around, silently, at the exact moment a guest is deciding.
We’ve written the demand side of this before. How a travel brand builds an agent that books without inventing a fare. This post is the mirror image: the supply side. Not the agent you build, but the plumbing that lets an agent you don’t control (ChatGPT, Gemini, a Sabre- or Amadeus-powered assistant) discover, price, and book your rooms correctly. One 2026 analysis of agent-readiness put the share of travel companies whose stack an AI agent can “query, compare, and book at machine speed” at around 11%. Whatever the precise number, the shape is right: most hotels are optimized for a human shopper who navigates a website, and an agent never sees the website. Here’s what the other 89% are missing, and where a Salesforce data foundation fits.
The demand is real; the trust is not, yet
Start with the tension, because it decides how aggressively to invest. On one side, AI-assisted trip planning has gone mainstream fast: Phocuswright found 56% of U.S. leisure travelers used AI for at least one trip, up from 43% just nine months earlier, one of the fastest behavior shifts the industry has seen. On the other side, travelers still don’t trust a machine with their card. Skift’s read is that only about 2% are willing to let AI book fully on their behalf; Expedia’s “AI Trust Gap” survey put the comfortable-booking figure at 8%, with 68% still preferring to complete the purchase with a trusted brand, and the top blockers being loss of control and payment privacy (each cited by well over half).
That gap is not a contradiction to resolve. It’s the market’s shape for the next couple of years, and it tells you what to build. Travelers are happy to let an agent discover and compare; they hesitate to let it transact. So the near-term value of agent-readiness isn’t “capture fully autonomous bookings.” It’s don’t get filtered out of the comparison, because the agent that assembles three options for a human to choose from is already the default, and if your rates and availability aren’t legible to it, you’re not one of the three. Even Google’s own launch is built around this: the traveler picks the room, reviews the terms, and approves the payment; the “agentic” work happens behind the scenes. The machine shortlists. The human still clicks buy. Being on the shortlist is the game.
What “agent-ready” means: three layers
Strip the marketing off “agent-ready” and it resolves into three concrete data problems. An agent has to be able to (1) see your inventory, (2) understand it precisely enough to map a request to the right room and rate, and (3) transact against it. Miss any layer and the agent gives up on you without a word.
Layer 1 (real-time ARI the agent can query. ARI is the industry’s shorthand for Availability, Rates, and Inventory) the three facts every channel needs. Today ARI flows from your PMS or CRS through a channel manager out to OTAs, metasearch, the GDS, and your booking engine, with reservations flowing back to the PMS as the one definitive record. The channel-manager API, the layer that already pushes ARI to channels without anyone logging into an extranet, is the natural attach point for an agent-facing feed. The requirement an agent adds is freshness to the query: a rate cached from this morning is a rate that no longer exists, and an agent that quotes it creates a commitment you may have to honor or an abandonment when the price changes at checkout.
Layer 2, machine-readable, room-level structure. This is where most properties fail. An agent matching “a quiet room with a king bed, late checkout, refundable” to your inventory needs your room types, rate plans, and, critically, your rules (refundability, change fees, restrictions) as structured data, not as prose on a landing page or a footnote in a confirmation email. The web’s answer is schema.org markup (Hotel, amenityFeature, aggregateRating, geo), and the adoption data shows the bottleneck plainly: one 2026 study found that while roughly 56% of hotels have some JSON-LD, only about a tenth have complete, “good” markup. The convergence worth internalizing: the same structured data that gets you recommended in an AI answer is what makes you bookable by an agent. Visibility and bookability are becoming one infrastructure problem.
Layer 3 (a path to transact. Discovery and comparison are worthless if the agent hits a wall at “book.” That means a programmatic booking-and-change surface and, increasingly, a way for the agent to pay) which is its own emerging standard, covered below. This is the layer where the machine goes from planner to booking system, and it’s the one that’s new.
An agent doesn’t browse your site and give up politely. It queries a feed, and if the feed is missing, stale, or unstructured, it routes to inventory it can read, an OTA’s, and you never see the request that didn’t come.
The rails the agents ride: MCP and agentic payments
Two standards are becoming the connective tissue, and both are worth knowing by name because your integration roadmap will run through them.
MCP for inventory. The Model Context Protocol (an open standard originally from Anthropic, now broadly adopted across assistant makers) is effectively a universal API connector designed for LLMs. Instead of every agent maker building a bespoke integration to your systems, you expose your ARI, rates, and policies through an MCP surface an agent can call for real-time context, and it works across ChatGPT, Claude, Gemini, and custom agents alike. This isn’t theoretical for hotels: Cendyn’s AI Connect pushes hotel ARI into those assistants over MCP, RateGain has described an MCP-enabled booking engine, Booking.com publishes an MCP server, and CRS vendors are shipping MCP connectors that let an agent run natural-language availability and rate-parity queries across thousands of properties. If you’ve been tracking MCP as the enterprise integration story, this is that same standard aimed at the distribution problem: one governed surface, many agents.
Agentic payments for the transaction. The reason an agent can’t simply “put in a card” is that nobody wants to hand a raw card number to an autonomous program. So a cluster of 2026 standards solve delegated payment the same way: the agent never holds the card. It presents a tokenized, scoped, revocable credential plus a signed mandate describing exactly what it’s allowed to buy, and the network authorizes it in real time within pre-set guardrails. The players: OpenAI and Stripe’s Agentic Commerce Protocol for the checkout flow; Google’s Agent Payments Protocol (AP2), launched with 60+ partners including the card networks, PayPal, and Salesforce, using signed Intent, Cart, and Payment mandates; Mastercard Agent Pay and Visa’s Trusted Agent Protocol binding tokenized credentials to a specific agent, merchant scope, and spend cap. This is the same machine-readable-checkout shift we mapped for agentic commerce on the retail side. Travel’s version is just governed by perishable inventory and per-rate rules instead of a product catalog.
Notably, Google’s hotel booking runs on its own commerce plumbing (reported as a “Universal Commerce Protocol”) with Amadeus as the first B2B lodging partner: a reminder that the GDS and travel-tech incumbents are positioning themselves as the bridge between your inventory and the consumer agents, which is a strategic choice you’ll have to make deliberately rather than by default.
The OTA question, and why the brands are (mostly) calm
Here’s the strategic wrinkle that should shape where you spend. The conventional fear is that AI agents disintermediate hotels. The more precise read, voiced by the people with the most to lose, is that agents disintermediate the middle first. Marriott’s CEO has said plainly that AI booking agents are a bigger threat to OTAs than to major brands with a mature direct channel and a loyalty base, and Marriott’s CFO has floated that AI-agent bookings could even come in cheaper than OTA commissions. The logic: a brand with a direct relationship and a real loyalty program (Bonvoy, Honors) has something the agent’s evaluation can weigh; a commoditized room sold through an OTA is exactly the kind of undifferentiated inventory an agent arbitrages in milliseconds.
But “mostly calm” is not “unconcerned.” In their early-2026 annual filings, both Marriott and Hilton formally named AI and large language models as a risk to direct bookings. Hilton’s language warned that “the entry of major technology platforms, such as large language models, into the internet travel bookings business” could “divert bookings away from our direct channels and increase our hotels’ cost of sales.” The first time that risk appears in a 10-K is a signal. And there’s a genuine counter-narrative worth taking seriously: because today’s AI answers frequently surface OTA listings, some analysts argue AI is strengthening the OTAs in the short run, leaving the independent hotel unseen at the decision moment. The strategic takeaway is not “agents will save direct booking.” It’s: agent-readiness is how you compete for the direct relationship in an agent-mediated world instead of ceding it back to the intermediary. That’s the same fight the loyalty program and revenue-management functions are already having, now moved to the distribution layer.
Where a unified guest profile earns its place
For a hotel group already on Salesforce, or considering it as the data spine, the use is not “Salesforce sells rooms to agents.” Be precise about this, because it’s easy to overclaim: there is no out-of-the-box Salesforce feature that publishes your ARI as an outbound feed to ChatGPT. What Salesforce is good at is the layer underneath the agent-facing surface: the governed, real-time guest and offer context that makes an agent interaction correct and personalized instead of generic.
- A unified guest profile as the grounding layer. Data 360 unifies reservations, loyalty status, stay history, and preferences into one profile, and its zero-copy approach can bring in operational data (a POS system, a rate store in a warehouse) without physically moving it. Salesforce’s own hospitality reference app demonstrates the pattern: a unified guest profile driving agent actions to raise booking likelihood, upsell, and loyalty. When an agent (yours, on the direct channel) knows this is a returning Bonvoy guest who always wants a high floor and a late checkout, the offer it assembles is one an OTA’s commodity view structurally cannot match. That is the direct-channel advantage, made legible.
- A first-party guest agent on the direct channel. Agentforce for hospitality handles the guest-service surface (reservation changes, cancellations, loyalty redemption) and the point of investing there is to make your own conversational channel good enough that a guest completes the booking with you rather than delegating it to a neutral agent that treats your room as interchangeable. One business-travel platform’s Agentforce deployment reportedly resolves around half of its chat cases with no human touch; the same capability pointed at booking is how a brand keeps the relationship.
- Alignment with the payment rails. Salesforce is among the AP2 launch partners, which matters less as a feature than as a signal: the platform holding your customer relationship is wiring into the same delegated-payment standards the consumer agents will transact on. The architecture to aim for is your governed offer-and-profile data behind an MCP or API surface an approved agent can call, a thing a consultancy builds on top of Data 360, not a checkbox you enable.
What to do
Audit the three layers before you build anything. Can an agent get your real-time ARI through your channel manager or CRS API, current to the query? Is your room-and-rate structure machine-readable (schema markup on the web, structured rate rules in the systems an agent would call) or does refundability live only in prose? And do you have a transaction path an agent can reach? Most hotels discover the failure is Layer 2: the data exists, but only in a form a human can read. That’s the project, and it’s unglamorous data work, not an AI project.
Then sequence it by the trust curve. Because travelers still want to approve the purchase, prioritize being cleanly discoverable and comparable (accurate, fresh, structured ARI exposed through the rails agents ride (MCP today, the payment protocols as they mature)) over chasing fully autonomous checkout the market isn’t ready for. Ground your own direct-channel agent in a unified guest profile so it can offer something an intermediary can’t. And decide deliberately whether the GDS and travel-tech bridges (Amadeus, Sabre) are your path to the consumer agents or a dependency you’d rather not deepen, because right now, defaulting means letting someone else own the connection to the buyer.
The properties that win the agentic transition won’t be the ones with the slickest chatbot. They’ll be the ones whose inventory an agent can read, price, and trust, because in a world where a machine assembles the shortlist, being unreadable isn’t neutral. It’s being invisible at the one moment the booking is decided.
Understanding the basics
What does it mean for a hotel to be “agent-ready”?
Agent-ready means an AI booking agent (ChatGPT, Google’s AI Mode, a Sabre- or Amadeus-powered assistant) can query your inventory, understand it precisely, and transact against it without a human navigating your website. Concretely it requires three layers: real-time Availability, Rates, and Inventory (ARI) an agent can pull fresh to the query; machine-readable, room-level structure including rate rules like refundability and change fees (via schema.org markup and structured data in your CRS/PMS, not prose); and a programmatic path to book and, increasingly, to pay. One 2026 analysis estimated only around 11% of travel companies currently clear that bar, most are still optimized for a human shopper an agent never becomes.
Will AI booking agents replace OTAs or hotel direct booking?
The people with the most at stake think agents threaten OTAs before hotels: Marriott’s leadership has argued that brands with a mature direct channel and real loyalty programs have differentiation an agent can weigh, while commoditized OTA inventory is exactly what an agent arbitrages. But both Marriott and Hilton named AI and large language models as a risk to direct bookings in their 2026 filings, and some analysts note AI currently strengthens OTAs by surfacing their listings first. The honest read: agent-readiness is how a hotel competes for the direct guest relationship in an agent-mediated market, not a guarantee, but the price of staying in the comparison.
How do AI agents pay for a hotel booking on a traveler’s behalf?
Through delegated-payment standards that never expose a raw card number to the agent. The agent presents a tokenized, scoped, revocable credential plus a signed “mandate” describing exactly what it may buy, and the card network authorizes it in real time within pre-set guardrails like a spend cap and merchant scope. The 2026 standards doing this include OpenAI and Stripe’s Agentic Commerce Protocol, Google’s Agent Payments Protocol (AP2), Mastercard Agent Pay, and Visa’s Trusted Agent Protocol. In practice today most consumer flows still keep a human approving the final payment. Full autonomy is where the plumbing is going, not where traveler trust is yet.
Where does Salesforce fit in agentic hotel distribution?
Underneath the agent-facing surface, not as the surface itself. There’s no out-of-the-box Salesforce feature that publishes your rates to external agents, but Data 360 unifies reservations, loyalty, and preferences into a real-time guest profile, with zero-copy access to operational data, that grounds an agent so its offers are personalized and correct rather than generic. Agentforce runs your first-party guest-service and booking agent on the direct channel, which is how a brand keeps the relationship instead of ceding it to a neutral agent. The pattern a consultancy builds is your governed offer-and-profile data behind an MCP or API surface an approved agent can call.
Trying to work out whether an AI agent can read, price, and book your inventory, and what the unglamorous data work to fix it looks like? Talk to us. Getting a real-time, governed data foundation under the guest experience, on Salesforce or alongside it, is exactly the architecture we do.
