When the AI Cancels, Nobody Calls the Kitchen
At 6am a chef in Sonoma buys eight pounds of halibut for a dinner that night. Table for eight, deposit on file, allergy note about shellfish, a guest who called three weeks ago and talked through the menu like he meant it. At 4pm the booking cancels itself. No call. No text from the guest. Just a notification: reservation updated by assistant. The fish is already bought. The prep is half done. There is no one on the other end of the line to say "we already bought the fish," because there is no line anymore. There's an agent that did what it was told to do, which was optimize the guest's evening, and it did that job well. It just didn't do the chef's job at all, because that was never its job.
This is the part of agentic booking that gets skipped in most of the talk about visibility and discovery. Everyone's worried about whether an AI can find you. Fewer people are worried about what happens after it finds you, books you, and then decides, on the guest's behalf, to change its mind. That second part is where the money actually moves.
Is agentic booking just about getting found?
No, and the industry is already past that question. A travel bot called Instinct, which works over SMS, WhatsApp, and voice, handles the whole loop: search, booking, rebooking, fare monitoring, credits, seat changes, all inside one conversation, no new app required. Trade press covering it in September called it proof that "frictionless" now means something specific: the agent doesn't stop at a recommendation, it executes the transaction and then keeps managing it after the fact. Meta's travel agent, Muse, does something similar for flights through the Duffel API, with accommodations on the roadmap. Google has already turned on agentic hotel booking inside its AI Mode search, letting a traveler complete a room reservation without ever landing on the hotel's own site.
None of that is about restaurants or dive boats or wine tastings yet. But the shape of it is obvious, and the shape is: discovery, booking, and post-booking changes are all becoming one continuous machine conversation. A guest doesn't call to cancel anymore. Their assistant does it, or doesn't even bother telling them it did.
For a hotel with two hundred rooms, one cancelled reservation is a rounding error. For a chef who bought fish for eight, a guide whose boat seats six, or a sommelier who blocked a private room for a tasting menu that took two days to plan, one cancellation is the whole shift. Independent operators live and die on small numbers. That's exactly the kind of business an agent optimizing for guest convenience has no reason to protect.
What deposits were actually for
The deposit was never about distrust of the guest. It was leverage. It was the one mechanism a small operator had to make a no-show cost something to somebody. A hotel can resell a cancelled room to the next walk-in. A private chef cannot resell a Tuesday night dinner for eight at 3pm on Tuesday. A dive guide cannot resell four seats on a boat that already left the dock. The deposit, the 48-hour cancellation window, the card on file, these exist because the inventory is non-refundable in a way that a hotel room simply isn't.
A human calling to cancel understands, on some level, that they're breaking a promise. They apologize. Sometimes they negotiate. Sometimes the operator waives the fee because the guest sounds sick, or because it's a regular, or because the operator has a gut read on the situation that no software has. That gut read, built over years of answering the phone, is worth actual money. It's the difference between eating the loss and not eating it.
An AI agent instructed to "cancel if the weather looks bad" or "rebook if a cheaper slot opens" doesn't carry any of that context. It doesn't know the fish is already bought. It doesn't know this is the third time this guest has done this. It knows the cancellation window listed on whatever page it read, if it read one at all, and it will act right up to the edge of that window and not a minute before, because waiting to the deadline is the economically rational move for the guest, not for you.
Do experience platforms become the new OTAs?
There's a cost detail buried in recent reporting that matters more than it sounds. AI assistants are billed per web search, and developers cap how many searches an agent runs per request. That means an assistant comparing dinner options doesn't have unlimited budget to go crawling ten independent websites. It has a strong incentive to hit one aggregator that already has structured menus, prices, and availability for two hundred restaurants in one query, rather than parsing your site, then the next chef's site, then the guide's PDF.
That's exactly the role OpenTable, Resy, Tripadvisor Experiences, and Airbnb Experiences already play for humans browsing, and it's the role they're positioned to keep playing for agents, for the same reason Booking.com keeps showing up in hotel searches: it's the cheap, known, already-structured answer. Marriott's CEO said publicly this year that AI agents are more likely to hurt OTAs than big hotel brands, because an agent optimizing for price and reliability will book direct when direct is easy to find. The flip side of that, which he also said plainly, is that independents without loyalty infrastructure or clean digital footprints are the ones left exposed. Swap "loyalty infrastructure" for "a phone number and a clear cancellation policy" and the same logic applies to a six-top tasting room.
So the aggregator risk for chefs and guides isn't really different from the OTA risk for hotels. It's a fee that gets added back into the transaction by an agent choosing the path of least resistance, on a booking the guest might have made directly if the agent had bothered to look.
What to actually do about it
First, stop keeping your cancellation and deposit policy in a PDF, a verbal explanation on the phone, or a line buried in a confirmation email. Put it in plain text on the page where your availability lives; make the terms something a parser can read in one pass, not something it has to infer. An agent that can read "48-hour cancellation, $75 deposit, non-refundable inside 24 hours" will factor that into its plan. An agent that finds nothing will assume flexible terms, because flexible is the safe default when nothing says otherwise.
Second, keep one human step in the loop for anything with real money on the line, a large party, a multi-day charter, a full buyout. That can be a required phone confirmation, a signed deposit form, anything that forces a human moment before the transaction locks. A recent piece on trust in travel bots raised the same concern from the guest's side: can a person reach a human, understand what the bot decided, and reverse it if it got something wrong? That door needs to swing both ways. You want it too.
Third, don't panic about the current numbers, because the current numbers are still small. One financial analysis of Booking Holdings this September found AI-driven bookings still under 1% of room nights, and traffic from AI tools to Booking.com also under 1%. The volume isn't there yet. What the same report found, though, is that when AI referrals do happen, they overwhelmingly route to Booking rather than direct sites. That's the part worth sitting with. The share is small today. The habit is already forming, and habits are hard to unwind once an agent's default routing gets baked into millions of conversations.
The honest version of this moment is that nobody, including the companies building these agents, fully knows yet whether they'll end up sending more direct business to small operators or simply adding a new layer of fees and cancellations on top of the old ones. What's clear is that the operators who write their terms down in plain language, keep a human reachable at the moments that matter, and check regularly how an assistant actually describes their business, which is one of the things you can do on bookaible, are the ones who'll have some say in which way it goes. The ones who don't will find out the hard way, the way the chef in Sonoma found out: a notification instead of a phone call, and eight pounds of halibut with nowhere to go.