When the guest is a machine, price alone won’t win the booking. See how governed data helps hotels compete for AI agent travel bookings.
In our family, my wife Ana finds hotels every time we go on vacation. Because the one time the job fell to me I booked the first place that met the right price, right dates, free cancellation. Done in ninety seconds. It met every condition except the ones I didn’t even think about, such as “not directly above a nightclub”. Ana does the work I skip: dozens of listings, prices compared, policies checked, reviews read (the most negative first).
She has a new competitor, though, and it books like me, not like her. AI agents have begun to compare, book and pay on travellers’ behalf — fast, literal, and supremely confident.
From the hotel’s side of the counter, the same booking looks strange. Picture the reservations desk of a small hotel on the Algarve coast. A booking lands: three nights, arriving Friday, breakfast included, free cancellation, under a set price. No phone call, no question about the sea view, no reply to the email offering a room upgrade for twenty euros a night. The booking was not made by a person. It was made by software acting for one, told to find a room that meets a set of conditions. The upsell that pays for the off-season went unread, because nobody was there to read it.
A new kind of buyer has arrived, and hotels have met one before. In the 2000s the online travel agencies commoditised distribution, taught travellers to sort by price and charged a commission for standing between the hotel and its guest. Hoteliers learned, slowly, to defend the direct relationship and to make the stay itself the product.
The agent is the next instalment of that story, with one twist: the buyer itself is now automated, and the toll is being levied again. OpenAI’s in-chat checkout charges a fee on purchases its agent completes — early reporting puts it at 4%. The travel agency charged the hotel to reach the traveller; the agent charges to reach the traveller’s software.
Travellers learned the same lesson from the other side. They screen on the platforms, then book direct — a phone call to the hotel — to dodge the markup. The agent closes that escape, because it is the screening and the booking at once, and it does both without ever lifting the phone.
The commoditisation runs further than hotels, and further than it has run before. When the buyer is an agent, every business whose product fits inside comparable fields — price, spec, delivery, terms — faces what the hotel faces. The numbers arrive as forecasts and deserve a forecast’s skepticism.
Adobe measured a 4,700% jump in generative-AI traffic to American retail sites in a single year, from a small base. Morgan Stanley expects agentic shoppers to account for around 10% to 20% of US online retail spend by 2030, with the market impact reaching as much as $385 billion. Gartner forecasts that by 2028, AI agents will intermediate 90% of B2B buying, channeling more than $15 trillion in spend through automated exchanges.
The first instinct, faced with a buyer that ranks on price, is to compete on the price it ranks on — cut the rate, match the cheapest listing, win the booking. That surrenders margin on the only thing the agent can weigh and gives away the things it cannot. An agent reads a rate, a date and a cancellation policy. It does not read the welcome, the room that smells of salt, the concierge who knows which kitchen is worth the walk. None of those has a field in the booking record, which is the very reason they hold their value while the rate is bid to the floor. The task of booking is shrinking, but the work of hosting is not.
Let the machine choose your hotel
Retreating to what the agent cannot value is half right and half dangerous. If everything comparable collapses to price, the agent never shortlists you, and the human never arrives to feel the welcome. So how the hotel reads to the agent — and inside the chatbot the traveller actually asks — comes to matter as much as how it reads to a person. Being chosen by the machine is the new price of being met by the guest.
A hotel can turn the same technology on the same problem. Your quiet advantages — a location set back from the road, communicating rooms for families, real distance to the beach (twenty metres, not “close”), flexible cancellation, a late checkout — count only if the agent can read them.
Governed data is the new defence
Ana digs them out of the reviews, worst first; the agent sees only what it is handed as clean, structured data. Publishing prices, availability, room types, policies, amenities, location and experiences in one consistent, machine-readable form — the web has a plain standard for it, JSON-LD — is the hotelier’s version of an old truth: governed data is what protects you when the buyer is a machine.
The upsell changes address, too. The twenty-euro upgrade nobody read in an email now has to live inside the offer the agent sees, put in front of the machine and the traveller in the flow, not kept for a check-in conversation the booking no longer passes through.
The second move is to wire that inventory into the agentic world now, while doing so is still a choice. The shift Gartner describes in business buying — data feeds and agent-ready catalogues becoming the entry ticket — is reaching travel. The hotels that connect their availability and terms to where the agents shop will be in the comparison; the ones that do not will be absent from it, which is worse than losing on price.
Tourists will win, too
For the traveller, none of this is a loss. The agent does Ana’s forty tabs in seconds, sidesteps the markups I never spot, and hands back a decision that is faster, cheaper and better informed. The buyer’s side of this is plainly good, which is the surest sign it will happen.
The hotel is also a buyer, and on that side the agent crosses onto the firm’s own ledger. Corporate agents already reorder supplies, renew software and settle invoices on a company’s behalf in early deployments. That is a convenience until one buys badly — pays a plausible but fraudulent supplier, waves through a renewal it ought to have queried.
The card networks are laying the rails for precisely this, testing agent-initiated payments inside preset limits. The limit travels with the agent; the responsibility does not. When the software spends the firm’s money, the firm answers for the result — the chargeback, the bad supplier, the contract nobody read. Authorship moves to the machine; accountability stays where it was.
Does the decision have an undo button?
The question facing a manager has changed. It is no longer whether an agent can perform a task — it usually can. The question is whether the firm can live with it performing the task badly. Two further questions settle that. Does the action have an undo? And does it move the needle? Sort what you let an agent do along those two axes and three zones appear. Reversible and small: let it run, no review — reorder the consumables. Reversible and large, or irreversible and small: act, then answer — the agent acts, but logged, capped and revocable, with the switch in your hand. Irreversible and large: you sign — the agent prepares and recommends, the commitment is yours, with no autonomy at all.
Delegate the execution, not the responsibility.
Delegate the execution, not the responsibility — what to let an AI agent do, by reversibility and size. Illustrative — schematic, not data.
Some businesses reach instead for a spending cap, and a cap watches the wrong axis. A small but irreversible action — that fraudulent supplier, that forced renewal — does more damage than a large reversible one and slips under any euro limit.
Governing agents by amount alone guards magnitude and ignores reversibility, the axis that actually carries the risk. Monday’s exercise is short. List what you have already handed to agents, drop each into the grid, and look hard at anything sitting top-right with no human signature against it. Those are the ones that will wake you at three in the morning.
The hotel that survived the travel agencies did not do it by being cheaper; it did it by owning the part of the visit the middleman could not sell. The agent sets the same test and marks it faster, and it asks for two things at once: be legible enough for the machine to choose you, and good enough for the guest to come back. Drawing that line — what to hand the machine, what to keep human — task by task, across the business, is the larger part of running one now. The agent can book the room, the way I would. Reading the room is still Ana’s job.
Most businesses are already facing this same question, just with a different machine. If you’re working out where your own AI agents belong on that grid — and where a human still has to sign — talk to us.
