Four rides out of 320 million
The numbers come from the companies themselves. Uber said on its Q3 earnings call that it handled roughly 320 million trips and orders during Halloween week, including four rides booked through AI agents and 500 Eats orders. Booking Holdings’ management said on its own call that referrals from large language models and personal assistants remain well below 1 percent of Booking.com room nights. PhocusWire reported the earnings-call commentary.
AI agents got a slot on the calls. The numbers attached to them were tiny.
Run the arithmetic. Four rides out of 320 million is 0.00000125 percent. One in 80 million. The 500 Uber Eats orders work out to about 1.6 per million. Put both together and you get 504 agent-fulfilled transactions out of roughly 320 million, which is 0.00016 percent, or about one in 635,000.
Nobody is claiming otherwise. The interesting part is the gap between capability and use. Agents can drive a checkout flow. Customers, for now, mostly do not ask them to.
Booking's under-one-percent
Booking Holdings’ management described the referral share as well below 1 percent of Booking.com room nights. Note what kind of statement that is. A ceiling is not a measurement. Below 1 percent could mean 0.9 percent, and it could mean 0.09 percent. The company did not publish the exact number, and I have not seen one from any other major travel platform.
For scale, Booking.com moves more than a billion room nights a year across its brands. Even at the top of that range, agent-mediated bookings remain a small side channel rather than a sales channel.
The Instinct exception
One name breaks the pattern. Instinct, an invite-only assistant platform, reports a run rate approaching $1 billion in annualized transaction volume, with roughly half of it tied to travel. That is about $500 million a year in travel bookings, or roughly $1.37 million a day. It is the only player in this set with numbers that sound like a business.
Two caveats before anyone gets excited. A run rate is an annualized snapshot, not audited revenue, and transaction volume is gross booking value, not commission. If such a platform took a 15 percent commission, that $500 million would translate to roughly $75 million a year in gross take. Instinct has not published a take rate, so treat that as arithmetic on my side, not a forecast.
The agents that did finish
A booking audit by Propellic tested 176 scripted requests. In all 176, the agent reached the payment step without being stopped by a login wall or a captcha. That is worth keeping next to Uber’s four rides, because it separates two different problems. Getting an agent through a booking funnel is an engineering task, and on this evidence it is largely solved. Getting a human to want that is not.
The audit measured navigation, not conversion. Reaching a payment page is not paying. Fraud checks, 3-D Secure and airline hold rules sit after the point where the test stopped.
United closes its front door
United Airlines banned third-party AI agents from booking through its app and website, citing account security and credential sharing. The reasoning is not complicated. An agent that logs in with your credentials looks, from the airline's side, exactly like you handing your password to a stranger. Until agents authenticate through delegated tokens instead of stored passwords, this objection will keep coming back, and it will keep being reasonable.
Europe gets this later
Meta launched its assistant Muse in the United States in September 2026, then expanded it to Japan and Singapore. No EU launch has been announced. Meta has also said it plans to charge transaction fees on business bookings the assistant facilitates, without publishing a rate. For a reader in Prague, Berlin or Madrid, the products generating these referral numbers are simply not available yet, which makes the European share of that sub-1-percent figure effectively zero for Muse-driven traffic.
The regulatory frame is already fixed, though. Article 50 transparency obligations under the EU AI Act have applied since 2 August 2026, so an assistant has to make clear that a user is talking to a machine, and generated content has to be labelled. The high-risk Annex III deadline, which much of the industry spent 2025 preparing for as an August 2026 event, moved to 2 December 2027 under the Digital Omnibus. Booking-related compliance pressure therefore arrives later than the roadmaps assumed.
GDPR is the nearer problem. An assistant that holds your payment details and passport data is a new processor in the chain, and the travel provider receives a request that looks like you but was assembled by a third party. Controllers need a lawful basis and a processor agreement for that. None of the earnings-call commentary addressed it.
What is still missing
Neither Uber nor Booking Holdings split its agent figures by country or by provider. The four rides could be anywhere. The under-1-percent referral share could be concentrated in one market. Until someone publishes that breakdown, the European slice of this story is unknown, and the only disclosed bound is a ceiling of one percent.
Does the AI Act force travel sites to accept bookings made by AI agents?
No. The transparency rules in force since 2 August 2026 require AI systems to tell people they are dealing with a machine. I found no provision that obliges a booking platform to accept agent traffic. Access is set by each company's own terms, which is how United could shut its app and website to third-party agents.
Do the Uber figures cover only outside agents, or also Uber's own assistant features?
Uber’s disclosure counts trips and orders fulfilled through AI agents. It does not break the total down by agent provider, by city or by country, so the split between first-party and third-party agents is not public.
Will customers pay more when an assistant makes the booking?
No published price difference exists yet. Meta has said it intends to charge businesses transaction fees for bookings its assistant facilitates, but has not named a rate. None of the travel companies on these calls disclosed agent-specific pricing, so any cost comparison today would be guesswork.
