Every travel executive I know can quote their look-to-book (L2B) ratio from memory. Very few can tell me their tokens per booking. That gap will matter much more than most boards expect—and soon.
On September 28, Instinct, a personal AI agent founded only last year, said it had raised $1 billion at a $10 billion valuation. It lives in your messages, plans trips, watches fares and even phones the hotel. Founder Noah Shinn told Patrick O'Shaughnessy that Instinct already runs at about $1 billion a year in transaction volume and that half of it is travel.
A week later Hark and Tab launched rival agents that also book travel. And then on October 14, Vayo came from Booking Holdings.
Instinct is free for travelers, and so is Meta's Muse, up to a usage limit. So, the question for our industry is what an agent like this costs to run, and who ends up paying for it.
The end of tokenmaxxing
Packy McCormick and Markie Wagner at Not Boring gave the idea a name I have used since Phocuswright Europe this June: return on tokens, or ROT.
ROT is the value of the output minus the cost of the tokens, divided by the cost of the tokens. Their best line: "Agents should do the thinking; code should do the doing."
For two years, nobody needed the formula. Model prices were falling fast, and vendors sold flat subscriptions. That period is over. This spring the large providers moved business customers to usage-based pricing, and Uber's CTO famously admitted the company had used its full 2026 budget for AI coding tools in four months. In June the Linux Foundation announced the Tokenomics Foundation to standardize how companies measure AI spend. And Booking.com is among its backers.
2 ledgers
In travel, ROT has two sides.
The first ledger belongs to the agent. Picture one that checks a fare every few hours for three weeks, compares forty hotels, calls the property and then books a $200 room. Last month Browser Use showed a lean agent running a Google Flights search in seven seconds, for less than half a cent.
Put simple numbers on it. An agent handles 1,000 requests, and 100 of them turn into bookings that stick, at $30 each. That is $3,000 in revenue. At 60 cents of model and tool cost per request, the bill is $600, because it is paid on all 1,000 requests, not just on the 100 that booked. Payment, distribution and support still come out of what is left.
The second ledger belongs to the supplier. I spent 11 years at Expedia, on both the buy and the sell side, and L2B was always a negotiation about who pays for the looking.
Agents will push those ratios much further because an agent doesn't get bored and doesn't go to sleep. Every check hits an airline or hotel system that earns nothing from the look. And agents don't only look. With free cancellation, an agent can book the best room, then the second best, then the third, hold all three and release the extras at the last minute. For the hotel, that is inventory held by a machine at no cost.
The split between looking and booking is already visible: Southwest now lets travelers search its fares inside ChatGPT, then sends them to Southwest.com to book.
Here the two ledgers meet. Right now, the only people paying for the look are the agents' investors. BTIG analyst Jake Fuller, quoted by PhocusWire, noted that neither Muse nor Instinct takes a commission cut (yet). Shinn has floated changing that and said boutique hotels would pay up to 30% per booking.
In that model, the supplier pays for the agent's looking as well as the booking. But an agent paid 30% by some hotels and nothing by others is no longer a neutral assistant. Travelers will want to know which room it is showing them and why.
Where the value is
Global distribution systems, online travel agencies and metasearch players hold something no agent startup has: years of abandoned sessions and query-to-conversion data. That is exactly what you need to build a demand quality score, a signal of how likely a given query is to turn into a booking. Whoever turns that into a product and sells it back to the agents creating the noise gets a new revenue line from the same problem that is eating margin today.
There is a quieter lever too: stop starting from zero. An agent that has to relearn my preferences on every trip pays again for work already done. Reusable traveler memory could cut that bill, though it still has to be measured. It also shapes distribution, because whoever controls what an agent knows about the traveler influences which options it shows. And memory never replaces a live check on price and availability. A cheaper answer is no use if the room cannot be booked.
Booking Holdings seems to agree. Vayo, the startup it launched on October 14 with up to $100 million planned behind it, builds a traveler profile from the confirmations in your inbox and learns from every trip. It is free in beta, and has not yet decided how it will monetize.
What boards should ask
Cost per token is a procurement number. Cost per completed trip, including the failed attempts and the fixes, is a strategy number.
I would want to know which parts of the flow really need a model, and which could run as plain code. I would want to know what happens to margin if usage doubles or the vendor reprices. I would ask how much we earn from a booking after paying for all the searches that went nowhere, and whether the agent still makes money when the traveler changes plans or plans with us and books somewhere else. On the supplier side I would ask how much of our search traffic is now machines and whether we charge anything for it.
Those rules are being written now. On October 6, Sierra and Meta announced the Personal Agent Protocol, nicknamed Poppy, an open standard for how personal agents identify themselves to businesses. Instinct, Shopify, Stripe and Walmart are launch partners, and United Airlines and Hertz are among the design partners. Once an agent identifies itself, a supplier can price the look. No hotel group is on the list yet, while Google has opened a waitlist for hotels on its own agent booking standard.
None of this makes personal agents a bad bet. But a big round doesn't prove a model works. The agents that last will be the ones that can show what their tokens actually build.
In travel we learned, slowly and at great cost, that a search is not a booking. Agents will teach us again, this time with an invoice for every look.
A convincing demo shows what an agent can do. Return on tokens asks whether we can afford to keep doing it.
About the author...
Eugenia Dunaeva is an independent director at
TravelAI and a VC and angel investor.