AI-referred traffic to U.S. travel sites grew 119% year over year in July, according to Adobe, and those visitors now book almost as often as everyone else. AI-referred traffic converted just 1% below non-AI traffic in July, compared with 47% below a year earlier. Chatbots are increasingly where trip planning begins, and we wanted to know whether the trip can end there too, with an agent that actually books, pays, and fixes things when they go wrong.
In March, we wrote that travel ranks last in agentic AI adoption. Airlines already use AI widely behind the scenes, with 63% running it in operations control, according to SITA, though mostly to recommend decisions rather than make them. The newer contest is on the consumer side, where personal AI agents, meaning software that acts on a person's behalf across apps and websites instead of only answering questions, are attracting serious money, and travel is emerging as one of their main use cases. Instinct raised $1 billion at a $10 billion valuation, and its founder has said more than half of the transactions on the platform are travel-related. Google switched on hotel booking inside AI Mode in August, and a month later Expedia agreed to sell hotel stays inside Meta's Muse. To see how far the technology has actually come, we ran the same trip through four personal AI agents, two generalists and two travel specialists, and timed each one from the first message to checkout. We haven't seen anyone publish a head-to-head like this one, and the results were clearer than we expected.
What happened when we tried to book
On October 19, 2026, we gave the generalists Instinct and Meta’s Muse and the travel specialists Miso and Odessia the identical prompt: “Can you find me a non-stop flight from NYC to SFO departing on 11/9 and returning on 11/13? Flights taking off from NY no earlier than 10 a.m. and flights arriving back from SF no later than 8 p.m.” We timed each agent from the first message to the moment it was ready to take payment and stopped there without purchasing. Because the tests ran back-to-back, the searches may have affected each other’s fares, but the fares all landed in the same ballpark. It's one trip and one prompt, so it's a snapshot rather than a benchmark, but the gaps were wide enough to be worth sharing.
The specialists were dramatically faster. Miso and Odessia each returned their first results over text in under 90 seconds, while Instinct took just under four minutes and Muse took seven and a half. Getting all the way to checkout took longer for every agent, as the breakdown below shows.
Miso was the strongest overall. It returned a top recommendation plus three alternatives in plain text and was ready to confirm inside the app in 5 minutes 53 seconds.
Odessia returned a single option and gave more only when asked. It reached payment in about 6 minutes 25 seconds but handed us off to an external website to check out ourselves.
Instinct returned six options as screenshots, recommended the cheapest, and was ready for payment in 8 minutes 42 seconds.
Muse returned four options as text and was the slowest by a wide margin, taking about 13 and a half minutes to reach payment.
Skift reported in September that Muse books flights through Duffel and pays with a Stripe virtual card, and our test showed Instinct running on the same setup, with Link by Stripe for payment and Duffel for flight inventory. Neither specialist disclosed its payment rails or airfare suppliers. That points to a pattern: generalists handling travel as one task among many reach for the same third-party building blocks, and vertical apps felt like systems that had done this many times before.
From search to delegation
Before 2023, travelers compared fares across browser tabs and OTAs, and from 2023 to 2025, AI could build a Lisbon itinerary that you still had to book yourself. In 2026, the traveler can set the goal and let an AI agent book, pay, and handle changes, with the human giving final approval. With a service like Miso, that can be as simple as a text that says "NYC to Austin Thursday and back on Monday, aisle seat only, use my status and points." Once the agent sits in front of the brand, the customer relationship moves one layer up the stack, along with the data and much of the pricing power that comes with it.
Two kinds of agents
The companies building here fall into two camps. Travel specialists like Miso, Odessia, Soar, and Mindtrip go deep on the trip itself. They've built, or are building, the ability to book through direct connections to the GDS, NDC, and suppliers, handle fares, points, and changes well, and often keep a human advisor behind the AI for exceptions and refunds. Generalists like Instinct and Meta's Muse run a person's whole life, with travel as one task among many, and usually book through browser automation or partner apps. Other players, such as Town and Google AI Mode, can take people partway through booking a trip but don’t yet handle all of it.
Each camp's strength is the other's weak spot. Generalists are shallow on refunds and irregular operations, while specialists are small and have to win a spot on someone's phone before they can win a booking. The capital gap is wide, too: Odessia's $6 million compares with $108 million in Town's latest round and $1 billion in Instinct's. Sequoia's recent backing of Odessia, a new travel AI agent, sharpens the question of whether travel ends up with a vertical winner or the generalists absorb it.
Our view is that vertical players can win, but not by out-delegating the generalists. Booking a work trip is a chore most travelers would happily hand off, but planning a vacation is different. A 2010 study of 1,530 Dutch adults found that people planning one were happier than those not going away, possibly because of the anticipation. A specialist that leans into that experience, surfacing offers and options a generalist is too broad or too busy to find, has a real place in the market. Our test pointed the same way, since the travel-specific agents were faster than the generalists and one handled the entire booking in-app. The harder problem is retention, since consumer travel is infrequent and switching costs are low, so the specialists that last will be the ones that give travelers a reason to come back between trips.
Trust is a ladder, not a yes or no
How much to hand an agent isn't a single decision. We think of it as six rungs: reading your email and calendar, suggesting options, holding a fare or building a cart, paying after you approve, acting alone within spending caps, and full autonomy with your logins and cards. Most of the risk sits in the top two rungs, and for travel today we think paying after the traveler approves is the right default. That payment layer is getting built right now through OpenAI's Agentic Commerce Protocol, Google's AP2, Visa Intelligent Commerce, and Mastercard Agent Pay, but checkout still lags. Google AI Mode books hotels, not flights, and OpenAI pulled back on Instant Checkout in ChatGPT months after launching it.
What this means for airlines
Airlines are starting behind. Adobe scored airline homepages at 42% readable by AI models, compared with 71% for hotels, and route and fare pages at 59%, so an agent can miss parts of the offer it's supposed to book. Startups like Bonafide are working on that gap by helping travel brands, airlines included, make their content accurate and bookable inside LLMs like ChatGPT and Gemini. If agents become the storefront, airlines risk disintermediation, and price-optimized bookings can bypass the loyalty programs carriers have spent decades building. There's also no settled answer yet for who owns the error when an agent misbooks or misquotes a fare.
Airlines still hold real advantages, though. They own the inventory and the servicing, especially once something goes wrong, which is the moment a general-purpose assistant is least equipped to handle. Disruption is the highest-value opening for agents, since rebooking a passenger before they reach the gate is worth far more than shaving a few dollars off a fare. Airlines can also set commission and loyalty terms for agent bookings now, before the big platforms set them on their behalf.
From the airline’s side, the real cost of an agent booking is the relationship rather than the fare. If an agent books through a path the carrier can’t see, the airline may not be able to message that passenger when their flight is canceled, rebook them, or sell them a bag or a seat. When an agent clicks “accept” on a fare’s rules, it’s unclear who agreed to what. We think airlines have more to fear from losing the customer behind the booking than from price.
We expect the shift to happen in stages. Over the next year, agents will layer on top of existing distribution rather than route around it, since NDC is making offers machine-readable and Muse already reaches airline inventory through Duffel. The year after, the payment rails should mature enough for agent-initiated purchases to be verifiable and trusted. Beyond that, we expect fully autonomous booking to arrive first in managed corporate travel, where Navan already gates bookings by company policy and newer players like SkyLink and BizTrip AI are building agents that book within corporate policy from the start. The next 12 to 24 months will likely decide who owns the booking moment.
What We're Watching
We're most interested in the parts of this market that general-purpose assistants can't easily do themselves. That starts with specialists selling booking and servicing to other agents as infrastructure, which is the model Soar and Voyagier are pursuing with MCP servers built to let Claude and other AI agents book real travel. Premium and corporate travelers are the natural first customers, since frequent flyers and policy-driven business travelers pay and come back. Agent identity, spending limits, and dispute handling are also still fragmented, which leaves room for a trust and payments layer built for travel.
If you're building where personal AI agents meet travel, reach out at inbound@sky-vc.com.
.jpeg)