Let’s talk about the company that got bought on Tuesday and switched off before the ink dried.
On the tenth of November, OpenAI bought a travel startup called Odessia. Same day, the service went dark. If you opened odessia.com you got a short notice: the company has been acquired, the product is no longer available. If you had a trip sitting in there — flights, hotels, the whole thing worked out over text messages — you were told to send an email.
That was the handover. Not a migration, not an export button. An inbox.
Here’s what Odessia was, because the numbers matter more than the announcement.
Founded by Francis Davidson, who co-founded Sonder before this. Raised six million dollars in June, led by Sequoia. Eight people in San Francisco. Opened to the public on the fourteenth of October.
Twenty-seven days. That’s how long the product was actually public.
In that time, and over the summer before it, tens of thousands of people used it. Davidson says they planned more than sixty million dollars of travel a year through it. That figure is the value of the trips, not the company’s revenue — but it’s a real signal. People trusted it with the expensive part of their year.
It ran on OpenAI’s own models. GPT-5.5 and Codex. And it held IATA accreditation, which is a licence from the airline industry’s own body that lets a company issue actual tickets and get access to private hotel rates.
That last line is the one to keep.
Now look at what OpenAI bought. Not revenue. Not a customer base with a subscription attached. Eight people, one founder, one working booking operation, and one licence.
So here’s the pattern, and it’s three moves.
First, let someone else prove the demand on your engine, at their cost, with their money, in public, where you can watch. Odessia did the expensive, unglamorous thing: it found out whether people will plan a whole trip by typing into a chat window. It found out that they will.
Second, buy the entity that already holds the permission. Not the idea — the accreditation, the relationships, the eight people who know which hotel rate is real.
Third, switch the product off. Keep the permission. Ship it back out as a feature.
It helps that the same day, OpenAI launched Businesses in ChatGPT — companies selling inside the chatbot. Once you have that, a travel agent doesn’t need to exist as a company. It needs to exist as a button. And nobody has to send an email to a founder about it.
This isn’t some sinister plot. It’s cheaper. Building those airline and hotel relationships from scratch takes years and people who know the industry. Buying eight of those people on a Tuesday takes a term sheet. The incentive does the work. That’s what makes it repeatable.
Now the part that should bother you if you run a company that lives inside someone else’s product.
Expedia fell as much as five percent on the news. Booking dropped three. Airbnb three and a half. Tripadvisor moved too.
Those are not strangers. Booking and Expedia were among the first travel platforms to put their own apps inside ChatGPT. KAYAK works with them. Delta Vacations has bought ads through a ChatGPT pilot. They handed over inventory, brand trust, and a customer route — and then the channel bought a competitor and their share prices moved on the announcement.
Being a supplier to the platform is not a hedge. It’s a position. And positions get repriced.
What people are missing is bigger than travel, so let me say it plainly.
The scarce thing in the agent economy is not intelligence. Planning a trip is now a solved party trick. Any model can draft you three itineraries before you finish your coffee.
The scarce thing is permission.
A licence granted by a body of humans. A rate a hotel gives one operator and not the next. An airline willing to let you issue a ticket on its metal with its name on it. Those things are not weights and they are not prompts. They are granted. By people. To an entity they can name, audit, and sue.
That’s why the price of an AI company now includes the boring paperwork. The moat moved off the demo and onto the licence. The demo takes a weekend. The licence takes a year and somebody’s signature.
And it explains the week. Airlines and retailers spending this same week blocking personal agents at the door, arguing over who is allowed to transact. Meanwhile the largest lab quietly bought its way past the door instead. Both moves are about the same asset: the right to act, held by someone who can be held responsible.
So what do you do about it. Four things, in order.
Find the licence in your category. Ask the ugly question: what part of this business cannot be copied by a weekend and a good prompt? If the answer is nothing, you don’t have a business. You have a demo with a bank account.
Hold the counterparty relationship in your own name. Your rates, your accreditation, your contract, your account with the supplier. The moment it lives inside someone’s reseller agreement, you are renting your own moat.
Keep the customer’s record portable, and keep it theirs. Odessia’s users now have an email address. Whatever you build, the person who paid you should be able to walk out with what they made.
And assume your engine supplier is a future competitor, because the demand signal you generate is a roadmap you’re handing over for free. Tens of thousands of people taught someone exactly where the money was. The lesson stayed with the company that owns the chat window.
One more thing, and this is for anyone building toward an exit.
Ask what your buyer actually wants. The thing, or the key. Because those are two very different endings, and only one of them leaves your product alive. If the buyer wants the key, your exit is a shutdown with a press release. That’s not a failure of the deal. It’s the deal working exactly as designed.
The principle worth keeping: authority is granted by a person. Never generated by a model. Which is why, in a market that can generate anything, the humans who can still say yes — and mean it, and answer for it — are the ones holding the only thing that stays scarce.
Go find your licence. That’s the work this week.