OpenAI Bought Odessia: A Travel Agency Run by Eight People and AI Agents

On November tenth, 2026, Francis Davidson, the founder of the startup Odessia, wrote that Odessia was acquired by OpenAI. Odessia was a travel agency where artificial intelligence did the work instead of human agents. The service opened to everyone on October fourteenth. Twenty-seven days passed between the launch and the sale. The company had fewer than ten people.

For OpenAI, the purchase itself is small. The price wasn't even disclosed. What's more interesting is something else. That same day, Davidson published a long piece about how he built the company. Almost every job there, from code to legal questions and finance, was handed to AI agents first. It's one of the most detailed accounts any founder has given of what a company looks like in practice when every person has a team of agents reporting to them.

This episode has four parts. Who Davidson is, and why his previous company collapsed. What Odessia could do. How, in his own words, eight people worked with agents. And what's worth taking for yourself, and what deserves some caution.

One caveat up front. Almost everything we know about how Odessia worked on the inside comes from the founder himself. Nobody has independently verified it. Whenever a number or a claim is his, I'll say so.

Part one. Sonder.

To understand Odessia, you have to start with Davidson's previous company. In 2012, he was a student at McGill University in Montreal, and he was subletting apartments to tourists. In 2014, that grew into a company that was first called Flatbook, and later Sonder.

The idea was simple. Sonder signed long-term leases on entire apartments and whole floors of buildings, furnished them, and rented them to guests like hotel rooms. No front desk, check-in through your phone. To a traveler, it looked like a convenient hotel. For the company, it was a very heavy structure: it paid rent to building owners for years, even when the apartments sat empty.

For a while, everything went up. In 2020, Sonder was valued at one point three billion dollars. In January 2022, the company went public on the Nasdaq by merging with a special purpose acquisition company, a so-called SPAC. The valuation had to be cut along the way, and the next day the trade publication Skift wrote that the debut was disappointing.

Then the problems started. In 2024, Sonder delayed its annual report and restated it because of errors in how it accounted for its leases. According to its filings, revenue in 2024 reached six hundred twenty-one million dollars, and the net loss was two hundred twenty-four million dollars. Rent alone cost about three hundred million dollars a year. By the end of 2024, the accumulated loss was approaching one and a half billion dollars.

A deal with Marriott was supposed to be the rescue. In August 2024, the largest hotel chain in the world signed a long-term agreement with Sonder, and about ten thousand five hundred Sonder rooms appeared in Marriott's system. But combining the technology of the two companies turned out to be very hard. Davidson later called that integration the hardest thing he had ever done in his life. In June 2025, he stepped down as CEO.

In November 2025, it was all over. Marriott terminated the agreement, citing Sonder's failure to meet its obligations. On November tenth, Sonder announced it was shutting down immediately and filed for bankruptcy with liquidation. According to press reports, guests were thrown out in the middle of their trips. Experts quoted by Commercial Observer say it was the long-term leases that killed the company. One of them said Sonder was on the verge of collapse even before the Marriott deal.

Davidson himself sums it up like this: he took Sonder from zero to six hundred million dollars in revenues, which took over a thousand people and a decade.

Remember those numbers: a thousand people, ten years, hundreds of millions of dollars in rent. He would build his next company in exactly the opposite way.

Part two. A fresh start.

After leaving Sonder, Davidson took a break. He told Inc. magazine that during that time he managed to read one novel. And in June 2026, he told Skift he was so happy to be moving on to something like just a blank new slate.

According to Skift, the idea for the new company was born partly in conversations with the venture firm Sequoia, where a close friend of his is a partner. One more detail that will matter at the end of the story: Davidson took part in OpenAI Grove. That's OpenAI's own program for technical people who are just starting to build a company. Applications opened in January 2026. According to the Solo Founders newsletter, that's also when work on Odessia's code began.

On June second, 2026, Odessia came out of stealth. On June tenth, Davidson announced: we raised six million dollars led by Sequoia. The same post had a video of him planning a trip to Mexico City in three minutes. Flights, hotels and a full itinerary that matches his preferences, all bookable on the spot. Free to use.

According to Skift, the name Odessia was meant to evoke a Greek goddess of travel. In one comment, Davidson joked: apparently, she's the goddess of travel. Inside the company, the agent was referred to as "she."

Notice the contrast with Sonder. Odessia had no long-term leases, no apartments, almost no offices and almost no employees. It had software, contracts with suppliers, and a small team.

Part three. What Odessia could do.

From the user's point of view, it was simple. You text Odessia a regular message, like you would a friend: where you want to go, when, what you like. The agent finds a flight, a hotel and experiences, puts together an itinerary, and books everything in that same conversation. You could export the finished trip to your calendar and send it to the people traveling with you. For travelers, the service was free. The company made money the same way a regular travel agency does, from commissions on bookings.

Here's an important detail that set Odessia apart from chatbots. The company had IATA accreditation, from the International Air Transport Association. That accreditation lets an agency issue airline tickets itself. On top of that, according to Skift, it gave access to private hotel rates and perks like free breakfast and room upgrades. Put simply, Odessia didn't send you off with a link to someone else's website. It was a travel agency itself.

Behind hotel search and booking was a supplier called Nuitée. It's a company that sells other businesses access to hotel rates and booking. According to Skift, Odessia wrote its own descriptions of hundreds of thousands of hotels and rated them on many attributes, right down to street noise. Odessia itself claimed its recommendations didn't depend on the size of the commission. There's no way to check that from the outside.

The hardest technical part, according to Davidson, wasn't the model. In a short video for LangChain, a company that makes tools for people building agents, he said the main challenge was assembling the right set of tools and the scaffolding around the model. The user sees a simple, pleasant conversation. But behind a single reply, the agent makes dozens of calls to different systems: searching flights, checking prices, looking at room availability. And all of that has to fit into a few seconds, because otherwise, as he puts it, people lose interest.

To track down the agent's mistakes, the team used a service called LangSmith. It collects detailed records of what the agent did in every conversation, the results of quality checks, and the code itself. Davidson says that thanks to this, even team members who hadn't built the agent from scratch could find the cause of a problem and propose a code fix themselves. Keep in mind that this is a promotional video from a vendor, and there isn't a single number in it. It was recorded in June, when, as Davidson himself says, the company didn't have any customers yet.

The founder doesn't say which models Odessia ran on. The publication TNW reports that they were OpenAI models, GPT five point five and Codex.

On October fourteenth, Odessia opened to everyone and called itself an always-on concierge for travel and experiences. Davidson opened his launch post by saying we're months away from AGI, and the leisure economy's about to explode. That's a bold claim, and he didn't back it up.

There's also a success number. According to Davidson, over sixty million dollars a year of travel was planned by Odessia over the past summer by tens of thousands of early adopters. Pay attention to the word "planned." That's not revenue, and it's not even the value of bookings. It's the value of trips people discussed with the agent. The company never disclosed its revenue.

Part four. How eight people worked with agents.

This is the heart of the story. Sources disagree on how many people Odessia had: seven, eight or nine. Davidson himself doesn't give a number. But he described in detail how they worked, and he broke it down into six principles. Let's go through each one.

Principle one: agents everywhere for everything.

Davidson writes that they got started at a time when agents started really working, and they went all-in. They burned the old playbooks, as he puts it. And, he stresses, we're not only talking about software engineering tasks. Strategy, partnerships, design, marketing, customer support, legal, finance, all of it.

The rule at the company was this: before doing pretty much anything manually, they first kicked things off with an agent. Their product launch video, he says, was created with an agent. When they needed to choose a technical partner, agents evaluated potential partner APIs by benchmarking their performance, putting together spreadsheets and slide decks with the results.

Davidson writes that his personal usage of agents skyrocketed twenty-five x through 2026.

Principle two: context maxxing.

This is probably the most important practical lesson. Davidson writes that agents need secure access to all the same information an employee has access to. In practice, that means dozens of plugins, MCPs and CLIs. Odessia's agents had access to every meeting recording, to the team chat app Slack, to email, to shared documents on Google Drive, and to all software tools used across the company. The connections were made through MCP. That's an open standard that lets you plug an AI agent into any outside program, roughly the way you plug a flash drive into a computer.

Here's what that looks like in practice. A single run, Davidson writes, will retrieve meeting transcripts, review a presentation in Drive, look at email exchanges, Slack messages, user feedback, and tap into a few systems of record. The whole, he says, is greater than the sum of the parts.

On top of that, the team created clean shared documentation for every decision. Why? So the agent understood not only the task, but the strategy behind it: why the company does things this way and not another way. And agents maintained this knowledge base and artifact history themselves.

It's worth pausing here. Most people use agents like a smart assistant with no memory: they explain the task from scratch every time. At Odessia, agents knew roughly as much about the company as an employee did. That probably explains why they could handle work far beyond programming. The cost of this approach is obvious too: an agent with access to email, documents and every system in the company calls for very serious attention to security. Davidson writes about secure access, but he doesn't give any details.

Principle three: voice prompts and skills.

Davidson admits openly that agents still need a lot of correction to produce excellent work. But writing prompts is too slow. Voice, he says, usually goes three to four x faster. No need to meticulously handcraft prompts. You can just say what's wrong.

And repeated corrections were turned into skills. In the world of agents, a skill is a written instruction that the agent loads on its own when it's relevant. For example, if you're telling an agent for the tenth time how to format a report, the smart move is to write that rule down once, so you never have to repeat it.

Principle four: iterations and frontier models.

This is where the boldest numbers are. Davidson gives an example: a task that may have taken, say, forty hours to do manually gets done with agents in two hours. That's a twenty-x productivity gain. Against that, he calls the cost of running the agent a rounding error.

His conclusion: the ROI is so good that the rational thing to do is to go unlimited. He says they saw productivity accelerate dramatically when they removed usage limits. The bottleneck isn't the agent. The bottleneck ends up being your capacity to consume agent outputs and drive decisions.

And one more conclusion that goes against the usual advice to save money. Davidson says a contrarian rule of thumb for 2026 has been that the more expensive a frontier model is per token, the higher its ROI. Cheaper models, however you prompt them, often fail at getting you to clear the quality bar, and you end up redoing the work. And every extra round of rework costs more than the difference in price. He also notes separately that, in his experience, closed models dominate the entire Pareto frontier at any level of intelligence. In other words, the closed models from big companies beat open ones at every level of capability.

This part deserves a sober look. Forty hours versus two is an example, not a measurement. There's no comparison with a control group. And the view that expensive closed models pay off best comes from someone whose company was bought that same day by a maker of closed models. That doesn't mean he's wrong. But it's a reason to test it on your own tasks rather than take it on faith.

Principle five: lead a team of agents.

As Davidson describes it, the workday at Odessia went like this: wake up to what the agents did overnight. A task takes several minutes to run, but you don't just stare at it. You can launch the next one. And every day, the team had daily standups to divide and conquer and hold each other accountable.

Gone are the days of quarterly roadmaps, he says. The main rule: the bottleneck should always be you, and never waiting on agents to do work. Everyone is now a manager of agents, he writes. They direct, then review completed work, give feedback, and iterate until the bar is met.

He also singles out agents that start work on their own, without a person launching them. It's also critical, he says, to have proactive agents that jump in to do work without being manually launched. At Odessia, those agents produced daily analytics reports. They investigated every error in production. They reviewed all customer feedback and traces, meaning the records of conversations with the agent. They matched accounts payable, checking supplier invoices against payments. They updated the task tracker Linear after the daily standup. And they ran overnight smoke tests to check that everything still worked.

Principle six: a frenetic pace.

You'd think higher productivity enabled by agents means less effort from us humans, but the opposite is true, Davidson writes. You hear about more and more startups doing six or seven days a week, and Odessia was no exception. An hour is so precious, because so much more can get done than before.

He calls it Jevons' paradox applied to our own work schedules. Jevons' paradox is an old observation from economics: when a resource starts being used more efficiently, total consumption of it doesn't fall, it rises. In the nineteenth century, that happened with coal when steam engines became more efficient. According to Davidson, the same thing happened with his team's working hours.

And that leads to his most quoted line: the best process is no process. He's always loved the concept of "touch it once" for email: you open a message, decide right away, and never come back to it. He started doing this for any product improvement or idea he wanted to explore. If someone has an idea for improving the product, it doesn't go into next quarter's plan. They launch an agent right away. Davidson says he's been able to dream of some ideas only to see them in the real world within hours or days. The ROI on planning and process has cratered, he writes. And in hindsight, it will seem crazy to us that we used to run the economy in such a manual, tedious way.

There's one detail that balances this picture a bit. According to a summary in the Solo Founders newsletter, Davidson is skeptical of the idea of a one-person company worth a billion dollars. In his view, artificial intelligence is still weak at naming, design and taste. That's a paraphrase, not a direct quote, but it does a good job of explaining why there was still a team of humans around the agents.

Part five. The deal.

On November tenth, Davidson announced that OpenAI was acquiring Odessia, and that he would be joining the team to focus on the future of travel. The price wasn't disclosed. OpenAI didn't put out an official press release. A company spokesperson told TechCrunch that OpenAI was happy to welcome Francis. TechCrunch calls the deal an acquihire, a purchase made for the team. OpenAI's Armin Hamrah wrote that from the first day of the Grove program, Davidson's passion for travel had energized Hamrah personally, and that Davidson would be leading the effort.

The Odessia service was shut down that same day. The website now has just a notice: the company was acquired, the service is no longer available, and for existing bookings you can contact support.

Why does OpenAI care? The company has been edging toward travel for a long time. In 2023, Expedia and Booking were among the first plugins for ChatGPT. In 2025, OpenAI released an agent that works in a browser on its own, and then built the Booking.com and Expedia apps right into ChatGPT. And in March 2026, according to The Information, OpenAI scaled back direct checkout inside ChatGPT: people searched for trips in the chat, but they preferred to buy on partner websites. Expedia's stock rose more than twelve percent that day.

By buying Odessia, OpenAI got someone who had already built a working, agent-run travel agency, with accreditation, suppliers and experience. That same day, OpenAI launched a business section in ChatGPT where companies can set up their own agents.

The market reacted right away. According to Investing.com, cited by TNW, Expedia's stock fell as much as five percent during the day, Booking as much as three percent, and Airbnb about three and a half percent. Something similar happened in September, when Meta showed off an agent that could book hotels and flights directly: back then, Expedia fell as much as seven percent. Investors are clearly afraid that agents will take customers away from online travel agencies. Meanwhile, the travel agencies themselves say traffic from agents is still small. A week before the deal, Expedia's CEO said agent traffic was still very small.

Part six. What it all means.

First. A small team with agents really can build a product that used to require a large staff. Less than a year, fewer than ten people, a working travel agency with real bookings. That's the most important part of this story. But look at what actually made Odessia a real business. IATA accreditation, a contract with a hotel supplier, Sequoia's money, the founder's reputation, ties to OpenAI. An agent can't write any of that. Code is getting cheap. What stays expensive is access, licenses and trust.

Second. The most useful part of Davidson's account isn't about models. It's about context. The agents worked well because they knew as much about the company as the employees did: meeting recordings, conversations, documents, and above all, written-down decisions with an explanation of why they were made. If you want agents to do more than write code, that's where to start.

Third. The economics flip. In Davidson's account, the scarcest resource isn't money for models. It's the attention of the people who assign the tasks and review the results. That's where the six-day week comes from: agents didn't free up the team's time, they made every hour more valuable. If you're building a team around agents, it's worth thinking about this ahead of time.

Fourth, about the numbers. Twenty-five x, forty hours versus two, sixty million dollars: all of these are claims from the founder himself, with no independent verification. The sixty million dollars is planned trips, not revenue. And the story didn't end with the agency growing. It ended with the agency being bought for its team and shut down on the day of the deal. For users, that's bad news: a service they relied on disappeared in a single day.

And finally. Odessia ran on OpenAI models and was a graduate of an OpenAI program. That's a common story for a startup, and for Davidson it ended well. But if your product is built entirely on a big company's models, you should understand that this company can see your market too, and it may decide to move in itself.

Here's what you can do as soon as tomorrow, with no team and no investors. Write your decisions down in one place your agents can see: what you decided, and why. Give tasks by voice, because it's faster. If you're correcting an agent on the same thing for the second or third time, turn that correction into a permanent instruction. And try handing an agent one recurring job that it does on its own, without you launching it: a morning report, or an overnight check that everything works.

Let me sum it up in three sentences.

Agents do the work, but access, licenses and trust are still earned by people.

The more an agent knows about your company, the more work you can hand it.

Listen carefully to founders' numbers about agents, but test them on your own tasks.

That is all for today. My name is Evgenii Arsentev, and I bring artificial intelligence into companies: automating routine work and cutting operating costs. Thanks for listening.