On May 27th, someone booked a vacation rental without a website, without an OTA, and without a human travel agent. They had a conversation with an AI. The AI found the property, pulled a live price, collected the guest’s details, and pushed the reservation directly into the property manager’s PMS. The whole thing took under a minute.
It was a test booking, not a live guest, but the infrastructure behind it is real and already working, and it points to something operators should be acting on now. The person who built it is Boris Pavlov, CEO and founder of OnSeason, and he joined a recent episode of The Check-In – with Sarah Nan DuPre and I – to walk through exactly what happened.
His central point: the most important thing a property manager can do to prepare has nothing to do with their website.
What actually happened on May 27th
For an AI agent to book a property, three things have to exist and connect. The property’s live data has to sit somewhere accessible. That data has to be pulled into a common format so the AI can compare one property to another fairly. And there has to be a way for the AI itself to read that data, which today means wrapping it in MCP, or model context protocol.
OnSeason built that stack, connecting property management systems to a database the AI can query, then exposing it through MCP. Once that infrastructure was in place, a booking through ChatGPT became possible.
“By the time I checked if it was in the PMS, it was already in the PMS,” Boris told us. “I don’t know how long it took, but it was seconds.”
Right now the booking shows up in the PMS as a direct booking, carrying the guest’s name, email, phone number, and party size, collected during the conversation with the AI. That’s partly a product decision and partly a reflection of where the commercial side of this still is. ChatGPT and Claude don’t have payment infrastructure fully built out yet for this kind of transaction, and no platform has published pricing for what it will charge to facilitate a booking. Once that’s resolved, Boris expects the channel attribution to change, but for now, every AI-agent booking is effectively direct.
Want to hear the full conversation?
This article is based on an episode of The Check-In podcast, where Leo Walton and Sarah Nan DuPre talk with the people shaping the short-term rental industry.
Listen on YouTube, Spotify, or Apple Podcasts.
Why your website isn’t the first thing to fix
The instinct in this industry has been to prioritise a direct booking website, even a basic one, as the fastest way to build a brand and reduce platform dependency. Boris doesn’t disagree with that as a long-term goal, but the conversation kept circling back to a different starting point: get the back office of your listing in order first, because that’s what an AI agent is actually going to read.
“You don’t need a website,” Boris said. “You don’t need anything except for a place that stores your data and transmits it… via the API. Every single operator needs to be working with a PMS that has a very robust and complete API.”
The reasoning is simple. An AI agent can only book what it can read, and the source of information for current agentic bookings is the PMS API. OnSeason pulls property data directly from the PMS, not from a listing page, which means the quality of that underlying data matters more than how the property is presented anywhere else.
That’s a real shift in priority. For years the advice has been website first, then everything else. Boris is arguing the order now runs the other way: get your data infrastructure right, because that’s what any future booking channel, agentic or otherwise, will depend on. He’s similarly cautious about operators rushing to build their own ChatGPT apps for their listings, technically possible today, in his view. Someone still has to find and install that app the same way someone has to find your website, and outside a small pool of early adopters, most travellers won’t know what a ChatGPT app is yet. The website and the SEO that drives traffic to it still do work an AI app can’t, so neither should be deprioritised. The data work underneath is what prepares you for the channel that comes after.
What the AI is actually reading
Once the data is flowing, the next question is what determines whether a property gets recommended at all. Boris pointed to two things doing most of the work: the structured amenities information and the written description, both of which carry more weight than reviews in his experience.
That’s a change from how most operators think about their listings. Reviews matter, but Boris noted they can be noisy or reference something irrelevant to what a given guest is actually looking for. Descriptions and amenity data are more stable, and they’re also the fields most hosts haven’t touched in years.
His advice on descriptions specifically: say who the property is actually for. “It doesn’t make sense if you just write all types of guests are the perfect guests for the property,” he said. “It’s not going to get picked up.”
Reviews still play a role, and an interesting one. Because OnSeason can read recent reviews, not just the aggregate score, it can factor in something a static listing never could. Boris gave the example of a review posted the day before mentioning nearby construction noise. A property with a strong rating overall could still get filtered out of a recommendation for a guest looking to relax, because the AI weighed that one recent, specific detail more heavily than the star rating.
The PMS gap that’s about to matter a lot more
OnSeason is connected to more than 40 property management systems, but Boris said only 11 support pulling the full range of data needed, things like photos and detailed descriptions, not just rates and availability. Some PMS providers don’t issue API keys to their own users at all, which means a property manager can’t access their own data programmatically even if they wanted to.
Every PMS can push rates and availability. Far fewer can push everything else that determines whether an AI actually recommends a property over a comparable one down the street. If you’re evaluating a PMS today, or wondering whether your current one is good enough, the completeness of its API is the question worth asking, not just whether it has one.
Boris expects some consolidation in the PMS market, though he’s not sure it will happen as fast as the underlying pressure suggests it should. “There’s too many PMSs out there,” he said, “and it’s become easier to develop new features and to even launch your own PMS now with AI coding tools… I hope we don’t end up with a lot of zombie software stacks that have managed to recruit a couple of operators and then struggle to scale.”
He also flagged a longer-term risk to PMS providers themselves: AI-native tools built for individual operational tasks, sitting on top of the PMS today, that could eventually replace parts of it. Not because operators build their own competing software, but because a dedicated AI agent for guest messaging, or pricing, or reporting, ends up doing that one job better than the general-purpose system it currently plugs into.
The timeline isn’t the point
None of the urgency around getting your data right means OTAs are going anywhere soon. Boris doesn’t expect agentic bookings to become the dominant way people book travel for around five years, comparing it to how travel agents didn’t disappear the moment the internet arrived. But he was equally clear the shift is accelerating with every new wave of technology, which is the reason to fix your data now rather than wait for the volume to show up.
Check whether your PMS API actually pushes the full picture, photos and descriptions included, not just rates and availability, and confirm you have access to your own key. Then rewrite your listing descriptions to say clearly who the property is for, and keep the amenity data behind it current. It’s unglamorous work next to a new website, which is exactly why most operators haven’t done it yet.
Boris came back to that same framing near the end of our conversation, and made the stakes explicit. “Having your back office in order, like you said, is becoming extremely important, because once agentic bookings are enabled at scale… that’s going to determine whether you get offered to travellers at all.”
The booking on May 27th took seconds because that back office was already in order. The next one will reward whoever did the same.
Protection that doesn’t care how the booking arrived
As new booking channels emerge, Truvi’s guest screening and damage protection still run automatically on every reservation, whatever channel it came through.
