Booking's AI wants the trip, even when Expedia got the booking! 😅

Booking Holdings has a new way to stay close to travellers who booked somewhere else. Vayo pulls travel confirmations from your inbox and keeps working on the trip. That is a different proposition from getting you back to a booking search box.

The company announced a planned US$100 million investment on October 14. I note the footnote: that investment depends on continued product progress and market conditions. This is a free, invite-only private beta with a waitlist, not $100 million already spent or a product everyone can use today.

Todd Bishop at GeekWire has a useful look at it. Vayo brings reservations together, watches for price drops and availability, and flags how a delayed flight could affect your hotel check-in or dinner. Importantly, those reservations can come from other platforms or directly from suppliers.

Booking does not have to win the original transaction to get involved.....

For mine, that is the commercial bit worth watching. A direct hotel booking could still end up being managed inside a Booking Holdings product. The supplier keeps that sale (unless the OTA decides to discount it - and then offer the swap... 🙃... but surely they wouldn’t do that, right!! 🤡), but Vayo has an opportunity to become the place the traveller returns to for the next decision. Direct booking and owning the ongoing relationship are not automatically the same thing.

The booking capability needs a little unpacking. GeekWire and PhocusWire report restaurant reservations are available through OpenTable. Flight and hotel booking inside Vayo is still to come; PhocusWire says those searches currently hand off to suppliers. Useful progress, but not yet the whole trip booked and managed autonomously.

There is also a fairly obvious trust question. Vayo promises to work beyond its parent’s brands, while its eventual business model is still unsettled. GeekWire reports subscriptions and booking commissions are possibilities. Being open to outside inventory is a start. How recommendations get ranked, and who pays when a booking happens, will matter too.

For hotels and other suppliers, this is another reason to separate winning a booking from keeping the customer relationship. Booking is building a way to participate in trips it did not originally sell.

Building the map of B2B AI in travel: is your company on it?

I'm putting together a map of the AI companies selling to the travel industry. Who's building what, and who they're built for.

If you build or sell B2B AI for travel, I'd like to review your company for it. It takes a few minutes to fill in, and you tell me what you do, who you sell to and where.

Submissions are reviewed. Filling in the form doesn't guarantee a spot on the map.

There's an optional section at the bottom if you'd also like to be considered for introductions when travel businesses ask me who can help. Skip it if it's not for you.

Submit your company

PROS wants AI agents closer to the money

A chatbot helping you find a flight is one thing. AI helping the airline decide what to sell you, and at what price, is a rather different conversation.

PROS announced Kairos this week, an airline retailing platform that brings its pricing science, airline data and AI into one workspace. The starting point is fairly practical: analysts can use natural language to get at information and tools, rather than bounce between applications. Databricks provides the data and AI foundation, with MCP connections intended to let specialist agents work together.

MCP is basically the wiring that lets an AI agent talk to other tools. Useful, but wiring alone doesn't tell you whether the thing makes good commercial decisions.

The more interesting bit for mine is what PROS says comes next. It plans to extend agents into offer optimisation, dynamic pricing, shopping, group sales and offer marketing. That puts AI much closer to the airline's revenue engine than another customer-service chatbot.

To be super clear, "plans to" is doing some work here. The announcement doesn't show those future agents delivering a revenue uplift, or give us airline deployment results to judge. This is a launch and a direction of travel, not proof that AI is now running the commercial department. Maybe hold off on retiring the revenue manager just yet.

The data bit is worth watching too. PROS says Kairos and Databricks can connect market, behavioural and operational data without requiring airlines to migrate everything to one platform. My read is that this is the harder, less photogenic part of agentic AI: giving the agent enough context to do something useful, and controlling what it's allowed to do with it.

If this works, the advantage isn't a nicer chat box. It's getting from a change in demand to a better offer with fewer manual steps. For airline commercial teams, I'd be asking which decisions the agent can actually make, which still need a human, and how you'll measure the effect on revenue and margin.

That is a more useful conversation than counting how many AI agents you've launched.

I'll be at the FACTS conference in Sydney at the end of November, talking about AI's impact on NDC.

Three and a half years of TripGenie: the lesson is not the model

Amy Wei, who led TripGenie at Trip.com, has written up what she learned and it's worth reading.

Trip.com was the first OTA to put an LLM assistant in its app, and nine people shipped the first version in nine days from February 2023.

The number I'd look at is the shift in what people asked.

Booking-related questions went from roughly 30% of interactions to roughly 60% over those years.

And a lot of what looked like booking questions were really requests for reassurance. Her team found more than half of travellers were using the AI to check what they were entitled to before booking a flight: lounge access, baggage rules, what the fare included. The information was public, just scattered. The AI's job was to put it in one place. (Read that sentence again)!

Two other points stood out:

• Conversation is not the universal interface. Their hotel comparison view cut the clicks to a decision by about 80%, and over half of those users picked the hotel the AI recommended. Sometimes a table beats another paragraph.

• Once the product acts for the traveller, she says "the AI got it wrong" is not an excuse. Her team moved from scoring single answers to scoring whole multi-turn conversations, and sorted failures into missing context, poor hand-offs and unclear next steps so each had an owner.

This is a rare look at what travellers do with an assistant after the novelty wears off. They are mostly trying to feel safe before they commit.

Your bank would like to get involved before you book

The bank used to turn up when you paid for the holiday. Now it would quite like to help you choose it.

Emirates NBD launched an AI Travel Assistant with Visa and Tripadvisor on October 5. It builds itineraries from Tripadvisor reviews and listings and surfaces Visa cardholder offers along the way. It is free and does not need a bank account or login.

The commercial move is the bit worth watching. The bank and card network are getting into the consideration stage instead of waiting at payment. There are no conversion results in the announcement, and planning an itinerary is not an agent completing a purchase.

Travel inside banking is a hot area right now. Unravel, Swifty, Via.ai, Odynn and others are all running to a similar door, with slightly different solutions. Unravel builds travel into banks' own apps. Swifty is an AI travel agent that Revolut has acquired. Via.ai offers a branded AI agent that financial institutions can use as a travel concierge. Odynn provides travel and rewards infrastructure for banks and fintechs.

Booked is not the same as allowed to board

An AI finding you a cheap flight is useful. An AI checking whether you can actually enter the country might be the more important job.

Intellivisa has raised $3 million in pre-seed funding for an AI travel-authorisation platform, according to Michael B. Baker at Business Travel News. United Airlines Ventures is among the investors listed in the report.

The company wants to sit in the post-booking flow of airlines, TMCs and OTAs. It says its agents can work out which visa or entry authorisation a traveller needs, complete forms and submit applications, with payment, filing and refunds handled as part of the process.

That is a very different proposition from asking a chatbot "do I need a visa?" and hoping it has not missed the bit that applies to your passport.

The commercial opportunity is in the gap after the sale. The traveller has a confirmation number, but the trip still has conditions attached. Nationality, route, purpose and documentation can matter. A booking system taking the money does not make those questions disappear.

For mine, this is where travel AI gets more useful and less forgiving at the same time. Removing paperwork is good. Filling the wrong form or telling someone they are cleared when they are not is a rather expensive way to save ten minutes.

The report describes company claims and a funding round, not independent evidence that the system gets every case right. It does not give processing success rates, error rates or the detail of how a traveller checks what is submitted. And an AI submitting an application is not the authority deciding whether someone can enter.

So I would want the fairly boring questions answered before putting it in front of customers. Which source establishes the current requirement? What does the traveller approve? What happens when the case does not fit neatly, or an application is rejected?

The wider point is worth keeping. AI travel does not have to be another recommendation engine. There is plenty of work around the trip people have already chosen. Help them complete that work properly, and you have something much more useful than a prettier itinerary.

The eye of the agentic storm: three October moves put agents at your front door

When your customer's agent turns up at your website, your contact centre or your checkout, will you know who sent it, what it is allowed to do, and whose interests it is serving? That is the question Gam Dias puts to every consumer business in his new piece, and his point is that the current calm is the warning, not the all-clear.

October brought three developments. On 6 October, Meta, Sierra and partners including Walmart and Stripe announced the Personal Agent Protocol, a draft standard for how a personal agent identifies itself to a business. Two days later Revolut's chief executive said its in-app assistant AIR will start shopping on Amazon for UK customers, with testing from December. And Meta's Muse, the agent people use to enrol kids in after-school clubs and chase appointments they have put off for months, already reports millions of users in the US.

Travel is already in this story, mostly as the industry blocking the door. Amazon blocked Muse in September, citing scraping, and similar blocks have been reported at Delta, United, Zillow, Yelp and eBay. Meta's reply: turning away a personal agent means turning away the customer behind it. Dias's sharper warning is for policies priced on human patience. An agent will happily book, rebook and cancel inside a free-cancellation window until it closes, and the same logic hits returns windows, price-match promises and loyalty points.

The demand numbers say the handover is gradual, not overnight. A Cover Genius and Gather survey of 1,392 consumers across seven countries found 78% open to AI helping them shop, but only 5% comfortable letting an agent buy on its own. People are handing over the errands first and keeping their hands on the money.

His advice is to use the quiet spell. Write down which actions a customer's agent may take, and what read and write mean for each of them, before agents start testing the gaps. An agent that can ask the same question five different ways will find every rule that was never settled.

Poppy: five steps for dealing with someone else's AI agent

A new open standard called Poppy (Personal Agent Protocol) is aimed at one question: when an AI agent turns up at your company on a customer's behalf, how do you know who it is and what it may do? Sierra announced it with Meta, Shopify, Stripe, Walmart and others. Here is the LinkedIn post with the walk-through.

The protocol runs in five steps.

1. Discovery. The company publishes a single file that tells agents how to deal with it: how customers sign in and what an agent is allowed to do.

2. Sessions. An agent says it is an agent and can begin as a guest, enough to check stock or a returns policy. The company knows what it is dealing with from the first request.

3. Sign-in. The customer signs in on the company's own page. The agent receives permission limited to what that customer approved, and nothing more.

4. One session. That same permission follows the agent across the website, the company's systems and its own agent, so before and after sign-in count as one visit.

5. The company picks the route. The agent might use the website with the permissions applied, connect through a technical interface, or be passed to the company's own agent.

It matters to me because I have automated a lot of the post-recording work for my podcast, and I have written a playbook on exactly how it is done. So I know what it is like to hand an agent real work in other people's software.

The Descript example. I have not confirmed that Descript supports or plans to support Poppy, so everything below is hypothetical.

Today my agent, Bax, works in Descript's screens much as I would. It builds the episode project in five scenes, removes filler words from the episode scene only, names the project, pulls the transcript and tells me when the edit is ready to check.

Enhanced. Descript could know an agent is working and what job it was given, rather than just seeing someone use the screens. The agent might also get a more dependable way in than working through the screens.

Made safer. If Descript let me approve a narrow job, such as editing a project, then exporting, publishing and billing could sit outside it. That would be a safeguard enforced by the product itself, on top of my own checks.

Inhibited. The same limits could get in the way. If the permitted actions left out the export step I want, I would have to approve it separately or finish it by hand, which would slow down a workflow that runs smoothly today.

Whether any of this arrives in the tools travel teams use is an open question. The same one applies to the companies we sell to. If a customer's agent walks up to your booking flow tomorrow, what would you want it to be allowed to do?

Source: Sierra, Introducing Personal Agent Protocol

Anyone else seen this?!

ChatGPT now creates little quizzes in the response on the fly to help refine the answer it gives.

These are responsive and clickable to get the user to exactly what they need fast.

See it in action here.

If you think someone (or everyone) you know or work with could grow from being more informed on the topic of ai + travel (or could use the training above) then please forward this email to them and they can click the button below:

Slack Group!

Podcasts and Sponsors

Podcasts now on Spotify and Apple Podcasts:

All podcasts are on Spotify and Apple Podcasts. Now is a great time to binge those you may have missed.

The latest released episode is my conversation with Antonio Picozzi, co-founder and CEO of dAirect. We talk about helping hotels get found through AI assistants without handing the customer straight back to an OTA, and why getting recommended is not the same as getting the booking.

Listen to the episode here.

Partner with Us - sponsorship info for the newsletter and podcast can be found here.

Most clicked last week was OpenAI's Radisson case study.

That is it - you have made it to the end of this edition. If you are thankful for this newsletter, you can buy me a coffee.

If I have missed something, you have a tip or any feedback at all, simply reply to this email and it will come straight to me.