On 5 November, Noa Eshed shared this after Amit Shamni’s WTM appearance:

Noa Eshed

Hear the episode: Bridgify’s HBX acquisition: Amit Shamni on AI and experiences.

AI agents are coming to buy. Are you ready to sell?

Travel businesses have spent years trying to get people onto their websites. What happens when the person sends an agent instead?

Olivia Moore and Justine Moore at a16z have a new piece on the B2A economy: business-to-agent. Their hypothesis is that AI’s bigger commercial shift is still ahead of us. Agents won’t just answer questions. They’ll choose products, use services and spend money for the person who sent them.

Being found is one job. Being chosen, booked and paid is a different one.

Their argument has four parts:

• Consumer agents become a new route to customers.

• New businesses emerge because agents can do work that previously needed a person.

• Existing businesses win by making themselves easier for agents to evaluate and use.

• Payments, identity, account access and communication need to catch up.

The piece isn’t a travel booking study, but its market map gives travel its own category. It names Soar, Arden, Veeza AI, Odessia and Lola among the examples. That’s the authors’ market snapshot, not proof that each product can complete a booking today. Odessia’s news this week makes the point rather neatly.

For mine, the useful questions are less about which logo made the map:

• Can an agent understand what you’re selling? Room types, inclusions, availability and the actual total price need to be clear, not buried in a lovely PDF.

• Can it finish the booking? A recommendation isn’t a sale if it gets stuck at login, payment or a form it can’t complete.

• Can it handle the trip changing? Cancellation terms, amendments and who helps when something goes wrong are part of the product too.

• Who owns the customer? The supplier may win the booking without winning the relationship.

GEO gets you into the conversation. B2A asks whether you can do the business once you’re there.

We might need to spend a bit less time admiring the homepage and a bit more time testing whether someone else’s agent can actually buy from us.

Travel businesses should also make sure they are offering the best offer they can. Otherwise, someone else might end up with the booking you did the hard work on.

PersonalOffer shows why that matters.

Want the right AI solution for your business? Contact me for a personal recommendation and an introduction to whoever might have the best one.

OpenAI bought the travel agent. Now what?

Odessia has found its home a lot faster than Odysseus did.

Francis Davidson announced that OpenAI has acquired Odessia, the AI travel agent he founded less than a year ago after Sonder’s collapse, and that he is joining the team to focus on the future of travel. OpenAI confirmed Davidson’s hire to TechCrunch. Terms weren’t disclosed.

His numbers: over $60M/year of travel planned through Odessia across the past summer, from tens of thousands of early adopters. Planned travel, not revenue or completed bookings. A Sky VC test put Odessia at the payment step in about 6 minutes 25 seconds, faster than the generalists in that particular flight test. It then handed the traveller to an external website to pay.

The road here, date by date:

• March 2023: Expedia and Kayak become the first travel plugins for ChatGPT.

• January 2025: Operator arrives, OpenAI’s first computer-using agent.

• July 2025: ChatGPT agent launches, with travel itineraries among its uses.

• October 2025: Apps come to ChatGPT, with Booking.com and Expedia among the pilot partners.

• March 2026: Reuters reports OpenAI is scaling back direct checkouts, focusing instead on checkout within third-party apps.

• June 2026: Odessia launches, backed by Sequoia.

• November 2026: Davidson announces the OpenAI deal.

Eugenia Dunaeva added some useful context in a comment: “Francis was in OpenAI’s Grove program from day 1, and Odessia was literally built inside OpenAI’s office, on GPT 5.5.” That’s Eugenia’s account of how closely the team already knew OpenAI, not an OpenAI announcement about the deal.

And what has Sam Altman actually said about travel? These are all his words from the November 2025 Conversations with Tyler interview😀 

• “You ask ChatGPT for the best hotel, not Google or something else. If ChatGPT were accepting payment to put a worse hotel above a better hotel, that’s probably catastrophic for your relationship with ChatGPT.”

• On a neutral transaction fee: “We’ll do that for travel at some point.”

• “Margins are going to go dramatically down on most goods and services, including things like hotel bookings. I’m happy about that.”

• “I think the way to monetize the world’s smartest model is certainly not hotel booking.”

For mine, buying a specialist and bringing in its founder says more about OpenAI’s interest in travel than another trip-planning demo would. It doesn’t tell us what the product will be.

The important catch: Odessia is already offline. OpenAI hasn’t announced a new travel product or a launch date. The deal is real. The next booking experience is still to be seen.

Claim your business in ChatGPT. Yes, really.

OpenAI has launched Businesses in ChatGPT in beta. Companies can claim their business, manage information and see how people engage with them. OpenAI puts the audience at 1.2 billion weekly users. Their number, but quite a shop window.

Why bother:

• Accuracy: once verified, you can update key details and add product feeds with pricing and availability. Better than hoping the model has guessed your rates correctly.

• Insight: Business Hub shows aggregated mentions, website clicks and discovery topics. It doesn’t give businesses people’s private conversations.

• Your own agent: eligible businesses can build and test a Business Agent. The conversation is clearly labelled and separate from ChatGPT’s independent answers.

• Advertising: businesses can sponsor their Business Agent as a destination for ChatGPT Ads. Paid placements are separate from the answers too.

How to claim:

1. Go to Business Hub through OpenAI’s announcement and claim your organisation.

2. Complete verification, update the details and add relevant product feeds.

3. Shopify merchants can connect through Business Hub or Shopify admin. Enterprise customers and advertisers can work through their OpenAI rep.

4. Check whether you’re eligible to build a Business Agent.

The distinction for travel: claiming is available to eligible businesses where ChatGPT operates. Business Agents are starting with eligible US ecommerce and online service businesses. Those are different gates. An Australian tour operator shouldn’t read this as an invitation to switch on an agent tomorrow.

But checking how your business is represented, and giving the platform accurate information? That looks like a sensible job for this week.

Is your AI travel business on the map?

We are mapping every company using AI to build for travel: the (in)complete landscape of AI + Travel. It will never be finished. Companies launch, raise, pivot and shut down every week.

Getting listed is how buyers, partners and this newsletter find you:

Step 1: Head to EverythingAIinTravel.com and sign in or create an account.

Step 2: Choose Submit profile (also linked as Submit an AI tool or service) and add your business details and logo.

Step 3: Follow the publishing steps. The listing offer uses code TC100 at checkout.

Step 4: Apply the code and check the total is zero before completing. If it isn’t, don’t pay. Reply to me and we’ll sort it out.

See the current map.

United says your agent can’t have the keys

OpenAI is buying a travel agent. United is restricting what one can do.

United’s own policy says unauthorized third-party agents cannot transact with the airline, directly or through agencies and aggregators. General trip research is still allowed, provided the tool doesn’t access United’s website or app. Even automated searching on those platforms needs permission.

• Giving your agent access doesn’t mean United permits it to book, buy or manage your travel.

• United says it has no authorized third-party agents yet. It wants secure, standardized integrations first.

• Bookings made by agents before 11 November won’t be cancelled, but United recommends checking they’re ticketed correctly.

CBS reports the airline is updating its booking and ticketing rules. That’s different from an all-AI ban. Accessibility tools aren’t affected.

For mine, this is the other half of B2A. Making an agent capable of buying is one job. Getting the supplier’s permission is another. Watch who closes the door and who signs an integration deal.

A nice conversation isn’t a booking

Via.ai connects a branded AI sales agent to your live inventory, booking journey and commercial rules. It can recommend the right trip, handle objections and close online, or hand a qualified opportunity to your team.

Via’s current production figures from one travel partner’s AI channel:

• 5x conversion versus the traditional website.

• 30% higher average basket value.

• More than 20,000 sales conversations a month.

Those are Via’s reported results, not a promise about your next deployment. But they are the kind of commercial measures travel companies should be asking for. A chatbot conversation and a sale are different things.

Via.ai’s Noam Shapira is an upcoming guest on the Everything AI in Travel podcast. Subscribe to find out the whole story.

Contact Ori Gal to learn more. #sponsored

The AI is cheap. The trip isn’t.

Who pays for the free travel agent? Eugenia Dunaeva asked that in PhocusWire. Nicholas Lombardi at Just Travel has published his answer with real measurements, which is a lot more useful than another founder saying the costs will disappear.

The headline: about 15 US cents in model calls per saved trip across 12 benchmark trips. Including the benchmark trips that went nowhere, 16 cents. Just Travel prices those calls at list rates even though startup credits cover its model bill today.

The rest of the bill is where it gets interesting:

• At scale, its estimated total is 61 cents to $1.62 per saved trip, including maps, places, servers and travel data.

• The cheaper end assumes 60% fewer paid map and place calls. That’s an assumption, not a measured saving.

• Its affiliate income is modelled. No affiliate program was live in the write-up. The $5 subscription is planned too.

• At a researched $12 commission, one booking in 79 trips covers today’s model cost. It does not cover the whole business.

• Counting every cost, the same $12 case needs roughly one booking in 7 to 20 saved trips, depending on scale and map-call savings.

Eugenia pushed back on the token-only comparison in the comments: maps and places are the bigger bill, and some of the business case depends on assumed savings and revenue still to be switched on.

For mine, that makes the piece better, not worse. The author has separated what was measured from what was assumed and made the spreadsheet available for scrutiny.

The useful question for your AI supplier isn’t “what does a token cost?” It’s what a completed trip costs after abandoned searches, changes, data calls and servicing. Then ask which part of the revenue is actually arriving today.

Almost no agent bookings. Still building an agent.

Airbnb and Expedia’s early-November earnings updates delivered a useful reality check. Both describe limited traffic or bookings from general AI agents. Both are positioning themselves for a much bigger role.

Airbnb plans to start rolling out its travel agent in the first half of 2027. Brian Chesky described:

• A generative interface that assembles maps, photos, comparisons, calendars and itineraries around your request.

• An ambition to cover every travel category, eventually booking through Airbnb’s app.

• Moving today’s partner link-offs into native bookings, with a bigger revenue cut and responsibility for customer service. No date for that change.

• Sponsored listings after the agent launches and settles.

His bar for a specialist beating a general agent: it must be “2, 3, 5X better”, not 20% better.

Expedia’s Ariane Gorin says agentic AI’s strength is discovery. Travellers still want a trusted brand when something goes wrong. Expedia is integrating with Google AI Mode and Meta Muse, testing OpenAI ads and supplying Qantas’s own agentic experience anyway.

Inside the business, Expedia reports more tangible results: its Vrbo voice agent resolved over 15% of inbound contacts in its first two months. Travellers using its activity agent were twice as likely to book an activity within seven days. That’s the company’s reported association, not proof that the agent alone caused the extra bookings.

So the public message is “not much has moved yet” while the product work says “don’t be caught standing still”.

For travel businesses, the fight is over the handoff: who brings the customer, who takes the booking, and who answers the phone when it goes wrong. A very familiar fight, with a new front door.

Airbnb’s less glamorous AI story might save more money

Away from the travel-agent announcements, Airbnb’s engineers have published a production benchmark with a rather good headline: frontier quality at about 1% of the cost.

Not a universal claim about every AI job. This was one classifier checking content against Airbnb’s pricing and payment transparency policy.

• Fine-tuning a small Gemma 4 E2B model lifted its F1 score from 0.563 to 0.902, alongside the best proprietary models in the test.

• Median response time stayed around 35 milliseconds. The product needs a result inside 200ms.

• At thousands of requests per minute, self-hosting cost roughly a hundredth of frontier API calls on this task.

• Below roughly 15-30 requests a minute, the frontier APIs were cheaper than a dedicated GPU.

The ingredients were high volume, a stable task and hundreds of thousands of labelled examples. Most small travel businesses don’t have that combination, so this isn’t a call to buy a GPU.

It is a reminder to stop asking which model is smartest and start asking which one does your particular job well enough, fast enough, at the right cost. Sometimes the boring classifier is the better business story.

Get found where travel search is going next

TypeHero scans your travel brand across Google, ChatGPT, Claude and Perplexity, then shows where visibility is leaking and what content to fix first.

Start with a free visibility scan. #sponsored

Instinct Diaries

I’ve got a personal agent from Instinct running my life these days.

What my AI agent actually did last week, 2-8 November:

• Sent Friday’s Everything AI in Travel newsletter and saved the new weekly template.

• Built more of the PersonalOffer infrastructure: 440 tests passed, with 15 skipped.

• Got the Michael Ryan episode onto Spotify for Saturday morning.

• Produced three York Lu clips and the approved Everything AI in Travel outro.

The miss: email checks stayed paused too long. That was caught and corrected on Sunday. Getting the job done includes noticing when the routine has stopped running.

Pass this on

If someone you work with would benefit from keeping up with AI + Travel, forward this email. They can subscribe here.

Got a tip or seen a story I’ve missed? Reply to this newsletter.

Slack Group!

The Slack group is full of the brightest minds in AI in travel. Join the Slack group here. I found my co-founder Adrian in this group of over 220 of the top voices in AI + Travel.

Talk in the Slack group was about disgruntled investors who didn’t get in on the Odessia deal (but no-one knows what the deal was - so maybe don’t be too upset yet 😅)

Most clicked last week

A dead heat: Reely.travel and the AI + Travel landscape map. Personalisation and finding out who’s building what. Fair enough.

Podcasts and Sponsors

The latest released episode is my conversation with Michael Ryan of Ink Digital: why “SEO is dead” keeps being wrong, what still works in AI search, and how to connect your marketing plan to the destinations that actually make you money. Listen to the episode here.

Michael’s view is that much of AI search visibility is still the same work good SEOs have done for years. The genuinely new part doesn’t excuse skipping the basics. One useful question for your team: where are competitors being cited while your brand is missing?

Partner with Us for newsletter and podcast sponsorship.

That’s it. If you’re thankful for this newsletter, you can always buy me a coffee.

I’ll put the most clicked story in next week’s edition. If I’ve missed something, you’ve got a tip or any feedback, reply and it comes straight to me. I’m doing this for you, so don’t be shy.

Glossary

This section is now monitored by AI to help keep it current.

AI (Artificial Intelligence): software that performs tasks such as recognising patterns, understanding language and making predictions or decisions.

KYA (Know Your Agent): checking an agent’s identity, who it works for and what it’s allowed to do. The AI-era cousin of KYC, Know Your Customer.

Generative AI: AI that produces new text, images, audio, video or code. LLM (Large Language Model): a model trained on large amounts of text to understand and generate language. Token: a chunk of text or other input that a model processes; often part of how model use is priced.

Agentic AI: AI that takes actions, such as searching, comparing and booking, within rules you set. An assistant can answer a question without being authorised or able to complete the job.

B2A (Business to Agent): designing and selling for the agent acting for your customer, not only the person. York Lu discussed the travel version on the podcast; a16z’s piece above takes a wider business view.

MCP (Model Context Protocol): a standard for connecting AI applications to tools and data. API (Application Programming Interface): a defined way for software systems to exchange requests and data. MCP uses APIs; they aren’t interchangeable names for the same thing.

UCP (Universal Commerce Protocol): an open standard for commerce interactions across platforms. AP2 (Agent Payments Protocol): a protocol for agent-led payments. ACP (Agentic Commerce Protocol): OpenAI and Stripe’s commerce protocol. PAP (Personal Agent Protocol): Sierra’s proposed protocol for personal agents working across services.

GEO (Generative Engine Optimisation): making your business understandable and visible in AI-generated answers. SEO (Search Engine Optimisation): improving visibility in search engines. UX (User Experience): what using a product is like for a person. Generative UI: an interface assembled for a particular task rather than a fixed page.

NDC (New Distribution Capability): the airline industry’s standard for richer retail offers. OTA (Online Travel Agency): Booking, Expedia, Trip.com and friends. TMC (Travel Management Company): corporate travel management, such as FCM, CWT and Amex GBT. DMO (Destination Marketing Organisation): the body that markets a destination.

Headless: separating the interface from the systems supplying content and services. An agent can become a new interface to those systems. Fine-tuning: further training a model on examples for a particular task. Classifier: a model that assigns categories or scores rather than writing a long answer. F1: a score balancing precision and recall; useful for the Airbnb comparison above, not a universal measure of intelligence.

The models and who owns them

Model names checked 12 November 2026. These are families and examples, not a ranking.

OpenAI: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. Model guide.

Google: Gemini 4 Argon; Gemma 4 is its open-model family used in the Airbnb story. Gemini 4.

Anthropic: Claude Opus, Sonnet and Haiku 5.5; Claude Fable and Mythos 5.2 were announced this week. Model announcements.

SpaceXSI: Grok 4.7. Model documentation.

Meta: Llama 4 Scout and Maverick, its open-weight Llama models. Llama 4.

DeepSeek: V4.1 Flash, including the October 17 update. Release notes.

Mistral AI: Mistral Large 4, also called Le Chonk. Model release.

Perplexity: Sonar, its grounded search model. Model documentation.

The assistants: they answer

ChatGPT (OpenAI), Copilot (Microsoft), Gemini (Google), Claude (Anthropic), Perplexity and Grok (SpaceXSI). Several can also use tools and take actions; the product name alone doesn’t tell you what permissions you’ve given it.

The personal agents: they do things

DOTS (OpenAI), Muse (Meta) and Instinct, the personal agent I use for research, email, calendar and getting things done. The practical difference is whether it can carry the job through, not whether it sounds helpful in the chat.