Anand Mishra: Welcome, everyone, to the AI in Shopping panel. My name is Anand Mishra. I'm the vice president of technology at ATPCO. ATPCO plays a neutral role in the industry, and that allows us to work with airlines and partners, and it gives us a broad view of what's changing in the industry from multiple angles. At ATPCO, we're investing in AI and data science, both ATPCO and 3Victors together, because we believe the next leap in travel is going to come from turning high-quality data into faster, more trustworthy decisions. The conference theme today is also data in motion, offers in action. So as data becomes easier to utilize and offers become more dynamic and contextual, AI and data science become critical in turning that motion into offers that you can trust. As the data flows through the ecosystem, multiple partners get impacted, and that's why we have invited industry partners with different perspectives to share with us what they're seeing. Before I ask my esteemed panelists to introduce themselves, I'm introducing some common terms that most of us have already seen, just to ground ourselves on what we mean by AI, LLM, agentic AI, and hallucination. I'd like to invite the panelists to introduce themselves. Guillaume, if you'd like to get us going.

Guillaume Dupont: Thanks, Anand. Guillaume Dupont, Amazon Web Services. AWS is the most comprehensive cloud platform. We're very happy to work with a lot of travel and hospitality partners. 21 of the top 30 airlines use AWS, as well as many industry providers like ATPCO, Alike, and Skyscanner. Thank you both for your partnership. I work at AWS setting the technical strategy for our North American airline group. I worked in the industry for a long time, with one exception during the COVID period, where I went to do cross-industry pricing and led a team of data scientists. Outside of that, I started my career at Amadeus working on revenue management, availability, and dynamic pricing. Pleasure to be here.

Anand Mishra: Laurie?

Laurie Garrow: Hi, my name is Laurie Garrow. I am a professor at Georgia Tech here in Atlanta. I'm also honored to run the Air Transportation Lab at GT, which in Atlanta is also known as ATL@GT. The history of that is that I actually inherited the center that was formerly the PODS Consortium, run by Peter Belobaba at MIT. It's really been an honor to take that over and lead a consortium of airline and other partners to look at the next generation of revenue management and pricing. The one thing I probably love working on in research the most is figuring out what passengers are going to do and how they decide different things. Phil?

Phil Donathy: It's wonderful to be here. I'm Phil Donathy. I'm product VP for most of Skyscanner's traveler-facing product, but mainly on the flights side. That's where I spend most of my time, and that's where I have most of my fun. The teams that I lead look after all of Skyscanner's connectivity to the industry. We have around 800 different API connections, many from people in this room, a search engine that sits on top of that, and obviously the Skyscanner product used by 150 million travelers a month. What a time it is to be alive in travel and tech. It feels like we're living in dog years. More has happened this year in terms of AI and the changes in the industry than I think has happened in the 15 years I've been in travel. What an exciting time to be here. I look forward to sharing some of what we've learned at Skyscanner. I'll pass to Professor Seaney.

Rick Seaney: Welcome, everybody. Thank you for coming out to the ATPCO show. My name is Rick Seaney. I'm the VP of innovation at 3Victors ATPCO. 3Victors was acquired by ATPCO in late 2023. My background is in high-performance supercomputing. As Phil mentioned, I'm also a professor at Baylor University, where I teach advanced cloud computing, and I get the privilege at ATPCO to essentially help drive AI strategy and where we're heading with AI. I'm looking to share a whole bunch of nuggets today, and hopefully we get some questions in as well.

Anand Mishra: Thank you all. So behind me is the Gartner Hype Cycle curve, and I'd like to start there and get each of your opinions on where you think AI in travel sits on the curve today. This is a question for the whole panel. Guillaume, would you like to share your views?

Guillaume Dupont: I think it depends on what exactly the definition of AI is, and that's why it was important to start with data. As I mentioned, I stepped away from the travel industry and the airline industry a few years ago, and a lot of industries are very jealous when they see what airlines have been able to achieve in their use of AI. For most carriers in the room, you have algorithms that take pricing decisions on the vast majority of your revenues. This is, in itself, a significant achievement that has existed for decades. So in my opinion, that would put airlines pretty far ahead, probably at four or five on the curve. Even if I look at other use cases, whether that's customer experience or operations, there are now a lot of mature use cases around handling irregular operations, rebooking passengers and crews, managing assets like gates and landing slots during irregular operations, that show a lot of maturity. Now, if we go back to agentic specifically, I see some levels of maturity in the way airlines are adopting this technology for customer experience. Where I think there's more debate is how that agentic technology is going to translate into business value and potentially business disruption. The prime example is agentic commerce and agentic shopping. I don't see clear consensus in the industry. I see a lot of voices for the disillusionment side, a lot of people expecting that travel shopping is going to be too complex for agents and that customers will not want to adopt this. I think that's very naive. Customers are already using these chatbots, and so something is poised to happen. I also see very loud voices at the peak of inflated expectations, assuming that many players in the value chain are going to fall into obsolescence. Again, I think that underestimates the value that these players provide, both in the way they engage with customers and technically on the data and aggregation side. So I expect that within the next year or so we'll move toward four and five. What it will take is the realization that models alone are not going to solve this problem anymore, but that it's more about what's around the model. For example, models are going to be pretty good at processing customer preferences, but how exactly is this data going to flow? How is memory going to be implemented? There's a lot of people working on this, but what's missing, I think, is some early successes to move away from two and three.

Anand Mishra: All right. Laurie?

Laurie Garrow: I'm going to take a slightly different perspective on this from the educational and workforce training side, teaching the next generation to be our OR analysts. On the education side, I feel that we are probably closer to one than two on the hype cycle, and I say that for two reasons. What would it take to get over two? The first is that right now I feel like I'm telling the students what to do. Use Claude to do programming. Use ChatGPT to answer these questions. We're just now starting to see different ways we can use AI, for instance in creating TA assistants so students can ask questions any time of the day. The second criteria I'd use is that I'm a professor who embraces AI. I actually offer my students extra credit if they can take my homework assignments and use AI to basically break the code and solve them. When I did that last fall, one out of 20 students was able to use AI to solve the homework assignments. When I get to almost the whole class, it'll be time for me to rethink what I'm doing, but also a sense that either the technology has matured enough or the students have developed the skill set they need to really interact with the technology moving forward.

Anand Mishra: That's a fundamental rethinking of how to approach your field. Interesting. Phil?

Phil Donathy: I'll tell you what is probably at the peak of inflated expectations: the stock market valuations of a bunch of these AI companies. To answer the question from the perspective of what we've seen at Skyscanner, I do have to refer back to the previous definitions of different types of AI. One that is definitely at the plateau of productivity, probably for the last three to four years, is the use of machine learning to rank things. Today Skyscanner has tens and tens of millions of traveler searches every day, triggering hundreds of millions of calls to partners. We ingest 100 billion prices. That is where you find the sweet spot for the use of machine learning to do things like be intelligent about which partners to call when, so that we can protect the industry from excessive look-to-book ratios. What we've also seen is that we can significantly improve the traveler experience and conversion by being intelligent through the use of models fed with huge amounts of data. The data is the key, and that's just working amazingly well for us and showing up in traveler experience. If we look at LLMs, a year ago I would have said that we were at the peak of inflated expectations, and there was a lot of what I call performative AI, just sticking a chatbot into the funnel to see what happens. We constantly run experiments. If you're on Skyscanner, you will be in an experiment. You might have a different experience from the person sitting next to you. We found that chatbots didn't work then, but what we are finding now, especially for the half of people who come to Skyscanner who don't know where they want to go or exactly when, is that for those more open-ended search queries, natural language search is starting to work very, very well in terms of helping the traveler get a better answer more quickly, and we measure that in terms of whether it actually improves conversion. There is one interesting thing, though, which is that it's not a chatbot experience. People like pictures, and people like prices, and they like clarity. So we use internally this phrase of generative UI, which is that if I've come asking for a skiing holiday, I probably want to see a different set of destinations and pictures and information versus if I've asked for a different type of query. There's something interesting about getting the user experience right that I don't think anybody has cracked yet, so that's maybe where we start to get into that productivity plateau. And then the one that's probably at peak of inflated expectations, at least for us right now, is this idea of agentic booking, just this idea of trusting an agent with your personal details, your payment details, to make the booking for you. I think the adoption of this will be patchy and will differ globally. What we're seeing, for example, is that travelers in India, just to pick one massive market, are, if I remember correctly, six times more likely to want to trust agentic booking than travelers in the UK or some other Western markets. So that's the one that's at the peak of inflated expectations, but I can see it rolling out in a patchy way over the next six to 12 months.

Anand Mishra: Rick?

Rick Seaney: Wow. What just shows with the three panelists here is that AI is chaos. It is absolute chaos. If you look at that graph, you'll notice the axes are emotions to infinity, or expectations, and time going to infinity. That graph is what I do every day. I go through all those emotions in a single day. I have a cup of coffee, check out some papers, watch a few videos, and I'm at the technology trigger. I have this peak of inflated expectations with caffeine helping a little. I go try it, it doesn't work right, and now I'm in the trough of disillusionment. Then I tweak it a little bit and I'm getting on the slope of enlightenment, and then somewhere there's the plateau. If I was going to say where I am on that graph, I'd have to stack all those graphs, average them together, and I'd be just on the upswing of the trough of disillusionment. Just the scale of the chaos. How many of us would like to do a full iOS update on our iPhones every day? Because that's what's happening with AI every day. It's continuously self-improving. It's getting so much better each day. I go to Claude, for example, and sometimes there are two updates a day. Depending on who you believe, 35 to 45 major features shipped in the first quarter for Anthropic alone, so it's chaos.

Anand Mishra: Phil, I'm really interested in getting deeper into the story you just presented on the booking flow. How is AI changing the way travelers discover and compare flights, and what does this really mean for meta search in 12 to 24 months?

Phil Donathy: It's probably worth splitting two different types of traveler intent. Certainly what we see on Skyscanner, again to give you a scale, is 150 million unique travelers every month. Half of them know exactly what they want to do. I use the example of having to be at this conference. I don't get to choose that I'm coming to Atlanta, I don't get to choose the dates. For that type of search, what's interesting is that the traditional form-filling approach, filling in origin, destination, dates, the travel industry has trained people to do that, and for that type of query, that's still working very well. When we try experiments to say, would you like to replace those traditional search controls with natural language text, it doesn't lead to any improved traveler experience. It's slower to get the answer and doesn't improve conversion. Throwing a chatbot at that type of defined searcher, we have not found to be helpful. What we are finding helpful is for a traveler with relatively undefined search intent. If you think about, I'd like a beach holiday in the Mediterranean sometime in June, for that type of query, natural language search seems to be working very, very well. But we haven't cracked it perfectly yet. We think there's more experimentation to do. So I expect that as an industry we will be able to do a better job of getting travelers answers to their undefined search queries. What will be required for that is the industry coming together to provide all of this, because it all sits on data. The AI is not magic. The AI will hallucinate stuff as well, by the way, if you don't give it high-quality, structured, accurate data. The one thing we need to figure out over the next 12 to 24 months is how we have the correct underlying data engine to power these types of experiences.

Anand Mishra: So Rick, to that point, if AI is only as good as the data that's driving it, what do you think are the biggest gaps on the data side in the travel industry?

Rick Seaney: There are so many gaps and so little time. In the area of coding, it's much more advanced in AI than it was five or six months ago, and we're getting huge benefits out of the coding aspects in all particular areas. And as Phil mentioned a couple of times, experimentation is really important, and having a culture at the company of experimentation. When we talk about what a consumer will have, I think that's the biggest change. Before, airlines and all the third parties that service airlines were way smarter than the consumer. They had revenue management techniques. They were trying to extract the most from them. Now, a consumer essentially has as many PhDs in shopping for air travel as they want on their shoulder, and they're going to be wanting to use those to ask things like, what's the best time to fly, when's the best time to buy my ticket, where do I go next, oh you read my email and I said I wanted to go backpacking three years ago, how do I find backpacking trips? All those things we've yet to even see surface yet. And when they do start to surface, we have to be ready as an industry to deal with them. Somebody earlier in the show mentioned they had, I think it was Sabre, one billion searches a day. We estimate somewhere around five billion shopping queries. That's a drop in the bucket of what it's going to be if everybody has an agent shopping on their behalf.

Phil Donathy: Can I build on that point? There's something really interesting there about the explosion of searches that have not yet led to an explosion in bookings. We're seeing easily 20 to 40% year-on-year growth in shopping transactions, and bookings at an industry level are maybe five or six percent year-on-year growth globally. There's a look-to-book challenge in there, which is very important from a cost perspective. Honestly, my plea whenever I talk to anybody in airline technology is there is one thing that travelers really hate, and they think the industry is doing it to them on purpose, which is inaccurate prices. People think it's bait-and-switch pricing. You hear all these memes about people clearing their cookies and all of this nonsense. As an industry, we have significantly improved the accuracy of prices, but we're going to need to find a way, as agent-driven shopping transaction volume continues to increase, to protect price accuracy and improve it for travelers. Based on data, that is the number one thing that leads to a traveler abandoning a booking: not having confidence in the accuracy of the price they're seeing.

Anand Mishra: Thank you. Laurie, forecasting is at the heart of revenue management. How much harder or easier does forecasting get in the world of AI shopping and agent-driven demand?

Laurie Garrow: One of the things I'm most excited about, hearing the panelists here and also from this morning, is that we were seeing agentic AI being used more on the operations side. I get more excited when I go to the primary customer intent side. If you think about what we do for segmentation right now, we really only have trip characteristics at the time of booking, sort of after the search has happened, that we can use for forecasting demand. So I know you're more likely leisure if you book in a party of two or more, if you're staying a Saturday night, if you have a certain length of stay. But how much more exciting would it be if we really understood destination choice? I want to go to a ski destination is where we should be defining our revenue management. If you think about where we need to push things, there's been a lot of discussion on, do we bring in one airport or multi-airport, do we bring in multiple search days. A lot of those assumptions we don't currently have in revenue management. This, to me, would suggest that we can do different forecasting or customization based on primary intent. For me, it's a tough problem, but I think it holds a lot of potential for us to do better segmentation as well as, formally, customization of pricing or price discounts in the future.

Anand Mishra: Guillaume, cloud was the headline for multiple decades, and AI is the current headline. What are you seeing in travel that's moving from pilot to production from some of your customers?

Guillaume Dupont: I do see a lot happening in the area of customer experience. Call center was probably the early emerging use case because it's such a friction point for the majority of airlines. Pretty early on we started to see customers, for example United Airlines, use AI to help call center agents summarize reservations, to be more proactive and offer better service faster. That was the early premise of AI. We're now looking at reinvention of customer experience through the lens of AI and the data that goes with that. I typically see at least multiple customers looking at collecting digital footprints from their engagement with an airline. That can be the way you use the app, looking at what you've tried to do in terms of digital interaction, looking at chat and call transcripts to really go in depth on what this customer has been trying to achieve, and then taking proactive steps to remove that friction. For example, did this customer try to change flights but fail? Those things happen every day within airlines. How can an airline proactively respond to this, either through the digital experience or even through the physical experience, to remediate those things? Look at technologies like geofencing to see where you are in the airport and facilitate the journey to the gate on the day of travel. These are really prime use cases that I see going into production. There are certain domains that I think are highly mature or getting more and more mature, in operations, in IROPS. And maybe broadly speaking, the way software is being produced is being reinvented as we speak. One of the most fundamental aspects of AI is enabling the speed of development and iteration, which in turn unlocks innovations across the board. That's what we're looking at.

Anand Mishra: Let's broaden that question for the whole panel. What do you predict will be structurally different because of AI in travel in the next 12 to 24 months? Rick, I'll start with you.

Rick Seaney: If we stay outside of robotics, which is coming but I think it's a little bit between three and five years before we see a bunch of robots running around, I think structurally what you'll find is that the human interface to a computer will move more to natural language. Essentially, the keyboard and mouse, and by the way the mouse was a great invention in the early '90s, will see that structural change dramatically on the consumer side. I think you'll see the enterprise actually slowly adopt agentic methods in a much slower fashion than consumers, because on the consumer side the bar is so low. If you've ever used Alexa or Siri, the bar is about as low as you can possibly be. So just having ChatGPT is like having a miracle worker when it first came out two and a half, three years ago. And I think adoption will be slow for enterprises because it's so chaotic. It's going to be hard for people within their organizations to change to a cycle of experimentation, dealing with chaos. What happens three versions down the road when a new version comes out? How do I upgrade? Am I stuck in this model from 2025, 2026, 2027? Can I upgrade? Those are all difficult questions we certainly don't know the answer to quite yet.

Phil Donathy: We're all travelers, and it's funny whenever you take a trip, we're all unique human beings with unique needs, and it's very difficult today to express those needs, the true intent of your travel, in any way that people search for travel today. I do see a step change in that. I see that people will much more be interacting through voice, through text, and they're going to expect much richer, much more personalized, much more contextually relevant responses to their searches. The human-computer interface will change. What is interesting is that the dynamics of the industry might also change in terms of where people are expressing those questions. Are they going to be doing it through Google with Gemini? Is ChatGPT going to become a truly giant consumer brand, a new gatekeeper at the top of the funnel for the industry? The other thing that will change, I hope, and somebody mentioned it on a panel earlier and it really resonated, is this idea that as an industry, we're very, very inefficient. People talk about millions or billions of shopping queries. Most of that work, that great compute that has been produced to generate a shopping result, isn't seen by a traveler because it's sitting on page three of the search results. And people don't get to page three. I'll give you one stat: 70-plus percent of all searches that happen on Skyscanner result in the traveler choosing to redirect from one of the top three itineraries. Partly I'll claim that's because we've done a good job of ranking them, ranked by best rather than cheapest. But also, people just don't have the patience to click to the second page. So I think as an industry over the next year, we will become much more efficient around using the feedback loop and the data to figure out where the traveler is not seeing the offer, and how to either not compute that if they're not going to see it, or how to make the offer better so that they do see it. That's a big shift for a lot of people in the industry.

Anand Mishra: Yeah, that connects to the data in motion aspect as well. That data also needs to flow back into the ecosystem.

Phil Donathy: AI is only as intelligent as the data that you feed it, and today we haven't got those rapid feedback loops in place across the entire industry. The place that actually does have it working very well is online travel agencies, who will make second-by-second pricing decisions off the back of real-time search ranking results. But collectively as an industry, we have a lot more to do.

Rick Seaney: The data in motion, the velocity of data, is going to change dramatically as well. Intra-day processing, we're still struggling with daily processing in many cases, but the velocity of data is certainly going to be one of the big architectural changes.

Anand Mishra: All right. Laurie?

Laurie Garrow: My vision, if I think about the consumer, is where I would love to see us move with AI. Not only the original intent, but looking at how it pushes me further to think about destinations. What I mean by that is one of the things I most enjoy right now, personally, about working with AI is learning things about new places I may be visiting that I didn't realize were tourist attractions, or that deferring by a week might save costs or the weather would be better. So thinking not only about a comprehensive travel experience as the flight and going to the airport, but really the package of destination, be it the hotel, tourist applications. For example, when I've worked with cities in the past, it's amazing that as tourists in Atlanta, most tourists go to one of 10 destinations. So just thinking about how to harness that data to really pull together a full trip experience. While we're probably not there yet, I do hope it's something we think about moving forward.

Guillaume Dupont: Generally aligned with what's been said. Extending a bit on what Phil was mentioning, I do think we'll see pretty significant changes in business process as a result of what AI is bringing. The parallel I'd like to take is just what happened when basic economy emerged in the industry five to ten years ago. Of course, many reasons for that, but one was the proliferation of the meta search channel. I fully expect that similar things are going to happen. I'm not sure exactly what it's going to be, whether that's flexibility rules or too many pending offers and unticketed PNRs requiring a positioning from a business perspective from airlines. I think we'll see initial agentic use cases becoming real within the timeframe you mentioned, one to two years. The one that's really, in my opinion, prime for disruption, is corporate travel. I have an agent that runs on my laptop, looks at my calendar, proactively sends me notes and preparations for meetings, and helps with data entry. I don't see why within the very near future, when I get a workshop invitation in a city, this agent is not able to proactively at least send me offers that meet my need, knowing that my corporate portal understands my loyalty programs, preferences, and policies. I do think as well, extrapolating on channel dynamics, during the COVID period we saw a lot of data partnerships between hotels and airlines sharing customer data. I think this is poised to happen again. One of the main value propositions of agentic travel from my perspective is finally looking at the whole experience. No longer booking flight first, then a hotel, then activities, then ground transportation. Bringing all of this together. Some players in the industry have organically high-quality data to do this. First-party providers have much less of it, and so we'll see aggregation from a first-party perspective for inventory and customer data. The last thing I'm actively working on that I think is already happening is a complete reinvention of early funnel marketing. SEO is becoming less and less relevant, and we're now looking at generative engine optimization. Going back to highly structured data, as Phil mentioned, there's a lot of implications as we peel the layer into how content gets created, how offers and the content within these offers is presented, processed, audited through continuous improvement. I would expect pretty significant changes again from a business perspective within the coming year or so.

Anand Mishra: Great. Thank you. So last question for the panel. This is a lightning round. Buy, sell, or hold on AI, ML, LLMs. Which would you buy? Which would you sell? Rick?

Rick Seaney: I'm going to be a buyer, except I don't know what tokens are going to cost in two years. But I'm definitely a buyer.

Phil Donathy: We talk about this a lot because we spend a lot of money on tokens, and that money was not in budgets two years ago and it is now. Skyscanner has decided we will be a buyer of AI. We are spending a lot more money on compute, which is probably making our friends at AWS very happy. There's a different question, which is okay, but how are you going to fund that? Because it's fun to play with AI. I personally burn a bunch of Claude tokens. It helps me to be a better product manager. But at the end of the day, this needs to have a return on investment. The way we think about it is we are going to be buyers of AI, but buyers with a very keen eye for return on investment. We expect that we're going to be able to increase revenue per employee as a result of using AI tools to accelerate the productivity of all of our employees, but primarily in our software business, so developer productivity, product management productivity, even designer productivity. A big buyer, but with a keen eye on the return on investment.

Laurie Garrow: Education has to buy. We have to invest in the future technology. What I find exciting on the educational front is that we have a lot of rules and regulations, certification programs, but students are demanding more value for their degree and want to make sure they get a job when they get out. We're going to see a lot of tension in terms of staying ahead of the curve and making sure we are prepping students to be able to go out and get jobs in this highly chaotic environment, whether that's being able to retrain or look at new things. But a very big buy, and experimentation to get ahead of the game.

Guillaume Dupont: Same, buying AI too. A lot of people in this room have been in technology. The hardest part used to be change management, and I'm just so amazed at the level of adoption that AI is seeing and how much more easily it's been embraced by society, in personal life but also by workers within corporations. This needs to be bought in and reinforced. I would hold a lot of LLMs, but as I was mentioning, models are less of the bottleneck now than they were two years ago. What needs to be explored now is what's around the model, all the data infrastructure, but also the user. One of the probably even harder challenges than technical challenges is how users are going to work with agentic commerce, what the user experience is going to look like, and how the trust they have in today's channels is going to translate into agents. I would buy more of machine learning. Many of us in data science were shifted to LLM and agentic when large language models came out, for valid reasons. The technology is so powerful. But the more time passes, the more we realize that these technologies are highly complementary. LLMs, as the name indicates, are great for natural language reasoning, so very good for customer experience, but they do not replace revenue management and a lot of the pricing decisions we have to make. Looking forward to seeing more and more the complements between these technologies.

Anand Mishra: Thank you. Thank you to the panel. It was a fascinating conversation, from chaos to potential, servicing experience for the traveler. We have an exciting road ahead of us. Please give a round of applause to the panel.

‍