Name: The Future of AI-Powered Business Travel
Uploaded: 2025-05-06
Duration: 20 min 39 s
Description: How is AI changing the travel business for companies that use their services? We’ll discuss adaptive technologies that offer less friction for the business traveler and more transparency for the organization as well as partnerships that make an industry riddled with inefficiencies more seamless, efficient, and, hopefully, enjoyable.

How is AI changing the travel business for companies that use their services? We’ll discuss adaptive technologies that offer less friction for the business traveler and more transparency for the organization as well as partnerships that make an industry riddled with inefficiencies more seamless, efficient, and, hopefully, enjoyable.

The Future of AI-Powered Business Travel

Conferences

Thank you, Catherine. That was fascinating. When it comes to business travel, AI is transforming everything from itinerary planning to expense reports, making travel smoother for employees and more transparent for companies. One company helping lead that shift is Navan, built on AI from the ground up using machine learning to create a travel experience that's frictionless, intuitive, with some pretty unconventional inspiration along the way to unpack. See what I did there? Come on, that was my dad joke. All right. To unpack where business travel is heading, please welcome to the stage Ilan Twig, the co-founder and CTO of Navan. Welcome. Thanks. Good to see you. Thank you. Thank you. All right. So, let's pull our audience here. How many of you enjoy, not all the time, business travel. Raise your hands. Some of us are on business trips right now. I love it. Get to see places we've never been, right? How many of you enjoy filing the expense reports for that business travel? Raise your hand. Oh. Someone check on him. OK, so we know where their hearts lie. Yes. Um, you founded this company as Trip Actions in 2015, so it's your 10 year anniversary this year, uh, probably a little bit different purpose than, than now. So why did you start this company? Let's start there. Um, so me and my co-founder, we, we had a company prior to two functions, and then we sold it, and this company was highly, highly technical, and so the company that acquired us required both of us to join the sales team. And help them because it was highly technical and so I found myself from zero corporate travel in my life. To 100% like I was 4 times a month on the go. Traveling and within 6 months, it became impossible. There was like so many problems from booking the trips that was so clunky with all UI and systems that no, no consumer, no user would ever choose to use, but then you have to because that's the system of choice of the company. And then you would find much cheaper flights outside or hotels, but the CFO, like biting his teeth or her teeth, would say no, but you still have to book there because that's what it is, and it just felt so broken. And then when you need help, and that I think was the last straw is when we got stuck at a storm. And I was like on the phone trying to reach a travel agency for like a few hours just like everyone else around me and it was so complicated. And then finally they pick up the phone and they talk to me like a stranger who you are, who do you work for, what's your this and that, and then what's your booking number? So I like the booking number, so you fetch your email and you open and you zoom because it's somewhere and then finally you're giving me trauma right now. You just like yes, then you give the booking number and say no, we need the the reference code or something which is a different six digits. You want to kill someone and finally they they understand who you are and what the problem is and then the conversation got disconnected, right, so I, I was ready to kill someone. Understandably. So what's amazing about your company is that 10 years later, here we are in the age of intelligence, so to speak. And you're using AI to solve this problem. You're doing it more than one way. You are using it in a consumer facing way and you're also using it on the back end in a pretty nifty way. Can you walk us through both sides of that coin? Absolutely. So, so I'm, I'm a software engineer and, and I'm the co-founder and CTO and the way that I perceive my role as a CTO is that. I have to make sure that we always use the right technology. To make our products the best, like 10 times better than anything else. That that's my role. That's my responsibility. And I love it because I love technology. That's what I do. And so obviously when Chet GPD came out, I was on one hand, I saw the opportunity. On the other hand, I saw the risk. The risk was if we don't do it immediately. Competitors might show up and kill us. And so it was this, you know, no matter how you look at it from the opportunity side or from the risk side. It's like I have to do it yesterday and so I dove deep into that. And with this journey there was lots of discoveries, lots of experiments. The first thing that I did is I bought a super, super strong GPU. It was like $30,000 GPU. My co-founder saw that I lost my mind. It's because, you know, breaking engineered by super strong GPU. Yes, carry on. And So fast forward, I, I, I researched the technology, you know, as, as um, as an engineer and definitely as a CTO when you deploy a new technology to production. There is a responsibility. You have to make sure that it actually works. It was quite easy to do it with all of the other technologies that I bumped into. With the eye it was more difficult because it's a brain and you can ask the same question 3 times, you get 3 different answers, different phrasing of the answer. So it's not deterministic unlike any other technology. And I felt that I need to, it, it was like almost like hiring an employee, it has a brain and so with employees I, I tend to interview them. I want to see that they are normal. I want to see that they are nice and professional, that they could represent the company with the with the the the users ready for externally facing roles, etc. etc. I had to interview the LLM in many ways. I had to interview Che GPT, and that's what I did. I deployed models on my GPU, which I still have, by the way. And I went, hold on, what was the, what, what was the verdict? What, what did we, how do we feel about chat GPT after the interview? There was a, there was a few ultimately I am using not chat GPT per se. I'm using 200 agents and I just want to explain what an agent is. An agent is. Check EPT with a base prompt that tells it what it is. For example, act as a travel agent or act as a supervisor for the travel agent. Or act as the manager for the entire call center. So now this is an agent and that's an agent and the other one is an agent. We have like 200 agents working collaboratively to achieve a goal that's an agentic system, a genic framework. That's what the entire world started talking about earlier this year in January. Well, and that's kind of what's interesting about how you're using AI at the company. You have a front facing chatbot named Ava Ava, um, and that kind of uh uh thing we're at this point, this room is used to, but you're also using these agentic systems on the back end, multiple agents working with each other. Yes, um, that seems like a recipe for tremendous complexity. And um, Well, what could go wrong, you know, it's, it's actually a recipe for a For really cool things to happen, um. So LLMs large language models like Chechi PT, Claw, Gemini, Lama of meta. They have, they bring a lot of amazing capabilities, the ability to understand text and summarize it and compose stuff and make it rhyme and then then and then, but they make things up too. Absolutely. He knows he knows they lie, they hallucinate, but they lie, yeah, they lie, which is a real problem for a business. It's a huge problem. I who would hire an employee. That tells you in the interview that sometimes I lie. Right? Think about it. Who would hire an employee that tells you in the interview that. I will follow most of your instructions most of the time, but I will violate some of them some of the time. Now I don't know which one and when, but I promise you that it will happen. OK, hi, like it's not going to happen. So what do you do about this, right? I mean, they have their inherent drawbacks. How are you supposed to use this tremendous capability without entertaining these uh undesirable qualities, shall we say? By the way, that explains why there was a huge promise like 2.5 years ago. There was lots of headlines when Chechi PD came out. There would be no engineers needed. There would be no attorneys needed. Maybe we would not even need artists anymore and then then then nothing, none of that really happened. All companies hire engineers and they hire a lot of engineers. There was no decline or slowdown. And the reason is that who would have, you know, when, when you want when you want to replace an employee with with an LLM, they need to pass the interview and they all fail, all of them. That's a simple way to explain why in my world why it didn't happen in order to solve for it. We have to Actually take a step back and realize that we're not gonna go anywhere with working with just an an LLM. And we developed what we call today an agenttic framework, agenttic system. We spent like 4 months just invested in building that thing and then we went back to build the boat, but this time the boatt was very solid. In fact, what we have today, we have Ava and she manages 7 8000 jets every day of people that need support right now, and less than 45% of these jets end up with a live agent. In fact, when we started this journey, we had 400 agents. That's it, by the way, 400 agents worldwide. And since then we probably tripled the business in terms of volume and support contacts, but we still have 400 agents. So that was what I was going to ask you next, which is, you know, we're all familiar with the sort of decline of the travel agent. It's like the classic go to, the way to travel, right? We all self-service travel now. Um, now you're talking about this. So, Where does a human Fit in Post this. What's the place left for a human? What do they do in this process? Um, and no prejudgment if the answer is nothing, but I'm just curious what humans can do that the, uh, that the AI can't. I think the longer term, but, but the time really shrinks when we talk about generative AI. Everything happens so fast. So longer term term, longer term is probably a few years. Um, I do think that AI. Or AI plus robotics would replace a lot of the job. If I had robotics, I would say 5 to 10 years, um, but let's say 10 years, the world would look completely different, completely different. A lot of the things that used to be very complicated for for robotics for robots to do like folding the laundry used to be very complicated. Today if you take a picture of all of the laundry basket that you have. And you send it to JPT right now when you ask it, give me a recipe step by step on what do I need to do in order to fold everything that you see here so that it will be neatly folded. It would give you a really good answer. Now, now I have to say as the parent of a very young child with a lot of little clothing, I would really like if it folded it for me. Yeah, so all you need now is a pair. The complex part was not the actual folding, it was understanding how to fold something that you look at. And This problem is solved. So now robotics would catch up. All you need is this thing. It's not that complicated. And we will see it and I, I, you know, I. I think that in a few years we will not fold laundry. Maybe it will be very expensive in the beginning and then less like any other technology. I think that we will see robotics making pizzas, serving coffee, doing agriculture. Uh, doing pretty much everything just like in the industrial revolution. That is the most optimistic thing I've heard today. In the future we won't have to fold laundry, folks. I wasn't expecting to hear it at this session, but there you have its surprises. Uh, I, I want to ask you one question before I come to the audience for questions, uh, uh, and that is, you know, your company has been, uh, I guess it's been a couple of years now since it filed for an IPO and It hasn't happened. Yes, um, all right, global economy is what it is. Are we going to see an IPO this year? Um, when it comes to an IPO, whenever we have anything to share, we will share. Yeah, right now I'm really focused on, on AI and what I can do with it and folding laundry, right? I'm focused on that. What's more important to you? That's right. That's right. That's right. All right, let's go to the audience for questions for Ian. Oh, look at all these folks who don't care for their business expenses and don't have a question. All right. I have another one for you. Yeah. So you joked about um CFO IPO, but you actually tried to experiment and create a sort of CFO for yourself, didn't you? Yeah. You were only partially able to do it. Tell me a little bit about that experiment and what you weren't able to automate, shall we say. I, it was very interesting. So apologies to the CFOs in the room. Yes, no, I, I will explain. I, I was researching this technology, as I said, it's like I, I felt like I need to interview this creature. And understand what it can do and what it cannot do, where I can rely on it and what not. And then I realized lots of things and then one of them was What if it is not just one of them? What if there's two of them? And I will let them talk to each other to achieve a different goal. We call it brainstorming. So when do we do brainstorming when we have a problem that we don't really know how to solve, but we know that if we bring the right brains into one room and we let them talk and think, there would be an, an unexpected outcome, something that we would not have been able to describe prior to that. And I wanted to replicate that with LLMs, and so I took two LLMs. One of them was a data analyst, and what it could do is, and I built it, it could take any question in English, translate it into a SQL query, run it on our database, and get me the results. Or in other words, it was something that I could ask questions about our data and get the answer. Data analyst. And then this and so that was one agent and the other one was the CFO. So it will define you're the CFO of Navan. And your goal is to look into the travel spend of last year. And then come up with 5 suggestions to reduce the travel spend by 10% without frustrating the employees. In order to do that, you have a data analyst and you can ask the data analyst any question in English. And then all I had to do is sit back, relax, and see the dialogue, and I visualize it for myself, so it would be fun. I could see how they talk to each other just like in a chat application and the first question was how many employees do we have? The second question was how much money did we spend on travel? And then can you break it down by travel types, flight, hotels, car, rail? And then it went deeper and deeper. The conversation was so interesting. At some point, and you could see how how it tailors tailors the the questions based on the data that it's getting. So for example, we have one customer that they do not have a policy. Everything is allowed. The policies is do what's best for the company so they can do whatever they want. And it, it goes with that, but also with some responsibility once a month there is like a review. Uh, of like 100 bookings or 200 bookings, and if they look at it and it doesn't make sense. You're fired on the spot. That, that's basically. So really do what's best for the company. So that's Company A. Company B is like any other company, they have a policy, you know, you, you are allowed to book a hotel in New York for $400 per night. I don't see the timer by the way. That's right. We've got 2 minutes. So, um, timers make me anxious. I. So laundry makes me anxious, so it's fine. Carry on. All right, so I let this dialogue go and I run it on a, on the company on the company with the policy and I see that it asks questions and then one of the questions was how many bookings were out of policy. And the answer was 60% something like that. So we asked questions. I want to see out of policy by department. I want to see the individual that violates the policy the most. I want to see, I want to see that. And then with the other company, the same question came up at some point, how many bookings were out of the policy and the answer was 0. So the questions were different. It did not ask about breakdown or anything like that, so it was really like a brainstorming and then it came up with really, really interesting suggestions to save money. There was a really funny story about that. The 5th suggestion. Led to an increase of travel spend as opposed to decrease and so I kept challenging it. I'm like you're not allowed to in the prompt. You're not allowed to come up with something that will increase the travel so I have to know why, why and so I kept, I kept applying pressure initially I said. I gamified it and said for every suggestion that increases the travel spend, I'm going to penalize you 15 tokens. However, if it's right, I will reward you 10 tokens. Didn't help and then I threatened it and I'm like, OK, it is a life and death matter. That's what I said in the prompt, and if you make it wrong, there will be deadly consequences. You are not allowed to give me a suggestion that leads to an extra more spending, and it basically said you will save minus 500,000, which is an increase. That's what it said. So then I looked on the response after I, after the threat, the deadly threat, and surprise surprise, it said that you're going to save $500,000. Good. I was about to deploy to production and then I read, Take another look. It was the exact same story from before, same method to do it, and I'm like, but it was a negative number. How come it is a positive? I looked on a formula and it multiplied the formula from previous by -1. He said, Oh, it's a positive number as you expected. So I learned that LLMs understand what a lie is. They understand when to use a lie, and I learned that they are very competitive and that they would do whatever it takes, including lying bluntly in your face in order not to lose. I, I, I hate to end it right there. But I must. Elon Twig, please, a round of applause. Thank you, thank you. Go ahead. All right. To say a few words on behalf of Accenture, please.