Indian online travel agency Ixigo has been on an interesting journey for the past decade, evolving from its roots as a metasearch to online travel agency.
It began by solving friction for train travelers in India but has developed and grown through a string of acquisitions and investments along the way, adding technology and inventory.
It was also an early mover when it came to experimenting with artificial intelligence.
In an interview with PhocusWire, Rajnish Kumar, co-founder and group chief product and technology officer, spoke of the company's early AI developments, the potential use cases for ChatGPT and current challenges.
Tell us about the importance of data, what you’re collecting and how you’re using in a meaningful way?
It’s important to understand the background for why it has become so important for us. When we started, we had a very different approach compared to most other online travel agencies (OTAs). The approach of most of the OTAs has been to build a business on top of transactions. You start by building a business where you say, "I'm going to sell some tickets and make some money." That’s the initial value proposition.
Our approach was more akin to the approach of a Google, a Facebook or WhatsApp. It's more of a utilitarian approach. We started by saying that we were going to solve some of the core problems people have. We might not make money on that, we might not sell tickets to start with. That really started the whole journey.
The initial approach was that we became a metasearch, which was quite a natural fit for our mindset. We didn't want to get into anything operationally heavy. Metasearch as a business was lending itself nicely to that, and the fact that India as a market is very price sensitive, it was good timing for us.
Then we identified a very large underserved market, which was the bottom of the pyramid in India. People were traveling in trains, and they had lots of issues which were unsolved. I think, at the core, Indian Railway Catering and Tourism Corporation (IRCTC) was only engaged in providing a ticketing platform, but everything else was not solved for.
For example, the classic use case we started off with was checking passenger name record (PNR) status. Thirty percent of tickets are wait-listed in India. When you book a train ticket it's a very painful journey for customers to keep checking their PNRs every now and then, up to very close to the day of departure. They need to see it’s confirmed, because if not, you want to cancel it in time so you don't lose a lot of money.
This process was painful, so essentially we created an app which would read [Short Message Service texts] coming from IRCTC and even airlines and OTAs, and there would be trip information in those SMSs. It was like TripIt for SMS because in India almost everything goes in SMS and now WhatsApp, not email.
So we pick that up and work it into a trip so people don't have to remember their 10 digit PNR number and go to a website to punch it in every time. We check that automatically, and as soon as your ticket gets confirmed or your waitlist moves we send a notification. That becomes extremely preemptive, and this is the secret story of how our initial growth hack worked. Because of that, people started sharing the app like crazy.
With this growth, data came to us. We got a lot of data about customers checking PNR status and how their PNR status was changing in real time. That data was valuable to us because we then built our first artificial intelligence application on top of that around 2014-15. We started predicting PNR status, which means that if you check your PNR, or if you check availability of a train before buying, we can tell you whether this ticket will get confirmed. That became one of the biggest features that became viral.
All the OTAs are now talking about mobile and the opportunities going forward. As someone who has been dealing with mobile for a long time, how do you see it being used in ways not currently being used?
In the early days, a lot of people figured [mobile] was only a platform for looking and not bookings. In India, that changes completely because a very small percentage of people have laptops. Everything they do happens on a mobile, so we leapfrogged compared to Europe and the U.S. What is changing now is that mobile is ubiquitous, and everything is happening on mobile. More and more what we'll see is that even during the trip, people will need support, and it could be a wide range of things, such as what to do in a city, what to do around a certain area and even foreign exchange.
Last year there was a lot of talk about metaverse and NFTs. Do you still see potential for them in travel?
Personally, I would not really bet big on them for the simple reason that if I have to put my effort into something, I would rather put it on ChatGPT or the large language models. That's an area that is extremely powerful and can be tapped in to, and there's enough to be done there.
You have been using AI for your app to help solve friction points for customers. Where do you see generative AI fitting best into travel?
We started with mostly predictive technology like the PNR status prediction, and we even built a trip planner around 2012-13. It was based on ancient AI models used by us to create a natural language understanding engine, and we were able to show you exactly the content you were looking for.
In hindsight, it looks as if ChatGPT is doing something very similar. We built that thing, and it looks like today it's very relevant, because I don't think people were expecting this to be as powerful in terms of the ability to understand what somebody is trying to say. What we built had lots of limitations because it was impossible to have a natural language understanding engine which has the capabilities of a large language model today.
Subscribe to our newsletter below
So we did not focus on that and started focusing instead on where AI was sufficiently advanced to give more accurate results, and that was predictive technologies. So our initial phase of evolution was where we built features to help customers, predicting PNR status and predicting train delays.
In India trains are not usually on time. Because of the scale, we reached a point where we figured out that we could build something like a Waze or Google Maps by using the location of the devices inside the trains to pinpoint the location of the train, which was one of the biggest pain points for customers. That was not the only problem. Once you could locate a train, you also need to know what happens to it once it gets to the final destination. If a train is running an hour late at one station, you need to know if it will catch up or get delayed further, so we built predictive technology for that.
You have these predictive AI models. How does ChatGPT or generative AI take things to the next stage?
The next stage of the evolution was to start using these predictive models for monetization, for example, Ixigo Flex and Ixigo Assured.
Initially we used to predict airfares by telling customers not to book because the price was going to go down. You can do that or you can underwrite that by selling a prize freeze. We looked at the cancellation behavior of customers, airlines, date of departure and many factors including the customer's own historical data to figure out how much premium we should charge to underwrite cancellation. All of that is today powered by AI.
So where does generative AI come into the picture? We started building a chatbot around 2015-16. We realized we could evolve that technology into a customer support chatbot because as a business we were transitioning from a metasearch model to a fully transactional model. Doing that had a cost, and the cost was customer support, so we had to make sure we had something that would create that operating leverage for us, which is why we transitioned that bot to a customer service bot.
When we started using generative AI, I think we got access in 2020 to GPT-3, we started using it for initial use cases mostly inside the organization. We never replaced our chatbot with that because it was not accurate.
The ability to mash ChatGPT up with the real-time APIs you have is where the plugin architecture is beautiful which is what I think travel companies can leverage really well.
Rajnish Kumar, Ixigo
Everything changed in 2022 when version 3.5 came, and we started playing with that, we realized it was way more powerful than version three and that’s when we started transitioning our chatbot. We started fine-tuning it with 10 years of data, which resulted in a significant bump in accuracy and handling rates. We’ve been able to reach a point where we're getting closer and closer to 90% handling rates over the chatbot. That's a big improvement because moving from 85 to 90 with this technology requires enormous amounts of data to train the models.
For the travel industry in general, what’s the greatest challenge in using ChatGPT or similar technology?
One of the issues with ChatGPT is that it is more or less static in nature. The amount of data it has is up to 2021 and everything is pre-trained on that data. The ability to mash that up with the real-time [application programming interfaces] you have is where the plug-in architecture is beautiful, which is what I think travel companies can leverage really well: if you combine the power of ChatGPT with real-time APIs that you have on flight status or more accurate real-time data on destination content or places of interest. You might ask for that data from ChatGPT, but that data is 2021 so take, for example, the best places to eat, maybe that place has shut down or the ratings are not correct or maybe it doesn't know the best restaurants that have come up since 2021.
So the challenge is that it's the data pre-2021, but there's an opportunity for travel companies to combine it with real-time APIs?
That's what we've been working on over the past few months. It's something we'll be showcasing very soon: the power of this tech when combined with real-time data and APIs that we have built over the past decade. The power of that is unlimited ... The trick in most of these cases is what you want ChatGPT to give you totally depends on prompt engineering. It’s how you train it to give you the outcome you want.
What are your current development priorities for Ixigo products?
We spent the past decade solving core customer pain areas, and as a result we have grown exponentially. We've reached a point where we have 60 million monthly active users (MAUs) and most of the downloads are organic. So for the past few years our focus has been on monetization, and we produced all sorts of booking funnels and transaction funnels and eventually added the embedded finance products to expand our margins. We're now seeing that operating leverage playing out and our EBITDA margins expanding significantly. The goal for us now as a business is the immense potential we still see to monetize the customer base.
Will you monetize in a similar way or do you have new plans?
We never had before the kind of intelligence that could understand customer behavior and then sell them the right product. If you know that somebody has a wait-listed ticket and that ticket is not going to be confirmed, which is a pretty realistic scenario, then there is the opportunity to cross-sell a bus ticket or upsell a flight ticket if the trip is too long. This is the kind of intelligence that we have been working on to scale the monetization and the cross-sell potential and new lines of product. It would be lame to be sitting on 60 million monthly active users and not have your own hotels product. It would be like leaving money on the table. So we'll have the full breadth of travel product, but more importantly we’ll have the intelligence to understand customer behavior and then use personalization and recommendation to sell the right kind things to people. That potential is really high.
The travel industry has talked about personalization for a long time. What does it mean to you and how can we achieve it?
That was the vision we presented at The Phocuswright Conference with the trip planner back in 2017. The idea was to create this virtual concierge who would know everything about you, all your tips, everything in the past and always be there for you. I see that that vision can be achieved now with this technology. It’s just getting closer and closer.
Can we borrow from other industries when it comes to the customer experience?
I think travel is by far the worst when it comes to the customer experience. What Amazon was doing 20 years ago travel is still struggling with now. That ability to put five products in a shopping cart, check out and get them delivered or, if you want to give them back, you can give them back. In travel that is not trivial. It's because of how this whole ecosystem was built, you have to do everything in a certain way, which is exactly the reason why we built Assured and Flex. I’d rather underwrite that with AI and give customers the experience they deserve.
Other examples are industries which are powered by pure play technology like the media and music industries.
When you wake up in the morning, what do you think your greatest challenges are?
The greatest challenges are also the reason why I wake up in the morning excited. I know there are problems I can solve today, and I can see real impact in terms of the user experience and margins in our business. That gives you a kick. In the recent past the biggest challenge we have all been facing, and still are, is how to balance the kind of work you need to do to fix things and deal with legacy issues with making sure you're also able to grow the business.
What has changed now is that while hiring the right people will always be a problem, 10 engineers is now like 100 engineers with ChatGPT or Copilot. What you could achieve in the past with 100 people you can achieve now with the right 10 to 15.
The cost of building software used to be directly proportional to the cost of engineers, and the cost of engineers has been going up exponentially in the past couple of decades. With this technology inflection there's going to be a flip side drop in the cost of building software. Software can be built very easily, which means any business which does not have a clear physical moat is ripe for disruption. So my biggest challenge is how do we build traditional moats in our business to make sure we are able to use this technology to create disruption and not get disrupted.
Phocuswright Europe 2023
Generative AI is set to revolutionize travel. Leaders from Trip.com
, Microsoft and Kayak tell us how the technology will change everything from the back of the house to the front.