It didn’t take long for generative artificial intelligence chatbots like ChatGPT to sweep the travel industry off its feet. When the technology’s shortcomings with facts and potential bias were exposed, companies developed plugins or other workarounds.
But as the industry looks to move beyond a whirlwind romance to something more permanent, it faces the same conundrum every romantic comedy heroine Hollywood ever created must overcome: How do you take the sweet-talking charmer who plays loose with facts and combine him with the smart, dependable – if a touch boring – fellow who will always be true?
In travel, as in romantic comedies, you might have to go around the world to test such a mix, which is just what the Oneworld airline alliance is attempting by bringing generative AI to its signature Round the World ticketing process. The alliance includes more than a dozen airlines such as American, British Airways, Iberia and Qatar.
With hundreds of possible destinations and too many flight combinations to count, around the world trips have remained largely the province of travel agents, even if a chatbot is ideal for testing all the different permutations that might come to mind.
“By making the process more accessible, we certainly aim to draw in more potential customers,” Oneworld CEO Rob Gurney said. “These are often trips of a lifetime for customers, and [travelers] want to explore options carefully. They welcome help in working out how to make their travel goals practical.”
Subscribe to our newsletter below
For its solution, Oneworld turned to Elemental Cognition. The startup was founded in 2015 by computer scientist David Ferrucci, a pioneer of commercial AI best known as the creator of IBM Watson. Much of his work since then has focused on developing an AI model that would combine a chatbot’s erudition with the logical reasoning a human would employ.
For Oneworld, the tool presents users with a global map and a chatbot dialogue box off to the side. It prompts the selection of a home city, and then the game begins as the user starts adding cities. Will you visit every continent? How many stops? The ticket can have up to 16 different flights with layovers of varying lengths so long as they are completed within a year. Once all the cities have been chosen, the tool presents the optimal route, though users can still make changes. (The base price depends on the number of continents a user chooses to visit. For one adult traveling economy class, base prices for six continents start at $6,899 plus taxes and fees.)
Oneworld said users are four times more likely to create bookable itineraries using the tool than with the standard interface, though Gurney won’t say how many of those have booked a trip.
“Early results look very promising,” he said, “and we expect that as customers become more familiar with the tool and utilize it to create and book Round the World itineraries, that usage of the tool to complete a booking will grow over time.”
To learn more about Elemental Cognition’s hybrid chatbot, PhocusWire spoke with Greg Burnham, a product manager at the company who, as it turned out, shares qualities with its creation by combining the exacting knowledge of a Princeton math major with a chatbot’s skill at explaining it simply. The conversation has been edited for brevity.
Let’s start with some background: How does your generative AI tool differ from ChatGPT and others like it?
Even before ChatGPT, we were using large language models. But that’s just one piece of our AI platform. The other main piece is what we call our reasoning engine. Our reasoning engine does precise, logical reasoning that is always on the rails, always colors inside the lines – and always gets the right answer.
That’s really different from large language models. Large language models typically, what they’re really good at doing, you ask them a question or give them a task, they will give you something that has the shape of an answer to that question: the right shape. But it might really struggle with the details.
A basic example: You can’t use a large language model as a calculator. It won’t multiply numbers correctly. It will give you something that has the shape of the right answer, the right number of digits, but when it comes to precise reasoning that needs to be trustworthy and reliably correct, that’s where our reasoning engine really stands out compared to just a pure, generative AI solution.
We combine the fluency and flexibility for conversational interactions that you get from the large language model with our own secret sauce of precision, reliability, always following the rules, the policies, that comes from our reasoning engine.
Are you saying ChatGPT is like a college student who tries to fake his way through an essay exam with fancy words instead of facts? There’s a term for that I don’t think I can use here.
Your words, not mine, but I think you’ve got the right idea. ChatGPT on its own is great when you just want that first draft, but you’re going to go in and clean it up and add any of the detail or fix anything that went wrong. But even in the Oneworld case, for example, there are some airfare rules there that you really have to get right. And the customer booking it isn’t an expert in those rules. They won’t be able to tell [if it’s wrong]. And that’s not going to work.
In my household, we feel accomplished if we manage to book a multi-city flight itinerary at a reasonable price that gets us home at a decent time without spending half a day in an airport lounge somewhere. It’s hard to imagine how much more complicated planning an around-the-world journey is. What variables did you account for? How many different options did you end up with?
There are three sources of complexity when trying to book one of these around-the-world itineraries. One is the [Oneworld] fare rules themselves. It’s a terrific product. It’s got a nice pricing structure. But in order to make it cost-effective, there are a lot of fare rules. You can’t have too many entries into a given continent. You can’t have too many segments within a continent. There’s a special rule about going in and out of Anchorage only once. There’s a lot of rules. That’s one source of complexity.
Source number two is the user’s preferences. It’s an exploratory process. It’s a trip of a lifetime. As they plan and research, they might change their mind.
There are 10 to the 34th – that’s 10 followed by 34 zeroes – possible itineraries that someone might request that are valid in Oneworld ... Our reasoning engine is able to, dynamically, on the fly, figure out what itinerary is going to work for them.
Greg Burnham - Elemental Cognition
The third source of complexity is flight availability. Not every city is connected to every other city on every day. As you’re putting in your dates, there might be a problem. You might request a city that isn’t connected via the Oneworld airlines to any other of the cities that you’ve requested, so there are problems with availability, period, and availability on a specific date.
What we wanted our system to do is make it so the customer doesn’t have to worry about any of that. They’re going to run into problems sometimes, but we want our system to present that in an easy-to-understand way: “Hey, given what you’ve requested, you’ve hit a problem, here are your options for getting out of that.” Otherwise, we just want them to have a seamless experience, exploring, trying things, settling into an itinerary that they’re happy with and clicking to book. That’s how we tried to construct the entire product.
Of those three complexities, which was the toughest?
What’s tough is the combination. The fare rules are complicated enough, but then you combine them with the preferences of the user, and maybe there’s a way to chart them an itinerary that meets all the rules, maybe not, and then you’ve got to explain what the problem is. Put that together with the flight availability on different dates, and that can change in real time, all of that comes together.
But our reasoning is really up to the task for that. Our whole platform is up to the task. We’re able to have smooth, conversational interactions with the customer, and then we’re able to reason through these really tricky, almost logical puzzles, if you will, where we’re able to find given what they requested an itinerary that minimizes travel distance, minimizes layover time, follows all the business rules and only flies on flights that are available. That problem collectively is what’s hard.
There are 10 to the 34th – that’s 10 followed by 34 zeroes – possible itineraries that someone might request that are valid in Oneworld ... Our reasoning engine is able to, dynamically, on the fly, figure out what itinerary is going to work for them.
How did you test it?
The answer to your question is “very thoroughly.” We had a lot of people. First it was friends and family, [then] we used some crowd workers to try to book their dream itinerary and built some metrics on things like where in the process people were hitting problems. We ultimately validate against the booking system to make sure it’s a bookable itinerary. … One of the great things about building our system is you can test all of the rules sort of individually and trust that they’ll be combined correctly.
What feedback have you gotten?
Anecdotally, one thing that stood out was one person said this was “better than a human” for interacting with. Having used it myself, it’s remarkable how engaging it is. You’re putting together a complicated plan here, and the system is very helpful, just guides you step by step, reacts to whatever requests you make in a way that I could feel that I might be burdening a travel agent if I was using a travel agent for a problem like that. [Laughter.]
It’s a bot, but it’s a smart bot. And it’s a patient bot, and it’s going to give me the right answers, and it’s going to react very easily with me. I just make a request and it says, “OK, here’s your updated itinerary.” Or, “Here’s your problem. Choose A, B or C to resolve the problem.” I can spend a lot of time just playing with it until I’m confident I’ve got an itinerary I really like. And that all happens in real time. We’ve gotten feedback to that effect.
If your boss, Dr. Ferrucci, came in and said you were needed to test this firsthand, what itinerary would you choose?
[Long pause.] I’ve got the places in the world I’d like to go to. I think it would be cool to go on a trip that goes to every continent. I’ve wanted to go to New Zealand. I studied Japanese in high school, but I’ve never been to Japan. I’ve never been to Greece – Greek beaches always seemed beautiful to me. And somewhere with hiking. Maybe in the Andes? You see how I’m stringing together all of these different places. I don’t know if any airline flies from Athens to Lima, but our system will tell me, “Oh, that’s not going to work, but if you add Bogotá that’s great.” Why not? Why not Colombia too?