It’s been nearly a year since the travel industry fell hard for ChatGPT and generative artificial intelligence.
By late February last year, three months after OpenAI launched the natural language processing tool, Trip.com created a chatbot in its app built on OpenAI's API and a month later Expedia Group and Kayak became the first travel companies to create plugins to integrate with ChatGPT. Very quickly, it seemed like everyone was marketing a new tool developed with the technology.
To learn more about how travel companies’ use of the technology has changed over the year, PhocusWire is reaching out to some that were early adopters in hopes that their lessons would prove useful to others. We begin with some of the startups that began using generative AI in the early days.
For some, it was a matter of honing what they had created to get more from their tools, both in terms of productivity and depth of insight. For others, it was about learning what works best — and what doesn’t — and focusing efforts on the former.
And in at least one case, it was about learning more than you might want about some travelers when they feel free to ask anything of an entity that can’t judge them. For them, it might be worth remembering: What's shared with a chatbot, doesn’t always stay with the chatbot.
In the coming weeks, check back for more stories in PhocusWire's series about travel's early adopters in generative AI and the lessons they share.
Magpie: Diving deep into analytics with GenAI
It was last February when Magpie, a content management system for tour and activity providers, released a tool built with ChatGPT’s application programming interface (API) designed to help tour and activity suppliers create marketing content optimized for online searches.
It was a hit with the company’s clients, who weren’t professional writers and sometimes struggled with effective product descriptions, said Magpie founder and CEO Christian Watts.
Subscribe to our newsletter below
“We haven’t stopped. It [the past year] has been about just tweaking that first product,” Watts said, rather modestly given that the “tweaks” include adding translations for more than 80 languages, along with a level of data analysis that would have been impractical when humans were responsible for poring over the material.
“We’re really focused [now] on reviews. I just feel like reviews are a good area for this stuff. There's so much data in reviews,” he said. “Hopefully, you have someone who sits and reads those reviews, but ... sometimes the message from so many reviews just gets lost because we're humans. Now you can take the last thousand reviews or the reviews over the last three months and you can do summaries and really find out some of the problems that are happening.”
He offered an example of a tour company bus driver named Joe who consistently gets bad marks. If one customer complains directly to the tour company, another on Tripadvisor, a third on Google, the problem could go unrecognized for a long time.
“You don’t find out until three months later that Joe the driver shouldn’t be driving a bus for whatever reason because you don’t triangulate [the different review sources],” Watts said. “But now you can just dump them all in [Magpie’s tool] and pull out trends like that.”
Looking forward, Watts expects to build on insights by going deeper into the analytics. The platform already has the clients' product descriptions, FAQs, listings on various websites and reviews from all sources. Binding that information together creates opportunities for good analytics.
“Not for the sake of it, but with insights,” Watts said. “So we can, say, find some things that maybe we're not mentioning in the product description. Maybe we give everyone a free ice cream at the end, and that's the number one thing everyone's talking about in the reviews. So let's mention free ice cream at the start of description. That may be a silly example, but it’s finding things like that with your insights that can make a big difference.”
Turneo: Using GenAI where it’s strongest
For Matija Marijan, the CEO and co-founder at Turneo, a year of using generative AI has proved a “huge learning curve.”
Turneo, one of PhocusWire’s Hot 25 Travel Startups for 2024, is an e-commerce platform for hotels and other travel brands that want to offer experiences to their guests. One of its first experiments with generative AI was using ChatGPT to create a travel chatbot to act as a virtual concierge, offering hotel guests bookable recommendations of local experiences.
“In the early days, we and our clients were all blown away with the first results from generative AI products,” Marijan said. “But once the novelty factor wore off, it boiled down to a simple question: Is this AI-feature solving a problem I have, and doing it in a way that is overall better?”
The answer wasn’t always yes. The company discontinued its chatbot even as it added products designed to simplify life for experience organizers and resellers so that “tedious tasks they previously had to deal with are now seamlessly handled by AI.” The impact on productivity has been great, Marijan said.
“What we learned, as we added it to our products and built new ones on top of it, is that certain things, such as creating itineraries with precise scheduling, are really hard for GPT,” he said. “We've since evolved these products to use GPT where it's strong — [such as] writing text — but rely on other forms of AI or human experts on things which generative AI is not good at.”
Now, for example, Turneo encourages hotel concierges to put together itineraries for guests. AI speeds up the writing and offers recommendations, but it’s a human author who delivers the hyper-personalized service.
“We remain huge believers in GenAI,” Marijan said. “Having been using it for a while, we're now much more clever in where we deploy it so that it delivers best results.”
D3x (formerly Akin): Never stop learning
D3x co-founder and CEO Jason Noronha and his team were playing around with ChatGPT more than a year ago when they realized they could use it to convert guest messages into API calls.
“That was mind-blowing — like, never before was that possible,” he recalled.
Noronha spoke from experience. Having previously co-founded India’s first backpacker hostel and then a small chain of hostels based in his native country, he had to develop a remote management system when he moved to New York to study at Columbia University. Guest communications proved one of the biggest challenges.
His new company — then called Akin, since rebranded to D3x — used ChatGPT-4 to create a personalized, multilingual AI concierge that could respond to emails and reviews from customers of the company’s hospitality clients.
Seeing what we could do with ChatGPT, we thought, ‘Well, that changes everything.’
Jason Noronha - D3x
“Seeing what we could do with ChatGPT, we thought, ‘Well, that changes everything.’ That changes all of the previous systems we always built with workflow.”
The specific path used to process data through a workflow request could get stuck in the eddies of human conversation. Before, if someone messaged a request for airport transportation, then asked about weather or toothpaste or extending another night, the system could get stuck on the ride to the airport.
ChatGPT’s natural language processing grasped what the guest wanted and accessed the right information. Better yet, with about 60% of queries being the simple sort that can be answered by a chatbot, the system helped free up a property’s employees to provide the human touch when needed.
So what’s different after a year of working with the system?
“We have numbers,” Noronha said. “It was an idea and a vision we were selling, but now we’ve actually installed it in probably 200 properties, and we’ve had some very large clients. So we’re able to tell you the efficacy of the software.”
The software keeps getting better too. By studying how human operators edit the generated AI response, they can train the models to give better responses. “We’re getting to the 90s now in terms of an overlap percentage, so that’s a good validation to say the response is useful.”
Twelve months in, not all the lessons from working with generative AI have been about the quality of the software. Some deal more with the human condition.
One odd thing they’ve noticed is how people feel freer about what they say once they realize they’re speaking to a chatbot, asking questions they might be embarrassed to ask a human. In one instance, it became clear in the transcript that the person was arriving early but didn’t want to pay to book an extra night. On realizing the conversation was with a chatbot, the questioner pressed for details. What kind of seating was in reception? Would anyone mind if he slept in the lobby?
Eventually, the AI told the questioner he might consider visiting the movie theater on the property and taking a snooze there.
“That was really strange, just watching that interaction,” Noronha said.
It gets better.
Finally, the traveler asked if the chatbot could share photos of the seating in the movie theater. The system didn’t have the photos — at least not at that time.
And that was a new lesson. Soon they would be adding photos of everything they could to the database — not so guests could find a comfortable place to skip out on paying for an extra night, but in anticipation of new questions that might arise.