The Next Thing: ChatGPT and Generative AI

Since its launch just 12 months ago, large language model-based artificial intelligence interface ChatGPT brought generative artificial intelligence technology—and its varied possibilities—to mainstream markets worldwide. To some, this innovation sounded alarms of mass layoffs and technological takeover; to others, it signaled the promise of faster and more human-centered innovations, the likes of which the business travel industry and the wider world had never seen.

In a presentation at the BTN Group’s Innovate conference in early October, Navan co-founder and CTO Ilan Twigg told the audience he had become so obsessed with the possibilities of generative AI in the months following the ChatGPT release that his therapist had “fired him” as patient. Twigg has replaced his therapist, but his fascination with generative AI continues. He referenced it as “powerful” at points and even “magical.”

Not everyone at the Innovate conference was as enamored as Twigg with generative AI’s current capabilities, but many agreed that its possibilities—and the increasing speed at which they would be realized—would change the managed travel industry forever.

What’s “Generative,” Anyway?

Getting to an era of generative AI hasn’t been a short journey—and plenty of experts say we’re just at the beginning, including McIndoe Risk Management founder and CEO Bruce McIndoe, whose background includes a master’s degree in computer science from Johns Hopkins University as well as work on the fundamentals of AI with a number of government agencies prior to entering the risk management world.

In its simplest form, McIndoe said, ChatGPT and its AI foundations are “just math,” performed by systems that tokenize words (in English and other Western languages) or characters (in languages like Chinese), which machines can recognize, associate with commonly corresponding words and information and, when prompted to engage on a topic, will communicate back with humanized linguistic patterns. While it sounds simple, the scale at which machines can perform this work makes it seem “magical,” to use Twigg’s word, but it doesn’t work without humans. At least not yet, said McIndoe.

Humans train machines to weigh associations, bake in variables and return information in what seems like spontaneous combinations. In its current form, McIndoe said, ChatGPT is still reliant on human prompts, searching a vast universe of tokenized words on a network and returning responses based on the number and weight of associated vectors. Without continuous prompts, however, he said the technology as it functions today would “collapse on itself,” pulling information that it generated and reformulating it over and over.

“We’re literally just stepping off very simple machine learning using neural networks,” he said, referring to those technologies that give weight to strong and weak information associations.

Even so, industries have seen generative AI use the foundation of those inputs with multiple machine learning algorithms and a layer of creative synthesis to analyze texts, write novels, produce digital images, compose music and code more spontaneous action into video games.

There are serious concerns within systems, however, with issues of factuality, bias, quality, plagiarism and intellectual property. ChatGPT, Bard and OpenAI produce much of its content “in the style of” and clearly based on the inputs of human artists and content owned by companies or individual creators. Questions of intellectual property, privacy and data ownership in the face of generative AI are real for every industry. Yet, so are the possibilities.

Experts Envision the AI Future

McIndoe, joined by Unlock Advisors founder Cara Whitehill, American Express Global Business Travel EVP of product and strategy Evan Konwiser, Cornerstone Systems CEO and co-founder Mat Orrego and Grapevine founder and CEO Jack Dow, analyzed some of the go-forward propositions for AI and managed travel during the Innovate session. Travel Tech Consulting principal Norm Rose moderated. Edited highlights of that discussion follow.

Advanced Search Tool – “A lot of knowledge in this industry is kind of buried in processes,” said Orrego. “And I think it’s in its early stages of evolution.” Orrego envisions a time when systems can identify that knowledge embedded within processes, curate the need-to-know information and then present that to an agent or account manager in a way “where you can interact with it and use it in a way that is helpful and speaks to the way people understand each other,” and not just in the way systems understand one another. “Because we lost a lot of industry knowledge during the pandemic,” as agents and other workers left the industry, and, he said, he sees tools like these as a new way to educate a new generation of workers. But, he emphasized, they won’t be able to do that without training. “Its magic is really our ability [to program] a tool to achieve what we want it to … and it has to be done in a way that is customized to the situation.”

Data and Analysis – “One thing that AI tools could be really helpful with is aggregating a lot of data from disparate sources,” said Whitehill. “You have stuff that goes into Tableau like credit card data, booking tool data, travel management company data. And you think about custom fields like traveler feedback and stuff that’s both on- and off-platform—your whole leakage problem. It’s complicated to analyze all those things. The opportunity to take AI-based tools or layers onto the Tableau platform versus spending a lot of that time trying to deduplicate, correlate and understand the data. That’s the kind of thing AI is really good at doing really quickly.”

Personalization vs. Contextualization – “We’re building hypotheses around what impact AI can make in terms of improving personalization and increasing conversion rates [for hotel attachment] or increasing customer satisfaction. There may be one bit of data that just isn’t there that’s relevant to that traveler,” said Dow. That piece of data may be something that AI could infer and bring to the table, said Konwiser.

“I think it’s a lot more about contextualization than personalization,” Konwiser said. “What I mean by that is understanding that you’re coming to Europe to [attend] this event is actually more important than who the person is. Or, when someone is trying to get a service, the idea that the flight was just canceled and the weather is bad is important to you. Context has to be in real time, updated on the fly, what’s going on with the person right now. It’s not going to live on the profile. Context is more powerful, and I’d like to see us spend more time on it as an industry.”

Added McIndoe, “We shouldn’t be looking to B-to-C innovation to bring amazing personalization and AI to what we need. There’s some real separation that we need to address in the context of [business travelers] accomplishing a task.”

The Future Booking Interface – Spoken language and chat interfaces are “at a key stage right now,” said Orrego. “Personally, it’s not my preferred way of interacting from a UI perspective, from an experience perspective. It’s not so natural right now that it knows, for example, things that are occurring within the environment. That’s just a personal perspective.”

“But you know who does prefer it? Everybody under the age of 25,” said Whitehill, answering her own question. “Because they’ve grown up using Siri and a lot of stuff and they’re used to voice engagement. ... Right now, voice chat isn’t that good. But it’s going to get better quickly. And this is where the whole concept of large language models is going to have to come into play—understanding needs in real time, being able to translate that context and curating content in a way that’s relevant.”

Accountability – As these AI-based engines roll out, trust will be a major component, and it will have to be multifaceted. The end user traveler will need to trust the systems, but so will the company in terms of the content they serve and the recommendations they issue. “You’ve got to wonder who’s training that thing,” said Konwiser. “Is my travel manager telling me that? I’m sure he wants me to do that, but I don’t know if that’s what I want to do.”

But even bigger issues for organizations to consider are bias and accountability. Will TMC-based or other third-party tools bias content to hotels, for example, where they earn commissions? What algorithms are really in play? Sometimes those are hard to decipher outside the box of AI. And what about in travel risk situations, where a non-preferred airline could be recommended? “Who’s accountable when the AI engine is giving answers that are wrong?” asked Konwiser. “That raises the stakes for how engineers in a room somewhere are coding things that have real-life implications. In an enterprise environment, you’ll have to be really on top of that.”

What Buyers Should Do Now

Travel buyers looking to bring large language model innovation to their programs need first to remember that LLM works by constantly consuming data to create and refresh the neural network associations that drive the fundamentals.

“If it’s free, you know the product is you, right? So if you’re out there using these free services, you are the product,” said Konwiser. That may be fine on an individual level for buyers who want to play around with query composition and how that affects AI responses. When it comes to company data, however, travel buyers using OpenAI or other no-cost systems could fall afoul of company policies.

But that shouldn’t make buyers hesitant to learn and experiment. Both startups and legacy players are creating compliant frameworks for corporates to use. Just be sure to understand where the data is going and how they are going to train the systems, “because once you give something, say goodbye. It’s gone,” said Konwiser.

Whitehill agreed with the critical nature of compliancy and governance in the face of large language models and AI, but she also saw the industry taking it in stride. “Vendors are developing with their own versions of those [large language] models that are trained on your proprietary data. That’s a good version. So don’t be spooked that it’s all bad and scary or too risky to put out there. There are private applications that can be used, but always validate the frameworks.”  

- Angelique Platas contributed to this report.