Brandon Strauss of CapTrav and KesselRun Corporate Travel Solutions explains why he’s in “watch and learn” mode when it comes to generative artificial intelligence.
Claude Shannon, a computer scientist and founder of information theory, published “A Mathematical Theory of Communication” in 1948, which defined how messages are carried from sender to receiver. This theory became critical to understanding communication flow, the meaning and intention of words, and, in practice, how language must be assimilated for artificial intelligence applications to work. Among other things, the theory states that any information transmitted from sender to receiver can be mathematically sorted between empirical information and semantic content. In other words, there is a scientific distinction between the information exchanged and the meaning of that information. One of the conclusions drawn is that while the two can be viewed as exclusive, in practice, a message would not really carry any meaningful information without the shared semantics. To a large degree, Shannon illustrates the key hurdles that generative artificial intelligence must overcome to be successful for corporate travel applications.
Industry suppliers and new technology providers today offer next-level machine learning capabilities that some inaccurately label generative AI. Applications that pull data out of reservations or apply historical travel patterns to support new bookings seem to be the next and better iteration of common functions, such as online booking, chat, reporting platforms and mobile applications. These examples improve on pre-existing machine learning capabilities but don’t reach the level of traditional artificial intelligence — much less the humanistic, semantics-level decision-making ability that defines generative AI. When we hear travel technologists reference the ability to “contextualize” intent with their products, it dramatically oversimplifies the semantics required under Shannon, does not paint an accurate description of services available and is not close to market-ready.
The problem of semantics in travel gets even more complex as you dig deeper. Not only is there a requirement to marry information to semantics, but the language itself must match to succeed. For a travel application, the language concerned here can be thought of as text, chat, phone or email. An email picked up by an agent and answered with a phone call potentially violates the Shannon theory, which means neither the information nor the semantics of the message is accurately transferred. Understanding how language has intention and conveys meaning is a hard problem to solve for a computer. Yet, intention, meaning and raw data are all real-world aspects that must be considered. Have you ever heard the phrase, “a difference that makes a difference”? It came from the idea that language and communication are well-defined and measurable, even if we don’t always understand how meaning and intention fit into a particular application.
Successful generative AI applications will understand the semantics of both business and leisure travel for a particular travel policy and individual. It will consider personal aspects of a traveler’s life, including every variable that a traveler may consider when making a travel decision. Generative AI will assimilate predictive analytics with semantics to provide the traveler with granular, sensible choices. Effective generative AI will proactively consider the multitude of variables that a traveler would consider as they anxiously watch a gathering storm through a conference room window 1,000 miles from home, for example.
Will my flight be canceled? Can I keep my hotel one more night? Will I make it to Billy’s softball game? I have that dinner tomorrow that I’d need to cancel. I need to be in Baltimore in three days. Perhaps wait a day and then just go to Maryland? I have a conference call at 3:00 p.m. tomorrow. I can’t sit in a middle seat because of my back issues.
Imagine the power of harnessing all this information using predictive analytics along with the interpretation of personal data, such as calendars and historical travel data, to provide proactive, semantic-driven choices while the worried traveler gazes from their desk chair.
Whereas most spend categories within organizations can probably utilize generative AI with readily available data points and fewer semantics, I think business travel will be challenging. Corporate travel has always been a stick-and-carrot type of spend category, where the company takes away liberties with policy while acquiescing for dozens of reasons using exception codes. Travel policy, traveler choice, a myriad of travel options and data privacy all play a role. The recent European Union Artificial Intelligence Act, which will put substantial limitations on the use of AI in Europe, may act as a world model for laws surrounding AI applications. Will these applications even be legal in Europe? What will California’s Consumer Privacy Act say about all of this? Successful applications of generative AI must overcome enormous data challenges while requiring access to very personal information.
Given the industry buzz and excitement around generative AI, it’s easy to envision how travel applications can be transformative. Every sea change must start with a ripple before it becomes a wave. Until then, I’ll continue to watch and try to learn. I get it. I just don’t see it, and I’m skeptical.
This Op Ed was created in collaboration with The Company Dime‘s Editorial Board of travel managers.
Thank you for your latest article. Your insights on the mislabeling of basic tracking and prioritization logic as generative AI really resonated with me. It seems like the term is being thrown around a lot these days, often more for its buzzword value than for accurately representing what it’s actually doing. Great for sales, but we know the truth about what is really happening under the hood.
That said, I will share my take on the use of generative AI in corporate travel. Its flexibility in shaping outputs, tracking context, and storing relational information in graph databases is something we can’t ignore. The advancements in Large Language Models, fueled by massive data, powerful GPUs, and extensive memory available today, are opening up new avenues for exploration that were impossible to explore just a few short years ago. It feels like the ’90s Internet Boom again, with everyone in our field experimenting to see what works and what doesn’t.
For me, the key question is: Does the use of an LLM or generative/non-generative AI create new value for the customer? Anything we build and deliver using generative AI should aim to lower costs, boost efficiency, and bring a bit of delight to travelers. It should aim to make processes simpler, our teams more effective, our customers’ teams more effective, and critically, allow us to automate what was formerly impossible to automate.
We can ask ourselves a few exploratory questions:
Does an AI agent really provide material value during the booking experience, or is it a nice to have that is flashy during sales but rarely used in reality? Travelers, and the generation they come from, will be the ultimate voters on its utility. Personally, I think it is very promising in certain contexts.
Will an Analytics AI agent generate deep enough insights to automate policy optimizations autonomously, and reflect those in the OBT and Mid-office simultaneously without a human’s approval? It will depend upon how well it’s been trained and what guardrails have been placed on it to keep it from going off the tracks. Not so closed a system that it’s useless, but not so open that it can accidentally ruin your company.
Behind the scenes, LLMs are becoming a vital tool, even engaging in their own LLM-to-LLM conversations while integrating with corporate travel APIs to gather data in real-time. These interactions will open up new opportunities for delivering value, things we are just now seeing pop up sporadically across travel products. The corporate travel sector is really diving into this technology, trying out various ideas. While many might not work out, the few that do could set new standards and push the whole industry forward.
Regarding the legal aspects, it’s a serious topic and my take is that there are no good answers. The rise of extremely powerful open-source free LLMs has created a bit of a Wild West situation in AI. I have deep concerns that there is only so much governments can really do. Case in point: ChatGPT is the best publicly accessible LLM at this time, but for many product problems, “Good Enough” is more than sufficient. That is the danger Open Source LLMs present for high-profile LLMs like ChatGPT/Bard/Claude 2/etc. It’s not the current state of technology that is worrying… it’s what happens in a few years due to Moore’s Law. Anyone will be able to download a 500 billion point LLM model for free that can locally perform as quickly and accurately as ChatGPT 3.5 does today and do it on their home PC. How will that usage be monitored, and more critically, enforced? I don’t have a plausible answer to those questions, but hopefully someone will figure that out in a way that does not stop progress.
I see that the future is blindingly bright for generative AI.
Interesting article, Brandon. In hotel technology, AI and (perhaps better) fast machine processing are already widely used and deployed – everything from chatbots, virtual concierges, loyalty, marketing, personalisation, forecasting & revenue management.
The nomenclature, as you point out, creates misunderstanding about how to use and benefit from AI. It’s important to debunk some misunderstandings. Many folks imagine some kind of malevolent intent or evil intelligence, but Thanos isn’t real — he’s a character in a Marvel movie. There’s also widespread concern about the impact on job security and productivity advantage/reliability, which needs teasing out just as the conversation about automation has enhanced service capabilities, not dehumanised them.
That was an interesting article. While I can’t claim to talk the same lingo and claim to be at your level of knowledge on the subject, I can fall back on my years of experience in the industry. The industry, as a whole, has done a terrible job of doing anything for the traveler beyond simply an order form with the online booking tool. The OBT takes the order and the TMC processes it. You’re done. The $ at stake don’t give a whole lot of wiggle room for investing in fancy AI/analytics.
There’s nothing out there I have seen to assist the traveler with anything beyond simple fare searching (lowest logical fare), a little hotel mapping and that’s about it. Granted I don’t get out much. For years the booking tools have talked about doing this and that, but very little materialized. The industry deserves a big fat F for failing to deliver much in the way of useful analytics for the consumer when making a decision. There’s potential as Brandon stated but there’s skepticism. Right now, it’s noise to me if I’m being negative -or – untapped potential if I’m being positive. I work for an analytics firm and colleagues will tell me that travel data is interesting and fun. It’s a pity we haven’t done more with it!
I think some travelers/bookers and travel agents would appreciate some good intelligence to help them make more informed decisions – not just after the fact, but at the time of making that decision when on that “click to book” screen!