In almost every webinar I’ve attended since last year, and even in a couple of presentations at the recent TechnoHotel Forum in Barcelona, there was always someone bringing up the argument of artificial intelligence in the hotel context. They all used the example of Netflix’s recommendation system as a clear and “applicable” example (how they know, I have no idea) in hotel technology. I’m fed up with snake oil salesmen who, without even understanding the technical foundations of what they so freely call “AI,” try to sell automations or logical business rule systems as if they were futuristic robots that instantly increase occupancy and RevPar.
With that premise, I’ll try to shed some light on what artificial intelligence is and isn’t, as well as its potential practical use cases in the industry.
As an appetizer, let’s start with a definition
More than three years ago, in an article similar in tone and content to this one, I wrote the following paragraph:
“Despite being a science fiction fan, my background as a (failed) engineer denies me the luxury of seeing AI (or any other technology) through a romantic lens. The practical approach is all that matters in business, and we are far (in the order of parsecs) from seeing a computer’s brain deceive us and make us believe that we are actually talking to another person (passing the Turing test, in short). There will be no Skynet or Voight-Kampff machine in the short term, my friends. However, I am surprised to discover how far behind we are in the application of AI in the entire tourism industry.”
(here’s the original >>)
I’m still surprised, although we have made some progress. Let’s decipher what this blessed “AI” is. Similar to human intelligence, which is no longer associated solely with logical abilities, in real life, AI is not a single element that drives the minds of robots like in the movies. The term “AI” encompasses various aspects and techniques such as neural networks, machine learning, deep learning, etc. Each of them has specific uses, and so far it is (almost) impossible to combine two or more in a single application, at the same time. In fact, the different methods of current AI can focus on only one task at a time. Among these methods, I would like to highlight three that are having (or will have) a profound impact on the industry’s economic results.
One is machine learning, which includes algorithms and programming methods that enable computers to detect patterns in data, predict outcomes… and learn from it to “improve their aim”. Another aspect is deep learning, a subset of machine learning, which powers recommendation systems like the famous Netflix. Simplifying it greatly, deep learning approaches data modeling in a “different, better” way than machine learning. Inspired by the structure of the human brain, it can perform much more complex tasks, such as governing driverless cars or facial recognition.
Finally, the method that is most used in the industry (for now) is natural language processing (NLP). It is the technique behind chatbots, allowing them to understand a written question in “natural language”—that is, similar to how humans communicate verbally—and eventually respond in the same manner. Hello ChatGPT!
Now, for any of these techniques to yield reliable results, it is essential to feed their algorithms with enormous amounts of data. In other words, big data tools are involved to manage massive databases with millions (or preferably billions) of records. This is a key ingredient because, without massive volumes of data, we cannot talk about artificial intelligence. Why? Because at the root of all this sci-fi fuss, there are nothing more than statistical techniques.
No Tits, No Algorithm
Let’s now see why it’s absurd or unethical (or just plain wrong, if there’s no ill intent) to mention “AI” when it’s actually something else (cheaper, providing much less value). For that, let’s use a couple of clichéd jokes from a high school teacher:
- If I have 3 chickens, Fran has 7 chickens, and Gustavo has 290 chickens… do we each have an average of 100 chickens?
- If a bureaucrat is sent to a street in El Raval to collect data on the religion professed by passersby, could the simpleton conclude that the city of Barcelona is predominantly Muslim?
As silly as they may seem, these examples point to a couple of basic statistical concepts that can ruin any algorithm: outliers (extreme or atypical values) and significant samples. With these elements, we can easily dismantle the arguments of the con artists. Let’s start with the popular tourist Netflix. For Netflix to recommend a movie or series I might like, the algorithm only needs a few choices from me to get an idea of what I like and suggest, within a limited time frame, a range of likely accurate options. But any user of the streaming service (as well as Spotify, with a similar recommendation system) will notice two problems as time goes on. First, recommendations will repeat, more and more often. Naturally, because no matter how large it is, the catalog of titles in a particular film genre is finite! In other words, the population is small, and the sample can’t be any smaller. The algorithm no longer has new samples to learn from and improve its results. On the contrary, it degrades over time. Second, sooner or later, someone (or myself, if I feel like it) will choose a movie on my account that, for whatever reason, has absolutely no relation to what I had been selecting so far… Hello, outlier!
In addition to turning the existing sample upside down, what used to be a compilation of androids, spaceships, and aliens will be contaminated with some rom-com or tearjerker. Goodbye, cool algorithm! At this point, I assume the reader has noticed where this is going: there’s nothing to automatically recommend if the choice menu is very limited, like the packages of a tour operator or even more so, the destinations and room types of a hotel chain.
There are academic studies that theorize the possibility of recommendation systems for itineraries/destinations: here’s one >>, this is another >>. And here’s one more >> (very interesting). But what’s the conclusion of these scholarly scientific works? The robot doesn’t work. In my opinion, something like that will be very useful when we can access hundreds of thousands or millions of destinations, located on hundreds of planets, distributed across several dozen galaxies. But for now, anyone who boasts about automated personalized selling of tourism products is a huge con artist.
Content personalization, yes, of course, any hotel chain can highlight the type of room or destination that interests me on their website, but that doesn’t require AI: a new invention called a cookie is just enough. The other area where I see too many con artists is in revenue management tools. Watch out for those who proudly announce that they automate price recommendations or predict demand by the minute… Maybe they could do it until 2019, but I highly doubt they’re currently in a position to repeat the feat if all they used as fuel for their forecasts was simply historical data.
Those who claim to combine too many data sources in their predictions are also not very reliable: the greater the complexity, the lower the accuracy of the algorithm’s results. I’m not going to waste time on those who lump together the terms “AI” and “marketing” in the same sales pitch. If I were a hotel decision-maker, I would cross them off the agenda.
Real AI in Production
Amidst all the sloppy marketing, I find certain initiatives commendable that – at least from what I know – actually put authentic AI algorithms to work. Among the chatbots, I can point out HiJiffy >>, which mercilessly defeated my BIFLIX automated RMS solution in the final of the 2020 Palladium/OnlyYOU startup competition >>. The good thing about this solution is that it has more long-term potential for improvement than many of its competitors because, unlike them, it uses a “collaborative” system (it also learns from community datasets, not just the client’s). Another product that I find interesting is the one proposed by my professor Brendan May, HERA by Hotel ResBot >>. With this tool, email responses are automated quickly and effectively, something that the front desk and the bottom line will truly appreciate.
Automated pricing is not for beginners, nor for those who want to leave everything in the hands of a robot, but in this field, it’s worth mentioning two prominent players in the industry: Lybra >>, which does what I unsuccessfully attempted in 2016 (combining more than two datasets in predictive ML algorithms), and Atomize >>, no idea how they do it, but they do it well. Where I don’t see big marketing investments is in the field of demand prediction. Is it still so difficult to obtain reliable results? Of course, with the pandemic and its consequences, it’s probably more challenging now… but technically, it’s within the reach of current data scientists, so either I missed something, or there’s a great opportunity for development in the market.
I also don’t see specific sentiment analysis tools >>, but in this case, it’s a cost problem: in fact, it’s only accessible to government entities, airlines, and large chains.
Conclusion
While we have made progress in implementing artificial intelligence techniques in the industry, we are still far from being able to argue that we benefit from it. If we’re still brandishing Excel as an indispensable management weapon while spending 10 minutes to record the check-in of a room with three guests, instead of aiming for the “artificial,” we should consider what kind of intelligence we are employing in our day-to-day operations.
If there’s one thing hoteliers should take away from this all, it’s this: let’s first implement analysis (business intelligence), then move on to advanced analysis (including revenue management), and from there, maybe we can turn to AI for real monetary benefit.
Thanks for reading!
Marcello Bresin