Blog
61% of marketers say AI is already boosting the bottom line. Are you seeing the same?
Learn more
AI and machine learning seem to be everywhere these days, and marketing is no exception. In a recent webinar with Skift, Epsilon-Conversant and United Airlines, we uncovered how travel marketers in particular have overcome these technology challenges to make waves in the industry.
As a former director of digital technologies at Holland America, an international cruise line, Therron Hofsetz was a pioneer in bringing AI to the travel industry, testing and learning the then nascent technology over the past decade. Now a senior director of digital experience at Epsilon, Hofsetz sat down with us to discuss his interactions with AI technology and how machine learning is affecting today’s marketers.
Let’s dig into your background a bit. How did you start working with machine learning technologies?
My introduction to machine learning actually started when I was planning a vacation back in 2010. My wife and I were in different cities at the time, and she was having a very different experience than me when trying to book the same flight through United Airlines.
Ten months later, I was at a quarterly business review with Oracle. They presented the United Airlines case study, and I finally understood what had been happening with our vacation planning experience. Their use of the machine-learning platform gave us different experiences because I was a frequent traveler and she was not.
Based on that meeting, I brought machine learning to Holland America. We started using machine learning toward the end of 2011, and our primary implementation was to personalise and upsell short-excursion sales on our web channel.
What are some ways that you’ve seen machine learning used in your career after that experience?
Taking machine learning to online and offline contexts has been powerful. It’s important to think beyond the channel you may want to pursue initially. At Holland America, for example, we decided to expand how we use machine learning when we started working with Epsilon.
The first use case was on the web, then email. After that we moved into a non-digital channel—the cabin. We did a personalised content drop by running passenger data through the machines to predict excursions each passenger was most likely to take. Then a PDF file would go to the ship, and the crew would deliver the printed version to each stateroom.
Clearly machine learning can shake up the old ways of doing things. What changes need to happen internally to support this shift?
Cultural change is really required to support this technology. There are so many different permutations based on individual profiles, so you can’t hold onto the “I need to proof this as final” mentality. You need to trust that the machines can produce the right thing based on the information you give them, or you won’t be able to scale fast enough.
We faced a lot of business-operations challenges with our email work. The marketing team was used to laying out the perfect design in a manual proofing process, and it was hard to let that go.
Building a culture that embraces machine learning is just as important as the technology. You have to move all of your upstream processing, such as content creation and curation, to a place where there is a high degree of confidence in the execution. To start, you can do a number of dry runs in batch sets to make sure it’s processing the way you expect.
What factors can make these processes challenging? What should marketers keep in mind as they develop processes around machine learning?
Content velocity becomes critical in order to service all of the micro-segments that are derived by the machines and algorithms. For example, a brand may need to build metadata that supports traditional rules around color palettes and image associations, so a machine can decide on the optimal combination.
It’s also a self-fulfilling cycle. As you learn what content people are clicking on and what they’re buying, you get a feedback loop. Then you can inform your content creation process earlier on, like picking certain types of images or words. By staying agile, you keep working in those iterations to enhance the process over time.
What advice can you give based on your own challenges with machine learning? How did you overcome them?
The first is that you need to be careful about measurement KPIs because you can easily over-count the decisions of the machine learning technology and over-inflate success.
In the first implementation we used three metrics to score how well the machine learning process was working—a click, an add-to-cart and a purchase. The way we did the original calculations triple-counted all of those things, so if a customer purchased an item, we also got credit for the click and the add-to-cart. That was clearly over-counting since the click and add-to-cart are implied if someone purchased.
The second is that it’s critical to tag content with metadata. The computers are getting better at cognitive learning, such as image recognition and sentiment analysis, but unless you provide that metadata in a way that a machine can really understand it, it’s difficult for them to know what is really happening.
We did a poor job of tagging our content when we first implemented the excursions in our recommendations at Holland. For example, if we had a rainy day or a sunny day image, the machine couldn’t give us insights on which one is better to use. Better tagging can help determine what attributes make an image successful, when to use that image and in what context.
When it came to measuring KPIs and tagging metadata, we had to go back to the business to update their recommendations based on what the algorithm needed to really learn.
What are marketers struggling with today related to bringing machine learning into their marketing programs?
There is a lot of buzz around AI and machine learning right now. It’s hard for marketers to cut through the noise and think about the true capabilities that apply to them when it comes to things like process optimisation and engagement.
Very few solutions out there focus purely on the marketer, which makes it even harder. The big industry players don’t seem to focus themselves too much. If you look at Amazon, Google or even Microsoft, their machine learning solutions could be used everywhere, not just for marketing purposes. Their cloud-based machine learning capabilities don’t directly translate to the way marketers think about their business.
Anyone exposed to this from the outside, however, really wants to understand tangible benefits. Say I’m a brand marketer focused on a certain discipline within marketing. How does machine learning benefit me?
And those pieces are not what Silicon Valley is really concentrating on. They’re looking to use it in more complex ways and have left behind those looking to use it more simply. They are certainly focusing computing power on bigger problems, like gene sequencing, which is far more complex than optimising subject lines for an email campaign.
What’s on the horizon? That is, where do you see AI and machine learning for marketers headed in the next five years?
The content creation and curation process will change significantly. To the chagrin of many creative types, insights derived from machine learning will become a bigger part of the creative process. There is concern that machines are taking the human side out of that process, but I don’t think that’s true. It actually moves that human side further up the funnel into that content creation process.
We’ll also see more embedded capabilities within platforms and products. When we met with United recently, they said that they are still relying on vendors for support, but that will change. Agility Harmony, for instance, shows subject line recommendations, next best offers and content optimisations built directly into the platform.
Most companies are pretty early in their stage of adoption right now. In that three-to-five-year window, we’re going to see a much higher rate of adoption.
What advice would you give to marketers looking to get started with machine learning in their marketing programs?
Start as soon as you can. Single channel, single data set, single decision, is probably the easiest place to start.
It can be overwhelming to think about the amount of information that is out there and the cultural change that it might take. You see this notion of always-on marketing and hyper-personalisation, but that’s a pretty mature set of use cases. Most companies are just starting to reach customers in a more personalised way.
