In our third and final installment examining how artificial intelligence (AI) can transform marketing automation and audience engagement, we’ll discuss the power that machine learning offers marketers in driving highly adaptive, personalized messaging experiences when there are hundreds or thousands of different options.
In our previous posts, we’ve been discussing ways to test and optimize message variations, both in the context-free scenario where we have no customer data and the context-rich scenario where we have lots of customer attribute and behavior data.
Now we want to turn our attention to scenarios where we have a set of messaging experience options so expansive that it’s impossible to explicitly test all the possibilities. Think of this as a form of figuring out the Next Best Action to take for a given audience.
To make this concrete, let’s consider a lead-generation campaign that consists of a series of email messages where each message presents a small number of products in a particular order. Here we’ll assume we have no past history for these new potential customers. The goal is to learn the most effective message sequence that leads to the most significant number of customer conversions on average for new contacts.
So why is this hard? Consider the volume of options.
Let’s say that the maximum number of emails we can send to a given contact is three, and that in each email sent, we present three products in a particular order. This means that in any email message series, nine products are presented in total in a specified order.
(Above: One possible product ordering of nine products from three different product categories)
How many possible orderings are there if there are nine products in total available to present? Lots! There are 9 factorial or 362,880 options. We clearly can’t test them all!
So what do we do?
The trick is to exploit the fact that the performance of these product orderings are related.
If two product orderings differ only slightly in a single email in the series, we should expect that their performance should be similar on average. If we add data about product categories as well, we have additional similarity structure that we can use in our models.
Using machine learning, our goal would be to intelligently sample the space of possible product orderings by modeling the correlation in performance between the available options. Here we are pulling ourselves out of a tricky situation — one that would otherwise be impossible to address if left to humans alone!
With rich models able to capture the underlying dependencies in complex messaging experiences, the possibilities for optimization and automation expand dramatically. Marketers can ask deeper questions about experience design and ultimately deliver more value to current and future customers.
Examples like these motivate our thinking at Motiva AI around the need for humans and machines to work as a team, with each bringing their unique advantages to the task at hand. We’re looking forward to delivering these next generation capabilities to our customers so that they may serve their customers in whole new ways.