In this series, we’re discussing ways in which artificial intelligence (AI) can transform marketing automation, allowing us to remove the guesswork, deliver a better customer experience, and yield improved ROI.
As we outlined previously, a core challenge arises in the definition of a marketing automation campaign. Today, a marketer must define a target population along with the message or sequence of messages to send. At Motiva.AI, our long-term goal is to replace this guesswork with machine learning that uncovers the best mappings between messaging experiences and customers over time, removing the need to manually define the relevant population. This would allow marketers to focus their energy on crafting compelling content and learning what most resonates with their customers.
In our last post, we described a first step toward this goal that discovers the most compelling message option to send to a given population. Motiva.AI applies a novel multi-armed bandit learning approach to achieve this end through adaptive experimentation over time. In that scenario, we assumed no knowledge about the individual customers.
In this post, we want to push the idea further. What if we could use all available customer data?
At a high level, leveraging customer data in a deeper way creates the opportunity to uncover meaningful populations that exhibit shared content preferences. This opens up new ways to understand customers, both individually and in larger segments, as well as tailor personalized messaging experiences.
So how do we accomplish that with machine learning?
In essence, the answer involves learning what relationships exist, if any, between customer attribute and behavior data and their responses to the available message options in an ongoing campaign.
Let’s take a specific example.
Imagine we have customer data for job role and industry. Let’s say we also have online behavior data that highlights their digital pathways through a website, including webpage dwell times and whitepaper downloads.
An adaptive marketing campaign would send regular batches of messages to customers and listen for responses as we described before. The primary difference now is that we’ll use the responses to learn a model that predicts a customer’s message preference based on the available attribute and behavior data. This might uncover significant relationships for example between the customer’s role, industry, and prior product preferences and the current message options promoting similar products. As those relationships grow stronger over the course of the ongoing campaign, the model would yield increasingly beneficial predictions for customers bearing similar attributes.
In machine learning terms, the change in the learning objective represents a shift from multi-armed bandit to contextual bandit learning. Here we’re bringing all available context into the learning task to support true customer-level, adaptive personalization of the experience.
By learning models that predict the likelihood of message engagement based on customer attribute and behavior data, we are automatically learning definitions of the underlying populations that are most likely interested in the associated message content. Coupled with an evolving model of message content similarity, the game changes. The possibility of continuous learning across campaigns without human intervention is within reach.
This is a great example of how bringing machine intelligence to a human team can create significant impact. It allows marketers the option over time to let AI adapt to emergent customer preferences dynamically — without manually having to guess at what will work with broadly defined populations. Customers get the highly tailored experience that they expect and respond to, which in turn helps increase marketing impact and ROI. In our next post, we’ll push these ideas even further into testing a more complex range of messaging experiences that could never be accomplished by humans alone.