Banking on AI: When Classifiers Meet Conversations

In 1992, MIT Professor Patrick Henry Winston defined Artificial Intelligence as: “Artificial Intelligence is the study of the computations that make it possible to perceive, reason, and act.”

This definition has proven to be pretty durable so, let’s step through an AI-based marketing application using Winston's words as our guide. 

Step 1: Perceive

First, we consume a regularly updated data set comprising demographic and account data plus interview response data for a bank’s customers.  Using algorithms, or computations, we classify each customer as to their likelihood of buying a particular product – like an auto loan, for example.  The classifier makes a few mistakes, but does a lot better than guessing, as illustrated below.  

Figure 1 - Classifier performance example

Figure 1 - Classifier performance example

Out of 3,006 auto loan buyer predictions, the algorithm made 711 mistakes.  And out of 6,994 predictions that customers were not auto loan buyers, the classifier only got 579 wrong.   When a customer with a high auto loan propensity arrives on the digital banking channels, like online or mobile banking, the machine perceives this customer as an opportunity to talk about auto loans. 

Step 2: Reason

The machine reasons that only customers with a relatively high propensity for an auto loan should be asked about their possible interest in an auto loan; a conversation which may well turn into an auto loan refinance conversation,  depending upon the direction the conversation takes.  The machine further reasons that these customers should not be interviewed about products for which they have a low propensity score. 

Step 3: Action

Action happens when the right interviews are served, completed, response-based offers are made, leads are automatically routed to the right people on the front lines of the banking institution, and the responses are fed back into the system to help the machine learn from its mistakes.  For example, if some high propensity auto loan prospects answered, “Not interested" or "not in-market”, during the interview, the machine will now look for common characteristics of this group and modify the classification function to correct for these mistakes and improve the next set of predictions.  That's how the machine learns from interview responses.  

And that, ladies and gentlemen, is what happens when classifiers meet conversations: machines learn! If you’d like to learn more about how to put classifiers and conversations to work for you – get in touch with us, we’d love you show you how!