How does A.I. select optimal prompts and call-to-actions at the best point in the conversation?

(3 min read)

For comparison, consider when typing a message on your smartphone.  If you type the letter “r” the autocomplete may select popular words such as “read” or “reply” because based on your previous interactions (and others who share their data with Google) the algorithm knows there is a good chance you will want to use one of these words.

Another comparison is Facebook multiple text ads which shows 125 combinations of headlines, ad text, and descriptions – testing out different combinations of ad copy based on what people are more likely to respond to.

Taking these two concepts together is similar to how RallySeller uses Q-learning to guess which prompt should come next, and what dialog patterns (or order of prompts) should be displayed to whom.  This is unique in that cumulative results are being considered as opposed to just getting clicks on the first message (at which point people may become unresponsive).

“Cumulative results” here refers to the number of taps on buttons, the length of conversations, and the number of taps on call-to-actions – additionally the bot is “penalized” with negative scores when it does not get engagement.

The “points” obtained and probabilities of getting a certain result are stored for later use with the next customer.

About Q-learning

Q-learning works with the concept of backpropagation meaning current actions (like the 2nd message sent) affect past actions as well (in this case the bot would determine which message is best for the 1st spot not only on how well it does but also on how well it does later on).

Another concept applied is the learning rate which doesn't just mow over old data when a change is detected.  In other words, you may have a successful sequence running that hits a few rough spots when people aren't responding, but these couple of instances aren't going to discredit all the previously learned data because the bot takes its time when gaining these insights.