Action arbitration with AI

U+ Bank wants to optimize the cross-selling of their credit cards on the web by using Pega Customer Decision Hub to show a credit card offer in a web banner when a customer logs in to their account. Customer Decision Hub uses the Predict Web Propensity prediction to arbitrate between the offers for which a customer is eligible. This demo explores how AI-based arbitration works and the advanced settings of the Predict Web Propensity prediction.

In Customer Decision Hub, the Next-Best-Action Designer contains the arbitration settings.

Arbitration aims to balance customer relevance with business priorities to decide which offer to show to the customer. To achieve this balance, the system multiplies the numerical values that represent propensity, context weighting, business value, and business levers to arrive at a prioritization value, which determines the top actions. Propensity is the predicted likelihood that a customer shows the target behavior, in this case, clicking a web banner.

In Customer Profile Viewer, you can examine the next best actions for the customer Troy.

For the current use case, the direction is Inbound, and the channel is the Web. TopOffers is the real-time container service that manages communication between Customer Decision Hub and the website of the bank.

When you request a decision for Troy, the Customer Profile Viewer shows you the offers for which Troy is eligible. Based on the engagement policy rules, Troy is eligible for two credit card offers: the Rewards Card and the Standard Card.

Initially, the model evidence is zero because the system has not yet captured any responses.

With zero evidence, the original model propensity is 0.5, or the flip of a coin. The final propensity that the system uses in the prioritization formula deviates from the original model propensity because it depends not only on the original model propensity but also on a mechanism that introduces noise while the evidence is low. The noise decreases while the model learns from the target and alternative responses, and the original model propensity and the final propensity converge. This mechanism assures that new actions receive exposure even when their models are still immature.

U+Bank uses the Predict Web Propensity prediction that comes with Customer Decision Hub out of the box to calculate the propensities. The Predict Web Propensity prediction calculates the final propensities for each combination of action and treatment in the inbound web channel.

A control group field determines a small percentage of customers that receive a random offer. During the production phase of the project, you can determine the impact of AI on the business by comparing the success rate of the offers that is based on AI, and the control group offers.

The target response has a Clicked label by default. For the alternative response, the label is NoResponse.

The Response timeout setting determines how long the system waits for a response from the customer after the impression. In a web scenario, the response timeout is 30 minutes by default, but an outbound channel requires a response timeout of several days to provide customers with enough time to respond to the message.

The Web Click Through Rate adaptive model configuration drives the prediction and supports decisions on both the customer and the account level.

For each credit card offer and treatment that the customer is eligible for, an adaptive model based on the Web Click Through Rate configuration calculates the likelihood that the customer clicks the banner.

Every hour, Adaptive Decision Manager updates the models to learn from the recent interactions.

The outcome of both decision requests for Troy is NoResponse.

When customer Troy logs into the website, Customer Decision Hub displays the Rewards Card offer, and Adaptive Decision Manager records an impression.

If Troy ignores the banner, Adaptive Decision Manager records NoResponse as the outcome after the response timeout elapses, in this case, 30 minutes. After the next update, the model calculates a lower propensity for Troy, as the new evidence that Troy is not interested in the offer weighs in.

Adaptive Decision Manager records a target response if Troy is interested in the Standard Card offer and clicks on the banner. After an update of the model, the propensity goes up.

You have reached the end of this video. What did it show you?

• How arbitration works.
• How to request a decision for a customer in Customer profile viewer.
• How to explore the original model propensity and model evidence.