The model aims to provide the recommendation for next best action for sales actions on Azure customers. Built with reinforcement learning framework, the model identifies the optimal policy by informing sellers on when to contact whom (customers) and what actions to take. The goal of this model is to accelerate the Azure adoption and consumption, through more efficient actions on selected customers in a timely fashion. For instance, recommending a new Azure service to the customers when the likelihood of adoption is high, or recommending a churn prevention call when the customer is at a high risk of churn. The model also searches for the more probable paths that nurture new Azure customers to the most deeply engaged customers, i.e., the path of adding a new service to the existing service combo.
※Reinforcement Learning Framework:
We define the pool of actions, states and rewards following the Azure business logics. There are six actions available to sales engagement. The customer states are defined by the usage level of each Azure services. The reward is defined as the potential change on normalizing unit of total Azure usages.
※Model Deployment Channel
We proposed two channels for model deployment. One can subscribe the email notification on their customers of interests, or can log-in to Azure Customer Portal to check the Next Best Action results.
We use RL to model this optimization problem for two reasons:
1.Actions on customers are not simply binary right (1) or wrong (2)
a.Not supervised learning, since we don’t have 1/0 label
b.Not unsupervised learning, since we do know the impact on customers
2.Actions on the same customer at different states may have different impact
a.There may not be one single best action on the customer throughout the customer life at Azure.
We have reached multiple teams that have relations with Field sellers.. We have identified various use cases with these teams and their sellers. The model will bring 5-10% increase in Azure service adoption and service consumption, and 5-10% decrease in churn in estimation.