Call prediction models are an innovative tool designed to introduce new efficiencies into a servicer’s call center by using activity and signals on a customer’s loan to help predict the top three reasons why a customer may be calling. With that information, support representatives can tailor each interaction to the individual caller. This helps servicers have more meaningful conversations that lead to positive outcomes, creating a better customer experience and helping agents resolve more calls faster.
ICE’s call prediction models are built with a focus on accuracy, privacy, and explainability, so servicers can feel confident proprietary data is being safeguarded – both theirs and their customers’. Here are how ICE’s models support those three core pillars.
Accuracy: Driving customer satisfaction
A call prediction model is not productive if it cannot reliably determine why a customer might be reaching out to a call center. Currently, ICE’s models achieve an 85% accuracy rate – meaning we can say with 85% confidence that at least one of the three reasons presented reasons for inquiry is why a customer is calling. And since incorrect predictions can lead to negative customer experiences, maintaining high accuracy ratings is paramount.
Model accuracy is driven by continuous training. When statistically significant model drift occurs, the models can be refined via a feedback loop based on real-world usage. Whenever the customer calls for a reason other than one of the three presented by the model, that response can be fed back into the model as part of its training process. This iterative training helps the model become more accurate over time, improving its prediction capabilities.
This same feedback loop also helps models better serve each servicer’s individual use cases. Since different companies have different processes for handling customer service calls, each model will be unique to the company using it. This helps models not only correctly predict calls more often, but also support the specific business needs of any given organization.
Privacy: Protecting proprietary data
Accuracy is essential, but it must not come at the expense of privacy. That is why ICE places a high priority on confidentiality in its call prediction models.
First and foremost, each company’s data is used solely to train its own model – no one else’s. This strict siloing of information is designed to keep your sensitive information strictly confidential. In fact, ICE’s call prediction models are completely agnostic of identifying information. They do not track customer information, nor do they collect any specific demographic details.
The same is true for the feedback loop used to train models. Any information that is fed back into the system for model refinement is completely scrubbed of all identifying information before being incorporated into the model. This means that as the models continue to improve, they do so without compromising privacy.
Explainability: Transparency in predictions
Regulatory scrutiny in this sphere is extremely high, and companies need to be able to explain how and why a model made the predictions that it did.
Unlike some black-box models which can be difficult to interpret, ICE’s call prediction models are globally explainable. In the event of an audit, the system would be able to provide a clear, transparent breakdown of its processes to show that its conclusions were free of bias. That is one key reason among many that demographic information is stripped out of data that enters the training feedback loop.
Furthermore, these models do not make any decisions for users or customers. Instead, they simply suggest possible outcomes (the top 3 reasons for a call) based on certain inputs (activity and signals on a loan). Whatever conclusion the model arrives at does not preclude a customer or an agent from any action, nor does it make them more likely to take any future action. The model does not replace human decision-making. It simply facilitates a conversation.
Conclusion
Call prediction models offer a lift to servicers who want to increase customer satisfaction and help agents have more productive, meaningful calls. With an emphasis on accuracy, privacy and explainability, these models provide a powerful, innovative tool to drive valuable interactions that help lead to positive outcomes for customers.