Background and objectives

Our client, one of the Big-6 energy retailers in the UK, was facing non-competitive costs-to-serve driven by high volumes of customer calls. After our client had tried to understand the drivers for the size and volatility of these calls, they recognised the need for more sophisticated machine learning / predictive analytics to understand the rationale for these calls and what action to take to reduce these. One of the objectives of their channel strategy was and is to increase the handling of customers on-line, but more insight was required to determine which customers would be most susceptible for on-line service and, equally important, for which type of interaction with the energy retailer. Because of our in-depth understanding of the energy retail market, our client gave us the mandate to conduct a proof of concept as we could “hit the ground running” without losing time to get to understand the underlying issues in the energy retail sector


Our approach

With our client we identified the best targets to develop a predictive model for are threefold: 1) which customers are most likely to call; 2) which customers do contribute most to the volatility of the customer contact volumes and; 3) which customers are most likely to generate repeat calls and call transfers. To develop the models for these predictions we used three categories of data: 1) customer specific data such as demographics, product / payment type, credit scores, etc.; 2) customer interaction data such as outbound calls, letters, emails, on-line clicking behaviour, meter reading submissions, billing data, payment data, balance data and; 3) baseline data on inbound customer contacts, i.e. calls and repeat calls. Together with a small client team we started identifying predictive signal in the data, also to determine which machine learning methodologies would be most suitable for the three types of predictions our client would be served with best.



In addition to the three predictive models we have developed, in determining if there is predictive power in the data provided, we identified several opportunities to improve the operations of our client: for example, the data showed a significant difference in number of repeat calls per team, which was a result of some team managers not managing the balance between average handling time and first contact resolution. To reduce the volume of repeat calls, which amounted to roughly 30% of the total number of calls, we developed software to predict which customer is most likely to generate repeat calls. With our software linked to our client’s CRM system, the contact agent sees a red flag when a customer with high repeat propensity calls and as a consequence the agent prioritises first contact resolution over average handling time. This agile application of contact centre targets does allow for a reduction in total number of inbound calls.