Large companies receive thousands of email messages every day. Most of these messages come from their customers and must be forwarded to the right support agents, depending on the customer needs. This sorting procedure can be done manually, which is cost- and time-inefficient, or automatically using natural language processing (NLP) techniques. However, each wrong classification may result in a message being forwarded to the wrong agent, creating further costs. For this reason a reliable estimation of the classification confidence is required. This poster describes how Centrica combined Conformal Prediction and NLP to reliably classify the topic and sentiment of incoming email.