Optimizing Customer-Agent Interactions with Natural Language Processing and Machine Learning

Efficient and successful customer service is an integral aspect of all businesses. In 2017, U.S. businesses lost $75 billion through poor customer service, where customers encountered unhelpful staff or spent too much time on unresolved issues. Customer experience management software companies analyze call center customer-agent transcriptions using methods such as sentiment analysis and topic modeling to improve their clients' customer service. However, these approaches are not optimized to account for the sequential nature of these customer-agent interactions. For example, while it is important to know how many customers cancel a service, businesses also need to understand how agents respond to a cancellation request and how certain actions may lead to a positive or negative outcome. To analyze the progression of conversations and understand actions that maximize positive outcomes, our research represents each contact center dialogue as a Markov decision process (MDP). For each conversation, we annotated whether the problem was resolved and whether the outcome was good or bad from a business perspective. We employed natural language processing (NLP) to extract the customer states and agent actions from call transcriptions. Our results identify and visualize the most frequent transcription sequences from successful conversations and estimate the expected probability of an outcome when an agent takes an action given a certain customer state. Such an approach may be used to develop programs to train agents for improved customer service in call centers.