Predicting Long-Term Payment Risks in Energy Customers Using Large Language Models and BERTopic on Initial Conversational Data

One of the critical challenges faced by service-oriented companies is the inability to reliably predict whether a potential customer will be a "low risk" or "high risk" client. This uncertainty can lead to inefficient allocation of resources, financial losses, and wasted effort, particularly when signing contracts with customers who may later fail to meet their payment obligations. While it is unrealistic to achieve a perfect prediction of such cases, adopting a probabilistic approach offers a promising pathway to mitigate risks and optimize decision-making. \\ This thesis addresses this problem through a case study utilizing a dataset provided by an energy-providing company. The dataset contains initial conversations between customers and company operators, offering valuable, albeit unstructured, insights into potential customer behavior. By leveraging advanced AI methodologies, including Large Language Models (LLMs) and BERTopic for topic modeling and conversational analysis, we aim to develop a predictive framework capable of identifying clients at higher risk of delayed or failed payments. Although the specific focus of this study is on the energy sector, the proposed methodology has broad applicability across various industries. By formulating a robust and scalable solution, we aim to enable service providers to make data-driven, probabilistic evaluations of customer profiles. The ultimate goal is to empower operators with actionable insights, enhancing their ability to predict and address potential customer risks effectively, thereby optimizing resource allocation and improving overall profitability.