<p>While manual, expert-based approaches to debt collection have long dominated the industry, the growing complexity and volume of data—resulting in information overload—highlight the need for machine-learning solutions. This paper reports on the development of a Smart Debt Collection System (SDCS) through a Design Science Research (DSR) approach, involving iterative design, field testing with a major financial institution, and feedback from stakeholders. Our research addresses the limitations of existing debt collection software, which primarily focuses on data gathering, administration, and reporting, but lacks automated decision-making capabilities. Our work makes two significant contributions. On the practical side, it introduces a system that not only automates decision-making processes but also dynamically adapts to debtor behavior, enabling more effective debt recovery strategies. On the research side, it proposes a novel system architecture integrating three key components: factorization (Markov process), optimization (deep neural network), and process specification (domain-specific language). This architecture not only underpins the SDCS but also provides a reusable framework for automating complex decision-making in other domains. The innovation of SDCS lies in its ability to automate decisions previously reliant on human expertise. This shift opens up new perspectives and possibilities in the field of debt collection, such as increased efficiency, improved accuracy, and scalability. The application of AI algorithms in this context has the potential to revolutionize the industry and enhance overall debt collection practices. Debt collection agencies can adopt our system to automate their processes, while researchers may apply the underlying approach to broader decision-making challenges.</p>