Business process management systems increasingly demand intelligent decision-making capabilities to optimize operational performance under uncertainty and in real time. Traditional approaches to decision-making in these systems rely on static rules or optimization models, which often lack flexibility and scalability. This thesis addresses these limitations by introducing a unified framework for automated decision-making in business process management systems, leveraging Deep Reinforcement Learning (DRL) and process-centric modeling techniques. The first contribution of this work is the Action-Evolution Petri Net (A-E PN) formalism, which integrates decision-making logic into business process models. This formalism enables the representation of dynamic assignment problems in a standardized and executable way, allowing DRL algorithms to be trained without additional implementation effort. The second contribution is a graph-based feature representation method, which overcomes the limitations of traditional vector-based representations in handling large or unbounded state spaces. This approach uses Graph Neural Networks (GNNs) to learn scalable and generalizable policies for complex decision-making problems. The proposed framework is implemented in GymPN, a Python library that supports modeling, training, and evaluation of DRL-based decision-making policies. GymPN extends the A-E PN formalism to handle partial observability and multiple decision points, making it suitable for modeling and solving complex optimization problems. Finally, the thesis explores the application of DRL to large-scale, stochastic environments through a specialized decision support system for a variant of the Dynamic Task Assignment Problem (DTAP) that is common in process management systems. Experimental results show that the proposed methods match or exceed the performance of the best alternative in multiple simulations parameterized on real-world event logs. Overall, this thesis advances the state of the art in automated decision-making for business process management systems by providing a unified and executable modeling framework for decision-making problems, scalable and generalizable DRL-based solving methods, and concrete tools for integrating AI-driven decision-making into process-aware information systems.