Next Best Action (NBA) systems help businesses decide which customer intervention to apply, such as a discount, promotion, or lifecycle action. A new arXiv paper describes the core challenges: these systems must maximize long-term economic value, satisfy complex operational constraints, and account for causal relationships in the data.
The paper frames this as a causal constrained offline reinforcement learning problem. Offline RL allows the system to learn from existing historical data, while the causal constraint ensures that the learned policy reflects true cause-and-effect rather than mere correlation. This is particularly important in business settings where interventions can have downstream effects.
By combining these elements, the proposed approach aims to provide a more principled way to build NBA systems. The abstract highlights the need for a unified treatment of value, constraints, and causality, suggesting a shift toward more robust and reliable action targeting in business applications.