Abstract
Artificial intelligence (AI) and operational research (OR) are increasingly combined in decision-making systems, but the literature remains fragmented across communities that classify contributions by learning technique, problem class, or solver component. This review proposes a functional taxonomy that classifies AI–OR integration according to the role that AI plays relative to the optimization process. Using an integrative review design with iterative taxonomy development, we identify five modes: predict-then-optimize, decision-focused learning, AI-enhanced optimization, learning-based optimization, and AI-augmented OR modeling and decision support. Each mode is defined formally in a common notation that specifies what is learned, which training signal is used, and which component retains decision authority at run time. These definitions are translated into an operational classification procedure, including a substitution test that separates learned guidance of OR algorithms from learned solvers, and are applied to a set of boundary cases. A cross-mode analysis compares guarantees, training signals, computational burden, and human involvement, and maps how the modes interact in composite systems. The analysis indicates that many open problems arise at the interfaces between modes, notably feasibility under uncertain constraints, error propagation in composite pipelines, semantic verification of generated models, and end-to-end evaluation. The framework provides a common vocabulary for designing and assessing hybrid AI–OR systems.