Artificial intelligence (AI)–enabled medication decision support can improve the identification and prioritization of prescribing risks, yet recommendations may be used selectively, ignored, or routinely overridden. These responses have different clinical meanings that cannot be determined from acceptance rates alone. This realist review examines how workflow conditions, professional trust, and organizational responses interact to shape medication decision-support use. Peer-reviewed literature published between 2018 and 2026 was identified through targeted scholarly searching and mapped against an initial programme theory. Context–mechanism–outcome reasoning was used to distinguish system characteristics, clinician interpretation, and observable actions. The evidence indicates that algorithmic performance, perceived relevance, interruption burden, professional accountability, and institutional implementation practices are distinct considerations in understanding medication-support behavior. It also identifies important differences between medication-alert studies, simulated human–AI decisions, and real-world implementation evaluations. However, complete record-level screening denominators and included-evidence-unit linkage were not recoverable from the search record, limiting final selection-flow reporting and confidence in a comprehensive synthesis. The explanatory account therefore remains conditional: useful uptake requires a combination of clinically relevant advice, workable opportunities for interpretation, and organizational arrangements that preserve professional judgment. Override and non-use require contextual evaluation because they may represent appropriate clinical reasoning, adaptation to workflow constraints, or failure of implementation.