Artificial intelligence is increasingly used to identify medication-related problems, generate alerts, and support medication review, yet the clinical value of a recommendation depends on more than whether it is technically correct. In complex polypharmacy, pharmacists may confront multiple plausible recommendations while simultaneously managing incomplete information, competing tasks, verification demands, and interprofessional constraints. This article develops a sociotechnical theory of recommendation prioritization that treats clinical attention as a limited, dynamically allocated resource. The theory distinguishes recommendation generation from patient relevance, attention capture, verification, professional judgment, and eventual action. It proposes that recommendation priority emerges from interactions among expected clinical consequence, patient specificity, actionability, verification burden, temporal urgency, workflow position, and calibrated professional trust. These interactions explain why increasing recommendation volume or predictive accuracy can fail to improve medication review and may, under some conditions, displace attention from higher-value problems. Pharmacist expertise is positioned as an interpretive and coordinating capability rather than a terminal approval step. AI-supported medication review should therefore be evaluated not only by predictive performance, but also by whether it allocates attention appropriately, preserves independent verification, fits clinical workflow, and supports accountable action. The resulting account provides testable conditions under which recommendation ranking should improve, fragment, or distort medication-review work.