Artificial intelligence is increasingly used to prioritize prescriptions, suppress low-value alerts, identify medication-error risk, and support pharmacist verification. However, technical detection performance is frequently discussed alongside workflow or patient outcomes as though these endpoints demonstrate equivalent clinical value. This evidence-mapping review examines where AI-assisted medication review has progressed beyond technical screening and where evidence remains limited across automation error, verification workload, pharmacist oversight, and patient-level consequences. Peer-reviewed evidence published during 2017–2026 was mapped by medication-review task, care setting, data input, AI approach, validation design, outcome class, automation-failure characteristics, verification burden, pharmacist role, and proximity of outcomes to patients. Methodological and reporting standards were used to preserve distinctions between model performance, early clinical evaluation, workflow effects, and patient consequences. Evidence is most developed for bounded detection and prioritization tasks, including prescription-risk identification, medication-alert targeting, atypical-order detection, and selection of patients for pharmacist review. Human–AI studies indicate that uncertainty presentation and erroneous algorithmic advice can alter pharmacist decisions and verification time. Patient-level evaluation is less mature and context dependent. The review therefore proposes that clinical value should be judged across separable layers: detection, verification burden, pharmacist actionability, and patient consequence. AI-assisted medication review should not be judged by discrimination performance alone. Evidence becomes more clinically informative when model outputs are evaluated within human verification, pharmacist action, organizational workflow, and patient-level outcome pathways, with explicit attention to error recovery and transferability.
