Human–AI collaboration in clinical pharmacy is often evaluated through average accuracy or efficiency, although the same system may reduce cognitive burden in one setting while increasing interruption, misplaced trust, or responsibility ambiguity in another. This realist review examined when, how, for whom, and under what conditions human–AI collaboration may improve medication-related decision quality, workload, and safety. We used a RAMESES-aligned realist-review design. Searches covered seven databases and supplementary citation, journal, web, and registry sources. After deduplication, 817 records were screened, 79 reports were sought, 76 full texts were assessed, and 21 reports representing 21 distinct study or document families were included. Evidence was selected for relevance and rigour, extracted into context–mechanism–outcome configurations, tested against contradictory cases, and synthesized into a refined programme theory. Collaboration appeared most promising when AI supplied complementary, task-specific information; pharmacists retained interpretive authority; outputs were integrated into workflow; and escalation routes remained explicit. Workload reduction could preserve cognitive capacity when low-value processing was removed, but poorly targeted alerts redistributed rather than reduced work. Trust was more appropriately calibrated when system limits, uncertainty, and accountability were visible. Incorrect recommendations, opaque explanations, weak organizational support, and unclear responsibility could promote automation bias, vigilance loss, or decision diffusion. workflow, professional authority, organizational support, and system reliability. The resulting programme theory is evidence-bounded and requires prospective testing in clinical-pharmacy settings before routine implementation.