An AI system helped pharmacists reach people. It did not yet prove better health.

Clinical AI and care delivery · Source date: November 2026 journal issue; indexed by November 2 · Evidence: Retrospective, propensity-matched quasi-experiment; historical controls · Primary source

A powerful way to evaluate AI is to begin with an actual service problem. Pharmacy teams know that many people eligible for medication therapy management do not take a consultation. Researchers compared a machine-learning targeting system used in 2025 with traditional assignment in a 2024 cohort. After propensity matching, the analysis included 183,888 patient pairs.

The share accepting a consultation was 21.2% with machine-learning targeting versus 12.6% in the historical approach—an 8.6-percentage-point absolute difference, or 68% relative increase. That is meaningful for reach and staff scheduling. But a historical comparison cannot fully remove changes in operations, outreach tactics or patient characteristics, and the measured endpoint was consultation acceptance—not adherence, avoidable hospitalization, adverse medication events, or mortality.

Reaching someone is the start of care. A prospective comparison could ask whether more completed consultations lead to better health and whether all eligible groups benefit. The historical comparison does not establish that the algorithm caused the difference.

causal AI effect.