Measuring Pharmacy Artificial Intelligence by Avoided Harm, Recovered Time, and Better Decisions

Artificial intelligence in pharmacy is often evaluated through discrimination, accuracy, alert yield, or task completion, yet these measures do not establish whether medication-related harm was avoided, professional time was genuinely recovered, or decisions became better. This article proposes a non-empirical value-evaluation framework for connecting technical performance to pharmacy-practice consequences without treating any single metric as sufficient. Value is defined as attributable, net, distributed, and sustained benefit relative to a specified medication-use problem, comparator, perspective, and time horizon. The framework separates seven evaluative layers: baseline need, technical fitness, pharmacy-task performance, human–AI and workflow interaction, consequence measurement, attribution, and net value. Three consequence domains are distinguished. Avoided harm requires evidence linking an AI signal to professional review, action, changed medication use, and a credible counterfactual clinical consequence. Recovered time requires deduction of verification, correction, training, maintenance, and coordination work, followed by assessment of how released capacity is used. Better decisions require evaluation of interpretation, uncertainty handling, appropriateness, timeliness, and final action rather than agreement with AI alone. Validation should combine task-specific performance studies, workflow observation, human-factors assessment, comparative outcome designs, economic evaluation, subgroup analysis, and postimplementation monitoring. The proposed framework does not supply universal thresholds, combine heterogeneous outcomes into a single score, or establish safety, cost-effectiveness, regulatory acceptability, or deployment readiness. Its original contribution is an evidence-bounded architecture for designing, interpreting, and governing pharmacy-AI value claims.