Detecting Workflow Drift before Pharmacy Algorithms Learn the Wrong Practice

Artificial intelligence increasingly participates in medication-related prioritization, alerting, verification, documentation, and clinical decision support. Yet deployed algorithms may continue learning while the pharmacy work that produces their inputs, labels, feedback, and outcomes is changing. This creates a risk that temporary workarounds, redistributed responsibilities, altered documentation, changing patient populations, or unsafe local practices become represented as legitimate practice. This article defines pharmacy workflow drift as a material temporal change in the tasks, professional roles, activity sequence, patient or case population, or documentation processes through which medication-related work is performed and represented. It proposes a Multidimensional Pharmacy Workflow-Drift Framework comprising a baseline workflow contract, five-dimensional observability layer, cross-dimensional coupling assessment, consequence layer, and learning-control layer. Detection is separated from diagnosis: a statistical or process signal initiates investigation but does not confirm drift, establish causation, or authorize model updating. The framework therefore connects drift localization to medication-safety, equity, human-factors, implementation, and accountability assessments before selecting a learning state such as continued operation, enhanced surveillance, learning freeze, restricted use, controlled updating, rollback, or retirement. Validation would require technical sensitivity, false-alarm assessment, causal plausibility, workflow localization, professional usability, subgroup stability, safety evaluation, rollback capability, and recurring local requalification. The framework is an original non-empirical governance synthesis. It does not provide validated thresholds, guarantee safer decisions, allocate legal liability, or establish regulatory or deployment readiness.