Artificial intelligence can attach probabilities, confidence scores, intervals, disagreement measures, or out-of-distribution signals to medication-related outputs, yet these numerical representations do not independently specify what a pharmacist, prescriber, or patient should do. The problem is not only whether uncertainty is estimated, but whether its source, decision relevance, consequences, and ownership are translated into a proportionate response. This article develops a proposed Uncertainty-Translation Framework for actionable medication advice. The framework separates algorithmic output from medication decision-making through five connected functions: uncertainty characterization; calibration, coverage, data-completeness, and scope checks; interpretation of the affected medication decision; selection of a bounded action class; and recipient-specific communication supported by an auditable governance record. Proposed action classes include proceeding with qualified advice, verifying data, checking an independent source, delaying action, communicating uncertainty, or escalating to an accountable professional. The framework treats these classes as conditional design hypotheses rather than clinical rules. Validation would require technical assessment of calibration and coverage; human-factors testing of comprehension, workload, and inappropriate reliance; evaluation of medication decisions and downstream consequences; subgroup and equity analyses; and prospective governance of versioning, overrides, incidents, and withdrawal. The framework cannot correct biased targets, invalid data, unsuitable model purposes, or deficient professional judgment. Its original contribution is to define uncertainty translation as an intermediate sociotechnical function between prediction and advice while preserving the boundaries between technical performance, clinical usefulness, professional authority, and validated benefit.