Digital pharmacy increasingly places algorithmic functions within medication workflows, yet the pharmacy workbench is still commonly treated as a physical workstation or software interface. This framing is insufficient because medication decisions emerge from interactions among patient context, prescription information, professional judgment, technical outputs, temporal pressures, communication, and organizational authority. This article proposes a sociotechnical architecture that reconceptualizes the pharmacy workbench as a human–AI coordination space rather than an automation endpoint. The architecture comprises six connected layers: patient context; prescription and medication information; bounded algorithmic functions; pharmacist judgment and action; coordination-state and temporal control; and escalation, governance, and learning. Algorithmic functions may detect, retrieve, prioritize, predict, summarize, explain, estimate uncertainty, or abstain, but their outputs remain distinct from medication decisions and professional authorization. Proposed coordination states make incomplete context, human–AI discordance, interruption, deferral, uncertainty, escalation, and closure visible rather than allowing them to remain implicit. Evaluation is organized across technical validity, clinical-task relevance, human–AI team performance, workflow fit, medication safety, equity, governance, external validity, and implementation sustainability. The architecture additionally requires task ownership, information provenance, resumption support, escalation acknowledgement, auditability, and controlled system change. It does not establish clinical effectiveness, reduced workload, improved medication safety, or deployment readiness. Its components, transition rules, and decision boundaries require empirical evaluation across pharmacy settings, medication tasks, professional roles, populations, and organizational conditions. The original contribution is an evidence-bounded architecture for studying and designing coordinated medication work without treating any model, alert, professional action, or data source as sufficient by itself.