The dispensing surface drives everything downstream — a cabinet-level vision model on an Omnicell XT has a fundamentally different image distribution, throughput profile, and validation scope than a central-pharmacy Swisslog BoxPicker or a 503B sterile-compounding line under USP <800>. Scoping surfaces up front also determines FDA 510(k) Class II device boundaries and whether the model falls inside an OEM-cleared device or as a downstream verification overlay. Most first-generation deployments fail when teams retrofit a retail-pharmacy model onto an inpatient cabinet workflow.

Sub-2-second verification is the published industry threshold Omnicell and BD design to, and it is not a soft target — once the verification step exceeds human working-memory patience, nurses and pharmacists develop workarounds that defeat the control entirely. Cloud round-trips from a nursing floor to a central data center typically consume 300–800ms before any inference runs, which compresses the model compute budget to almost nothing. The latency decision is made once at architecture time and constrains every subsequent choice.

Industry data puts baseline dispensing errors at roughly 5% of all dispensed medications (ASHP 2023) and AI verification reduces escape rate by up to 85% in benchmark deployments — but the remaining escape rate is where real patient harm concentrates, because staff trust in the AI silences the human double-check. A per-segment miss-rate target, tied to acuity and drug risk class, is the only defensible way to set the threshold. Flag rates above ~5% on high-volume cabinets drive alarm fatigue fast.

AHRQ PSNet puts the average preventable ADE cost at $5,857 inpatient and $125K+ for sentinel events; Vizient benchmarking and the Stanford Health Care case study put annual formulary-optimization savings at $1.2M–$4.8M per health system, with Stanford reporting $3.1M in Year 1 through therapeutic substitution and 340B contract optimization. Framing the program against a dollar-denominated budget changes how pharmacy, finance, and IT prioritize verification accuracy versus formulary-ML breadth. Without it, teams tend to over-invest in one half and starve the other.

HIPAA at 45 CFR 160/164 and the HITECH breach-notification rule make the inference runtime a covered-entity boundary the moment images or audit logs carry patient identifiers — which is almost always the case at the cabinet. Routing verification images or order data to a public cloud endpoint expands the BAA surface and creates new breach liability. Scope reduction — keeping PHI on-node for verification and de-identifying only the formulary-ML layer — is an architectural decision, not a policy afterthought.

Automated dispensing cabinets are FDA Class II medical devices requiring 510(k) clearance, and FDA's March 2026 draft guidance on AI/ML-enabled pharmacy automation explicitly addresses predetermined change control plans for the inference component. Treating an AI verification overlay as "just software" when it is in fact gating a device-level safety control is the fastest way to end up with an FDA Form 483 observation. The classification decision governs validation rigor, design history file requirements, and software change control under IEC 62304 and ISO 13485.