ASHP 2025: Pharmacists Must Own AI Validation and Governance

• ASHP’s 2025 statement replaces the 2020 guidance, explicitly addressing generative AI, large language models (LLMs), and agentic systems and shifting expectations toward pharmacist-led validation, oversight, and ongoing surveillance across the medication-use process.

• Pharmacy leaders, informaticists, and Pharmacy and Therapeutics (P&T) committees must prioritize vendor transparency, local validation, audit trails, monitoring for model drift, and clear policies for sanctioned generative tools to manage shadow AI and operational liability.

• Governance and liability shift: ASHP expects pharmacists to own validation, monitoring, and policy for AI in the medication‑use process. Meeting that bar will likely require a formal AI governance program, versioned test sets, audit trails, and clear human sign‑off rules to manage liability for misuse and non‑use.

• Operational reality—agents: EHR‑embedded agents may reduce documentation and forecasting time but will shift pharmacists toward verification and audit tasks. It will require revised workflows, defined review levels, training to counter automation bias, and post‑go‑live monitoring to avoid added workload and rubber‑stamp approvals.

• Structural and Shadow AI: To meet ASHP expectations, pharmacies need cleaner data pipelines across EHR, IV compounding, and ADC systems plus sanctioned, privacy‑compliant generative tools to displace unsanctioned use. Budget and talent investments are necessary, and uneven rollout could widen gaps between flagship and satellite sites.

Treat the ASHP 2025 statement as an immediate governance priority; pharmacists must own AI validation, monitoring, and policy within 6–24 months.

Local validation: pharmacy informaticists and safety pharmacists should build versioned test sets from electronic health record (EHR) orders, allergies/labs, automated dispensing cabinet (ADC) transactions, barcode scans, and IV gravimetric/vision logs. They must measure alert precision/recall, false‑negative dose‑range rates, and inventory forecast error in a non‑production environment.

• Post‑go‑live surveillance: teams should track override rates, alert fire rates, near‑miss captures, ADC stockouts/backorders, and compounding verification mismatches. Informatics maintains dashboards and medication safety reviews weekly trends; Pharmacy and Therapeutics (P&T) receives quarterly summaries. Predefined thresholds for case‑mix or formulary shifts should trigger revalidation.

• Integration and outputs: EHR‑embedded clinical decision support (CDS) may produce queued orders, draft notes, and dosing checks; IV workflow systems send verification signals to pumps and scales; ADC platforms generate restock recommendations. Data flows use HL7 and FHIR APIs, and humans must verify outputs before order release or medication handoff.

• Regulatory boundaries: the FDA treats image/signal analysis and autonomous dose calculators as Software as a Medical Device (SaMD); guideline‑based order sets and drug–drug alerts can be Non‑Device CDS if clinicians can independently review outputs. The ONC HTI‑1 rule requires disclosure and risk management; NIST frameworks are voluntary.