AI tools can tighten medicine inventories and reduce shortages in pharmacy chains, but most evidence comes from simulations and proprietary pilots, making real-world gains contingent on data quality, governance and workflow redesign.

Main Finding

AI-driven methods (machine-learning forecasting, mathematical optimization, simulation, reinforcement learning, and automation) can materially improve pharmacy inventory performance — lower average inventory costs, better forecast accuracy, reduced waste and manual work, and improved shortage planning — but realized value depends heavily on data quality, process redesign, governance, interpretability, and alignment with procurement/resilience policies. AI is most valuable as an augment to human judgment within a staged implementation roadmap rather than as an immediate replacement for established operational practices.

Key Points

• Scope and evidence base
  • Systematic review of literature from Jan 2016–May 2026 across Scopus, Web of Science, PubMed, IEEE Xplore, and Google Scholar. English only.
  • 433 records → 312 screened after duplicate removal → 78 full texts assessed → 35 studies included; 10 additional regulatory/industry documents used for context.
  • Studies rated on a five‑item quality checklist; 11 high, 16 medium, 8 exploratory/early-stage.
• Dominant technical approaches
  • Forecasting and prediction (largest cluster), optimization under uncertainty (MILP, MIP), reinforcement learning (e.g., PPO), attention-based graph neural networks, simulation and multi-agent models (including LLM-based agents for shortage response).
  • Hybrid architectures that pair ML forecasting + prescriptive optimization + simulation yielded strongest actionable results.
• Typical reported outcomes (selected study highlights)
  • Improved logistical financial outcome by 34% in a Dutch outpatient cold‑chain MILP case (Potters et al., 2024).
  • Attention-based graph neural network improved supply/inventory prediction in large networks (Ahn et al., 2024).
  • RL and classical policies both lowered average costs vs human baseline; classical policies often more robust and less costly to run (Stranieri et al., 2025).
  • ShortageSim multi-agent model reduced resolution‑lag by 83% vs baseline on FDA shortage events (Cui et al., 2025).
  • Backup-supplier network design reduced expected shortages from 10% to 4% (Tucker & Daskin, 2021).
• Data, metrics, and limitations
  • Datasets heterogeneous and rarely standardized or shared; high-value work uses internal enterprise dispensing, replenishment, supplier lead times, shortage notices.
  • Five metric families dominate: forecast accuracy (MAE/RMSE), service performance (stock‑out, fill rate), inventory efficiency (holding cost, expiry/waste), resilience (time‑to‑recovery), workflow impact (staff hours, manual touches).
  • Many studies simulation-based; limited multi‑site, real-world deployment evidence and few public benchmark datasets — replication and cross-study comparisons are constrained.
• Managerial lessons
  • Start with visibility and data readiness (item master, supplier mapping, expiry capture) before automation.
  • Prefer interpretable, forecast-plus-policy or MILP solutions early; reserve RL for well-justified complexity.
  • Combine AI with redundancy and procurement rules (buffers, supplier diversification) — AI is a resilience lever, not a substitute.
  • Governance controls needed: intended‑use statements, auditable data lineage, drift monitoring, human override.
  • Staged change management: dashboards → exception detection → recommended actions → semi/auto execution.

Data & Methods

• Review protocol
  • Databases: Scopus, Web of Science, PubMed, IEEE Xplore, Google Scholar.
  • Time window: January 2016 – May 2026. English-language only.
  • Search terms combined pharmacy/pharmaceutical supply topics with AI/optimization terms (forecasting, RL, inventory control).
• Inclusion/exclusion criteria
  • Included: operationally focused studies on pharmacy/hospital medication inventory or pharmaceutical distribution using AI/ML/advanced optimization with extractable methods and managerial implications.
  • Excluded: molecular drug discovery, purely clinical decision support without operational implications, non-English records, editorials without extractable methods.
• Screening and coding
  • Flow: 433 records → 121 duplicates removed → 312 title/abstract screened → 78 full texts → 35 studies included.
  • Coded studies plus 10 regulatory/industry sources for context.
  • Five‑item quality checklist: setting clarity, data description adequacy, method transparency, metric appropriateness, managerial usefulness.
• Analytical synthesis
  • Thematic coding for analytical approach (forecasting, optimization, RL, simulation), operational setting (hospital pharmacy, outpatient, network-level), datasets used, and outcome metrics.
  • Comparative table of representative studies summarizing context, technique, data/metrics, and operational result.

Implications for AI Economics

• Value drivers and measurable economic impacts
  • Direct cost impacts: lower holding costs and expiry/waste, improved working capital through more efficient inventory levels, and reduced emergency procurement costs during shortages.
  • Indirect value: pharmacist time redeployed toward clinical tasks, fewer manual errors, improved regulatory compliance, and better population-level service continuity.
  • Quantified examples show large potential uplifts (e.g., 34% logistical improvement; 83% shortage‑resolution lag reduction) but these are context‑specific and often simulation or single-case based.
• Cost–benefit and complexity premium
  • Economically, organizations should require that more complex AI (e.g., deep RL) deliver an expected "complexity premium" that exceeds added computation, validation, maintenance, and governance costs.
  • Simpler, interpretable models (forecast + rule/policy, MILP) often yield acceptable robustness and lower operationalization cost — higher expected ROI in many settings.
• Resilience vs. efficiency trade-offs
  • Optimization that minimizes average cost may increase risk of shortages; regulators and managers value resilience and patient-safety constraints, implying multi-objective optimization and potential willingness to accept higher inventory costs for lower shortage risk.
  • Economic evaluation should include downside externalities (clinical risk, reputational or regulatory penalties) and multiplier effects from shortages.
• Governance, compliance and regulatory economics
  • EU AI Act, NIST, WHO guidance raise compliance costs (documentation, monitoring, transparency) that must be included in deployment business cases, especially for multi-jurisdictional vendors/health systems.
  • Auditable models and data lineage reduce legal/operational risk but increase upfront implementation cost; these are economically justified where shortage/exposure costs are high.
• Market and procurement implications
  • Procurement and contracting should be aligned: AI-enabled forecasting can change purchase timing and volumes, affecting supplier production planning and contract structures (e.g., more dynamic or contingent purchasing).
  • There is economic value in combining AI with contracting for supplier diversification, buffer commitments, and contingent pricing to internalize shortage risk.
• Research and policy gaps relevant to AI economists
  • Need for empirical, multi-site cost-effectiveness and ROI studies from real deployments (not just simulations).
  • Standardized benchmark datasets and shared evaluation protocols would enable cross-study economic comparison and better market signals.
  • Investigate distributional effects: which hospitals/regions or patient groups benefit or are disadvantaged by automated allocation rules.
  • Study optimal trade-offs between redundancy (buffer costs) and predictive accuracy (investment in AI) under different shortage probability regimes.
• Practical metrics for economic evaluation
  • Use a balanced scorecard for valuation: stock-outs/shortages (patient-safety cost proxies), expiry/waste (direct cost), working capital (inventory value), procurement/emergency spend reduction, and labor cost changes (pharmacist hours redeployed).
  • Include compliance and monitoring costs as recurring operational expenditures in NPV/IRR models.

Concluding note for AI economists: the field shows promising operational economics but is still maturing. Economic models assessing AI in pharmacy inventory must capture implementation, governance, and resilience costs as well as savings; prefer staged investments, rigorous field-level trials, and concerted efforts to create shared benchmarks to move from promising simulations to generalizable economic conclusions.