Agentic inventory intelligence

High-cost oncology stock sitting across an ERP, several separate dispensing exports, and spreadsheets — with no shared drug codes or units between them, so reconciling what was actually ordered against what was actually dispensed meant a human mapping the same drug by three different names by hand. Ordering decisions for some of the hospital's most expensive and most time-sensitive stock depended on one buyer's accumulated judgement, held in their head rather than in any system — a single point of failure the pharmacy itself was uncomfortable with long before this engagement started. Stockouts on oncology drugs delay treatment; overstock on the same drugs ties up a budget line the pharmacy can't easily recover. Getting the order right mattered more here than in a typical ward, and the existing tooling gave nobody a way to see the full picture at once.

Deployed inside the hospital's own infrastructure, reading directly from the ERP and dispensing systems already in place rather than requiring a parallel data export to an outside platform — patient-adjacent operational data for a public hospital pharmacy doesn't leave the building to get forecast. See on-premise vs. cloud AI for healthcare for the tradeoffs that decision actually turns on. The scope was deliberately bounded to non-clinical operations — stock, dispensing volumes, and procurement — not clinical decision-making, which keeps the system a supply-chain tool rather than a prescribing one.

An agentic workflow built from five registered tools working in sequence. Ingest pulls ERP stock and purchase orders plus multi-source dispensing exports on a schedule. Harmonisation maps every source's units and drug codes to one canonical schema, so the same drug is the same row no matter which system it came from. Forecasting runs multiple algorithms against that harmonised history and attaches a confidence interval to each prediction rather than a single point estimate. Recommendation then proposes an order two ways — a statistical model and a second model trained to mirror the buyer's own historical pattern — so the pharmacist sees where the two agree and where they diverge. Disruption monitoring watches public supply alerts and supplier email for early signals of a shortage before it hits the hospital's own stock count. All five are registered modules an agent orchestrates in sequence, not a single monolithic forecasting model — ingest, harmonisation, forecasting, recommendation and monitoring can each be inspected, re-tuned or replaced on its own without the others needing to change.

Every recommendation is shown with its reasoning attached — which algorithm produced it, what confidence interval it carries, and where the buyer-mirrored model disagrees with the statistical one — never a bare number the pharmacist has to take on faith. Nothing is auto-selected or auto-ordered: the system proposes, the pharmacist decides, consistent with the boundary described in human-in-the-loop design for healthcare AI. Where the forecast's confidence interval is too wide to support a specific recommendation, the system is built to say so rather than produce a falsely precise number — the same discipline covered in AI hallucination mitigation in regulated settings, applied to a forecast rather than a generated passage. Every recommendation and every pharmacist decision on it is logged to an immutable record, per audit trail requirements for AI systems.

One explainable operating picture for the pharmacy in place of three disconnected systems and a buyer's private judgement. Stock, dispensing, and procurement now read from the same canonical schema, so a reconciliation that used to mean manually matching drug names across exports is now a lookup. Recommendations arrive with reasoning attached and are never auto-selected — the pharmacist stays accountable for every order, with the forecasting and cross-referencing work already done rather than left for them to assemble from three disconnected sources under time pressure.

The pharmacy team operates the full workflow day to day. schema_harmonise · demand_forecast · disruption_monitor — a supply-operations stack that transfers to any provider setting with a similar ingest problem, not just oncology.