Procurement intelligence that protects margin, strengthens control and keeps care supplied

Most procurement AI conversations start with sourcing bots and end with nothing deployed. The jobs that actually get automated are smaller and less glamorous: classifying 400,000 PO lines, pulling FOC slabs and renewal dates out of 300 agreements, chasing 60 vendors for stock availability every week, triaging invoice exceptions, and forecasting next quarter's consumables demand.

Each has a clear input, a checkable output, and a person who currently does it by hand. That makes them ideal for a model or an agent with a human approving the result. We have run every one of them inside a multi-hospital group using Microsoft 365 Copilot agents, Power Platform and a small amount of Python; none required a new system.

The precondition is data the model can trust. Classification on an unclean vendor master produces confident but incorrect results. Fix the master data, agree the taxonomy, then let the model do the volume.

Most hospital groups have already run the big tenders. The remaining value sits where a tender never touches: free-of-charge quantities promised but not received, bonus stock credited to the wrong entity, consignment usage billed at list rather than contract price, and the 15–30% of PO lines that never reference a contract.

Our reviews across pharma, consumables and devices consistently find 8–12% of addressed spend in these four pockets. None requires a new supplier; all require data that reconciles agreement terms to receipts and invoices at line level.

Load agreements with their commercial terms, match every receipt and invoice to the agreement line, and report the variance by supplier and entity every month. Once that routine exists, recovery conversations are short, because the evidence is theirs as much as yours.

Two hospitals buying the same stent at the same unit price can have a 20% difference in total cost. One receives bonus units, 90-day terms and free proctoring; the other pays list, 30-day terms and a service contract. A unit-price benchmark calls them equal.

Total-cost benchmarking converts every commercial term into a per-unit equivalent: FOC becomes a discount, extended terms a financing value, bundled services an avoided cost. The result is one comparable number per SKU per supplier per entity.

A negotiation pack showing the supplier's own terms across your entities, converted to total cost, moves the discussion from price to structure, which is where suppliers have room to move.

Every spend cube looks convincing on a slide. The ones that survive a CFO's question reconcile in three places: total spend ties to the general ledger within an explained tolerance; supplier totals tie to accounts payable; category totals follow a classification rule someone can defend.

The hard work is master data. Vendor names spelled four ways, items described eleven ways, categories chosen by whoever raised the requisition. A cube built on that is challenged the first time a category head sees their number.

Build the classification rules before the charts, document them, and publish the tolerance. "We explain 97% of spend and here is the 3%" is more credible than a claim of 100%.

When we map requisition-to-PO in a hospital group, hands-on work usually totals under two hours. The remaining days sit in queues: an approver on leave, a specification clarification by email, a budget check in Finance, a re-key into the ERP.

Three changes remove most of it: a delegation-of-authority matrix with alternates and value thresholds that reflect actual risk; catalogue and contract coverage so routine items need no sourcing step; and a single intake channel for requisitions so nothing arrives through informal messages.

The fourth lever is automation, but only after the first three. Automating a twelve-step process gives you a fast twelve-step process. Simplify first, then let the AI practice take the volume.

Consolidation delivers scale pricing, fewer suppliers and consistent controls when three conditions hold: entities genuinely buy comparable items, the central team has authority over specifications as well as prices, and sites experience the service as faster, not slower.

Where those fail, groups end up with a central team negotiating frameworks nobody uses while sites buy locally through maverick channels. The compliance report looks worse than before the GPO existed.

Decide which categories are group-led, site-led or hybrid; how clinical standardisation is governed; how savings are credited to entities; and what service levels the centre commits to. Answer those before the organisation chart.

Write-offs are rarely a surprise to the data; they are a surprise to the people who did not look. Three indicators, tracked monthly by item class and location, predict almost all of them: stock cover above six months, value aged beyond 180 days, and unreconciled consignment movements.

Stock aged beyond 180 days in a hospital store is usually a specification change, a clinician preference that moved, or a bulk purchase that chased a discount. Each has an owner, and each is recoverable at 180 days rather than at expiry.

A monthly inventory health report with those three numbers, valued and ranked by location, turns the annual write-off into a weekly operational decision.

Consignment is the right model for high-value implants: the hospital pays on use, the supplier carries the stock. It only works when three records agree: what the supplier says was placed, what the system says was used, and what the invoice says was billed.

In most groups those records live in three places and are rarely reconciled. Usage captured by hand in theatre, invoices priced at list rather than contract, and replenishment counted by the supplier's representative. The variance is rarely in the hospital's favour.

The fix is procedural before it is technical: a single usage capture point, contract prices loaded against every consignment item, and a monthly three-way reconciliation with the variance reported by supplier. Then automate the matching.

Procurement value is usually reported as cost avoidance that Finance quietly discounts. Functions that change that conversation report value Finance can find in the P&L: rebates received, marketing and training funds secured, equipment placed at zero capital, consignment models that move stock off the balance sheet.

Suppliers in devices and pharma have budgets for exactly these programmes and would rather fund growth than cut price. A procurement team that understands the supplier's commercial model can structure agreements that pay both sides.

The discipline required is the same as savings tracking: a value register, Finance validation rules agreed in advance, reporting in the same cadence as the MIS pack.