Before#
The purchasing workflow required reviewing roughly 1,247 active SKUs per reorder cycle. Each decision involved checking current stock by location, inbound purchase orders, recent and longer-term sales velocity, stockout-distorted history, supplier and pack constraints, landed cost, lead time, air-versus-sea freight, committed capital, related component SKUs, and open purchase orders.
A reorder-point report can rank urgency. It cannot allocate a limited budget across competing opportunities.
System approach#
- Normalize the decision surface. A purchasing view over multiple source systems; the intelligence layer computes more than 87 derived features capturing availability, inbound state, velocity, risk, economics, and provenance.
- Simulate inventory through time. A day-by-day simulation estimates when stock runs out and how candidate orders with different lead times change that trajectory.
- Value scarce capital. Candidates are scored on urgency and economic return - profit per dollar and capital lockup. The allocation stage chooses among SKUs under a fixed budget instead of generating independent unconstrained recommendations.
- Compare logistics choices. Air freight preserves sales but consumes margin; sea freight protects margin but may arrive too late. The tradeoff is explicit.
- Preserve provenance. Recommendations carry the facts and calculations behind them, and outputs connect to existing purchase-order and ticket workflows.
Allocation walkthrough#
Consider three sample products under a $10,000 budget constraint. Independent reorder quantities total $18,900 - the budget forces a decision, which is exactly what an unconstrained per-SKU report cannot make.
| Product | Position | Independent want | Allocated | Rationale |
|---|---|---|---|---|
| A | high margin, stockout in 9 days | $5,400 by sea | $5,400 (air) | only air arrives before the stockout; margin survives the freight premium |
| B | stable demand, bridge stock healthy | $7,200 | $4,600 (sea) | sea is viable; partial top-up covers the lead-time window at lower cost |
| C | high apparent velocity | $6,300 | $0 | velocity trace shows one bulk order, not organic demand - rejected with evidence |
Representative walkthrough. The rejected alternative illustrates the operational constraint: ranking by demand alone would over-allocate to non-urgent sea freight.
Hard problems#
- Historical sales understate demand after stockouts.
- Bulk orders create velocity outliers that masquerade as demand.
- A listing pack may consume multiple physical components.
- Inventory exists across warehouse, FBA, inbound, and supplier states.
- The globally best order set differs from each SKU’s independent optimum.
- Lead-time uncertainty makes false precision dangerous.
- Operator trust requires showing why a recommendation exists.
Human control#
This is decision support and controlled execution, not an autonomous buyer. Recommendations carry provenance and explicit assumptions; supplier grouping and existing-PO awareness prevent duplication; and the operational handoff keeps a person in the loop for commitments of real money.