Unattended Retail Replenishment

How should each vending machine, smart cooler, or micro-market decide what it gets filled with? Not from a planogram set once at headquarters. A machine in a hospital lobby, one in a gym, and one on a factory floor sell different things at different hours, and treating them as one average wastes both shelf space and route time. The answer we design toward is simple to state: every location learns its own market. Each point of sale gets its own item-level demand model, and the refill list for each visit is generated from what that specific location actually sells.

How it works#

Unattended retail is the per-unit inventory problem in its purest form: many small “warehouses”, each with its own demand curve, restocked on a route with limited van capacity. The design we bring has three layers.

Per-location, per-item forecasting. Sales telemetry from each machine becomes a demand history for that location, and each item at each location gets its own forecast. Seasonality here is often intraday and weekly, the office building empties on Fridays, the gym peaks in the evening, and models learn that per location instead of assuming it.

Dynamic par levels instead of fixed planograms. The stocking target per slot follows the location’s forecast and its restock frequency, the same logic as dynamic safety stock per SKU: thin buffers where demand is steady, thicker ones where it is spiky, and product-mix recommendations when a slot consistently underperforms at one location but not others.

Recommendations inside the operation you already run. The output is a refill list per machine per visit, returned into your existing route-planning and ERP environment through an API layer, not a new system your operators must adopt.

What we’ve built#

We say this plainly: our published work is the underlying per-unit capability, proven in adjacent domains, not a deployed vending system, and we would rather you know exactly which is which.

• Item-level inventory intelligence for ERP environments: designed and specified item-level demand forecasting with dynamic safety stock and purchasing recommendations, as a cloud API integrating into Priority ERP, with a success KPI of at least a 10% reduction in “purchasing error”.
• Demand forecasting built and validated on real data: multi-region product-family forecasts for a global water-technology manufacturer, and a three-year flagship engagement where monthly forecasts were graded against actuals, including dedicated models for distinct customer segments, which is the same statistical discipline as distinct locations.

From the wider industry: More Retail in India wired ML demand forecasting directly into automated store ordering, lifting forecast accuracy from 24% to 76% and in-stock rates from 80% to 90%, per an AWS vendor case study. The lesson transfers whole: the value arrives when the forecast writes the refill list.