.png)
The key priority of the project was to equip the client with a machine learning–driven forecasting and inventory optimisation process.
A mid-sized retail chain based in the US needed help to tackle persistent inventory imbalances and lost sales opportunities. The client already had a modern POS and ERP system in place but relied heavily on manual forecasts. These forecasts often led to overstock in some regions and stockouts in others, hurting both revenue and margins.
As part of the client’s decision to explore AI capabilities, they sought a solution to integrate their sales, promotions, and supply data into a single forecasting engine—capable of predicting demand by location and automating replenishment approvals for high-confidence scenarios.
The client’s inventory decisions were managed by store managers using spreadsheets and historical sales. As the business operated in multiple regions, the first requirement was to integrate POS sales, inventory levels, and promotional data into a single system to create an “Approved Demand Plan” for the entire chain.
Data Integration: The “Import Sales & Inventory Data” feature was used to consolidate POS, stock levels, and promotional calendars into the ML forecasting module. APIs were used to fetch data in near real time from multiple systems.
Another requirement was to design the replenishment approval process so that ordering decisions for each store were based on forecasted demand for that specific period—independent of other locations or past cycles.
Forecast-Driven Replenishment Workflow:
1. Implemented amount-based approval hierarchy for replenishment orders generated by ML forecasts.
2. Allowed first two replenishment cycles in each pilot store to bypass approval logic for faster testing and validation.
The approval process needed to follow a clear threshold-based hierarchy, allowing faster approvals for smaller orders and routing larger, high-impact orders for executive sign-off.
Dual Baseline Model:
1. Version one created using aggregated historical sales data, designated as the “Original Baseline Forecast.”
2. Version two initiated at zero, designated as the “Current Forecast,” updated for each period.
3. Any new forecast triggered an approval workflow where the approval level was determined by:
Forecasted Order Value for the Current Cycle = Current Forecast – Original Baseline Forecast
Solution resulted in a fully automated, ML-driven replenishment workflow integrated with the client’s POS and ERP systems.
The ML forecasting engine became the central system for demand planning, replenishment estimation, and store-level inventory tracking—while store managers gained access to actionable order recommendations directly within their familiar ERP interface, without needing to handle raw forecasting models.
The replenishment approval workflow design allowed the client to execute location-specific ordering quickly, bypassing unnecessary approval layers for smaller orders. This enabled store managers to act faster on high-confidence replenishment needs, reducing reliance on head office sign-offs and ensuring products stayed on shelves when customers wanted them.
.png)
.png)
