AI for Multi-Warehouse Inventory Allocation: Shipping From the Right Location

A home fitness equipment retailer with three warehouses (New Jersey, Texas, and Nevada) was shipping 38% of their orders from the wrong warehouse. By wrong, I mean a warehouse that was not the closest one to the customer. Orders from California were shipping from New Jersey because that is where the inventory happened to be. The average shipping cost on those cross-country shipments was $14.80 compared to $6.20 for orders fulfilled from the nearest warehouse.

After implementing an AI-based inventory allocation system that positioned stock based on predicted regional demand, their ship-from-nearest rate improved from 62% to 84%. Annual shipping costs dropped by roughly $1.2 million on about $28 million in revenue. The improvement came not from adding warehouse capacity but from smarter decisions about where to place existing inventory.

The Core Problem: Demand Is Not Evenly Distributed

If you sell nationally from multiple warehouses and allocate inventory evenly across locations, you are almost certainly misallocating. Demand for specific products varies significantly by geography. A retailer selling outdoor apparel might see 4x the demand for rain jackets in the Pacific Northwest compared to Arizona, while sun protection products show the opposite pattern.

The naive approach is to allocate based on each warehouse's share of historical total demand. If your East Coast warehouse handled 45% of orders last quarter, you put 45% of new inventory there. This ignores product-level geographic variation. The overall order distribution might be 45/30/25 across your three warehouses, but a specific SKU might have demand that splits 20/50/30 due to regional preferences, local weather patterns, or demographic differences.

What the AI Allocation Model Optimizes

The allocation model solves a constrained optimization problem. It wants to minimize total fulfillment cost (shipping cost plus handling cost) while maintaining adequate service levels (in-stock rate and delivery speed) across all regions, respecting warehouse capacity constraints, and accounting for inventory in transit and on order.

The key input is a demand forecast broken down by SKU and region. Rather than forecasting total demand and then splitting it across warehouses, the model forecasts demand at each warehouse's service area independently. A customer in Denver might be serviced by either the Texas or Nevada warehouse depending on the product, current stock levels, and carrier rates, so the model considers overlapping service areas.

Shipping cost matrices map the actual cost of shipping each product (based on weight and dimensions) from each warehouse to each delivery zone. These are not simple distance calculations. Carrier pricing includes zone-based rate structures, dimensional weight charges, and volume discounts that vary by warehouse. The model uses your actual negotiated rates for accurate cost optimization.

The output is an allocation recommendation for each incoming replenishment shipment. Instead of splitting a 1,000-unit purchase order evenly across three warehouses, the model might recommend 250 to New Jersey, 450 to Texas, and 300 to Nevada based on the current demand forecast, existing stock at each location, and the cost-minimizing delivery scenario.

Transfer Logic: When to Move Inventory Between Warehouses

Sometimes the initial allocation is wrong, or demand patterns shift mid-season. The model also generates transfer recommendations, suggesting when it is cost-effective to move inventory from one warehouse to another. This calculation compares the cost of the inter-warehouse transfer (freight plus handling at both ends) against the expected savings from fulfilling future orders from a closer location.

A transfer makes sense when the receiving warehouse is projected to stock out within the lead time for a new replenishment order, the sending warehouse has excess inventory beyond its projected needs for the next 30 days, and the transfer cost is less than the aggregate shipping cost savings on the orders that would otherwise ship from a farther location.

The math on transfers often works for lightweight, high-velocity products. Moving a pallet of skincare products from New Jersey to Nevada might cost $180 in LTL freight, but if it prevents 200 orders from shipping cross-country at an incremental $8.60 each, the savings are $1,720 minus the $180 transfer cost. For heavy or bulky items, the transfer cost is higher and the math works less often.

Handling Split Shipments

Multi-item orders create a secondary optimization problem. If a customer orders three items and one is only available at the East Coast warehouse while the other two are at the West Coast warehouse, you have three options: ship everything from East Coast (two items travel farther than necessary), split the shipment into two packages (higher total shipping cost but faster delivery for the two West Coast items), or wait until all items are available at one location (slower overall delivery).

The AI model evaluates these options for each multi-item order based on the actual cost difference, the customer's delivery expectations (standard vs. expedited), and whether the customer has shown sensitivity to split shipments in the past (some customers complain about receiving multiple packages). The optimal choice varies by order and there is no single best policy.

Reducing split shipments is often a bigger savings opportunity than optimizing single-item allocation. A housewares retailer found that 22% of their multi-item orders were shipping as split shipments at an average incremental cost of $7.40 per order. By adjusting inventory allocation to ensure that commonly co-purchased items were stocked at the same warehouse, they reduced split shipments to 13% and saved $340,000 annually.

Data Requirements and Implementation

Building an effective allocation model requires several data sets: historical orders with customer location (zip code level), warehouse fulfillment data (which warehouse shipped each order), carrier rate tables for all active carriers and service levels, current inventory positions at each warehouse updated at least daily, and purchase order data for incoming replenishment.

The model itself can be a linear programming solver for the allocation optimization, combined with a demand forecasting model (gradient boosting or neural network) for the regional demand predictions. Open-source solvers like Google OR-Tools or PuLP handle the optimization well, and the demand forecasting can use any standard ML framework.

Integration with your WMS (Warehouse Management System) and OMS (Order Management System) is essential. The allocation recommendations need to flow into your receiving process so incoming inventory is directed to the right warehouse. The order routing logic in your OMS needs to consider the allocation model's warehouse priorities when assigning fulfillment locations to new orders.

For ecommerce retailers operating multiple fulfillment centers, inventory allocation is one of those problems that feels manageable until you do the math. The difference between a naive even-split allocation and an optimized one typically ranges from 15-25% in total shipping cost savings. On a $20 million business spending 8% of revenue on shipping, that is $240,000 to $400,000 per year, which is enough to fund the allocation system many times over and still have money left for other improvements.