How does AI allocation and replenishment work across stores and channels?
Allocation decides where stock goes first; replenishment decides what follows. AI makes both decisions at size and location level. Here is how it works, what it needs and where it struggles.
KEY TAKEAWAYS Summary by the editors
- Allocation is the initial distribution of a delivery to stores, web and partner accounts; replenishment is the ongoing top-up of stock based on what sells.
- AI allocation forecasts demand at style, colour, size and location level and then optimises quantities under constraints such as pack sizes, store capacity and transport schedules.
- Zara's store replenishment model, documented in Interfaces in 2010, chose shipment quantities by linking stock levels to demand and was reported to have increased sales by 3 to 4 percent.
- Because each size and colour is a separate GTIN, allocation models can work at variant level, but only if sales and stock are captured at that level in every channel.
- Allocation models must account for stockouts and display effects, otherwise they send too little to the stores that sell out fastest.
AI allocation and replenishment work by forecasting demand for each product variant at each location, then calculating how much stock to send so that expected sales are maximised within the available inventory and logistics constraints. Allocation handles the first distribution of a delivery; replenishment repeats the calculation as sales come in, moving stock towards where it is selling.
What is the difference between allocation and replenishment?
The two decisions use similar models but face different information. Initial allocation happens before or at launch, when there is little or no sales history for the product, so the model relies on store profiles, similar past products and planned demand. Replenishment happens in season and can use actual sales, which makes forecasts more accurate but leaves less stock to play with.
| Aspect | Initial allocation | Replenishment |
|---|---|---|
| Timing | Before or at product launch | Weekly, daily or more often during the season |
| Main uncertainty | Which stores and channels will want the product | How quickly current sales will continue |
| Main data | Store profiles, analogous products, buy plan | Recent sales and stock by size and location |
| Typical constraint | Total buy quantity, pack sizes, launch presentation minimums | Warehouse stock left, transport schedules, store space |
| Common error | Spreading stock evenly regardless of local demand | Reacting to stockout-depressed sales as if demand were low |
How does an AI allocation model decide quantities?
Rule-based allocation typically uses store grades (A, B, C stores) and fixed ratios. AI-based allocation replaces fixed ratios with a forecast for each store, product and size, and then solves an optimisation problem. A typical sequence:
- Estimate demand for each variant at each location for the coming period, with an uncertainty range.
- Set the presentation minimum: the stock a store needs to show a product credibly, often a full size run.
- Compare expected marginal sales of one more unit in each location and assign units where they add most.
- Respect constraints: available stock, pack sizes, store capacity and delivery days.
- Hold back a reserve in the warehouse when uncertainty is high, so stock can follow early sales.
The best-documented public example is Zara. Researchers working with the company built a model to determine each shipment from its central warehouses to its stores, explicitly modelling the link between stock levels and demand. According to the paper in Interfaces (2010), the process increased sales by 3 to 4 percent, with estimated additional revenues of about 233 million US dollars in 2007 and 353 million in 2008. The lesson is that how much is displayed in a store influences how much sells there, and the model has to reflect that.
Why does variant-level data matter?
Allocation decisions are made per size and colour, not per style. GS1, the barcode standards body, specifies that each product variation (each size, each colour and each combination) needs its own GTIN. That makes variant-level allocation technically possible everywhere a barcode is scanned. The practical gap is that many businesses still plan at style level, or do not receive variant-level sales and stock from partner stores and marketplaces.
Stockouts distort the data further. Research by Akchen and Caro on a footwear retailer found that nearly 25 percent of demand left unmet by stockouts shifted to adjacent sizes. Raw sales therefore overstate demand for neighbouring sizes and understate it for the sizes that ran out. Replenishment models need to flag periods of zero or low stock so they do not interpret a sold-out size as an unpopular one.
How does AI handle omnichannel stock?
When stores, e-commerce and wholesale draw on shared stock, allocation becomes a question of which channel gets the next unit. Online demand is usually higher volume and easier to forecast, while store demand is spread across many locations. Some retailers fulfil online orders from store stock, which adds a further option. An AI model can compare the expected margin of each use, but only if all channels report sales and stock in near real time.
Near-real-time signals are the domain of demand sensing. DHL describes demand sensing as short-term forecasting that uses real-time data such as point-of-sale activity, shipments, promotions, weather and social trends, refreshed daily or even hourly. For replenishment, faster signals mainly help with reallocating stock between locations during the season.
What does a business need to implement it?
- Accurate stock by location and variant, ideally with regular counts or item-level tracking.
- Sales history by location and variant, including web and, where possible, partner sell-out.
- Store attributes: size, climate, customer profile, space and fixture capacity.
- Logistics rules: delivery days, pack sizes, lead times and transfer costs.
- Governance: clear rules for when allocators can override the model and how overrides are reviewed.
Large retailers are folding allocation into broader planning programmes. Marks & Spencer, for example, announced in January 2024 an end-to-end planning platform for clothing and home whose scope includes forecasting, allocation and replenishment alongside financial and range planning.
What are the limits?
AI allocation is strongest for products with steady demand and many locations. It struggles with very low volumes per store, where a single sale changes the picture, and with one-off events the model has never seen. Transfers between stores cost money and staff time, so models that recommend constant movement can look good on paper and fail in operations. The benchmark is not perfect allocation but fewer stockouts of core sizes and less end-of-season imbalance between locations, measured against a clear baseline.
A sensible rollout starts with replenishment of carry-over lines in a subset of stores, where history is richest and results can be compared with a control group within weeks. Initial allocation of new products, which depends on analogous items and store profiles, is usually the second step once the data foundations are proven.
Frequently asked questions
What is AI allocation in retail?
AI allocation uses demand forecasts by product, size and location to decide how much of a delivery each store or channel should receive. It replaces fixed store grades and ratios with location-specific estimates and an optimisation that respects stock and logistics constraints.
How often should fashion stock be replenished?
It depends on logistics and volume. Fast-moving retailers replenish stores several times a week, while others work weekly. AI models can recalculate at any frequency, but transport schedules and costs usually set the practical rhythm.
Why do allocation models under-supply best-selling stores?
Because when a size sells out, recorded sales stop. A model trained on raw sales sees lower sales and sends less. Flagging stockout periods and estimating lost demand corrects this.
Can AI allocation work for wholesale accounts?
Yes, for replenishment programmes where the brand sees partner sales and stock at variant level. Without sell-out data from the retailer, the brand can only allocate based on order history.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- London Business School: Zara uses operations research to reengineer its global distribution system
- GS1: How many GS1 GTINs do I need when I have a product with many sizes and colours?
- UCL Discovery: Akchen and Caro, On Size Substitution and Its Role in Assortment and Inventory Planning
- DHL: Supply Chain Demand Sensing (glossary)
- Just Style: M&S to digitally transform clothing and home end-to-end planning system
