A retail chain reduced stockouts by 55% and optimised inventory costs by 30% with AI-powered reordering

Store managers placed manual reorders based on gut feel and historical averages. Fast-selling items ran out regularly, while slow movers accumulated excess stock. There was no centralised view of inventory across locations. Seasonal demand shifts caught the team off guard every quarter, and markdowns on excess inventory eroded margins.

The chain operated eleven stores across four cities, carrying between 3,000 and 5,000 active SKUs per location depending on floor size. Each store manager was responsible for their own inventory, monitoring stock levels, deciding when to reorder, and managing the relationship with the central buying team. The process was almost entirely intuition-based. Store managers ordered what they thought would sell, informed by their memory of previous weeks and a rough sense of upcoming events or seasons.

The structural problem with this approach was that individual managers had good local knowledge but no access to aggregated data. A manager who had worked at a location for two years had a strong intuitive sense of which products moved on which days of the week. But they had no visibility into what was selling at other locations, no way to see whether a slow-moving product was performing better elsewhere, and no data-driven signal for how much to increase an order ahead of a local event or a national marketing campaign.

Stockouts were the most visible problem. Fast-selling items in popular sizes, colours, or variants ran out regularly, sometimes within days of a reorder arriving. When a product was out of stock, the chain lost the sale entirely: customers who could not find their size or preferred variant often did not visit a second store to look for it, they simply bought from a competitor. The stockout problem was worst on high-velocity, low-margin items where the manager's reorder quantity was calibrated for typical demand and could not accommodate the spike that followed a promotion or a competitor closure nearby.

On the other side of the inventory problem, slow-moving items accumulated excess stock that tied up working capital and eventually required markdown clearance. The chain ran a quarterly markdown cycle that cleared unsold stock at 30-50% discounts, eroding the margins on products that had simply been over-ordered. The total markdown cost in the previous year represented a significant percentage of gross margin, a cost that could have been substantially avoided with better demand forecasting.

We deployed an AI demand forecasting model trained on two years of POS data, seasonal trends, local events, and weather patterns. The system generates reorder recommendations daily for each store and SKU, accounting for lead times and minimum order quantities. A central dashboard gives the merchandising team real-time visibility into stock levels, reorder status, and demand forecasts across all locations. Reorder approvals happen in one click.

The foundation of the solution was a clean, unified data model. The chain's eleven stores used the same POS system, but the data had never been aggregated or analysed centrally. We built an extraction pipeline that pulled two years of transaction-level sales data, enriched it with product metadata (category, supplier, lead time, minimum order quantity, cost and selling price), store metadata (location, size, catchment demographics), and external data sources (local event calendars, weather data, school holiday schedules, and national promotions from the chain's marketing calendar).

The forecasting model is a gradient boosted tree ensemble trained on this enriched dataset, with separate model instances per product category to account for the different demand dynamics of fashion items, consumables, and seasonal goods. The model generates daily demand forecasts for each SKU at each location, with a 28-day forward horizon updated every night using the previous day's sales data. Forecast accuracy is measured weekly against actuals, and the model retrains monthly on the full updated dataset.

Reorder recommendations are generated nightly from the forecasts. For each SKU at each location, the system calculates current stock level, forecasted demand over the supplier's lead time plus a safety stock buffer, the minimum order quantity, and any outstanding purchase orders already in transit. When projected stock will fall below the safety stock level within the lead time window, the system generates a recommended purchase order with the specific quantity calculated to bring stock to the optimal level without over-ordering.

Recommendations are surfaced through a central dashboard accessible to the merchandising team and, in a simplified view, to individual store managers. The dashboard shows each store's recommended orders for the next three days, flagged by urgency. The merchandising team reviews and approves orders with a single click, the system generates the formatted purchase order and sends it directly to the supplier via email or EDI integration. Managers who disagree with a recommendation can adjust the quantity before approving, and the model incorporates these overrides as feedback signals for future forecasting.

The central inventory view gives the merchandising team, for the first time, a consolidated picture of stock levels, sell-through rates, and reorder status across all eleven stores. They can identify which SKUs are outperforming forecast across all locations (signal to negotiate larger orders), which are underperforming (signal to reduce future buys), and which show strong performance in some locations that could inform ranging decisions in others.

Stockouts dropped by 55%. Inventory carrying costs fell by 30%. Markdowns on excess stock decreased significantly. The merchandising team shifted from reactive ordering to strategic planning, using demand forecasts to negotiate better supplier terms.

The stockout reduction was the most commercially significant outcome. Across the chain, stockout incidents, defined as a SKU reaching zero stock for more than 24 hours, dropped by 55% in the first six months of operation compared to the same period the previous year. The improvement was concentrated in the high-velocity items that had previously been most prone to running out: bestselling sizes in core product categories, fast-moving seasonal items, and products that spiked after promotional activity. These were precisely the items where stockouts were most costly, because they represented the highest-demand, highest-margin sales opportunities.

Inventory carrying costs fell by 30% as excess stock accumulation reduced. The forecasting model's ability to right-size orders to actual demand, rather than erring on the high side to avoid stockouts, which is the natural tendency of manual orderers, meant that less stock sat unsold in the back room. Working capital tied up in inventory freed up, which the finance team redirected to a small expansion of the chain's product range in two high-performing categories.

Markdowns in the quarterly clearance cycle decreased substantially. The chain's markdown volume fell by around 40% compared to the pre-deployment average, reflecting the reduction in over-ordering of slow-moving items. The products that did go to markdown were predominantly end-of-range items being discontinued by the supplier rather than excess stock from poor ordering decisions.

The merchandising team's working pattern changed significantly. Before deployment, the team operated reactively, responding to stockout calls from store managers, processing urgent reorders under time pressure, and manually reviewing store-by-store sales spreadsheets to identify trends. After deployment, they had a forward-looking, data-driven view of inventory across all stores. This shift enabled a new kind of strategic work: using the demand forecast data to build more accurate buying plans for the following season, identifying supplier lead time issues before they caused stockouts, and negotiating better terms with suppliers using concrete sell-through data as leverage.