AI Demand & Stock Forecasting — 31% Less Overstock for a European Online Retailer
E-commerce
Visual-AI-Labs replaced spreadsheet-based reordering with an AI demand and stock forecasting system in two 30-day cycles, cutting overstock by 31% and stockouts by 44%.
- −31% — Overstock (tied-up capital)
- −44% — Stockout rate on top-1,000 SKUs
- −75% — Time spent on manual reorder planning
- 86% — Forecast accuracy (top SKUs, 4-week horizon)
The problem
Reordering across the retailer’s three brands ran on a shared spreadsheet, reviewed weekly by two planners using last year’s sales and gut feel. Seasonal spikes and promo-driven demand were routinely missed: best-sellers went out of stock mid-campaign while slower SKUs piled up in the warehouse. Planners spent most of their week updating the spreadsheet rather than deciding what to actually reorder.
What Visual-AI-Labs built
Visual-AI-Labs delivered a demand and stock forecasting system in two 30-day cycles. Cycle 1 connected to the retailer’s sales history, current stock levels and planned promotions, and shipped per-SKU demand forecasts with confidence intervals and an automated reorder-point engine that flags what to reorder and when. Cycle 2 added supplier lead-time modelling, a "what-if" view for upcoming promotions, and a planner dashboard that replaces the spreadsheet with an actionable, always-current view.
- Per-SKU demand forecasts with confidence intervals, refreshed daily
- Automated reorder-point engine tied to real supplier lead times
- "What-if" promo simulation to preview stock impact before a campaign launches
- Connector to the existing ERP/WMS — no change to how orders are physically placed
- Planner dashboard replacing the shared spreadsheet, with exception alerts
Results
Within 90 days, overstock (measured as tied-up working capital in slow-moving inventory) fell 31%, and stockouts on the top 1,000 SKUs dropped 44%. Forecast accuracy on a 4-week horizon reached 86% for top-selling SKUs. Planners cut manual reorder-planning time by roughly 75%, redirecting that time to supplier negotiation and new-SKU planning instead of spreadsheet upkeep.
Tech & process
Visual-AI-Labs ran two successive 30-day cycles with weekly demos, validating forecasts against held-out historical data before any reorder recommendation went live. The forecasting models blend statistical time-series methods with promo and seasonality signals; the reorder engine only ever proposes actions inside the retailer’s existing ERP/WMS — planners retain final approval on every purchase order in the first months of operation.
Discuss a similar project with Visual-AI-Labs →
FAQ
Does the AI place orders automatically?
By default, no — it proposes reorder recommendations that a planner approves. Full auto-approval for low-risk SKUs can be enabled once trust is established.
How does it handle new SKUs with no sales history?
New SKUs are seeded from similar-product cohorts and blended in as their own sales history accumulates, rather than left unforecasted.
What about promotions and seasonal spikes?
Planned promotions and known seasonal patterns are fed into the model explicitly, and the "what-if" view lets planners preview stock impact before a campaign launches.
Did the retailer need a new ERP or WMS?
No — the forecasting and reorder engine connect to the existing ERP/WMS via API; no replacement was required.
How accurate are the forecasts?
Around 86% on a 4-week horizon for top-selling SKUs in this project; accuracy is monitored continuously and models are retrained as new data arrives.
How long until results were measurable?
Reorder recommendations went live within the first 30-day cycle; the full overstock and stockout improvements were measured at the 90-day mark.
Does this work for a single-brand retailer too?
Yes — the multi-brand setup added complexity Visual-AI-Labs specifically accounted for; a single-brand catalogue is a simpler starting point.