IKEA

Managing demand planning across 450+ stores and e-commerce operations in 54 markets means IKEA forecasts potentially billions of individual product needs each year. For a retailer of this scale, forecast accuracy is directly tied to customer satisfaction and operating cost. The legacy statistical system generated forecasts top-down — starting at global level and disaggregating to country and store — relying entirely on historical sales patterns. This approach struggled to capture localised demand signals or real-time shifts in buying behaviour. The consequence was a chronic tension between overstock, which drives up logistics costs and suppresses margins, and stockouts, which degrade availability and customer experience. Roughly 8% of AI-generated forecasts required manual correction by planners, adding operational friction and introducing human inconsistency at scale.

IKEA built Demand Sensing, an internally developed AI forecasting platform described as its largest supply chain AI deployment to date. The system applies machine learning and predictive analytics to up to 200 data sources per product — a substantial expansion beyond historical sales data. Inputs include weather forecasts, festival and public holiday calendars, salary-cycle purchasing effects, and seasonal demand patterns. Critically, the architecture inverts the legacy model: rather than disaggregating global forecasts downward, Demand Sensing constructs forecasts bottom-up from individual store level, then aggregates to market, country, region, and global views. The platform covers both in-store and online channels in a unified demand signal, with a flexible forecast horizon running from daily resolution out to four months. The initial rollout was piloted in Norway before broader deployment.

Demand Sensing delivered a measurable step-change in forecast quality. The headline outcome was a forecast acceptance rate rising from 92% to 98%, with the manual override rate falling from 8% to 2%. Beyond the headline numbers, the bottom-up architecture resolved a specific accuracy failure in the legacy system: a localised sales spike — for example in the Furuset store in Oslo — previously propagated upward and inflated national forecasts incorrectly. Under Demand Sensing, that signal stays scoped to the relevant store. Qualitative outcomes include reduced emergency markdowns on overstocked product, lower logistics waste, and a leaner planning workflow for supply chain teams. Cost savings from improved inventory accuracy are designed to flow through to end-customer pricing.

• A bottom-up, store-level forecast architecture prevents local demand noise from contaminating regional or national planning — a structural fix that statistical disaggregation cannot replicate.
• Integrating 200+ external data sources (weather, festivals, salary cycles) allows the model to anticipate demand shifts that purely transactional history will always lag.
• Omnichannel signal unification — treating online and in-store demand as a single input — is a prerequisite for accurate planning in modern retail; siloed channel forecasts create inventory mismatches.
• Piloting in a single market before global rollout reduces deployment risk and allows the model to be validated against a known baseline before scaling.
• Reducing manual overrides from 8% to 2% is not just an efficiency gain — it also removes planner-introduced inconsistency from the forecast, improving downstream supply chain decisions.