API Group reduces over-stock 8.5% and improves on-time delivery 11% with AI inventory optimisation

API Group, a large printing business operating across the US and UK, faced a demand planning problem driven by its largest customer. That customer's materials forecasts were frequently and dramatically inaccurate, forcing API Group to hold high stock — yet it still could not always meet delivery on time. The consequences compounded: capital tied up in warehouse stock, overtime costs to rush correct colours and styles into production, wastage from surplus inventory, and deteriorating customer satisfaction.

API Group engaged Kortical to own the full data science lifecycle — from raw historical data through to a deployed machine learning application. Kortical's AI Cloud platform was used to build a highly tuned time series forecasting model: feature engineering was handled via Kortical's SDK to generate the necessary lag variables, then the platform's AutoML engine generated and ranked tens of thousands of candidate models per week on a leaderboard, surfacing the highest-performing configuration. Optimisation techniques were then layered on top of the forecasting model to quantify the trade-off between two competing objectives — reducing overstock versus improving on-time delivery rate. Three distinct inventory strategies (lean, reliable, and balanced) were modelled. A human-in-the-loop UI allowed demand planners to overlay their own domain knowledge onto the ML forecasts; considerations such as storage capacity, supplier lead times, seasonal demand shifts, and upcoming campaigns can be encoded into the solution. The entire solution was delivered by a lean team of one data scientist and one domain expert.

API Group adopted the balanced inventory strategy, achieving simultaneous improvement across both dimensions that had previously been in direct tension:

• 11% increase in on-time delivery — improving fulfilment reliability and customer satisfaction
• 8.5% reduction in over-stock — releasing working capital previously tied up in warehouse inventory

For comparison, the lean strategy would have cut overstock by 35% but introduced significant delivery risk; the reliable strategy would have pushed on-time delivery up by 15.6% but reduced overstock by only 0.4%. The balanced approach delivered meaningful gains on both metrics without sacrificing either. The demand planning team adopted the solution directly through the forecast UI, enabling ongoing human-AI collaboration rather than a black-box handoff.

• Frame the problem as a trade-off, not a single objective — modelling lean, reliable, and balanced strategies shows the impact of reducing overstock versus improving on-time delivery.
• AutoML at scale accelerates model selection — generating and ranking tens of thousands of candidate models helps surface the best-performing model.
• Human-in-the-loop design matters — a UI that lets planners overlay domain knowledge on ML forecasts bridges the gap between algorithmic output and operational reality.
• Lean delivery teams are viable — a single data scientist paired with a domain expert is sufficient when the platform handles the heavy lifting of model generation and evaluation.
• Address the root data quality problem — when a customer's forecasts are the source of variability, building an internal ML forecast to try to beat them is worth testing.