A national logistics company operating 15 distribution centers and a fleet of 800 vehicles faced two persistent problems: demand forecasting errors that caused either excess inventory or stockouts at distribution centers, and route planning that relied on static schedules rather than real-time conditions. Both problems were costing the company approximately $12 million annually in excess inventory, emergency shipments, and inefficient routing.

The Architecture

The system combines demand forecasting, inventory optimization, and dynamic route planning.

Demand forecasting - A time-series forecasting model predicts demand for each product at each distribution center for the next 14 days. Inputs include historical sales data, seasonal patterns, weather forecasts, promotional calendars, and leading indicators (web traffic, order pipeline data). The model produces point forecasts and prediction intervals, enabling risk-aware inventory decisions.

Inventory optimization - Based on demand forecasts and prediction intervals, the system calculates optimal stock levels at each distribution center. The optimization considers: service level targets (99.5% fill rate for critical items, 95% for standard), lead times from suppliers, transfer times between distribution centers, and holding costs. It generates replenishment orders and inter-DC transfer recommendations daily.

Dynamic route planning - The route optimization engine plans delivery routes considering real-time factors: traffic conditions, weather, vehicle capacity constraints, customer time windows, driver hours-of-service regulations, and priority deliveries. Routes are planned the evening before and re-optimized in real-time as conditions change during the delivery day.

Exception management - An LLM monitors the full system for exceptions that require human attention: demand forecast errors exceeding thresholds, stockout risks that cannot be resolved through normal replenishment, delivery delays that will miss customer commitments, and unusual patterns that may indicate data quality issues.

Key Lessons

Weather was the most impactful external signal for demand forecasting. Incorporating weather forecasts improved demand prediction accuracy by 15% for weather-sensitive product categories. A cold snap increased demand for heating supplies; predicted storms triggered pre-positioning of emergency supplies. The improvement was not from sophisticated weather modeling but from simple weather-category interaction features.

Inter-DC transfers were the hidden optimization opportunity. The biggest cost savings came not from better demand forecasting (which improved incrementally) but from optimizing transfers between distribution centers. The AI system identified that 20% of emergency shipments from suppliers could have been fulfilled from excess inventory at another DC at one-third the cost.

Driver adoption required co-design. The first route optimization system was technically optimal but impractical. Routes ignored driver preferences (familiar territories, lunch timing), rest stop availability, and practical constraints like narrow streets that could not accommodate large vehicles. Co-designing the system with drivers and incorporating their constraints produced routes that were 5% less theoretically optimal but had 95% driver compliance versus 60% for the purely optimal routes.

Real-time re-optimization had diminishing returns. Continuous re-optimization disrupted driver expectations and customer communication. The team settled on three optimization points per day: morning (pre-departure), midday (adjust for actual conditions), and emergency-only in the afternoon. This balanced optimization value with operational stability.

Results

Demand forecast accuracy improved from 72% to 88% (measured as MAPE). Excess inventory decreased by 25%, freeing $8 million in working capital. Stockout incidents decreased by 40%. Fleet utilization improved by 12% through better route optimization. Emergency shipment costs decreased by 55%. Total annual savings exceeded $8 million.