Most networks re-plan once a day — the ones that win re-plan every time something changes
The Challenge
Operators of large fleets, multi-location inventory, and variable-demand labor plan against static assumptions or infrequent batch updates, and losses accumulate invisibly across thousands of daily micro-decisions. A 5,000-vehicle fleet re-optimized routes once a day and ran at 64% utilization while fuel was wasted on stale plans. A 500-store retailer's overnight inventory batch left a gap between stock events and the truth, producing phantom stock, stockouts, and fulfillment losses. A manufacturer spent 15 hours a week on messy, manual order preparation and revenue forecasting. A 100-site preschool operator, where labor is effectively the only variable cost, needed hours of manual data pulls to see which sites were overstaffed and could not manage labor tightly against mandated ratios.
The people responsible for these decisions are also working blind. Dispatchers, store managers, and site directors make hundreds of judgment calls a day on stale information, and intuition alone cannot process the number of variables involved. Finance sees the cost of the inefficiency only in aggregate, weeks later. The pattern recurs wherever demand is variable and resources are costly: hospital staffing, hospitality scheduling, retail and pharmacy inventory, distribution networks, and delivery ETAs.
- Stale daily plans — a 5,000-vehicle fleet re-optimized once a day and ran at just 64% utilization while fuel burned on routes that no longer matched reality.
- Phantom stock — an overnight batch left a 500-store retailer's systems out of sync with the shelf, producing stockouts in the gap.
- Manual forecasting drag — a manufacturer spent 15 hours a week hand-building order and revenue forecasts, leaving no time to act on them.
- Invisible labor waste — a 100-site operator needed hours of manual data pulls just to see which sites were overstaffed against mandated ratios.
The Approach
Machine-learning demand models are trained on historical and real-time data and paired with an optimization layer that acts on the forecast continuously. For fleets, a constraint-based engine combining genetic algorithms and reinforcement learning balances fuel, time, and delivery windows, recalculating only the routes hit by traffic, weather, or closures and pushing updates straight to drivers with predictive ETAs to the command center.
For retail, SKU-level demand forecasts per location drive replenishment while an AI routing layer picks the optimal fulfillment node for every order, moving inventory from overnight batch to sub-second latency. For labor, a model predicts hour-by-hour attendance and an optimization model recommends staffing that maximizes site-level margin within regulatory ratios; site leaders accept, override, or give feedback that retrains the model. In several cases the solution is pure classical ML and operations research with no generative AI involved, and because the models learn from the organization's own history, value typically appears within the first two months.
Demand forecasting (ML) · Constraint-based route optimization · Reinforcement learning · SKU-level replenishment modeling · Human-in-the-loop retraining
AI Role
AI models forecast demand (fuel/routing needs, SKU-level stock, hour-by-hour staffing) from historical and real-time data, and an optimization layer acts on that forecast automatically — rerouting, replenishing, or adjusting staffing continuously rather than on a fixed batch cycle. AI Automation Scope: Continuous, real-time optimization across an entire fleet, store network, or staffing model.
.webp)
