.png)
An AI forecasting platform cut route planning time from 6 hours to 22 minutes across a fleet of 340 vehicles operating across 14 distribution centres.
Overview
A regional logistics operator running 340 vehicles across 14 distribution centres in the food service and grocery delivery sector was planning daily routes manually using a combination of a legacy routing tool built in 2014 and dispatcher experience. Route planning started at 4am each day and took until 10am — 6 hours of planning before the first vehicle left a depot. The process required 22 dispatchers working simultaneously across the 14 sites, each producing routes using local knowledge and intuition rather than a consistent, optimised methodology.
On-time delivery performance was running at 81% — below the 92% contractual commitment the operator had made to its three largest grocery chain customers. Two of those customers had triggered penalty clauses in the previous quarter, generating $340,000 in contract deductions. A third had issued a formal performance improvement notice with a 90-day window before contract termination review.
Verttx built a unified demand forecasting and route optimisation platform integrating all 14 depot operations into a single planning environment. The platform went live 10 weeks after the initial discovery call. Daily planning time dropped from 6 hours to 22 minutes. On-time delivery rate reached 96% within the first full month of operation.
The core planning problem was that demand and routing were being treated as separate processes. Each morning, dispatchers would receive the day's confirmed orders — which continued arriving until 7am, creating a moving planning target — and then build routes from scratch using the legacy routing tool, which had no predictive capability and no integration with live traffic or weather data. Routes were optimised for distance, not for delivery time windows, vehicle capacity utilisation, or driver hours compliance under Department of Transportation regulations.
The result was systematic inefficiency across all four dimensions. Vehicle capacity utilisation averaged 67% across the fleet — vehicles were leaving depots with significant unused capacity because the manual planning process could not efficiently consolidate orders across proximity clusters. Fuel cost per delivery was $4.80 against an industry benchmark of $3.90 for comparable route density. Driver overtime was running at 22% above contracted hours, generating $1.2 million in annual overtime costs that were not present in the original contract pricing model. And on-time performance was 11 percentage points below contractual commitment.
The dispatchers were not the problem — they were experienced logistics professionals doing the best they could with inadequate tools. The problem was structural: no system that requires 22 humans to manually build routes for 340 vehicles in a 6-hour window can consistently produce optimised outcomes, particularly when demand patterns are variable and real-world conditions change between planning and execution.
The platform's foundation is a 72-hour rolling demand forecast covering order volume, geographic distribution, time window density, and vehicle type requirements for each of the 14 depot catchment areas. The forecasting model was trained on 38 months of historical order data covering 4.2 million deliveries, and incorporates 14 external signal categories including day-of-week and seasonality patterns, local school and public holiday calendars, weather forecasts from a commercial meteorological API, regional event schedules that affect traffic and demand, and promotional calendar data from the three grocery chain customers provided via a direct data feed.
Forecast accuracy at the 24-hour horizon reaches 91.3% mean absolute percentage accuracy across the depot network — sufficient to begin vehicle pre-positioning and driver scheduling 24 hours before order confirmation, which was the critical operational change that made the 22-minute planning window possible. Dispatchers no longer build routes from a standing start at 4am. They review and approve a system-generated plan that has been developing since the previous afternoon.
The routing engine solves a vehicle routing problem with time windows (VRPTW) — one of the computationally hardest classes of optimisation problem in logistics — using a hybrid metaheuristic combining adaptive large neighbourhood search with a machine learning heuristic initialiser trained on the operator's own historical route data. The ML initialiser generates a high-quality initial solution in milliseconds by recognising route patterns from similar historical planning scenarios, which the metaheuristic then refines. This hybrid approach produces routes within 3.2% of theoretical optimum for the operator's typical planning scenarios — a commercially viable approximation that pure metaheuristic approaches take 4-6 hours to reach on comparable problem sizes.
The engine optimises simultaneously across five constraints: delivery time window compliance, vehicle weight and volume capacity, driver hours under DOT regulations, depot departure sequencing, and fuel efficiency weighted by real-time traffic conditions sourced from a HERE Maps commercial routing API. Routes are dynamically re-optimised throughout the day as new orders arrive, cancellations occur, or traffic incidents are detected — producing updated plans within 90 seconds of any triggering event.
The platform was designed around dispatcher acceptance from the first design session. Dispatchers at all 14 depots participated in a structured discovery workshop where Verttx mapped their current decision-making process, identified the cases where they regularly overrode tool recommendations based on local knowledge, and built those override patterns into the system's constraint model. The interface gives dispatchers full visibility into the optimisation rationale for every route — why two stops were consolidated, why a vehicle was assigned to a specific run, why a time window was sequenced as it was. Dispatchers can override any system recommendation with a single action, and each override is logged and analysed to identify whether it represents a local knowledge signal the system should learn from or a dispatcher preference that does not improve outcomes.
Within six weeks of go-live, dispatcher override rates had fallen from 34% of system recommendations to 8% — indicating that the system's outputs had earned operational trust across the depot network.
Daily planning time across the 14-depot network fell from 6 hours to 22 minutes. The 4am planning start was moved to 5:30am. Dispatchers now review and approve system-generated plans rather than building routes from scratch — a change that reduced planning headcount requirements from 22 dispatchers to 14 across the network, with the 8 redeployed to exception management and driver support roles that had been under-resourced.
On-time delivery performance reached 96% within the first full month — 4 percentage points above the contractual commitment and 15 points above the pre-implementation baseline. All three grocery chain customers formally acknowledged the improvement. The customer who had issued the performance improvement notice withdrew it at the 90-day review. The two customers who had triggered penalty clauses in the previous quarter issued no deductions in the first two quarters post-implementation, recovering $340,000 in contract value that had been at risk annually.
Vehicle capacity utilisation improved from 67% to 84% — allowing the operator to serve the same delivery volume with 28 fewer vehicle movements per day, directly reducing fuel consumption. Fleet fuel costs fell by 18%, producing annual savings of $1.14 million against the pre-implementation baseline. Driver overtime fell from 22% above contracted hours to 6%, reducing annual overtime costs by $890,000. The combined operational saving from fuel and overtime reduction alone returned the full platform investment in 4.2 months.
The forecasting models, the route optimization engine, all depot integrations, the dispatcher interface, and the full monitoring infrastructure were transferred to the operator's engineering team at handover — the system is owned and operated entirely by the client, with no dependency on Verttx for daily operations or future development.
We had experienced dispatchers doing their best with a tool that was never built for the scale we were operating at. The routes the platform generates in 22 minutes are better than what we were producing in 6 hours. The on-time rate speaks for itself — we went from defending ourselves to customers to being ahead of our commitments. — Head of Fleet Operations, Regional Logistics Operator
.png)
Daily route planning time fell from 6 hours to 22 minutes across 14 depots. On-time delivery rate improved from 81% to 96% within the first full month — 4 points above contractual commitment. Vehicle capacity utilisation improved from 67% to 84%. Fleet fuel costs fell 18%, saving $1.14 million annually. Driver overtime costs fell by $890,000. The platform returned its full investment in 4.2 months. The customer who had issued a performance improvement notice withdrew it at the 90-day review.
.png)
.png)
