AI for logistics routing and fleet optimization

AI route optimisation generates meaningful ROI at fleet sizes of 10 vehicles or more for high-frequency delivery operations (daily multi-stop routes). The economics improve significantly above 25–30 vehicles where dispatch complexity justifies the platform cost and where the aggregate fuel and time savings exceed subscription costs with comfortable margin. For fleets below 10 vehicles or operations with simple, repeated routes, the optimisation benefit versus a good manual planning process may not justify enterprise platform costs — evaluate simpler tools or per-route pricing models.

This is a genuine limitation of pure algorithmic optimisation. Experienced drivers know that the back entrance to a particular customer saves 15 minutes versus the mapped front entrance, that a specific road floods in rain, or that a customer is only reachable by a smaller vehicle despite map data suggesting standard access. Modern AI routing platforms address this through driver feedback mechanisms — reporting route data issues, marking custom waypoints — that feed corrections back into the routing model. The best implementations treat driver knowledge as training data rather than competition with the algorithm. Over 6–12 months, AI models incorporating driver feedback typically outperform those without it by 8–12% on delivery time accuracy.

Integration capability varies significantly by platform. Enterprise platforms like Oracle TMS, SAP TM, and Manhattan Associates provide native bi-directional integration with major WMS systems. Mid-market platforms like Locus and OptimoRoute provide REST APIs suitable for custom integration with most modern WMS platforms. Legacy TMS systems with limited API surfaces are the most common integration challenge — they may require ETL batch processes rather than real-time API integration, reducing the effectiveness of dynamic re-routing capabilities that depend on real-time order data. Assess your TMS/WMS integration surface in platform evaluation.

Effective predictive maintenance requires: OBD-II/J1939 telematics data from each vehicle (engine RPM, coolant temperature, fuel consumption, D2C codes), mileage and operating hours, maintenance history (what was replaced and when), and ideally component-level sensor data for critical systems (brake wear sensors, tyre pressure monitoring). Minimum viable datasets start with telematics and maintenance history — models trained on these can predict 60–70% of major component failures with 2–4 week advance warning. Additional sensor data improves accuracy. The AI models require 12–18 months of fleet-specific data to achieve reliable failure prediction for most component types.

Same-day and on-demand delivery requires continuous re-optimisation as new orders arrive throughout the day — a dynamic vehicle routing problem where the order set is not fixed. AI optimisers for this scenario use online algorithms that insert new orders into existing routes with minimal disruption to scheduled deliveries, dynamically re-assign orders between drivers based on proximity and capacity, and predict demand patterns to position drivers in high-likelihood order zones before orders arrive. Platforms purpose-built for on-demand logistics (Onfleet, Bringg, Locus) handle this more effectively than route planning tools designed for batch daily planning.

The most common implementation challenges are: data quality (incomplete order records, inaccurate customer addresses, missing historical stop times); driver adoption (resistance to algorithm-generated routes, especially from experienced dispatchers who view AI as threatening their expertise); integration complexity with legacy TMS/WMS systems; and initial model calibration periods where AI recommendations are less accurate than experienced human dispatchers until the model has accumulated sufficient operational data. The human change management challenges consistently outweigh the technical challenges in retrospective analyses of logistics AI deployments.

Carbon reduction from AI route optimisation is calculated from fuel consumption reduction data using DEFRA or EPA emissions factors for the relevant fuel type and vehicle class (typically 2.68 kg CO2e per litre of diesel for heavy vehicles). A 15% fuel reduction for a 50-vehicle fleet running 200km/day on diesel translates to approximately 290 tonnes CO2e annual reduction — a measurable contribution to Scope 1 emission reduction targets. More sophisticated carbon accounting also captures the maintenance benefit (fewer breakdowns = fewer recovery vehicle trips) and optimised load utilisation (more deliveries per vehicle trip = lower per-delivery emissions intensity). This data is increasingly required for customer sustainability reporting and ESG disclosures.

Territory optimisation is the strategic-level problem of defining how to divide a service area into delivery territories assigned to specific vehicles or drivers — a prerequisite for effective daily route optimisation. AI territory optimisation uses historical delivery density data, driver home locations, customer service level requirements, and road network analysis to design territories that balance workload, minimise inter-territory crossover, and keep drivers close to their home base. It is most valuable when expanding into new geographies, restructuring after acquiring delivery operations, or rebalancing territories that have become uneven through organic growth. Re-running territory optimisation annually typically recovers 5–10% efficiency gains lost to organic imbalance accumulation.