Logistics teams hear "AI route optimization" and picture autonomous dispatch with no humans in the loop. Reality on the dock floor is different: planners need explainability, drivers need reliable mobile workflows, and customers need exception handling when a truck breaks down.
This article distills what worked on a regional freight engagement — 180 daily loads, spreadsheet dispatch, 42% cost-per-load improvement — and what we would not recommend despite vendor slide decks.
Start with constraints dispatchers already respect
Vehicle routing with time windows (VRPTW) only helps if the model encodes real constraints: hours-of-service, dock appointments, preferred drivers, equipment type, and customer blackout windows. Optimizers that ignore these get ignored by dispatchers.
We shadowed planners for ten days before tuning weights. Trust beat theoretical optimality.
Explain every recommendation in plain language
"Why this truck?" must have an answer: return-leg pairing, backhaul opportunity, HOS buffer, appointment slack. Black-box routes die in the first week. UI copy matters as much as OR-Tools configuration.
Human-in-the-loop beats full autonomy
AI drafts the board in minutes; humans approve exceptions, customer promises, and carrier relationships. Full autonomy sounds futuristic until weather closes a hub and someone needs to rewrite twelve routes before drivers arrive.
Automation should remove spreadsheet rebuilds — not remove judgment.
Measure cost per load weekly, not vanity AI metrics
Track empty miles, on-time delivery, spot-market response time, dispatcher hours saved, and fully loaded cost per mile. Model accuracy is intermediate; business outcomes are the scoreboard.
Ferro Logistics saw ROI inside one peak season because finance and ops agreed on metrics before launch.
Key takeaways
- Encode real dispatch constraints — HOS, docks, drivers — or planners will ignore the tool.
- Explain route decisions; trust beats opaque optimality.
- Keep humans approving exceptions; automate spreadsheet rebuilds.
- Score on cost per load and empty miles, not model accuracy alone.
Logistics AI automation succeeds when dispatchers feel faster, not replaced. If you are evaluating route optimization software, ask vendors how they handle explainability and exception workflows — not just demo map animations.