Agentic AI in fleet operations: what actually works

We spent a year putting AI agents into live fleet operations — dispatch, maintenance, driver communication. Here is what survived contact with real drivers, real trucks and real Monday mornings.

Start with the phone calls

Every fleet we looked at ran on the same hidden system: phone calls and chat messages between dispatchers and drivers. Where is the truck, why is it late, who takes the extra stop. The first agent that earned its place did nothing clever. It answered those questions from live telematics before a human had to ask them.

That taught us the rule we now apply everywhere: an agent is worth deploying when it removes an interruption, not when it demonstrates intelligence.

Three agents that stayed

• Dispatch assistant. Proposes reassignments when a trip slips, with the reason attached. The dispatcher confirms with one tap or ignores it. Acceptance rate is the metric, not accuracy.
• Maintenance watcher. Reads inspection results, fault codes and service schedules, and opens a work order draft with the parts list. Nobody wanted it to book the workshop on its own, so it doesn’t.
• Driver comms. Turns a plan change into a short message in the driver’s language, and turns the driver’s voice note back into a structured status. This one had the fastest adoption, because drivers hate typing.

Two that didn’t

A fully autonomous route optimiser. It produced better routes on paper and worse days in practice, because it did not know what the dispatcher knew about that customer, that gate, that driver. We kept the optimiser and put a human back in front of it.

A chat interface over the whole platform. People asked it one question, got a good answer, and went back to the screen they trusted. The screen is the product. The agent should live inside it.

Fewer phone calls. Fewer surprises. One screen that tells you the truth about your operation.

What we would tell a fleet starting today

Instrument first. An agent is only as good as the events it can see. Pick one interruption that happens fifty times a day and remove it. Measure minutes saved per dispatcher, not model scores. And keep the human in the loop wherever a wrong action costs money, fuel or trust.

By Bassem Fawaz · #agentic-ai #fleet-operations #cooriroo