Since 2024, an interesting split has appeared in regional delivery fleets. On one side, AI route optimization systems are spreading fast, with vendor pages promising shorter mileage and less time on the road. On the other side, dispatchers with a decade or more on the job look at the routes the system produces and shake their heads — the route looks shorter on paper, they say, but it won't necessarily run faster. The disagreement is worth taking seriously, because the real question isn't whether the technology works. It's what route optimization actually saves, and whose money it saves.

Start with a rough calculation built on stated assumptions. Take a regional fleet running 40 deliveries a day with 12 trucks, each truck averaging 180 kilometers daily. If AI route optimization genuinely cut mileage by 8 percent, the fleet would save roughly 173 kilometers a day, or about 4,500 kilometers a month on a 26-day schedule. At a rough per-kilometer operating cost, that adds up to several thousand dollars a month — a meaningful sum over a year. The problem is that the 8 percent figure comes from vendors measuring under ideal conditions. Once real-world variables enter the picture — unloading waits, customer receiving windows, last-minute orders — the math gets much less flattering.

Route Optimization on Paper Isn't Always Fastest on the Road

In the second half of 2024, a regional distributor in East China rolled out an AI dispatch system. The routes it produced were indeed shorter on paper: total mileage came down, and theoretical driving time came down with it. But after three months, the fleet's on-time rate had actually dropped. The post-mortem was instructive. The system optimized for map distance but never accounted for unloading waits. Customer A's warehouse only accepted two trucks in the morning, and the system scheduled three in the same window — the third truck waited over an hour at the gate. Customer B's receiving window ran from 2 to 3 p.m., but the system routed a truck there in the morning to save mileage, and the driver simply had to wait.

This points to a key distinction. AI route optimization optimizes distance on a map. A veteran dispatcher optimizes certainty inside time windows. What the dispatcher carries in their head isn't a map — it's which customer can unload at what hour, which warehouse gate jams up, which road is impassable after 3 p.m. No system generates that knowledge on its own. It all has to be fed in, entry by entry, by a human. Until that feeding happens, the more "optimal" the system's routes look, the more waiting time they may create in practice.

Another case makes the point even better. A dispatcher with 12 years on the job once swapped the order of two customers whose receiving windows conflicted. The system flagged the new route as a detour — a dozen extra kilometers. But after the swap, neither truck had to wait at a customer gate, saving 40 minutes of idle time between them. Mileage went up, time went down, and one driver even squeezed in an extra delivery that day. By mileage the decision lost money. By time it made money. The trouble is that many AI route optimization scorecards only measure mileage, not waiting.

Getting the Experience Into the Route Optimization System, in Three Steps

That brings us to a practical question: how does the knowledge in a veteran dispatcher's head actually get into an AI system? From the fleets I've seen do this well, it comes down to three steps.

Step one is structuring it. Don't try to do it all at once. Start by writing down the three most valuable kinds of information: how long each customer takes to unload, each customer's receiving window, and which stretches of road jam up at which hours. The dispatcher spends ten minutes at the end of each day noting the day's exceptions. After a month, that's a solid local knowledge base.

Step two is validating. Don't roll the system's routes out to the whole fleet at once. Have a veteran dispatcher ride along with the new routes for two weeks, scoring them daily against what actually happened on the road. Keep what the system got right, flag what it got wrong — and every flag is more data to feed in. The worst mistake at this stage is trusting the system blindly and being afraid to correct it when it goes off track.

Step three is keeping the people. After the system goes live, the dispatcher role doesn't disappear — it changes. Instead of planning routes every day, the dispatcher reviews routes and fills gaps in the data. The outcome every fleet should fear is this: three years after go-live, the veteran has retired, the new hire only knows how to click a mouse, and nobody remembers how the road is supposed to be run. Technology can be upgraded. Judgment has to be handed down by a person.

When to Trust the System, When to Trust the Veteran

Based on what I've observed over the past two years, the dividing line is actually fairly clear. Here's a framework for making the call:

Situation The AI system fits better The veteran's experience fits better
Routes Fixed routes, high volume, many stops Frequent last-minute orders, routes change daily
Customer time windows Flexible windows, adjustable Strict windows that conflict with each other
Road complexity Simple network, predictable congestion Construction, restrictions, seasonal congestion
Data foundation Customer and unloading-time data already structured Know-how still in people's heads, not in the system

This framework points to a counterintuitive conclusion: the greatest value of AI route optimization may not be producing more optimal routes, but turning the veteran dispatcher's knowledge into an asset the whole fleet owns. The real gains appear after dispatchers feed in, entry by entry, the kind of knowledge like "which customer can unload at what hour." And the biggest risk runs in the opposite direction — not that the system calculates a wrong route, but that the fleet gradually loses the ability to plan routes at all. When the veteran retires, the experience leaves with them, and the system is left holding an algorithm that was never properly fed.

So what does route optimization actually save? The honest answer depends on where your fleet stands. For a fleet with fixed routes and complete data, the mileage savings are real. For a fleet with complex time windows and experience that hasn't yet entered the system, the first thing to save is waiting time, not mileage. Vendor-advertised savings figures can serve as a reference, but treating them directly as your own expected return is a good way to be disappointed.

The real question in this contest isn't who wins. Systems will keep getting stronger — that much is a trend. But whether the experience survives is the fleet's own choice. AI route optimization and the veteran dispatcher aren't an either-or question. They're a question of how to get the second into the first. Answer that well, and what you save isn't just fuel money — it's the thing a fleet can least afford to lose: people who know how the road should be run.

By Dr. Chen, Supply Chain Researcher