Here is the verdict first on AI delivery route planning: let AI draft the routes, and let veteran drivers fix only the exceptions. Neither side runs the whole show.
AI route planning and driver experience are not a choice between two options. They are a division of labor. I have seen this division work in three projects. And every project that let one side take over completely ran into trouble on go-live night.
What AI is good at, and what veteran drivers are good at
Draw the boundary first. If the boundary is unclear, everything after it turns into arguments.
AI is good at calculation. Twenty stops, three trucks, each with weight and volume limits, plus delivery time windows — the number of combinations is beyond what a human brain can work through. AI produces a shortest-distance first draft in minutes. When live traffic changes, it re-plans immediately. It never forgets loading constraints: overweight, over-volume, or mixing chilled and ambient goods — it enforces the rules every time.
Veteran drivers are good at the things that are not on any map. One residential compound bans trucks after 4 p.m. One customer needs a call to the gate guard, Mr. Wang, before unloading. One road looks open but floods every time it rains. One office building allows only 15 minutes at the unloading bay. None of this lives in any system. It lives in drivers' heads.
I have seen this pitfall three times: the project team sent out AI-planned routes untouched, and on day one three deliveries ran late. Every one of them was a "the system did not know" problem — a customer who wanted a call 30 minutes before arrival, a shortcut that is one-way during the day but which the AI treated as two-way. The AI did not miscalculate. The information was never in the input.
So the test is simple: anything that can be written as a rule goes to AI; anything that lives only in someone's memory stays with the driver. If you cannot tell those two apart, the project will fail.
Rolling out AI delivery route planning: four steps, no skipping
Step one, turn driver experience into rules. Pick the two or three drivers who know the routes best, ride along with each for two days, and write down every "road you cannot take, compound you cannot enter, customer you must call first." What can become a system rule goes into the rule base. What cannot becomes a note attached to the stop.
Step two, AI drafts, drivers fix only the exceptions. Every day's routes come from AI first. Before departure, drivers review them and change only the stops with problems, not the good ones. Every change needs a written reason. Review the reasons once a week. Any change that keeps recurring means a rule is missing from the rule base. Add it.
Step three, set a red line: if drivers change more than 30% of the stops, the AI rule base has a serious gap. Stop and fix the rules instead of forcing the trucks out. I set this number at a same-city delivery client. In their first week the change rate was over 40%. We stopped for half a day, added 11 rules, and in the second week the change rate fell below 10%.
Step four, review once a week. Look at three numbers: on-time rate, total mileage, and number of driver changes. On-time rate drops — check the rules. Mileage rises — check the constraints. Changes are frequent — check the inputs. Keep the review to 30 minutes, standing up.
Do the math: what is the saved mileage worth?
The table below is a hypothetical example, using volume levels common for a small-to-mid-size delivery fleet. It is a scenario illustration, not any real company's data:
| Item | Hypothetical example |
|---|---|
| Fleet size | 10 trucks, one trip each per day |
| Daily total mileage before AI optimization | 1,200 km |
| Daily total mileage after AI optimization | 1,080 km, a 10% saving |
| Mileage saved over 22 working days a month | 2,640 km |
| At 3 yuan per km all-in cost | about 7,920 yuan saved per month |
The saving is not only fuel and mileage. Drivers finish more than an hour earlier each day, late deliveries fall, and customer complaints fall with them. When I showed this math to a client, his exact words were "we should have done this last year." But one warning: the 10% saving is the number in this hypothetical example, not a promise. How much your fleet saves depends on how messy your current routes are — a fleet that already plans well might save only 3%.
Go-live night checklist — tick every box
I have seen this pitfall three times, always the same kind of miss on go-live night. So tick every item on this list. If one is missing, do not go live.
- Does the rule base include restricted hours and restricted roads? If not, do not dispatch.
- Does every stop carry the customer phone number and unloading notes? If the driver arrives and cannot reach anyone, the trip is wasted.
- Is someone reviewing the AI output? For the first month after go-live, someone must look at every route before dispatch.
- Is the route-change entry easy for drivers to use? If it is buried, drivers will not bother, and they will drive the wrong route instead.
- Who re-plans when something goes wrong — the dispatcher, or the system automatically? If responsibility is unclear, everyone waits for everyone else on the day it happens.
- Was historical data loaded? If the AI has never seen your last 3 months of orders, its routes are theory, not practice.
Once every box is ticked, watch for the three most common failure points.
The first: treating AI like a fortune teller. AI can only calculate from the data you give it. If the data has no road-closure information, the route it produces is wrong. When data quality is poor, the more accurate the AI is, the faster it goes wrong.
The second: drivers rejecting AI wholesale. Some veteran drivers think "I have driven 10 years, I do not need to be taught," and rewrite the whole draft without looking at it. Then the AI never learns anything, and three months later the project exists in name only. The fix is a "reason for change" field: drivers may change a route, but they must leave the reason. Those reasons are what make the rule base grow.
The third: removing human review on day one. The first two weeks after go-live are the most fragile. The rule base is still incomplete, and driver trust is not built yet. Someone must watch it for those two weeks, and only let go once the change rate stabilizes.
One last point: in AI delivery route planning, technology is only half the job. The other half is moving what is inside veteran drivers' heads into the system, a little at a time. The more you move in, the more accurate the AI gets. If you cannot move it in, the AI is just an expensive calculator. And the verdict again: AI drafts, drivers fix only the exceptions. Follow that division of labor, and three months from now your routes will run far more smoothly than they do today.
By Lin Xiao, WMS Implementation Consultant







