AI in transport planning: what a mid-sized transport company can do with it

AI in transport planning works best as a suggestion engine: every morning the system draws up a first trip plan based on orders, time windows, driving hours and what the on-board computers measured yesterday, and the planner reviews it, moves things around and decides. The gain is in fewer hours spent on standard trips and a few percentage points fewer empty kilometres. It does not save you a planner, and that is not the goal either. For a transport company of fifty to three hundred and fifty people, planning is the second step and not the first: start with the document flow, and move on to planning once your people understand why that first step works.

This piece is about that second step. What a planner actually does in a day, what software can take over and what it cannot, which data you need, how to measure whether it pays off, and what happens to the knowledge that currently sits in your planners' heads.

Why the empty kilometre does not disappear by itself

Statistics Netherlands (CBS) tracks how many kilometres Dutch trucks drive and how many of those are loaded (table 83077NED). In 2025 they drove 6,685 million kilometres in total, of which 4,946 million carried cargo. That means 26.0 percent of all kilometres were empty. In 2019 it was 26.4 percent. For domestic haulage for hire or reward it is around 30 percent, and that figure has barely moved over the same six years.

Route optimisation has been around for much longer than six years. That says nothing bad about the software. It does say something about where the problem lies. An empty kilometre rarely happens because a route was badly calculated. It happens because no return load was known at the time of planning, because a customer moved their time window, because a driver had reached their driving hours, or because the order that would have completed the trip came in at half past three. Those are planning decisions under uncertainty, and that is where the planner sits.

Planning at a screen on the work floor. Most empty kilometres arise at the moment of planning.

Who this piece is for

Anyone searching for AI in transport planning mainly finds vendors of planning software explaining what their package can do. That is useful if you already know you are going to buy a package. It does not help with the question an owner or managing director has to answer first: is my organisation ready for this, what does it deliver, and what happens to my people?

According to CBS (table 81589NED), the Netherlands has almost 20,000 companies in road freight transport in the third quarter of 2026. The vast majority are one-person businesses. About 390 have fifty or more people. That is the group this piece is written for: large enough for its own planning department with two to five planners, a TMS and on-board computers in the trucks, and too small for its own data department.

In that group, AI is still rarely used for planning. In December 2025, CBS reported that of the companies using AI, 35 percent do so for marketing or sales and 4 percent for logistics purposes. In the transport and storage sector, seven percent of companies used any form of AI at all in 2025. If you start carefully now, you are not behind. You have time to do it properly.

What a planner actually does in a day

On paper, building a trip plan looks like a calculation: this many orders, this many trucks, this many addresses, find the shortest combination. In reality most of the day is something else. A planner guards rules, reads people and catches exceptions.

Driving and rest times. Under the rules as summarised by the Dutch government, a driver may drive at most 9 hours a day, 10 hours twice a week, at most 56 hours a week and 90 hours per two weeks. After 4.5 hours of driving comes a 45-minute break, and daily rest is at least 11 hours, which may be reduced to 9 hours three times between two weekly rest periods. Software can calculate this. What software does not know is that a driver stood at a loading point for two hours yesterday and therefore starts today differently from what the system says, unless the on-board computer passed that on properly.

Customer time windows. On paper a customer has a window from 7:00 to 11:00. In practice the planner knows you have to be there before half past eight because the receiving bay is full after that, and that another customer never meant their window that strictly and is fine with an hour later.

Rush orders and late changes. The order that comes in at quarter past four for tomorrow morning, the customer who cancels a load while the truck is already on the road, the loading point that calls to say the goods will not be ready until two o'clock. This is the moment a plan either comes together or falls apart.

Charters. When do you hire a charter, and from whom? That depends on price, reliability, and whether your own truck earns more elsewhere tomorrow. That trade-off is usually written down nowhere.

People. Which driver knows the customer with the awkward entrance, who wants to be home on time on Fridays, who prefers not to drive abroad. That is not sentiment. A driver who is satisfied stays, and in a market where CBS still counted 7,800 open vacancies in land transport in the second quarter of 2026 (table 80472NED), that is worth money.

The left column is the part that takes a lot of time and needs little judgement. The right column is the part you hired a planner for. A good system takes over as much of the left column as possible, so that more time is left for the right one.

When you are ready for the planning step

Our piece on AI in logistics sets out the order we recommend: document flow first, then planning. There are two reasons for that. With documents you can check the output yourself, so your people learn how it works and where it goes wrong without a customer waiting on it. And planning touches the most sensitive part of your business. A mistake on an invoice you fix tomorrow. A mistake in the plan is standing at a customer's closed door tonight.

You are ready for the planning step when you can answer yes to the following four questions.

Does the first application run without anyone having to watch it? If the document flow still needs daily attention, nobody has room for something new.

Can your own people explain why that first application works? This is about your own people, the vendor does not count here. If the answer is that it just works, no capability has been built yet, only a subscription.

Do you record actual times, and not only planned times? A model learns from what really happened. If your TMS only knows the plan and not the actual arrival, departure and waiting time, it learns from your own assumptions.

Is there a planner who wants to own this? Someone who wants to think along about what the system suggests, and who is given the time to do so. Without that person, you do not start.

If you are not there yet, that is fine. It means the first step is not finished.

Which data you need

Almost every transport company of this size already has the data. The problem is that the sources do not connect. For a suggestion engine you need four things.

Order history from your TMS. Preferably two to three years, so that seasons, public holidays and the departure or arrival of a large customer are in it. A model trained on six months of data only knows the half year it was trained on.

Actual times from the on-board computer or telematics. Arrival, departure, waiting time per loading and unloading address, and the route actually driven. This is where the real value is, because this is where it shows that one particular loading point always runs an hour late.

Driving hours per driver. From the tachograph or the system that reads that data. Without it, a system suggests trips that are not allowed.

What the planner changed, and why. You probably do not have this now, and it is the most important of the four. Every time a planner adjusts a suggestion, that is information about something the system did not know. Record that correction with a short reason.

You do not need to buy a new system or replace your TMS for this. What you do need is to write down what each field means and which source wins when two systems contradict each other. How to approach that without a migration project is described in getting your data in order for AI.

Two trucks on the motorway. What the on-board computer measures en route is worth more than what the plan assumed.

What to automate first

Start small. The most tempting idea is a system that builds the plan overnight and has it ready in the morning. It is also the idea on which most projects get stuck, because the planner does not trust the output, checks everything again and ends up spending more time than before.

What does work is building up in three phases. Each phase is reversible: if it does not work, you simply plan the way you plan now.

Phase one: running in the shadow. For a few weeks the system makes a plan every day, and the planner plans as always. Nobody uses the suggestions. At the end of each day you put them side by side. Where do they differ, and who was right? This costs almost nothing and gives you the most honest test there is.

Phase two: suggestion as a starting point. The planner starts the day with the suggestion instead of an empty screen, and adjusts what is wrong. Every adjustment is recorded with a reason in a few words: customer wants earlier, driver knows the address, loading point runs late. You measure how much of the suggestion stays as it is.

Phase three: standard trips without intervention. Only when the suggestion for a certain type of trip has stood unchanged for weeks do you let that type of trip be planned automatically. Fixed trips to fixed customers on fixed days are the logical candidate. Everything else stays a suggestion.

A full autopilot has no place in this plan. The planner keeps the controls, also in phase three. More on why planning is different from other processes in this respect can be found in AI in operations planning.

A worked example to fill in with your own numbers

This example is made up and is meant as a method of calculation. Replace every number with your own company's.

Suppose a transport company with 80 trucks that each drive an average of 110,000 kilometres a year. That is 8.8 million kilometres a year. If 26 percent of that is empty, equal to the national figure, that comes to 2,288,000 empty kilometres.

Suppose better planning lowers the share of empty kilometres by one percentage point for the same work. That is 88,000 fewer kilometres a year. At 0.90 euros of variable cost per kilometre, for fuel, tyres, maintenance and tolls, that is 79,200 euros a year. Use your own cost per kilometre here, because it varies widely between companies.

Then the planning hours. Suppose three planners each spend one hour less per day on standard trips. Over 230 working days that is 690 hours a year. Those hours do not come off the payroll, and they should not. They go to rush orders, finding return loads and customer contact, the work where a planner earns money.

One percentage point sounds modest, and that is deliberate. Do not count on a drop from 26 to 10 percent. Part of the empty kilometres arise because cargo happens to be in one place and the truck needs to be in another, and no planning solves that.

How to measure whether it pays off

Start measuring before you change anything. Three months of baseline is enough to see the normal fluctuations. After that, measure the same points in the same way.

Look at all six at once. Fewer empty kilometres with more late deliveries means you have shifted the cost to your customer. And measure per type of trip, not only for the company as a whole, because an improvement on fixed trips can be cancelled out by a bad week in rush orders.

What happens to the planner's knowledge

This is where the resistance sits, and that resistance is justified. A planner with fifteen years of experience knows things that are written down nowhere. If a system is presented as something that will take over their work, they will distrust it, with good reason. They also know they are the one who has to fix it when it goes wrong.

It works the other way round if you position the planner as the person who teaches the system what it does not know. Every adjustment with a reason is a piece of knowledge that used to sit only in their head. After a few months there are hundreds of those reasons in the log. The system can use part of them. The rest becomes readable for a new planner, for a stand-in during the holidays, and for you as the owner.

That last point is a risk many companies underestimate. If your best planner leaves tomorrow, how much of the planning goes with them? A suggestion engine with a well-kept log makes you less dependent on one person, without that person becoming less important. How to hold on to that kind of knowledge structurally is described in anchoring AI knowledge.

Two agreements make the difference. The first: the time that is freed up stays with the planner and does not disappear into a new norm of more orders per hour. Otherwise cooperating well gets punished. The second: the planner is at the table from the first week, and so does not join only when something needs testing. Why that matters so much with existing staff is explained in getting more out of the people you already have.

Where it goes wrong in practice

The system learns from the plan instead of from reality. If only planned times are recorded, the model repeats your own mistakes with more confidence.

The adjustments disappear. A planner who sees that their corrections are needed again tomorrow stops correcting. Then the system loses exactly the information that should have made it better.

Moving to phase three too early. A few good weeks in summer say little about December. Only let a type of trip be planned automatically once the suggestion has also held up through a busy period.

Nobody inside the company understands the model. An outside party can build it. If that party leaves and nobody can explain why the system suggests a trip, you come to a standstill at the moment it matters.

Where to start

Which planning package you choose comes later. First you want to know where your organisation stands now: how far along the document flow is, whether your data connects, and whether there are people who can carry this. The AI-Foto is a free self-scan of ten minutes and twelve questions. Afterwards you see where your organisation stands, and whether planning is next or something else needs to happen first.

About this page

The figures on empty kilometres come from CBS table 83077NED (road transport, vehicle kilometres total and loaded, 2019 to 2025). The number of companies in road freight transport comes from CBS table 81589NED (third quarter of 2026), the vacancies from CBS table 80472NED (second quarter of 2026). The figures on AI use come from the CBS news release of December 2025 on its ICT survey. The driving and rest times are as summarised by the Dutch government; the full rules and exceptions are set out in the European regulation. The worked example is made up and intended as a method of calculation, not a promise. This is the situation as of 5 October 2026.