What does the procedure do?
A description in sentences, not in model names: which decision is being prepared, on which data, within which limits.
Anyone using AI in dispatch must be able to answer two questions: who is liable when it gets it wrong — and what does it expressly not do? The second answer is the more credible one.
These four questions now appear in almost every tender. Whoever cannot answer them is not screened out because of the AI, but because of the missing information.
A description in sentences, not in model names: which decision is being prepared, on which data, within which limits.
At which point does a human step in, and can they overrule the result without bypassing the system?
Which proposal arose when and on what basis — and whether that can still be reconstructed afterwards.
Who carries the decision when the proposal was wrong: the vendor, the operator, the user? That is a contractual question, not a technical one.
Traceability is not a documentation task at the end, but a design decision at the start. Three properties carry almost all the requirements.
Radius, deadline, quota, price step: where a comprehensible rule decides, the explanation is the rule itself. A model explains itself less easily.
Every value carries where it came from — measured, calculated or from a model. That is half the transparency obligation at the same time.
What is autonomous is the preparing and the asking. Whatever binds is confirmed by a person — and that step is in the log.
Most procedures in transport planning are not high-risk applications within the meaning of the Regulation: they are about loads, vehicles and prices, not about people. Two areas nevertheless deserve a close look — anything that assesses employees, and anything that automatically excludes business partners from orders.
And regardless of the classification, the transparency obligation applies: whoever speaks to a system must know that it is not a human. That is little effort and a lot of trust — and it is the point at which good vendors decided early.
This page is a practitioner's assessment and does not replace legal advice. What it is meant to do: ask the right questions early, while the answers are still cheap.
Three deadlines are set. Each is listed here with the legal act it rests on.
As at 27 August 2026. Deadlines shift — the ETS2 start was once set for 2027. What stands here is the direction, not legal advice: anyone basing a decision on it should check the current status first.
Proposing yes, deciding no. An award commits money and carries a name — it belongs to a person, not in a model. What happens when this refusal is missing is worked through on a reverse auction.
A prediction that nobody holds against reality is decoration. We deliver the follow-up measurement with it, or we do not deliver the forecast.
A calculated position goes in a field of its own. How it looks on the outside is a display question; that the program knows the difference is not.
Every value going out carries its method and source with it. Model values are marked as such — even when a round number would look better.
Whoever has not committed does not get planned in. A chain built on an uncommitted carrier falls apart with the first refusal and blocks loads until then.
A model that does not beat the existing planning is not progress — even if it sounds more modern and presents better.
Every one of these refusals has been paid for dearly once — by us or by someone we watched. An estimated position that passed as a measured one made an empty run of over a thousand kilometres look like twelve. A check that reported “passed” on an internal error did the wrong thing unnoticed for days.
That is why there is a list of noes here instead of a list of yeses. Capabilities can be demonstrated; limits hold even when nobody is watching.
We go through the points that come up in a tender — and say where we say no.