What AI does today — and what it does not
AI in logistics is usually written about in the future tense. This is what actually runs.
The classic case: an order arrives by mail
It happens in every company, and everything else hangs off it: whatever enters the system wrongly here only surfaces in the invoice.
- Mail in the inboxA customer writes the way they write: prose, PDF, attachment, sometimes a call.
- Fields recognisedSender, loading point, date, weight — from any format, without a per-customer template.
- Checked against master dataDoes the company know this delivery point? Does the date fit the agreed window?
- Question back when unclearAn email back instead of a guessed line. This is where it is decided whether the order is right.
- Order in the systemComplete and checked — and every field says where it came from.
- Confirmation outThe customer knows what is booked, without having to ask.
What the case really costs
Not the two minutes of typing. What is expensive is the wrong delivery point running through three stations until goods-in notices.
Why the question back is the core
A model that guesses produces cases nobody reports. One that asks produces an email.
The order arrives as a mail, a PDF, an attachment — every customer writes differently.
What AI changes in logistics — and what it does not
The honest answer is not “everything” but a list: what moves is where somebody reads, compares or retypes — not where somebody answers for a decision.
| Step | How it runs today | What AI makes of it | How you measure it |
|---|---|---|---|
| Order intake | Mail, PDF, attachment, sometimes a phone call. Every customer writes differently — one rule each, or a person. | Reads every format, maps the fields and asks instead of guessing. The question back is the part that counts. | Time from arrival to a finished order in the system. |
| Planning | Experience, phone, spreadsheet. Good — but not explainable, and away on holiday. | A proposal with its reasoning. The human decides; the difference goes back into the model. | Empty running and orders per trip — worked through. |
| Consignment note after upload | The carrier uploads it, a person looks at every one: damage noted, signature there, whose? | Reads it, spots noted damage and holds the signatory against the carrier that was booked — which catches a hand-off to a sub. | Days from upload to the credit note. |
| Billing and credit notes | Matching has run automatically for years. The catch sits higher: every new case needs its own workflow — written, tested, maintained. | The model builds the workflow itself. The step where the expensive part ends. | Time until a new case runs through. |
| Metrics | The report shows the figure. What follows, everybody works out — or nobody does. | The reading stands next to the number, with its conclusion and its recipient. | How many reports lead to an action — drag it. |
“Can” is not “does”: order intake, matching and the reading are running. The damage check on the note is still a person’s job; the hurdles are handwriting and sloppy paperwork. The self-built workflow is the next step.
What changes is the reading, the comparing and the preparing. What does not change is who signs.
Today a person reads, compares and retypes at five points.
Three things work, three limits come with them
Three capabilities that hold up today — and for each the line where a human stays.
An agent without boundaries is not a more capable agent. It is one whose mistakes you notice only once they are posted.
Reading documents
Consignment notes, invoices, order emails — even without a known layout. Limit: handwriting. A human guesses there too.
Spotting anomalies
The trip outside the pattern. The customer whose behaviour tips. Limit: the model finds. Why, comes from a phone call.
Making suggestions
Which carrier, which sequence, which price. Limit: useful as a suggestion, a liability problem as a silent decision.
A carrier invoice arrives — no known template, free layout.
Agentic AI in two sentences
The term gets used for anything that does more than one step. What it means fits in two sentences — the second is the more important one.
What it means
The model does not execute one step but assembles a sequence itself, uses tools and checks whether it is making progress.
What it does not mean
That the machine chooses the goal. Which tools, which values, where to stop — the frame still comes from outside.
Legal framework: the EU AI Act has been enforced since 2 August 2026 and the Article 50 transparency duty applies; high-risk was postponed to late 2027 by the Digital Omnibus. What counts now: knowing which system runs where. More under Future.
A shipment needs a carrier. The goal is set by the dispatcher — not by the model.
Worked example: the consignment note after delivery
The carrier uploads it, nothing gets retyped. Three questions are interesting — the third finds what nobody looks for.
What the machine can do
Read it — reliable on clean paperwork. Detect damage — when the remark is in the field. Check the signatory — a name other than the booked carrier means it was passed on.
Context
The third is the sharpest: nobody finds it while leafing through, because nobody is looking. The damage check itself is still done by a human today.
The consignment note triggers money both ways: once released the sub-carrier gets credited — and the customer usually pays only once it is on file.
The carrier uploads the consignment note. Nothing is retyped — nor was it before.
Where the best solution sits today
What is named is the pattern, not the vendor — a product name ages, an idea does not.
Bounded autonomy
The pattern that wins: clearly defined authority, a clean escalation path. Fully autonomous systems mostly do not scale.
Feedback, not output
Hold every result against something — note against order, recommendation against the closed deal. Without that nothing learns, it only emits.
Supply the source
A figure with a provenance can be checked; one without has to be believed. That is the most effective answer to hallucination.