Which forwarding tasks can be automated today?

Tasks that repeat, rest on text or a document and have a clear correctness criterion: reading orders and transport documents, classifying and attaching them to the right shipment, chasing routine status updates, preparing recurring reports. What they share is that a human reviews the result anyway.

What will AI not do in freight forwarding?

It will not decide which carrier to trust with a load under time pressure, judge whether a carrier who was late twice deserves a third chance, or negotiate a rate. In those tasks the cost of a mistake is asymmetric and the inputs contain knowledge nobody ever wrote down.

Will AI replace the forwarder?

Not within any foreseeable horizon. It removes the administrative layer that has grown around commercial decisions, not the decisions themselves. In practice it changes how many shipments one person handles rather than how many people you need.

How do I work out whether automation pays off?

Three numbers: how many times a day someone performs the task, how long one repetition takes and what an hour of that person's time costs. Their product is the upper bound on savings — the real figure lands lower, because some cases return to a human.

What happens when AI misreads a document?

Extracted values should land in the form as a proposal shown next to the original, not save themselves. A correction is then one field. A system that writes without confirmation moves risk somewhere nobody is watching it.

Where should we start?

With one task somebody repeats dozens of times a day and whose result can be checked in a second. One well-chosen process reveals the real scale of the saving and gives you a better basis for the next decision than any demo.

Next.js

So many promises have accumulated around artificial intelligence in transport that the conversation usually starts with a misunderstanding. One side pictures a system that runs the forwarding desk by itself. The other remembers a project from two years ago after which nothing changed.

The truth is less spectacular and more useful: AI handles work that repeats, rests on text or a document and has a clear correctness criterion. It handles everything else badly. The whole skill lies in knowing which side of that line a given task falls on — before you spend money on it.

What works today

Reading documents and orders

Orders arrive in every conceivable form: a PDF attachment, the body of an email, a message from a freight exchange, a photo of a printout taken on a phone. Pulling places, dates, cargo description and rate out of them is a task where language models are genuinely good today — including on forms they have never seen before.

The same applies to transport documents. Recognising a consignment note, reading the numbers off it and matching it to the right shipment is work a human does flawlessly and very slowly.

Why this works: the result can be verified in a second, by glancing at the original next to it. A mistake costs little because it is caught immediately.

Classification and matching

Which shipment this scan belongs to. Which counterparty in the database matches the name on an invoice. Whether this message is a new order, a change of date or a status question. These decisions are simple, but taken hundreds of times a day — which is exactly what makes them expensive.

Chasing status updates

A significant part of a forwarder's day is asking people about things not yet known: whether the driver arrived, whether the warehouse accepted, when unloading will happen. The routine part of that correspondence can be taken over, leaving the unusual cases to a person.

Preparing reports

A report someone assembles by hand every week is almost always a candidate for automation. Newer systems let you describe what you need in words instead of configuring filters — which, for a non-technical user, is the difference between a report that exists and a report you have to ask someone for.

What works conditionally

Planning and load matching. An algorithm can search available vehicles and loads, take the driver's remaining driving time and the vehicle's parameters into account, and propose an option. It does this faster than a person and without skipping possibilities.

What it cannot take into account is everything absent from the database: that this driver does not want to run this lane, that unloading at this consignee always takes longer than the booked slot, that this client has to be handled differently from what the spreadsheet suggests.

That is why sensible deployments treat this as a proposal with a rationale, not a decision. An automaton that dispatches trucks on its own is blind in this industry, and proves it quickly.

What AI will not do

It will not decide which carrier to trust with a load under time pressure. It will not judge whether a carrier who was late twice deserves a third chance. It will not negotiate a rate.

These are not technically difficult tasks. They are tasks where the cost of a mistake is asymmetric — one bad decision about entrusting a load costs more than a hundred good ones saved — and where the inputs contain knowledge nobody ever wrote down.

One more thing worth stating plainly, because industry material tends to skip it: automating a process the company cannot describe is not a software project. If two people in the same team do the same thing differently and both believe they are right, that is something to settle before a deployment, not during it.

How to work out whether it pays off

Not from the technology end, but from the cost that is supposed to disappear. You need three numbers:

How many times a day someone performs the task
How long one repetition takes
What an hour of that person's time costs the company

Their product is the upper bound on the saving. The real figure lands lower, because some cases return to a human and someone has to supervise the output.

A simple example. Entering an order takes twelve minutes, fifteen arrive daily, an hour of a forwarder's time costs the company the equivalent of roughly twenty euros. That is three hours a day, or about twelve hundred euros a month. If automation cuts it to two minutes per order, you recover most of that — and only then is it worth discussing the cost of the project.

This example is not hypothetical. Across our deployments, automating order intake cuts processing time by an average of 80 % — from a dozen-odd minutes to two or three, measured from the message arriving to a confirmed record in the system.

It is worth being precise about what that number covers. It covers one task: entering an order. It does not mean a forwarding desk works five times faster, because order entry is only part of its day. It does mean that this particular administrative work stops being the bottleneck — and that growing shipment volume no longer requires growing the team in proportion.

What changes for the team

The most common question after a deployment is not "does it work" but "what do those people do now". The answer is less dramatic than headlines about replacing workers suggest.

Retyping and searching disappear. What remains is what actually creates value in forwarding: talking to the client, negotiating, deciding who to entrust a load to, reacting when something goes wrong. In practice it is not the number of people that changes, but the number of shipments one person handles — and in an industry running on low-teens margins, that is the only durable lever.

There is also a side effect few mention beforehand: the work becomes less fragile to absence. Once knowledge about a shipment stops living solely in the head of the person running it, a holiday stops being an operational event. For small teams that often matters more than any time saving.

The human in the loop, and not out of caution

The rule "AI proposes, a human confirms" is often read as legal cover. It is something else: the precondition for being able to measure quality at all.

When every proposal passes through acceptance, you know what share gets corrected, on which cases, and whether it improves over time. Without that measurement you cannot responsibly widen the scope of automation, because you do not know where you stand.

The thresholds at which a system acts without confirmation are set later and on data — not upfront and on a vendor's assurance.

Three mistakes we see most often

Starting with the hardest process. It is usually the one that hurts most — and the worst candidate for a first project, because it has the most exceptions. Better to start with something dull and repetitive where the effect shows within a month.

Automating a mess. If the counterparty list holds the same company in three versions and documents sit in one person's inbox, automation will speed up the production of mess. Order in the data is a precondition, not a side effect.

No baseline measurement. Without knowing how long the task took before, you cannot say whether anything improved. Measuring takes a week and is the one thing that cannot be done after the fact.

Where to start

With one task somebody repeats dozens of times a day and whose result can be checked in a second. One well-chosen process will tell you more about what is possible than any demo — and give you a basis for deciding on the next one.

In forwarding that first process is almost always order and document handling, because that is where the repetitions concentrate. Planning tends to come second, but only once the input data is in order.

If you would like to run these numbers on one of your own tasks, describe it to us. We need three figures and we will tell you plainly whether automation will pay for itself or whether you should start somewhere else.

AI in freight forwarding: what works today and what is still a promise