AI does not belong where your value is

In international freight, quoting a shipment is the work. It is not the paperwork before the work: it is the work. Every dollar counts, and the difference between a good quote and an acceptable one is made by someone who knows the route, the consolidation, the season and the client.

So when a freight forwarder starts thinking about AI, the first idea is always the same: have the AI quote on its own. The client enters the requirements, the system returns a price. It is the demo everyone wants to see.

It is also, almost always, the worst part of the process to automate.

Automating what you do best is not a saving

If the value of your service is a finely built quote — personalised, optimised, shaped by someone who understands the operation — then automating the quote does not cut your costs. It levels you with the rest of the market. You are using technology to make yourself replaceable at the one thing that made you worth choosing.

A model can produce a quote that looks reasonable. What it cannot know is that this particular client will accept a longer transit if the price drops, that this origin gets complicated in high season, or that it is worth consolidating with another shipment leaving next week. None of that is in the requirements the client typed in: it lives in the head of the salesperson who handles the account.

The conclusion is not “no AI”. It is AI somewhere else.

There is plenty to automate around the value

That same forwarder has, around the quote, an enormous amount of friction worth taking off people’s hands:

• Clients asking again and again where their shipment is, waiting on an answer someone has to go and look up in a system.
• General questions answered the same way every time.
• Quote requests that arrive incomplete, forcing three rounds of back and forth before the real work can even start.

That is where AI earns its keep. It answers on shipments in transit, handles the general questions, and when someone asks for a quote it does what it does best: it gathers the full set of details, organises the request and hands it over as an opportunity to the salesperson who owns that account.

The salesperson does not get a notification that somebody asked something. They get the request with the details already gathered, and sit down directly to the fine work. The quote comes out better, and it comes out sooner.

AI did not replace the value: it gave time back to it.

Four signs your problem is not an AI problem

That case yields a rule that reaches well beyond freight. Be sceptical when:

The rule is deterministic. If the decision can be written as a table or a set of conditions, you do not need a model: you need the conditions written down. A probabilistic model solving something with an exact answer is more expensive, slower and less reliable.

The volume does not justify it. AI carries a permanent running cost. If the process happens twenty times a month, the cost of running it, measuring it and correcting it never pays for itself.

Errors are expensive and verification is not cheap. A model gets things wrong; that is part of the deal. The question is what happens when it does. If nobody is going to review it and the error reaches the client directly, the problem is not technical: it is in the design of the process.

The data is not there. Much of what a business knows lives in loose spreadsheets and in people’s heads. No AI is going to invent it for you. When the process is undefined, automating it does not fix it: it speeds it up. A bad process with AI is a bad process, more expensive, every month.

And the cost nobody budgets for

There is a difference between AI and traditional software that changes how the decision should be made, and it is barely discussed.

Software is expensive to build once and cheap to run. If you picked the wrong problem, you lost the up-front investment and the bill ends there.

AI is cheap to start and runs forever, at a cost that grows with usage. If you picked the wrong problem, you did not lose an investment: you took out a subscription to your mistake.

On top of that sits the cost that almost never makes it into the budget: evaluation. Working on demo day means nothing. Someone has to measure whether it still works, review the cases where it fails, and correct them. That work does not end. It sinks AI projects far more often than the price of tokens does.

So

The potential is real, and it is enormous. Which is precisely why the decision does not get made on enthusiasm.

Before asking whether AI can do it — it almost always can, in some form — ask where your value sits. That part gets protected. Everything around it, stealing its time, is where AI moves the needle.