Talk of artificial intelligence in foreign trade tends to split into two sentences. One says the customs broker, the logistics clerk, and the trade manager will be replaced in a few years. The other says this is a world of documents and signatures, so no model will ever sit at the desk. On the floor, neither holds. What will change is not the names of the professions. It is which part of the work is slow, which part is expensive, and which part can no longer be defended.
The visible job and the invisible one
From the outside an import file looks like one motion: goods arrive, papers are gathered, a regime is chosen, tax is calculated, the line is closed. From the inside the file is a pile of comparisons. Invoice line against packing list, packing list against bill of lading, bill of lading against the order, the order against the last shipment. An HS code is not written once. It is weighed again with origin, value, unit of measure, and delivery term. The first honest place for AI is here: reading text, matching lines, marking drift. That is not a decision. It is less fog in front of one.
I have been in this work since I was sixteen. Import, export, transit, warehouse, inland customs. When volume rises, people do not get tired in the way textbooks describe. They get faster. When they get faster, the last declaration sticks to the next one. A model can be a second eye against fatigue. It cannot take fatigue’s place. A second eye does not take the legal seat of the first.
Liability is not inferred. It is carried.
A customs declaration is not a suggestion. It is a statement given in the importer’s name, and it can be opened later in post-clearance. A wrong tariff, a wrong value, a wrong origin is not closed by a model saying it is ninety percent sure. Being sure and being liable are different things. So what AI will change in trade is not the romance of automation. It is the division of labor. The machine produces candidates. The human selects, reasons, and signs. If the selection is recorded, the system matures. If it is not, you have a fast guess.
The same limit sits on the document. A footnote on a bill of lading, a line on a certificate of origin, a discount on an invoice can mislead a model. Trade paper is not written like a novel. It is incomplete, scattered, sometimes in two languages, sometimes in two units. A good system does not hide that. It does not say “I understood.” It shows what it did not. In foreign trade the most expensive sentence is confidence laid over uncertainty.
What will not change
Delivery terms, payment terms, finance, and the cash cycle stay independent of the model. The CIF value of a good and the cost that lands in the company’s cash are not the same thing. Demurrage, storage, FX, the day VAT is locked, a later assessment from a wrong classification — these do not appear on a slide as “process efficiency.” AI can read that layer too. Reading it is not the same as making the cash decision.
What will change is this: a good operator will hunt fewer papers and manage more exceptions. A weak operation will multiply the same errors faster. Technology enlarges discipline. Without discipline, speed is a virtue and the result is a cost. AI in foreign trade should be discussed after that sentence. Before it, what is usually being discussed is not a product. It is an expectation.