An HS-code argument often opens like a word game. Someone asks what the good is. The answer is a brand, a catalogue line, sometimes a photograph. The tariff does not work that way. The Harmonized System is not a naming book. It is a classification tree. Material, function, essential character, mixture, set, the degree to which a part completes a machine — these come before the name the market uses. Asking AI for “the code of this good” is like asking a lawyer whether you will win the case. A sentence can be formed. Liability does not sit where the sentence was formed.
A candidate is not a ruling
A good model can read the invoice description, the technical sheet, and earlier classifications, then produce three or five candidate headings. That is useful. It is even necessary. In volume, people write the first code that comes to mind. A model can recall a rare note, a split, an exception. The benefit ends there, and the real work begins. The choice among candidates is made by the General Interpretative Rules, the chapter notes, and national practice. The choice is not a search result. It is a reason.
The same good can fall elsewhere by use. Is a part an accessory of a machine, or an article in its own right? Is a textile a fabric, apparel, or babywear? Is a food in shell, shelled, or roasted? These questions are not answered from a catalogue. They are answered in the tariff’s own language. AI can imitate the language. It does not administer the tariff. Administration returns to the same language in post-clearance.
The break I have seen on the floor
The error I have seen most often at the import desk is the last file’s code moving onto the next. The goods are “similar.” The line is “almost” the same. The unit has changed, the mix has changed, the origin has changed; the code stays. People do this from fatigue. A model can do it from statistics. Both are dangerous. The tariff does not walk by resemblance. It walks by ruling.
In import files handled under power of attorney on a UPS line, in a vehicle-import group, in micro-export files where ECCN and HS sit side by side, the same lesson repeats: a code is not a label. It sets the duty, the document, sometimes the export-control path. 7113.11.00 and 6111.20.90 are not the same kind of sentence. One is precious-metal jewelry. The other is cotton babywear. A model can tell them apart. Telling them apart is not declaring in your place.
The rule for working together
Human and AI are not rivals in HS classification. They become rivals when the order is wrong. The right order is this: the model reads and proposes; the human opens the note, applies the rule, writes the reason; the system keeps the reason. A reason that is not kept disappears on the next shipment. A disappeared reason becomes “that is what we thought at the time” when an assessment arrives a year later. In foreign trade, memory is not intent. It is a record.
That is why using AI as a search engine for HS codes is wrong. Using it as a preparation layer is right. A preparation layer does not take the human’s seat. It narrows the blind spot. The blind spot does not close. The tariff is a text that opens again with every new good. A machine may read the text. The name bound to the text still belongs to the person who declares.