AI and HS Classification: Where the Machine Has to Stop

The Harmonized System settles its own interpretive hierarchy in its first rule, and the rule is narrower than most users assume. GRI 1 provides that the titles of sections, chapters and sub-chapters are for ease of reference only, and that for legal purposes classification shall be determined according to the terms of the headings and any relative section or chapter notes.

Two consequences follow that models systematically miss. First, the chapter title is not evidence. A product that obviously belongs to the subject matter of a chapter may be excluded from it by a chapter note, and the note governs. Second, the operative text is often exclusionary and located elsewhere in the schedule — a note in one section removing goods to another. A reasoning process built on resemblance between a product description and a heading description will not go looking for the note that defeats the resemblance, because nothing in the description points at it.

GRI 3 then handles the genuinely contested case, where goods are prima facie classifiable under two or more headings, and it does so as an ordered sequence: 3(a) the heading providing the most specific description; failing that, 3(b) the material or component giving the goods their essential character; failing that, 3(c) the heading last in numerical order among those equally meriting consideration. This is a decision procedure, not a weighing of considerations — and a determination that does not state which limb it rests on has not shown its reasoning.

The limits are real, and dismissing the tooling on account of them would be an expensive mistake. Classification work contains a large volume of tractable labour that machines handle well.

Retrieval. Finding every note, ruling and explanatory text that bears on a candidate heading is a search problem across a large corpus, and it is the step humans skip when busy. A system that surfaces the exclusionary note is doing the highest-value part of the job.

Consistency. Firms misclassify most often by drift — the same product coded differently across shipments, sites or years, usually because different people classified it. Detecting that a consignment has been coded differently from the last thirty identical consignments is mechanical, and it is one of the strongest audit-defence signals a firm can hold.

Change tracking. Tariff schedules, notes and rates move. Knowing that a heading relevant to a firm's catalogue was amended is monitoring work, and monitoring is exactly what does not survive being someone's Tuesday afternoon task.

Assembling the file. A candidate heading, the notes considered, the GRI limb relied on, the competing heading rejected and why — presented for a human to accept or reject — converts hours of research into minutes of judgement.

Classification is often confused with origin, and a single output claiming to resolve both is resolving neither. Under CAROTAR 2020, an importer claiming a preferential rate must possess origin information as indicated in the prescribed form and exercise reasonable care as to its accuracy and truthfulness, and must furnish it on the proper officer's requisition. A certificate of origin is not, by itself, discharge of that obligation.

Two separate evidentiary burdens therefore sit on the same consignment — the code, determined under the GRI; and the origin claim, supported by information the importer must actually hold. Systems that present a single confident answer covering both are compressing away the distinction the law draws.

TradeWatch treats classification as a prepared determination rather than an answer: candidate headings with the notes and GRI limb that produced them, drift flagged against the firm's own history, origin information tracked as a separate requirement, and anything unresolved surfaced as unresolved rather than defaulted. Kanan Labs prepares a readiness packet. It does not file Shipping Bills and holds no customs credentials — your licensed CHA files. Final HS classification requires human review.