The six GRI rules, in order, and the two that actually bite
The most common mistake in tariff classification is treating it as a matching problem.
You have a product. You have a schedule with forty-odd thousand codes. You find the one whose description sounds most like your product. Done.
That's not how it works, and the reason is the General Rules of Interpretation. They sit at the front of every tariff schedule in the world, they're six paragraphs long, and most people who classify goods for a living have never read them end to end.
Six rules, applied in order
The GRI are the WCO's procedure for arriving at a code. The word order is doing real work here. You don't get to pick whichever rule produces the answer you'd prefer. You apply Rule 1, and you only move to Rule 2 if Rule 1 genuinely fails to resolve it.
Rule 1 says classification is determined by the terms of the headings and any relative section or chapter notes. That's the rule. Everything below it is what you do when this one doesn't settle the question — which, for most goods, it does.
Rule 2 handles incomplete or unfinished articles that have the essential character of the finished thing (a bicycle in a flat box is still a bicycle), and extends headings for a material to mixtures of that material with others.
Rule 3 is what to do when two or more headings both seem to apply. Three sub-rules, applied in strict sequence: 3(a) the most specific description wins over the more general; 3(b) if that doesn't decide it, classify by the material or component that gives the goods their essential character; 3(c) if that still doesn't decide it, take the heading that comes last in numerical order. That last one is a tie-breaker, not a judgement, and it exists precisely so that the procedure always terminates.
Rule 4 covers goods that fit nothing: classify with the most akin goods. Genuinely rare, and if you're reaching for it you've probably missed something upstream.
Rule 5 deals with cases, containers and packing materials — camera cases, instrument cases, the packaging itself.
Rule 6 applies the whole of the above one level down, to subheadings, and only compares subheadings at the same level. That constraint matters more than it looks: you can't jump from a one-dash subheading to a two-dash one under a different parent.
The two that decide arguments
Read enough binding rulings and a pattern shows up fast. Rules 4 and 5 barely appear. Rule 2 appears occasionally. Almost every contested case turns on Rule 1 or Rule 3(b).
Rule 1 bites because of the notes. People read the heading text and stop there. But a section or chapter note can exclude your product from a heading that describes it perfectly, and the note carries exactly the same legal force as the heading. Section XVI has notes that reroute whole categories of machinery. Chapter 39 has notes that decide whether a plastic article belongs there or in the chapter for the thing it is. Chapter 61 and 62 hinge on notes about knitted versus woven that override what the garment obviously is in commercial terms.
More classifications are wrong because somebody didn't read a section note than for any other single reason. It isn't close.
Rule 3(b) bites because "essential character" is a judgement call and the rule declines to define it. A steel tool with a plastic handle. A textile bag with a leather trim. A gift set of a shaving brush, a bowl and a soap. A device that's part electronic and part mechanical. The Explanatory Notes offer weight, quantity, value and role in relation to use as possible indicators, and then say which of these matters depends on the goods.
Which is honest, and also means two competent professionals can reach different codes and both hold defensible positions. That's not a flaw in the system. It's why authorities publish binding rulings at all — the ruling isn't reciting the rule, it's recording where somebody drew the line on a specific set of facts.
Worth noting that 3(b) is also where the money hides. The duty difference between the two candidate headings on a composite good is routinely several percentage points, and occasionally the difference between a normal rate and a trade remedy.
Why this matters for anything automated
If a system hands you a code without telling you which rule it applied, it hasn't classified anything. It's done a lookup and dressed it up.
The output that's actually useful reads more like: this heading, under GRI 1, because chapter note 2(b) excludes the alternative you'd otherwise reach. Or: this heading under GRI 3(b), essential character given by the steel body, consistent with these three prior rulings on comparable articles.
That's a position. You can take it to a customs officer, you can defend it at audit, and if it's wrong you can see exactly where the reasoning failed and fix that step rather than retraining the whole thing.
A bare ten-digit number with a confidence score is not a position. It's a guess with a decimal point attached. And the confidence score is particularly misleading here, because the cases where a model is most confident are the easy Rule 1 cases nobody was going to get wrong, while the 3(b) composites — the ones where the money is — produce middling confidence on both candidate headings and no indication that a judgement is being made at all.
There's a design consequence in that. If you're building or buying this, the interesting question isn't the average accuracy. It's whether the system can identify that it's in a 3(b) situation and escalate, rather than silently picking one side of a genuine judgement call.
The practical upshot
If you classify at any volume, the questions to ask about your existing process are the same ones to ask of any tool.
Does it read the section and chapter notes, or only the heading text? Does it tell you which rule it applied and why? Can it show you prior cases where the same line was drawn, and can it tell the difference between a case that's on point and one that merely uses similar words? And when the answer is genuinely arguable, does it say so, or does it pick one and move on?
The schedule lists the codes. The rules decide which one is yours.