Classify AI in practice: three consignments, one decisive attribute each
Classify AI in practice: three consignments, one decisive attribute each
A declarant in a logistics company does not choose the cargo. On the day in question, three consignments were waiting for declarations:
- car multimedia and navigation devices with 5-inch touch screens;
- e-cigarette pods, supplied without e-liquid;
- textile gnome figures for a seasonal décor retailer.
Three unrelated products, one shared problem. Each of them has a near-twin under a different code: a head unit that looks like the navigation device but has no dominant function; a cartridge that looks like the pod but has no heating coil; a gnome that looks like the doll but has no face. The commercial name on the invoice - 'multimedia device', 'pod', 'decorative figure' - fits both twins. What separates them is a single attribute the invoice does not state, and different rules make each attribute decisive: the principal-function rule of Note 3 to Section XVI, the incomplete-goods rule of GRI 2(a), and a product-specific Explanatory Note.
The declarant ran all three throughClassify AI's Finder mode - the automated classification tool, which works through the nomenclature and proposes a code with as few questions as possible, asking only when a legally decisive fact is missing. The input is simply the product description as it stands on the invoice. Across the three consignments, Finder asked exactly one question.
Case 1: the navigation device - which function is principal?
The product was entered as: "Car multimedia and navigation device, 5 inch touch screen, GPS, FM."
The description suggests an infotainment unit. Finder identified Chapter 85 and set out the decisive fork. If GPS navigation is the principal function, the device is radio navigational aid apparatus of heading 8526 - the ground on which the Court of Justice, in case C-268/18, classified a multifunctional in-vehicle GPS device under 8526 91 20. If instead the functions are equally integrated and none dominates, no principal function can be identified and GRI 3(c) applies - the route that leads modern multifunction head units to heading 8528, where Regulation (EU) 2023/2656 classified a multifunction dashboard apparatus.
Then the one question of the day: is this principally a GPS navigation device, a combined head unit with no dominant function, or principally a multimedia device with GPS as a secondary feature? The declarant confirmed navigation is the principal function - and that answer resolved the classification: 8526 91 20 00, radio navigational receivers, under Note 3 to Section XVI, with the CN Explanatory Notes to heading 8526 expressly covering GPS receivers. Even Regulation (EC) No 1231/2007, which reached the same code for a combined GPS and PDA device, got there by a different rule - GRI 3(c), because neither function was principal - which is exactly why the question, not the product name, decides.
Finder stated the alternative openly in its assumptions: change the answer, change the code.
Case 2: the e-cigarette pods - is a heating coil built in?
The product was entered as: "E-cigarette pod with built-in heating coil and mouthpiece, top-fill, no battery, supplied without e-liquid."
No questions this time - the decisive attribute was on the invoice, and Finder named it at every level: the split between subheadings 8543 40 and 8543 90 is the presence or absence of an integrated heating coil. With the coil, the pod is an incomplete electronic cigarette that already has the essential character of the complete vaporising device, and under GRI 2(a) it is classified as that device - 8543 40 00 00 - even without battery or e-liquid. Without the coil, an outwardly similar cartridge is a part under 8543 90, by Note 2(b) to Section XVI.
The evidence stack shows the source ladder working in order. The binding EU instrument is Commission Implementing Regulation (EU) 2023/808, which Finder cited. The interpretive clincher on the coil line comes from the World Customs Organization's classification opinions on subheadings 8543.40 and 8543.90 - a matched pair whose only material difference is the coil: one covers the pod with it, the other the cartridge without it. Finder cited the pair and applied the contrast. It also correctly excluded heading 2404: with no e-liquid, nicotine or substitute in the consignment, this is the delivery device, not the consumable.
For a declarant, this case carries the sharpest compliance message of the three. Which obligations attach to vaping goods - excise, product regulation, notification duties - depends on what the goods objectively are, and that is settled by the same analysis as the code. A pod declared as a part that customs later finds to be an incomplete device means post-clearance recovery of duties and, depending on national law, penalties - not a one-line correction.
Case 3: the gnomes - face, arms and feet?
The product was entered as: "Textile gnome figure for home decoration, visible face, both arms and feet."
Seasonal décor is classic logistics cargo, and its classification practice has long been divided: some customs authorities classified textile gnomes as made-up textile articles under heading 6307, others as dolls under 9503. Anyone who has searched the EBTI database for such goods knows the raw result - a page of gnome rulings pointing both ways, each to be opened and compared by hand. Here that work arrived done: Finder retrieved the rulings and split them by the feature that separates them - stylised figures without human features on the 6307 side, figures with a face, arms and feet on the 9503 side.
The rule behind that split is recent. In July 2026 the Commission addressed the divide with a Combined Nomenclature Explanatory Note specifically on gnomes of textile (C/2026/4225, OJ C, 31 July 2026). Explanatory Notes are not legally binding, but they are an important aid to the uniform interpretation of the nomenclature - and this one supplies the decision rule practice was missing: a textile figure representing a human being, with some visible facial features and both hands or arms and feet or legs, is a doll of subheading 9503 00 21; a figure without that feature set stays in 6307; one defined by festivity symbols goes to 9505. Committee discussions remain open for gnomes outside the note's feature set - and Finder correctly distinguished them, because this product is squarely inside it.
The product met all three criteria, and Finder concluded 9503 00 21 90 - distinguishing the stuffed-toys line (9503 00 41, articles for play, not decoration), and handling the remaining subdivision the right way: for the alternative TARIC line within 9503 00 21, Finder declared the assumption that this is an ordinary commercially produced article and named the code that would apply if it were instead a hand-made doll of folklore character.
The declarant's lesson from this case is about containers, not codes: a shipment of "identical" gnomes can legally split across three headings by features nobody put on the packing list. A gnome whose face hides under the hat is not classified like its cousin with a visible one.
What the three cases have in common
In each case, the commercial description named the product but not the decisive attribute:
- "Multimedia device" - but the principal function decides, and it is navigation: 8526 91 20 00, not 8528.
- "Pod" - but the built-in coil decides, and it makes the pod an incomplete device: 8543 40 00 00, not a part of 8543 90.
- "Decorative figure" - but the face, arms and feet decide, and they make it a doll: 9503 00 21 90, not 6307.
And each attribute was named by a different kind of source: a Court of Justice judgment, a binding classification regulation backed by a matched pair of WCO opinions, and an Explanatory Note published weeks before the shipment. Classification is not a static lookup - the sources move, and the July 2026 note in case 3 gave goods that had divided customs authorities for years their first common decision rule.
Finder asked one question across the three consignments - the one whose answer flipped a code - and classified the other two in single passes, stating its assumptions and the alternative code wherever a fact stayed open. The declarant could see, for each consignment, not just the proposed code but the exact attribute it stands on.
Why this workflow matters
To be clear about where this fits: the repeat lanes of a declarant's day stay on the codes already proven in the file. This workflow is for the other consignments - the new product, the ambiguous invoice line, the shipment that looks like last week's but may be its twin. Those are the ones that produce customs queries, and each of these three products shows why: the near-twin under a different code is one missing attribute away.
For goods that ship repeatedly - and seasonal décor and vaping products are exactly that - the professional next step is Binding Tariff Information. A BTI application stands on precisely the material a session like this assembles: the governing judgment or regulation, the interpretive notes and opinions, and the closest issued rulings. A BTI binds only its holder, so the rulings cited along the way are comparison material, not protection - the protection is your own decision, applied for before the season, not during the dispute.
The decision, and the responsibility for it, stay with the person - not the machine. The declarant chooses the code on the declaration and answers for it;Classify AI supplies the evidence and shows which attribute the decision turns on. That attribute still has to be true of the actual goods in the container - a multimedia-first head unit, a cartridge without a coil or a gnome without feet each carries a different code. What changes is the starting position: the decisive question is on the screen before the declaration is filed, not in the query from customs afterwards.
