Blog

August 12, 2026

12 min read

Automatic HS coding: how it actually works (and what it cannot do)

What happens between a product description and a proposed HS code: the four facts that decide, the search over the nomenclature, the GRI, and why a serious tool shows no confidence percentage.

Automatic classification is not a dictionary that translates a product name into a number. It is a legal argument, built quickly, that you still have to be able to defend.

"Automatic HS code" suggests a lookup: you type a product, a code comes out. That is not what happens, and understanding the difference is what separates operators who use these tools well from operators who get caught by them. This article opens the box: what the machine reads, what it searches, what it can prove, and where it stops.

A sentence in, a justified tariff heading out

1. Why an HS code cannot be looked up

The Moroccan Integrated Tariff is not a list of products. It is a legal hierarchy of roughly 21,000 lines, organised in sections and chapters, governed by the General Rules of Interpretation (GRI) and qualified by section and chapter notes that add and remove products from headings.

Three consequences follow, and every one of them breaks the "search engine" mental model.

Your words are not the tariff's words

The nomenclature ignores commercial vocabulary. Seat returns results where gaming armchair returns none. No amount of keyword matching fixes a query written in the language of a catalogue rather than the language of the tariff.

Most products have no heading of their own

The tariff names families, not articles. A product with no explicit heading falls into a residual "other" line of the relevant chapter, and that line's wording contains none of the words you would think to type.

The notes can overrule the obvious

A heading whose wording fits your product perfectly can still be closed to it by a chapter note. That exclusion is not in the heading, it is pages away, and it is legally decisive.

So "automatic" cannot mean matching. It has to mean reproducing the reasoning a classifier performs, fast, and with its sources attached.

2. The four facts that decide, and the ones that never do

Before anything can be searched, the description has to be turned into classification-relevant facts. There are essentially four.

What actually moves a classification

Constituent material: polypropylene, not plastic; solid oak, not wood. The dominant material often decides the chapter.
Function: what the article does, and in what setting: domestic, professional, industrial.
Presentation: assembled or knocked down, in bulk or put up for retail sale, alone or as a set.
State: raw, processed, frozen, dried.

What adds nothing

Country of origin. It decides duty rates and preferential regimes, never the code.
Brand and commercial reference.
Price and quantity.
Dimensions, unless the nomenclature itself sets a threshold for that product.

This is why Transinaut's composer accepts 5 to 200 characters and no more. The limit is not a technical constraint, it is an editorial one: a dense sentence carrying material, function and presentation outperforms a paragraph of marketing copy, because everything past those four facts is noise the model has to discard.

Too vague: HS code for a phone? The tool has to guess, or ask.

Usable: Mobile phone with touchscreen, sold with charger and earphones in one package.

3. What runs between your sentence and the answer

1Extraction

The description is read for the four facts above. Whatever the tool understood is echoed back in the introduction of the answer, which makes that introduction the most useful paragraph on the page. If the material or the function is wrong there, everything below it is wrong too, and one corrective sentence fixes more than a rewritten request.

2Search over the nomenclature

Candidate headings are retrieved from the tariff itself, over code digits and over the words of the official descriptions, not over a product catalogue. This is the step that finds residual "other" headings, and it is where an automatic tool is strongest against a human with a PDF: it does not skip a line because it is tired.

3Weighing the candidates against the rules

Two or three headings usually survive. They are then held against the GRI and the notes:

GRI 1: the wording of the heading and the section and chapter notes come first. Everything else is subordinate.
GRI 2(a): an incomplete or unassembled article is classified as the finished one, provided it has its essential character. A chair shipped knocked down is a chair.
GRI 3(b): for composite goods and sets, the material or component that gives the article its essential character decides.
GRI 3(c): where two headings remain equally valid, the one last in numerical order wins.

On Advanced and Expert research levels this is not a single sweep: the assistant runs several passes and marks "Pass 2" when it starts a round of verification. There is no progress bar because the number of passes is not known in advance, and a deep analysis regularly runs past a minute.

4Writing the answer with its sources

Each proposed code arrives with a justification and with the references it rests on: the explanatory notes and the classification rulings attached to that line. The references are the point. A code without them is a number; a code with them is an argument you can put in front of ADII.

4. Reading a structured answer

Every classification answer has the same four parts, and each one has a job.

Introduction

What the tool took from your description and the angle it is taking. Your checkpoint for misunderstandings.

Proposed codes

One card per code: the code, its justification (collapsed, expandable) and its references. A copy button, and a button that saves the code to your library.

Questions to narrow it down

Short questions aimed at exactly what is still missing. One click sends them, and that is the fastest route from a 4-digit heading to a full code.

Conclusion

The summary and, where it applies, the reservations: missing information, a borderline heading, a note you should read yourself.

When an answer contains codes, a table lists code, description, import duties, unit, notes and rulings for each one, and every code opens its full official page.

An answer can also legitimately contain no code at all. That is not a failure: it means the description does not support a responsible classification, and the conclusion then carries the questions that would.

Function, material, method of manufacture: what the tool has to read

5. A worked example: an office chair

The description

Office chair, swivel, gas-lift height adjustment, steel frame, polyester upholstery, on castors, shipped knocked down.

118 characters. Material, function, presentation, state: all four are there.

The reasoning

Function: a seat. Chapter 94, heading 9401 (seats and their parts).
Presentation: knocked down. GRI 2(a): an unassembled article with the essential character of the finished one is classified as the finished one. It stays a seat; it does not become a box of parts.
The deciding feature: swivel, with variable height adjustment. That is its own subheading: 9401.30.
The trap avoided: the castors, on their own, would be classified as castors with mountings of base metal (heading 8302). Shipped as part of the chair, they follow the chair. Classifying a component instead of the article is one of the most common errors on knocked-down goods.

6. What automation is good at, and where it stops

Genuinely better than manual

Consistency. The same product described the same way gets the same treatment, on the four hundredth line as on the first.
Coverage. It reads every candidate line, including the residual headings a human skims past.
Vocabulary. It translates trade language into the nomenclature's language, which is where most manual searches fail.
Retrieval of sources. Notes and rulings arrive with the code instead of being looked up afterwards, which is to say: usually never.

Where it stops

Genuinely composite products. Two materials with no dominant one is a GRI 3(b) judgement call, and it is a judgement.
Thin descriptions. Too little detail and the answer becomes a plausible guess. The follow-up questions exist for precisely this.
Novel articles. A product category the tariff has not caught up with cannot be resolved by reading the tariff harder.
The national digits. Digits 7 to 10 follow policy, not physics, and they are the least standardised part of the code.
Liability. The tool proposes; you declare. If the classification is wrong, the reassessment and the penalty are yours. See the terms of use.

7. Automatic does not mean unverified

The loop that makes an automatic classification defensible is short, and every step you skip is paid for later, at the moment the code has to be justified.

The order that works

Describe precisely: material, function, presentation, state.
Let the tool propose, and read its introduction before its conclusion.
Answer the suggested questions rather than rewriting the request from scratch.
Open the official code page and read the full wording and the notes yourself.
Compare the surviving candidates: two to four codes, side by side, at the same digit length.
Save the code with a note: the reference, the code you rejected, and why.

FAQ

Does ADII accept a classification produced automatically?

Customs accept, or challenge, a code on a declaration, and they do not care how you arrived at it. What matters is whether you can justify it with the wording, the notes and the rulings. That is exactly what an automatic tool should be judged on: not the speed, the justification it leaves you with.

When should I ask for a binding ruling instead?

Binding Tariff Information is a formal request to ADII for an official position, and it takes weeks. It is worth it when the product is unusual or new, when two headings stay defensible after verification, or when volumes and duty rates make a reassessment expensive. For ordinary catalogue products it is overkill.

Can I classify from a photo?

You can attach one image per message, 5 MB maximum (PNG, JPEG, GIF, WEBP, SVG), and it counts against a separate monthly image quota: 50 on Starter, 500 on Pro, none on Free. But a photo shows shape, not material or function: send it with a line of description, framed against a neutral background, composition labels legible.

Why do I get a different answer when I ask twice?

Usually because the two descriptions were not the same, or because the conversation carried context from an earlier product. Start a new chat per product: within one conversation the earlier description keeps colouring the classification, which helps while you refine one article and hurts the moment you switch.

Does the code expire?

The six international digits are revised by the WCO on a five-year cycle. The Moroccan digits can change with each annual revision of the Integrated Tariff, usually in January. Keep the date of the decision next to the code, because it tells you what needs re-checking before the next campaign.