Blog

August 12, 2026

13 min read

Classifying a whole catalogue: 500 SKUs without 500 classifications

The method for coding a full catalogue: collapse SKUs into classification families, write descriptions that hold, budget your monthly quota, and verify by sample instead of by faith.

A catalogue of 500 references is almost never 500 classification problems. It is usually forty, repeated.

Coding one product is a task. Coding a catalogue is a project, and the operators who treat it as five hundred repetitions of the same task burn their month, their patience and their budget before they reach the third page of the price list. What follows is the method that makes a catalogue tractable, with the real limits of the tools you will be using.

Five hundred references, forty classification decisions

1. Why a catalogue collapses

An HS code is decided by four facts: constituent material, function, presentation and state. Everything else on a product line is commercially essential and classification-irrelevant:

Changes the code

The material: polypropylene vs stainless steel.
The function: a kitchen scale vs a bathroom scale.
The presentation: loose, in a set, put up for retail sale, knocked down.
The state: raw, processed, frozen, dried.
A size or capacity threshold, but only where the nomenclature itself sets one.

Does not change the code

Colour, finish, pattern.
Size and volume, absent an explicit threshold.
Internal reference, EAN, supplier.
Brand, packaging design, retail price.
Country of origin, which governs duties and preferential regimes, never the code.

Apply that to a real catalogue and the arithmetic changes shape. A housewares importer with 480 references:

In the catalogue	Lines	Classification decisions
Melamine plates, 6 diameters × 5 colours	30	1
Stainless steel cutlery, 4 pieces × 3 finishes	12	1 to 2
Vacuum flasks, 3 capacities × 4 colours	12	1
Kitchen scales, mechanical and electronic	6	2
Melamine plates sold as a 12-piece set	4	1, and it is not the plates' code
…	…	…
Total	480	≈ 46

Forty-six decisions is a fortnight of careful work. Four hundred and eighty is a quarter, and it will not be more accurate. It will be less, because nobody stays sharp for four hundred and eighty judgement calls.

2. Build the working file before you open the tool

The classification lives in a spreadsheet, not in a chat history. The tool is where decisions get made; the file is what survives an audit two years later.

Column	Why it exists
`ref`	Your internal reference.
`label`	The commercial label, as it appears in the catalogue.
`material`	Exact constituent material.
`function`	What it does, in what setting.
`presentation`	Loose, set, knocked down, retail packaging.
`state`	Raw, processed, frozen, dried.
`family`	The family key, described below. One value shared by every variant.
`code`	The code retained, at its full national length.
`rejected`	The candidate code you ruled out.
`reason`	Why. One sentence, written the day you decided.
`sources`	The explanatory note or ruling you leaned on.
`duty`	The import duty rate for the retained line.
`checked`	`assistant`, `code page`, `comparison`, `professional`.
`date`	The date of the decision.

The family key is mechanical: concatenate the three facts that matter and lowercase them.

melamine|tableware-plate|loose        -> family 1
melamine|tableware-plate|retail-set   -> family 2
stainless-steel|cutlery-fork|loose    -> family 3

Sort by that column and your catalogue has already answered the only question that matters at this stage: how many real decisions am I facing?

3. Order the families by what an error would cost

Do not work alphabetically. Work by exposure: units per year × value × duty rate, and pull anything under a restriction or a defence measure to the top.

Three tiers is enough

Tier 1, verify to the bottom. High volume, high duty, or a duty gap between two plausible codes. These get the deepest research level, the comparison table, the notes read in full, and a professional opinion where the file is worth it.
Tier 2, verify normally. Ordinary products, moderate duty, no ambiguity after the first answer.
Tier 3, sample. Long tail, low volume, low duty. Classify the family, verify one in five, and record that you sampled.

A catalogue where every line got the same amount of attention is a catalogue where the wrong lines got most of it.

4. Classify family by family

1Write the description as a template, not a sentence

The composer takes 5 to 200 characters. Fill them with the four facts in a fixed order and your descriptions become comparable across the whole catalogue:

[state] [material] [article], [function], [presentation]

Moulded melamine plate, tableware, sold loose is 45 characters, and nothing in it is noise. Reuse the same skeleton for every family: consistent inputs are what make inconsistent outputs visible.

2One new chat per family

The assistant takes the whole conversation into account. That is exactly what you want while you refine one family, and exactly what you do not want when you move to the next: the previous description keeps colouring the classification. New chat between families, every time.

3Match the research level to the tier

Basic for an unambiguous family. Advanced as soon as two headings compete for it. Expert for tier 1 and for disputed files. It is the slowest mode and it spends the same quota as any other, so keep it for the families that deserve it.

Running a whole catalogue on Expert does not improve the obvious families. It just makes the project three times longer.

4Answer the suggested questions instead of rewriting

The questions under an answer target precisely the fact that is missing to go one level more precise. One click beats a freshly written message, which usually walks back onto a path already ruled out. On a catalogue, this is where most of your time is either saved or lost.

5Save one code per family, with its note

Save the retained code with the reference, the rejected candidate and the reason. Save it per family, not per SKU: 46 saved codes with notes are usable; 480 duplicates are a second catalogue to maintain. It also keeps you inside the Starter plan's 500-saved-code ceiling with room to spare.

5. Budget the quota before you start

Requests are counted monthly, and a family costs more than one: the first answer, plus one or two clarifications. Plan on one to three requests per family.

Plan	Requests / month	Families you can realistically clear	Fits
Free	3	1 or 2	A pilot on your hardest family, to see whether the method holds.
Starter	300	100 to 250	A 500-line catalogue that collapses to forty or fifty families, with room for re-runs.
Pro	1,000	330 and up	Several catalogues, or one catalogue where most families need Expert.

Two things not to forget in the budget: image classifications are a separate quota (50 on Starter, 500 on Pro, none on Free), and Expert costs the same request as Basic, because the difference is time, not allowance.

The sample is drawn where an error costs the most

6. Verify by sample, and prove that you did

What verification means, concretely

Open the official code page for every retained code and read the whole wording, including what follows the comma.
Put close candidates through the comparison: two to four codes, side by side, same digit length, notes and rulings visible. Included in every plan.
Read the notes before settling, not after. The distinction between two neighbouring lines is nearly always there.
Where the ambiguity survives verification, escalate: a customs professional for a tier 1 file, a binding ruling from ADII where volumes justify the wait.

Sampling, done defensibly

Sampling is legitimate; undocumented sampling is not. In the checked column, record what each line actually received. An auditor reading assistant on a tier 3 line and comparison + professional on your five biggest families sees a risk-based process. The same file with an empty column looks like guesswork, whatever the codes say.

7. A realistic two-week plan for 500 lines

Days	Work	Output
1-2	Export the catalogue, fill material / function / presentation / state from the technical sheets.	A file where the gaps are visible.
3	Build the family keys, sort, count. Rank the families by exposure.	≈ 45 families in three tiers.
4-6	Classify tiers 2 and 3, one chat per family, Basic and Advanced.	A retained code per family.
7-8	Classify tier 1 on Expert, with the comparison open.	Codes plus written reasons.
9	Verify: code pages, notes, comparisons. Sample tier 3.	The `checked` and `sources` columns filled.
10	Escalate what survived, save the library, hand the file to your declarant.	A catalogue that can be justified line by line.

The single biggest saving in that table is on days 1 to 3, where no classification happens at all.

FAQ

Can I send an Excel file or a CSV of my catalogue?

Not today: there is no bulk import and no API on any plan. The workflow is one conversation per family, with your spreadsheet as the record. If your volumes fall outside the published plans, contact us rather than assuming a limit.

Do colour and size variants each need their own code?

Normally no. Colour never decides a classification, and size only does where the nomenclature sets an explicit threshold for that product: some textiles, some cables, some machines. Check whether a threshold exists for your family; if it does, the family splits at the threshold, not at every size.

Should I classify per SKU anyway, for my ERP?

Fill every SKU line in the ERP, yes, but by propagating the family's code, not by re-deciding it. The decision lives at family level with one note; the SKUs inherit it. That is also what makes a correction cheap: one family fixed, thirty lines updated.

How long does a 500-line catalogue actually take?

Two weeks of focused work for a catalogue that collapses to forty or fifty families, most of it spent on the technical sheets and the verification rather than on the classification itself. A catalogue with many composite products, sets or borderline materials can double that.

Who is liable if a family is misclassified?

You are. The tool proposes and documents; the declaration is yours, and so is the reassessment if the code is wrong, across every SKU in the family, which is precisely why the family-level decision deserves the written reason. See the terms of use.