Can ChatGPT Find the Right HS Code? How to Check Before You File

As a first draft, yes. ChatGPT can name likely headings and a plausible code from a good product description. It cannot be your final answer: the importer of record is responsible for the code filed. Check any suggestion against the official tariff, the chapter notes and, where the stakes are high, a binding ruling.

The first four to six digits are the more dependable part of a chatbot's answer; the national digits after them are where it is most likely to guess. The method below treats ChatGPT as a research assistant that proposes candidates and names the facts that would settle them, then makes you prove the choice with the tariff text. It includes a ten-product test you can run on your own goods, so you measure the error rate for your products instead of trusting anyone's claim, including mine.

Follow me on Instagram@sagnikteaches

How a code is built, and which digits a chatbot can be trusted with

The Harmonized System (HS) is an international nomenclature that groups goods into chapters (two digits), headings (four digits) and subheadings (six digits). Most customs territories then add their own digits, commonly taking the code to eight or ten, where duty rates, statistical splits and import controls are set. So a code has layers:

Connect on LinkedInSagnik Bhattacharya

• Chapter (2 digits): the broad material or product family, such as ceramics or knitted clothing.
• Heading (4 digits): a specific class within it.
• Subheading (6 digits): the international split, shared worldwide.
• National digits (7 to 10): your tariff's own detail, where the duty rate often changes.

Classification follows the General Rules of Interpretation, six rules applied in order. In practice, the first rule does most of the work: the wording of the headings and the section and chapter notes decide, and the notes can exclude a product from a heading that looks like an obvious fit. A chat model has seen plenty of tariff text and can often name the right chapter and heading. What it cannot do is look up today's national tariff, see your product, or accept legal responsibility. Where it falls over is predictable: the notes, the national digits, and goods that sit between two headings.

Subscribe on YouTube@sagnikteaches

A ten-product test: measure the error rate on your own goods

Published claims about chatbot accuracy vary and none of them is about your catalogue. A test on ten of your own products takes about an hour and tells you exactly how much first-draft help to expect. I have not run it on your goods, so no accuracy figure is claimed here; the point is that you can get a real one.

1. Choose ten products whose correct code you already know from a cleared entry, a broker's classification or a ruling. Include two or three awkward ones: items made of more than one material, sets, parts and accessories.
2. Write a buyer's description for each: just what a supplier listing or product page would say, with no code and no hints.
3. Ask ChatGPT in a fresh chat for each one, using the prompt below, so earlier answers cannot contaminate the next.
4. Score each answer at four levels: right chapter, right heading, right six-digit subheading, right national code. Mark each level yes or no.
5. Count wrong answers at each level and divide by ten.

The scoring sheet looks like this. The values below are invented to show how to read the result; yours will differ.

Five rows are shown; a full test has ten. Suppose the ten rows were scored like this: chapter correct 10 of 10, heading 8 of 10, six digits 7 of 10, national code 5 of 10. The error rate then grows with each layer: 0 per cent at chapter, 20 per cent at heading, 30 per cent at six digits and 50 per cent at the national code. The shape matters more than the figures. If your result looks like that, you can use ChatGPT to find the neighbourhood and nothing more. If your error rate at heading level is above about one in ten, stop using it for first drafts of that product type.

Rerun the test every few months, and after a model change, since results can shift. The same ten products make a handy regression test.

The description sheet that decides the quality of the answer

Most wrong codes begin with a thin description. "Water bottle" fits at least three headings. Fill in these fields for each product before asking anything:

• What it is and what it does, in a sentence, including who uses it.
• Material and composition, with percentages by weight where it is a mix.
• How it is made: knitted or woven, cast or machined, moulded, vacuum-insulated, printed.
• Form when imported: complete, unassembled, a part, a set sold together.
• Dimensions and capacity where they distinguish classes.
• Electrical and technical features: voltage, wireless function, battery.
• How it is sold: retail pack, bulk, as a set.

Ask your supplier for these in writing, since they will also help you with customs questions later; the tutorial on supplier emails for importers includes wording for asking. Do not paste a supplier's confidential specification or a customer's private design into a consumer chat without checking your agreements and the plan's data controls first.

A prompt that returns candidates, exclusions and open questions

I import goods and need a first-draft customs classification. I will verify your answer against the official tariff, so be explicit about uncertainty.

Product description:
[paste description sheet]

Please return:
1. Up to three candidate headings (4 digits), each with the reason it could apply, quoted from the heading wording as you remember it.
2. For each candidate, any section or chapter note or exclusion that could rule it out.
3. Your best 6-digit subheading and your confidence (high, medium, low).
4. The questions about the product that would change the answer.
5. Do NOT give national 8 or 10 digit codes, duty rates or tax figures. I will look those up in my own tariff.

Two choices in that prompt matter. First, it bans national digits and duty rates, because those are the layers where invented detail slips in. Second, it asks for the questions that would change the answer, which is often the most valuable output: it tells you what to ask your supplier next.

Three products worked through

The stoneware mug: a clean first draft

A printed stoneware mug for hot drinks. The material (ceramic, not porcelain) and the use (tableware) point to the heading for ceramic tableware and kitchenware other than porcelain, heading 6912. This is the kind of product a first draft should get right, because no other heading competes with it. Your own check takes five minutes. Open the tariff at heading 6912 and read the heading text; check the chapter notes on ceramics for any exclusion; confirm the mug is not porcelain, which sits in heading 6911 (a supplier's material declaration settles that); then find the national subheading under 6912 in your own tariff. For a product like this the AI saved you time and little more.

The insulated bottle: where the obvious answer is probably wrong

A double-walled, vacuum-insulated stainless steel bottle. The tempting answer is the heading for household articles of stainless steel (7323). But the international nomenclature has a heading for vacuum flasks and other vacuum vessels (9617), and published customs rulings on insulated stainless steel bottles have placed them there; one advance ruling, for example, rests on the first interpretation rule and the explanatory notes to that heading. The facts that decide it are about construction: is there a vacuum between the walls, and is the whole item a vacuum vessel? A single-walled steel bottle with a coating belongs in the steel article heading.

That is why the prompt asks for the questions that would change the answer. A reply to this product might look like this (illustrative):

Candidates: 7323 (table, kitchen or other household articles of iron or steel), 9617 (vacuum flasks and other vacuum vessels), 3924 (plastic household articles, if the body were plastic).
Best 6-digit: 7323.93 - confidence MEDIUM.
Questions that would change the answer: Is there a vacuum between the walls? Is the lid or cup included? Is the body stainless steel?

If the supplier confirms a vacuum between the walls, the 9617 candidate moves to the top, and your best step is to search the ruling database of the customs authority you import into for the same kind of product and read how it was classified, or to apply for an advance ruling if the volume justifies it. This is what "verify against rulings" means: the AI found the fork, and a ruling settled it. An importer who accepted "7323.93, confidence medium" on a pallet of 5,000 bottles could pay the wrong duty on every unit until someone noticed.

Knitted gloves: the material test that sends you to a different chapter

A pair of cotton gloves. In the international nomenclature, knitted or crocheted gloves sit in chapter 61 (heading 6116) and gloves that are not knitted, such as woven ones, in chapter 62 (heading 6216). So the one fact "how is the fabric made" moves the product across a chapter. Ask the supplier for the fabric construction; if the description says "jersey" or "stretch cotton", it is likely knitted, but do not assume. After that, the six-digit split is by fibre (cotton, synthetic and so on), which is where composition percentages matter. A reply that says "cotton gloves, 6116.92" with no mention of knitted versus woven has guessed at the key fact. The national digits after 6116.92 are then yours to look up.

Five more traps to put in your ten-product test

A test of ten easy items proves little. Fill at least half of your ten with products where the answer turns on one fact. Here are five patterns that catch first drafts out, each with the question to ask and the rule that usually settles it.

• The flat-pack bookcase. It arrives unassembled in several cartons. A chatbot often reaches for a heading about parts of furniture. The interpretation rules treat an unassembled article, presented together, as the complete article, so the product is classified as furniture. The question to ask: does the shipment contain everything needed to assemble it, or only some parts?
• The gift set. A candle, a ceramic holder and a card, boxed for retail. A set put up for retail sale is classified by the component that gives it its essential character, when the goods are not covered by a single heading. The question: which item is worth the most, takes the most space or has the main role? Write down the answer and the reasoning, since this is a judgement call.
• The cotton tote bag. Textile bags can sit in a heading for containers and cases or in one for other made-up textile articles, depending on features such as a stiffened shape, a lining or an outer surface of a particular material. The question: what is the outer surface, and does it have an inner lining or pockets? A chatbot that does not ask has chosen at random.
• The LED desk lamp with a USB charger. The product combines lighting and a charging function. Which function gives the product its essential character, or is its principal one, usually decides. The question: what is it mainly sold and used for?
• The phone case. Material (leather, plastic, silicone, textile) and construction (a wallet with card slots or a simple shell) can place it in different headings, and rulings differ between them. The question: what is the outer material and does it hold cards or other items?

Score each on the same four levels. If your scores follow the shape of the illustrative table above, the chapter will mostly be right and the heading is where first drafts diverge, which is also where the duty rate tends to change. If your own scores show the opposite, trust your scores. For each missed item, add the missing fact to the description sheet, so the sheet gets better with every miss. After a month you will have a reusable library of descriptions for the products you buy repeatedly, each with its verified code and the date it was checked.

Four checks before a code goes on an entry

1. Read the heading and the notes. Open the official tariff for your import territory. Read the full text of the heading, the subheading, and the section and chapter notes, plus the explanatory notes where available. The notes are where the exclusions live, and chatbots paraphrase them poorly.
2. Search for rulings and precedents. Look in the customs authority's published ruling database for similar goods. A ruling on a close match is strong evidence. Where there is none and the duty difference between candidate codes is large, consider applying for a binding or advance ruling. Ask a customs broker how long that takes in your market.
3. Compare with what was cleared before. If you or a competitor's goods of the same kind were cleared under a code, check why. A prior entry is a clue, not proof, since mistakes repeat.
4. Write down the decision. Record the product facts, the candidates considered, the reason for the pick, the date and the tariff edition. If your entries are ever questioned, this file shows reasonable care, and it lets a broker or a colleague pick up the thread.

Where the money is large, have a licensed customs broker confirm the code. The tutorial on customs brokers using AI for client updates shows what a broker's own workflow looks like, which helps when you decide what to ask them to check.

Who is liable when the code is wrong

The importer of record is the party that declares the goods to customs and normally carries the legal responsibility for the declaration, including the classification, the value and the origin. Using a broker transfers some of the work, but depending on the territory and the contract, not all of the risk. The consequences of an incorrect code can include underpaid duty collected later, penalties, delayed clearance and, for controlled goods, licensing problems. The details vary by territory, so ask a broker or adviser what applies to you.

This is a practical argument for the test above. If a chatbot suggestion becomes your filed code, the responsibility is yours, and so is the evidence trail. A reasonable routine: the AI proposes, you verify at the tariff, the broker confirms any code above an agreed duty or volume threshold, and the decision is recorded.

Where first-draft classification goes wrong

• Stale codes. HS 2028 replaces HS 2022 on 1 January 2028, according to the World Customs Organization, and national tariffs change in between. The model's memory is not today's tariff, and a code drafted now may be split or moved by the new edition. Always check the edition in force on your entry date, and recheck repeat products before the first shipment of 2028.
• Invented national digits. Eight- and ten-digit codes look authoritative and are often wrong. Ban them in the prompt, as above.
• Sets and kits. A gift set of a candle, a mug and a notebook is classified by a rule for sets and by which item gives the set its character, not by averaging. Treat sets as a broker question until you know the rule.
• Parts and accessories. A part might go with the machine, or into its own heading, depending on the notes. Chatbots often default to the machine.
• Marketing words. "Smart", "eco", "premium", "professional" mean nothing to a tariff. Strip them from the description.
• Using the supplier's code as the answer. It is a starting point. It may have been chosen for the supplier's own export paperwork.

A short before and after shows the first one. Before: "Wireless speaker with fabric cover, 20W, from the supplier's page." After: "Portable wireless speaker, Bluetooth, battery, 20W, plastic body with a woven fabric grille, sold in a retail box with a USB cable; no microphone, no screen." The second gives a classifier everything it needs; the first invites a guess at "audio equipment" in general.

Fitting classification into the rest of your costing

Once the code is checked, its duty rate feeds your cost per unit. A $4.20 unit cost with duty at 6.5 per cent on the goods value adds about $0.27 per unit, while at 2.7 per cent it adds about $0.11; on 5,000 units the difference between two candidate codes is about $800. These are illustrative rates, so use your own tariff's. The point is that classification has a price, and that price is the argument for spending a broker's fee on borderline goods. The landed-cost calculator tutorial shows how to build that into a sheet and test a change in duty rate in seconds, and the importer compliance checklist covers the other documents a shipment needs besides the code.

If you would like to set up a repeatable routine for your catalogue, including the description sheet and the scoring test, that is what my AI implementation consultation covers; you can book a 1:1 call here. Nothing here is customs or legal advice; confirm classification with a licensed customs professional for the territory you import into.

Questions importers ask about AI and tariff codes

Is the HS code on my supplier's invoice good enough?

Treat it as a clue, not evidence. Suppliers pick codes for their own export paperwork, and the same goods can be coded differently across suppliers. You, as importer, file the code and answer for it, so check it against the tariff and the product facts yourself, or have a broker confirm it.

Why do the first six digits match but the rest differ between countries?

The first six digits come from the international Harmonized System. Each customs territory then adds its own digits, usually taking the code to eight or ten, for national duty rates and statistics. A chat assistant is far more likely to be right at six digits than at ten, so always check the national extension against your own tariff.

Should I give ChatGPT photos of the product?

A photo helps with construction and material, but it cannot show composition percentages, whether a fabric is knitted or woven, or what is inside. Give it a written specification as well. Never upload images or documents that show a customer's confidential design.

Can a wrong HS code get me in trouble even if AI suggested it?

Yes. The importer of record is responsible for the declaration, and 'the chatbot said so' is not a defence. Consequences can include back duty, penalties and delays. Keep the evidence of how you checked the code, and use a binding ruling where the money at stake is large.

How often do codes change?

The international nomenclature is revised every five or six years: HS 2022 is the edition in force, and the eighth edition, HS 2028, takes effect on 1 January 2028. National tariffs also change in between. A code that was right on last year's shipment may not exist or may have been split. Check which edition and which national tariff version applies on your entry date.

Further reads

• Proforma Invoice vs Purchase Order: Let AI Catch the Mismatches — Catch mismatches between invoice wording and the code you file.
• Supply Chain AI Software: What a Small Importer Needs First — What a small importer needs before buying supply chain software.
• Find Overseas Suppliers With AI: Shortlists, RFQs and Red Flags — Get product specifications from suppliers that classification needs.
• Pre-Shipment Inspection Checklist: AI Drafts Your Inspector Brief — Check the goods match the description before they ship.
• Alibaba Sourcing With AI: Vet Suppliers and Plan a Sample Order — Request the specification facts you need before ordering.
• Is It Safe to Put Customer Data Into ChatGPT? — Decide what supplier and customer data to keep out of a chat.
• Export Declaration: What AI Can Pre-Check Before You File — Use AI to catch mismatched values, weights and descriptions across your shipping papers, and know which decisions stay with you or your broker.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: World Customs Organization pages on the Harmonized System and the HS 2028 edition; the General Rules of Interpretation and the heading texts for 3924, 4202, 6116, 6216, 6307, 6911, 6912, 7323 and 9617 as published in a national tariff schedule; a published advance tariff ruling placing an insulated stainless steel bottle in heading 9617. All checked 5 November 2026. The product examples and scores are illustrative and not measured results.