Customs tariff classification with an internal AI: CN code proposals with reasons

What an internal AI does in tariff classification

An internal AI for customs tariff classification reads the description, material, function and datasheet of a product, proposes the 8-digit code of the EU's Combined Nomenclature, explains why it chose that chapter, heading and subheading, and quotes the nomenclature text and notes it relied on. A person in your customs or export team checks the proposal and records the code with the reasoning; where the case stays doubtful, the AI prepares the description for an application for binding tariff information. The model needed is small enough to share a server with other office tasks.

We wrote this for owners, managing directors and the people who run export, customs or master data at manufacturers and wholesalers that classify their own products, often without a customs specialist on the staff. The first part explains the nomenclature and the work in plain words; the hardware sits in one section with one table; notes for the IT team come at the end. Customs points are general information on EU rules as of October 2026; the customs authority decides on a classification, and your customs adviser or freight forwarder confirms the code for a doubtful case.

The task is familiar to anyone who maintains an article master. Every new product needs a tariff code before the first shipment out of the EU or the first import of a component. The code decides the duty rate and what the declaration says about the goods. Many companies copy the code from a similar product, take the supplier's code without checking it or leave the field empty until the freight forwarder asks. That works until a customs check finds a wrong code.

The Combined Nomenclature in plain words

The Combined Nomenclature was established by Council Regulation (EEC) No 2658/87 and is, in the words of the Commission's customs page, the basis for declaring goods at import or export and for statistics on trade between EU countries; it determines the customs duty that applies. Each subdivision is a CN code of eight digits. The first six digits come from the Harmonized System of the World Customs Organization; the last two are the EU's own subdivision.

Structure as shown on the Combined Nomenclature page of the European Commission (taxation-customs.ec.europa.eu), read 25 October 2026.

Three kinds of text decide a classification. The nomenclature itself, with its headings, its section and chapter notes and the general rules for its interpretation in its preliminary provisions, is law. The explanatory notes to the CN explain the scope of the headings; the Commission states that "they are not legally binding", and a reasoned proposal still cites them, because they show how a heading is meant. And earlier decisions, such as binding tariff information issued to other companies and classification regulations, show how a product of a similar kind was classified.

The nomenclature changes every year. Annex I to Regulation 2658/87 is updated annually, including changes agreed at international level in the Harmonized System, and published as a separate regulation. The 2026 version is Commission Implementing Regulation (EU) 2025/1926 of 22 September 2025, applicable from 1 January 2026. A code that was right last year can be wrong this year.

How the AI proposes an HS and CN code

The AI does not guess a code from memory. It searches the current texts and reasons from them, step by step:

1. Product data. The system collects what the article master holds: description, material and its shares, function, how the product works, intended use, datasheet, photos, the supplier's code if one is given.
2. Search. It searches the current CN, the explanatory notes and your own earlier decisions for the headings that match, using the method of retrieval on documents explained in our guide to RAG on company data.
3. Reasoning. The model writes its proposal from the top down: section and chapter, heading, subheading, CN subheading, each step with the text that supports it and the alternatives it ruled out.
4. Gaps. Where the decision turns on a fact the data does not contain, such as the share of a material or whether a part is designed only for one machine, the model names the missing fact instead of picking a code.
5. Decision. A person checks the proposal, adds missing facts, decides and records the code with the reasoning, the date and the CN version.

Our description of a classification record; the record is the company's own documentation of how a code was chosen.

The value lies in the record. A customs officer, an auditor or a new colleague can follow why a code was chosen, and when the nomenclature changes, the record shows which texts the decision rested on. The AI makes errors of the kind a person makes: it can choose the wrong heading where two seem to fit, or take a function for granted that the product does not have. That is why each step carries its source text and why a person decides.

Binding tariff information: the reliable route

Binding tariff information, BTI, is "a legal decision issued by EU countries' customs authorities on the tariff classification of a product", as the Commission's EBTI page describes it, under Articles 22 to 37 of the Union Customs Code. BTI decisions "are generally valid for 3 years throughout the EU" and are "binding on all EU customs administrations and the holder of the decision". They can be annulled, revoked or cease to be valid before the three years end, and they cannot be amended. The Commission's general information on applications states that BTI decisions are free of charge, though customs may pass on specific costs such as an expert analysis.

You apply through the EU Customs Trader Portal, or the national portal in the Member States that run their own system, with one application for each type of product. The Commission's information sheet on BTI under the Union Customs Code (2019) adds that, once the application is accepted, the customs authority has 120 days to issue the decision, and that the holder must declare the BTI in the customs declaration. All valid and invalid BTI decisions are published in the public EBTI database, with confidential information protected.

The AI helps with the application, which asks for exactly what a good classification record contains.

Commission general information on EBTI applications (boxes and their content; Box 10 is treated as confidential), read 25 October 2026; the third column is our description.

A BTI is worth its effort for products you ship often, products where a wrong code costs duty, and products where your classification and the supplier's differ. For the others, the internal record with its reasoning is the documentation. Earlier BTI decisions found in the public database can go into the AI's search library, so its proposals cite how similar products were classified.

We set up classification assistants on company AI servers under our Private AI/ML service; send us a sample of your article master and the codes you use today, and our team will show you how proposals with reasoning would look for your products.

When the nomenclature changes

Each new CN version needs a check of your existing codes. The AI can compare the codes in your article master with the new version, list the codes that no longer exist or whose text has changed, and propose new codes with reasoning for those articles. A person decides each one before the new year's first declaration. BTI decisions held by your company may cease to be valid when the legislation changes, so the same check covers them.

Classification regulations and new BTI decisions published during the year can change the picture for a product as well. Adding them to the search library keeps proposals current; the record of each decision shows which version of the texts it used.

Export documents and export control

The CN code appears on the export declaration and on the documents that go with a shipment, and the same code feeds the statistics on trade between EU countries. Using one recorded code per article across invoice, packing list and declaration avoids the contradictions a customs check notices.

The tariff code does not answer whether a product needs an export licence. Dual-use controls work with their own list and their own entry numbers, which our guide to the end-user certificate for NVIDIA GPUs explains for our own deliveries. A classification assistant can flag products that may fall under such controls for the export control officer, but that check is a separate decision.

Why product data stays in the company

A good classification needs the facts that companies keep to themselves: compositions, formulas, functions of unreleased products, supplier names and the share of each material. The BTI application treats the commercial data in Box 10 as confidential for the same reason. With the model on your own server, the article master and the reasoning records stay on hardware you control. Our article on trade secrets, NDAs and AI tools covers the business side of that choice.

Which model and which server

Classification needs a model that reads a page of product data and a few pages of nomenclature text and reasons over them, so the requests are short. Qwen3.8-27B, 27 billion parameters under the Apache 2.0 licence by its Hugging Face model card, is a good fit. NVIDIA publishes an NVFP4 build of it with 21.92 GB of weights (20.42 GiB), tested by NVIDIA on GB300; the official FP8 build holds 30.9 GB (28.75 GiB).

How much memory. By our users-per-card rule, 90 per cent of the memory the driver reports less 3 GiB per card and less the weights, a 32 GB RTX PRO 4500 Server Edition keeps 25.7 GiB, as our guide to Qwen3.8-27B on the RTX PRO 4500 Server Edition derives. The NVFP4 build leaves 5.28 GiB, and with an FP8 cache a request of 8,192 tokens needs about 0.39 GiB, so one card handles 13 classification requests at once. A 96 GB RTX PRO 6000 Server Edition keeps 83.04 GiB and runs the FP8 build with 138 requests of 8,192 tokens or 47 of 32,768.

How fast. In its review published on 14 October 2026, StorageReview measured with vLLM 0.27.1 153 output tokens per second on one RTX PRO 4500 Server Edition for an NVFP4 build of Qwen3.8-27B at 32 parallel requests of 8K tokens in and 1K out; the article does not name the checkpoint. If a proposal with its reasoning takes 400 output tokens, an assumption, that is about 0.38 proposals per second, or about 1,380 an hour by our estimate. Rechecking a catalogue of 5,000 articles at the change of the CN version would take under four hours of one card by the same estimate.

Configurations from our AI servers page and our guide to buying a 4× RTX PRO 4500 Server Edition server; requests by our users-per-card rule (Qwen3.8-27B NVFP4 20.42 GiB and FP8 28.75 GiB with FP8 cache; gpt-oss-120b 60.8 GiB with 16-bit cache); power by our estimate (cards, two processors and 15 per cent for the rest), confirmed by a measurement on the delivered system.

Four RTX PRO 4500 Server Edition cards. A 2U server with four single-slot 32 GB cards, two processors and at least 256 GB of system memory, at about 1.4 kW at the wall by our estimate with the cards at 165 W and two 280 W processors. Classification takes one card; the others run document search, order entry or a second office task, as our guide to AI order entry from emails and PDFs describes. Our guide to buying a 4× RTX PRO 4500 Server Edition server lists the parts.

Two RTX PRO 6000 Server Edition cards. The entry server on our AI servers page: two NVIDIA RTX PRO 6000 Blackwell Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives, about 1.8 kW. One card runs the classification model and other document work, the other a larger assistant for the whole company.

Both servers belong in a server room or a data centre, as our guide to a GPU server in an office explains. Without a server room, our Private AI/ML service runs the platform on dedicated hardware in a Tier-3 data centre in Lithuania, inside the EU.

Notes for the IT team

1. Versioned library. Load the CN of each year as its own version, with the explanatory notes and your saved BTI decisions, and pass the version valid on the declaration date with each request; keep earlier versions for the records.
2. Structured output. Request the classification record as JSON against a fixed schema; vLLM accepts a JSON schema with each request, and its documentation states that "the output will follow the JSON schema". Check by software that the proposed code exists in the CN version used.
3. NVFP4 on the RTX PRO 4500. NVIDIA tested its NVFP4 build on GB300; StorageReview ran an NVFP4 build of the model on the RTX PRO 4500 Server Edition. Test the exact checkpoint and serving engine on the target card before production.
4. Master data link. Write decided codes back to the ERP through one interface, with the person and date; the model never writes to the ERP itself.
5. Network. vLLM's --api-key protects only paths under /v1, /v2, /inference and /cohere (vLLM 0.32.0 security guide), so keep the engine on an internal network behind an LLM gateway with keys and a query log.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with RTX PRO 4500 Server Edition or RTX PRO 6000 Server Edition cards, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, sets up the model, the nomenclature library and the link to your article master. Tell us how many articles you classify and which systems hold your product data, and we propose a setup.

FAQ