AI speeds up HS code classification when it acts as an assistant to the specialist: it extracts product characteristics from invoices, specifications, and product cards, suggests several code candidates with reasoning based on the General Rules for the Interpretation of the Harmonized System and the notes, and shows what information is missing. The final code is chosen by a person: under the EAEU Customs Code, classification in customs declarations is performed by the declarant, and the customs authority checks its accuracy. So the project goal is not “a neural network that knows the code,” but a process in which every decision can be explained to customs.
Prepared by AIrassvet and verified against open sources on October 1, 2026. It is intended for importers and exporters, international trade department staff, customs brokers, distributors, and sellers who import goods. The rules are cited according to the texts of the EAEU Customs Code, Federal Law No. 289-FZ, and the Administrative Offenses Code of the Russian Federation in ConsultantPlus as of the verification date. This is not legal or customs advice: confirm the classification of a specific product with a specialist or through a binding tariff information ruling. The AIrassvet section at the end describes the publisher’s services.
Contents
- Why HS classification is a good fit for AI
- Legal framework: who is responsible for the code
- How AI applies the interpretation rules
- What product data is needed
- Solution architecture
- Where AI helps in the international trade process
- Off-the-shelf service or custom system
- Risks and limitations
- Step-by-step implementation
- How to measure quality
- Frequently asked questions
- Primary sources and verification method
Why HS classification is a good fit for AI
Code selection is a text-heavy task that relies on a large reference system. The specialist reads the product description, identifies the material, purpose, and operating principle, then looks for the appropriate commodity item among the nomenclature headings, checks the section and chapter notes, explanations, and classification decisions in practice. This task has properties that are well suited to language models:
- Semantic search. The supplier’s description rarely matches the wording of the commodity item: “smart Wi-Fi socket” has to be mapped to the language of the nomenclature.
- Lots of unstructured documents. Invoices, specifications, and product passports in different languages.
- Repetition. A distributor’s or manufacturer’s assortment is largely repetitive, and past decisions are a valuable knowledge base.
But the cost of an error is high: the wrong code can change the duty rate, document requirements, and trigger additional assessments and penalties. So AI should accelerate analysis, not replace the specialist’s responsibility.
Legal framework: who is responsible for the code
| Rule | What it establishes | Why it matters for AI |
|---|---|---|
| Decision of the EEC Council dated 09/14/2021 No. 80 | Approves the common EAEU HS and the Common Customs Tariff; the nomenclature includes the General Rules for Interpretation and notes and is regularly amended by EEC decisions | The reference database for AI must be updated together with the nomenclature; the model must not rely on “memory” |
| Article 20 of the EAEU Customs Code | Declarants perform classification when filing the declaration; customs authorities verify its accuracy and, if an error is found, classify the goods themselves; if exact data is unavailable, the code may be determined at no less than the first 4 digits | Responsibility for the code remains with the company; AI should show at what level the code is determined with confidence |
| Articles 23–25 of the EAEU Customs Code | A binding classification ruling is issued no later than 90 calendar days from the date the application is registered and is valid for 3 years unless national law sets a longer term | For disputed and high-volume items, it is safer to obtain a customs ruling than to rely on any algorithm |
| Article 18 of Federal Law No. 289-FZ | In Russia, a binding classification ruling is valid for five years from the date it is issued | Adopted rulings should be stored in the knowledge base and used as a priority source |
| Part 2 of Article 16.2 of the Administrative Offenses Code of the Russian Federation | Declaring false information about the classification code, if it was or could have been used as a basis for exemption from payment or underpayment of duties and taxes above the established threshold, results in a fine for legal entities ranging from one-half to twice the amount of the payments due | Every AI suggestion must be reviewed by a human and retain its justification |
Key takeaway: AI does not change the responsibility structure. It helps prepare a well-founded decision faster, which the specialist approves and can defend.
How AI applies the interpretation rules
General Rules for the Interpretation of the HS (GRI) are six rules within the HS that determine how a product is assigned to a heading and then to a subheading. A good AI assistant follows the specialist’s logic and shows it, rather than returning a bare code.
| Step | What the specialist does | What AI does |
|---|---|---|
| 1. Understand the product | Identifies material, purpose, operating principle, and completeness | Extracts characteristics from documents and creates a list of missing information |
| 2. Find candidate headings | Searches for headings by text and notes (GRI 1) | Finds candidates through semantic search across heading texts, notes, and explanations |
| 3. Resolve competition | Applies rules for incomplete goods, mixtures, sets, and choosing between headings (GRI 2–3) | Explains which rule applies and why one candidate is more specific than another |
| 4. Choose the subheading | Compares subheadings at the same level (GRI 6) | Walks the code tree to 10 digits and marks the levels where information is missing |
| 5. Check practice | Cross-checks with classification rulings and binding rulings | Finds similar rulings in the database and shows discrepancies |
| 6. Record | Approves the code and justification | Saves the reasoning chain, sources, and nomenclature version |
Rule for the model: no code without a reference to the text of the heading or note. If AI cannot cite the basis, it must return “expert review needed” instead of guessing.
What product data is needed
The quality of classification depends more on the description than on the model. A description like “spare parts” or “plastic products” does not allow either a person or AI to classify the product. Minimum set of characteristics:
- Product name and commercial description in Russian and, if necessary, in the supplier’s language.
- Material and its share for composite products.
- Purpose and scope of use: what it is used for, and in what equipment.
- Operating principle for machines, instruments, and electronics.
- Contents and packaging: set, spare part, incomplete product.
- Technical parametersthat determine the subheading: power, voltage, size, composition.
- Photos and technical documentationif the description is ambiguous.
AI is already useful at this stage: it reads invoices and specifications, including scans, and generates a request to the supplier for missing characteristics. For more on document recognition, see the article OCR and Document Recognition for Business.
Solution Architecture
- Reference database. The current version of the EEU Customs Code with notes, explanations, classification rulings, as well as your own declaration history and any preliminary rulings received. Each document is stored with a date and version.
- Extracting characteristics. OCR and a language model turn supplier documents into a structured product card.
- Finding candidates. Semantic search across the database (RAG) finds tariff lines, notes, and similar past decisions. For more on the approach, see the article RAG System for Business.
- Reasoning under the General Rules of Interpretation. The model compares candidates against the rules and returns 2–3 options with justification, a confidence level, and a list of missing information.
- Control rules. Deterministic checks: the code exists in the current version, matches the previous decision for the same item number, and has not changed without cause.
- Specialist approval. An interface where the specialist reviews candidates and sources, selects the code, and records the justification.
- Log and training. Approved decisions expand the database, and discrepancies between the AI and the specialist are reviewed to improve the instructions.
If your product catalog is maintained in 1C or an ERP system, it is convenient to store the code and justification in the item master record. A similar task of matching descriptions is covered in the article Matching Product Catalogs with AI in 1C.
Where AI Helps in the Foreign Trade Process
| Task | What AI Does | Who Makes the Decision |
|---|---|---|
| Classifying new items | Code candidates with justification under the General Rules of Interpretation | Classification specialist |
| Assortment review | Finds items with suspicious or outdated codes after product catalog changes | Specialist, broker if needed |
| Preparing descriptions for the declaration | Creates a description with the characteristics that matter for the code | Declarant |
| Supplier requests | Prepares a list of missing characteristics | Purchasing manager |
| Verifying shipping documents | Compares the invoice, specification, and product card | Foreign trade specialist |
| Preparing an application for a preliminary ruling | Compiles the description and technical information for the disputed item | Specialist, lawyer |
For supply-related scenarios, see the article AI for Procurement and Tenders.
Off-the-Shelf Service or a Custom System
There are online services on the market that use AI to assign an HS code based on a product description. They are convenient for one-time tasks and testing hypotheses. A custom system makes sense when:
- the assortment is large and regularly updated;
- integration with 1C, ERP, and your declaration history matters;
- you need your own control rules and a justification log;
- product descriptions and contracts cannot be shared with external services.
Regardless of the option you choose, test the service’s accuracy on your own sample set: accuracy claims on vendor websites do not replace testing on your assortment. If the documents contain trade secrets, consider deploying the model within your company’s environment — a comparison of approaches is in the article Cloud or On-Prem LLM: TCO.
Risks and Limitations
| Risk | How It Appears | How to Reduce It |
|---|---|---|
| Confident error | The model gives a plausible but incorrect code | Only candidates with a cited basis, specialist approval |
| Outdated nomenclature | The code has been changed or removed by an EEC decision | Regular reference data updates, code check against the current version |
| Incomplete description | Code selected by name only, without specifications | Required card fields; answer “expert review needed” when data is incomplete |
| Adopting someone else’s practice | A decision based on a similar but different product | Showing differences from the solutions found |
| Data leak | Contracts and specifications are sent to an external service | On-premises deployment, anonymization, and a service provider agreement |
| Hallucinations in the rationale | A reference to a nonexistent note | Citation verification against the database — more in the article RAG accuracy and hallucinations |
Step-by-step implementation
Step 1. Choose a process area
Start with one workflow: new items from suppliers, checking the current assortment, or preparing product descriptions. Record how much time is currently spent on one item and how often customs adjusts the codes.
Step 2. Build a benchmark sample
Take several hundred items with confirmed codes: released declarations without adjustments, preliminary decisions, items reviewed by an expert. Add complex cases: sets, spare parts, multi-function products.
Step 3. Prepare the reference database
Upload the current version of the HS code nomenclature with notes and explanations, your own decisions, and history. Assign someone to update it after classification changes.
Step 4. Configure and measure
Compare AI suggestions against the benchmark at the 4-, 6-, and 10-digit levels. Separately assess how often the system honestly says “expert review needed.”
Step 5. Pilot alongside a specialist
The AI suggests candidates, the specialist makes the decision. Track time per item and discrepancies, review them, and refine the instructions.
Step 6. Integration
Connect the system to item master records in 1C or an ERP, to the customs declaration preparation process, and to the database of preliminary decisions.
How to measure quality
| Metric | How to calculate | Why it matters |
|---|---|---|
| 4-digit accuracy | Matches to the benchmark by HS heading / all items in the sample | Whether the group and heading are correct |
| 10-digit accuracy | Exact matches / all items | Code readiness for declaration |
| Candidate hit rate | Benchmark code among 3 candidates / all items | How useful the suggestion is for the specialist |
| Share of honest refusals | “Expert review needed” for incomplete descriptions / all incomplete descriptions | Safety: the system does not guess |
| Time per item | Median before and after implementation | Time savings for the specialist |
| Customs adjustments | Changed codes / all declared items during the period | Overall process quality |
Calculate the metrics on your sample before launch and regularly afterward: there is no universal “AI accuracy for HS codes”; it depends on the assortment and the quality of the descriptions.
Frequently asked questions
Can a neural network determine the HS code on its own?
A neural network can suggest a code and provide a rationale, but it cannot take responsibility. Under the EAEU Customs Code, the declarant performs the classification when filing the declaration, and customs verifies it. That is why AI is used as an assistant, while a specialist approves the code.
What accuracy does AI provide when selecting an HS code?
It depends on the assortment, the quality of the descriptions, and the reference database. Test accuracy on your own sample with confirmed codes separately at the 4-digit and 10-digit levels. Vendor claims about accuracy do not replace this test.
What are the penalties for an incorrect HS code?
If an incorrect code caused or could have caused underpayment of duties and taxes above the statutory threshold, Part 2 of Article 16.2 of the RF Code of Administrative Offenses provides for a fine for legal entities from one-half to twice the amount of payments due. In addition, customs may reclassify the goods and assess additional payments.
How is AI different from searching for a code in a reference directory?
A directory search finds items by keywords, while AI matches the meaning of the description with the wording of headings and notes, extracts characteristics from documents, explains the choice based on interpretation rules, and identifies what information is missing.
When is it better to obtain a preliminary classification ruling?
For disputed, high-value, or high-volume imported items. Under the EAEU Customs Code, the ruling is issued no later than 90 calendar days from the date the application is registered, and in Russia it is valid for five years. AI can help prepare the product description for the application.
Can contracts and specifications be uploaded to a public AI service?
That depends on the service terms and trade secret requirements. For confidential documents, companies use models deployed within the company’s environment or services with verified data-processing terms, and only product characteristics are sent in the prompt.
Primary sources and verification method
Verified on October 1, 2026:
- EEC: EAEU HS and EAEU CET — approved by EEC Council Decision No. 80 dated 09/14/2021, including the General Rules for Interpretation and notes.
- EEC: GRI Rule 6 — classification at the subheading level.
- EAEU Customs Code, Article 20 — classification by the declarant, customs verification, code of at least 4 digits.
- EAEU Customs Code, Article 25 — 90 calendar days for issuance and 3 years of validity for a preliminary decision.
- Law No. 289-FZ, Article 18 — five-year validity of a preliminary decision in Russia.
- Code of Administrative Offenses of the Russian Federation, Article 16.2 — liability for false declaration.
The step sequence, architecture, task tables, risks, and metrics are based on an editorial analysis by AIrassvet. Online code-selection services were not tested for this article, and their accuracy claims are not cited.
How AIrassvet helps automate product classification
If the foreign trade team spends a lot of time analyzing descriptions and searching for the code, automation is best started with one workflow and a benchmark sample. On the AIrassvet website, which also uses the name “Vladimir Lomtev’s AI Studio,” the directions applicable to this task are described:
- Recognizing invoices, specifications, and product data sheets with computer vision and extracting product attributes.
- A RAG-powered search base for item master data, notes, explanations, and your past decisions, with source citations.
- Deploying the model on company servers if contracts and specifications must stay inside your secure environment.
- Audit the classification process, prepare a reference dataset, integrate with 1C or ERP, and test and train your team.
The first step is to choose one workflow, such as new items from suppliers, describe the current process, data sources, constraints, and the criterion you will use to decide whether the pilot was successful.
Result: AI prepares the justification, and the specialist approves the code
AI for selecting HS codes is useful where there are many descriptions, documents, and repetitive product assortments: it extracts attributes, suggests candidates based on interpretation rules, and finds similar decisions. The legal framework sets the boundary: the declarant classifies, customs verifies, and an inaccurate code can lead to liability.
Start with a reference dataset, measure accuracy at the 4-digit and 10-digit levels and the share of valid rejections, keep the reference guide up to date, and retain the justification for each decision. That way, AI will speed up foreign trade operations without increasing customs risk.