HS code classification: AI suggestions for human review

The Harmonized System is the shared language of customs tariffs. More than 200 countries and economies use it. Codes are international to six digits, and national tariff schedules can add further digits.

Every new product raises the same question: which code? And for a company that imports or ships thousands of products, the answer is usually already in its own records:

• Your broker classified your earlier shipments, and those codes sit in your entries and your ERP.
• Your compliance team made calls on the products that could fit two headings, sometimes backed by a ruling.
• New products are mostly variations of old ones: a new color, a new size, a new supplier.

WCO: the Harmonized System.

Suppose you have 20,000 reviewed products and another 2,000 to classify. That is an illustrative workflow, not a measured result: train on the reviewed items and use confidence to help prioritize the next batch.

StayCharted AI Model Trainer learns from the codes you've filed, on the products you sell. It suggests codes consistent with your past decisions, says how sure it is, and leaves the decision with a person.

Keep your labels aligned with the standard’s definitions and release. A model can repeat past mistakes as well as useful patterns; review the history before training.

1. Export your classified products to CSV or Excel: one row per product, with the columns that describe it (description, material, intended use, your product category) and the code you filed. A model can read several columns together.
2. Choose the level to train on. The six-digit subheading is the same in every country that uses the HS. Your full national tariff line adds digits; train on it if your history is consistent at that level.
3. Check the data before training. Data quality finds duplicate rows, and codes with too few examples.
4. Train, then read the report. Accuracy is measured on products the model never saw during training, with accuracy per code and the codes it mixes up.
5. Suggest codes for new products. Upload a file with the code column blank, or call the model from your systems through the Prediction API. Every row gets a suggested code and a confidence.
6. Send the uncertain ones to a person. Pick a confidence cutoff from the report. Everything below it goes to a Review Queue for your compliance team, or onto the list you send your broker. Their answers become training examples for the next version.

Confidence helps prioritize review; it does not guarantee correctness. Check a sample of high-confidence answers too. Review Queue and Prediction API availability depends on your plan; see current plans. CSV/Excel export and the API are workflow options, not a native PIM or ERP integration.

• It suggests; a person decides. The importer stays responsible for the classification declared to customs. Use the model to prepare and prioritize work for your compliance team or broker, not to replace them.
• It learns your past decisions, including the wrong ones. Wrong codes in your history can teach the model the wrong patterns. Review evaluation errors and source records for inconsistent history.
• It reads text, not the legal notes. HS classification follows General Rules of Interpretation, section and chapter notes, and rulings. The model learns patterns from your descriptions; it doesn't apply those rules.
• Codes change. HS 2028 takes effect on 1 January 2028. Have your compliance team review the WCO amendments, relevant correlation tables and national schedules before updating labels and retraining.

Online retailers and marketplaces selling across borders need a code for every product they ship, and their catalogs change daily. Most of those products are new variants of products already classified.

A model trained on the catalog you've already classified suggests codes for new listings as they arrive, from the title, description and your own category. The products it's unsure about go to a person first. And the same catalog history can train a second model that places supplier products into your own category tree. More on this in e-commerce product categorization.

An assistant can be a useful starting point when you have a few products to explore. It does not automatically know your approved labels or company conventions unless you supply that context. A trained model learns from labeled examples and gives you held-out results to inspect.

Our assistant-versus-trained-model benchmarks cover other tasks, including banking messages and contract clauses. We have not published a benchmark for this taxonomy. Those results do not establish accuracy on your product codes. Compare approaches on representative items from your own catalog before relying on them.

• Start with the codes you use. Begin with the codes represented in your reviewed data. Our published benchmarks go up to 100 categories in one model. For a wide range, train one model per chapter or product line, or design and validate a separate chapter-then-code workflow.
• Check rare codes. A category with only a few examples may be difficult to learn and evaluate. The report shows which ones.
• Descriptions decide it. "Bag" leaves out important facts; including material and intended use gives reviewers more information but may not settle the classification. Include material and use where you have them.

Can AI assign HS codes automatically?

It can suggest them, with a confidence score, from your own classification history. The importer remains responsible for the declared code, so a person should confirm, especially for uncertain or high-value items.

Is a suggested code a binding classification?

No. Only customs authorities issue binding rulings. A suggestion is a starting point for your compliance team or broker.

Should we train on six digits or the full tariff line?

The HS defines the international six-digit level; use the correct edition and jurisdiction. The full national line is more specific; train on it if your history is consistent at that level.

What happens when the HS changes in 2028?

Have your compliance team review the WCO changes, relevant correlation tables and the applicable national schedule. Resolve changed codes before retraining; a correlation is not a substitute for a classification decision.

Is my product data used to train anyone else's model?

No. Each model is trained on your examples for your use and is not pooled with anyone else's data.