ProcessLogistics
AI HS code classification that stands up to a customs audit
Tariff classification is legal reasoning, not a lookup. AI can draft candidate codes for thousands of SKUs, but each decision must follow the General Interpretative Rules in order, cite the notes and rulings it relied on and carry a reviewer's sign-off. Here is a five-step process that produces codes a customs authority can audit, with routing rules and a worked hypothetical.
On this page
- Why tariff classification resists simple automation
- The legal method: GIR 1 to GIR 6, applied in order
- How a SKU moves through an AI-assisted classification pipeline
- Five steps from product record to defensible tariff code
- Routing rules for AI-proposed tariff codes
- Worked example: a smart thermostat with a Wi-Fi module
- Keeping a classified catalog current
- Questions and answers
- Sources
Why tariff classification resists simple automation
The Harmonized System is maintained by the World Customs Organization, identifies each commodity group with a six-digit code and is used by more than 200 countries and economies as the basis of their tariffs1.
National schedules add digits of their own. The EU Combined Nomenclature extends the HS to eight digits and is republished every year2, with TARIC measures on top, while the United States classifies under the Harmonized Tariff Schedule (HTSUS)9. The HS itself is revised in review cycles, and the next edition, HS 2028, enters into force on 1 January 20283. Each revision can move products between headings, so last year's correct catalog may be wrong next year.
The deeper difficulty is that codes are defined by legal text, not product names. Two near-identical devices can land in different headings because of one component or the way they are put up for sale. Responsibility stays with the trader: in the United States the importer of record must use reasonable care when declaring classification4, and in the EU the person lodging a declaration answers for its accuracy under the Union Customs Code7. An AI system therefore has to produce the reasoning, not just the answer.
The legal method: GIR 1 to GIR 6, applied in order
Customs authorities apply the General Interpretative Rules in sequence, with the section and chapter notes. The WCO Explanatory Notes are not legal text, but authorities and courts treat them as an important aid to interpretation, so a rationale that cites them is stronger.
- GIR 1
- Classification follows the terms of the headings and any relevant section or chapter notes; titles are for reference only. Most goods are settled here, and the notes often resolve apparent conflicts.
- GIR 2
- Extends a heading to unfinished or unassembled articles that have the essential character of the finished one (Rule 2(a)), and to mixtures and combinations of materials (Rule 2(b)).
- GIR 3
- For goods classifiable under two or more headings: the most specific description wins (Rule 3(a)); otherwise composite goods and retail sets follow the component giving essential character (Rule 3(b)); failing both, the heading last in numerical order (Rule 3(c)).
- GIR 4
- Goods the earlier rules cannot place go to the heading for the goods they are most akin to. A model proposing it should always trigger specialist review.
- GIR 5
- Specially shaped cases presented with an article, and packing normally used for the goods, are generally classified with the goods.
- GIR 6
- Applies the same logic at subheading level, comparing only subheadings at the same level and using subheading notes.
How a SKU moves through an AI-assisted classification pipeline
- Classification-ready record
The attributes a classifier needs: material, composition, function, use and presentation.
- Retrieve legal text
Search tariff text, notes, Explanatory Notes and published rulings for the SKU.
- Draft candidates and GIR
The model proposes headings, applies the rules in order and cites every passage it used.
- Route by risk
Confidence, precedent and duty spread decide whether a person must review the SKU.
- Analyst or specialist
A qualified reviewer accepts, edits or escalates, and may recommend binding advice.
- Recorded decision
Code, rule applied, sources, reviewer, date and tariff version are stored together.
Five steps from product record to defensible tariff code
Build a classification-ready product record
Classification turns on facts catalogs rarely hold: material composition by weight or value, function, intended use, presentation for sale and whether the item ships as a set. Define the attribute schema per product family, fill gaps from specifications and supplier data, and block classification when a required attribute is missing rather than letting the model guess.
Generate candidates with retrieval over legal sources
Give the model the tariff text, notes, Explanatory Notes and published rulings, such as the EU's EBTI database5 and CBP's CROSS rulings search6, and require it to quote the passages it relies on. Ask for two or three candidate headings with reasons for and against each, and discard any candidate without a cited passage.
Route by confidence, precedent and duty impact
A model can be confidently wrong about a note it never retrieved, so confidence alone is a weak signal. Combine it with precedent, the duty difference between competing codes and any flags for licensing, trade remedies or export controls, as in the routing table below.
Record the rationale for every decision
Store the final code with the deciding rule, the headings rejected, the notes and rulings cited, the attributes relied on, the reviewer, the date and the tariff version. This file is what you show an auditor or customs officer, and how you find every affected SKU when the tariff changes.
Seek binding advice where the stakes justify it
For high-volume or high-duty items with genuine ambiguity, apply for a binding decision. An EU Binding Tariff Information decision is valid for three years from the date it takes effect and binds customs only as against its holder7; US binding rulings are issued under 19 CFR Part 1778. Feed each decision back into the retrieval corpus.
Routing rules for AI-proposed tariff codes
| Signal | Auto-accept | Analyst review | Specialist or binding ruling |
|---|---|---|---|
| Agreement between candidates | One heading clearly ahead and stable across reruns | Two plausible headings in the same chapter | Plausible headings in different chapters or sections |
| Precedent | Identical item already classified internally or covered by a ruling you hold | Similar items ruled on, but the facts differ | No precedent, or rulings that point in different directions |
| Duty spread between candidates | None or negligible | Material, but not decisive for the product's margin | Large enough to change sourcing or pricing decisions |
| Regulatory flags | None attached to any candidate | A license or certificate requirement attached to one candidate | Trade remedies, dual-use controls or sanctions exposure attached to a candidate |
| Product novelty | Known variant of an existing family | New family built on familiar technology | New technology or a multi-function device |
Illustrative thresholds. Set your own with your broker or customs counsel.
Worked example: a smart thermostat with a Wi-Fi module
Keeping a classified catalog current
Questions and answers
Who is liable if an AI system assigns the wrong tariff code?
The importer or declarant, not the software vendor. In the United States the importer of record must use reasonable care when declaring classification4; in the EU, the person lodging the declaration is responsible for its accuracy7. AI does not reduce that duty, but a recorded rationale, a documented review step and evidence of regular audits are what demonstrate care if a code is challenged.
Can AI help with rules of origin as well as classification?
Yes, but origin needs different evidence: the bill of materials, where each input was made and the rule in the trade agreement you want to use. Many of those rules are written as changes in tariff classification between inputs and the finished good, so accurate component codes come first. AI can match agreement rules to bills of materials and flag missing supplier declarations; a specialist should still sign the determination.
How often should a classified product catalog be rechecked?
Whenever the tariff changes, the product changes or a relevant ruling appears, plus a regular sampling audit. In the EU that means at least the annual Combined Nomenclature update; worldwide it means every HS edition, with HS 2028 the next one3. Event-driven checks catch most problems; the audit catches drift in how the model or reviewers apply the notes.
Can we rely on a binding ruling issued to another company?
Not as a binding decision. An EU BTI binds customs only as against its holder7, and a US ruling applies to the facts described in the request. Rulings issued to others are still good evidence of how authorities reason, so they belong in the retrieval corpus, but where the duty at stake is significant you should hold a decision of your own.
Sources
- What is the Harmonized System (HS)? — World Customs Organization · checked 10 October 2026
- Combined Nomenclature — European Commission, Taxation and Customs Union · checked 10 October 2026
- HS Nomenclature 2028 Edition — World Customs Organization · checked 10 October 2026
- 19 U.S. Code § 1484 – Entry of merchandise — Legal Information Institute, Cornell Law School · checked 10 October 2026
- European Binding Tariff Information (EBTI) database — European Commission · checked 10 October 2026
- Customs Rulings Online Search System (CROSS) — U.S. Customs and Border Protection · checked 10 October 2026
- Regulation (EU) No 952/2013 laying down the Union Customs Code — EUR-Lex · checked 10 October 2026
- 19 CFR Part 177 – Administrative rulings — Electronic Code of Federal Regulations · checked 10 October 2026
- Harmonized Tariff Schedule of the United States — U.S. International Trade Commission · checked 10 October 2026