Can AI Help With Apparel Tariff Classification?

Apparel classification is an evidence problem before it is a language-model problem.

A short product name—“women’s jacket,” “knit set,” or “linen top”—does not contain enough information to determine treatment under the Harmonized Tariff Schedule of the United States. Construction, fiber by weight, knit or woven structure, garment identity, component relationships, intended use, and legal notes can change the analysis. The applicable schedule and additional duties can also change.

AI can help assemble and compare a research packet. It should not be allowed to invent missing product facts, silently choose an HTS code, or present a duty rate as final.

Freeze the physical product record

Create one versioned classification record for the exact imported article. Include:

• style and variant identifiers;
• complete product description;
• knit, crocheted, woven, nonwoven, coated, or other construction;
• fiber composition by weight with source and verification state;
• garment type and intended use;
• coverage, length, openings, fasteners, pockets, lining, panels, collar, sleeves, waistband, and distinctive construction;
• gender or age category only where the legal provision requires the relevant distinction;
• whether pieces are imported, packaged, and sold together;
• trims, coatings, reinforcement, protective features, or accessories;
• country-of-origin evidence stored separately from classification;
• front, back, interior, closure, hem, label, and detail images;
• physical sample location;
• manufacturer specification and bill of materials;
• date, HTS edition, researcher, and reviewer.

Do not ask AI to infer weave, fiber percentage, component weight, or construction from a marketing image. Record “unknown,” stop the analysis, and obtain supplier documentation, a sample, or testing as appropriate.

Research in an authoritative order

Start with the current official HTS, including section and chapter notes. The U.S. International Trade Commission publishes the schedule and maintains a current search tool, archive, training, and change information. On August 31, 2026, the USITC page identified 2026 Revision 17 as published August 24; that date is evidence of why every analysis needs an edition and access date, not a permanent shortcut.

Next, search Customs and Border Protection’s CROSS database for published rulings involving materially similar facts. Similar vocabulary is not enough. Compare construction, fiber, use, components, and the legal reasoning. Record ruling number, date, status, factual similarities, factual differences, headings considered, and whether the ruling has been modified, revoked, or referenced.

CBP says CROSS is a searchable collection of rulings and notes that the collection is still being completed. Search results are research evidence, not a guarantee that every relevant ruling was found.

Give AI bounded research tasks

Appropriate assistance can include:

• extracting known facts from approved specifications into a fixed schema;
• identifying missing fields without filling them;
• generating search synonyms for a researcher;
• comparing the factual fields of selected rulings;
• checking that citations, dates, and links are present;
• producing a question list for a customs professional;
• monitoring a saved list of official sources for changes.

Prohibit:

• generating a code from a product title alone;
• inventing a fiber, construction, use, ruling, note, or rate;
• treating a retailer description as legal evidence;
• copying confidential supplier files into an unapproved model;
• making origin, valuation, trade-remedy, quota, admissibility, or preference decisions;
• filing an entry or changing master data without authorized human approval.

Every AI statement should point to the source field or official text that supports it. If the evidence packet changes, rerun the research instead of editing only the conclusion.

Separate classification from adjacent determinations

An HTS classification does not, by itself, decide country of origin, customs value, marking, admissibility, special-program eligibility, additional duties, quota, forced-labor compliance, or product-safety requirements. Maintain separate owners and records for those determinations, then connect them at the import transaction.

Duty rates shown in a schedule or ruling may not be the full amount payable for a current entry. CBP’s binding-ruling guidance states that binding classification advice comes from its Office of Regulations and Rulings and notes that tariff classifications are binding while duty rates are not. Qualified customs review must use current facts and law.

A reproducible readiness audit

FashionMember created four fictional apparel records in content/data/FM-034-classification-intake.json. The script scripts/fm034-classification-intake.php checks whether a packet has core product facts, construction detail, a physical sample, a dated HTS edition, logged CROSS research, and an assigned customs reviewer.

Two fictional records are research-ready for expert review. One is held because the fiber breakdown is pending and no CROSS research is logged. Another is held because construction detail and a physical sample are missing.

“Research-ready” does not mean classified. The script contains no tariff-code field and returns no duty rate. It does not determine origin, value, admissibility, trade-program eligibility, marking, or binding treatment.

Create a human approval record

The final file should show:

1. the exact imported product and evidence version;
2. schedule edition and notes reviewed;
3. rulings and official guidance considered;
4. alternatives considered and why they were rejected;
5. unresolved facts and assumptions;
6. classification selected by the authorized reviewer;
7. separate origin, value, and additional-duty records;
8. effective and recheck dates;
9. broker instructions and acknowledgment;
10. whether a binding ruling should be requested.

CBP’s Binding Ruling Program accepts detailed requests, and a sample may be provided. A binding ruling is a formal process, not a prompt result. When facts are ambiguous, stakes are material, or treatment is novel, discuss the ruling path with qualified customs counsel or a broker.

Monitor after the first entry

Classification governance continues after setup. Recheck when product construction, fiber, component set, use, supplier, origin facts, packaging, or import program changes. Monitor HTS revisions, Customs Bulletins, Federal Register notices, and ruling modifications.

Reconcile the approved classification record with purchase orders, commercial invoices, broker entries, and landed-cost systems. Investigate overrides. Preserve enough documentation to explain what the importer knew and which sources were used at the time.

AI can reduce clerical friction in that evidence chain. Its useful output is a clearer, more complete packet for accountable review—not a confident code detached from the garment.

Sources and verification

• USITC Harmonized Tariff Information — official current HTS access, revision announcements, archive, training, and tariff resources.
• USITC HTS Search — official searchable schedule; users must check the current edition, notes, and applicable provisions.
• CBP Binding Ruling Program — official explanation of binding classification advice, request routes, samples, CROSS, and related publications.
• CBP CROSS — official Customs Rulings Online Search System for published rulings.
• CBP CROSS access guidance — official description of the database, collections, search functions, and completeness caveat.
• NIST AI RMF Core — official voluntary framework used for source traceability, evaluation, monitoring, and human accountability principles.