Back to Blog
Compliance

How AI-Powered HS Code Classification Actually Works (And Its Limits)

AI can classify products into HS codes with impressive accuracy — but it's not magic. Here's what's happening under the hood, when to trust it, and when human review is essential.

Engineering Team December 20, 2025 11 min read
HS Codes AI Classification Trade Compliance

The Classification Challenge

The Harmonized System contains over 5,000 six-digit headings and 200,000+ national tariff lines. Classifying a product correctly requires understanding not just what it is, but what it's made of, what it's used for, and sometimes how it was manufactured.

Human customs brokers spend years learning classification rules. Can AI replicate this expertise? Partially — and understanding the boundaries is critical for compliance.

How Our Classifier Works

ComplianceGrid's /v1/hs/classify endpoint uses a multi-stage approach:

Stage 1: Description Normalization

The input description is cleaned, expanded, and enriched. "SS kitchen knife" becomes "stainless steel kitchen knife for food preparation." This stage uses an LLM to resolve abbreviations, infer materials, and add context.

Stage 2: Candidate Retrieval

We search our indexed HTS database (sourced from USITC) using semantic similarity to find the top 20-50 candidate headings and subheadings. This narrows the search space from 200K+ tariff lines to a manageable set.

Stage 3: GRI-Guided Classification

The General Rules of Interpretation (GRI) are the legal framework for HS classification. Our system applies these rules in order:

  • GRI 1: Classification by terms of headings and section/chapter notes
  • GRI 2(a): Incomplete or unfinished articles classified as complete if they have the essential character
  • GRI 2(b): Mixtures and combinations of materials
  • GRI 3: When two or more headings apply, use the most specific heading
  • GRI 4-6: Residual rules for goods not classifiable by the above

Stage 4: Confidence Scoring

The system outputs a confidence score based on how clearly the product maps to a single tariff line. HIGH confidence means there's one obvious classification. LOW means multiple headings could apply, and human review is recommended.

When to Trust AI Classification

AI classification works well for:

  • Clearly defined, single-material products (e.g., "cotton t-shirt", "stainless steel bolt")
  • Products with unambiguous intended use
  • Commodities with established classification precedent

When Human Review Is Essential

Always involve a licensed customs broker or classification specialist for:

  • Multi-material or composite products
  • Products with dual-use potential (civilian and military applications)
  • High-value shipments where misclassification has significant duty implications
  • Novel products without classification precedent
  • Any classification flagged as LOW confidence

Practical Recommendation

Use AI classification as a first pass to narrow down candidates and speed up the workflow. Then have qualified personnel validate the result, especially for production classifications that will appear on customs entries. The AI gets you 80% of the way in seconds — the remaining 20% requires human expertise.