AI in Trade Compliance: Practical Guide for Importers
What changed: Clarified the September 2025 ATLAS benchmark, its test set and limits, removed unsupported accuracy comparisons, and corrected the relevant penalty and documentation claims.
It speeds up the research steps that used to take a specialist hours: suggesting HTS codes with the reasoning behind them, reading invoices and packing lists into structured data, matching names against restricted-party lists, and checking new tariff rules against the products a company imports. Done well, it applies the classification rules in order and asks for facts it is missing rather than guessing. It does not replace the licensed customs broker who decides the code on an entry, and general-purpose chatbots are not built for the classification step at all.
In this post
- Can general AI tools like ChatGPT or Copilot classify HTS codes?
- Why do general AI tools fail at HTS classification?
- What does CBP's ruling HQ H350722 say about AI classification tools?
- How does GingerControl build AI classification?
- Where else is AI already working in trade compliance?
- What should importers ask any AI compliance vendor?
- Does the EU AI Act apply to customs classification?
- Frequently Asked Questions
- Putting AI to work on classification
- References
Can general AI tools like ChatGPT or Copilot classify HTS codes?
Use a chatbot's code as a research candidate and verify it against the product facts, current HTS, GRI, Notes and relevant CBP rulings before filing. In Flexify.AI's ATLAS paper published September 22, 2025, GPT-5-Thinking matched 25.0% of 10-digit codes and 55.5% at 6 digits on 200 CROSS-derived test examples; Gemini-2.5-Pro-Thinking matched 13.5% and 31.0%, and the authors' fine-tuned Atlas model matched 40.0% and 57.5% [1]. These historical results apply to that test, not every current chatbot or classification service. A code put on an entry remains the importer's responsibility.
Why do general AI tools fail at HTS classification?
- They skip the legal method. HTS classification follows GRI 1 to 6 in sequence, with Section and Chapter Notes that include or exclude whole groups of goods. GRI 3(b) essential character is weighed by the factors in the Explanatory Note: the nature of each material or component, its bulk, quantity, weight or value, and its role in the use of the goods [5]. A general model knows about these rules; it does not apply them as rules.
- They invent codes. A model writes the most likely next text. A 10-digit number in the right format can still be a code that does not exist, or one for a different product.
- They do not read rulings first. CBP's CROSS database holds the rulings that show how CBP classified similar goods. A general chatbot does not consult it before answering.
- They answer instead of asking. Classification often turns on a fact the description leaves out: the share of each material, how the product is used, what is in a set. A chatbot answers anyway.
- They leave no usable record. If CBP sends a CF-28 or audits, the importer has to show reasonable care. A chat transcript is not a documented classification method.
Compare bulk tariff code validation systems by how they test current codes and route the exceptions for review.
What does CBP's ruling HQ H350722 say about AI classification tools?
On January 16, 2026, CBP issued HQ H350722 on an online platform that offered AI-generated HTSUS classifications, among other services, from an unlicensed company [2]. The points that matter for anyone using or building such tools:
- Six digits versus ten. CBP found that giving classifications at the 6-digit HS level is not customs business, while giving 8- or 10-digit HTSUS classifications for goods that have been or will be imported is, and requires a licensed broker.
- The link to an entry decides. CBP contrasted a general classification database open to any product, which it had found permissible with a meaningful disclaimer (HQ H272798), with classifying specific goods a client has bought and will import, which it had found to be customs business (HQ H290535) [3].
- A disclaimer alone does not cure it where the tool gives specific subheadings for specific goods to be entered.
- A tool is not a person under the broker regulations, so the classification decision for imported goods has to be made by a licensed customs broker, however capable the software.
The practical reading for importers: AI output is research for a licensed broker to review, not a code to file on its own.
How does GingerControl build AI classification?
GingerControl's Classification Researcher, in the GingerLite app, works the way the method above asks:
- It narrows candidates and asks. It narrows the candidate codes by asking about the facts that separate them before it settles; "Ginger doesn't guess. It asks."
- It reads rulings during the classification. Similar CROSS rulings inform the decision, not a citation added afterwards, and the code is determined in stages, 4, 6, 8 and 10 digits.
- It records the reasoning. Each case ends with a report: the GRI applied, the Section and Chapter Notes, the rulings read and, for a composite product, which component gives it its essential character and why. One case takes 3 to 5 minutes with its report.
GingerControl's classifications are for general reference, educational and planning purposes; before a code goes on an entry, your licensed customs broker reviews it. GingerControl is not a customs broker and does not file entries. For whole catalogs, GingerOS classifies in batches and records the reasoning for your approval; the Classification + Tariff Calculation API, for enterprise systems, returns codes and U.S. duty layers without a reasoning report.
Where else is AI already working in trade compliance?
- Documents to data. Models read invoices, packing lists and bills of lading into structured fields for a broker to review. HQ H350722 also found that an unlicensed company's OCR tool preparing entry data was customs business, so who runs this step matters [2].
- Screening. Name matching that handles spelling variants and transliterations cuts false alerts in restricted-party screening.
- Rule changes against products. Checking each new tariff notice against the HTS codes a company imports, so only relevant changes reach the team. GingerControl's Compliance Radar does this for U.S. notices.
The Compliance Radar alerts matched to your HTS codes workflow ties each policy notice to the product codes in the monitored catalog.
What should importers ask any AI compliance vendor?
| Question | Why it matters | GingerControl's answer |
|---|---|---|
| Who decides the code that goes on the entry? | After HQ H350722, a licensed broker | Your licensed broker reviews each code; GingerControl does not file |
| Does it apply GRI and the Notes, and ask for missing facts? | Text matching fails on composite and multi-use goods | The Classification Researcher and GingerOS ask before they settle |
| Are rulings read before the code is chosen? | Precedent should shape the answer | Yes, in the Classification Researcher |
| Does each code come with its reasoning? | Reasonable care needs a record | Yes in the Classification Researcher and GingerOS; the API returns codes and duty layers only |
| What does the accuracy figure measure? | Digit level, test set and who ran it change the meaning | 99.86% at the 10-digit level on GingerControl's own internal test of 1,000+ products, a first-party test |
| Which duty layers does it cover? | A base rate alone understates duty | U.S. base rate, Section 301, Section 232, other Chapter 99 lines, MPF and HMF; AD/CVD is not calculated |
| Does your data train the model? | Product and supplier data are sensitive | Customer data never trains any model |
Does the EU AI Act apply to customs classification?
Probably not as a high-risk use. The EU AI Act (Regulation (EU) 2024/1689) lists its high-risk uses in Annex III; classifying goods for customs is not among them, although border control uses involving people are [4]. Its general duties, such as AI literacy for staff, have applied to companies deploying AI in the EU since February 2, 2025, and the high-risk rules for Annex III uses, first due on August 2, 2026, now apply from December 2, 2027 under the AI Omnibus, in force since July 27, 2026 [7]. For an importer into the U.S., the rules that bite are U.S. ones: reasonable care under 19 U.S.C. 1484 and a licensed broker deciding the code [6].
Frequently Asked Questions
Is AI accurate enough for HTS classification?
For research, purpose-built tools can help; a code still needs review before filing. The September 22, 2025 ATLAS paper reported 25.0% at 10 digits and 55.5% at 6 digits for GPT-5-Thinking, and 40.0% and 57.5% for its fine-tuned Atlas model, on 200 CROSS-derived test examples. GingerControl's 99.86% at 10 digits comes from its own internal test of 1,000+ products, a first-party test. The datasets differ, so these figures are not a direct comparison or a measure of current tools generally.
Can I use ChatGPT or Copilot for HTS classification?
Verify any chatbot-generated code before filing: check that the code exists in the current HTS, that the product facts support it, and that the applicable GRI, Notes and rulings have been considered. If a materially incorrect code goes on an entry through negligence, gross negligence or fraud, the importer can face exposure under 19 U.S.C. 1592.
What did CBP rule about AI classification tools in HQ H350722?
That giving 8- or 10-digit HTSUS classifications for goods being imported is customs business requiring a licensed broker, while 6-digit classifications are not; that a disclaimer does not cure specific classifications for specific entries; and that a software tool is not a person who can make the classification decision.
Does using AI help meet CBP's reasonable care standard?
It can support it when the output is reviewed by qualified people, the reasoning behind each code is kept, and the process is documented. A tool that leaves no record of why a code was chosen does not help, and neither does a chatbot transcript.
Who should decide the final HTS code when AI is used?
A licensed customs broker, for goods that will be entered. AI output is research that narrows the candidates and records the reasoning; the broker reviews it and decides what goes on the entry.
Does the EU AI Act treat customs classification as high risk?
Customs classification of goods is not on the Act's Annex III list of high-risk uses. Companies deploying AI in the EU still have the Act's general duties, and importers into the U.S. answer to U.S. reasonable-care rules.
Putting AI to work on classification
Use AI for the research and keep a licensed broker on the decision. GingerControl's Classification Researcher gives each product a 10-digit code with its reasoning, in the GingerLite app at $150 a month for 300 credits. Open GingerLite.
For a company-wide system, GingerControl builds GingerOS, the AI trade compliance platform that runs on your process and the data you have, ERP or not. It starts with a Prototype Week, the step before a quote. See how a project runs.
References
- ATLAS: Benchmarking and Adapting LLMs for Global Trade via HTS Classification, arXiv 2509.18400
Data cited: Historical 200-example CROSS-derived test; GPT-5-Thinking 25.0%/55.5%, Gemini-2.5-Pro-Thinking 13.5%/31.0%, fine-tuned Atlas 40.0%/57.5% at 10/6 digits; authors at Flexify.AI; GingerControl not tested
Source: arXiv 2509.18400
Published: September 22, 2025 - CBP Headquarters Ruling HQ H350722
Data cited: 6-digit versus 8- and 10-digit classifications; link to a prospective entry; disclaimers; a tool is not a person under 19 CFR 111.1; OCR entry-data tool
Source: CBP HQ H350722
Published: January 16, 2026 - CBP Headquarters Rulings HQ H272798 (January 26, 2017) and HQ H290535 (September 29, 2022)
Data cited: a general classification database is not customs business when a meaningful disclaimer is implemented (H272798); specific subheadings for specific goods need a license even with a disclaimer (H290535)
Source: CBP HQ H272798 and CBP HQ H290535 - Regulation (EU) 2024/1689, Artificial Intelligence Act
Data cited: Annex III list of high-risk uses
Source: AI Act Service Desk, Annex III
Published: July 12, 2024 - CBP Headquarters Ruling HQ 950636
Data cited: Explanatory Note VIII to GRI 3(b) factors for essential character
Source: CBP HQ 950636
Published: January 16, 1992 - 19 U.S.C. 1484, Entry of merchandise
Data cited: the importer of record classifies and values goods using reasonable care
Source: 19 U.S.C. 1484, Legal Information Institute - European Commission, AI Act: application timeline
Data cited: prohibited practices and AI literacy obligations applying from February 2, 2025; AI Omnibus in force July 27, 2026; Annex III high-risk rules from December 2, 2027
Source: European Commission, Regulatory framework for AI
Shortlisting AI for trade compliance?
Hand this to GingerOS.
Pick a 30-minute slot on the next page with Chen Cui, our CEO. Bring one notice or one SKU list (an ERP export, a spreadsheet or an email will do) and see what GingerOS does with it.
Reviewed by Michael Weick, LCB / CCS




