Build vs. Buy: Should You Build Your Own HTS Classification System?
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.
For most importers and platforms, buy rather than build: an HTS classifier has to apply the GRI and the Section and Chapter Notes, use CBP rulings and keep up with a schedule that changed 32 times in 2025. The ATLAS study published September 22, 2025 found that a model fine-tuned on more than 18,000 CBP rulings matched 40% of 10-digit codes, and GPT-5-Thinking 25%. Build only if classification is your product. For a manufacturer with 20,000 SKUs that wants a system shaped to its own process, a third route is to have one fitted rather than built from scratch. This is a historical result for that test set, not a current accuracy ceiling.
In this post
- Should you build your own HTS classification system?
- When does building your own classifier make sense?
- What does building an HTS classifier actually require?
- How well do general AI models classify on their own?
- Build vs buy vs fitted: how do the three compare?
- What does GingerControl offer instead of a build?
- Putting the build-or-buy decision in front of your team
- References
Build vs buy for HTS classification is the choice between writing your own classifier and using someone else's tool, API or system. GingerControl sells the buy options: the Classification + Tariff Calculation API for enterprise systems, the GingerLite app for brokers and small businesses, and GingerOS, a system fitted to a company's own process. I'm Chen Cui, a GingerControl co-founder; we built a classifier, so this post is what we learned about the build side.
Should you build your own HTS classification system?
In most cases, no: build only when classification is the product you sell and you can staff both engineers and licensed classification expertise for as long as the product lives. Everyone else gets further by buying a tool or API, or by having a system fitted to their process.
When does building your own classifier make sense?
Building makes sense when three things are true at once:
- Classification is your product, not a step in someone else's workflow.
- You can staff it permanently with engineers and people who know the GRI, the notes and CBP rulings.
- You can carry the legal line: if your product gives others codes beyond six digits for goods they will import, CBP treats that as customs business (HQ H290535, HQ H350722) [1]. See can AI tools legally classify HTS codes? for the detail.
If only the first is true, an API inside your product is usually the better route: see embedded HS classification for SaaS.
What does building an HTS classifier actually require?
Building one requires the data, the legal reasoning, a test set and an update process; the model is the smallest part.
| Component | What it involves | Source or cost |
|---|---|---|
| The schedule | Every line, rate, note and footnote, revision by revision | USITC HTS and its free API at hts.usitc.gov/reststop [2] |
| Updates | 32 revisions in 2025; 21 in 2026 by October 9; HS 2028 with 299 sets of changes on January 1, 2028 | USITC release list; WCO [2] [3] |
| Precedent | CBP rulings, more than 222,000 searchable in October 2026 | CBP CROSS [4] |
| Interpretation | WCO Explanatory Notes, "the official interpretation of the Harmonized System at the international level" | WCO Trade Tools, 450 euros per user a year [3] |
| Legal reasoning | GRI 1 to 6, Section and Chapter Notes, essential character for composites | People who know them |
| A question step | Asking for the fact that splits two candidate codes | Product design |
| A test set | Your own products with broker-confirmed codes, scored at 6, 8 and 10 digits | Your broker's time |
| A review step and a record | Who approved each code, with what reasoning, kept five years | 19 CFR 163.4 [6] |
If you only need schedule data in your systems, the free USITC API may be all you need; HTS lookup API vs manual classification explains it.
How well do general AI models classify on their own?
The ATLAS paper published September 22, 2025 tested 200 CROSS-derived examples: its fine-tuned Atlas model matched 40% of 10-digit codes and GPT-5-Thinking matched 25% [5]. These are historical results for those models and that dataset, not a current ceiling for LLM classification.
Source: Yuvraj and Devarakonda, ATLAS, arXiv 2509.18400, September 22, 2025
See the data
| Model | 10-digit exact match |
|---|---|
| Atlas (fine-tuned LLaMA-3.3-70B) | 40% |
| GPT-5-Thinking | 25% |
| Gemini-2.5-Pro-Thinking | 13.5% |
| DeepSeek-R1 | 2.5% |
At 6 digits the same models did better (57.5% for the fine-tuned model, 55.5% for GPT-5-Thinking). The study's authors work at Flexify.AI and built the fine-tuned model themselves; they tested no commercial classification vendor, so read it as historical evidence about the tested models on that dataset. It does not establish current product performance. How vendors state their own accuracy is in HTS classification providers compared on accuracy.
Build vs buy vs fitted: how do the three compare?
Building gives you full control and full responsibility; buying gives you speed and someone else's roadmap; a fitted system gives you a system shaped to your process without writing it yourself.
| Build your own | Buy a tool or API | Have a system fitted | |
|---|---|---|---|
| Time to first code | Long: data, rules, test set first | Days | A prototype in days, then phased go-live |
| Who keeps it current | Your team, every revision | The vendor | The vendor, inside your process |
| Fit to your workflow | Exact | The vendor's workflow | Your workflow |
| Legal line | Yours to manage | The vendor's terms plus your review | Your review, built into the flow |
| Best for | Companies whose product is classification | Teams with a standard workflow | Manufacturers and groups with their own process |
Quotable insight: Building an HTS classifier is not a model problem; it is a maintenance problem that never ends. The schedule changed 32 times in 2025, CBP adds rulings every week, and HS 2028 is coming. Whoever owns the classifier owns that stream of change, so the build-or-buy question is really "who do we want keeping up with the tariff every week?"
The cost side of buying is worked out in the ROI of automated HTS classification, and APIs are compared in the best HTS classification APIs. If the real question is who on your team should own classification, see customs broker vs in-house classification.
What does GingerControl offer instead of a build?
GingerControl offers three buy options, matched to who is buying:
- For enterprise systems: the Classification + Tariff Calculation API, which returns a 10-digit HTS code with its U.S. duty layers in one call, 120 items per call by default; it starts with a call.
- For brokers and small businesses: the GingerLite app, with the Classification Researcher, at $150 a month for 300 credits.
- For manufacturers and multi-entity groups: GingerOS, which runs on your process and the data you have, ERP or not. Without an ERP, we set up the data storage and propose a data flow for you; messy data gets cleaned and organized first. It starts with the Prototype Week: one call, then a clickable prototype of your own flow in three days, and a quote after; how GingerOS connects covers the technical side.
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, does not file entries, and does not give legal advice.
Putting the build-or-buy decision in front of your team
Before anyone writes code, list who will keep the classifier current, who reviews its codes and how its answers reach your ERP; if those answers point to your own process, start with a GingerOS call. See how GingerOS fits your process
References
- U.S. Customs and Border Protection, rulings HQ H290535 (September 29, 2022) and HQ H350722 (January 16, 2026)
Data cited: classifying goods for others is customs business; six digits is not; an AI tool deriving subheadings beyond six digits for goods that will be entered conducts customs business
Source: CBP CROSS, HQ H290535 and HQ H350722 - U.S. International Trade Commission, HTS REST API and release list (read October 11, 2026)
Data cited: free public API at hts.usitc.gov/reststop; 32 revisions to the 2025 HTS; Revision 21 of the 2026 HTS effective October 9, 2026
Source: USITC HTS release list - World Customs Organization, HS 2028 news (January 21, 2026), Explanatory Notes page and WCO Trade Tools pricing (read October 11, 2026)
Data cited: HS 2028 in force January 1, 2028, 299 sets of changes; Explanatory Notes are "the official interpretation of the Harmonized System at the international level"; Harmonized System subscription 450 euros per single user per year
Source: WCO HS 2028 amendments, WCO Explanatory Notes and WCO Trade Tools pricing - U.S. Customs and Border Protection, CROSS statistics feed (read October 11, 2026)
Data cited: 222,008 searchable rulings
Source: CBP CROSS - Yuvraj and Devarakonda, "ATLAS: Benchmarking and Adapting LLMs for Global Trade via Harmonized Tariff Code Classification", arXiv 2509.18400 (September 22, 2025)
Data cited: dataset of 18,731 CROSS rulings (200 test); 10-digit and 6-digit exact match: Atlas 40.0% and 57.5%, GPT-5-Thinking 25.0% and 55.5%, Gemini-2.5-Pro-Thinking 13.5% and 31.0%, DeepSeek-R1 2.5% and 26.5%; authors at Flexify.AI; no commercial vendor tested
Source: arXiv 2509.18400 - eCFR and U.S. Code: 19 CFR 163.4(a), 19 U.S.C. 1484(a)(1), 19 U.S.C. 1592(c)(3) (read October 11, 2026)
Data cited: records kept "5 years from the date of entry"; the importer of record files "the declared value, classification and rate of duty" using reasonable care; negligence penalty up to the lesser of the domestic value or "two times the lawful duties, taxes, and fees" lost
Source: 19 CFR 163.4, eCFR, 19 U.S.C. 1484, U.S. Code and 19 U.S.C. 1592, U.S. Code
Questions people ask
Should we build our own HTS classification model with machine learning?
Only if classification is your product. A useful classifier must apply the GRI and the Section and Chapter Notes, use CBP rulings, keep up with a schedule that changed 32 times in 2025, and leave a reason a broker can review. For a manufacturer with 20,000 SKUs, buying a tool or API, or having a system fitted to your process, is usually faster and cheaper to keep right.
How accurate are general AI models at HTS classification?
Evaluate the model on your own products before relying on it. In the ATLAS study published September 22, 2025 on 200 CBP rulings, GPT-5-Thinking matched 25% of codes at 10 digits and 55.5% at 6 digits, and a model fine-tuned on more than 18,000 rulings reached 40% and 57.5%. For a platform team planning to wrap a general model, budget for the legal reasoning and review, not just the model. This is a historical result for that test set, not a current accuracy ceiling.
What data do you need to build an HTS classifier?
The current HTS with every revision (the USITC's free API provides it), Section and Chapter Notes, CBP's CROSS rulings (more than 222,000 searchable in October 2026), ideally the WCO Explanatory Notes (paid), and a test set of your own products with broker-confirmed codes. For a team classifying 5,000 SKUs, the test set is the part most often skipped.
How often does a homemade HTS classifier need updating?
Continuously. By our count of the USITC's release list, the 2025 HTS had 32 revisions and the 2026 HTS had 21 by October 9, 2026, and HS 2028 brings 299 sets of changes on January 1, 2028. For an in-house tool covering 10,000 codes, each revision means checking which codes moved and retesting.
Our product data is messy. Should we build a system to clean it first?
Not necessarily. Cleaning the data is the first step of a fitted system, not a separate build. GingerOS runs on your process and the data you have, ERP or not: GingerControl cleans and organizes messy data first, then builds the AI system, and it reads emails, PDFs, Excel files and images. For a group with 15,000 parts, start with a Prototype Week.
Is it cheaper to build or buy HTS classification?
Buying is usually cheaper unless classification is your product, because the cost of building is not the model but keeping it right: schedule revisions, rulings, notes, testing and review. For a platform with 50,000 calls a month, price a build as a permanent team, then compare it with API quotes and public prices such as SimplyDuty's 0.10 pounds per API call, read October 11, 2026.
Can we fine-tune an open model on CBP rulings to classify HTS codes?
You can, and the ATLAS study published September 22, 2025 did exactly that: a LLaMA-3.3-70B model fine-tuned on CROSS rulings matched 40% of 10-digit codes and 57.5% of 6-digit codes on its test set. For a team aiming at entry-ready codes, that still needs the GRI, the notes, a question step for missing facts and broker review on top. This is a historical result for that test set, not a current accuracy ceiling.
Who is liable if our homemade HTS classifier gets a code wrong?
The importer of record. It files the classification using reasonable care under 19 U.S.C. 1484, and a revenue-affecting negligent violation is capped at the lesser of domestic value or twice the lawful duties, taxes and fees lost or potentially lost under 19 U.S.C. 1592, whatever tool produced the code. For an importer paying $2 million a year in duty, the classifier's record of how each code was reached matters as much as its accuracy.
Classifying SKUs one ruling at a time?
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




