HS classification was supposed to be the easy item on the roadmap. The platform already had the trade data and the customers asking for it. How hard could an engine be?
Harder than it looked. The in-house build keeps running into the same wall: tricky products, interpretation rules, rulings that shift the answer, and accuracy that never gets to the level its customers expect. Meanwhile, the customer requests keep coming.
Then leadership makes the call. Stop building it in-house. Find someone who already has.
Finding a vendor was easy. Finding one accurate enough for their customers was not.
Why AI tariff classification matters to trade-tech platforms
Two platforms, two very different customer bases, and the same gap. The first is a global trade data provider with 1,000+ customers. It already sells HS classification reference data and is moving into compliance workflow software, where it wants to own content and software rather than sell services.
The second is a customs filing platform with ~4,000 users in 14 countries, most of them customs brokers, and it is expanding into new markets. It does not offer HS classification today, even though its brokers need exactly that.
| Global trade data provider | Customs filing platform | |
|---|---|---|
| Reach | 1,000+ customers worldwide | ~4,000 users, 14 countries |
| End users | Compliance classifier teams of 10–20 | Customs brokers |
| Classification today | Reference data, no classification engine | None |
| How it would ship | Classification section in its own dashboard | REST API inside its own product |
| Rulings that matter | US (CBP) and EU (EBTI) | Local rulings in each country it serves |
For the data provider, classification raises contract value and helps it compete with the large trade management suites. For the customs filing platform, it fills the most obvious gap in a broker’s day.
- 1 Platform UI the customer’s own branding
- 2 TariffWolf API SKU, description, import country
- 3 HS code + justification with the cited rulings
- 4 Classifier or broker approves the final decision stays with a person
Both needed the same capability. Building it themselves was proving harder than anyone expected.
The hidden cost of manual HS code lookup outside the platform
A classification gap in the middle of the workflow
The customs filing platform’s brokers classify elsewhere and bring the code back. The data provider sells classification content, but its customers’ classifiers still turn it into a decision by hand, item by item, often chasing spec data held by engineering.
Product classification accuracy customers can defend
A code without a reason gets overruled. The data provider’s end customers already have codes they trust, and they challenge any answer that doesn’t hold up. Items that turn on interpretation rules or essential character make it harder still.
Catalogs too large for one-by-one classification
A careful, question-by-question flow works for a team of 15–20 people classifying 50–100 items each per week. It does not work for a customer onboarding 50,000–70,000 SKUs, or the 500,000-item catalogs some customers want fully automated. By TariffWolf’s estimate, a large catalog is about three months of manual work.
Our end customers are extremely particular about accuracy. They already have codes they trust, and anything we give them has to hold up.
Cost of inaction: Before embedding classification, neither platform could return an HS code inside its own product. For a large customer catalog, that meant roughly three months of manual classification (TariffWolf estimate).
The breaking point
For both platforms, the trigger was growth. The data provider was moving from selling data into compliance software, where classification is the feature that lifts contract value. The customs filing platform was expanding into new countries, and its brokers kept asking for the one tool it didn’t have.
Building it in-house meant tariff schedules, legal notes, interpretation rules and rulings for every country, kept current indefinitely, and the accuracy still fell short of what end customers expect. Leadership asked the teams to look outside. Both landed on the same answer: embed an engine through an API and keep the experience in their own product.
Then a spreadsheet formula returned an HS code, and a paragraph explaining it.
Why both platforms chose TariffWolf for automated tariff classification
The data provider ranked its criteria in order: accuracy, latency, scale (as many as 10,000 classifications an hour), and only then price. The customs filing platform wanted something simple to wire in: a REST API that takes a SKU, a description and the import country, and returns a 6-digit HS code in one call.
For the data provider, the “aha” was a Google Sheet: one formula called the API and returned a code with a written justification, so the team tested its own SKUs with no integration work. For the customs filing platform, it was an industrial compressor cooler skid narrowed step by step to 8414.80, with the justification and matching CBP rulings alongside. The country manager recognized those rulings as the equivalent of the classification rulings published by the customs authority in their own country.
Who decides: TariffWolf proposes a code and shows its reasoning. The platform’s end user, a compliance classifier or a customs broker, reviews it, selects the final national digits and approves it. For US codes, a person always picks the last digits of the 10-digit number.
| Criteria | Manual classification outside the platform | TariffWolf embedded via API |
|---|---|---|
| Large catalog | ~3 months (TariffWolf estimate) | 3–5 days with automated bulk passes (projected) |
| Explanation of the code | In the classifier’s notes, if anywhere | Written justification citing legal notes and rulings |
| Consistency | Depends on who classifies | Same input, same code |
| Thin product data | A best guess | A follow-up question, or no code at all |
Our customs brokers need exactly this kind of tool.
How the workflow changed
Trade compliance automation, proven one test sheet at a time
- Call 1: Sample dataThe data provider shares 15 sample items and tests through the Google Sheets interface before any API work.
- Call 2: First version testedAbout 80 classifications from three prospects. The first version reaches 82% accuracy, and some answers are not yet consistent from run to run.
- Final versionThe engine is rebuilt to rely far less on AI: a fixed sequence of questions, checked against legal notes and then CBP and EU rulings, returning up to three ranked codes.
- Call 3: First measurable winThe final version reaches ~97% on the data provider’s benchmark set and is demoed live.
Effort from the platforms: the evaluation ran in a Google Sheet, with no integration work.
Integrations: REST API with a JSON payload; a classification section is planned inside the data provider’s dashboard.
What was harder than expected: The first version leaned on AI for the classification itself and stopped at 82%. In the final version, a language model only reads the input and writes the justification; the classification follows a fixed sequence of questions.
So what did the numbers look like in the final version?
~97% product classification accuracy on a partner benchmark
These are evaluation results, not production figures. They come from TariffWolf testing on the data provider’s own benchmark set.
| Metric | Before | After | Change | Timeframe |
|---|---|---|---|---|
| Classification accuracy | 82% (first version) | ~97% (final version) | +15 points | Calls 2 → 3 |
| Consistency | Answers could vary run to run | Same input, same code | By design | First → final version |
| Time per item | Varies | 20–45 s (+~30 s per question) | Measured | Call 2 |
| Large-catalog effort | ~3 months | 3–5 days | Projected | TariffWolf estimate |
The unexpected win: The data provider’s benchmark set proved so useful that it now benchmarks every TariffWolf release.
Classification inside our own compliance software lets us grow contract value with customers we already have, and compete with the big suites.
The outcome
- Accuracy: from 82% in the first version to ~97% in the final version, on a design partner’s benchmark set.
- Capacity (projected): a ~3-month catalog effort reduced to 3–5 days with automated bulk passes.
- Consistency: same input, same code, with a written rationale and cited rulings for every item, and no code when the data is too thin.
- Pipeline: ~5 of the data provider’s end customers ready for beta.
TariffWolf turned HS classification from a manual task done outside the platform into a structured, repeatable capability inside it, with classifiers and brokers still making the call.
What’s next: scaling to tens of thousands of classifications
Next is volume. As end customers bring full catalogs, the data provider plans to run tens of thousands of classifications through the API in bulk mode, against the bar it set on day one: as many as 10,000 classifications an hour. High-confidence codes go to review, follow-up questions arrive with suggested answers, and items without enough data go back for more. The customs filing platform is taking the integration to its product owner as it expands into new markets.
3 tips for embedding AI tariff classification in your product
- Set the accuracy bar before your customers see anything, and test it on their real, messy data.
- Ask for the reason, not just the code. A rationale with cited rulings is what lets your customer accept or challenge the answer.
- Plan for bulk, but send uncertain items to people. No code beats a wrong code.