Unis, LLC Customer story
Clear a post de minimis backlog with one API call.
Two vendors made Unis choose between a fast wrong answer and a slow right one. One call that returns the code and the full tariff stack ended the trade-off.
- Daily classifications1.5K12K+Items cleared per day
- Latency per item (p50)~60s30sProduction median, one endpoint
- Re-classification rework22%4%Flipped by internal QA per batch
- Coverage per callHTS + full tariffMFN, 301, 232, 122 and split codes
Both prior classification vendors were retired at the end of week eight.
Results by week eight. Daily classifications: 1.5K to 12K+ (Items cleared per day); Latency per item (p50): ~60s to 30s (Production median, one endpoint); Re-classification rework: 22% to 4% (Flipped by internal QA per batch); Coverage per call: HTS + full tariff (MFN, 301, 232, 122 and split codes).
What happened at Unis?
A four-week integration replaced two failing vendors with one call that returns the code and the duty.
Unis, LLC is a nationwide US freight forwarder and 3PL with 50+ facilities. When de minimis was withdrawn for the goods its e-commerce shippers move, every parcel needed a real HTS code and duty, and catalogs of 5,000 to 50,000 items landed on the classification team. Two vendors forced a choice between fast and wrong or slow and incomplete. GingerControl integrated its Classification + Tariff Calculation API in four weeks: one call returns the HTS code, the full tariff stack and split-code components. By week six Unis cleared 12,000+ classifications a day.
Unis, LLCPublished June 3, 2026- Engagement
- API integration + custom extension
- Industry
- Freight forwarding and 3PL, US
- Footprint
- 50+ facilities, nationwide
- Build window
- 4 weeks
- Serves via Unis
- Royal Mail (UK) and Chunghwa Post (Taiwan)
Case file for Unis, LLC: Engagement API integration + custom extension; Industry Freight forwarding and 3PL, US; Footprint 50+ facilities, nationwide; Build window 4 weeks; Serves via Unis Royal Mail (UK) and Chunghwa Post (Taiwan).
Why couldn’t two vendors keep up?
One was fast but overridden about half the time. The other took about a minute per item and returned only the code.
- After de minimis ended, new customers onboarded catalogs of 5,000 to 50,000 items at a time.
- Every classification needed a tariff calculation, and every calculation had to hold up if CBP asked.
- Unis split traffic between the two vendors by shipment risk. Neither worked.
“The choice every vendor was offering us was wrong or slow. Neither of those is a choice we can make at the volume we are running.”
New catalogs arriving at 5,000 to 50,000 items each
| Vendor | Speed | Returns | Problem |
|---|---|---|---|
| Fast classifier | Under a second | HTS code | Overridden about half the time |
| Accurate classifier | About a minute per item | HTS code | Tariffs and Chapter 99 by hand |
Table of the two classification vendors Unis used before: a fast classifier answering in under a second but overridden about half the time, and an accurate classifier taking about a minute per item, both returning only an HTS code.
Was the trade-off unavoidable?
No. The time went to rework and to tariff logic the classification-only APIs left to the caller.
- The fast vendor guessed on mixed materials, kits and sets: about 47% of its outputs were wrong or below the confidence Unis could ship.
- The accurate vendor’s 60 to 90 seconds per item meant four to seven hours for a 5,000-item catalog, before any tariff work.
- Neither handled split-code products like watches, so about 1 in 15 items went to a manual queue.
“We did not need a faster wrong answer. We needed the full answer, in one call, in a time we could budget.”
Measured during the Unis vendor evaluation. Lower latency is better, higher accuracy is better.
- Full tariff stack
- HTS code only
Scatter chart of accuracy on the Unis QA sample against latency per item: fast classifier about 53% in under a second, accurate classifier about 85% at 60 seconds, senior broker by hand about 88% at 4 to 7 minutes, and the Classification + Tariff Calculation API 96% at 30 seconds with the full tariff stack.
What does one call return?
The HTS code, the general and special rates and every Chapter 99 entry, in the same response.
- Two required fields in, description and country of origin, so every Unis feed could use it: customer feeds, emailed Excel, SFTP drops and EDI.
- Split-code products such as Chapter 91 watches return a components array, each with its own code and tariffs. The composite math is done on our side.
- Batch calls return results in the caller’s item_id order, with a total, succeeded and failed summary.
- The custom extension, steel and aluminum pour countries for Section 232 and an optional reasoning trace, sits on the same endpoint and is in the public OpenAPI spec.
- Ten production days in shadow mode against the senior team’s codes on the hardest categories before going live.
POST/openapi/v1/tariff
{
"description": "Cotton knit short sleeve T-shirt",
"country_of_origin": "DE",
"extra": {
"steel_pour_country": "IT",
"aluminum_pour_country": "FR"
}
}{
"hts_code": "6109.10.0012",
"tariffs": {
"general_rate": "16.5%",
"special_rate": "Free",
"Section 301": [],
"Section 232 - Metals": [],
"Section 122": [
{
"code": "9903.03.01",
"rate": "10%"
}
]
}
}{
"hts_code": "9101.11.40",
"tariffs": null,
"components": [
{
"hts_code": "…",
"tariffs": { … }
}
]
}Example from the API manual, not Unis data
The Classification + Tariff Calculation API call POST /openapi/v1/tariff, using the example from the API manual: a request with a description, country of origin and steel and aluminum pour countries, a response with the HTS code, general and special rates and Chapter 99 entries for Sections 301, 232 and 122, and a split-code product returning a components array.
What changed by week eight?
12,000+ classifications a day, rework down from 22% to 4%, and both prior vendors retired.
Four weeks before to eight weeks after go-live. Before, both vendors together topped out near 1,500 a day.
Line chart of daily classifications from 4 weeks before to 8 weeks after go-live: about 1,500 a day before, rising to about 12,000 a day by week eight.
Before, and 8 weeks after the API replaced both prior vendors.
- Manual classification and rework
- Tariff stacking by hand
- Document and customer chasing
- Brokerage and exception work
Bar chart of the classification team’s time. Before: manual classification and rework 65%, tariff stacking by hand 18%, document and customer chasing 12%, brokerage and exception work 5%. After: 8%, 5%, 15% and 72%.
- The tariff side needed no build: Unis’s ABI filing pipeline and customer duty estimator read the response as it comes.
- Watches for one e-commerce client come back as case, movement, strap and battery, each with its own code and rate, in one call.
“We have seen a lot of classification APIs. This is the first one we would describe as actually finished.”
| Metric | Before | After | Change |
|---|---|---|---|
| Classifications cleared per day | 1,500 | 12,000+ | 8x throughput |
| Median latency per item | ~60s | 30s | 50% faster |
| Re-classification rework | 22% | 4% | 82% reduction |
| Tariff coverage per call | HTS only | HTS + full tariff | 1 call vs 3+ |
| Split-code products | Manual, ~1 in 15 items | Composite duty in the response | Manual queue gone |
Before and after: Classifications cleared per day, 1,500 before, 12,000+ after (8x throughput); Median latency per item, ~60s before, 30s after (50% faster); Re-classification rework, 22% before, 4% after (82% reduction); Tariff coverage per call, HTS only before, HTS + full tariff after (1 call vs 3+); Split-code products, Manual, ~1 in 15 items before, Composite duty in the response after (Manual queue gone).
Why was this more than a vendor swap?
Three design calls: code and tariff in one call, split codes as part of the contract, and room for custom fields from day one.
- Classification and tariff are one callEvery other vendor on the shortlist stopped at the HTS code. Folding the tariff in saved Unis a whole integration layer.Decided
- Split codes are part of the contractChapter 91 and similar products don’t classify as single units. A components array with per-part tariffs fit the real product mix.Decided
- Custom fields were designed inPour countries and the reasoning trace went on the same endpoint, documented in the public spec, not on a customer-only branch.Decided
Decision log: Classification and tariff are one call; Split codes are part of the contract; Custom fields were designed in.