AI copilot for US customs classification
Find the line in the tariff book. Bring the ruling that proves it.
TariffPilot reads a plain product description and walks the Harmonized Tariff Schedule to a 10-digit code, with calibrated confidence and the CROSS rulings behind it. Clear cases clear in seconds; hard cases reach a licensed broker with the evidence already assembled.
Pre-launch · in development · no card, just an email“Stainless-steel insulated water bottle, 750 ml, double-wall vacuum, screw cap, for retail sale”
- 7323Table, kitchen or other household articles, of iron or steel
- 7323.93Of stainless steel
-
7323.93.0080 Other kitchen ware Match · 0.92
cites CROSS N305649 · N292417 · H301619
-
9617.00.1000 Vacuum vessels, capacity ≤ 1 liter Broker review
conflicting heading · both ruling lines attached · 0.41
Block 01 · The problem
Classification is slow, expert-scarce, and high-stakes
Every product imported into the US needs a 10-digit HTS code. The wrong one means overpaid duties or penalties, and the expertise to get it right is stretched thin.
01
Precedent is buried
Roughly 200,000 historical CROSS rulings encode how CBP actually classifies products, but for a working broker they are practically unsearchable case by case.
02
Tariffs keep moving
New trade actions and annual HTS revisions constantly invalidate prior classifications. A code that was right last quarter can be wrong, and expensive, today.
03
Full automation fails
Published benchmarks show even fine-tuned 70B models reach only ~40% accuracy at the 10-digit level. A tool that answers everything confidently is a liability, not a product.
Block 02 · How it works
Triage, not oracle
TariffPilot auto-classifies the high-confidence majority and hands licensed brokers the hard cases, ranked candidates and precedent already attached.
Describe the product
Paste a plain-English product description, one item or a whole line sheet. Materials, use, and packaging are all the engine needs to start.
> "women's knit cardigan, 60% cotton 40% acrylic"Ranked codes, cited rulings
A retrieval engine over the full CROSS corpus and current HTS schedule returns candidate codes, each with a calibrated confidence score and the specific rulings that support it.
6110.20.2079 · 0.88 · cites N312846Brokers take the hard cases
Low-confidence items enter a review queue for a licensed broker, evidence pre-assembled. Every correction feeds back to make the engine sharper.
2 of 14 held · precedent attachedBlock 03 · What's inside
Built for people who answer to CBP
Decision support for customs brokers, e-commerce importers, and trade-compliance teams. The licensed professional stays in charge; TariffPilot does the digging.
Calibrated confidence
Every candidate code carries a probability that means what it says. Confidence thresholds decide what auto-clears and what a human sees, and you set the thresholds.
Cited CROSS precedents
No black-box answers. Each suggestion links the actual CBP rulings that support it, so a broker can verify the reasoning in minutes instead of hunting for it in hours.
Tariff-change alerts
Watch-lists that flag when a new ruling, trade action, or HTS revision touches a code in your catalog, before it shows up as a duty bill or a penalty.
Bulk import roadmap
CSV and ERP import to classify an entire catalog in one pass, with the same triage split: the routine bulk cleared, the exceptions queued for review.
Block 04 · Status
In development, honestly
TariffPilot is pre-launch, so this page makes no claims about users, logos, or revenue. What exists today is the plan: corpus ingestion and a retrieval baseline first, then calibrated triage with a cited-answer interface, then a small pilot with working brokers.
●CROSS corpus + HTS schedule ingestion
●Calibrated triage & cited-answer UI
●Broker pilot & correction loop
what you import is plenty
Block 05 · Questions
Fair questions
Is this legal or customs advice?
No. TariffPilot is decision support. It surfaces candidate codes, confidence, and the precedent behind them; classification decisions and filings remain with you and your licensed customs broker. That division of labor is deliberate and built into the product.
Why not just automate classification fully?
Because the evidence says it doesn't work. On public benchmarks, even fine-tuned large models get roughly 40% of 10-digit codes right. Triage is the honest architecture: automate the cases the model is demonstrably good at, and give experts a head start on the rest.
Who is TariffPilot for?
Small and mid-size customs brokerages first, then e-commerce importers and in-house trade-compliance teams. If you classify products against the US HTS on a regular basis, it is built for your desk.
When can I use it?
It is in active development; the waitlist is open now. Early-access invitations will go out in small batches to people who classify for a living, so telling us what you import genuinely moves you up the list.