Text, image or URL
The input can be a product description, a photograph of the item or its packaging, or a product page URL — whichever your system already holds.
Customs & Compliance
An HS code is only useful if you know how much to trust it.
Send product text, an image or a URL and get back a code with a confidence score. High-confidence items file straight through; everything else is routed to a person, which is what makes automating classification safe at volume.
What it returns
A code, and a number that says how much to trust it.
An API that returns only a code forces a choice between trusting all of it and checking all of it. The confidence score is what makes the middle possible.
A code and a confidence score
The classification and the model's own confidence in it, returned together so the caller can decide what to do next.
Routing on the score
A threshold you set: low-confidence items go to a reviewer, the rest file straight through. The threshold is a business decision, not a fixed property of the model.
Fits the clearance workflow
The code lands on the same shipment record as the description checks and the duty calculation, so nothing is re-keyed between systems.
What to expect from it
Be suspicious of any accuracy number, including ours.
Published benchmarks are modest
The strongest published agent on the HSCodeComp dataset reached 46.8% exact match at ten digits against 95.0% for human experts — around 82% agreement at chapter level. Anyone claiming near-perfect ten-digit accuracy is measuring something else.
Every benchmark measures the wrong thing
These figures measure agreement with expert labels, not acceptance by a customs authority. No published benchmark reports acceptance rates, so a high score is not evidence your declarations will clear.
Depth changes the answer
Accuracy at chapter level and accuracy at ten digits differ by a wide margin. Decide which one your process actually needs before choosing a threshold.
Test it on your own catalogue
Benchmarks are drawn from other people's products. The only number that matters is the one measured on your own declarations — the methodology is set out here.
Test the classifier on your own product data.
Send us a sample of your catalogue and we will run it, show you the confidence distribution and the error pattern, and tell you what share of your volume could file unattended at a threshold you choose.
Book a free automation review