Tariff Classification Intelligence
Qubere classifies every line item by working GRI 1 through 6 in order, citing the actual CBP CROSS rulings behind the call, checking it against your account's own prior broker decisions, and routing anything short of genuinely confident straight to a human — before it ever reaches a filing.
The problem
What a classification lead says about AI tariff tools:
"A model can guess an HTS code. Can it show its work under GRI 1 through 6, the way a broker has to when CBP asks?"
Legal reasoning
"Half the 'ruling citations' these tools give me don't exist when I look them up. That's worse than no citation."
Hallucinated rulings
"We've classified this exact part before. Why does the tool re-derive it from scratch every single time?"
No institutional memory
How Qubere addresses it
GRI legal reasoning
Every classification works the General Rules of Interpretation in order — heading and section/chapter notes first (GRI 1), only moving to GRI 2–6 when GRI 1 doesn't resolve it cleanly. For composite goods, essential character (GRI 3(b)) is captured as its own input, not buried in prose.
Verified CBP CROSS citations
The model is instructed to cite a CROSS ruling only when genuinely confident it applies — and every citation it returns is looked up against the real CBP CROSS rulings database before being shown. A ruling number that doesn't resolve is dropped, not surfaced with false confidence.
Account precedent memory
Before classifying, the agent retrieves your account's own prior human-approved decisions and filing outcomes for related products — hybrid vector + full-text search, scoped strictly to your account. Matched precedent is shown to the reviewer as evidence, tagged by its source, not folded silently into a number.
Confidence-gated auto-approval
Each account sets its own auto-approval policy by confidence threshold. A line item that clears it becomes AUTO_VERIFIED and can flow straight to filing; anything short goes to the Classification review queue for a broker. "Approved" stays a human decision either way — the agent's own state is never named that.
| Line 1 · connector housing | 8536.69.4056 | 96% | auto-verified |
| Line 2 · BMS control board | 8537.10.9170 | 61% | needs review |
| Line 3 · wiring harness | 8544.30.0000 | unverified | low confidence |
Guided classification
Describe the product and Qubere steps through chapter, heading, subheading, tariff line and statistical suffix. At each level the model sees the full list of sibling options, plus the section and chapter notes at the heading step, and recommends one with a confidence. High confidence continues on its own; otherwise you confirm or pick another option.
Objection handling
| "How is this different from an LLM that just guesses an HTS code?" | GRI 1–6 applied in order, with the driving rule stated explicitly; CROSS ruling citations verified against the real database, not asserted; confidence that gates auto-approval rather than being decorative. |
| "What stops the model from fabricating a ruling number?" | Every cited CROSS ruling is looked up before it's shown. A number that doesn't resolve is dropped from the citation list, never surfaced with confidence attached. |
| "Does it learn from our own past decisions, or start from zero every time?" | Account-scoped precedent memory retrieves prior human-approved classifications and filing outcomes for related products and shows them as cited evidence — tagged by authority, never silently averaged in. |
| "Can the AI approve its own classification?" | No. The agent's own state is AUTO_VERIFIED, distinct from human APPROVED, and only clears to auto-flow at or above the account's configured confidence threshold. |
The close
A classification you could defend to CBP yourself, because the reasoning is already written down.
Next step: bring a part your team has classified before. We'll run it through, show the GRI reasoning, click into the verified ruling, and show your own prior decision cited back to you.
Full detail: docs/apps/customs/sales/ai-and-document-intelligence.md