Tariff Classification Intelligence

A 10-digit HTS code, with the legal argument attached.
Not a lookup. A ruling.

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.

6
General Rules of Interpretation, applied in strict order
1
confidence score — never inflated to clear an auto-approval bar
0
fabricated CROSS ruling numbers — unverifiable citations are dropped

The problem

"It gives me a code. It doesn't give me a defense."

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

A legal argument, a verified citation, a memory of your own decisions.

  • GRI 1 through 6 applied in strict order — the model must state which rule actually drove the call, not cite GRI 6 out of habit.
  • Every CBP CROSS ruling cited is looked up and verified against the real rulings database before it's shown — an unverifiable citation is dropped, not kept.
  • Prior broker-approved classifications for the same account are retrieved and cited as evidence, tagged by how much authority they carry.
  • Confidence gates auto-approval — a vague or low-confidence call routes to Needs Review, it doesn't clear the bar on hope.
  • "Approved" stays a human word: the agent's own output is AUTO_VERIFIED, never APPROVED, until a broker signs off.

GRI legal reasoning

The code comes with the rule that produced it.

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.

  • 10-digit HTSUS code, description, and general (non-preferential) duty rate from the current HTSUS revision
  • GRI citations recorded as which rule(s) actually drove the decision — not a boilerplate list
  • A full legal rationale, written the way a broker would defend the call to CBP
  • Confidence reflects genuine classification certainty — an ambiguous product description caps confidence, it doesn't get rounded up
How to demo
Classify a composite product with two plausible headings — show GRI 3(b) essential-character reasoning in the rationale, not a coin flip between headings.
Classification/SHP-2026-000144 · line 2
SR
Sarah
HTS code8537.10.9170
Duty rate2.7% (general)
GRI appliedGRI 1 · GRI 3(b)
Confidence61%
Legal rationale
Board incorporates a control PCB and a battery pack in one housing; no single heading covers the whole. Essential character (GRI 3(b)) is the control function — classified as electrical control apparatus of heading 8537, not as a battery of 8507.

Verified CBP CROSS citations

A cited ruling is a ruling that actually exists.

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.

  • No unverified ruling number ever reaches the reviewer as a citation
  • A fabricated or misremembered ruling is worse than no citation — so the system removes it rather than keep it
  • Verification runs per line item, alongside the classification call itself
  • On a classification case, a reviewer can search CROSS and attach a ruling as evidence (supports, challenges or context only); a revoked ruling blocks a case it was supporting, and the account chooses whether a modified ruling warns or blocks
How to demo
Classify a product with a well-known prior ruling — click through the cited CROSS number to the actual ruling text. "It didn't just say a number, it checked the number."
Classification/Rulings
SR
Sarah
NY N309562verified · resolves
HQ H301987verified · resolves
Rejected during verification
NY N412209no match → dropped, not shown
CBP CROSS ingestion service · lookup-on-cite

Account precedent memory

It remembers what your brokers already decided.

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.

  • Human broker approvals and confirmed filing outcomes carry higher authority than a bare model inference
  • Cited precedent is the real record — content, source type, and confidence — not just a count
  • Memory retrieval degrades gracefully: if it fails, classification still proceeds on GRI reasoning alone, it never blocks a line item
How to demo
Classify a part your team has approved before — open the case in Comply › Classification and show "Informed by N prior account decisions" with the actual prior determination quoted.
Classification/Case Review
SR
Sarah
HS Code8537.10.9170
ProductBMS control board
Confidence92%
Informed by 2 prior account decisions
Broker approved: Same board family classified 8537.10.9170 on PO-88213, June 2026.
Filed: Entry 305-4471xxx-x liquidated at 8537.10.9170, no rate advance.

Confidence-gated auto-approval

The bar clears on genuine confidence, or it doesn't clear.

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.

  • Low-confidence fallback candidates are labelled unverified and forced to review — they never masquerade as a normal classification
  • Every decision keeps its full evidence bundle: GRI citations, verified rulings, account precedent, reasoning chain
  • Per-account policy id recorded on the decision, so a policy change is itself auditable
How to demo
Classify a well-known part (auto-verified) next to an ambiguous one (needs review) — same pipeline, different outcome, driven only by the confidence number.
Classification/Review Queue
SR
Sarah
Line 1 · connector housing8536.69.405696%auto-verified
Line 2 · BMS control board8537.10.917061%needs review
Line 3 · wiring harness8544.30.0000unverifiedlow confidence
Line 3 — fallback candidate
DB candidate used as unverified low-confidence suggestion. Requires human broker classification review before filing.

Guided classification

Not sure of the code? Walk the tariff, one level at a time.

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.

  • The classification path stays visible, and any earlier step can be changed
  • Only options actually offered can be picked, and a final code is always a stored tariff line, shown with its General rate
  • A research aid: it creates no case, records no decision and approves nothing; it sits alongside the automated classifier, not in place of it
How to demo
Open HTS Lookup, click "Not sure of the code? Describe your product", and describe an ambiguous part. Show a confirm-or-pick step, then change an earlier step in the path.
HTS Lookup/Guided Classification
SR
Sarah
Classification path
Chapter85
Heading8537
Subheadingnext step
Suggested · illustrative
Use the suggestion, or choose another option from the list.

Objection handling

Every hard question, answered directly.

"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

"Ask it for the code. It gives you the code, the rule, the ruling, and the precedent — and tells you when it isn't sure."

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