AI & Document Intelligence

Every field traces to a page and a box.
The AI can't launder a decision.

Qubere turns an uploaded PDF into filing-ready facts where every value points back to a page, a bounding box, and the original document (SHA-256-verified on every read) — and puts an assistant in front of the platform that answers only with real, linked records and runs every action through the same permission and audit checks as the UI.

4
confidence signals, never merged into one
130+
assistant tools · RBAC-gated before the model sees them
1
AI surface to audit — the old copilot was deleted

The problem

"It sounds confident and it's guessing."

What a compliance leader says about AI trade tools:

"Extraction tools give me a JSON blob and no way to check it. When it's wrong on a value, I find out at liquidation."

No verification

"Every AI chatbot bolted onto a compliance tool is a liability. If the model classifies a product, who's accountable?"

Accountability

"I want to ask 'which shipments are at risk, what's the dollar exposure' — and I need the number to be right."

Trust the answer

How Qubere addresses it

Evidence, not prose. Same permissions, same audit log.

  • Every extracted fact carries its page number and bounding box, chained all the way back to the SHA-256-verified original.
  • Parser confidence, OCR confidence, model confidence, and validation status are four separate things — "not measured" is never shown as zero.
  • The assistant is a client of the existing app layer — its "value at risk" is the same computation the dashboard runs.
  • Write actions are permission-gated, audit-logged as assistant-initiated, and behind a confirmation step the button UI doesn't even have.
  • Origin safety: the model must quote the legal origin position verbatim — it may not substitute a manufacturer country.

Field-level provenance

Every declared value shows you where on the document it came from.

Customs field → extraction field → parser element → page + bounding box → canonical artifact → original. Absence is preserved — a null page means the parser didn't report one, never a default. A bounding box is never synthesised.

  • Layout-aware parsing (tables, sections, OCR fallback) via Google Document AI
  • Tables and line items carry the same page + bounding box as every other field, each tagged with its reading-order position on the document
  • Line extraction is checked for completeness: dropped invoice line quantities are recovered, line sums are verified against the header totals, and a value that couldn't be read shows as missing rather than 0
  • Logos, signatures, seals, stamps, handwritten notes, QR codes and barcodes are captured as structured, located objects — not just noted in passing
  • Cross-document reconciliation blocks filing on a value conflict; the same fact repeated across pages of one document collapses into a single entry
  • Deterministic checks — duplicate identifiers, date ordering, package-count totals — run alongside the model's own reported validations and force review on a real discrepancy
  • Every flagged field carries a machine-readable reason — low confidence, missing on the source document, or a cross-document conflict — not just a "needs review" flag
  • Extraction, classification, normalization, and compliance-audit agents route through a shared multi-provider model layer — no single vendor is a point of failure for document intelligence
How to demo
On an extracted field, click through to the source — the highlight on the page image. "No other extraction tool in this space shows you where a declared value came from."
Trade Documents/commercial_invoice_88.pdf
SR
Sarah
ValueUSD 142,400.00
Sourcepage 1 · bbox [412, 690, 566, 712]
Element ref#/texts/47
Reconciled againstpacking_list_88.pdf line-sum · match
Confidence — four signals, never averaged
Parser (mean_score)0.97 · grade A
OCRnull · no OCR pass (born-digital)
Extraction model0.93
Deterministic validationPASS
Why this field needs a look
Verification stateNEEDS_REVIEW
Reason codeLOW_CONFIDENCE

Durable processing & quality gate

Upload is instant. Parsing survives a restart.

The upload request validates, stores the immutable original, hashes it, creates one durable run, and returns. A parse becomes the active version only if it passes an objective gate — text coverage, blank-page counts, parser warnings. No invented score.

  • Insufficient text → automatic OCR escalation → then a person
  • Reprocess creates a new version; the old one is kept for audit
  • Client email, when enabled: auto-attach policy tiers Off/Confident/Aggressive, an LLM fallback pass under Aggressive, broker review for the rest
  • Bulk upload is capped at 20 files per request, rejected with a structured 422 rather than a vague failure — protects intake throughput from an unbounded batch instead of just raising a limit
How to demo
Upload a scanned/image PDF — watch it drive the OCR escalation path and land in Needs Review. "It doesn't pretend a blank parse is a good one."
Trade Documents/Processing
SR
Sarah
Processing state machine
QUEUED → SUBMITTED  · profile STANDARD
POLLING → SUCCEEDED  · text coverage 41%
Quality gate: RETRY_WITH_OCR  · <60% coverage
New run v2 · profile OCR_FALLBACK → SUCCEEDED  · coverage 88%
PASS_WITH_WARNINGS → active version v2
immutable original · SHA-256 on every readmalware scan: advisory

Multi-document splitting

One PDF, several documents in it — split only when a human says so.

A 40-page carrier packet is often an invoice, packing list, and BOL stapled together. Header-keyword detection flags this as an advisory suggestion only — it never auto-splits, and every existing field and flow on the original document row is untouched until a person confirms.

  • Page-by-page keyword matching (invoice, packing list, BOL, AWB, certificates of origin, and more) — a missed split is safe, a false-positive split inside one real document is the failure mode this is tuned against
  • A person confirms explicit page ranges, or reuses the stored suggestion, via the split action — nothing splits automatically
  • Split creates new, additive child documents; the parent is never mutated or deleted
  • Each child re-enters the exact same parser pipeline as any other upload — no special-cased extraction logic for split documents
How to demo
Upload a merged carrier packet — show the multi-document suggestion, then confirm the split and watch each child document classify and extract independently.
Trade Documents/carrier_packet_0451.pdf
SR
Sarah
Pages 1–3COMMERCIAL_INVOICE · confidence 2/2 header hits
Pages 4–6PACKING_LIST · confidence 1/1 header hit
Pages 7–42BILL_OF_LADING · confidence 2/2 header hits
Original document unchanged until confirmed — split children re-enter the standard parser pipeline

Extraction accuracy feedback loop

Every human correction becomes ground truth, not a discarded click.

Every field-review verdict — approve, edit, reject, select an alternate — used to vanish the moment the review status updated. Now each one is recorded against the parser provider/version and confidence signals that produced it, so accuracy is measured from real outcomes, not assumed.

  • One recorded event per verdict: confirmed match, corrected, rejected, or alternate selected
  • Recording is best-effort and non-blocking — it can never fail or slow down the actual review action
  • A new report aggregates accuracy by field, document type, and parser provider/version — which fields are wrong most often, and whether the confidence score is actually calibrated to that
How to demo
Correct a low-confidence field in review, then open the Document Extraction Accuracy report and show that correction landing as a real data point tied to the parser version that produced it.
Reports/Document Extraction Accuracy
SR
Sarah
invoiceTotalqubere.parser/2CONFIRMED_MATCH
hsCodequbere.parser/1CORRECTED
countryOfOriginqubere.parser/2REJECTED
Recorded from the same review verdict that already updates FieldApproval/HydrationCandidate — no separate workflow

Qubere Magic Marker

Fix a field by pointing at it, not retyping it.

A wrong extracted value used to mean deleting it and typing the correction by hand off the page image. Now a reviewer clicks Mark on document, drags a box around the real text on the page, and that text fills the field — the correction comes straight from the source, not from memory or a second read of the PDF.

  • One click switches the review panel to the document view with the target field already in focus
  • Drag-select on the PDF canvas reads the text under the box and drops it straight into the field editor
  • Works on any field, on any document already rendered as a PDF — no separate annotation tool or export step
How to demo
Open a field with a wrong value, click Mark on document, drag a box over the correct text on the page — the field fills itself. "No retyping, no copy-paste from a PDF viewer."
Trade Documents/commercial_invoice_88.pdf
SR
Sarah
FieldIncoterm
Current valueCIF (likely wrong)
Region selectedpage 1 · bbox [188, 244, 236, 262]
Text under selectionFOB
Field after fillFOB
Region text fills the editor directly — still requires the normal save/confirm step

Group by HS Code

One row per HS code, computed from the documents — not summed by hand.

Customs wants one quantity, weight, volume, and value per HS code. Operators used to total that off the extracted line items in a spreadsheet. The shipment's line-items table now has a By Line / Group by HS Code toggle that consolidates every line sharing a code, straight from whatever documents were extracted for the shipment.

  • Each group lists the original line numbers it was built from, so you can see which articles share a code
  • Quantity, net/gross weight, volume, and total value are summed per group, with unit and currency
  • Never a silent wrong total: lines with different weight or volume units are not summed, and a value total missing a line's amount is flagged as a partial sum
  • Lines with no HTS code collect into an Unclassified group instead of disappearing
How to demo
Open a shipment with several lines sharing a code, switch to Group by HS Code, and show the Lines column. "That's the one number per HS code you used to build by hand."
Shipments/Line items
SR
Sarah
HS codeLinesQtyNet / GrossVolumeTotal value
8537.10.91701, 3, 712084.0/96.0 KG0.62 CBMUSD 18,600
8544.42.90902, 530042.0/46.0 KG0.30 CBMUSD 9,150
9403.20.00804, 640Mixed units — not summed0.55 CBMUSD 3,200 (partial)
Illustrative data. Amber = a total the system declined to present as complete

Ingest channels

Portal, API, or email — every document lands in the same pipeline.

Documents arrive from people, from systems, and from inboxes. Each channel stores the original and hands it to the same parse, classify, and extract pipeline — there is no separate, weaker path for an email attachment or an API upload.

  • Portal upload: drag-and-drop from the shipment or the documents hub
  • Intake API: an ERP, TMS, or partner posts a document URL (or a batch of up to 25) with an API key; the response returns the document ID and any candidate shipment matches, and nothing is enqueued if any item fails validation
  • Inbound email: attachments are downloaded, scanned, stored, and processed exactly like a manual upload; sender policy is checked first
  • Mail from a sender nobody has registered is quarantined, not attributed to an account until an admin releases it
How to demo
Send an invoice by email, then upload one in the portal — show both landing as documents with identical processing status. "Same pipeline, whichever door it came through."
Trade Documents/Inbound
SR
Sarah
FileChannelStatus
invoice_8841.pdfPortal uploadprocessing
packing_list.pdfIntake APIprocessing
bol_0451.pdfInbound emailprocessing
scan_0012.pdfInbound email · unregistered senderquarantined
Illustrative data. A quarantined attachment is stored and scanned but not attributed to an account until released

Document types & decision score

It knows what kind of document it is — and shows how much to trust each field.

Classification runs against a catalog of about 25 customs document types in six categories, each with keywords and a required-for-filing flag. Every extracted field then gets a single 0–100 decision score built from four independent signals, so a reviewer sees why a value is trusted, not just a bare percentage.

  • Catalog spans commercial (invoice, packing list, PO), customs/CBP (7501, 3461, ISF), partner-agency (FDA, EPA, USDA, FCC), transport (B/L, AWB), post-entry (PSC, drawback), and compliance documents
  • Score = semantic mapping 40% + deterministic validation 30% + extraction 20% + cross-document corroboration 10%
  • The weights are fixed, not statistically calibrated — whether a score tracks real accuracy is what the extraction accuracy report measures
How to demo
Open a field in review and show its score next to the four contributing signals, then the accuracy report. "The score is a rule we publish; the report is how we check it."
Trade Documents/commercial_invoice_88.pdf
SR
Sarah
Signals behind the score
Semantic mapping (40%)95
Validation (30%)100
Extraction (20%)80
Corroboration (10%)40
Illustrative values: 95×0.4 + 100×0.3 + 80×0.2 + 40×0.1 = 88

Malware screening & quarantine

A file is never reported clean unless a scanner actually ran.

Documents come from outside the organisation. Every upload gets a verdict — CLEAN, NOT_SCANNED, or QUARANTINE — with the reason recorded in the audit trail, so it is never ambiguous afterward whether a given document was scanned.

  • ClamAV scanning when configured, with an optional VirusTotal hash pre-check
  • No scanner configured means NOT_SCANNED, said plainly — never a fabricated "clean"
  • In block mode, a scanner error quarantines the file (fail-closed); in advisory mode it is stored and recorded as unscanned
  • A quarantine inbox lists held files with bulk actions
How to demo
Open the quarantine inbox, show the recorded reason on a held file, and point at the scan mode setting. "We tell you what was checked and what wasn't."
Trade Documents/Quarantine
SR
Sarah
FileVerdictReason
invoice_new.pdfQUARANTINEFlagged by scanner
scan_0012.pdfQUARANTINEScanner error (fail-closed)
Illustrative data. The verdict and reason are written to the audit trail for every upload

Documents beyond extraction

Ask the customer for a document, control what they see, bill from the paperwork.

Once a document is on file it keeps working: a broker can request a missing one from the customer, choose which documents the customer portal shows, and turn a supplier invoice into a draft vendor bill.

  • Request a document: pick a type and a recipient email; it opens a request on the shipment with a due date and a message, and provisions the recipient's portal access
  • Per-document visibility: internal or customer, each change written to the audit log
  • Customer document package: when the Entry Proof is published, the customer-visible documents plus the CBP 7501 are bundled into one download; if nothing is customer-visible, no package is made
  • Vendor bill from a document: a classified vendor invoice drafts a bill (vendor, number, dates, currency, totals, lines), gated by the billing permission
How to demo
Request a missing signed invoice from the customer, flip one document to customer-visible, then draft a vendor bill from a supplier invoice. "Three follow-ups, no retyping."
Shipments/Documents
SR
Sarah
DocumentVisibilityAction
Commercial invoiceCustomer—
Broker worksheetInternal—
Freight invoiceInternalDraft vendor bill
Illustrative data. Visibility changes are audit-logged; the customer package only includes customer-visible documents

Qubere Learns

Correct it once per account — it stops making the same mistake for that account.

A field-review verdict that actually corrects a value (not just approves it) is durably remembered per account + document type + field, the same institutional-memory mechanism already feeding the HTS, origin, valuation, and filing agents. The next extraction for that account retrieves its own correction history and folds it into the prompt.

  • Triggered only on a real correction, dispatched durably via Inngest so a dropped serverless invocation can't lose it
  • Retrieval degrades silently on failure — a memory-lookup problem never blocks or slows extraction
  • The persisted agent decision records how many memories were used and flags Account Institutional Memory as a data source when it was
How to demo
Correct the same field for one account twice, then reprocess a similar document for that account — show the decision's data sources listing Account Institutional Memory. "It remembers this customer's paperwork."
Trade Documents/commercial_invoice_212.pdf
SR
Sarah
Account Institutional Memory used
Memories retrieved2
Applied tocommercial_invoice → Incoterm
Remembered"Incoterm is reliably FOB (corrected by a human reviewer)"
Extracted valueFOB
Data sourcesGemini 2.5 Flash Vision, Account Institutional Memory
Same mechanism already feeding the HTS classification, origin, and filing agents — extended to document extraction

Trade Data Normalization

Extracted exactly as written, reconciled into compliant, standardized values.

The extraction agent is deliberately forbidden from normalizing anything — it captures currency symbols, units, dates, and codes exactly as written so nothing is silently altered before a human can see it. A separate, versioned normalizer registry then reconciles each field to its canonical form as it's promoted onto the shipment record, with a matching validator confirming the result is actually compliant.

  • Currency & Incoterms → ISO-3 / ISO-3166 codes, each checked by a dedicated validator (isoCurrencyValidator, isoIncotermValidator)
  • Country & state → full ISO-3166-1 alpha-2 name-to-code mapping, not a hardcoded shortlist that silently drops unmapped countries
  • Dates, money, HTS codes, party names, and identifiers → each with its own deterministic normalizer, versioned in one registry so every promotion is reproducible
How to demo
Upload an invoice with "United Kingdom", "fob", and "03/14/2026" — show them land on the shipment record as GB, FOB, and 2026-03-14, with the original document text still visible as the source evidence.
Trade Documents/commercial_invoice_212.pdf
SR
Sarah
Country of Origin
Extracted (verbatim)"United Kingdom"
NormalizedGB
Validatoriso2CountryValidator ✓
Incoterm
Extracted (verbatim)"fob"
NormalizedFOB
ValidatorisoIncotermValidator ✓
Original extracted text stays on record as evidence — normalization never overwrites the source

Missing-data & document resolution

It knows which documents this entry type actually needs — and closes the loop when they show up.

Required documents are looked up by the shipment's CBP entry type, not a generic checklist, so a Type 01 consumption entry and a Type 86 entry never get the same missing-doc list. Each gap becomes a real, actionable exception — and it clears itself the moment the right document lands.

  • Entry-type-specific requirements resolve from a shared lookup so the shipment table, the exception list, and the filing gate can never disagree on what's missing
  • Every missing-doc exception carries an upload prompt naming the exact document needed — not a generic "documents missing" flag
  • A document upload auto-resolves any customer request linked to it — no one has to remember to go close the loop by hand
How to demo
Open a shipment with a Packing List Missing exception and show the AI prompt to request it from the counterparty. Upload the packing list — watch the exception and the linked request both resolve automatically.
Shipment SHP-2026-000144
SR
Sarah
Packing List Missing
Required forentry type 01 · customs entry filing
ActionUpload the Packing List to clear this requirement
On upload → linked customer request auto-resolves · exception clears without a manual step

Document Rules Studio

Account-specific document logic, tested before it's ever live.

Every account has its own document quirks — a broker who always wants a specific invoice field flagged, a client whose packing lists never carry a certain code. Rules Studio lets an admin author that logic as a typed, versioned rule instead of a code change, and nothing goes live until it's been dry-run against a real document.

  • Every rule version moves through DRAFT → TESTING → APPROVED → PUBLISHED → RETIRED — only one version is ever live per rule
  • Conditions match ALL or ANY of a field list against EXISTS / NOT_EXISTS / EQUALS / NOT_EQUALS / CONTAINS / GT — no free-form scripting a reviewer can't audit
  • Document types and condition fields are picked from dropdowns, and the API rejects any value that could never match — no silent typo rules
  • Deactivate is the kill switch for a live rule; Activate brings it back; Delete is only offered once a rule is no longer live
  • Test Version runs a draft against one real document and shows the result without changing anything on that document
  • Published rules evaluate automatically as part of the document intelligence agent's classification pass
  • Lives in the Doc Studio sidebar item — Overview, Document Rules Studio, Field Precedence and Simulator tabs in one place
How to demo
Create a draft rule, click Test Version against a real document to show the dry-run result, then publish it and reprocess a matching document to watch it fire for real.
Doc Studio/Document Rules
SR
Sarah
Conditions — match ALL
signatureDetectedEQUALS false
documentTypeEQUALS PACKING_LIST
Version history
v3PUBLISHED · live
v2RETIRED
Test runs against one real document and changes nothing — only Publish makes a version live

Source-of-truth precedence

When two documents disagree, the account decides which one wins — not a coin flip.

An invoice and a packing list can report a different weight for the same shipment. Instead of always freezing on a manual review, an account can publish a precedence policy — "packing list wins on weight" — and the reconciliation engine applies it automatically, every time, for that field.

  • Policies are DRAFT until published; only one PUBLISHED policy is ever live per account + field
  • Deactivate, Activate and Delete from the policy page — activating a policy retires the field's other live one; a live policy can't be deleted; past resolutions keep their policy version
  • The field is chosen from the fields reconciliation actually compares, so a misspelled key can't create a policy that never applies
  • No policy for a field → falls through to the existing manual review queue, unchanged
  • A policy that can't unambiguously pick a winner among the candidates returns REVIEW_REQUIRED rather than guessing
  • Every resolution is idempotent and written once per shipment + field, so re-running reconciliation never double-applies a resolution
How to demo
Show a weight conflict resolving on its own the moment a published precedence policy exists for that field — then show a field with no policy still landing in manual review exactly as before.
Doc Studio/Field Precedence
SR
Sarah
commercial_invoice_88.pdf412 kg
packing_list_88.pdf438 kg · selected
Policy applied
FieldgrossWeight
RulePACKING_LIST source always wins
StatusRESOLVED · written once
No policy for a field → unchanged manual review queue, not a guess

Field-level explanation

"Why is this the value?" answered in one click, from the real chain.

Next to any extracted field is an explain control that reads the same provenance chain a person would have to click through by hand — source document, page, confidence signals, and any precedence or reconciliation decision that touched it — and states it as a short, sourced answer instead of prose.

  • Built from the same extraction, confidence, and precedence records shown elsewhere — nothing is inferred or invented for the explanation
  • Line items get the same control: an info icon on each line's HTS code, country of origin and unit price explains that line, not the whole document
  • Product Intelligence raises one open decision per shipment rather than one per document, so a reviewer isn't asked the same question repeatedly
  • Surfaces as a popover directly on the field review panel, no separate page or tool
How to demo
Click the explain icon on a resolved field and show the popover naming the exact source document, page, and (if it applies) the precedence policy that decided it.
Shipment SHP-2026-000144/Field Review
SR
Sarah
Why this value?
Value438 kg
Sourcepacking_list_88.pdf · page 1
Decided byField Precedence policy · PACKING_LIST wins
Confidenceparser 0.95 · model 0.91
Same provenance and precedence records shown elsewhere — nothing new is computed for the explanation

Correction-learning rule suggestions

If reviewers keep fixing the same field, it says so — and offers a draft rule.

The recorded extraction-accuracy events (every approve/edit/reject/select-alternate) are aggregated by field and document type. When a field crosses a real repeat-correction threshold, it surfaces as a suggestion — never an auto-applied change.

  • Surfaces only once a field/document-type pair has at least 5 recorded events and a combined correction+rejection rate above 40%
  • Accept creates a DRAFT-only FLAG_FOR_REVIEW rule in Document Rules Studio — it still has to go through Test & Publish like any other rule, never auto-published
  • Also suggests field mappings from correction history: when the same header was mapped to the same field across at least 3 distinct documents, Accept creates a DRAFT FIELD_MAPPING rule — accepting twice never duplicates it
  • Dismiss is durably recorded per field + document type, so a suggestion an admin already decided not to act on doesn't keep resurfacing (mapping suggestions are keyed per header → field)
How to demo
Open the Correction Suggestions panel, show a repeat-correction pattern, click Flag for review and land directly on the new draft rule — then dismiss a different suggestion and reload to show it stays dismissed.
Doc Studio/Document Rules/Suggestions
SR
Sarah
hsCode · COMMERCIAL_INVOICE
Sample size11 events
Corrected + rejected rate63%
Accept creates a DRAFT rule only · Dismiss is remembered per field + document type

Cross-shipment duplicate detection

A document that already exists on another shipment gets flagged, not silently re-filed.

Every inbound attachment — whether it lands through normal processing or is released out of quarantine — is checked against documents already on file elsewhere in the account. A match surfaces as an advisory notification, without ever blocking storage or parsing.

  • Fires on both the standard attachment-processing path and quarantine release — one detection rule, not two
  • Advisory only: storage, parsing, and release all continue regardless of the result
  • Routed to every role-holder with document-update permission on the account, deduplicated so the same match doesn't spam the same person twice
How to demo
Upload the same invoice against two different shipments — the second upload stores and parses normally, but document-update role holders get a "matches N document(s) already on file" notification.
Notifications
AC
Alice · Acme
TypeDOCUMENT_POSSIBLE_DUPLICATE
Message"commercial_invoice_88.pdf" matches 1 document(s) already on file.
EntityShipmentDocument
Routed toaccount role-holders with document.update
Storage, parsing, and quarantine release all proceed regardless of this notification

Relevance-ranked full-text search

Search the document index like a search engine, not a filter form.

A GIN-indexed, weighted full-text vector sits on every document, so a search term ranks the best match first instead of just filtering. Filenames and identifiers weigh higher than incidental text, and the same query still works even against a database that hasn't run the latest migration.

  • Weighted A/B/D full-text vector, generated at the database level so it's always in sync with the row
  • Relevance becomes a real, selectable sort column — and the default the moment a search term is present with no explicit sort
  • A live-vector fallback keeps ranked search working on databases mid-migration, instead of hard-failing
How to demo
Search the document index for a partial invoice number or filename fragment — the closest match lands at the top, not buried in upload-date order.
Trade Documents/Search
AC
Alice · Acme
DocumentRankShipment
commercial_invoice_88.pdf0.91SHP-000144
acme-manufacturing-rate-card.csv0.44SHP-000112
packing_list_88.pdf0.31SHP-000144
ts_rank over a GIN-indexed, weighted tsvector — falls back to a live vector on pre-migration databases

Canonical file naming

"invoice_final_v3_scan (2).pdf" becomes a name you can actually file by.

Alongside the untouched original filename, every document now carries a generated, canonical display name — {documentType}_{shipmentNumber}_{YYYYMMDD}{ext} — built the moment the document intelligence agent classifies it. The messy inbound name stays on record; the canonical name is what people actually see.

  • Purely additive — the original fileName is never overwritten or lost
  • Wired into classification, the documents API/query, and the document list UI in one pass
  • A backfill script applies canonical names to documents uploaded before this shipped, so the index isn't split between old and new naming
How to demo
Upload a file named something meaningless like scan0007.pdf — after classification, the document list shows COMMERCIAL_INVOICE_SHP-2026-000144_20260927.pdf instead.
Trade Documents
AC
Alice · Acme
Original filenamescan0007.pdf
Canonical nameCOMMERCIAL_INVOICE_SHP-2026-000144_20260927.pdf
Built fromdocumentType + shipmentNumber + classification date
Original filename retained on the record — canonical name is display-only

Ask Qubere · the assistant

The answer is a table of real shipments — or there is no answer.

~70 tools across shipments, filing readiness, screening, decisions, exceptions, rulings, drawback. Any claim that names a specific record renders as a linked reference. If the model can't back a number with real rows, it doesn't state the number.

  • "Value at risk" is the same computation the Command Center runs
  • Follow-ups compose: "now just the ones assigned to Priya"
How to demo
Ask "which shipments are at risk and what's the dollar exposure?" in /chat, then open /app/dashboard — show the identical figure. That match is the whole trust argument.
✦ Ask Qubere
JT
Joe · Target
Which shipments are at risk and what's the dollar exposure?
called get_value_at_risk() · list_shipments({ healthStatus })
4 shipments below the readiness threshold, $199,949 declared value at risk.
SHP-2026-00000728% ready$96,120critical
SHP-2026-00000215% ready$61,400critical
SHP-2026-00001952% ready$42,429at risk
now just the ones missing a packing list

Guardrails & write actions

It can act — and it can't talk its way past a permission.

The registry is filtered to what the caller may use before it's offered to the model; the orchestrator re-checks the tool name before executing. A PLANNER asking "show all shipments" gets the same row-level filter the Shipments page applies — same code path.

  • create_shipment, classify_product, approve/reject_decision, resolve_exception — each permission-gated
  • Every write → AuditLog, tagged assistant-initiated, behind an explicit confirm
  • 6 tool rounds / turn · 15 q/min per user · shared daily token ceiling
How to demo
Log in as a PLANNER and as an ADMIN; ask the same question; show the different result sets. Then ask it to create a shipment — watch it ask only for the importer, show a summary card, and wait for a yes.
✦ Ask Qubere
JT
Joe · Target
approve the classification decision on SHP-2026-000144 line 2
Confirm this action
Actionapprove_decision · DEC-88213
HTS8537.10.9170 · GRI 1
Permissiondecisions.approve ✓
On approve → AuditLog: DECISION_APPROVED · source: assistant · DEC-88213

Origin safety

It will not substitute a manufacturer country for an origin it doesn't have.

A dedicated tool surfaces a product's legal country-of-origin position. The system prompt forces the model to quote it verbatim and forbids falling back to a manufacturing, supplier, ship-from, or export country — even when the tool's own physical-fact fields mention one.

  • Retrieved document content enters the context in a labelled envelope — never an instruction
  • The model never sees SQL, an internal API, the page DOM, or an account id
How to demo
Ask "what's the country of origin for product X?" on a product with no approved determination. It says so — it doesn't guess. This is Reasonable Care, enforced at the AI layer.
✦ Ask Qubere
AC
Alice · Acme
What's the country of origin for the BMS control board?
called get_product_origin_position()
There is no approved country-of-origin determination on record for this product. I can't infer one from the manufacturing country (Vietnam) or the supplier — origin is a legal determination, not a physical fact.

Physical facts on file: manufactured in Vietnam · supplier Hanoi Precision · last CoO review: none.

Interactive chat & opt-in verbose audit

It asks before it acts — with real buttons, not a text-parsing hack.

An ask_user control tool renders 1–4 labelled options (or a multi-select, or free text) as clickable cards, replacing an earlier regex-on-assistant-text confirm hack. 15 new write tools give the chat web-app parity — each a thin wrapper over the same route the equivalent screen calls, so tenancy, RBAC, validation and audit stay owned by the route, not the model.

  • create/update_shipment, create/update_product, create/update_party, create_customs_filing_draft, record_duty_payment, run_drawback_match and more — every write requires an ask_user confirmation first
  • Opt-in Account.copilotVerboseAudit (default off) logs the assistant's answer and each tool call's args/results, size-capped, onto the existing AuditLog — queryable via query_chat_audit
  • System prompt, credentials and raw document text are never persisted, on any setting
How to demo
Ask it to create a shipment with no importer named — watch ask_user render as buttons, not prose. Then turn on verbose audit and run query_chat_audit to show the full turn on record.
✦ Ask Qubere
JT
Joe · Target
create a shipment for the Ruixin PO arriving next week
Which importer is this for?
ask_user → elicitation event · answer folds into history, no re-ask

Objection handling

Every hard question, answered directly.

"Which model? Are you sending our data to OpenAI?"Reasoning is Gemini (3.6 Flash default); Claude is wired as an alternative for the assistant and a few advisory routes. Parsing is Google Document AI. Model choice is per-surface config, and provenance records which model actually ran. No OpenAI.
"What if the AI is down?"Every workflow is fully usable without it. With no model key, the assistant route says so rather than answering from nothing; agent calls fail closed rather than fabricating.
"Can it read our whole database?"No. It calls a fixed tool registry, each running with the caller's real session context. The model never sees SQL, an internal API, the DOM, or a tenant id.

The close

"Ask it for value-at-risk. Then open the dashboard. Same number — because it's the same code."

Evidence you can click. An assistant that can't outrun its own audit log.

Next step: upload one of your commercial invoices on a screen-share. We'll show you the extraction, click into the provenance, and let you try to break the assistant.

Full detail: docs/apps/customs/sales/ai-and-document-intelligence.md