Qubere AI Copilot

One assistant, woven through
every screen. Grounded in your data, gated by your permissions.

Qubere AI Copilot is the same assistant behind Command Center, Ask Qubere chat, and the Help Center — not a chatbot bolted on top, but a client of the app layer itself: same RBAC, same audit log, same tenant scoping the UI already enforces.

130+
tools, RBAC-filtered before the model ever sees them
1
AI surface to audit — the old copilot was deleted
0
write actions without an explicit confirm

The problem

Every AI copilot in this space is a liability wearing a chat bubble.

What a compliance leader says about "AI assistants" bolted onto trade software:

"It answers confidently and I have no way to check whether the number is real or invented."

No grounding

"If it can create a shipment or approve a classification, who's accountable — and where's the record?"

No accountability

"A junior analyst asks it something and gets back data they were never supposed to see."

No permission model

How Qubere addresses it

It's a client of the app, not a layer on top of it.

  • Every tool call runs through the exact route the equivalent screen uses — tenancy, RBAC, validation, and audit stay owned by that route, never re-implemented for the model.
  • The tool registry is filtered to what the caller may use before it's offered to the model — a PLANNER and an ADMIN asking the same question get different result sets, same as the UI. Tools follow the sidebar: the CBAM tools only appear for users who can open the CBAM nav item.
  • A persistent per-client knowledge graph grounds it in real rows, not a context window rebuilt from scratch every session.
  • Every write is permission-gated, audit-logged as assistant-initiated, and behind an explicit confirmation the button UI doesn't even require.

Grounded answers

"Value at risk" from the Copilot is the same number as the dashboard.

Any claim that names a specific record renders as a linked reference. If the model can't back a number with real rows returned by a tool, it doesn't state the number — it says so.

  • 130+ tools across shipments, products, parties, filing readiness, screening, decisions, exceptions, rulings, post-entry, customs valuation and CBAM
  • Follow-ups compose: "now just the ones assigned to Priya"
  • Origin safety: quotes the legal country-of-origin position verbatim — never substitutes a manufacturing country
How to demo
Ask "which shipments are at risk and what's the dollar exposure?" from the Command Center Copilot, then open the dashboard — show the identical figure. That match is the whole trust argument.
Command Center/✦ Copilot
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
Same figure a PLANNER sees on /app/dashboard, filtered to their own scope

Confirmed actions

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

An ask_user control tool renders 1–4 labelled options (or free text) as clickable cards. 15 write tools give the Copilot web-app parity — create/update a shipment, product, or party; approve or reject a decision; record a duty payment — each a thin wrapper over the same route the screen calls.

  • Every write tool's description requires an ask_user confirmation first
  • On approve → AuditLog: DECISION_APPROVED · source: assistant
  • CBP transmission, user/permission management and settings stay off the tool list entirely
How to demo
Ask it to create a shipment with no importer named — watch it ask via clickable buttons, show a summary card, and wait for a yes before writing anything.
✦ Copilot
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 ✓

Persistent grounding

It stops re-discovering your account every session.

A per-client knowledge graph — importer, approved classifications, screened parties, open findings, recent entries — is injected into the prompt instead of rebuilt from scratch each turn. The same graph, rendered customer-safe, is what a client sees in their portal as the Compliance Profile.

  • Solid edge = a DB row · dashed edge = an LLM summary · ring color = fresh / stale / pending
  • Rebuilt incrementally off the write paths that already exist — not a nightly batch job
  • Ships off by default per account; a broker turns it on when ready to demo
How to demo
Ask "which of Target's suppliers isn't screened?" — an aggregate with no dedicated endpoint. It answers by walking the graph. See the full grounding & staleness story in its own deck.
Clients/Target Corp/Assistant knowledge
SR
Sarah
47 facts · HTSUS rev 14 · 2 stale · rebuilt 4m ago
Shenzhen X (supplier)screened · fresh
8481.80 (classification)3 inputs moved
AD/CVD findingrate n/a · pending
ENTRY-TGTreadiness score 88
Account.assistantGraphEnabledcustomer-safe view → Partner Portal

Two surfaces, one engine

Full workspace where you need power. A locked-down widget where you don't.

The same assistant engine serves two purposes: the full Ask Qubere workspace for account-scoped, tool-driven work, and a compact, read-only widget in the Help Center for product questions. One codebase, one prompt-and-tool framework, two deliberately different capability sets.

Copilot workspace
Full tool registry, RBAC-scoped, can read and (with confirmation) write account data.
Help Center widget
Read-only by construction — answers only from product-help search with cited documentation sources, redirects account questions to the full workspace. Global search and Support can also hand a question off to Ask Qubere chat.
How to demo
Ask the Help Center widget "what's my value at risk?" — it declines and points to the full Copilot instead of guessing.
Help Center
JT
Joe · Target
what's my value at risk right now?
I can't check your account from here — this widget only answers product questions. Open ✦ Ask Qubere from the sidebar for anything account-specific.
how does filing readiness get calculated?
called search_product_help("filing readiness")
From the readiness guide: it checks missing documents, open exceptions, and required-field completeness per shipment.

Guardrails

It can't talk its way past a permission.

The orchestrator re-checks the tool name before executing, on top of the registry filter applied before the model ever sees the list. Rate limits and token ceilings bound blast radius; retrieved document content enters the context in a labelled envelope, never as an instruction.

  • The model never sees SQL, an internal API, the page DOM, or an account id
  • 6 tool rounds / turn · 15 q/min per user · shared daily token ceiling
  • Opt-in verbose audit (off by default) — logs the answer and each tool call's args/results, size-capped, on the existing AuditLog
  • Model routing now spans Gemini, Claude, OpenAI, and DeepSeek behind one client interface — an admin can pin a specific model per account, client, or user, with a clear precedence order over the surface default
  • If a pinned provider is unavailable, the assistant says so plainly instead of failing silent or generic — it never pretends the platform default answered on the pinned model's behalf
How to demo
Log in as a PLANNER and as an ADMIN; ask the same question; show the different result sets. Then turn on verbose audit and query it back with query_chat_audit.
Evidence & Reporting/Audit History
AC
Alice · Acme
TimeActionEntityActorSrc
12:41:07DECISION_APPROVEDDEC-88213J. Kesslerassistant
12:39:55SHIPMENT_CREATEDSHP-2026-000201J. Tranassistant
11:22:03COPILOT_QUERYvalue_at_riskJ. Tranchat
registry filtered pre-modelorchestrator re-checks tool name

Objection handling

Every hard question, answered directly.

"Which model? Are you sending our data to OpenAI?"Reasoning routes across Gemini, Claude, OpenAI, and DeepSeek behind one provider-agnostic client; Parsing is Google Document AI. Model choice is per-surface config by default, and an admin can pin a specific model per account, client, or user. Provenance records which model actually ran.
"Isn't this just the chat we saw already?"Same engine, yes — that's the point. Copilot is the branding for one assistant that shows up in Command Center, the full chat workspace, and a deliberately locked-down Help Center widget. One thing to secure and audit, not three.
"What if it's down?"Every workflow is fully usable without it. With no model key, the assistant route says so rather than answering from nothing; tool calls fail closed rather than fabricating. If your pinned provider specifically is unavailable, it tells you that too — it degrades honestly, it doesn't fail silently or pretend the fallback model was your pin.

The close

"Ask it. Confirm it. Audit it. Same permissions as the person asking."

One assistant, gated by your RBAC, grounded in your data, everywhere it shows up.

Command CenterAsk Qubere ChatHelp Center WidgetAssistant Knowledge GraphRBAC-gated ToolsConfirmed Write ActionsAssistant-initiated Audit

Next step: log in as two different roles and ask the Copilot the same question in each. The different answers are the pitch.

Deep dive on document provenance & chat internals: ai-and-document-intelligence.html · grounding & staleness: graft-inspired-grounding.html