Case Study

Agentic Cross-Border Trade Compliance Advisor

Nigeria → China corridor — a $13B trade lane where Form M errors and compliance ambiguity routinely stall shipments.

Agent in Action

$ corridor-agent --transaction TX-001
 
[✓] Document extraction    COMPLETE
[✓] Supplier verification  VERIFIED
[⚠] Sanctions screening    FATF flag
[✓] Regulatory retrieval   4 citations
 
VERDICT: COMPLIANCE_READY
Refs: CBN/TED/FEM/PUB/FPC/01/003
      NCS Circular 01/2022

The Four Tools

Document extraction

Parses Form M, PAAR, and supplier invoices to extract transaction details.

Supplier verification

Cross-checks supplier identity and history against known-good and flagged registries.

Sanctions / AML screening

Screens counterparties against FATF and sanctions lists, flagging matches for review.

Regulatory retrieval (RAG)

Retrieves and cites specific CBN and NCS circulars relevant to the transaction.

Architecture Decision: State Machine, Not Free-Form Agent

Corridor uses a state-machine agent rather than an autonomous free-form loop. Compliance workflows have a fixed, known sequence of checks — a state machine makes that sequence explicit, auditable, and impossible to skip. A free-form agent could technically reach the same verdict, but couldn't guarantee it checked everything it was supposed to.

Governance

The deployment rule: "automate the workflow, not the accountability." The agent verifies, screens, and cites — but any verdict short of a clean pass routes to a human reviewer. No transaction is auto-blocked or auto-cleared by the model alone when the outcome is ambiguous.

Three Demo Scenarios

Clean transaction

All checks pass — verdict is COMPLIANCE_READY with full citation trail.

Missing documentation

Extraction succeeds but required fields are absent — verdict is MISSING_INFORMATION, with a list of what's needed.

Sanctions flag

A counterparty partially matches a watch list — verdict is REVIEW_REQUIRED, routed to a human rather than auto-approved or auto-blocked.