Invoice Exception Triage Agent

Inspiration

Accounts payable teams spend too much time checking routine invoice exceptions across supplier data, purchase orders, duplicate history, tax IDs, payment terms, and currency rules. Clean invoices should not wait in a manual queue, but risky invoices should never be approved without accountability.

What it does

Invoice Exception Triage Agent evaluates normalized invoice JSON and returns a deterministic triage response: risk level, decision, plain-English reasons, recommended action, audit summary, and machine-readable flags for UiPath routing. Clean invoices can be auto-approved, medium-risk invoices go to AP analyst review, and high-risk invoices are escalated to human approval.

How we built it

The coded agent is implemented in Python with deterministic risk rules and a FastAPI wrapper for /triage/invoice. The repository includes a dependency-light CLI fallback, mock ERP data, sample invoices, reproducible expected outputs, pytest coverage, and an OpenAPI contract. UiPath Studio Web API Workflow was configured and debug-tested against the coded endpoint. The broader UiPath flow is documented for Maestro BPMN, API Workflow, Action Center, and Automation Cloud governance.

UiPath workflow

The intended orchestration is: Start -> Receive invoice -> Call triage API -> Decision gateway -> Auto approve / AP review / Human escalation -> Audit report -> End. The high-risk path creates a human-in-the-loop approval task before ERP posting or payment.

What we learned

Agentic business workflows need both automation and governance. The coded agent makes fast, repeatable decisions, while UiPath provides orchestration, human-in-the-loop controls, audit visibility, and operational monitoring.

Codex usage

OpenAI Codex helped scaffold and harden the code, tests, OpenAPI contract, UiPath documentation, demo materials, and submission evidence. Its output is integrated into a working repository, not just a README.

Limitations

The prototype uses JSON mock ERP data for reproducibility. It does not claim a production Maestro or Action Center deployment, and production use would require hosting, authentication, secret management, ERP integration, and tenant-specific governance.

Built With

• automation
• openai
• studio
• uipath