InvoiceShield Case

InvoiceShield Case was inspired by a simple finance problem: clean invoices can be automated, but risky invoice exceptions should not be blindly pushed through. Bank account mismatches, suspected duplicates, PO variances, and missing goods receipts are exactly the moments where automation needs governance, evidence, and human accountability. InvoiceShield turns those risky AP invoice exceptions into governed UiPath Maestro Cases. AI agents investigate the invoice, deterministic rules decide the route, a finance human owns the final money decision, and the mock ERP write is blocked until that human signs off. The case then closes with an audit summary and traceable evidence. We built the project around the idea that agents should recommend, not approve payments. Maestro handles orchestration and routing, while agents provide investigation, scoring, and explanation. The solution includes a Case Decision Agent, an Audit Summary Agent, and a coded Deep Investigation Agent using LangGraph. A Python policy layer checks vendor data, purchase orders, goods receipts, invoice history, duplicates, bank mismatches, and weak fraud signals. We learned how important it is to separate AI judgment from deterministic governance. The strongest design was not “let an agent run everything,” but a controlled case flow where rules, humans, guardrails, and agents each have a clear role. The main challenge was working with Maestro Case in an early preview environment and connecting cloud case orchestration with a realistic invoice intake flow. To keep the prototype reliable, we built a local harness with mock invoice data, deterministic tests, stress tests, and an eval suite to prove the routing and guardrails work.