Inspiration

Sri Lankan corporates generate rich cash flow data every day, but treasury teams still manage it manually because enterprise treasury optimization is priced as a costly add-on most companies never license.

What it does

Treasury Copilot Agent reasons toward a treasury goal, like maintaining a liquidity buffer while maximizing yield and proposes concrete cash actions with clear rationale. Every proposal is approved by a human before execution, and rejections shape future proposals.

How we built it

An agentic loop orchestrated by an LLM, calling internal tools. Mock services to mirror realistic integration.

Challenges we ran into

Designing genuine judgment-making instead of a linear pipeline handling conflicting signals, low-confidence forecasts, and ambiguous data without just following fixed rules.

Accomplishments that we're proud of

A human-in-the-loop gate that actually learns from rejections, and an architecture grounded in real infrastructure.

What we learned

Building a truly agentic system means designing explicit judgment and failure-handling logic, not just wrapping a model and calling it autonomous.

What's next for Treasury Copilot Agent

Pilot integration with a real ERP and bank sandbox, and extending the same reasoning pattern to other treasury workflows like reconciliation and vendor payments.