PayPilot AI: Agentic Payment Operations

About the Project

Managing payment operations can become complicated, especially when businesses have to deal with multiple transactions, payment statuses, refunds, and invoices. Many of these tasks still require manual effort and switching between different tools.

We wanted to explore how AI could make these workflows simpler. That is how we came up with PayPilot AI, an AI-powered payment operations assistant built around PayPal.

With PayPilot AI, users can interact with the system using natural language instead of manually navigating every payment operation. They can ask the assistant to create orders, capture payments, initiate refunds, create invoices, and check transaction statuses.

How It Works

We built PayPilot AI around a combination of an LLM and a deterministic backend. The LLM helps understand what the user is asking, while the backend handles validation and makes sure the requested operation is safe to proceed with.

When a user submits a request, the system identifies the intended operation and extracts the required information. If something is missing or unclear, it asks the user for clarification rather than making assumptions.

One challenge we focused on was handling references such as "capture that order" or "refund the previous payment." The system uses verified transaction records and session context to resolve these references. If it cannot identify the correct transaction, it asks the user instead of guessing.

Before a financial operation is executed, PayPilot AI creates a pending action that requires explicit human approval. This keeps the user in control rather than allowing the AI to independently execute financial transactions.

We also built transaction tracking, audit logging, and persistent storage to make the workflow easier to review and maintain.

Technology Used

  • Python and FastAPI for the backend
  • LLM integration for natural-language understanding
  • PayPal REST APIs for payment operations
  • SQLite for transaction and workflow persistence
  • HTML, CSS, and JavaScript for the frontend
  • Pytest for automated testing

Challenges and What We Learned

One of the biggest challenges was making the AI reliable when dealing with financial operations.

An AI model might misunderstand an amount, confuse two transactions, or assume that a payment has been completed when it has only been proposed. We wanted to avoid these situations.

We addressed this by separating language understanding from financial execution. The AI interprets the request, but deterministic backend logic validates the information, checks transaction context, and enforces the approval process.

We also worked on handling incomplete requests, preventing duplicate actions, maintaining session context, and recording important workflow events.

Through this project, we learned that building an agentic AI application is not just about making an LLM perform tasks. It is equally important to know when the system should stop, ask questions, and let a human make the final decision.

What's Next

We see PayPilot AI as a starting point for making payment operations more conversational, accessible, and easier to manage. Our next focus is strengthening execution reliability, improving contextual interactions, and making the platform more production-ready.

Our goal is simple: make payment operations easier to manage without losing the control and trust that financial workflows require.

Built With

Share this project:

Updates

Submission history