B2B finance teams often know that an invoice is overdue, but the real challenge is understanding what happened, why recovery stalled, what evidence exists, what action is safest, and when a human must intervene.

We built Vasooli AI to solve this problem. Vasooli AI is an AI-powered B2B Revenue Recovery Intelligence and Decision System that helps finance teams investigate overdue invoices, understand customer payment commitments, evaluate evidence, calculate recovery priority, and recommend safe recovery actions.

The system follows a complete workflow: Understand, Investigate, Reason, Human Approval, Action, Payment Verification, and Audit.

Vasooli AI can understand natural-language and Hinglish payment commitments, such as a customer promising to transfer ₹1 lakh by Friday. It extracts the promised amount and date and uses this information during the recovery decision process.

The system also provides an explainable recovery priority score. For our flagship case, invoice #INV-2026-8891, the dynamically calculated priority score is 77/100 based on aging severity, exposure, broken promises, customer reliability, evidence quality, and MSMED statutory signals.

A major feature is human-in-the-loop governance. The AI does not independently execute sensitive recovery actions. It recommends an action, and a finance manager must review and authorize it before the action is dispatched.

After approval, the system can generate a Razorpay payment link using the sandbox simulation workflow and verify the payment result. Recovery actions and important decisions are also recorded in the audit trail.

We also built a safety case where the system refuses sensitive escalation when the required evidence is missing. This safety decision is enforced by the backend workflow rather than being only frontend text.

Technically, the project uses React and TypeScript for the frontend, Node.js and Express for the backend, Prisma and SQLite for data management, REST APIs for communication, AI and agent workflows for investigation and reasoning, and Razorpay sandbox simulation for payment processing.

One of our biggest challenges was maintaining consistency across the complete recovery lifecycle. We solved this by connecting the different application views and workflow stages to the same underlying case data and APIs.

Another challenge was understanding payment commitments from conversational text, especially Hinglish messages. We designed the system to extract useful information such as promised amount and target date from these conversations.

Through this project, we learned that AI in financial workflows should not only be intelligent. It should also be explainable, evidence-driven, secure, auditable, and governed by humans when sensitive actions are involved.

Vasooli AI transforms an overdue invoice from a simple outstanding balance into a complete, traceable, and safe revenue recovery workflow.