India’s First AI-Powered GST Tax Intelligence System

By the time Andhra Pradesh's Commercial Tax Department brought us in, it was processing thousands of GST returns and millions of e-invoice records every day, across twenty-six commercial tax divisions. Officers were sampling a fraction of those filings manually — the rest moved through the system unread. The sophisticated fraud — circular trading networks, invoice factories, manipulated input tax credit claims — operated in everything the team never had time to look at. The department had the data. It did not have the time.

The GST AI Tax Officer is an agentic system that sits on top of the department's existing infrastructure. Data sources reach the system through Model Context Protocol (MCP) connectors — one each for GSTN filings, the e-invoice portal, UPI transaction telemetry, DISCOM-linked GST registrations, and cross-state data exchange. Underneath the agent layer, a PageIndex retrieval substrate indexes millions of historical filings, supporting documents, and prior case evidence so every agent has instant grounded context for any return it inspects. Three specialist agents run in parallel against the stream. The Anomaly Agent flags mismatched HSN codes, sudden volume spikes, and outlier filing patterns. The Graph Agent maps dealer-to-dealer relationships across millions of returns and surfaces circular-trading rings and invoice factories — patterns that only resolve as a network, never as a single filing. The ITC Verification Agent cross-references claimed input tax credits against the upstream supplier's actual filings. When any agent fires, an orchestration layer packages the return with its full evidence chain — the filing itself, the supplier graph, the PageIndex-retrieved supporting documents, and the rationale for the flag — and routes it to a tax officer for one decision: dismiss, audit, or escalate. The officer never has to leave their workspace.

Two things, honestly. First, we shipped the Anomaly Agent first and underweighted the Graph Agent — circular trading shows up most clearly as a network of dealer relationships, not as individual filings, and we spent the first eight weeks chasing flagged returns the Graph Agent could have caught in days had we started there. Second, we built the officer workspace for desktop and then discovered the field reality: GST verification often happens on the move, at a dealer's premises, with a phone. The next deployment ships with a mobile evidence-review flow alongside the desktop console — not after it.