From Fraud Score to Fraud Investigation: Building an Agentic Fraud Investigator with TigerGraph

Fraud detection is often framed as a classification problem:
transaction → model → fraud probability

That is useful, but it is not enough for an investigator.
A real investigation asks different questions:

• What is connected to this transaction?
• Is the same device associated with other cards?
• Has the customer or card appeared in previous investigations?
• Does the transaction belong to a suspicious temporal sequence?
• What evidence supports fraud?
• What evidence supports legitimacy?
• What information is still missing?
• Should we act now or gather more evidence?
• What action is allowed under policy?
• Can every final claim be traced back to evidence? For Hacker House Goa 2026, we built an agentic fraud investigation system using TigerGraph, GraphRAG and LangGraph to explore exactly that problem. The result is less like a fraud classifier and more like a bounded digital investigator.