Fraud Detection Using Agentic AI — AI skill for Claude Code

🛡️ Insurance Claims Fraud Detection Agent

An AI-powered fraud detection system that combines a traditional machine learning classifier with an LLM-based agentic reasoning layer — built to explore how structured ML predictions and generative AI can work together in a real-world insurance workflow.

What it does

1. A claims adjuster (or automated intake system) enters structured claim details — amount, policy tenure, prior claims history, filing delay, claim type, provider, and member age.
2. A Random Forest classifier scores the claim's fraud risk based on patterns learned from historical claims data. 3.Claude (Anthropic)acts as an agentic explanation layer — it takes the model's risk score and the specific triggered risk factors, reasons over them, and produces a clear, plain-language justification a human adjuster can actually act on, along with a recommended next step.

The goal is a system where the ML model handles pattern detection at scale, and the LLM agent handles the human-facing reasoning and communication — each doing what it's best at.

---

Architecture

Claim Input (Streamlit form)
        │
        ▼
Random Forest Classifier ──► Fraud Risk Score (%)
        │
        ▼
Risk Factor Detection (rule-based thresholds)
        │
        ▼
Claude API (agentic reasoning) ──► Plain-language explanation + recommended action
        │
        ▼
Displayed to the adjuster in the UI

---

Tech Stack

Layer	Tool
ML Model	Scikit-learn (Random Forest, class-balanced)
Agentic Reasoning	Anthropic Claude API
Frontend	Streamlit
Data	Synthetic claims dataset (5,000 records, ~3% fraud rate)

---

Why these choices?

• Random Forest — robust to the class imbalance inherent in real fraud detection (fraud is rare), and its feature importances give interpretable risk signals to feed into the explanation layer.
• Claude as the reasoning layer — rather than hard-coding explanation templates, an LLM can