Agent Fraud Detection System · Insurance Use Case

Detect Fraud Before
the Loss Is Confirmed

A 3-layer detection pipeline — rule-based controls, behavioral anomaly scoring, and supervised machine learning — that turns raw transactional data into an investigation-ready fraud case queue.

Detection Pipeline

Rule-Based Engine
Layer 1

Deterministic · 100% Traceable

Anomaly Detection
Layer 2

Unsupervised · No Labels Required

Supervised Model
Layer 3

Adaptive · Self-improving

Investigation-Ready Case Queue

The Problem We Solve

How the Fraud Happens

Premium collection leakage — agent collects but never remits

Premium Collected

Agent collects from client

Receipt Unverified

No payer ID recorded

Funds Diverted

Never remitted to company

Payment Overdue

Policy at risk, loan taken out

Policy Lapses

Loss discovered in audit

Our Approach

3-Layer Detection Pipeline

Rule-Based Engine

Explainable

Deterministic business rules derived from EDA and domain knowledge. Every alert is 100% traceable to a specific data point — fully auditable.

Anomaly Detection

No Labels Needed

Unsupervised multi-signal scoring that catches unusual behavioral patterns — no labeled data required, works from day one.

Supervised Model

Adaptive

Probability-based fraud prediction trained on labeled cases. Learns continuously through a human-in-the-loop retraining cycle.

Raw Data
Rule Checks
Anomaly Score
Risk Prediction
Case Queue

Data Foundation

What Data Does the System Use?

6 source tables — each contributing a different dimension of fraud signal

Policy Master

Core policy lifecycle — status, coverage period, premium amount, and assigned agent

Payment Receipts

Premium collection records — payment channel, amount, and payer identity verification

Policy Loans

Loans taken against active policies — amount, date, and whether the agent is the receiver

Payment Due Dates

Scheduled premium due dates used to calculate overdue gaps and reconciliation delays

Status History

Month-over-month policy status snapshots for detecting rapid In-force → Lapsed transitions

Agent Network

Family relationship graph among agents — used to detect coordinated network-level fraud

Key Advantages

Why This Approach Works

Fully Explainable

Every alert links back to a rule, field, and threshold. No black box — investigators always know why.

Works from Day One

Rule-based and anomaly layers need zero labeled fraud data. No cold-start problem.

Human-in-the-Loop

Investigators review and label cases. Verdicts feed back into the model — continuously improving accuracy.

Audit-Ready Evidence

Each case dossier contains receipts, loans, status history, and rule evidence for compliance reporting.

Ready to explore the platform?

Walk through each module with mock data — or import your own CSV files to validate against real agent portfolios.