InsuranceFraudAi
Cost-sensitive decision tree · fully explainable

Flag the claims worth investigating — before the payout leaves.

Insurance fraud drains billions each year, and it follows rule-like patterns — a brand-new policy, a late-night incident, a suspiciously large claim. InsuranceFraudAi learns those rule chains with a cost-sensitive decision tree and raises a Flag for Review on the riskiest files, with a transparent reason for every call.

Interactive

Score a claim in real time

Adjust the claim variables — the model re-scores instantly, entirely in your browser, and shows the exact decision path it took. Move the operating point to trade precision against recall.

Claim details

Load a scenario, or adjust any field to see the model re-score live.

Fraud probability
36%
Flag for Review
0%Decision threshold · 35%100%

Operating point

threshold 0.35

Lower the threshold to catch more fraud (higher recall); raise it to reduce false alarms (higher precision).

Why this decision

The exact rule chain the claim followed through the tree — every prediction is fully auditable.

  1. Police Report Filed: “No”
  2. CCTV / MOBESE Footage: “Yes”
  3. Incident Hour 22 h (> 7 h)
  4. Adjuster Suspicion Score 61 /100 (≤ 66 /100)
  5. Claim Amount ₺46,000 (> ₺37,879)
Policy Management

Policy Renewal Reminders

Automatically notify customers before their policies expire. Choose notification channels and set up automated reminders to improve retention.

Ayşe Demir · POL-4471
Expiry Date: August 19, 2026
13 days left

Notification Method

Days Before Expiry14 days

Message Preview

Dear Ayşe Demir, your policy POL-4471 expires on August 19, 2026. Please renew to keep your coverage active without interruption.

Case Processing

Incident Report Analyzer

Upload the mutual incident statements to automatically run a situation assessment, estimate each party's share, and generate an expert situation report.

Upload Both Party Statements

Drag & drop files here, or click to browse

Situation A Report
Situation B Report

Upload Both Party Statements

Model performance

Precision, recall & the cost trade-off

Fraud is rare and false negatives are expensive, so accuracy is the wrong yardstick. These are the numbers that matter, measured on a held-out test set the model never saw during training.

Precision / recall vs decision threshold

Every threshold is a different business choice. There is no free lunch — catching more fraud always costs precision.

PrecisionRecallhigher threshold → more precision, less recall

Confusion matrix at the balanced operating point

Where the model is right, and — just as important — how it is wrong.

Held-out test set · threshold 0.35Precision 30% · Recall 69%
129
True positives
Fraud correctly flagged
59
False negatives
Missed fraud — costliest error
303
False positives
Clean claims flagged for review
1009
True negatives
Clean claims auto-approved

What drives the model

Feature importance from total impurity reduction. The adjuster suspicion score, policy age and claim size dominate — matching how a human investigator triages.

Police Report Filed
21.3%
CCTV / MOBESE Footage
20.1%
Adjuster Suspicion Score
16.1%
Incident Hour
14.0%
Policy Age
13.2%
Witness Present
10.8%
Claim Amount
3.7%
Days to Report
0.9%

Share of total Gini impurity reduction attributed to each feature across the tree. Ranks (not scaled input values) — robust to feature units.

Under the hood

How InsuranceFraudAi works

A deliberately simple, auditable pipeline. No black boxes — every prediction can be traced to a rule chain and defended to a regulator.

CART decision tree

A classification tree learns axis-aligned rule chains that mirror how fraud actually unfolds — exactly the structure investigators reason about.

Cost-sensitive learning

Class weights during training plus a cost-optimal threshold at inference encode that a missed fraud hurts far more than a false alarm.

Tunable operating point

Slide the threshold to move along the precision/recall curve and hit whatever risk appetite the fraud team is working to.

Explainable by design

Every score ships with the decision path and the top risk factors, so an analyst can justify opening — or closing — an investigation.

Training data
4,500 synthetic claims, seeded & reproducible, ~12% fraud base rate
Tree shape
depth 6 · 41 leaves · evaluated on 1,500 held-out claims
Runtime
Inference is pure tree traversal — microseconds, on the edge or in the browser