Robotic Process Automation — Process Mining

Insurance claims process mining using Python and PM4Py to discover workflow patterns, detect anomalies, and visualize the end-to-end claim lifecycle with Directly-Follows Graphs and Heuristic Miner process maps.

Notebook: Open in Google Colab

---

Background

This project analyzes an insurance claims event log containing 22,625 unique claims and 131,233 total events across 6 distinct activities. The goal is to apply process mining techniques to understand how claims flow through the system, identify deviations from the expected process, and surface opportunities for automation and control improvements.

The claims process follows this general lifecycle:

First Notification of Loss (FNOL) → Assign Claim → Set Reserve → Decide Claim → Payment Sent → Close Claim

---

Dataset

Field	Description
case_id	Unique identifier for each insurance claim
activity_name	Process activity (e.g., FNOL, Assign Claim, Decide Claim)
timestamp	Date and time the activity occurred
claim_amount	Dollar value of the claim
car_make	Vehicle manufacturer
car_model	Vehicle model
car_year	Vehicle year
type_of_accident	Category of accident (e.g., Head-on, Rear-end)

---

Analysis

Key Metrics

Question	Result
Unique insurance claims	22,625
Total claim events	131,233
Unique activities	6
Claims with payment sent	18,108
Payment Sent → Decide Claim flows (anomaly)	107
Process always ends with Close Claim	False
Most common path: Set Reserve before Decide Claim	True

Key Findings

• Standardized early process: Every claim runs through the same FNOL → Assign → Set Reserve → Decide sequence, indicating a highly structured intake workflow — but also raises the question of whether low-risk claims need all steps.
• Payment before decision (control gap): 107 cases show payment being issued before a formal Decide Claim activity is recorded, signaling a governance and auditability risk.
• Not all claims close cleanly: A small number of cases end at Payment Sent without reaching Close Claim, pointing to potential data quality issues or incomplete process execution.
• 18,108 of 22,625 claims (80%) resulted in payment; the remaining ~20% were rejected or closed without payment.
• Automation opportunity: Given how linear and repeatable this process is, straight-through processing for straightforward, low-risk claims is a natural next step.

---

Process Maps

The full interactive process maps are rendered in the Colab notebook linked above.

Directly-Follows Graph (DFG)

Shows the frequency of transitions between activities across all 22,625 cases.

Heuristic Miner (HM)

Shows the most common ("happy path") flow through the process, filtering out noise.

---

Tools & Libraries

Tool	Purpose
Python	Core analysis language
pandas	Data loading and exploration
PM4Py	Process mining — event log conversion, DFG, Heuristic Miner
Google Colab	Interactive analysis environment
matplotlib	Supporting visualizations