To Catch a Thief: Explainable AI in Insurance Fraud Detection

White lies (inflated claims) cost the insurance industry billions of dollars every year. After
investing heavily to automate workflows (from policy subscription to claims processing),
digitization has ironically made fraud easier to commit and harder to catch. To an industry
drowning in data and paying out millions per day for fraudulent claims, artificial intelligence
and machine learning offer new hope.
The case introduces what it calls “explainable AI” seen through the eyes of a senior
operations executive at Shift, an insurtech unicorn company whose algorithm is used by
global insurers such as Generali France and Mitsui Sumitomo to fight fraudulent claims. The
focus is on strategy making (following a private equity funding round) and algorithm-level
decisions.
With an anonymized dataset of more than 10,000 claims and a guided coding exercise in
statistical computing softwares R and Python, students are able to backtest their strategies
on historical data. Beyond the exercise there is ample material to drive case discussion.

The purpose is to introduce the strategic objectives of an artificial intelligence model, highlighting the
concept of explainability – the reasons behind machine predictions – in order to reach the ultimate goal of the technology, which in the case of Shift is to enable its special investigative units to detect a
maximum number of potentially fraudulent claims.
The case can be taught through a process of Q&A, class discussion and (group)
assignments. Assignment questions lend structure to class discussion, i.e., introducing the
concept of explainability in data science and AI, the industry-wide problem of insurance fraud
and impediments to detection, and exploring how aligning strategic priorities can improve
the usefulness of a data exercise.