AI-Driven Micro-Insurance for Boutique Start-Ups: Is the Gain in the Algorithm or in the Data?
1 viewsAbstract
Boutique start-ups (micro-enterprises employing fewer than ten people) rank among the least insured commercial risks in the markets, thus artificial intelligence is presented as the solution. The following study poses a narrower question: when an AI-enabled scheme achieves better results than a conventional one, which element of "AI" is really responsible for the performance. A portfolio of 8,000 boutique start-ups is simulated using a compound Poisson loss model with lognormal severity along with an unobservable heterogeneity term that reflects the firm's private information. Five rating and distribution regimes are evaluated on identical data: community rating, traditional tier rating, a generalized linear model, a gradient boosting model trained on digitally observable data, and the same model combined with parametric settlement. Demand follows constant absolute risk aversion, while prices and purchase decisions converge toward a fixed point, making participation and adverse selection outcomes instead of assumptions. Three findings emerge. Community rating collapses: participation drops from 43.8% to 18.7%, and the remaining pool carries twice the risk of the population. Digital risk-based rating roughly doubles coverage, lowers the average premium from $599 to $310 and reduces the expense ratio by half, from 23.4% to 11.2%. Most importantly, a decomposition reveals that essentially all of the improvement in risk classification stems from the breadth of digitally observable data rather than from the learning algorithm: the data-access effect is +0.118 Gini while the algorithm effect falls between −0.015 and +0.007. Machine learning is best regarded as a replacement for the actuarial capacity that micro-insurers cannot afford, not as a superior statistical technology.
Keywords
Adverse Selection, Machine Learning, Micro-Insurance, Parametric Insurance, Risk Classification