Non-technical losses (NTL), particularly electricity theft, remain a persistent challenge for distribution operators, adversely affecting network efficiency, operational planning, and economic performance. Traditional deterministic detection methods, although transparent and easily interpretable, are limited in their ability to capture nonlinear consumption patterns and dynamic behavioral deviations. This paper presents an integrated deterministic–AI approach that combines threshold-based indicators with Kolmogorov–Arnold Networks (KAN), a recent explainable neural architecture capable of modeling complex nonlinear relations between physical and statistical indicators. The approach was applied to low-voltage distribution data collected from the distribution operator, covering more than one hundred consumers and twelve representative monthly load profiles. Comparative analysis demonstrated that the hybrid deterministic–KAN method improved detection accuracy from 78% to 92%, while reducing false positives by approximately 10–12%. Beyond its technical results, the paper contributes to the interdisciplinary education of future engineers by bridging electrical engineering, mathematics, and artificial intelligence. Compared to our previous study focused exclusively on deterministic NTL detection, this research introduces a transparent and adaptive approach that enhances both operational decision-making and academic training in the context of Europe’s energy digitalization.