This study investigates the use of interpretable XGBoost for predicting the severity of workers' compensation insurance claims. Although the XGBoost standard has been widely applied in insurance analytics due to its strong predictive performance, its lack of transparency limits its practical adoption in actuarial practice and regulatory contexts. To address this, we evaluate XGBoost with a maximum tree depth of two (XGB2), a simplified and interpretable variant that naturally resembles a GAMI. Experiments conducted on three datasets - WorkComp, WCC, and ALP - demonstrate that XGBoost slightly outperforms XGB2 in most cases, but the performance gap is relatively small. The interpretability analysis of XGB2 on the WCC dataset highlights the dominant role of medical and indemnity costs, along with contextual organizational factors, in determining claim severity. The results confirm that XGB2 provides not only competitive predictive accuracy but also the transparency required to support fair pricing and actuarial decision-making in the insurance sector. © 2025 IEEE.