We present the concept of a hybrid and hierarchical XAI 2.0 system, the core of which is the Kolmogorov-Arnold Networks (KAN). This architecture overcomes the limitations of traditional neural networks and post-hoc methods by integrating the intrinsic interpretability of KAN with fuzzy logic, hybridization techniques, and multi-level analysis. We consider both the advantages of KAN (accuracy, interpretability) and their disadvantages (computational complexity). The proposed multi-level structure combines KAN with other models (MLP, CNN) and uses hierarchical approaches (HKAN, Multi-Exit KAN) to create adaptive and fully transparent systems capable of incremental learning without catastrophic forgetting.