Deep Learning (DL) is driving AI-based technologies to unpredictable achievements by automatically extracting rich data representations through an End-to-End (E2E) learning process. However, E2E models are intrinsically black boxes, hindering the possibility to understand and analyze the model decision process. This further leads to: (i) models that are highly susceptible to imperceptible perturbations in the data distribution, and (ii) models requiring huge amount of data to be trained, or even just finetuned. Recently, several contributions in literature have shown that these issues can be mitigated by employing concept-based network architecture. In general, concept-based architectures provide the final classification based on further classes which are predicted by previous layers of the network (concept-based networks). This allows to explain the prediction of the main class in terms of the other classes, to interact with the model, and also to check whether the predictions are consistent with the domain-knowledge. Latest contributions, moreover, have also shown that these models achieve classification performance that are better than any interpretable model and at least inline with black-box E2E neural networks. In this paper, we propose for the first time to comprehensively analyze concept-based approaches and, more importantly, to show their benefits when compared to standard E2E multi-class learning.