In the design engineering process of assembly modeling, existing computer-aided design (CAD) models of parts are combined to build new products, termed assemblies. Due to the enormous variety of parts available, designs of inexperienced designers can easily show unusual part combinations stemming from unfamiliarity or a lack of suitable alternatives. This paper addresses the challenge of handling anomalies in CAD assemblies with a twofold goal: first, we aim to identify such anomalies, and second, to provide suggestions for alternative parts to correct these anomalies. We employ a graph-based representation of CAD assemblies and utilize graph neural networks (GNNs) to develop models for the two respective tasks modeled as node classification problems. The models are evaluated both separately and in an end-to-end (E2E) fashion, i.e., considering the task of handling anomalies as a whole. Our experiments demonstrate their effectiveness in improving the quality of CAD assemblies with minimal human intervention.