<p>Municipal solid waste (MSW) generation is projected to reach 3.8 billion tonnes annually by 2050, placing increasing pressure on waste management systems and resource recovery pathways. Although machine learning (ML) and deep learning (DL) methods are widely applied across individual MSW tasks, no prior review has systematically linked algorithm performance to enhanced resource recovery (ERR) outcomes across the full waste management chain. This study addresses that gap through a structured, PRISMA-compliant review of 155 peer-reviewed studies published predominantly within the last five years. The review covers artificial neural networks (ANNs), support vector machines (SVMs), ensemble tree-based methods, convolutional neural networks (CNNs), and recurrent architectures, evaluated across waste generation prediction, sorting and recycling, biological and thermochemical treatment, emissions monitoring, and collection and routing. It was found that three factors consistently govern model performance; data type, dataset scale, and deployment context. Ensemble tree-based models excel on structured tabular data; CNNs lead in image-based sorting tasks, exceeding 98% accuracy in deployed systems; and LSTM variants are most effective for time-series forecasting. Hybrid and ensemble architectures consistently outperform their standalone counterparts across all domains. Key limitations include geographic concentration of training data in high-income countries, a notable simulation-to-reality performance gap in routing applications, and an unresolved accuracy–interpretability trade-off with regulatory implications. While AI shows strong potential for enhancing ERR across the MSW chain, its real-world deployment depends on standardized open datasets, wider use of explainable AI (XAI), physics-informed modelling for thermochemical processes, and greater emphasis on field-scale validation.</p> Graphical Abstract <p></p>