Herbal medicine has long been characterized by multi-component, multi-target, and multi-pathway therapeutic actions, but its complex mechanisms remain difficult to explain using conventional single-target pharmacology. Network pharmacology provides a systems-level framework for decoding these effects by integrating herbal compound databases, disease-associated gene resources, protein–protein interaction networks, and graph-theoretical algorithms. This review summarizes the most recent advances in herbal network pharmacology, with emphasis on core databases, network topology metrics, hub-target identification, and multiscale interactome analysis. It further discusses how artificial intelligence, especially graph neural networks, knowledge graph embedding, and explainable AI, is improving herb target prediction and mechanistic interpretation. The integration of single-cell transcriptomics, spatial metabolomics, molecular docking, and transfer-learning frameworks has expanded the field from static pathway mapping toward cell-type-specific and spatially resolved systems pharmacology. Despite these advances, major limitations remain, including overreliance on in silico predictions, insufficient wet-laboratory validation, neglect of dose–response relationships, inadequate consideration of pharmacokinetics and physiological barriers, and poor reproducibility caused by extract heterogeneity. Future development should move toward quantitative, dynamic, and experimentally validated systems pharmacology, supported by standardized reporting guidelines and rigorous biophysical, omics-based, and phenotypic validation. Such advances may help transform herbal medicine into a more precise, reproducible, and internationally acceptable therapeutic resource.