The rapid proliferation of cloud-native applications, hybrid work models, and bandwidth-intensive services has fundamentally challenged the static nature of traditional Wide Area Networks (WAN). Software-Defined WAN (SD-WAN) introduced a centralized control plane to decouple network software from hardware, yet the manual definition of steering policies often fails to account for the highly volatile nature of internet transport circuits. This review examines the paradigm shift toward AI-based Traffic Engineering (TE) within SD-WAN architectures. By leveraging Machine Learning (ML) and Deep Learning (DL) models, SD-WAN controllers can now transition from reactive threshold-based switching to proactive, predictive traffic steering. We categorize current methodologies, focusing on the use of Reinforcement Learning (RL) for dynamic path optimization and Long Short-Term Memory (LSTM) networks for forecasting link congestion. This article explores how AI-driven TE optimizes Quality of Experience (QoE) for mission-critical applications—such as VoIP and real-time video—by analyzing multi-dimensional telemetry including jitter, latency, and packet loss in real-time. Furthermore, the review addresses the critical challenges of model interpretability in network operations, the \"cold start\" problem in new deployments, and the necessity for federated learning to ensure data privacy across multi-tenant SD-WAN environments. By synthesizing recent academic breakthroughs and industrial implementations, this paper provides a strategic roadmap for building \"Self-Driving WANs.\" The findings suggest that AI-integrated traffic engineering not only reduces operational expenditure (OPEX) by automating complex routing decisions but also provides the \"Cognitive Intelligence\" required to manage the unpredictable performance of commodity internet underlays in a global digital economy.