Abstract: High-fidelity computational fluid dynamics (CFD) and wind-tunnel testing are central to conventional automotive aerodynamic shape optimization, but their relatively high computational and experimental costs limit large-scale exploration during the early design stage. This paper reviews recent progress in artificial intelligence (AI) for automotive aerodynamic shape optimization from the perspective of an intelligent computer-aided engineering (CAE) closed loop. We first summarize the external-flow problem, including aerodynamic objectives such as drag, lift distribution, crosswind stability, and thermal management, as well as geometric, packaging, regulatory, and manufacturing constraints. These factors define an optimization problem involving multiple objectives, multiple constraints, and substantial computational cost. We then organize the literature along three technical lines: forward rapid evaluation, inverse shape design, and agent-based workflow scheduling. For forward surrogate modeling, we review the evolution from low-dimensional parametric geometry representations to direct geometric representations and summarize the corresponding model routes. For inverse design, we compare surrogate-assisted classical optimization with AI-native generative inverse design. For agent-based design, we trace the development from computational fluid dynamics workflow automation and sequential design validation to multi-round vision-physics-decision loops. Based on this review, and with attention to the need for further improvement in tool-chain integration, engineering-constraint representation, and high-fidelity feedback, we construct a conceptual Trusted Intelligent CAE Closed-loop Framework for early-stage automotive design. The framework consists of design-task organization, an intelligent design optimization inner loop, and a high-fidelity verification and knowledge-feedback outer loop. It integrates three-dimensional geometry generation and local editing, physics-based AI surrogate models, agent-based decision-making, and human review into a unified workflow. Trustworthiness is understood as traceable geometry, checkable physical prediction, reviewable engineering constraints, and auditable key decisions, rather than fully autonomous reliability. Finally, we summarize further research needs concerning geometry interfaces, locally controllable generation, adaptation to multiple objectives and operating conditions, engineering-constraint integration, uncertainty-aware verification, and high-fidelity feedback updating. Future directions related to physics-aware modeling, engineering-oriented geometric integration, and expert-level decision-making are also discussed.