The convergence of Artificial Intelligence (AI) with edge computing offers an opportunity to change the trajectories of time-sensitive systems by allowing inference to occur closer to the source of data rather than processing everything in a centralized cloud. Significant progress in our ability to improve model design (particularly MobileNet and EfficientNet), use of hardware accelerators (such as NVIDIA Jetson, and Google Edge TPU), and the ability to compress models in innovative ways with techniques like, quantization, pruning, and knowledge distillation has made on-device AI feasible in practice. Emerging frameworks for edge deployment of AI models (such as TensorFlow Lite, PyTorch Mobile, etc.), and new training paradigms (e.g., federated learning), further facilitate on-device computing while minimizing transferring raw data to the cloud. In practice, edge AI allows systems to make autonomous decisions faster, while operating under severe resource constraints with intermittent connectivity. This paper reviews many aspects of AI at the edge. In particular, it discusses the fundamentals of edge computing, the requirements associated with real-time systems, methods for deploying and optimizing models on devices, algorithmic strategies (for example, distributed learning and/or incremental learning), real-world applications, and challenges moving forward. The content in this paper is illustrated by recent advances and examples of usage in the real world, which demonstrate that edge AI will enable increasingly efficient deployment of high-impact AI applications with the added benefits of low latency, privacy protection, and scalability.