Abstract: Driven by advances in deep learning and computational power, artificial intelligence in chemistry is evolving from single-property prediction toward workflow-integrated systems embedded in real-world research and development processes. This article outlines four key themes: data and representation, prediction and generation, experimentation and closed-loop integration, as well as large-model agents and trustworthy evaluation. It surveys multimodal and weakly labeled data management, self-supervised pre-training and 3D equivariant modeling, diffusion-based and other generative paradigms with generate-evaluate iteration, and planning-control loops on programmable experimental platforms. The discussion highlights the critical role of uncertainty calibration, out-of-distribution robustness, and safety governance for practical deployment, concluding with an outlook on advancing automated chemical discovery through open standards and collaborative infrastructure.