Abstract: This paper systematically reviews AI applications in architecture and planning design worldwide, and innovatively categorizes them into universal efficiency tools and specialized in-depth tools. Addressing the pervasive 'hallucination' problem in current general large language models and the consequent dilemmas of industry restructuring and value erosion, this paper proposes that designers should enhance human-AI collaboration literacy, universities should foster interdisciplinary talents, and enterprises should scale up R&D of vertical domain models, to advance the paradigm shift of technology-empowered human-AI collaboration. Using DeepSeek-7B as the base model, this study performs fine-tuning, constructs a high-quality database with over 10 million words of industry literature and expert knowledge, and develops a vertical-domain intelligent Q&A system based on the RetrievalAugmented Generation (RAG) architecture. Comparative evaluations on typical industry tasks covering cultural tourism operation, financing-oriented design and real estate economic calculation demonstrate that the fine-tuned model outperforms general large models in answer structure logic, strategic dimension completeness and professional knowledge accuracy.