Building code consultation is a critical and inefficient task for architecture, which needs highly fragmented information, interpretive ambiguity and over-reliance on expert experts. While general-purpose AI tools such as ChatGPT are readily available, they have hallucinations (create unfamiliar code by using no documentation) and the domain-specific advice practitioners need. This thesis explores AI-enhanced code consultation through a mixed-methods case study in Chinese regulation. Experiments conducted by four large-scale architecture firms revealed the following key issues: senior expert bottlenecks, the gap between regulatory text and practical application, multi-domain consultation complexity, and the verification requirements for professional accountability. These findings reveal four hierarchical requirements: reliable source attribution, embedded expert interpretation, context-aware multi-area support, and workflow integration.