High-quality data is paramount for developing AI. The process of building this data often presents many unexpected challenges. This presentation shares our comprehensive experience in overcoming the numerous challenges encountered during the development of AI for tongue diagnosis. This approach goes beyond simply training labeled data to create and evaluate AI algorithms. It also aims to help researchers understand the characteristics of the fields in which AI is being utilized, develop data collection methodologies tailored to the specific characteristics of those fields, and understand the process of solving these problems.