Some of the main processes involved in the design and development of AI-mediated speaking assessments using ChatGPT, from tasks to rating, are outlined. First, the underlying GPT (Generative Pre-trained Transformer), a large language model, and its affordances for developing speaking assessments are discussed. From this discussion, some general principles in effectively employing this new tool to design and implement speaking assessments are derived, and three different methods of customizing ChatGPT to develop speaking tests are presented. Finally, some potential pitfalls and the technical limitations of the system are considered.