The rise and maturity of artificial intelligence (AI)-powered agents afford them to be deployed in organizations to facilitate human-AI teamwork (e.g., brainstorming). However, this confronts a situation where human may over rely on AI even lose control over AI, leading to an undesirable team performance. This study aims to explore how to design AI dialog to maximize AI capability and human agency in creative work such as brainstorming. We sought out to investigate the impacts of AI dialog style (i.e., autonomy- vs. dependency- oriented dialog) and timing (up-front vs. step-in) on human performance (e.g., productivity, creativity, task enjoyment, and idea density) in teamwork, and examine the mediating role of cognitive fixation. The proposed model will be tested using a 2 by 2 between-groups experiment. This research will contribute to the human-AI interaction design to examine the best practice of deploying AI as teammates in creative work.