This paper explores how different human-AI collaboration protocols (HAI-CPs) influence performance in individual human+AI ‘centaur’ settings as well as in ‘computer in the group’ settings. Establishing effective HAI-CPs is essential to advancing Human Work Interaction Design (HWID), especially as industries move from the question of AI adoption to the challenge of integrating AI sustainably and effectively into collaborative workflows. Based on an experiment with logic puzzles supported by a simulated GPT-like AI, this study assesses AI’s impact on decision-making, particularly regarding automation bias (the degree to which AI misleads users) and algorithmic aversion (the rejection of accurate AI recommendations). Our findings reveal that AI significantly enhances the cognitive performance of both centaurs and groups (‘Bigae’), acting as a performance leveler by improving outcomes for lower performers more than high performers. Centaurs and Bigae’s reported perceptions and observed error rates suggest that collective intelligence leads to less reliance on AI in group settings, relegating the machine to a more peripheral (adjunct) role in the decision-making process. While collectives are still improved by adding AI systems as additional average-performing teammates, deliberation with peers still appears to be the most powerful booster of human performance, and collective intelligence still outperforms individual intelligence, even when this is supported by AI.