Generative AI Agents and Coordination at Work

Miros Zohrehvand, Anil R. Doshi, and Piyush Gulati

Working paper

Advances in large language models make it feasible to delegate managerial work to AI, yet research on generative AI in organizations has focused overwhelmingly on individual task productivity. We ask when delegating a bounded managerial role to AI improves team performance. In a pre-registered, role-differentiated online experiment, three-person teams (a Captain who can only manage through communication, plus two specialized workers) play four rounds of a space-mining game; teams are randomly assigned a human or an AI Captain, and the rounds vary information demand and task interdependence. We theorize the delegated role as information-to-action conversion: turning dispersed, incomplete information into timely directives and resource choices. AI-managed teams roughly double the performance of human-managed teams, and the advantage widens under high information demand, consistent with our first hypothesis. Process evidence shows AI managers issuing more directive plans, committing fewer coordination errors of every diagnostic class, and being experienced by workers as clearer and more timely. The advantage narrows where actions must be correctly sequenced (the contingency our second, pre-registered hypothesis targets), though the present sample is underpowered to resolve that interaction. The findings position the manager's architecture as an organization design variable and identify scope conditions for delegating managerial functions to AI.