Atlas Support
ResearchEconomicsWritten and edited by:

The Principal–AI Agent Problem: Governance in the Age of AI Agents and Principal–AI Agent Theory

Atlas Support’s Economic Research team introduces a paper that uses agency theory to explain how an AI agent’s actions can diverge from the intentions of the person delegating work.

Where divergence comes from

Even if the AI pursues no private interest, a principal can struggle to communicate objectives and constraints, observe actions, or intervene in time. The paper calls the resulting divergence “AI agency slack” and examines the costs of preventing, detecting, and correcting it, along with the losses that remain.

Designing oversight after delegation

The discussion centers on shareholders delegating management decisions to AI and extends to workplaces, public administration, professional services, and asset management. It considers how instructions, monitoring, approval, and stop authority can balance the gains from delegation with the cost of oversight.