The Principal-A(I)gency Problem: What happens when agentic AI powers the agent?

Sammanfattning

The Principal-Agent problem has been a cornerstone of New Institutional Economics school of thought for the past half-century, providing a comprehensive framework for academics and practitioners alike in how to bridge the often-present misalignment of interests between a delegator of a task and the delegatee, whose interests are frequently incongruent with the former. However, the advent of Artificial Intelligence in the enterprise world raises the question of whether this framework can sufficiently accommodate a non-human agent. This study explores what implications an agentic AI system has for the Principal-Agent relationship, and whether the framework requires revision in the context of autonomous AI agents. It examines this in the empirical context of the Swedish Electricity Industry, an environment chosen for its widespread importance in society, and its delicate market structure, where physical electricity demand always must equal supply. Drawing on 21 semi-structured interviews with traders, consultants, executives, and institutional representatives across the Swedish electricity ecosystem, the study's results show that practitioners still consider near- or full agentic AI as non-implementable in today's environment, but a likely outcome in the long-term. W e propose a two-level principal-AI-agent framework, building upon conventional agency theory, in which regulators act as a second principal representing the societal interest, and argue that corrigibility - the agent's structural openness to correction post error - as well as trust, will be prerequisite conditions for a successful agentification of the electricity markets.