From Agency Costs to Delegation Costs: Revisiting Agency Theory in the Age of Agentic AI
At its core, Agency Theory examines the relationship between a principal and an agent. Shareholders delegate authority to executives, boards delegate authority to management, and clients delegate authority to advisors. The challenge arises because the principal cannot perfectly observe what the agent knows, what actions are being taken, or whether those actions remain aligned with the principal's interests. Agency Theory provided a framework for understanding these tensions through concepts such as information asymmetry, incentive alignment, monitoring mechanisms, and agency costs.
Consider a seemingly straightforward business scenario. A senior executive asks an AI system to evaluate a potential market expansion opportunity. A supervising agent decomposes the request into multiple tasks. Research agents gather market intelligence. Analytical agents assess competitive dynamics and forecast outcomes. Planning agents formulate recommendations. External tools provide additional data and validation. Eventually, a recommendation is presented to the executive. While the workflow appears efficient, something important has changed. The executive's original intent has passed through multiple layers of interpretation, decomposition, summarization, and optimization before reaching its final form.
The structure of delegation has evolved. Traditional Agency Theory largely focused on a principal-agent relationship. Agentic ecosystems increasingly resemble principal-agent-network relationships. Authority is no longer delegated to a single actor. Instead, it is distributed across interconnected agents, tools, workflows, and decision pathways. The fundamental governance question remains the same as it was in 1989: how do we ensure that delegated actions remain aligned with original intent? What has changed are the mechanisms through which misalignment can occur.
In traditional organizational settings, misalignment often resulted from divergent incentives. Managers may pursue objectives that differ from those of shareholders. Advisors may possess information unavailable to clients. Departments may optimize local performance at the expense of enterprise outcomes. In agentic systems, the challenge is different. AI agents do not possess personal ambition, career aspirations, or financial incentives. However, they do operate within objectives, prompts, constraints, evaluation metrics, and feedback mechanisms. These elements effectively become the operating logic that guides their behavior. While the source of misalignment differs, the governance challenge remains surprisingly familiar.
One of the most important insights from Agency Theory is the concept of information asymmetry. Agents often possess information unavailable to principals, creating uncertainty and governance risk. In agentic systems, a related phenomenon may emerge that can be described as context asymmetry. As information moves through chains of agents, it is repeatedly compressed, summarized, translated, and reformulated. Important nuances may be lost. Assumptions may become implicit. Context that was obvious at one stage may disappear entirely at another. The issue is no longer simply hidden information. It is the gradual transformation of context itself as work flows through multiple layers of delegation.
This observation suggests that organizations may need a complementary concept to Agency Costs. Agency Costs describe the resources required to align interests and monitor delegated authority. Agentic ecosystems may introduce what could be called Delegation Costs. These costs do not arise because agents have conflicting motivations. Instead, they emerge because every layer of delegation introduces opportunities for divergence from the original intent.
The first of these costs is Intent Translation Cost. Human objectives must be translated into instructions, prompts, goals, constraints, and policies. Every translation creates opportunities for ambiguity. What appears clear to a human decision maker may be interpreted differently by downstream agents. The second is Context Loss Cost. As tasks are passed between agents and tools, information is compressed and filtered. Critical details may be omitted, even when each individual step appears reasonable. The third is Coordination Cost. Multiple agents must synchronize assumptions, outputs, and dependencies. Success at an individual agent level does not guarantee success at a system level. The fourth is Objective Drift Cost. Agents may faithfully pursue assigned objectives while gradually moving away from broader organizational intent. Finally, Verification Cost emerges because humans must increasingly validate outputs generated through complex chains of delegated activity whose internal processes may not be fully visible.
Taken together, these costs form a governance burden that may become increasingly significant as organizations adopt agentic architectures. This does not imply that agentic systems are inherently problematic. Rather, it suggests that governance frameworks designed for single decision makers may be insufficient when decisions emerge from networks of interacting agents.
This perspective also highlights a potential gap in contemporary AI governance discussions. Much of the current focus is rightly directed toward transparency, explainability, fairness, privacy, accountability, and risk management. However, agentic systems may require an additional governance lens centered on delegation itself. Organizations may need mechanisms that preserve intent across delegation layers, maintain context integrity, detect objective drift, coordinate agent interactions, and ensure meaningful verification. In other words, governance may need to focus not only on what individual models do, but also on how decisions propagate through networks of delegated actors.
Perhaps the most intriguing implication concerns accountability. As agentic systems evolve, decision making becomes increasingly distributed. Recommendations and actions may emerge from the combined contributions of multiple agents, tools, and data sources. Yet accountability remains singular. Boards, regulators, customers, and stakeholders will still seek a responsible party when outcomes are questioned. Distributed execution does not eliminate accountability. It merely makes accountability harder to trace. This creates a governance tension that may define much of the future conversation around agentic AI.
The enduring relevance of Eisenhardt's work lies not in its specific assumptions about human actors, but in its broader examination of delegated authority under conditions of uncertainty and incomplete information. The technologies have changed dramatically since 1989. The governance questions have not. Who is acting on whose behalf? How do we know their actions remain aligned with original intent? Who is accountable when they are not? These questions sat at the heart of Agency Theory more than three decades ago. They may prove equally important as organizations navigate the emerging realities of agentic AI ecosystems.
This article is inspired by Kathleen M. Eisenhardt's seminal paper, "Agency Theory: An Assessment and Review" (Academy of Management Review, 1989). The ideas presented here explore whether some of the theory's foundational insights may provide a useful lens for understanding governance challenges in emerging agentic AI ecosystems.
Comments