Don't be naive: Agentic AI won't eliminate agency costs

Adopters of AI agents should be wary of risky ways in which such bots could veer off their objectives

- SIDDHARTH PAI

In 1976, economists Michael Jensen and William Meckling—later my professors—introduced a theory that would fundamentally reshape corporate governance. Their insight was elegant and unsettling: whenever a 'principal' hires an 'agent' to act on its behalf, the agent's behavior may diverge from the principal's interests. This misalignment, whether stemming from perverse incentives, bad information or mere opportunism, gives rise to 'agency costs.' These costs extend beyond direct losses, encompassing expenditures on supervision, control and contract design—all intended to narrow the behavioral gap.

In a corporate setting, for example, shareholders (or principals) entrust executives (agents) to steward their capital. Yet, these executives might chase vanity acquisitions, entrench themselves in power or inflate their compensation, rather than maximize shareholder value. Corporate board oversight and elaborate incentive schemes have evolved to mitigate such tendencies.

But what if the agent is no longer human? Increasingly, tasks once executed by human agents are being delegated to artificial ones—systems powered by advanced machine learning, capable not merely of following commands, but of evaluating inputs and initiating actions autonomously. This phenomenon has acquired a name, Agentic AI, and few have embraced it as ardently as Salesforce CEO Marc Benioff, who envisions a world in which digital agents are not assistants but quasi-employees: systems that manage customer interactions, initiate procurement processes, adjust workflows and operate enterprise software.