Why AI Agents Cannot Govern Themselves: A Representation-Based Explanation of Enterprise Agent Failure

Artificial intelligence is rapidly evolving from content generation toward autonomous and semi-autonomous agents capable of retrieving information, invoking tools, coordinating workflows, and executing operational tasks. As enterprises deploy AI agents into customer service, operations, software engineering, cybersecurity, finance, and governance processes, a fundamental question emerges: can AI agents govern themselves? This preprint argues that enterprise agent failure is fundamentally a representation problem rather than solely a model, alignment, or autonomy problem. AI agents do not operate directly on organizational reality; they operate on machine-legible representations of customers, processes, policies, identities, authorities, risks, and contexts. When those representations are incomplete, stale, fragmented, contested, or institutionally illegitimate, agents may appear competent while acting on a distorted version of reality. The paper introduces a representation-based explanation of enterprise agent failure by integrating three complementary frameworks: • The Representation Economy – a framework explaining how value, trust, governance, and coordination increasingly depend on machine-legible representations of reality. • SENSE–CORE–DRIVER – an enterprise AI architecture that separates perception (SENSE), cognition (CORE), and governed execution (DRIVER). • Digital Anthropology for Enterprise AI – an approach for understanding how human reality, tacit knowledge, exceptions, and contextual work practices become machine-legible. The paper develops a taxonomy of enterprise agent failure, examines why agents cannot independently determine representational adequacy, legitimate their own authority, create meaningful recourse, or distinguish formal process from lived organizational reality, and proposes practical governance principles for enterprise deployment. The central argument is that the future of enterprise AI governance will depend less on creating self-governing agents and more on designing institutions in which agents operate within the limits of represented reality, delegated authority, runtime verification, and recoverable execution. This work extends and applies the Representation Economy, SENSE–CORE–DRIVER, and Digital Anthropology research program to the emerging domain of enterprise AI agent governance.