Organizing for Autonomy: How Agentic AI Reshapes AI Governance in Large Enterprises

Abstract

Large organizations are increasingly experimenting with and deploying new AI systems with Agentic capabilities or fully agentic systems. These systems consist of autonomous capabilities which decide and act within organizational workflows. Whilst this development entails, traditional IT and AI governance frameworks are founded on the assumption that these systems are stable, the frameworks consider decision rights, compliance, oversight, and accountability at the level of individual systems. This creates a gap between the characteristics and capabilities of agentic AI systems and current governance approaches. Where non-human identities, delegated autonomy, runtime action, and platform-based orchestration become central governance concerns. This thesis examines the research question: “How does agentic AI reshape AI governance within large organizations, and what governance adaptations does this require?.” A qualitative research approach has been adopted which is based on a literature review, semi-structured expert interviews, and case studies. Eight interviews were conducted to gather empirical data among experts with roles in strategy, responsible AI, technical architecture, platform development, and business management. The scope of this thesis is to large organizations and focuses on governance challenges, risks and requirements linked to the deployment of agentic AI. The findings indicated the reforming of AI governance through agentic AI by distancing itself from ex-ante approval of limited systems and shifting towards a continuous governance approach. The agentic systems operate through delegation, tool use, and platform-controlled execution. The emergent governance domains in this thesis are non-human identity governance, delegated autonomy governance, runtime action governance, forensic auditability, and platform-based governance. This thesis further proposes layered architecture and continuous approach of agentic governance. Additionally, this thesis proposes the theoretical contribution of the Agentic Governance Control Plane (AGCP).