Last updated:
The governance of autonomous AI systems capable of executing independent actions (e.g., transactions, code deployment) distinct from Predictive AI (which provides insights) and Generative AI (which creates content).
Unlike standard AI governance, which focuses on model accuracy, bias, and content moderation, Agentic Governance focuses on permissions and operational bounds. It addresses the "active risk" of agents - systems that can interact with the real world. Key controls include "meaningful human control" mechanisms, automated kill-switches, and "permissioning" architectures that limit an agent’s authority (e.g., read-only access vs. write access) based on its risk level.
Real world example:
A financial services firm deploys an autonomous customer service agent authorized to process refunds. Agentic governance protocols restrict the agent to a maximum refund limit of $50 per transaction. If the agent attempts to process a $5,000 refund due to a hallucination or error, the governance layer (the "guardrail") automatically blocks the action and flags it for human review, preventing financial loss.
