A governance framework for building trustworthy agentic AI for public sector and regulated organizations

This article presents a governance framework for deploying trustworthy agentic AI systems in public sector and regulated organizations.

• Agentic AI systems differ from traditional AI through autonomous decision-making and multi-step workflow capabilities
• Four-scope classification model: No agency (read-only), Prescribed (human approval required), Supervised (autonomous within boundaries), Full agency (continuous autonomy)
• Six security dimensions: Identity context, Data/memory/state protection, Audit/logging, Agent/FM controls, Agency boundaries/policies, Orchestration
• Memory integrity requires expiration policies, read-only enforcement for lower scopes, session isolation, and integrity validation
• Comprehensive logging must capture both API-level actions and decision-context reasoning for full auditability
• Amazon Bedrock is FedRAMP High authorized in GovCloud and FedRAMP Moderate in commercial regions
• Framework aligns with ISO/IEC 42001:2023 compliance standard across all six security dimensions

Organizations should classify existing AI systems by scope, evaluate controls against security dimensions, and build governance incrementally starting with lower-autonomy systems before deploying higher-scope agents.