AI Governance RACI Matrix: Who Owns What

As enterprises move AI agents from pilot to production, governance decisions accumulate faster than informal ownership can track. Who approves a new tool-call permission? Who is accountable when an agent takes an unauthorized action? Who signs off on a runtime policy change versus who simply executes it? In most organizations these questions are answered informally, if at all, with responsibility spread unevenly across security, platform engineering, compliance, and business unit leadership.

A RACI matrix gives each AI governance function a documented owner. Responsible identifies who performs the work. Accountable identifies who is answerable for the outcome and should be a single role per function, not a committee. Consulted covers parties whose input is required before a decision is made. Informed covers parties who need to know after a decision has been made. NIST's AI RMF Playbook recommends mapping specific internal roles, including risk owners, technical staff, and legal or compliance functions, to each risk management function precisely to reduce this kind of diffusion.

Illustrative AI Governance RACI Matrix

The following framework is illustrative rather than prescriptive. It shows one reasonable way to apply the four RACI roles to the core functions of the AI agent governance lifecycle; the specific assignment for any organization should reflect its own structure, risk tolerance, and regulatory exposure.

The Four Roles Defined

Function-by-Function Assignment

Where Runtime Enforcement Intersects with Governance Accountability

Policy definition and runtime enforcement are frequently owned by different teams, and this split is often where accountability breaks down. Policy definition determines what an agent may do in principle, informed by risk tolerance and regulatory obligation. Runtime enforcement is the technical mechanism, blocking or allowing an agent's tool calls and API access, that determines what an agent may actually do in practice. When these two functions are not explicitly linked through RACI assignment, enforcement can drift from policy intent, either because engineering implements controls faster than governance can review them, or because policy changes are never reflected in runtime configuration.

OWASP's Top 10 for LLM Applications identifies excessive agency and insecure plugin design as risk categories tied directly to under-governed tool-call permissions and autonomous action scope. Agent identity and permissions management raises a related but distinct set of questions. It functions similarly to identity and access management for human or service accounts, but requires dynamic, context-dependent scoping that most legacy IAM models were not built to handle. Tool-call authorization sits at the intersection of security and engineering and often warrants a shared Responsible role between the two, with a single Accountable owner designated to prevent gaps.

Principles for Assigning AI Governance RACI Roles

• Assign exactly one Accountable role per governance function to prevent diffusion of ownership.
• Separate Responsible, who executes the work, from Accountable, who answers for the outcome.
• Position legal and compliance as Consulted for regulatory interpretation, not as Responsible for technical enforcement.
• Give platform and security engineering Responsible status for runtime enforcement and tool-call authorization, without defaulting them to Accountable status absent governance sign-off.
• Designate incident response accountability separately from the technical team executing containment.
• Revisit RACI assignments whenever agent capabilities or permission scopes change.

Standards Provide Reference Points, Not a Ready-Made Matrix

No public framework provides a plug-and-play RACI matrix for AI agent runtime governance. NIST's AI RMF organizes risk management into four functions, Govern, Map, Measure, and Manage, with Govern addressing organizational roles and policy oversight directly. NIST's Generative AI Profile, published as NIST AI 600-1, extends this with governance actions specific to generative AI, including human oversight and accountability during deployment. ISO/IEC 42001:2023, the first international AI management system standard, requires organizations to document roles and responsibilities for AI system development, deployment, and monitoring as a condition of certification.

The EU AI Act creates a legal distinction between providers, who develop AI systems, and deployers, who use them, with each carrying separate risk management and governance duties. Providers of high-risk systems must maintain technical documentation and logging to support post-market monitoring and audits. These frameworks establish governance principles, and in the EU AI Act's case, legal obligations, but the specific internal RACI assignment for runtime enforcement, tool-call authorization, and incident response remains an organization-specific decision informed by, rather than dictated by, these standards.