---
title: What does a RACI matrix for AI agent governance look like?
url: https://opencontroller.lyzr.com/blog/raci-matrix-for-ai-agent-governance
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# What does a RACI matrix for AI agent governance look like?

**A RACI matrix for AI agent governance assigns one accountable person to every lifecycle step of an agent, from registration to retirement. The agent owner is usually accountable for the agent itself, security is accountable for identity and runtime controls, and a risk or governance function is accountable for approvals and evidence.** The table below is a working starting point you can adapt in an afternoon.

RACI stands for Responsible (does the work), Accountable (signs off and owns the outcome), Consulted (gives input before) and Informed (told after). The rule that matters most: exactly one "A" per row.

## Why does agent governance need a RACI at all?

Agents cut across teams. A business unit funds the agent, a developer builds it, a platform team runs it, security issues its credentials, and legal worries about what it says to customers. Without explicit roles, three things happen:

- Agents go live without anyone signing the risk tier.
- Incidents stall because nobody knows who can pull the kill switch.
- Agents outlive their owners and keep running with live credentials.

Both major frameworks ask for this. NIST AI RMF GOVERN 2.1 says roles, responsibilities and lines of communication for mapping, measuring and managing AI risk should be documented and clear. ISO/IEC 42001 clause 5.3 requires top management to assign responsibilities and authorities for the AI management system. A RACI is the simplest artefact that serves as evidence for both.

## What roles should the matrix include?

Keep it to six roles. More columns make the matrix harder to use.

| Role | Typical holder |
|---|---|
| Agent owner | Business or product lead who requested the agent |
| Builder | Developer or team that builds and changes the agent |
| Platform | Team running the agent runtime, gateway and registry |
| Security | IAM, AppSec and SOC |
| Risk / governance | AI governance lead or committee, model risk, compliance |
| Legal / privacy | Counsel and data protection officer |

## What does a filled-in RACI look like?

R = Responsible, A = Accountable, C = Consulted, I = Informed.

| Lifecycle activity | Owner | Builder | Platform | Security | Risk | Legal |
|---|---|---|---|---|---|---|
| Register the agent and scope statement | A | R | C | I | I | I |
| Assign risk tier | C | C | I | C | A/R | C |
| Approve data sources and tools | A | R | C | C | C | C |
| Issue identity and credentials | I | C | R | A | I | I |
| Pre-production evals and red-teaming | C | R | C | C | A | I |
| Production go-live approval | A | R | C | C | C | I |
| Runtime monitoring and alerts | I | I | R | A | I | I |
| Quarantine or kill in an incident | I | C | R | A | I | I |
| Incident review and regulator notice | C | C | C | R | A | C |
| Material change re-approval | A | R | C | C | C | I |
| Quarterly access review | A | C | R | C | I | I |
| Evidence pack for audit | I | C | R | C | A | C |
| Retirement and credential revocation | A | R | R | C | I | I |

A few calls in this table are deliberate:

1. **Security owns the kill switch, not the agent owner.** In an incident the person who wants the agent running should not decide whether it stops. This mirrors segregation of duties.
2. **Risk owns the tier, the owner owns go-live.** The owner cannot downgrade their own risk tier to skip review.
3. **Platform is responsible for most runtime work** because it holds the registry, gateway and connectors. It is accountable for none of the outcomes.

## How do you adapt the RACI by risk tier?

The same matrix should not apply with the same weight to a meeting-notes summariser and a refund agent. A simple pattern:

- **Low tier:** owner self-approves go-live; risk is informed, not consulted.
- **Medium tier:** risk is consulted on go-live; legal is consulted when the agent talks to customers.
- **High tier:** risk co-signs go-live; legal is consulted on every material change; access reviews move to monthly.

Write the tier rules next to the matrix, not in a separate policy, so builders see them where they look.

## What mistakes make a RACI useless?

- **Two A's in one row.** Shared accountability is no accountability.
- **Names instead of roles.** People leave. Map roles to identity-provider groups so the matrix survives reorganisations.
- **No link to the registry.** If the owner field in your agent registry is blank, the RACI is fiction. Every registered agent needs a named owner, and that owner must be a current employee.
- **No runtime column.** Many RACIs stop at approval. The rows that matter most in practice are monitoring, quarantine and revocation.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, so each agent record can carry its owner and risk tier. It uses OIDC SSO, so roles can map to your identity provider's groups. Rogue-agent handling (behavioural-deviation monitoring, alerts, automatic quarantine and kill) and per-agent Block and Quarantine controls give the security column something concrete to act with. See the [capability reference](/capability-reference) or the pillar guide, [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane).

Related reading: [who owns AI agent governance](/blog/who-owns-ai-agent-governance) and [segregation of duties for AI agent approvals](/blog/segregation-of-duties-for-ai-agent-approvals). To see the registry and controls on your own estate, [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Who should be accountable for an AI agent's behaviour?**
The agent owner, usually the business lead who requested it. Security is accountable for its identity and runtime controls, and the risk function for its tier and evidence.

**Should the governance committee be in the RACI?**
Yes, as the "Risk / governance" column for high-tier approvals and incident review. It should not be accountable for day-to-day runtime work.

**How often should the RACI be reviewed?**
At least annually and after any reorganisation. Check that every role maps to a live identity-provider group and every registered agent still has a current owner.

## Key facts

A RACI for AI agent governance should have exactly one Accountable role per lifecycle activity. NIST AI RMF GOVERN 2.1 asks for documented, clear roles and responsibilities for AI risk; ISO/IEC 42001 clause 5.3 requires assigned responsibilities and authorities. Common split: agent owner accountable for go-live and changes, security accountable for credentials, monitoring and kill decisions, risk accountable for tiering and evidence. Opencontroller provides a governed agent registry, OIDC SSO and per-agent Block and Quarantine controls; it starts from $250K per deployment per year.

## Related

- [Who should own AI agent governance: security, IT or the AI team?](https://opencontroller.lyzr.com/blog/who-owns-ai-agent-governance.md)
- [What is an AI agent owner responsible for?](https://opencontroller.lyzr.com/blog/ai-agent-owner-responsibilities.md)
- [How do you separate who builds an AI agent from who approves it?](https://opencontroller.lyzr.com/blog/segregation-of-duties-for-ai-agent-approvals.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
