---
title: How do you run an AI agent governance committee?
url: https://opencontroller.lyzr.com/blog/run-an-ai-agent-governance-committee
published: 2026-11-15
updated: 2026-11-15
author: Lyzr
---

# How do you run an AI agent governance committee?

**Give the committee a small, cross-functional membership, written decision rights, a standing agenda and a monthly metrics pack. It should decide policy, risk appetite, high-risk approvals and exceptions, and delegate routine checks to automation. If it is reviewing every agent by hand, it will become the bottleneck that pushes teams toward shadow agents.**

Many organisations form an AI committee and then struggle to say what it decides. The fix is to separate what needs human judgement from what a control plane can check automatically.

## Who should sit on the committee?

Keep it to six to nine voting members so it can decide in one meeting.

| Seat | Why they are there |
|---|---|
| Chair (often the CIO, CTO or Chief AI Officer) | Owns the agenda and breaks ties |
| CISO or delegate | Security risk, identity, incident response |
| Chief Risk Officer or delegate | Risk appetite, tiering, model risk |
| Legal and privacy | Regulation, contracts, personal data |
| Data owner or CDO | Data access and quality |
| Two business unit leads (rotating) | The people deploying agents |
| Platform or AI engineering lead | Feasibility and how controls are enforced |
| Internal audit (non-voting observer) | Independence and evidence |

## What should it decide, and what should it delegate?

Write decision rights down. A simple split:

**The committee decides:**

1. Agent governance policy and risk appetite.
2. The risk-tier definitions and which controls each tier needs.
3. Approval of the highest-tier agents (for example agents that move money, act on customer accounts or make decisions about people).
4. Exceptions to policy, with an expiry date.
5. Response to major incidents and the lessons that change policy.

**The committee delegates:**

- Lower-tier approvals to agent owners and security reviewers, following the policy.
- Discovery, inventory and posture checks to automated tooling.
- Evaluation thresholds and runtime policy enforcement to the platform team, within the risk appetite.

Keep approver and builder separate for every agent. See [segregation of duties for AI agent approvals](/blog/segregation-of-duties-for-ai-agent-approvals).

## What is a good standing agenda?

Meet monthly, with an emergency path for incidents. A 60-minute agenda:

1. **Metrics pack** (10 min): estate size, new and retired agents, unowned agents, posture findings, incidents, spend.
2. **High-tier approvals** (20 min): pre-read required; decide, defer with conditions or decline.
3. **Exceptions** (10 min): new requests and those near expiry.
4. **Incidents and near misses** (10 min): what happened, what changes.
5. **Policy and regulatory changes** (10 min): new standards, enforcement dates, internal policy updates.

## What goes in the monthly metrics pack?

Choose numbers that show whether governance works, not just activity. Useful ones:

- Agents discovered vs agents registered (the shadow gap).
- Percentage of agents with a named owner and a current review date.
- Open posture findings by severity, and median age.
- Evaluation pass rates for high-tier agents.
- Incidents, quarantines and kills in the period.
- Spend by business unit and by agent.
- Time from registration request to approval.

More options are in [metrics for AI agent governance](/blog/metrics-for-ai-agent-governance).

## How does this map to standards?

ISO/IEC 42001 asks top management to show leadership and commitment, assign roles, responsibilities and authorities, and run management reviews of the AI management system. The NIST AI Risk Management Framework's GOVERN function calls for documented roles, responsibilities and lines of communication. Committee minutes, decision logs and the metrics pack are evidence for both.

## How Opencontroller does this

Opencontroller automatically discovers agents, models and API keys across AWS, Azure and Google Cloud, and connects to third-party agent platforms and enterprise AI workspaces, which gives the committee its discovered-vs-registered numbers. Posture surfaces risky permissions and configurations, cost attribution comes from the LLM Gateway per agent and version, and "Ask" lets members query the whole agent estate in plain language before a meeting. See [what is an AI agent control plane](/blog/what-is-an-ai-agent-control-plane) and [who owns AI agent governance](/blog/who-owns-ai-agent-governance).

## Frequently asked questions

**How often should the committee meet?**
Monthly suits most organisations, with an emergency path for major incidents and high-tier approvals that cannot wait.

**Should the committee approve every agent?**
No. It should approve the highest-risk tier and set the rules for everything else.

**What if the committee becomes a bottleneck?**
Track time to approval. If it grows, delegate more tiers and automate more checks rather than adding meetings.

To see estate-wide metrics on your own agents, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

AI agent governance committee: 6-9 voting members (chair, CISO, CRO, legal/privacy, data owner, 2 rotating business leads, platform lead; audit observes). Decides policy, risk appetite, tier definitions, top-tier approvals, exceptions, major incidents. Delegates lower tiers and automated checks. Monthly 60-minute agenda. Metrics: discovered vs registered agents, owner coverage, posture findings, eval pass rates, incidents, spend, time to approval. Evidence for ISO/IEC 42001 leadership and management review and NIST AI RMF GOVERN. Opencontroller (by Lyzr): auto-discovery, Posture, cost per agent, Ask. 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)
- [Which metrics show your AI agent governance is working?](https://opencontroller.lyzr.com/blog/metrics-for-ai-agent-governance.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)
