---
title: What roles should an AI agent governance console have?
url: https://opencontroller.lyzr.com/blog/roles-inside-an-ai-governance-console
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What roles should an AI agent governance console have?

**Most enterprises need five roles in an AI agent governance console: platform admin, security reviewer, agent owner, finance viewer and read-only auditor. Map each role to an identity provider group, scope agent owners to their own agents, and make sure nobody can both approve and operate the same high-risk agent.**

The question usually comes up the day single sign-on goes live. Everyone can log in, and suddenly a developer can see every agent in the company, or a finance analyst can see prompts they should never read. Getting roles right early prevents that, and it produces clean evidence for auditors later. The identity side of agents themselves is covered in the [on-behalf-of delegation pillar](/blog/ai-agent-identity-on-behalf-of-delegation). This post is about the humans who use the console.

## Which roles do you need?

| Role | Typical holder | Can do | Should not do |
|---|---|---|---|
| Platform admin | AI platform or IT team | Configure connectors, SSO, gateway routes, global policies | Approve their own policy exceptions |
| Security reviewer | Security, risk or AI governance team | Review posture findings, approve agents and MCP servers, quarantine or block agents | Change agent code or gateway routes |
| Agent owner | Product or engineering lead per agent | See and edit their own agents, versions, evals and costs; request approval | See other teams' prompts and traces |
| Finance viewer | FinOps or finance business partner | See spend by agent, team and version; set budgets | See prompt or response content |
| Auditor (read-only) | Internal audit, second line, external auditor | Read inventory, policies, approvals and evidence exports | Change anything |

Some organisations add two more: a **business unit admin** who manages agents for one division, and an **incident responder** who can kill or quarantine any agent during an incident but cannot change policy.

## How should you scope what each role sees?

Role names are the easy part. Scope is where mistakes happen.

1. **Scope by ownership, not by role alone.** An agent owner role should apply to the agents that owner's group owns. Without that, "agent owner" quietly becomes "read everything".
2. **Separate metadata from content.** Seeing that an agent exists, its cost and its risk tier is low sensitivity. Seeing prompts, responses and tool outputs can expose personal or confidential data. Make content access a separate permission.
3. **Separate configuration from operation.** Changing a gateway route or a policy is configuration. Blocking an agent is operation. They need different approvals.
4. **Make break-glass explicit.** Give incident responders a time-limited elevated role, and log every use.

## What separation-of-duties rules should you enforce?

These rules map to the separation of duties control family that auditors already test (for example AC-5 in [NIST SP 800-53 Rev. 5](https://csrc.nist.gov/pubs/sp/800/53/r5/upd1/final)):

- The person who registers or changes a high-risk agent cannot approve it.
- The person who writes a policy exception cannot approve it.
- Platform admins cannot delete audit logs or evidence exports.
- Finance viewers cannot raise their own budget limits without a second approver.

More detail on approval chains is in [segregation of duties for AI agent approvals](/blog/segregation-of-duties-for-ai-agent-approvals).

## How do you map roles to your identity provider?

1. Create one identity provider group per role, plus one group per agent-owning team.
2. Send group membership to the console in the SSO token (group or role claims), or provision it with SCIM. The trade-offs are in [SCIM vs SSO group claims for roles](/blog/scim-vs-sso-group-claims-for-roles).
3. Keep the mapping table under change control: group name, console role, scope.
4. Review membership quarterly, and on every joiner, mover and leaver event.
5. Disable local accounts once SSO works, except one sealed break-glass account. See [SSO-only access for AI governance tools](/blog/sso-only-access-for-ai-governance-tools).

## Rollout checklist

- [ ] Five core roles defined, with a written "can do / cannot do" list.
- [ ] Agent owner role scoped to owned agents only.
- [ ] Prompt and response content behind a separate permission.
- [ ] Separation-of-duties rules configured and tested with two test users.
- [ ] Role-to-group mapping documented and approved.
- [ ] Test log-ins done as each role, including a user who should be denied.
- [ ] Quarterly access review scheduled, with evidence stored.

## How Opencontroller does this

Opencontroller supports OIDC single sign-on, so console access can run through your existing identity provider and its groups. The governed registry records each agent's owner, versions, tools and MCP servers, which gives you the ownership data needed to scope agent owners. Policies are written as code with an exception path, so who changed what is reviewable. During rollout, Lyzr engineers embed with your team to set up SSO and the role mapping. [Book a demo](https://www.lyzr.ai/book-demo/) to walk through a role design for your organisation.

## Frequently asked questions

**Should developers have access to the governance console at all?**
Yes, as agent owners scoped to their own agents. Developers who can see their agent's evals, costs and findings fix problems faster than ones who wait for a security ticket.

**Can one person hold the platform admin and security reviewer roles?**
In a small team it happens, but then require a second approver for high-risk agent approvals and policy exceptions, and record it as a known exception.

**Who should be able to kill an agent?**
Security reviewers and incident responders, plus the agent's owner for their own agent. Every kill or quarantine should be logged with who did it and why.

## Key facts

Recommended AI governance console roles: platform admin (configure connectors, SSO, routes, global policies), security reviewer (approve agents and MCP servers, quarantine/block), agent owner (own agents only), finance viewer (spend and budgets, no content), auditor read-only. Optional: business unit admin, incident responder with time-limited break-glass. Separation-of-duties rules: registrant cannot approve high-risk agent; exception author cannot approve exception; admins cannot delete audit logs; finance viewers need second approver to raise budgets. Maps to NIST SP 800-53 AC-5. Map roles to IdP groups via SSO claims or SCIM; quarterly access review. Opencontroller supports OIDC SSO, a governed registry with owners and versions, policy as code; Lyzr engineers embed for rollout. Pricing starts from $250K per deployment per year.

## Related

- [How do you map identity provider groups to admin and viewer roles in an AI governance tool?](https://opencontroller.lyzr.com/blog/map-idp-groups-to-roles-in-ai-governance-tools.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)
- [Should SCIM or SSO group claims assign roles in an AI governance console?](https://opencontroller.lyzr.com/blog/scim-vs-sso-group-claims-for-roles.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
