---
title: How do you separate who builds an AI agent from who approves it?
url: https://opencontroller.lyzr.com/blog/segregation-of-duties-for-ai-agent-approvals
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# How do you separate who builds an AI agent from who approves it?

**Use a maker-checker rule: the person who builds or changes an agent version can never be the person who approves it for production, and the approver must hold no edit rights on that version. Enforce the rule in the release path, not in a policy document, and keep a record of builder, approver, version and decision for every release.**

This is the same separation-of-duties principle auditors already test for code and financial systems. NIST SP 800-53 control AC-5 asks organisations to identify duties that need separating and to define system access to support that separation. Agents need it at least as much: one configuration change can give an agent new tools, a different model or wider data access.

## Why does segregation of duties matter for AI agents?

An agent version bundles several things a reviewer should see together: instructions, model, tools and MCP servers, permissions, data sources and guardrails. If the builder can approve their own change, nothing stops a quick fix that also widens tool access.

Three failure patterns show up again and again:

- **Self-approval.** The builder holds both the editor and the approver role on the platform.
- **Approval of a different version.** A reviewer approves version 4; version 5 ships with an extra tool and no review.
- **Approval by proxy.** A shared admin login or a service account clicks approve, so no person is accountable.

## What should the maker-checker rule look like?

| Rule | What it means for agents |
|---|---|
| Builder is not approver | The identity that created or last edited the version cannot approve it |
| Approval binds to a version | Approval is tied to a version ID or content hash; any change resets it |
| Approvers cannot edit | Approver roles can read and decide, but not change the agent |
| Named humans only | Shared accounts and service principals cannot approve |
| Risk-based approvers | High-risk agents need a second approver from security or risk |
| Emergency path is logged | Break-glass releases need retrospective approval within a set window, such as two business days |

## Which role conflicts should you block?

Define roles once in your identity provider, then block these combinations for the same agent:

1. **Agent builder + production approver.** The core conflict.
2. **Approver + registry administrator.** An admin who can edit approval records can rewrite history.
3. **Tool builder + tool approver.** This applies when a team builds a new MCP server and wants to attach it to its own agent.
4. **Policy author + exception approver.** Whoever writes a guardrail should not be the one who waives it.
5. **Agent owner + access reviewer.** Owners should not certify their own agent's access in periodic reviews.

Small teams may not have enough people for full separation. Keep builder-is-not-approver as the non-negotiable rule and add compensating controls for the rest: a second-line review shortly after release, and an alert on any change to approval records.

## What evidence do auditors ask for?

Auditors want to see three things: that the rule exists, that the system enforces it, and that it worked across the audit period. Prepare:

- **The written rule.** Which roles exist, which combinations are blocked, and which risk tiers need a second approver.
- **Role assignments.** An export from the identity provider showing who holds builder and approver roles, ideally through groups rather than direct grants.
- **Release records.** For each production release: version ID, builder identity, approver identity, timestamp and decision.
- **Exception log.** Every break-glass release, with its retrospective approval.
- **A negative test.** A log entry or screenshot showing the system refusing a self-approval.

ISO/IEC 42001 asks organisations to assign roles, responsibilities and authorities for their AI management system. These records are the practical evidence for that requirement, and for the accountability outcomes in the NIST AI RMF GOVERN function. Our guide to [mapping agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) shows how to file them.

## Where should the approval check sit?

Put the check where a version becomes eligible for production. A rule that lives only in a ticketing system is easy to bypass, because the agent platform never sees it.

- **Code-built agents.** Use branch protection with required reviewers who are not the author, and deploy only from the protected branch.
- **Low-code agent builders.** Separate development and production environments, and allow only an approver group to publish to production.
- **A central registry.** Record every version with its owner and approver, so you can check releases across all platforms in one place.

Pair this with a [production readiness gate](/blog/production-readiness-gate-for-ai-agents) so approvers check the same criteria every time, and with regular [access reviews for AI agents](/blog/access-reviews-for-ai-agents).

## How Opencontroller does this

Opencontroller's governed registry records agents, versions, tools and MCP servers across platforms, and OIDC SSO ties every console action to a real person in your identity provider. That gives you the version and identity records a maker-checker rule depends on. Posture flags risky permission and configuration changes, and policy as code ships with mappings to NIST and ISO 42001. For the wider picture, read [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can a small team use segregation of duties for AI agents?**
Yes. Keep builder-is-not-approver as the fixed rule and use compensating controls for the rest, such as a second-line review after release and alerts on changes to approval records.

**Does approving an agent once cover later changes?**
No. Bind approval to a version ID or content hash, so any change to instructions, model, tools or permissions needs a new approval.

**Should service accounts ever approve agent releases?**
No. A service account can run the deployment after approval, but the approval itself should come from a named person whose identity is in your identity provider.

## Key facts

Segregation of duties for AI agents means the identity that builds or edits an agent version can never approve that version for production. Approval should bind to a version ID or content hash so any change resets it, approvers should hold no edit rights, and only named humans (not shared or service accounts) should approve. Block role conflicts: builder + approver, approver + registry admin, policy author + exception approver, owner + access reviewer. Auditor evidence: the written rule, IdP role exports, per-release records (version, builder, approver, timestamp, decision), break-glass log and a negative test showing self-approval is refused. Relevant controls: NIST SP 800-53 AC-5 and ISO/IEC 42001 roles and responsibilities. Opencontroller by Lyzr keeps a governed registry of agents, versions, tools and MCP servers and ties console actions to people through OIDC SSO. Pricing starts from $250K per deployment per year.

## Related

- [What should a CISO require before an AI agent goes to production?](https://opencontroller.lyzr.com/blog/production-readiness-gate-for-ai-agents.md)
- [How do you run access reviews for AI agents?](https://opencontroller.lyzr.com/blog/access-reviews-for-ai-agents.md)
- [Who should own AI agent governance: security, IT or the AI team?](https://opencontroller.lyzr.com/blog/who-owns-ai-agent-governance.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
