---
title: How do you run access reviews for AI agents?
url: https://opencontroller.lyzr.com/blog/access-reviews-for-ai-agents
published: 2026-11-12
updated: 2026-11-16
author: Lyzr
---

# How do you run access reviews for AI agents?

**Review every AI agent's identities, permissions, tools and data access at least quarterly, with its business owner and technical owner signing off. Show reviewers what the agent actually used in the period, not just what it was granted, and revoke anything unused. Keep the review record and remediation tickets as evidence.**

Most companies already review human access every quarter. Agents are usually missing from that process. They hold service accounts, API keys and OAuth grants, often broader than any employee's, and nobody is asked to confirm they still need them. The OWASP Non-Human Identities Top 10 (2025) ranks Improper Offboarding first (NHI1:2025) and lists Overprivileged NHI (NHI5:2025) and NHI Reuse (NHI9:2025) too. A regular access review is the routine control for all three.

## What should a reviewer see for each agent?

A review is only as good as the information on the screen. Give reviewers these fields.

| Field | Why it matters |
|---|---|
| Agent name, version and purpose | The reviewer must know what the agent is for to judge its access |
| Business owner and technical owner | Someone accountable must exist and still be employed |
| Identities and credentials | Service accounts, API keys, OAuth clients, federated identities |
| Permissions and scopes granted | The thing being reviewed |
| Permissions actually used in the period | Unused grants are the first to remove |
| Tools and MCP servers connected | Tools are access too; see [limiting which tools an agent can call](/blog/limit-which-tools-each-ai-agent-can-call) |
| Data sources and classifications reached | Sensitive data raises the bar for approval |
| Models used | Unapproved models are a finding |
| Users the agent acts for | Delegated access needs its own check |
| Last activity date | Idle agents are candidates for retirement |
| Open findings | Risky configuration flagged since the last review |

The "used versus granted" comparison does most of the work. Cloud providers expose last-used data for many permissions, and gateway logs show which tools and models were called.

## Who should sign off?

Split the decision by question.

1. **Business owner:** Is this agent still needed for its stated purpose?
2. **Technical owner:** Is each permission still required, and scoped as tightly as possible?
3. **Resource owner:** For sensitive systems, such as finance or HR data, does the system owner accept this agent's access?
4. **Security or IAM team:** Are findings resolved, and does the review meet policy?

No single person should approve all four, especially the developer who built the agent.

## When should access be revoked without waiting for a review?

Define triggers in policy so action does not depend on the next cycle.

- **No activity for 90 days.** Disable credentials, then retire the agent if nobody objects.
- **Owner leaves or changes role.** Reassign or disable within a set number of days. See [what happens to AI agents when the owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves).
- **A new version adds tools or scopes.** Treat it as a new access request.
- **A permission unused for a full review period.** Remove it.
- **A security finding,** such as a leaked key or a wildcard trust policy.

## How often should you review?

| Agent type | Suggested cycle |
|---|---|
| Writes to business systems, moves money, or reaches regulated data | Quarterly |
| Read-only on internal data | Every six months |
| Experimental, sandbox only | At promotion to production, then by its new tier |

Match whatever cycle your auditors already expect for privileged human accounts.

## What evidence should you keep?

Access reviews appear in most control frameworks, including NIST SP 800-53 account management (AC-2) and ISO/IEC 27001:2022 control 5.18 on access rights. Auditors usually ask for:

1. The list of agents in scope and how it was produced.
2. Each reviewer's decision, with date and identity.
3. Remediation tickets for every revoke decision, with closure dates.
4. Exceptions, with a named approver and expiry.

See [SOC 2 evidence for AI agents](/blog/soc-2-evidence-for-ai-agents) for how this fits a wider evidence pack.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud and keeps a governed registry of agents, versions, tools and MCP servers, so the review list starts complete. Posture detects risky permissions and configuration, and Ask lets reviewers query the whole agent estate in plain language. Read the [identity and delegation guide](/blog/ai-agent-identity-on-behalf-of-delegation), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**How often should AI agent access be reviewed?**
Quarterly for agents that write to business systems or reach sensitive data; every six months for read-only agents.

**Can I reuse my human access review process?**
The workflow, yes. But reviewers need agent-specific fields such as tools, models and delegated users.

**What if nobody claims an agent during the review?**
Disable its credentials, record the decision, and retire it if no owner appears within a set period.

## Key facts

An AI agent access review should show reviewers, for each agent: owner, purpose, identities and credentials, permissions and scopes, tools and MCP servers, data reached, models used, last activity and actual usage of each permission. Sign-off comes from the business owner (still needed?), the technical owner (scoped correctly?) and the resource owner for sensitive systems. Revoke triggers include no activity for 90 days, owner departure, unused permissions and unapproved tools. Keep the review record, decisions and remediation tickets as evidence; access review requirements appear in NIST SP 800-53 AC-2 and ISO/IEC 27001:2022 control 5.18. Opencontroller provides auto-discovery, a governed registry of agents, versions, tools and MCP servers, Posture for risky permissions and Ask for questions across the agent estate. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [What happens to AI agents when their owner leaves the company?](https://opencontroller.lyzr.com/blog/what-happens-to-ai-agents-when-the-owner-leaves.md)
- [What does least privilege look like for an AI agent?](https://opencontroller.lyzr.com/blog/least-privilege-checklist-for-ai-agents.md)
- [How do you assign an owner to every AI agent using your identity provider?](https://opencontroller.lyzr.com/blog/assign-owners-to-ai-agents-with-your-identity-provider.md)
- [More on Agent identity and access](https://opencontroller.lyzr.com/blog/cluster/agent-identity-and-access)
