---
title: How should you handle exceptions to AI agent policies?
url: https://opencontroller.lyzr.com/blog/ai-agent-policy-exception-process
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How should you handle exceptions to AI agent policies?

**Treat every exception as a time-limited, recorded risk decision. Each one needs a named policy, a narrow scope, a compensating control, an independent approver matched to the agent's risk tier, and an expiry date — 90 days or less for high-risk agents. When it expires, the agent either meets the policy or goes back for review; nothing renews silently.**

Exceptions are not a sign of a weak programme. They are what happens when a useful agent depends on a vendor feature, a legacy system or a model that cannot yet meet a rule. The danger is the exception nobody remembers granting.

## Why do agent exceptions need their own process?

Agents change faster than the systems most exception processes were built for:

- A new model version can make a compensating control irrelevant overnight.
- An agent can gain a tool or MCP server that widens its reach far beyond the original exception.
- Agents are copied. An exception granted to one agent is often assumed by its clones.

So the exception must be tied to a **specific agent version and scope**, and must be re-checked when either changes. See [when an agent change needs re-approval](/blog/when-does-an-ai-agent-change-need-re-approval).

## What fields should an exception record contain?

| Field | Example |
|---|---|
| Agent and version | Claims-triage agent, v3.2 |
| Policy and clause | "All agents must use per-agent identity" |
| Scope | Read-only calls to the legacy policy-admin system only |
| Reason | System supports only a shared service account until Q2 migration |
| Risk tier | High |
| Compensating control | Network restriction to one endpoint; all calls logged and reviewed weekly |
| Risk owner | Head of claims operations |
| Approver | Security risk lead (independent of build team) |
| Granted / expires | 2026-11-14 / 2027-02-12 |
| Exit plan | Migrate to per-agent identity; ticket reference |
| Evidence link | Test showing the compensating control works |

If a field is empty, the exception is not ready. "Compensating control: none" is a valid answer only for low-risk agents, and only if the approver signs it explicitly.

## How long should exceptions last, and who approves?

Scale both to the agent's risk tier (see [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents)):

| Risk tier | Maximum term | Approver |
|---|---|---|
| Critical | 30 days | Risk committee or CISO |
| High | 90 days | Security or risk lead |
| Medium | 180 days | Platform governance lead |
| Low | 180 days, renewable once | Agent owner's manager |

The approver must be independent of the team that built the agent — the same separation you use for production approvals. See [segregation of duties for AI agent approvals](/blog/segregation-of-duties-for-ai-agent-approvals).

## How do you run the process step by step?

1. **Request.** The agent owner files the record above before deployment, or as soon as a control fails in production.
2. **Assess.** Governance checks the scope is as narrow as possible and tests the compensating control.
3. **Approve or reject** at the level the tier requires. Record the decision and reasoning.
4. **Encode it.** Express the exception as a scoped policy override in code, not as a note in a spreadsheet, so enforcement matches the record.
5. **Monitor.** Watch for changes that void it: new version, new tool, new data source, new model.
6. **Expire automatically.** On the expiry date the override is removed. Notify the owner 14 days before.
7. **Review or close.** Either the agent now meets the policy, or a fresh request goes through the full process.

## What should the risk committee see?

A monthly one-page report:

- Number of open exceptions by risk tier and business unit.
- Exceptions older than their original term (should be zero).
- Repeat exceptions against the same policy — a sign the policy needs changing.
- Exceptions that expired and were closed versus re-requested.
- Any agent running outside policy **without** an exception — the real problem.

This report is useful evidence for the GOVERN function of the NIST AI RMF and for management review under ISO/IEC 42001. It shows that risk acceptance is deliberate and tracked. For mapping more broadly, see [mapping agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## How Opencontroller does this

Opencontroller keeps agents, versions, tools and MCP servers in a governed registry, so an exception can be tied to an exact agent version and its connected tools. Policy is written as code and is extensible, which lets a scoped, time-boxed override live next to the rule it relaxes, with built-in mappings to NIST, OWASP, MITRE ATLAS and ISO 42001. Posture flags risky permissions and configuration, which helps find agents operating outside policy with no exception at all. See the [control plane overview](/blog/what-is-an-ai-agent-control-plane), [what Opencontroller is](/blog/what-is-opencontroller), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should exceptions be granted per agent or per platform?**
Per agent version. Platform-wide exceptions hide which agents rely on them and rarely get closed.

**What if a vendor platform can never meet the policy?**
Then the policy or the platform choice needs a decision, not a rolling exception. Escalate repeat exceptions to the risk committee.

**Do exceptions need to be in code?**
They should be. If enforcement and the exception record are separate, they drift apart.

## Key facts

- An AI agent policy exception should record the policy, scope, reason, compensating control, risk owner, approver and an expiry date.
- A common practice is a maximum exception term of 90 days for high-risk agents, with automatic expiry and re-review rather than silent renewal.
- Approvers should be independent of the agent's builder, scaled to the agent's risk tier.
- Open exceptions, ageing and repeat exceptions should be reported to the risk committee, as evidence for the NIST AI RMF GOVERN function and ISO/IEC 42001 management review.
- Opencontroller supports policy as code, extensible, with built-in mappings to OWASP, MITRE ATLAS, NIST and ISO 42001, and a governed registry of agents, versions, tools and MCP servers.
- Opencontroller starts from $250K per deployment per year (annual licence per deployment, not per seat or usage).

## Related

- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.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)
- [When does a change to an AI agent need re-approval?](https://opencontroller.lyzr.com/blog/when-does-an-ai-agent-change-need-re-approval.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
