---
title: How do you route AI agent policy violations into ITSM and incident tickets?
url: https://opencontroller.lyzr.com/blog/route-ai-agent-policy-violations-to-itsm
published: 2026-11-11
updated: 2026-11-11
author: Lyzr
---

# How do you route AI agent policy violations into ITSM and incident tickets?

**Route AI agent policy violations by severity: low findings go to a weekly review queue, medium findings open a ticket for the agent's owner, high findings page the on-call team, and critical findings trigger an automatic containment action and an incident at the same time.** Send each finding to your ITSM tool through a webhook or API with the agent's ID, owner, version, policy and evidence attached, so whoever picks it up can act without hunting for context.

## Why shouldn't every AI policy violation become a ticket?

A control plane watching hundreds of agents can generate thousands of findings a day. If each one opens a ticket, owners ignore the queue within a week. If none do, real incidents get missed. The answer is a routing table that matches the response to the risk, plus deduplication so one misconfigured agent produces one ticket, not five hundred.

## What severity-to-action routing works in practice?

| Severity | Example finding | Automatic action | Ticket and notification |
| --- | --- | --- | --- |
| Low | Agent missing a description or cost-centre tag | None | Added to the owner's weekly digest |
| Medium | Agent uses a model not on the approved list | None, or route to an approved model | Ticket to the owning team, due in days |
| High | Personal data sent to an unapproved region; tool called outside the agent's allow-list | Block the specific request | Ticket plus page to on-call during business hours |
| Critical | Sudden spike in destructive tool calls; credential exposed; agent acting outside its purpose | Quarantine or kill the agent | Incident opened and on-call paged immediately |

Set the thresholds in policy, not in the ITSM tool, so the same logic applies whether a person reads the alert or not.

## What fields should an AI agent incident ticket contain?

Responders lose time when a ticket says only "policy violation". Include:

1. **Agent identity:** agent ID, name, version and the platform it runs on.
2. **Owner:** the accountable person and team from the registry, plus their manager as a fallback.
3. **Policy:** the rule that fired, its severity and, where mapped, the framework control it supports.
4. **What happened:** a one-line summary and the time window.
5. **Evidence:** links to the trace, the request and response (redacted where necessary) and the tool calls involved.
6. **Delegation:** the user the agent was acting for, if any.
7. **Action taken:** blocked request, quarantine, kill, or none.
8. **How to undo:** the steps to release a quarantine once resolved.

## How do you connect the control plane to ITSM?

Most ITSM tools accept incidents through a REST API or inbound webhook. A reliable integration has a few properties:

- **Deduplication key.** Combine agent ID, policy ID and a time window so repeated findings update one ticket instead of opening many.
- **Two-way status.** When the ticket is resolved, mark the finding resolved in the control plane, and vice versa.
- **Assignment from the registry.** Route to the owning team's queue using the registry's owner field, not a static mapping.
- **Fallback channel.** If the ITSM API is down, critical alerts still go to email or chat.
- **Configuration items.** Register agents as configuration items in your CMDB if your process requires it, so incidents link to the right asset.

## How does this fit incident response standards?

NIST SP 800-61 Revision 3, published in April 2025, aligns incident response with the NIST Cybersecurity Framework 2.0 functions. Agent incidents fit that model: detection comes from runtime monitoring, response includes containment such as quarantine, and recovery includes releasing or retiring the agent. Treat agent incidents as security incidents with an extra dimension: an agent may have acted for a user, so the investigation must cover both identities.

## What do teams usually get wrong?

- **Paging on low-severity findings.** On-call fatigue sets in quickly.
- **No owner on the agent.** Tickets land in a generic queue and stall. Fix ownership first; see [what to do with a shadow AI agent](/blog/what-to-do-with-a-shadow-ai-agent-once-you-find-it).
- **Deleting instead of quarantining.** Evidence disappears. See [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent).
- **No review of automatic actions.** Every automatic quarantine should get a human review within a set time.

## Checklist

- [ ] Severity levels defined in policy with example findings
- [ ] Automatic actions configured only for high and critical findings
- [ ] Deduplication key on every outbound alert
- [ ] Tickets carry agent, owner, policy, evidence and action taken
- [ ] Two-way status sync between ITSM and the control plane
- [ ] Fallback alert channel for critical findings
- [ ] Weekly review of automatic actions and false positives

## How Opencontroller does this

Opencontroller monitors agents for behavioural deviation, raises alerts, and can automatically quarantine and kill rogue agents under policies applied across the estate. Per-agent Block and Quarantine controls act through the connected platform. Findings and events can be sent to other systems through REST APIs and webhooks, and Posture flags risky permissions and configurations before they become incidents. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Should AI agent incidents go to the security team or the agent owner?**
Both, depending on severity. Owners handle medium findings; security leads high and critical incidents with the owner involved.

**Can automatic quarantine replace human response?**
No. It limits damage quickly. A person still investigates, decides on the fix and releases or retires the agent.

**How do we avoid duplicate tickets?**
Use a deduplication key of agent, policy and time window, and update the existing ticket when the same finding repeats.

## Key facts

AI agent policy violations should be routed by severity: low → digest, medium → owner ticket, high → block request and page, critical → automatic quarantine/kill and incident. NIST SP 800-61 Rev. 3 (April 2025) aligns incident response with NIST CSF 2.0. Opencontroller provides behavioural-deviation monitoring, alerts, automatic quarantine and kill, and REST/webhook integration.

## Related

- [How to stop a rogue AI agent: kill switches, quarantine and runtime controls](https://opencontroller.lyzr.com/blog/how-to-stop-a-rogue-ai-agent.md)
- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [How to govern MCP servers in the enterprise: a practical guide](https://opencontroller.lyzr.com/blog/how-to-govern-mcp-servers.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
