AI Agent Incident Response Runbook and Report Template
Sooner or later an agent will move the wrong date, answer from the wrong record or show someone data they should not see. The first hour goes better if the report and the steps already exist. This template gives you both.
Fill in
AI Agent Incident Response Runbook Template
AI Agent Incident Report
Company: Acme Precision Ltd Incident ID: AI-2026-004 Reported: 2026-11-03 by S. Okafor, Planning Severity: Medium Type: Wrong data written
Agent and tool
update_promise_date via purchasing-assistant-mcp
What happened
Promise dates on 14 purchase order lines were moved out by 7 days after a vague request.
Impact
14 PO lines for 3 vendors. No financial postings affected.
Containment
Write tool disabled at 10:20. Dates restored from audit log before planning run.
Root cause
Tool accepted a bulk update without a record cap. Prompt did not require confirmation of scope.
Response checklist
| Step | Time | Owner | Done |
|---|---|---|---|
| Disable affected tool or revoke service account | |||
| Preserve logs, prompts and tool arguments | |||
| Identify and list affected records | |||
| Restore or correct data with approval | |||
| Notify data owner and, if required, regulators or customers | |||
| Add the case to the evaluation test set |
Corrective actions
| Action | Owner | Due date | Status |
|---|---|---|---|
Approvals
A starting point to adapt to your own incident process, not legal advice. Notification duties for personal data or regulated records depend on your jurisdiction and contracts, so check them with your counsel or compliance team.
Stress test my agent controls
Optionally request a review of your agent guardrails and incident readiness.
Optional. The tool and every export work without this. No spam.
How to use the template
Open it the moment an incident is reported. Fill in the agent and tool, a plain description of what happened and the records affected. Record what you did to contain it and when. Work through the response checklist in order, then write the root cause and corrective actions once the facts are known. Export to PDF or Word for the record and share with the data owner and security.
Containment first
The fastest containment for an agent is to disable the single tool or revoke the service account, not to switch off everything. That is why a per-tool register and a separate account for write paths matter. Preserve evidence before you change anything: the prompt, the tool arguments, the generated query and the ERP transaction log. Then restore data from the audit trail with the same approval you would require for a normal correction.
- Disable the tool, not the whole service, where possible
- Preserve logs before restarting anything
- Restore using the recorded before values
Root cause categories
Most agent incidents fall into a few causes: a tool with too much reach, a prompt or description that did not require confirmation, an ambiguous user request the agent resolved by guessing, a data mapping error, or a role that granted more than intended. Write the cause at the system level so the corrective action changes a control, such as a record cap, a required confirmation or a narrower role, rather than blaming the model.
Common mistakes
Teams often fix the single record and miss the pattern, fail to add the case to the test set so it can recur, or delay notifying the data owner while investigating. Review the corrective actions at an agreed date and close them formally. Keep the tone factual: the goal is a stronger control, not a person to blame.
Frequently Asked Questions
What counts as an AI agent incident?
Any unexpected write, exposure of data to someone not entitled to it, action taken without required approval, or business decision made on a wrong answer. Near misses are worth recording too.
How fast should we contain an incident?
As fast as you can safely disable the affected tool or account. Containment should come before root cause analysis, with logs preserved first.
Do we have to notify regulators or customers?
It depends on the data involved, your jurisdiction and your contracts. Personal data exposure often has legal timelines. Confirm with counsel or your compliance team and record the decision in the report.
Should incidents become test cases?
Yes. Add each incident as a case to your evaluation set so a future change that brings the failure back is caught before release.
Netray builds ERP agents with read-only defaults, visible queries and approval on write-back to keep incidents small. If you want a review of your controls, we can help.
Related Tools
Agent Action Audit Log Worksheet
Record agent actions, the tool called, the ERP record affected and the approver, with a computed latency and exception count.
Developer ToolsAI Agent ERP Access Request Form Template
A fill-in form that records which ERP roles, objects and limits an AI agent service account needs, with approval signatures.
Developer ToolsMCP Tool Catalog and Risk Worksheet
Register each MCP tool your agent exposes, the ERP object it touches and a computed risk score, then export to Excel.
AI Agents & AutomationAI Agent Evaluation Test Case Worksheet
Log evaluation questions for an ERP agent with expected answers, the tool it should call and scored results, with a computed pass rate.
More Compliance & Contracts tools
See all- SAP SoD Residual Risk Score Calculator
- SAP Firefighter ID Usage Review Log
- IBM i (AS/400) User Profile Access Review Worksheet
- SyteLine Security and User Access Audit Checklist
- Infor XA (MAPICS) User Access Review Worksheet
- China ERP Data Security and Compliance Assessment
- ERP Vendor Data Processing Agreement Outline
- GDPR ERP Data Migration Checklist
- EU Machinery Regulation 2023/1230 Readiness Checklist