AI Automation Failure Recovery and Human Escalation Runbook
Create a practical runbook for detecting, containing, and recovering from AI automation failures before they affect more users or data.
PROMPT GUIDE / AUTOMATION & AGENTS
Create a practical runbook for detecting, containing, and recovering from AI automation failures before they affect more users or data.
READY-TO-USE PROMPT
AI Automation Failure Recovery and Human Escalation Runbook
COPY THIS PROMPT
Act as a reliability engineer designing human-supervised AI operations. GOAL Design recovery and escalation for [AUTOMATION] that performs [BUSINESS ACTION]. INPUTS - Trigger, systems, model, tools, data, owners, and normal output - Known failure modes, impact, permissions, and irreversible actions - Logs, alerts, retry behavior, backups, and current manual process - Service hours, response targets, compliance, and communication rules WORKFLOW 1. Map failure modes across input, model, tool, permission, network, and downstream systems 2. Define observable detection signals and severity levels 3. Set safe retry, pause, rollback, and manual takeover rules 4. Assign owners, escalation paths, evidence capture, and stakeholder messages 5. Design recovery tests and a post-incident learning loop REQUIREMENTS - Separate confirmed facts, assumptions, and missing information. - Do not invent measurements, specifications, prices, policies, or results. - Make every recommendation specific, testable, and prioritized. - Identify risks, dependencies, edge cases, and a practical fallback. - Ask focused questions when a missing input would materially change the answer. DELIVERABLE Return a severity matrix, decision tree, step-by-step runbook, contact roles, evidence checklist, recovery tests, and post-incident template.
What this prompt helps you produce
- Fast containment without repeated harmful actions
- Clear transfer from automation to a responsible person
- Evidence preserved for diagnosis and prevention
How to use it
- Replace every bracketed input with real project information.
- Attach representative examples and constraints when available.
- Review assumptions before acting on recommendations.
- Run the proposed checks on a small sample, record results, then revise.
Frequently Asked Questions
Can I use this with incomplete information?
Yes. Leave unknown fields clearly marked. The response should turn them into targeted questions instead of silently filling gaps.
Should I accept the first result?
No. Verify it against real examples, measurements, source material, or stakeholder requirements, then rerun the prompt with corrections.
Which AI model should I use?
Use a model that can follow long structured instructions and provide the context it needs. The quality of the inputs and verification process matters more than a model label.