# Plan what to do when AI gets something wrong at work

An escalation path for AI errors: how to spot, contain, report, fix, and learn from mistakes made with AI tools.

- Topic: Working with AI
- Framework: Few-Shot
- Level: Advanced
- Tags: Legal, Write, Plan, Fill-in Template
- Source: https://www.thinkwithprompt.com/prompts/plan-what-to-do-when-ai-gets-something-wrong-at-work

## Prompt

```text
Create an escalation path for errors caused by AI in our work.

**Where we use AI:** [PROCESSES and tools - e.g. "AI drafts customer emails", "AI agent updates CRM records", "AI summaries used in reports"]
**Types of errors we worry about:** [e.g. "wrong figures sent to a client", "confidential data shared", "incorrect advice", "agent changed the wrong records"]
**Teams and roles available:** [e.g. manager, IT, security, legal, compliance, communications, the tool owner]
**Existing incident process:** [DESCRIBE, or "none"]

Return:
1. **Severity levels** - 3 or 4 levels with plain definitions and an example of each from my work (e.g. Low: internal draft error caught before sending; Critical: personal data exposed or wrong information sent to a customer or regulator).
2. **First actions** - what the person who spots the error does in the first hour: stop, contain (unsend, pause the tool, revoke access), preserve evidence, report.
3. **Escalation table:** Severity | Who to notify | How (Teams or Slack, email, phone) | Within what time | Who decides next steps.
4. **Fix and follow-up** - correcting the output, telling affected people, and checking for the same error elsewhere.
5. **Learning** - a short review template: what happened, why the check did not catch it, what we change (prompt, review step, permissions, training).
6. A one-paragraph version to share with the team.

Make it blame-free so people report errors early. Where a data breach or regulatory notification could be involved, say to contact security or legal immediately - do not set notification deadlines yourself; mark them [CONFIRM WITH LEGAL].
```

## What to fill in

- DESCRIBE, or "none"
- CONFIRM WITH LEGAL

Free to copy and use. From ThinkWithPrompt (https://www.thinkwithprompt.com/), a library of 2,663 curated AI prompts.