AI Incident Response Process Flowchart
What to do when an AI feature gives a harmful, wrong or leaked answer: contain it fast, find whether the prompt, model, data or a tool caused it, fix it and add the case to the tests.
4 teams11 steps2 decisionsFree download
Free to download. Edit it in Flow Studio to change the teams and steps, ask AI for changes, or export to PowerPoint, Word or draw.io.
Step by step
The same flow as a numbered list, ready for an SOP, a BRD or a ticket.
| # | Step | Who | Next |
|---|---|---|---|
| 1 | Harmful or wrong AI output reportedStartUser report, monitoring alert or red-team findingReporter · step 2 | Reporter | step 2 |
| 2 | Triage severitySev by harm, reach and whether data left the companyOn-call engineer · step 3 | On-call engineer | step 3 |
| 3 | Serious?DecisionOn-call engineer · Yes: step 4, No: step 5 | On-call engineer | Yes: step 4, No: step 5 |
| 4 | Switch the feature off or to fallbackFeature kill switch; keep tracesOn-call engineer · step 6 | On-call engineer | step 6 |
| 5 | Queue for the next fixAI team · step 8 | AI team | step 8 |
| 6 | Personal data exposed?DecisionOn-call engineer · Yes: step 7, No: step 8 | On-call engineer | Yes: step 7, No: step 8 |
| 7 | Assess and notifyRegulator and affected people within legal deadlines (e.g. GDPR 72 h)Privacy and comms · step 8 | Privacy and comms | step 8 |
| 8 | Find the cause from the tracePrompt, model version, retrieved data, injected input, tool permissionAI team · step 9 | AI team | step 9 |
| 9 | Fix and add a test casePatch prompt, data or guardrail; add the case to the eval suiteAI team · step 10 | AI team | step 10 |
| 10 | Re-enable after evals passOn-call engineer · step 11 | On-call engineer | step 11 |
| 11 | Closed with write-upEndAI team · Done | AI team | Done |
When to use it
Use it next to your normal incident process for any AI feature in production, so the on-call team knows how AI failures differ and what to switch off.
Who's involved
- Reporter1
- On-call engineer5
- AI team4
- Privacy and comms1
Yours works differently?
Describe your own process in a few sentences and Flow Studio draws it, with your teams and your rules.
Describe your processarrow_forwardFor the technical team
The build decisions behind this flow. Each step above also carries its technical detail.
- 01Kill switch per feature via config; fallback copy or human queue when off.
- 02Trace store keyed by request id: prompt version, model id, retrieved chunks, tool calls, output, user feedback.
- 03Root-cause buckets: prompt regression, model version change, bad or poisoned source data, prompt injection, tool permission, missing guardrail.
- 04Post-incident: add the case to the eval suite and block release until it passes.
How to get it right
- 1Build a kill switch for each AI feature before launch: off, or fall back to a safe non-AI path.
- 2Keep the exact inputs, retrieved passages, tool calls and outputs for the incident. Without them you can't find the cause.
- 3Every incident ends with a new test case, so the same failure can't ship again.
- 4If personal data may have leaked, bring in privacy straight away: notification deadlines can be as short as 72 hours.
Questions
What counts as an AI incident?
Any output or action that causes or could cause harm: unsafe or offensive content, confident wrong answers people relied on, leaked personal or confidential data, an agent taking an unintended action, or a jailbreak or prompt-injection exploit.
How is AI incident response different from normal incidents?
The cause is often not a code bug but a prompt, a model update, retrieved data or a malicious input. You need the full trace of the request to find it, and the fix usually comes with a new evaluation case.
More process flow templates
- AI feature integrationTaking an AI feature from idea to production: pick the use case, test models against real examples, add guardrails, pass a security review, and roll out in stages with monitoring.
- RAG pipelineHow retrieval-augmented generation answers questions from your own documents: documents are split, embedded and indexed, then each question pulls the best passages and the model answers with citations.
- AI agent with tools and MCPHow an agentic AI system works through a task: it plans, calls tools through MCP servers with scoped access, asks a person before risky actions, checks results and loops until the goal is met.
- MCP server integrationConnecting an internal system to AI assistants through a Model Context Protocol server: design the tools, secure them, test them, review them and roll them out with monitoring.