A hallucinated answer from the AI reached the customer's end customer and caused a complaint. How do you handle the incident?

Incident steps

1. Contain (hours): disable auto-send for that category, or switch to HITL; patch with a guardrail if the pattern is clear.
2. Assess scope: search logs for similar outputs (same topic, same failure pattern) to see how many users were affected.
3. Communicate: notify the customer's owner quickly with what happened, the impact, the immediate actions and the next update time. No blame, no minimising.
4. Root cause (traces):
   ◦ Was the right document retrieved?
   ◦ Did the policy change recently (stale index)?
   ◦ Did the prompt allow answering without sources?
   ◦ Did a guardrail (groundedness check) exist?
5. Fix: the specific issue + systemic improvements (abstain when sources are missing, citation checks, confidence thresholds).
6. Prevent: add the case and its variants to the eval set; monitoring for similar patterns.
7. Post-mortem document: timeline, root cause, contributing factors, action items with owners and dates.

Interview signal: calm ownership, transparency, and systemic fixes, not just patching one prompt.

Related

• Previous: Q31. A customer asks, "How much will this cost us per month in model usage?" How do you estimate it?
• Next: Q33. The customer wants to compare your solution against a competitor and their internal team's build. How do you approach the bake-off?
• NeMo Guardrails
• Guardrails AI
• DeepEval