← Rescue patterns
Case pattern
AI automation that hallucinated in prod
Anonymized pattern we see often. Same playbook: diagnose, stabilize, harden, hand over.
Before
An ops bot drafted emails and CRM updates with confident wrong answers—and no evals to catch it.
How we approached it
- 01Map where the bot can write or send without approval
- 02Add human-in-the-loop on high-stakes actions
- 03Ship a small eval suite for the failure modes you’ve already seen
- 04Monitor false actions and tighten prompts/tools from real data
After
We added guardrails, human-in-the-loop checks, and a regression eval suite before messages left the system.
Outcome
Fewer false actions, measurable eval coverage, and a safer automation loop.