Engineering note · Technical support engineering · by Ayman Sbeiti — I support high-trust software platforms · Hiring?
Engineering playbook · AI workflow
Reviewing an AI assistant's wrong answers
Core principle
Configuring the assistant took hours. Reading its wrong answers and fixing the content underneath them was the real project.
Real-world example
Configuring Intercom's Fin against our knowledge base during the support-platform migration was quick: point the assistant at the content, set its boundaries, turn it on. What that bought us was an assistant exactly as trustworthy as the knowledge base underneath it, which at that point was not trustworthy enough. The real work was a review loop: reading the assistant's answers, finding the inaccurate and hallucinated ones, and tracing each back to its cause in the content.
Why it happens
The assistant rarely failed on its own. Wrong answers traced back to the content it was reading, and fixing an answer almost always meant fixing the underlying article rather than adjusting the assistant. Content debt became AI debt the moment the assistant went live.
What I now check
- Review the assistant's answers the way you'd review a new agent's: systematically, against the source of truth, with the failures logged and traced. Spot-checking until you feel reassured doesn't count.
- Treat the boundary decisions as the trust decisions: what the assistant must hand off to a human matters more to customers than how much it can answer.
Production takeaway
An assistant that says less, correctly, beats one that says more, confidently.