Reliability
Failure taxonomy
A written, shared vocabulary for the specific ways your AI system goes wrong.
"The AI is being weird sometimes" cannot be prioritised, assigned or fixed. "Tool selection error on multi-item requests" can.
A failure taxonomy comes from reading real production transcripts and grouping what actually went wrong into named categories. Once failures have names they can be counted, ranked by cost, and argued about productively.
It is unglamorous work and it is usually the highest-leverage thing an AI team can do — because it turns a vague sense of unreliability into a prioritised list.
Related