we shipped an AI feature last quarter expecting it to deflect support tickets. it did the opposite for the first six weeks. turns out when something is deterministic, users blame themselves when it breaks ("i must've clicked wrong"). when it's AI, they blame the product and they tell you about it. every weird output became a ticket, because now there was something to report. the fix wasn't a better model. it was setting expectations in the UI before the output, and giving people an...