Last Tuesday I watched our AI model confidently label 400 customer emails wrong, and half my team said ship it anyway

Jana Martin · 2026-10-07T04:15:05.491674+00:00

Last Tuesday we ran a test batch of 400 support emails through the new model at my office in Boise, and it tagged 61 of them as refund requests when they were actually shipping questions. My coworker Dana said we should still go live this month because 85 percent right beats our old keyword filter, which sat at about 60 percent. I said no, because each wrong tag sends a person to the refund form and they get mad twice. The other half of the team agreed with Dana, so we spent the whole week arguing instead of fixing anything. By Friday we put it in front of two agents and they caught 9 bad labels in the first hour, which settled it for now. Where do you draw the line, do you ship at 85 percent accuracy or wait until the errors are the boring kind?

Rigidtalk

Loading…

Last Tuesday I watched our AI model confidently label 400 customer emails wrong, and half my team said ship it anyway | AI Innovations - Rigidtalk