What I Learned From 100 AI Support Tickets
After reading a hundred support threads, the failure patterns were boring and repeatable. So are the fixes.
I went through a hundred support tickets tagged AI-related and expected exotic failures. They weren't. Eighty percent were the same three things: a prompt that assumed the model knew private context it didn't have, an automation with no human checkpoint, and output nobody reviewed before it shipped.
The context one is the big one. People ask the model to 'use our policy' as if it read the handbook. It didn't. The fix is mundane: give it the document, or connect the retrieval, or stop pretending the model is telepathic. Every ticket like this closed the moment we fed it the actual source.
The no-checkpoint automation was the scary one. A flow that emails customers with AI-generated text and nobody reads it is a reputation incident waiting to happen. Adding a single human approval step killed the worst class of complaints overnight.
The lesson isn't 'AI is unreliable.' It's that the failures are process failures wearing an AI costume. The tickets weren't about the model being wrong; they were about us skipping the steps we'd do for any other draft. Read the output. Give it context. Keep a human in the loop. Boring, and it works.