Vent: My AI customer service bot in Portland kept gaslighting my own clients
I run a small plumbing company and last month I set up an AI chatbot to handle after-hours calls, thinking it would save me from angry voicemails. Turns out it was confidently telling people their leak could wait until Monday, even when they said water was pooling in their basement. One customer quoted it back to me, saying 'your system said it's low priority,' and that's when I realized I had to test the bot with worst-case scenarios before trusting it. I've since added a keyword override for 'emergency' and 'leak' that routes straight to my cell, but I'm wondering if anyone else has had to put training wheels like that on their AI tools, or if I'm just overthinking it?