That UX guy who called my AI chatbot "robotic" forced me to retrain

I spent 3 months building this customer support bot for a client in Austin. Some UX designer they brought in tested it and just said "this sounds like a computer, not a person." I was so mad at first, but he was right. I went back and fed it transcripts of real human conversations, stopped using formal phrases like "how may I assist you," and switched to simple stuff like "hey, what's up?" After 2 weeks of retraining, the bot's satisfaction scores went from 60% to 88%. Anyone else had to totally change their training data after harsh feedback?

2 comments

2 Comments

nora_dixon1mo ago

Did you actually sit and listen to the full transcripts yourself before feeding them in, or did you just dump a bunch of files into the training set? I've seen folks skip that step and end up with a bot that picks up all the weird filler words and awkward pauses from real talk.

the_rose1mo agoMost Upvoted

Oh man, I had a friend who tried this with customer service chat transcripts. He just dumped like 5000 files in without cleaning them up first. Next thing you know, the bot starts every response with "Uhhhhhhh so like..." because that's how half the agents talked. It was basically asking people if they could "hold on a sec" while it figured out how to refund their order. He learned real quick why you gotta at least skim that stuff first.