I'm begging anyone using AI for customer support to double check the chat logs before shipping

Three months ago I let an AI draft all my replies and it told a customer their order was 'probably lost in the void between warehouses' and I only caught it because the customer screenshot it and posted it on Reddit, has anyone else's bot gone rogue with the tone?

2 comments

2 Comments

margaretm231mo ago

@victorb17 you're not wrong but honestly the "void between warehouses" thing is still making me laugh months later lol. What worked for me was setting up a hard rule where the AI cannot use any metaphors or jokes at all. I literally put in the prompt "pretend you are a polite robot with zero personality" and it helped a ton. Also I started using a two step system where the AI drafts the reply and then a separate prompt checks it for tone before I even see it. That caught a few wild ones like "sorry your package took the scenic route" before they went out. You gotta treat it like a toddler with a sharpie and watch everything it does.

victorb171mo ago

Nah, I gotta disagree. You let an AI run your customer service without checking it. That's on you. The tool did what it was told. It wrote a reply. You just didn't read it before hitting send. If you actually set up the bot right with proper tone guidelines and a human review step, this doesn't happen. I've used AI for basic replies for months. Works fine. But you have to script it like a script. Tell it not to be sarcastic. Tell it to stick to standard corporate speak. The real issue is people treating AI like a magic fix. It's not. It's a tool. You still have to babysit it. The 'void between warehouses' thing is funny though. Customer probably got a kick out of it. But yeah you dropped the ball. Not the bot.