What's the solution here? An intermediate classifier to catch irrelevant commands? Seems wasteful.
It's almost like the solution needs to be a fine-tuned model that has been trained on a lot of previous customer support interactions, and shut down/redirect anything strange to a human representative.
Then I ask, why bother using a GPT? It has so much loaded knowledge that is detrimental to it's narrow goal.
I'm all for chatbots, as a lot of questions & issues can be resolved using them very quickly.
Can they though? Generally when I chat with customer service it’s because I need a change which cannot (or cannot easily) be done myself.
Giving chatbots the power to make drastic alterations to accounts could potentially cause a lot of problems.