I’ve been working with a fintech and sales group on harnessing AI chatbots to guide clients and customers through a process. What keeps coming up is that, much of the time, it’s preferable to walk a customer through a stochastic decisioning tree so that they reach a logical conclusion — or a purchase decision.
The trouble is that GPT-4 isn’t going to start a conversation with an end in view. A traditional programd chatbot will do exactly that. It asks what you’re looking for and narrows in on what, precisely, you want to do.
What a human does well is both parts. A good salesperson takes you through a series of questions with multiple-choice answers until you’ve configured your ideal product, and can also riff with you off-script when you ask something with a more complicated answer than a menu allows. The scripted bot has the first skill and none of the second. The language model has the second in abundance and, so far, none of the first. A mix of classic bot and AI is ideal in these situations. How to do that? The learning continues.
So far, I’ve failed to create a prompt that instructs an LLM to work toward a consensus over multiple back-and-forths. I’ve tried a number of ways to language-program and system-prompt-engineer exactly that kind of result. No luck yet. The model seems happy to answer and reluctant to steer.
On the other hand, I think it will be possible to take a traditional chat platform like Rasa — an open source framework “for automating text and voice-based conversations” — and offer the user multiple-choice responses plus a category like “other.” Pick “other” and the conversation diverts from the dumb chatbot to an AI chat. I’m reasonably certain this will work. Once the AI chat has done its part, a function could call the dumb chatbot back to complete its stochastic progression through its programming.
The timing helps. Last week OpenAI added “function calling” to GPT-4 and GPT-3.5-turbo: developers describe their functions to the model, which has been tuned to “detect when a function needs to be called” and to reply “with JSON that adheres to the function signature.” Among the uses OpenAI lists are chatbots that answer by calling external tools. A hand-off back to a scripted flow is exactly that kind of tool.
