Insight of the Week: Task automation beats question answering

By Kerry Robinson

"Any sufficiently advanced technology is indistinguishable from magic" - Arthur. C. Clarke

I don't know about you, but when I first used ChatGPT, it certainly felt like magic. Even though I understood exactly how the magic trick worked, it was impossible not to be impressed.

I spent months exploring the capabilities and weaknesses. Understanding what it could and couldn't do.

It was clear to me that what we had was a step-change from the kinds of AI and machine learning that went before. But it was also clear to me that we would need to work hard to get the best out of it.

The first solutions we launched were based on question answering - what's not known as RAG: Retrieval Augmented Generation. My new boss came up with the pithy name: faqGPT, and our demonstrations garnered lots of interest.

But there are two challenges with question answering in contact center scenarios:

1. Customers don't just have questions, they want to get stuff done. So RAG is rarely enough.
2. Question answering is actually a lot harder than getting stuff done.

Why is question answering harder? Well, on the one hand, there are a lot of different questions that might be asked, so you need to bring a lot of good knowledge to the table, and many businesses knowledge bases are messy and out of data.

But the other issue is knowing when you're right. How can you be sure that the answer you gave to an agent, or directly to a customer, is correct? It's literally one of the hardest problems in AI. Generative AI is great at chatting, and summarizing and translating, but it literally has no logic built in. Right and wrong are all shades of grey to ChatGPT, and similar models. It's all statistics.

When we move on from simple RAG based question answering systems, and focus on task automation, things get more complex in one way, but a lot easier, and safer, in another.

Task based automation is more complex, because there's more structure to the interaction. We're not just getting a question and providing an answer. We're negotiating a fine line between the customer's need, the realities of data in a CRM, the business process, and the need for security and compliance:

• Before we check a balance, we need to authenticate.
• Before we get order details we need to find it.
• Before we finalize a booking, we need to calculate the price.
• Before we process the transaction, we need confirmation from the customer.

Imposing this structure is easy in a simple menu based IVR system, but much harder to achieve with Generative AI. But it is doable. As you'll have seen from the examples I shared in my previous article on Extreme self-service with GPT.?

It also takes more powerful models like GPT4 which are slower and more expensive. Or you need to do fine-tuning of smaller, cheaper, faster models, which I covered in my article on The power of AI fine tuning.

But once you've cracked the code, it becomes easier to stay on track. Because now we have constraints!

• You can't get the order details unless the authentication token is valid. And the order number is correct.
• You can't calculate the price unless the line-items are valid.
• The customer won't say yes, if the summary is wrong.

That's the beauty of using a Gen AI powered voicebot or chatbot to get stuff done for a customer, rather than just using it to answer questions: if it gets a question wrong, you might not know, and that could have bad consequences. If it doesn't follow the process to complete a task, it won't be able to complete it, so it'll either have to work with the customer to fix the problem or escalate to a senior agent to pick up the pieces. Either way, no real harm done.

So as seductive as question answering is. And as elegant as RAG architectures are. The real opportunity in voicebots and chatbots, in the short term, at least, is in automating tasks. In voice, as well as chat.

But don't be limited by your preconceptions of what's possible: Gen AI is not IVR ?- it's capable of many if not most of the things your junior agents do today, which has serious consequences for your contact center, and could mean the beginning of the end for agent churn. ??

Kerry Robinson is an Oxford physicist with a Master's in Artificial Intelligence. Kerry is a technologist, scientist, and lover of data with over 20 years of experience in conversational AI. He combines business, customer experience, and technical expertise to deliver IVR, voice, and chatbot strategy and keep Waterfield Tech buzzing.

Subscribe to Kerry's Weekly AI Insights