A Human and AI hybrid approach for complex AI-based Conversational Commerce
On this trip, I was able to see a remarkable real life AI and Technology powered call center, where the humans oversee and assist a smart technology that talks to customers directly.
I literally watched an AI based machine “sell” a vacation package for about $700 to a customer who had no idea that they were talking to a machine (and not a human). It was literally amazing. This was not a call to simply “answer a question” or “recommend a product.” No, this was a complete sales process starting with an introduction of a loyalty program, then describing a vacation product, then making the customer feel comfortable, and eventually even taking their credit card information. And all of this was done in a completely natural tone and manner.
How is this possible? This firm is using a very different approach to AI and human interaction. Normally, AI or ML systems use large training sets to learn or infer the decision making patterns of humans and then the machines independently simulate those decisions. This is effectively supervised machine learning, which is a core building block in AI applications.
However, the AI-Pros system is different. It uses a more interactive approach. First, the basic speech patterns are captured from live operators who are very skilled at doing their job, i.e. selling this particular vacation package. Then segments from the conversation can be called on demand and used interactively during new conversations with new customers. At this point, humans assist the machines to guide the conversation at a meta level, instead of directly speaking into the conversation.
Even though humans are interactively assisting the conversation, this is a breakthrough approach compared to the way that the giants like Google and Amazon approach the problem. Today, Amazon and Google employ AI chatbots without any human interaction or oversight in the conversation. Note that the end-goal is the same for both approaches, which is to eventually have the computer talk with a person so naturally that they could even sell you a product or tutor you in a calculus class. However the machine-human assisted model might actually get there faster.
Sometimes, what seems like the most direct path might seem shorter, but in some cases like this, the more indirect path might lead to the end-goal faster.
Standard AI/ML Approach:a) Person ← learn from → AI (training)
b) AI ←→ talk with customer (execution)
a) Person ← learn from → AI (training)
Alternate: AI/ML/Person Interactive Approach
b) Person + AI ←→ talk with customer
In this alternate approach, the AI does not need to be as sophisticated in the beginning and yet the system is good enough to operate in real life execution. This is really important because the AI can become more sophisticated during the operation over time, on top of what can be learned in the training. As a result the traditional approach will actually not progress as far nor as fast as this alternative approach over time.
As a parallel, let’s consider Tesla’s approach to autonomous driving vs. driving assistance. While Tesla could have waited until their autonomous driving was perfectly ready, they instead introduced a version where the AI and driver both need to participate. In a similar manner, this has allowed Tesla to introduce the feature much earlier. And the millions of miles of assisted driving that Tesla has driven so far will all be used to later improve the AI’s eventual objective of completely automated driving.
What we are now seeing in this AI-based selling product is a new approach where the AI systems and humans work interactively. And amazingly, it already works. As a result, not only are we seeing better results today, but also we are likely to also see much better results once it eventually develops a pure AI-based machine using this approach.
This is an important case and learning lesson for anyone developing AI applications in the real world for very challenging problems.