Neural Networks to the Rescue
As we discussed in our previous blog, chatbots have a big problem: they are expensive to build, hard to maintain, and have a very low success rate in answering customer care questions. Why would anyone go to great lengths, and great expense, to build something that doesn’t really work very well? They wouldn’t, of course.
Introducing Neural Networks
Simply put, Neural Networks are a form of computing that simulates how the human brain is thought to work. It consists of an extensive network of neurons that communicate and transfer an input signal to an output. In a customer service setting, the input is all the dialogue interaction between the agent and customer plus all context, and the output is the agent’s response.
While Neural Networks have been around for decades, their enormous computational and data requirements made them an impractical approach until recently. Neural Networks take thousands of examples of solved problems and predict answers to similar problems in the future. This process is sometimes called Deep Learning, which is a popular name for Neural Networks with millions or even billions of neuron connections.
Neural Networks are great at certain tasks like recognizing faces (the iPhone X) or searching for objectionable content (Google). And they have been shown to have excellent accuracy in customer care applications. For instance, at True AI, initial customer results using neural networks are that 60–80% of our recommended replies are used by the agent.
But is even that good enough?
Not really. You don’t want to put a system that fails 20–40% of the time in front of a customer. That will just lead to loads of complaints.
Chat To The Agents, Not To The Customer
But suppose the chatbot doesn’t speak directly to the customer.
Suppose the neural network makes recommendations to the customer care agent instead. That’s a different story. If most of the time the agent is presented with the correct response and just has to click ACCEPT to send it on to the chat user, then we are looking at an enormous reduction in the time it takes to reply to a chat question. And because the agent will edit or discard any wrong response proposed, the chat user will experience genuine human-level accuracy in response to their chat questions. To the customer/user, it just works. They never realise that they are talking (mostly) to an AI. To the agent, it frees up valuable time for him or her up to focus on more complex and interesting issues rather than typing routine and repetitive answer to frequent questions.
Neural Networks improve the quality of service too
Not only will a Neural Network make the chat more efficient, it will improve the quality of responses. This is done by weighting the historical chats from experienced agents more heavily when building the Neural Network. So when an inexperienced agent is asked a question, the Neural Network proposes answers based on ones given by the best agents. It is like having your most experienced agent sitting next to all your other agents. This significantly reduces the training required to onboard new agents and provides for a superior customer experience since everyone gets the benefit of the experienced agent’s responses.
We have been referring here to neural networks, but we really mean a very new and specialised kind of neural network employing End-to-end reasoning. The best neural network for customer service does NOT merely classify questions and serve up canned or template responses (if it is a payment enquiry, use the “get account number” template, if it is a “power outage” enquiry use the “get location” template, etc.).
Rather, the best and most natural response is GENERATED word by word, tailored to the precise question that was asked, rather than a one-size-fits all templated answer for each kind of question.
More on the advantages of End-to-end reasoning in our next blog!
