Chatbots: A Black Eye?
Some chatbot approaches are already legacy
As we said in our last post, Chatbots are hot.
Most people think of Chatbots as a fully automated chat interaction. As such they obtain 100% automation.
But not all automation is created equal! An automated mistake is no help to a frustrated customer. Quite the contrary, receiving an irrelevant reply infuriates customers. And poor experience is a major contributor to customer churn, making this an expensive problem.
Consider this actual conversation with an IBM chatbot for ordering pizza:
This is an example of a non-linear dialogue: the human wants to discuss toppings but the chatbot only wants to discuss delivery. Much of human dialogue is not linear. We sometimes wish to accomplish multiple tasks at once or frequently our memory is prompted by something we have just said or heard, resulting in a desire to change the subject. This is normal and comfortable for humans.
In contrast, linear dialogues seem unnatural to us and can make us feel constrained and uncomfortable. We experience “linear” dialogues whenever we reach a telephone IVR that directs us to press 1 for this or press 2 for that. If what we want isn’t listed, we are out of luck. Telephone IVR has been called the most hated technology in history. So if chatbots are constrained to be linear like IVR, they will also be unpopular with customers.
But it gets worse. Chatbots based on explicit rules (expert systems) or decision trees are complicated to construct. And they can be a nightmare to maintain. As time goes on the knowledge must be updated, and if not, the bot will give incorrect answers. For instance, according to Wired Magazine, Facebook’s Bot for Messenger was recently discontinued after the social media giant reported a 70% failure rate for its chatbots.
There are a lot of lousy chatbots out there today, some being pushed by the mega-brands in tech. Most of them employ the decades old legacy approach of decision trees with a static knowledge base that doesn’t learn over time. There is a real danger that the AI industry will get a black eye as companies spend a fortune on this legacy technology and then find it doesn’t work very well. For more examples on failing bots and a good explanation of why they fail, check out this site.
Is there an AI technology that really works in practice for customer care? Our research shows that neural networks perform much better than traditional decision trees. Next time we will examine artificial neural networks and machine learning, which offer us a way forward.
