You can find every opinion under the sun on chatbots: claiming that they are the wave of the future, that chatbots will replace apps and make the 800-number obsolete, chatbots make life easier, or that apps are better than chatbots for certain tasks. Certainly they will revolutionize search, reduce wait times to reach a human, and lower the cost of the workforce needed to serve the customer, with a side of novelty.
Chatbots can be designed for anything from automated customer support to e-commerce, interactive experiences, and entertainment. They are continuously being revamped to streamline the customer experience and meet the hastening human need for speed.
Bots can, among many other tasks:
- check shipping orders and order food, a car, or flowers
- manage your calendar, book travel and hotels
- give you the news, check weather and traffic
- help you shop and access product info
- play your favorite music
- give answers to questions depending on their vocabulary
Businesses are trying to leverage Facebook’s one billion users to reach their customers en masse. They hope to connect with customers in new ways to enable self-service within interaction channels that Millennials are innately comfortable with and that Gen Xers and Boomers have become accustomed to. But any new method by business to communicate with customers should be for the customer's benefit and to improve both customer success and the customer relationship.
Ever since In Search of Excellence released 30 years ago, there is no dispute that the customer is the most important focus of your business, perhaps even rivaling innovation. Customers want to be served swiftly and accurately, so while the evolution of chatbots has been colorful, in general they still have a quite a way to go.
What is especially significant is that bots are no longer pretending to be human as they have for decades, but are being presented simply as what they are: and humans are okay with that. There has been a shift in society where we are comfortable with talking to a machine, knowing we are talking to a machine, and to a large extent accepting that machine’s limitations. And it is a relief, because a by-product of the charade was that expectations were too high, resulting in frustration when service fell short and we were either misunderstood, not understood at all, or served the wrong content.
Now we almost assume when we seek help on a website, via chat or email, that we are talking to a bot – which in itself helps reduce disillusionment. When once they were alienating, because bots have now come clean as a UI, expectations have been adjusted. So now, rather than expect performance matching our own human experience from a bot masquerading as a human (and be disappointed), we know that we will not get human capability. And there’s an added curiosity as to how much the bot can actually do. Some even help you along with that, saying things like, “Here are some things you can ask me.”
The floodgates are open
The innovations in machine learning incorporating deep neural networks and artificial intelligence are advancing chatbot technology by leaps and bounds.1 Chatbots are able to be programmed for a very wide variety of queries through “supervised learning,”2 with myriad possible answers – but their vocabulary is still limited. To a certain extent they can extrapolate their existing knowledge based on past interactions to draw conclusions on data points they have never seen before, but the further away they get from the data set they were trained on, the more likely mistakes will be made.
Some chatbots are even dipping into the Internet of Things – linking to lights, electrical switches, and thermostats. But the collective wisdom that’s developing alongside the bots available today is that they shouldn’t be programmed to do everything; that the more restrained their specialization, the more successful they’ll be.
Bots are programmed for a specific purpose to answer anticipated, level one customer service inquiries (the first stop when users contact customer support agents). These are easy to map out. The bot’s confidence in those responses goes up or down based on whether there was a successful outcome that satisfied the customer, raising or lowering the rankings assigned to the answers, and the bot then knows the most effective response to give in the future when the same query is presented. This is called positive and negative reinforcement, and is at the heart of what is referred to as “unsupervised learning.”3
- Neural networks (i.e., artificial neural networks, one type of artificial intelligence) are a type of computation model loosely based on the human body’s own biological neural networks, enabling a machine to learn by analyzing observed data patterns. A neural network can be deep or shallow, referring to the depth of the layers within the network which are made up of computation points that are akin to human neurons. This technology is used in speech recognition and natural language processing, among other things.
- Supervised learning, a form of machine learning, is a system by which developers train the chatbots with an algorithm and give that algorithm a data set that produces correct answers to given queries. In other words, humans “teach” them their function and vocabulary, the chatbots “learn” to recognize the words and sentences in commands from humans that will trigger them to actually give the right answer or perform the requested action. Developers know what the inputs and outputs are, for example for a specific question, they set the chatbot to respond a certain way, and then ask the machine to come up with approximations of correct outputs. There is a control set that the chatbot is trained on, i.e., questions and answers, and once the machine knows those it is also able to predict the answer to questions it has not heard before to a certain degree of accuracy.
- Unsupervised learning involves an automatic feedback cycle on the part of the chatbot, which classifies whether the interaction was successful or not, and that feedback influences future behavior.