Part 1, Introduction: Chatbots beyond the hype
In 2016, Microsoft CEO Satya Nadella stated, “Bots are the new apps.” This proclamation sparked a Silicon Valley frenzy over chatbots, and led to headlines dubbing 2016 the “year of the bot.” Two years later, however, after numerous embarrassing public bot failures (let us all remember Tay’s racist meltdown, T-mobile’s failed service bot, and BabyQ’s unpatriotic statements), many executives are wondering when and how the year of the bot will actually manifest as an enterprise tool.
In a recent large-scale survey on Enterprise Chatbots NewtonX collected insights from 1,000 senior executives at large enterprises globally to take the pulse of the chatbot market. The feedback revealed equivocation: 61% of executives reported that they “Do Not Know” if chatbots can provide a strategic advantage to their business. To dive into why executives are unsure about this technology, NewtonX conducted follow-up interviews with 50 thought leaders in the space, including executives at companies that have chatbots as part of their product offerings and companies that use chatbots as part of a strategic initiative. The insights gleaned from this study informed the structure and data referenced in this three part guide to chatbots in 2018.
One reason for the chatbot failures of 2016 was that their intent was not clearly defined. While Forrester found that customers want their chatbots to be “polite, caring, intelligent and funny,” NewtonX data found that personal characteristics of chatbots are secondary to usability: 93% of consumers said the number one most important factor in whether or not they will interact with a chatbot is whether or not it “expedites a process.”
This is what was missing in the Chatbots 1.0 of 2016. While they certainly had the personal touch, their functionality was severely lacking, particularly from an enterprise perspective. As TechCrunch wrote in 2016, “Despite [the]hoopla, it’s hard to find a single chatbot that’s actually a really good product.” Today, however, while enterprise chatbots may be less flashy than their predecessors, they do have precise business applications that improve efficiency, scalability, and the customer experience. In part 2 of this Executive’s Guide to Chatbots in 2018 we will outline what these business applications are, and what benefits they offer.
In order to understand functionality, however, it’s important to first understand how chatbots work, how they were developed, and what their abilities are today.
A Quick History of the Bot
The first chatbot was built in 1966 at MIT. Dubbed “Eliza,” the bot was created to elucidate the superficial, scripted nature of human conversation. The bot simulated conversation through scripts by using a technique called pattern matching — wherein the bot would identify pieces of conversation that directly mapped to a response. For instance Eliza could be programmed to respond to “Do you know what the Statue of Liberty is?” by mapping “What is the Statue of Liberty” to “The Statue of Liberty is a sculpture on Liberty Island in New York City,” and mapping “Do you know what is” to “What is.” The chatbot could respond to any text that it could classify into one of its associated patterns.
Eliza spawned many chatbot descendants, the most notable of which are outlined here. She was also the product of a legacy of events that were necessary for the creation of chatbots:
1950 – The Turing Test
In 1950, Alan Turing wrote a paper on computing machinery and intelligence, in which he posed the question, “Can machines think?” To answer this question, he created a thought experiment in which there are two rooms, one with a human and one with a machine. A judge sits in a third room, and communicates with the other two rooms by means of a computer terminal in order to determine which room contains the human, and which the machine. If the machine successfully fools the judge into thinking it’s a human 50% or more the time then it passes what is now known as The Turing Test. In the end, Turing changed his question from “Can machines think?” to “Can machines act like thinking things can?”
1997 – Jabberwacky is released to the public on the world wide web
Jabberwacky was created in the 1980s by Rollo Carpenter, a British programmer, with the specific intent of passing the Turing test. While the chatbot was technically very similar to Eliza, it differed in two notable ways: the first, is that it had a very short working memory, meaning it could repeat back what users said to it with tiny additions. For instance, if a user said, “I’m sad,” the bot could respond, “I’m sad too! Why are you sad?”. It’s also notable for being the first Internet chatbot available to the general public.
2001 – SmarterChild
SmarterChild was the first chatbot created for utility rather than novelty. It was released via AOL Messenger and MSN Messenger, and had databases of information about things like the weather, news, and movie times. It could also perform basic calculations including converting between Farenheit and Celsius.
2011 – Watson by IBM
Watson was the first highly publicized chatbot capable of answering questions posted in natural language. It was propelled to fame in 2011 when it beat Jeopardy champions Ken Jennings & Brad Rutter. Watson was one of the first bots to be able to interact with natural language beyond scripts. Instead of translating phrases into scripted responses, the bot parses questions into different keywords and sentence fragments in order to find statistically related phrases. Today, it’s used in myriad industries, from law, to the diamond industry, to government.
2016 – Tay by Microsoft
Tay was released by Microsoft on Twitter at the height of the chatbot hype, and was revoked after just 16 hours. Tay was one of the first (of many) chatbots that utilized machine learning to improve its capabilities over time based on interactions with users. Unfortunately, there were no limits placed on this learning, so Tay ended up tweeting racist content and drug references. The incident was reminiscent of Watson, which ended up using profanity after reading entries from Urban Dictionary.
The Present: How 2018 Bots Work
There are different types of chatbots in use today, and some function more efficiently than others. For instance, some chatbots offer button options within the chat for quick responses, while other rely solely on NLP. Almost all enterprise bots, however, use an algorithmic hierarchy for categorization and response. This means that they’ve been trained to understand phrases and combinations of keywords as belonging to a certain category (such as “greeting”) and then respond accordingly. If they do not understand a query, they will either defer to a human who can enter the conversation, or offer a canned deflection.
We will explore the use cases for these bots, the training data that they’re built on, and how they use machine learning to improve over time in Part 2 of this Executive’s Guide to Chatbots in 2018, “Chatbots Today: The Three Primary Use Cases.”