With a growing use of general purpose AI in business, we are seeing a rising interest in chatbots. For those unacquainted with this term, a chatbot is an automated text agent with pre-defined conversation flows. Conversational AI is employed in an effort to understand a person’s intent and take action based on their words.
Whether the person speaking with the conversational agent is a customer or an employee, risks arise with how a chatbot interprets human statements, requests, and questions. When these misinterpretations occur, the errors can cascade and limit the utility of the chatbot. As with other new and disruptive technologies, many deployments of chatbots have neglected to include provisions for managing the security of data, lacked operational tools to fit within established project management practices, and ultimately failed to deliver business value.
The AI deep learning models that underlie chatbots are often opaque. We know WHAT the AI does; but we don’t know HOW it does it. Conversational AI is powered by the introduction of randomness into its development. These remarkable tools learn by trial and error, just as we do. And like humans, they are prone to error. That’s why the major discipline in the field of AI is called “machine learning.” Not knowing HOW or WHY an AI tool works is hindering adoption in business.
Today, most AIs are largely inscrutable — and cannot be validated. For example, a recently created AI-powered hiring recommendation product yielded an outcome that indicated gender and racial biases. Its AI was trained using 10-year old data, which pre-dated today’s standards of fairness. This example underlines why understandable processes and the proper use of data is essential. Guidance from intelligent machines will require greater control and transparency to be viable.
MAKE CHATBOTS SAFE
How then can the advantages of conversational AI be introduced into business processes while retaining acceptable operational reliability?
Jonathan Shea, Vice President of Information Technology at MacroGenics speaks to a gradual adoption of cognitive AI within the highly regulated pharmaceutical industry.
“In a prior role with a strong customer service channel, we leveraged voice prompts and employed voice analytics to match customers to the best agent based on their personality traits. This drove an enhanced customer experience. I anticipate AI to greatly improve matching and other service capabilities. Even when AI can’t fully automate the customer experience, I expect to use it to enhance support for customer service representatives.”
To have a successful chatbot program that delivers value for business stakeholders, you should manage them as you would any other form of software in your system portfolio. In that light, chatbots have some unique risks which are fortunately easy to manage. Here are three risks and recommended ways to overcome them.
SAY THE RIGHT THING
A few early conversational bots tried to learn from human responses in real-time. In 2016, a Twitter bot created by Microsoft even learned undesirable behaviors and offended millions of observers. As with any customer engagement system, errors can rapidly cascade through systems, negating goodwill and increasing the costs of running our organizations.
Instead, you should specify approved wording for each machine response (also known as an utterance). Even with this control, personalized responses are possible. To create engaging conversation, employ the same human editors you depend upon for advertising and internet sites.
PREVENT CATASTROPHIC ERRORS
This may be the highest risk of conversational interfaces. If a customer wishes to pay a third party (think of financial applications such as PayPal’s Venmo, or another mobile payment service), the lack of menu driven visual verification could lead to a payment going to the wrong person. Even the amount could be incorrectly captured by a poorly designed chatbot and require human customer service to remedy the situation.
To prevent these outcomes, manage chatbot projects with the same rigor that you require for every other system that engages with customers or employees. In one use case for current online services, there may already be an automated workflow to execute transactions accurately and efficiently. Conversational agents should be governed by your existing information management policy and operate within your established business controls.
MAKE IT DO SOMETHING VERY USEFUL
Today’s bots can fail to understand the intent and specific details within a human communication. Despite advancements in natural language understanding (NLU), chatbots can become dumbfounded and unable to assist the person.
Successful businesses are built on personalizing customer experience and providing solutions that go beyond solving pain points. Design your chatbot to deliver what your employees, stakeholders, and customers crave.
You might choose a service they already have, such as the ability to check-in to a hotel. Imagine your customers not having to install an app on their smartphone and still be able to check-in using an SMS text chatbot. The ability to make a customer’s experience easier is one of the best cases for extending current methods of engagement with well-designed chatbots.
While chatbots employ the most advanced AI available today, they are simply clever computer programs. To achieve success and avoid mishaps, corporate technology leaders should purposefully govern chatbot development and operations.
 “Companies Grapple with AI’s Opaque Decision-Making Process,” The Wall Street Journal, Sara Castellanos, May 2, 2018, blogs.wsj.com/cio/2018/05/02/companies-grapple-with-ais-opaque-decision-making-process
 “Had a good conversation with an intelligent bot? The story of a naughty artificial intelligence,” Jack Crawford, October 9, 2016, linkedin.com/pulse/had-good-conversation-intelligent-bot-story-naughty-jack-c-crawford
Copyright (c) 2016, Jack C Crawford, All rights reserved