Done right, conversational AI can boost customer experience. But as with all things, the key is doing it right. Because the costs of doing conversational AI wrong can be greater than never using it in the first place.

We asked experts for their advice on how to design effective conversational AI.

Start With Natural Language Processing

"Effective conversational AI starts with natural language processing and the magic of cognitive resources (primarily pattern recognition and analytics) and applies it to intent recognition," said Dan Miller, founder of Opus Research. "By rapidly recognizing what a person wants to do and matching it to tasks such as form filling or navigating web-based resources, it can accelerate the processes involved in completing a desired task."

The success of conversational AI is case-driven, Miller added. "Think how wonderful it would be for a chatbot or voicebot to understand the simple statement, ‘I need my first COVID shot,’ and carry out a successful conversation that culminates in successful booking of an appointment at a local pharmacy or other vaccination venue."

Effectiveness is measured by the system’s ability to understand and even anticipate an individual’s intent based on what they type or say along with a variety of unstructured data or metadata, Miller added. “In the above case, at some point the conversational AI would ingest a person’s location, age, insurance status and other eligibility indicators and navigate them through the process of form filling and booking expeditiously.”

When an effective conversational AI can’t do something, or reaches a dead-end, it knows to transfer to a live person to help, Miller said. In doing so, the virtual assistant (conversational AI) remains available as a coach or nudge that helps the live agent expedite task completion as well.

Related Article: Conversational AI Needs Conversation Design

Think Like a Customer

“Set up your AI to respond to questions that customers would ask,” said Donna Fluss, president of DMG Consulting. "A traditional IVR uses a logic tree and will typically go through several questions before the customer gets the answer for something as simple as “how to I file (an insurance) claim." Rather than asking four to five subsequent questions, as a traditional IVR would, a conversational AI system will go straight to the answer.

Ideally, that is the case no matter how the question is asked. However, that is only if the conversational AI is sufficiently trained and continuously tweaked using machine learning, Fluss said. So the first time the conversational AI is asked the same question in a different way, it may refer the call to a live agent, but once trained on the new way of the question being asked, the conversational AI would respond correctly.

Fluss added that the conversational AI should use the same repository of information, regardless of the channel used, so it provides the same response with each channel.

Related Article: Why Your Approach to Chatbot and IVR Projects Is All Wrong

Learning Opportunities

Circumvent Dead Ends

"Nothing is more frustrating than reaching a dead-end or being told to dial a different number or use a different platform,” agreed Mitch Mason, offering manager for IBM Watson. “By integrating this technology more deeply into business operations, systems using conversational AI can seek help from human agents when stumped, without deferring to an entirely separate help line.”

Businesses should offer multiple, alternative options when the initial option doesn’t seem to be what the customer is seeking, Mason added. “Conversational AI systems should be developed to offer a similar level of flexibility while communicating with customers. Even if a virtual assistant gives the wrong response, the ability to offer multiple suggestions provides your customers alternative paths forward, rather than hitting a frustrating dead end. Disambiguation features can further reduce the risk of creating a dead end for your customers by prompting the AI to clarify the meaning and intent of messages before responding.”

Mason said conversational AI should focus on real customer problems, using real world data, from historic conversations or even interactions between customers and service agents. Investing time here is critical to making the best, most personalized experience possible for customers.

Related Article: How to Make Conversational AI Smarter

Should Go Without Saying, But Use Quality Data

“The most important element of effective conversational AI is large volumes of quality data,” said Jen Snell, vice president of product strategy and marketing at Verint Intelligent Self-Service. “As many companies have realized, simple point solution bots may serve as a quick fix but can be detrimental to the customer experience and do not scale to more mature, strategic solutions. Quality conversational AI requires access to quality data and the ability to improve solutions with that data to achieve better context, understanding and outcomes. If your conversational AI isn’t able to make sense of your data and isn’t always improving, it’s failing your company and customers.”

Related Article: Natural Language Processing Is Hitting Its Stride

Avoid Robotic Responses

“Infuse empathy and humanity into bots’ framework: The tone and cadence of a chatbot’s communication flow shouldn’t feel robotic,” said Vikram Khandpur, senior vice president of product partnerships and integration at Sinch. “Chatbots are often the first point-of-contact for customers, so it’s critical that their tone is clear, consistent and empathetic to maximize the customer experience. Utilizing an authentic, personalized tone when developing a chatbot’s conversational design can help to mimic the assistance of a service agent without the long-waited connection times.”