by Clint Boulton

What makes a great chatbot? Laser focus on customers

Feature
Mar 19, 2020
Artificial IntelligenceDigital TransformationIT Strategy

Chatbot success hinges on the right NLP algorithms and regular upkeep, experts say. Hereu2019s how to ensure your chatbot initiative pays off.

artificial intelligence / machine learning
Credit: Thinkstock

From restaurants to transportation, chatbots that mimic human speech while facilitating tasks on behalf of humans are seemingly everywhere.

Twenty-five percent of customer service and support operations will integrate some form of virtual customer assistant or chatbot technology this year, up from less than 2 percent in 2017, according to Gartner. Processing customer requests on websites, mobile apps, consumer messaging apps and social networks are among the top uses, the research firm says.

Chatbot success hinges on a deep understanding of your customers, says HyperGiant CEO Ben Lamm, who built chatbots for the likes of TGI Fridays, Whole Foods, and Shake Shack as the CEO of Conversable, a maker of chatbot software. The technology must be designed and deployed with clear customer experience mandates, goals and key performance indicators that show value for the user, he says.

“Whether that’s more convenient purchases, streamlined conversational support, or elevating your experience of a live event, the chatbot needs to take the customer experience to a new level of engagement,” says Lamm, who sold Conversable to LivePerson in 2018.

Here product owners who’ve built chatbots share the secrets of their success, including precautions companies should take when building them.

Train chatbot conducts customer experience

Improved customer engagement is afoot at Amtrak, where intelligent virtual assistant Ask Julie offers convenient access to information from Amtrak’s website to the transportation giant’s more than 30 million annual train passengers. The experience is catching on, with more than 156,000 people trying Ask Julie each month.

One reason is that the software, provided by Verint’s Next IT consultancy, operates with a better working vocabulary, says Allen Sebrell, Amtrak’s senior manager of digital intelligence and strategy.

Sebrell attributes Ask Julie’s success over previous iterations to vastly improved natural language processing (NLP), which breaks down words and concepts in a sentence to improve machine reading comprehension and intent processing. Simply put, Ask Julie understands much more than her Amtrak predecessors.

“The error rate, or the ‘I don’t know rate,’ is under 5 percent, but it used to be pretty high,” Sebrell says. “You have to make sure the NLP capability is on point because if you don’t have an engine that recognizes all kinds of questions, you’ll never get close to delivering the right message.”

Amtrak’s chatbot team regularly refines and augments Ask Julie with fresh content, ensuring a consistent user experience even in the face of staff turnover and attrition, says Sebrell, who adds that ensuring senior leadership and other stakeholders understand the value proposition for investing in an intelligent virtual assistant is key.

To improve Ask Julie, Sebrell and his team review raw conversations and “I don’t know” responses to Amtrak’s 18 major topics, which are subdivided into over 60 groups of data, comprising more than 1,000 “knowledge units.” Eventually, Ask Julie will enable passengers to purchase train tickets, not just ask how to do so.

A virtual flight planner

Colombia’s national airline Avianca offers Carla, a Facebook Messenger chatbot that enables passengers to access information about flight status and weather forecasts, locate luggage, request seat changes and check-in from their smartphones. Customers can also ask Carla about ticket refunds and use it to provide real-time feedback to Avianca’s customer service.

“Carla was designed to help passengers answer some of the more basic questions versus trying to do everything for them,” according to Rob Harles, managing director at Accenture Interactive, which co-developed the chatbot with Avianca.

Carla runs on Amazon Web Services, which enables it to scale easily during peak periods without affecting customer experience, Harles says. Since launching in 2016, Carla has racked up tens of thousands of unique users and fielded millions of interactions.

6 tips for building a better chatbot

To help you understand the basics of how to get started building a chatbot, Conversable outlines its process in the following broad strokes:

  • Design the conversation:Identify the best use cases in customers’ existing operations and write out the conversation flows.
  • Build the conversation:Think of this as the version 1.0 of the final product. Test-drive to make sure the experience is up to par, and revise conversation flows as necessary.
  • System integration:Use webhooks to make sure the data needed for each conversation flow is available. For example, if someone wants to know the price of a product, or how many calories are in a menu item, you will need to be able to pull in that data on-demand during the conversation.
  • Learning:Continue to train and tune the chatbot through supervised machine learning, in which the chatbot’s algorithm is trained to infer solutions based on examples of desired outcomes.
  • Expansion:People often leap from one topic to the next during a conversation. Identifying a relationship between one conversation and another and linking the flows saves customers a lot of time and ensures the chatbot can stick with the user.
  • Advanced AI: Continuous improvement is key. As more data is collected, incorporate business rules to make the conversation more intelligent.

Ultimately, a great chatbot requires thinking through the business use case and customer journey and a willingness to course correct when and where necessary, Lamm says, adding that you can’t simply wave an AI wand at chatbots and watch them go.

“I don’t care how ‘smart’ your AI is,” Lamm says. “AI is not magic and it doesn’t mean you can skip the usual rigor applied to core business functions and IT projects.”

Chatbot pitfalls

Not every chatbot proves successful and experts offer examples of common challenges that can thwart chatbots.

The language laggard. A chatbot that receives poor language training will not be able to answer customer questions because it won’t recognize the words or shorthand your customers use, or the meaning of certain words used together, says Forrester Research analyst Charles Betz. “Language is complex and having a system that can account for all elements of your sentence and context takes far more steps than we immediately think of when breaking down a sentence,” Betz explains. “Training must compensate for this complexity.”

The ocean boiler. Chatbots that are designed to work anywhere from customer support for giant product portfolios to e-commerce don’t serve businesses and their customers well, says Lamm. “Narrower applications of conversational AI ensure the experience is accurate, consistent and scalable,” he adds.

The rush job. Some clients rush out a chatbot because it’s cool and because they want to automate something to avoid dealing with customers, Harles says. To check against this instinct, Harles asks clients to take a step back and understand the fundamental pain points they have and the discrete tasks they are trying to accomplish. Is the task best performed by algorithms, machine learning software or a good old-fashioned human?