How to Make Your Chatbot More Human-Like
Continuing Evolution of Healthcare AI Improves Consumer Experience
// By Althea Fung //
Chatbots — a computer program or artificial intelligence (AI) that has text or voice conversations — are redefining customer service, especially in healthcare. A 2017 Juniper Research study found that healthcare providers that use chatbots can expect average time savings of just over four minutes per inquiry, with average cost savings in the range of $0.50-$0.70 per interaction.
While the success of the chatbot is often calculated in saved nickels and dimes, a bot’s ability to provide a “human-like” experience is also an important indicator of success. One of the most significant challenges facing bots is how to make the conversation feel natural.
For Brian Gresh, president of Loyal, an Atlanta-based healthcare tech company that created the chatbot platform Guide, the first step in addressing the “human speak” challenge is understanding what the user wants.
“Chatbots use natural language processing (NLP), which trains computers to identify the intent of the user. NLP breaks that down into words and phrases to understand the person’s intent, and directs them to the right content,” Gresh says. “To do that, you have to feed different conversations and data into the AI and through that process train the AI to understand the intent of the questions.”
To understand what the end user wants, Loyal works with organizations like Adventist HealthCare to first identify what problems the bot is helping the end user to solve. “There’s real value in having a vendor that can take past learnings from other clients to help us get out the door,” says Richard Rinaudot, digital marketing director at Adventist HealthCare.
Adventist HealthCare, an integrated health system in the suburbs of Washington, D.C., began looking into chatbots as a way to provide 24-hour support for basic questions that the organization regularly receives.
Extension of Customer Service
“The role of our digital channels is to connect a patient to a doctor, location, or a service that can best treat them or their family in the most effective and efficient way possible,” says Rinaudot. “We saw the chatbot fitting in with that mission. Every day we get all sorts of inquiries through our digital channels, whether it’s via contact forms, or social media inquiries through our Facebook page, that can only be staffed during regular business hours. We saw the chatbot as an extension to provide that customer service 24/7.”
Building the Intent Library, Training the AI
Once Adventist HealthCare identified the purpose of the bot, Loyal worked on defining the data pool by which answers will surface. Loyal uses what Gresh calls an “intent library,” which matches words and phrases to meanings and intents. The intent library is used to train the chatbot to understand better what visitors are looking for. Over time, the bot will recognize the users’ questions to surface the correct responses and direct them to the right content.
“With the intent library, they have knowledge around frequently asked questions and typical interactions with the bot, so they can guide us to provide customized responses based on our operations. There’s real value in having a vendor that can take past learnings from other clients to help us get out the door,” Rinaudot says. He adds that Adventist HealthCare leveraged the interactions from contact forms, social media, and interactions visitors and patients have with frontline staff to field their chatbot response repository.
Though this process is somewhat similar to a search engine, Gresh says, chatbots should be more sophisticated in surfacing information.
“You can think of a chatbot as having search-like functionality, but the process is different. We are taking data and constantly feeding it into the model,” says Gresh. “The AI model learns based on the input. We supervise what it’s learning to see how it responds to the input. If the way it is learning is not appropriate, we can push it in the right direction.”
Rinaudot adds, “With any bot you can feed it 100 questions, but when you go live it’s inevitable there’s going to be that 101st question you’re not going to have an answer to. It’s a constant learning process, where we are surfacing questions, researching answers, and improving the coverage of knowledge and the extent of the knowledge on a weekly basis.”
Engagement with the Bot on the Rise
Adventist HealthCare soft-launched the chatbot in select areas of its website during the third quarter of 2018 to gauge the viability of the product before a full launch in the beginning of 2019. The launch of the chatbot on the site coincided with Adventist’s website redesign launch. Rinaudot says that engagement with the bot has increased more than 10 times since the full launch. Also, he says staff are turning to the bot to improve patient relations.
“I think a lot of departments are seeing the opportunity presented by the bot to impact their operation, whether it’s reducing call volume or providing that around-the-clock service to potential customers,” he says. Rinaudot says the system plans to extend the chatbot service to offer Live Chat during normal business hours. Gresh says Live Chat functionality can easily be made available to organizations that can field the service.
Identifying Content Gaps Streamlines Navigation
Adventist HealthCare hasn’t yet done customer surveying, but Loyal does provide post-session surveys. Gresh says some of his clients have taken the web analytics and user test responses to fill content gaps they didn’t know about on their site. Adventist HealthCare is tailoring the initial intent based on where on the site the user interacts with the chatbot. This method, Rinaudot says, is helping their customers complete tasks online quicker.
“If [the user] is on a medical group page, we may assume that they’d like to make an appointment. If they’re in the billing area, we assume they probably want to make a bill payment. What we’re able to do is seed these options on the initial launch of the chat to put them at the fingertips of the user,” Rinaudot says. “What we’re noticing is about 50 percent of our conversations are initiating based on those intents. I think we’re doing a pretty effective job navigating users to the tools and resources they’re looking for.”
Guide is a HIPAA-compliant platform that has the ability to collect protected health information to schedule appointments and complete transactions. But Gresh says they provide de-identified user sessions to clients.
“With this data, you can really start to understand what the users are looking for and what are their pain points beyond what typical web analytics gives you. Now you’re actually conversing with your consumer. We won’t always answer every question correctly, but people are willing to engage, rephrase questions, and connect,” he says.
Althea A. Fung is a digital content strategist and healthcare journalist. She is a senior editor at NewYork-Presbyterian.