Sergey Nivens - Fotolia
The first foray that design software company Autodesk Inc. made into the world of customer service chatbots was a simple, two-legged robot character. Essentially, it helped customers get activation codes for software.
But the company didn't continue with the bot because it didn't have a persona that lived up to the Autodesk brand, which aims to be smart and innovative. In February 2017, the company created an AI virtual assistant called Ava that has deeper functionality
"If you're just trying to build an automation engine, you don't need to think about persona," said Rachael Rekart, director of machine assistance at Autodesk. "If you're trying to build a customer engagement tool, persona matters."
Enterprises find limitations of chatbots
Customer service chatbots have risen in popularity in recent years. They are seen as one of the few areas of AI that's truly ready for enterprise use and can be fairly simple to implement.
Several platforms, including Facebook, enable companies to build chatbots using their conversation engine, which means users don't have to have expertise programming conversational agents.
But these types of customer service chatbots can only get you so far. In particular, they tend to struggle the more that users try to engage in conversation with them. And as enterprises find chatbot limitations, some are looking to go a step further and build virtual assistants that can address more complex customer problems.
In a presentation at the IBM Think conference in Las Vegas, Rekart said Autodesk's AI virtual assistant Ava handles about 30% of the company's total customer service calls. The bot assistant handles about 45% of these calls end to end, 25% are just people chatting with the bot for fun and the other 30% have to be kicked up to a human agent. The Ava virtual assistant has a 90% customer satisfaction rate.
Rekart described this iteration as more successful than their initial Otto chatbot because the team invested time to develop a comprehensive persona. The Ava agent has a female face and voice, and the types of things she says line up more with the company's brand. Responses were initially written by the product developers, but Rekart hired a creative writer to go back over the responses to rewrite them using a single voice.
Zeus Kerravala, founder and principal analyst at ZK Research, discusses customer service chatbots.
The team also put a lot of effort into making Ava appear and sound diverse in order to reach out to as broad an array of users as possible, and to plan for a future in which the background of the typical Autodesk user might change.
"Invest in a cohesive persona out of the gate," Rekart said. "You're trying to create a cohesive experience and you're doing that through dialogue."
At consulting firm Ernst and Young, the human resources team is also looking to take an existing chatbot and turn it into more of an AI virtual assistant. The HR department operates a customer service chatbot that helps employees find information on how to do certain things in the company's self-service portal.
Steve Gill, the company's director of HR systems, said this simple tool has saved Ernst and Young millions of dollars per year by reducing call volume to the company's support center. But it could do even more.
He and his team are currently investigating how to turn this tool into a virtual assistant that can actually take care of issues for employees on its own rather than simply directing employees to resources where they can do it themselves. This way, if an employee wants to change his address or perform other simple HR tasks, he can have the agent take care of it for him without having to go into the source system.
AI virtual assistant tools surpass search
Virtual agents aren't just replacing basic customer service chatbots. At the IBM Think conference, David Almendros, AI director at Barcelona, Spain-based bank CaixaBank, said his team developed an agent to build upon an existing natural language search tool.
"We realized that chat is better than search because people talk more naturally and you can have a conversation," he said.
In 2014, the bank developed a natural language processing search tool using Watson on the back end. The search tool supported customer service agents working in the bank's foreign trade department. If a customer called up with a question related to trade, the agent could ask the search tool for information on the topic.
But the team wanted a better tool that could do more than deliver search results. The search tool delivered the results users were looking for 70% of the time, but Almendros figured a virtual assistant could do better because then users wouldn't have to try to match search terms to existing FAQ pages or internal policy documents, which the search tool typically used.
In 2016, the bank worked with IBM on a Watson-based virtual agent to field employee customer service questions. Today, the agent delivers more than 60% accurate responses, with the number climbing over 75% if you take out the ambiguous entries from users, like deposit or help with credit cards.
And the system will get better. When it is unable to satisfactorily answer an employee's question, the support call goes to a human agent. The information that the human agent ultimately gives the caller is then passed on to the AI virtual assistant. In this way, it gets trained on the fly using real-world, human interactions.
This setup, in which the agent handles routine and simple queries while humans tackle more complex problems, is a good example of how humans can work alongside AI virtual assistant tools to deliver higher value for users, Almendros said.
"This is not a fight between a machine and a human," he said. "It's intended to empower a human and reduce the low-value tasks. We want to remove those things from them."