When Jennifer Hewit was tasked with modernizing the service desk at Credit Suisse Group, one of the first things...
she considered was how to introduce chat as a new way for employees and service desk agents to communicate.
After doing some market research, Hewit, director of digital and cognitive services at the bank, decided to invest in IPsoft Amelia, a virtual agent built on machine learning. She believed the virtual agent would do more than open up a new channel of communication; Amelia could also automate simple, repetitive tasks, freeing up time for service desk agents to focus on more complex problems.
Eighteen months in, Hewit can now reflect on what it was like to get Amelia off the ground. Here are three things she learned along the way.
Start with a bot vision
Hewit started by breaking down the top 10 business processes that live agents solve in 10 minutes; she put those into Visio-like diagrams or flowcharts built on a yes/no construct. Then she tapped experienced service desk agents to do the work, but soon discovered that additional skillsets were needed to bring Amelia to life.
"As I went on this journey, I recognized that if I continue to only have service desk experts be the primary people building the brain, then Amelia would sound like an IT person," she said. "And if I put myself in the position of our users, half the time, they don't understand how we communicate about IT."
It was a moment of clarity that resulted in Hewit seeking out experts "who were more humanistic and user-design focused," she said. She contracted out a neuroscientist and computational linguist to work in partnership with the service desk agents and focus on design as well as Amelia's vocabulary.
The two contractors proved so successful that Hewit has since hired them on as full-time employees and is currently looking to bring in similar skills. But, she stressed, it's a talent investment not every company will need to make.
"I had a very strong vision around what I wanted Amelia to be for Credit Suisse. I wanted her to be more humanistic and to drive customer experience," she said. "When I talk to other people who are bringing Amelia in, they get it and why I did that, but it doesn't necessarily mean it always aligns to their vision or the personality that they're designing for their organization."
Find the right message
Hewit also began promoting Amelia by "proudly going around saying, 'We're bringing artificial intelligence to the bank,'" she said. But she soon realized the tactic wasn't working.
"There's generally some resistance around the things that people don't understand. And artificial intelligence is definitely one of those things," she said. "I knew that I was going to fight adoption if I didn't change my communications strategy."
Jennifer Hewitdirector of digital and cognitive services, Credit Suisse
So, Hewit switched gears and began to highlight how Amelia could solve a real pain point for employees by ushering in a new channel for communication: Employees could now chat with, rather than call, the service desk when they needed assistance. As a first point of contact, Amelia filters through chat requests, solving the ones it can and passing along the ones it can't to a live agent.
While employee adoption is important for any technology to take root, it's doubly important for AI technologies like virtual agents. The interactions between Amelia and employees generate valuable data that's used to refine, reinforce and continue training the system, according to Hewit.
Indeed, in an effort to build up Amelia's effectiveness, Hewit did something she said she wouldn't recommend to those building customer-facing agents. She rolled Amelia out knowing the virtual agent would only be able to understand what the user was asking for 23% of the time.
"No matter how much you think you've designed a business problem to accommodate how you think a user is going to speak to Amelia, it's not until you go live and you start getting into that real conversation that you're able to actually drive the effectiveness," she said.
Conversations between Amelia and employees are recorded. The data is fed back into the system and, with the help of machine learning, is used as a form of training. Five months after Hewit rolled Amelia out, the virtual agent was able to understand users 83% of the time; today, Amelia is 87% effective. "I don't think we can improve upon that much because, if you think about when humans talk to each other, we don't understand the context of the conversation 100% of the time," Hewit said.
Prepare to manage expectations
Despite the virtual agent's progress, not all 75,000 Credit Suisse employees have accepted Amelia with open arms. "The moment this channel comes alive, there is an expectation that she can do everything," Hewit said.
Hewit had several conversations with employees to explain why Amelia can't do more. She describes the virtual agent as a baby who still has to be taught. And she explains that, while an employee might rank a process as critical, she has to figure out how to drive the biggest impacts across the organization as a whole.
"So, I have very mixed reviews," she said. "I think she's doing great in the organization. I get lots of feedback on the capability, and people see the opportunities. But it's always a challenge managing human beings' expectations."
She's also cognizant of the fear that artificial intelligence is going to automate human workers out of a job. When talking to service desk agents, she explains that Amelia is an augmentation -- not automation -- tool.
"I call it AI for intelligent augmentation. So, AI for IA," she said. "And what I mean by that is AI is all about the machine; it's all about the technology. But if you talk about it in partnership with intelligent augmentation, it's humans and machines doing more together."