Artificial intelligence applications have been depicted in films and popular entertainment for years, ranging from Stanley Kubrick’s HAL (Heuristically Programmed Algorithmic Computer) in “2001: A Space Odyssey,” to the eerily believable “Her,” in which Joaquin Phoenix falls in love with an AI virtual assistant.
The current reality for the retirement plan industry is a little less exciting: AI applications are being used as tools to improve productivity. AI technologies allow companies to automate repetitive processes, interact with plan participants more efficiently and analyze their plan-related data more effectively. These uses lack sci-fi buzz, but they illustrate how AI is gaining traction in the plan industry.
Language processing applications like ChatGPT and Bard have received recent publicity because of that technology’s ability to provide well-organized responses in plain English. Those capabilities can work well in participant-facing applications, where inquiries often follow identifiable paths. While chatbots may seem ubiquitous online, they can result in frustrating dead-end interactions with systems that do not provide the desired results. On the other side of the coin, AI is helping some companies avoid that problem and get better results from chatbot technology.
Alight’s Ask Lisa
Geoffrey Peterson, vice president of data and analytics at Alight Solutions, which is based in Lincolnshire, Illinois, says the company launched its Ask Lisa chatbot in 2017. To date, the system has supported more than 26 million interactions across Alight’s client base while maintaining a natural language understanding of more than 95%, an industry-leading metric, he says.
Ask Lisa deploys a symbolic natural language model for question intent detection, Peterson explains. The model consists of more than 3,500 concepts that serve to deliver responses across questions on health, wealth, human resources and additional benefit-related queries.
“In addition to the symbolic model, we are continually evaluating new AI technologies such as GPT-based large language models to address additional business opportunities, such as enhancing our call center with AI-enabled product enhancements to help elevate the experience for our clients and their employees,” Peterson says.
The software underlying Ask Lisa is based on a combination of Alight’s in-house technology and collaboration with vendors. Peterson, who joined the company in February, explains the Ask Lisa symbolic model is licensed through an external vendor, but the content, tuning and performance-management is based on Alight architecture, data and capabilities.
Ask Lisa’s adoption varies across Alight’s clients, says Peterson. “For example, some of our clients implemented Lisa when we went live in 2017, while others took longer to implement, as they wanted to analyze each response that Ask Lisa offered..”
Some Alight clients are raising concerns about large language models like ChatGPT, Peterson adds, but “any new large language model-based capabilities used to enhance Ask Lisa will only be released once we are comfortable that they perform well within the same quality and performance framework [we currently have].”
Voya PAL
In November 2021, Voya Financial launched Voya PAL, an intelligent customer service chatbot. According to Rajat Kalia, Voya’s chief technology officer for workspace solutions, the chatbot leverages a combination of machine learning and conversational artificial intelligence, which is enabled by natural language processing.
“As more customers engage with this digital assistant and as it continues to consume more data and learn more scenarios, it’s essentially becoming smarter, and it becomes even more in-tune to the customers’ needs,” Kalia says.
Customers can choose to work with PAL either with or without logging in. Those who choose to log in have access to additional capabilities, such as making account changes and getting assistance with certain transactions. “The way they interact with it feels very similar to how their experience would be if they were actually talking to a live agent, over the phone or through a chat,” Kalia says.
PAL is producing the desired results, according to Kalia: From its launch in November 2021 through year-end 2022, participants interacted with the system more than 190,000 times. In PAL’s initial months, the interactions produced a resolution rate of about 70%, which increased to 74% by year-end 2022. The chatbot has been picking up momentum and gaining popularity through April. “In the first four months of this year, we’re just shy of over 125,000 interactions and we’ve already seen the self-service resolution rate increase to about 86%,” Kalia notes.
Voya also plans to use PAL with its customer service agents, who will be able to engage with customers while the chatbot is working in parallel to navigate systems and locate desired information. This will “bring together all of the information that the CSA needs to help quickly and efficiently resolve the customer’s inquiry,” Kalia says.
A major challenge to models like PAL is the time required, Kalia adds. The training period depends on the technology’s usage, how much data the model consumes and the desired accuracy level. Because PAL is a supervised machine learning model, training can proceed more quickly, and PAL’s performance results are in line with expectations, he explains.
RiXtrema’s 401kAI
AI technologies for plan advisers are also coming online. New York City-based RiXtrema Inc., a fintech company focused on the financial advisor market, recently launched 401kAI, which is designed to increase the efficiency of advisers’ plan research and business development efforts. 401kAI uses a combination of ChatGPT and RiXtrema’s algorithm.
401kAI “saves a ton of time for advisers in two tasks that are ubiquitous in every practice,” says Daniel Satchkov, RiXtrema’s founder and president. “One is reading and understanding, comprehending—basically summarization of data. The second one is communication. Communication could be writing emails or creating a calling script.”
Satchkov cites the example of an adviser researching a plan being considered as a prospect to approach. Analyzing the plan’s public data takes an hour, at least. 401kAI has access to more than 850,000 plans and can review a plan’s information and summarize the most important points in a one-page document instantly. It can then use that information to create customized emails and call pitches.
Interest has been strong since the company hosted several introductory webinars in mid-May. In the first two days after the launch, RiXtrema received 232 demo requests, a number they usually receive in six months.
“While you need to cut through the hype of this AI thing, you do need to realize that it is a very powerful machine that can save you 90% of your time on many mundane tasks,” says Satchkov. “It can take away the least interesting part of your work and make you work on the actual human side, on the interaction, on communication, on meeting clients, on working with clients.”