<img src="//bat.bing.com/action/0?ti=5739614&amp;Ver=2" height="0" width="0" style="display:none; visibility: hidden;">

Around here, we take birthdays seriously.

Sofia birthday

Which is why we’re celebrating Sofia, our AI personal benefits assistant’s birthday all month long. Over the last four years, we’ve all become extremely attached to Sofia, and warmly call her a fellow Solver. She works alongside our clients and member services team to help connect our users to the right benefits information at the right place and at the right time.

Sofia is not just a chatbot, but a personalized benefits concierge.


Our team has come a long way with developing her skill set so she can understand user inquiries and respond to deliver a streamlined, empathetic benefits experience.

To get more insights about how AI is innovating benefits, we sat down with Sony SungChu, Businessolver’s SVP of Science and Innovation (aka, the “father” of Sofia).

You just celebrated Sofia’s fourth birthday. What are you most proud of?

It’s been such a journey working on and with Sofia. I am most proud of seeing Sofia help our customers with more and more types of inquiries and empowering users to communicate with Sofia the way they want to. For example, using expressions they are familiar with, using verbal communication not just text, and having Sofia accessible on their mobile app as well as call center phone lines are all extremely valuable innovations in her development.

How have you further developed her knowledge base?

In terms of increasing Sofia’s knowledge base, at first, we started with COBRA and dependent verification. Now we are very close to reaching feature parity with our Benefitsolver web application. In fact, customers don’t even need to click around the website to navigate Benefitsolver anymore; they can simply ask Sofia and she will help them get to the right place. She can even help them complete a task like uploading a document or verifying a dependent. Also, one of Sofia’s most popular skills is her ability to educate. People ask Sofia about benefits terminology like, “What is a deductible?” More importantly, during the pandemic, Sofia helped inform millions of users with information and news about COVID-19 workplace policy and vaccines.

Why is empathy important to the development of AI and how does this approach separate Sofia from other AI technologies in the market?

We obsess over our users’ needs. That inherently motivates us to think about how they want to interact with their benefits. Benefits are very personal, but they are also extremely complex. We want to create a personalized benefits experience that is both streamlined and easy to use so that our users can get the help they need and elect the right benefits to keep themselves and their families healthy and happy.

We are constantly looking to refine that personal interaction. The way we teach Sofia new skills comes from our desire to invent and simplify each user’s unique journey as they learn about their benefits, elect them, use them, and make decisions across that spectrum.

This above anything else guides Sofia’s design, research roadmap, and functionality. Whenever we release new features, we are already critically thinking about the next milestone and how we will push the envelope again in our next iteration to help our customers reach new heights.

How has the market reacted to Sofia? What are other organizations trying to do that you see as a direct correlation to what Sofia is trying to do?

In the market, there is certainly more awareness on the differences between a chatbot and a personal benefits assistant. Whenever we describe what Sofia can do to help our clients and their employees, it becomes clear that Sofia can do way more than just answer questions. Sofia can help you do things like update your information, recommend plans to enroll in, and check up on the status of transactions for you. That’s way more than a simple chatbot.

One of the fair criticisms of AI is that they can’t or it’s difficult to develop them with DEI in mind. Is that a concern for Sofia, and if so, how are you managing for it?

We continuously invest R&D time to improve on our processes to identify and deal with potential biases in Sofia. In general, we look at four types of biases:

  • Reporting bias, which happens when we don’t consider every cohort based on use cases.
  • Selection bias, which occurs when our data underrepresents certain groups.
  • Extrapolation bias, which happens when we over generalize and assume that what’s true for one group applies to others.
  • Implicit bias, which occurs when we train AI based on our own assumptions. Like assuming everyone likes science jokes.

It takes continuous analysis of our user base to be able to safeguard against these types of biases; but this process is part of our dedication to our clients and users. We want to first seek to understand who uses Sofia. Then consider how they want to use Sofia, not how we want them to use Sofia. This process ensures that the bidirectional communication between our users and Sofia is considerate and aware of the diversity, equity, and inclusion our clients and users expect to experience.

What is the future of AI and benefits? Where do you see this technology going in 20 years?

We believe that customer service and virtual assistants are on course to converge. The greater the convergence, the more helpful Sofia is. Essentially, we can imagine a future where virtual assistants will eventually be the primary UI in many applications, because an application like Sofia should be able to simplify and enhance the way we interact with systems.

For example, if I want to make a payment, instead of going to a website and navigating to a payment page, I should be able to say, “Sofia, make a payment.” And boom, payment done. Based on my information, she should know what the context is, whether it’s a COBRA payment or another type of payment. Then from there, she can help me make the payment without me ever having to click around Benefitsolver.

This is particularly true of systems where the processes that have the end user as an actor are complex, as is the case in benefits administration and health care.

Natural language processing (NLP), computer vision, and information search and retrieval give us a chance to do something like the example above about making a payment. Ideally, we will have the ability to communicate with systems using our natural language and behind the scenes, Sofia has access to a set of tools that help her serve up the right content or provide the right service.

For example, these tools could be using bidirectional encoder representations from transformers (or BERT) as a way to find relevant information in documents or computer vision to classify documents. Or using optical character recognition (OCR) to read information and provide feedback or access an algorithm for making a recommendation.

Another way to think about this is that virtual assistants are the UI to the set of machine learning tools that power decision making, workflows, and information sharing between users and systems.

In the short term, we want to introduce more personalized services (meaning we collect data about you—this is the explicit data) and more intelligent services. To do this, we can make inferences using our models for things like recommendations, prescriptive action, sentiment analysis, and so on using implicit data or a mix of both.

Watch an interview with Kristen Hahn and Bruce Gillis below for more insight into Sofia. 

HubSpot Video

Want more blogs like this sent directly to your inbox? Subscribe below. 

New call-to-action

View all Posts by Rae Shanahan