When I think of GPT-4, it’s this amorphous LLM with more than a trillion parameters. Did you consider that when you chose it? The trend seems to be toward more targeted, domain-specific LLMs that require less processing power to train and run. "When we started out, ChatGPT was the best option. As of right now, I still do believe for our purposes — document summarization — it’s still the best in the market. I do agree with you that there will be some tools released that are more targeted…, especially as it relates to medical information. So, I’ll be interested to evaluate those as they come on the market to see if they’ll be a better fit for us. We have options, which is fantastic. Right now, ChatGPT is still the one that works very well for us."
ChatGPT is a cloud service, so how do you ensure that your data isn’t used to train other models for others to use? How do keep your data safe? "We have contractual and security obligations from Microsoft. They worked with us to ensure our installation of ChatGPT was locked down. The information we use in our instance of ChatGPT cannot be used as part of the public domain and they cannot use our learnings for any purpose other than to service Sedgwick."
How did you prepare your team for rolling out generative AI? Who did you bring onboard, such as prompt engineers, or was it all internal training of existing employees? "We trained. Segwick has a fairly entrepreneurial environment when it comes to our innovation and digital sector. I’m lucky enough to have lead that group for the last several years. Like we’ve done in the past, I took some really, truly tremendous souls in our department already and throughout our IT organization and said, ‘Are you guys up for learning something new?’
"When you tell a great IT person that they’re going to get to learn brand new technology and can think outside of the box to create solutions we’ve never done before — there’s a line for it, because this is why IT people get out of bed. This is what we love to do. We love to be creative. We love to find solutions.
"There was a learning curve. We didn’t know a whole lot about how to prompt engineer, and so we did partner strategically with Microsoft to understand how they approached their own prompt engineering. They came and did some virtual — they call it Level 300 — classes to help us prepare for that upskilling in-house. And it’s gone very well.
"In terms of the people, I’ll take a creative and curious person over someone with years of experience any day. The creativity it takes to build some of these solutions — it’s fascinating and it’s out of the box. And we’re challenging our people to notice how we’re using AI...and then say, 'What did we just learn from that?' So, people who are super willing to learn from what they’re seeing out there is what I’m really looking for.
"For example, earlier this year, I was doing my online shopping at a local grocery store, Publix, which I love. At the end of my experience, Publix said, ‘Hey, it looks like you’re making this. Do you want a recipe for it?’ I was like, whoa! What just happened. How in this world did they know what I was about to do?
"I took that to my team and said, look at this. Look at how this works. And, so we said, can we do something on the claims side similar to that where we asked clients, ‘Are you trying to do this?’
"There are so many different applications you can glean from just paying attention to the outside world. So, I want people who are just willing to pay attention."
How many people do you have on your AI team? How many were prompt engineers? How many were security folks, etc? "We have dedicated versus designated resources. Dedicate, we have eight people. All of them went through prompt engineering training, and not all of them do that on a day-to-day basis. Some are analysts. Some are programmers. Some are leaders. But then we also have designated contacts throughout the rest of the company who are part of our data science programs, our security teams, in our claims admin roles who also went through a lot of the training...to learn about the technology. We partner with those groups on a daily basis to make sure we stay in line with our corporate policies."
Where do you see genAI making the biggest impact in the next year, and the next five years? "In the next year, I think it will be efficiency gains. In the next five years, I think it will transform the way we handle claims and impact our customer service experience more profoundly. The low-hanging fruit, that’s efficiency gains. But what we can learn from it in terms of creating a white-glove customer service experience or providing more guidance to our teams in terms of what they can possibly do next in a claim, that’s what will transform the industry."
What does that look like? "...GenAI is just one tool in a tool box, but if you can combine that widely and tactfully with other things like data science programs and predictive modeling, possibly even design engines and external sources like fraud detection, then you can move the needle in terms of having a prescriptive interaction with your examiners.
"I can’t wait for someone to be able to walk in the door and say, ‘Hey, last night we got this document and it looks like there’s already enough information to approve this claim.’ Do a review of that and see if you can move the claim forward. That would be tremendous. How much work and red tape does that take off of an examiner’s desk?
"I think that’s a win both from the experience and efficiency perspective."
Have you been able to determine an ROI on your AI investments? "We have a lot of models. We should be able to do some ROI quantification after we have six months under our belts. We’ll be able to say if this is truly giving back the returns we anticipated. There will be a first phase ROI, where all we do is take the document summarization. Then there’ll be a secondary ROI, where we look to enable the examiners with more intelligent tools. So, the first phase ROI is what we’ll be able to calculate over the next couple of months."
What were some of the biggest hurdles you experienced in rolling out genAI? "I think we still have a lot of trepidation about how safe and secure genAI is because it’s brand new. A lot of our clients are saying, ‘Wait, I want to do a full evaluation before using this.’ So an understanding of a new technology is always the first hurdle [along with] making sure that everything is as safe and secure as we would need.
"The thing that wasn’t a hurdle was the enthusiasm for the use of this tool — not only from an IT perspective where people were excited about how we’d be able to use it, but our business partners were equally excited and said, ‘I have this idea or that idea.’ So we worked really hard to create an ideation funnel so that we could take all of these ideas from our 33,000 people around the globe and say, ‘Is that a real, legitimate use case. And, if so, let’s get it on a list.'
"So the enthusiasm is what I’ve been most impressed with and I think that’s the wave that will propel us forward in the future."
Are your employees concerned about AI replacing jobs? "Actually, no. That’s not our targeted use for this. While we want to drive efficiency, I think right now we are looking for ways to change the interaction with our claimants. We want our examiners to have more capacity to pick up the phone and interact with claimants in a way that they haven’t had time to do in the recent past because of demand. So how do we actually help them in doing their jobs — not eliminating them. How do we give them time to redirect the case into engagement. That’s where I expect to have a really positive outcome from the customer perspective."