Sedgwick, a third-party insurance claims management provider operating in 80 countries, receives about 1.7 million pages of digital claims-related documents a day. The documents then go through an arduous vetting process by examiners who must decide whether they’re valid and how they should be handled.
One claim can take weeks to adjudicate.
In April, Sedgwick unveiled a generative artificial intelligence (genAI) tool called Sidekick to help with document summarization, data classification, and analysis. Sidekick uses OpenAI’s GPT-4, giving the company “an unlimited number of large language models to be created for varying purposes.”
In July, Sedgwick piloted the first application of the tool in its production environment, automating processes in a bid to improve employee efficiency and bolster customer satisfaction. To date, the genAI technology has combed through 14,000 documents, and has been “shockingly good” at accurately spitting out summaries.
Leah Cooper, global chief digital officer at Sedgwick, led the genAI rollout at the company and not only built a team to fine-tune the genAI applications but also worked to educate employees outside of IT. Their ideas on the potential uses for genAI has stirred new excitement at the company, she said.
Cooper spoke with Computerworld about how companies can find employees best suited to work with genAI and what they should do to properly implement the technology. An edited version of that conversation follows.
How have you addressed the rollout of AI with your team? “We’re trying to wrap our heads around a digital strategy, and I keep telling them, ‘It’s OK. We’re going to baby step to a solid digital strategy.’ I have to keep reminding them to bite off small bites; you don’t have to try to eat the elephant [all at once].”
What does Sedgwick do? “We have 134 solutions, everything from workers comp and property…to employee absent leave. We, as a global company, cover so much other than just being a third-party administrator of claims, but that’s a big part of who we are. Claims go across people, property, things.
“We are the world’s largest provider of claims management. We’re the provider of technology-enabled risk benefit and integrated business solutions.
“That’s the nice way of saying we help businesses and their employees through some of the worse times of their life, and we do it with care. That’s our approach to everything we do; it’s people first.”
What is third-party claims administration? “At the end of the day, over 20,000 people call us because something negative has happened in their lives. Basically, 20,000 people a day pick up the phone and say, ‘I need help.’ How can we help facilitate claims management of this. It might be on the casualty side. It might be on the damaged property side. It might be worker’s compensation. Someone has been hurt on the job, so it might be their manager calling in to say this person was hurt in this manner; let’s get this person’s claim under way.
“We can help them through that entire process…to make sure they get paid while they’re out of work. This is the kind of administration services we provide, in addition to dozens of others.”
There’s probably not a consumer alive who’s not been frustrated at one time or another with customer service. Tell me about the problem you were facing with your customer service reps. “One of the things we want to make sure from our colleagues’ perspective, is give them the best possible tools to support them in doing their job. If we have the opportunity to introduce digital solutions that happened to be AI-enabled, that’s even better. That means we’re finding new ways to support our colleagues administering claims or answering the phone in ways we’ve never been able to do before.
“What we hope to be able to do with some types of AI, like generative AI, we hope we can support these colleagues by taking the administrative burden off their desks by automating certain processes and gaining the efficiencies and customer satisfaction from doing that.”
“That’s one of the things people have a big concern about; if you turn [genAI] on with a real business process, how do you make sure that it’s doing it right. Well, you basically grade it over its use in 10,000 instances. Is it getting an A+? Is it getting a B-? Right now, we’re in A+ territory.”
Can you explain why and how you chose to roll out generative AI? “Back this time last year, when no one knew a whole lot about generative AI, something came out called ChatGPT that made artificial intelligence very personal to millions of people in a matter of days. It was exciting because people who’d never had access to this kind of AI now had it on their phones in real time and they can see the benefits of it.
“People are calling generative AI an ‘iPhone moment’ because the iPhone changed the way we communicate. I honestly believe generative AI will have a similar impact in how it transforms the way we do business. And it’s super exciting.
“That’s what made us sit up and pay attention. We said, ‘Is there a real use case for this technology at Sedgwick? And, more importantly, can we do it and still adhere to our data privacy policies and keep our digital documents in house?’
“We never want it to leave our ecosystem. That’s the gold standard. You must be able to keep it in-house from a security perspective. So, those were the first two things we set out to discover…as we looked at building our prototype.”
What are your use cases? “The first use case that came up was in document summarization. It was low-hanging fruit, but…low-hanging fruit is still fruit, so you don’t pass it up. It can take a multi-page document, and you can ask it to summarize it and give you what you’re looking for from a business process perspective.
“So, we set out to engineer a use case for summarizing the medical documents for the claims process. We were able to basically get inside the head of an examiner to say what is it you look for in these documents and clone that through prompt engineering and come back with a result…that was as good, if not better, as a human being could do — and do it in a matter of seconds. That was our first use case and we began that pilot program in our production environment in July. So, I feel we were a little ahead of the curve in terms of a practical application.
“At the same time, we were still learning. So, we developed policies around the responsible use of generative AI, and how we would take every effort to govern against the abuse of AI or results that contained bias. All of those things should be part of the thought process as we make something operational.
“So, we rolled out the use case and on average it was taking eight to 10 minutes of relatively short documents – six pages or so. We’d take an average document…and say, what are the subjective, objective, assessment, and plan details we can glean from this in a summarized format? We found the result were shockingly good.
“We’ve had several months to tweak that now and get it as perfect as possible, and we’ve now processed well over 14,000 documents. As we do that, we ask our examiners to grade that [result] and ask, ‘What do you think. Did it get it right?’ So far, it has been overwhelmingly positive in terms of the feedback we get.
“One of the things we did is we kept the examiner involved during this evaluation process and we did that for a couple of reasons. Number one, before we begin using this on a regular basis, we needed to prove it was accurate and it was doing the job well. We did this specifically on the worker’s compensation claims we were supporting.
“Not only did it allow us to evaluate that accuracy, but we gained support for our product by creating faith from our end users that it was getting it right. That’s one of the things people have a big concern about; if you turn [genAI] on with a real business process, how do you make sure that it’s doing it right? Well, you basically grade it over its use in 10,000 instances. Is it getting an A+? Is it getting a B-? Right now, we’re in A+ territory.”
Are you operational with the document summarization AI for all of your examiners? “We have rolled it out to over 500 users internally. We have thousands of users. The next step to expanding our use is to say to the business side, ‘Do you feel comfortable with us automating this now? Can we create these summarizations automatically and put that information into the claims file?’ We’re anticipating we’ll get that sign off from the business side in the next couple of months, because what we’re doing next is focusing on the integration. I don’t want people to have to do this review. I want to be able to do that automatically; that’s how you drive the efficiency.”
What’s next on your AI roadmap? “If we can build an interface, first from our imaging system and then into our Sidekick product that leverages genAI — and then back down into our claims system by pushing the summary into the claims system — then we move into the next phase. First, we roll it out to every user in the worker’s compensation space, but we also set it up for an actual digital impact. By that, I don’t mean just the efficiency of it; I mean, can we turn around and say, ‘I just learned this from the document that was summarized?’
“Is there a way to provide guidance to our examiners using our data science program with predictive models using possibly decision triage rules and fraud detection checks. Can we give prescriptive recommendations to our examiner as to what they should do next? What that would mean is there’s no more queue, there’s no more waiting on an examiner to go through dozens of documents every single day. Instead, they walk in and say, ‘This claim is ready for the next step; this claim needs X.’ So, we change the way the technology interacts with the claims handling process. That’s where we’re focusing a lot of the attention in 2024.”
On average, how many claims do adjusters go through each day? “I can tell you on average across all our claims that go through our imaging system, we receive 1.7 million pages of documents a day. Now if you average that out to each being six or seven pages, it’s still a lot that we’re reviewing in support of out clients. Now the end-user impact — even if we did nothing except driving the efficiency of that summarization — is still huge in our game.”
So your next phase is building an interface —an interface to what and for what purpose? “There are really three big components [on our backend]. The first is our imaging system, which is where all of our digital documents flow into. Second, it’s Sidekick. That’s an application we developed internally that allows us to develop different [language] models that each have their own set of prompt engineering. So, for example, I can say, run the worker compensation documents summary. Or I can say, run the property estimation summary, or whatever it is. We can have an unlimited number of models throughout the company. Based on the type of document and line of business, we can get targeted summarizations back to our business.
“But, we have to know strategically where to shoot those because we have admin systems around the world. Some that do property adjustment, some that do workers comp claim administration, some that do leave of absence management, along with dozens of others. So with Sidekick, we need to know how to move from just a document summary to feeding that information back downstream to the right place. So, you’ve got the imaging system, you’ve got Sidekick and its connection out to ChatGPT and you’ve got the claims system.”
So, you’re using ChatGPT for your genAI system. Did you consider using another product or going with an open-source language model? “We actually looked at several different platforms when we were starting out. We already had a strategic relationship with Microsoft, because we had rolled out a cloud-based application a few years ago around the globe. That was called Smart.ly, and it enabled us to take in new claim information from any point in 48 different languages on four continents. Because we already had that Azure platform and environment set up, we were able to work with them to get ChatGPT installed in our ecosystem in a way that would not allow our own data and documents to go elsewhere.
“So we could just leverage what was already there. For us, that was a slam-dunk win, because we already had the environment set up to do this. That didn’t mean we didn’t build this in such a way to allow us to evaluate other genAI products out there on the market. I know Google Bart has some really impressive capabilities. I’m also waiting on a release of something called BioGPT.
“There are so many different models that are going to hit the market in the next 12 months. We had to think proactively and architect this in such a way that Sidekick could easily switch over and interface with different tools if we could bring those safely into our environment.”