Skip to content
The InsurTech Geek PodcastEpisodesTopicsGuestsBooksAboutSponsorContactSubscribe
Stan Smith
← Episode guide Episode 165

Advancing Workers' Comp Through AI Partnerships & Predictive Intelligence

Stan Smith
Gradient AI
Published Sep 4, 2025
Runtime 44 minutes
Host James Benham
YouTube Spotify Amazon iHeart Apple

Stan Smith

Hosts James Benham & Rob Galbraith are joined by Stan Smith from Gradient AI. Stan shares his expertise in advancing workers' comp through AI partnerships and predictive intelligence. Discover how Gradient's partnership with Terra gives adjusters earlier insight into high-risk claims, reduces costs, and improves outcomes for injured workers.

Watch here

The full episode

Press play for the whole conversation.

Sponsored by
Terra
1

Stan Smith from Gradient AI

Stan Smith from Gradient AI shares his expertise in advancing workers' comp through AI partnerships and predictive intelligence. Discover how Gradient's partnership with Terra gives adjusters earlier insight into high-risk claims, reduces costs, and improves outcomes for injured workers.

2

Why This Episode Matters

Episode 165 of the InsurTech Geek Podcast welcomes back one of its earliest and most forward-thinking guests - Stan Smith, the visionary behind Gradient AI. With over two decades of experience in predictive analytics and machine learning, Stan returns to dive deep into how AI is reshaping the workers' compensation space. In this conversation, hosts James Benham and Rob Galbraith guide listeners through an evolution of technology, business logic, and intelligent decision-making at scale.

This episode is a must-listen for anyone involved in claims, core systems, insurance innovation, or AI application in regulated industries.

3

From Supply Chains to Claims: Stan Smith's Predictive Journey

Stan's story begins not in insurance but in supply chain optimization, where he helped large organizations like United Technologies understand supplier risk. Following 9/11, Gradient's early machine learning models could predict supplier insolvency before businesses were even aware, marking a turning point in how data was used for foresight, not just hindsight.

This background laid the foundation for Gradient AI's powerful shift into insurance - especially workers' comp.

"We were able to predict which suppliers might go out of business - before the companies even realized it." - Stan Smith

4

Debunking Legacy Insurance Myths with Data

One of the most striking stories Stan shares revolves around the frequent flyer myth in claims - i.e., the belief that a claimant with multiple past claims is always a costly case.

Gradient's models showed otherwise. In fact, 70% of these claims closed quickly with minimal cost, despite the adjusters' gut instincts saying otherwise.

"High frequency is not necessarily a driver of high severity." - Stan Smith

These revelations underscore why relying on AI-driven pattern recognition - and not just human intuition - is essential for modern claims management.

5

Embedding AI Into the Core of Workflows

A core theme of the episode is the "next-best-action" framework. Instead of overwhelming adjusters with data or abstract predictions, Gradient integrates recommendations within the workflow. The system highlights what to do next (e.g., assign a nurse case manager) and - critically - explains why.

"If they have to look somewhere else, they're not going to do it." - Stan Smith

This makes AI not just informative, but actionable - a crucial differentiator in an industry constrained by time and regulation.

6

Partnership Spotlight: Terra + Gradient AI

A major announcement in the episode is the partnership between Terra and Gradient AI, aligning a cloud-native core system with an AI-native intelligence engine.

Why it matters: Terra's speed in deployment eliminates the delays often seen with legacy core system rollouts. Gradient's predictive insights enhance claim outcomes by supporting timely, informed decisions. Together, they aim to provide full-circle intelligence from the inside out.

"We want to be decision support, but in a really specific way." - Stan Smith

7

The LLM Wave: Disruption or Integration?

Stan weighs in on the rise of large language models (LLMs) like ChatGPT and Anthropic's Claude. While acknowledging their value - particularly for summarization and document interpretation - he emphasizes that verticalized AI still holds the key for real, industry-specific transformation.

"Technology moves faster every year - and that's just the new normal." - Stan Smith

8

Key Takeaways

Predictive analytics can outperform human assumptions in complex insurance scenarios. Next-best-action models embedded in the adjuster's UI improve both adoption and effectiveness. Terra + Gradient AI offer a powerful partnership between cloud-native core systems and AI-native decision tools. LLMs are powerful, but vertical AI with industry-specific data will lead the next phase of meaningful automation.

1
10 Ways to Improve Claims Management Practical strategies to enhance claims efficiency - ideal for anyone interested in the real-world application of insurtech innovation.
2
What is Reinsurance? A foundational read for understanding how insurers protect themselves - relevant to any discussion on regulation and risk.
9

Past Episode Recommendation

1
Episode 3 - AI, Predictive Analytics, Artificial Intelligence with Stan Smith Stan's first appearance on the podcast - a perfect companion to this episode showing his evolution and Gradient AI's growth.
10

Connect with Stan Smith

Stan Smith Gradient AI Podcast

The full transcript

▸ Read it · 8,449 words

Transcribed automatically from the episode audio, so some names and terms may be misheard.

James BenhamTerra, this week's episode sponsor, is revolutionizing the world of workers' compensation claims and policy management. Terra's cutting edge platform empowers insurers with comprehensive benchmarking tools, robust management systems, and an integrated ancillary services marketplace. Discover how Terra is driving innovation and efficiency in workers' comp at terra.insure. That's terra.insure. And we are back with another great episode of the InsurTech Geek. I'm James Benham, your InsurTech Geek. With me, my cohost, the most interesting man in insurance, the illustrious Rob Galbraith. Rob, how's it going, buddy?

Rob GalbraithIt's going. It's going, James. And, I actually got some exciting news earlier this month. It's been, I would say, five year journey, but, got word that my book is actually translated into Mongolian and will be for sale in Mongolia next year. So

James Benhamend of insurance as we know it is

Rob Galbraithgonna be in Mongolia. It'll be in Mongolia, both on a a physical copy and an ebook. Yes. And, you know, there was a a couple individuals that reached out to me when the book got released in 2019, and they said they wanted to embark on this. And it was something that, you know, we thought for a long time was gonna happen, and you know, the pandemic hit and other priorities and stuff came up. Right? So it was something I thought was just a a missed opportunity, and I actually had no idea that they had a a team of people that have been working diligently, kinda low key over the past several years and finally got it across the finish line.

So I'm thrilled. Well,

James Benhamit's it's the language of of the great conqueror. Right? That's right. Genghis Khan. Right? And interesting, you know, random facts about Mongolia, Rob. More horses than people in Mongolia. Wow. Okay. I could see that. I I There's 4,000,000 there's there's 4,000,000 people, 70,000,000 livestock. Wow. Including horses, goats, camels, yaks, but, horses are so dominant. And I actually have a friend who who flew to Mongolia a couple of years ago and did a great horse race across Mongolia. Oh, wow. It's it's, an epic journey. It's it's it's quite dangerous, actually.

Not because your people are gonna attack you, but because you're just enduring the elements. You know? And and there's a lot of accidents that happen on horses. And so it's it's very interesting. Two hundred and fifty sunny days a year, one of the sunniest places on Earth. And I don't know if you knew this, Mongolia, birthplace of dinosaurs. So it was, the oldest and first dinosaur eggs ever discovered were found in the Gobi Desert. So there's all kinds of, like, random stuff about, Mongolia. Really interesting place.

I've always wanted to go. I've been to China. I I I have not been to Mongolia, but, when I do, I know that I can bring some copies of your Mongolian book and teach people about the wonders of the end of insurance as we know it. That's right. That's right. Yep. It's amazing. That's right. Congratulations. That's great. And with us, a guy who does not speak Mongolian, He speak he he he speaks a a form of English known as wicked awesome. He's wicked smart. He thinks things are wicked piss.

He pox his. He's from Massachusetts slash New Hampshire, the great, the amazing, Stan Smith. Stan, how's it going? Good, James. Thanks for the introduction.

Stan SmithBy the way, I'm I'm a military brat, so I've lived all over. So I've been here for a long time, and New England boss is my home, but, wasn't wasn't born there. Wasn't born there. I don't have that accent. I I missed that opportunity. So that's that's You did. You know, when you're when you're from the United States military

James BenhamYeah. You you're you're from the military. You're not from any one place. That's right. One of my longtime teammates here at JBK, her dad ended up being the chief of staff of the air force, which is really cool. Really an amazing guy. He's now president of Texas A and M. And whenever I would hear people ask her where she was from, she would go, United States Air Force.

Stan SmithThat's right with the answer. That's right with I'm the

James Benhamfrom the United States Air Force. Right. Because, you know, she did high school in Virginia. She grew up in Germany. And then, you know, I mean, it was Yep. You know, they were Texans. They all ended up back in Texas, but that's funny. That's funny. Well, I still think of you as as wicked smart. So Thank you. Even if you don't park the cotton, the odds. So you you still wicked smart. We're we're I'm excited to have this conversation. For for all of our listeners out there that are regulars, and we've got we've got quite a few of you who, who listen quite a bit.

Stan was on episode three. When I first started this show, I only knew let's just say, I I had a a a more limited scope of of people that I knew in the space, but I was really honored to know two people, Skip Breitel and Stan Smith. And Skip introduced me to Stan. And so my very first episode was just me talking about the show. My very second episode was with Skip Brechtl. And I can just, you know, say very briefly, you know, he passed passed away back in February and was a a very dear friend of me and Stan, and he actually introduced me to Stan.

Yep. And then strongly encouraged a a friendship there in in the way that's that's that Skip did. He just he just kept harassing you by you call him yet? You working with him yet? You get you know, he would just keep asking you questions till it's all you finally, you do it. Anyway, rest in peace, Skip. We love you. Rest in die. And then right after that was was Stan. Now I had Skip on the show just a a a couple months before he passed away, and I'm about to have some videos that I released.

Stan's already seen them that I did an interview series with Skip, and so I'm excited about that. But, you know, it be because of that relationship, Stan and I have been working together for some time and have been have been you know, we we build core systems here at JBK and Terra, and Stan builds core AI systems that predict outcomes and recommend treatments and provide really detailed AI driven analysis, not just linear regression models, but but really rather rather complex systems that analyze tens and now hundreds of millions of records.

And then look for a correlation and causality. He's one of the guys, that and it's what what's been neat is, you know, we've been working together, I think, for seven years. Yep. And he's one of the guys that has has been delivering real AI for a long time, not just, you know, AI in name only where it was like an expert system disguising as an AI system, or is it it wasn't just an interface to chat GBT. Like like like, he was building and training his own models for a very long time with his team.

And so I'll just give, like, a really brief rundown on Stan because it's been quite a few episodes. You know? This is this is this has been a 100 and, I don't know, a 150 or so episodes since then. And so I think it'd be it'd it'd warrant a real a brief reintroduction. He he was educated at Dartmouth, as they say. Yeah. He he he is from the United States military, grew up everywhere. Then, you know, for eight years, he was VP of sales of Matrix one, then went on to VP of field ops at Agile Software, then was CEO of Open Ratings Inc, then six years as CEO of Semtech.

And then, then for 2012 to 2018, and this is where it really got interesting and important, he was a predictive analytics practice leader at Milliman. And while there while there, this kind of this idea spun up, and then, it spun out into this company, called Gradient AI. And so since 2018, Danau, he's founder and and CEO of Gradient AI, which is a really dedicated AI and analytics firm headquartered up there in Massachusetts. And so it's been really neat, tracking this journey. So, Stan, again, I thought I would kinda give people some of your a little bit of your background so you didn't have to.

Thank you. Let's let's just go back. We we talked about this a 150 episodes ago. I wanna talk about it again. That journey that you took through technology and software and ratings and and then into predictive analytics at a at really an an analytics firm, like, what was that about for you? You know, was this an intentional path through multiple disciplines that led you to this, or what was what was what was your thought there? I mean, sometimes in our careers, we tend to just take the opportunities that are presented to us.

Sometimes there's an intentional strategy. So walk walk me through the the story arc. So so the the intentional path you know, I was in the software business a number of times. This is, I think, my sixth startup when it goes back to numbers.

Stan SmithThe intentional path was when I I started this company, Open Ratings, back in about twenty five years ago. And the idea there was that in the supply chain, that everybody had automated a lot of things around it, but there was a missing link, which was what's going on with the supplier that's behind your supply chain. So many of these companies like United Technology would have 60 to 100,000 suppliers. They couldn't possibly know all of them, couldn't know the latest and greatest about all these companies, what's going on, what's going well, what's not going well, so forth and so on.

So it was a missing link. We assessed that with the market. And the value we're trying to deliver was, can we help these companies better manage their business by better managing their supply chain? Because if the supply chain doesn't deliver, a lot of these companies can't deliver their end product. And what came out of all that stuff was they all had lots of data. It was messy. It was incomplete. It was error prone in some cases, but they all had snippets of information. If you could figure out a way to pull it all together, you could actually put a very concise set of information from company A about how their suppliers were performing and then bring in companies B through Z and onward and build a very comprehensive view of how these suppliers were performing for multiple customers without sharing any of the customer information.

But what was their last transaction like with this supplier? Did they complete their transaction? Was it on time? Was it the quality there? And so forth and so on. And the only way to do this was a lot of database and things, but we also discovered this thing called machine learning. You know, when that problem opportunity was presented to me, the only way to pull this together was this term called machine learning, which was really the early stage of AI. It's one and the same, but it was more of a generic description, which is pattern recognition using math and large amounts of data.

And that's how we went to market with that company. And we were able to do really interesting things, ask data new questions. One of the new questions we asked, I think, was the most important moment of that company's trajectory was after nineeleven occurred. A lot of companies in the aerospace industry stopped flying planes. That meant a lot less consumption of parts. That meant a lot of the suppliers started becoming insolvent and going bankrupt, and that was a big problem for these aerospace companies. They came to us and said, Hey, can you take this performance data and other third party data and predict when these small suppliers might go out of business before we would actually realize they might be at risk?

And it turned out we could. So that's the power of AI, is you can actually take a dataset and ask it a new question, throw new data at it as well, which is what we did. We threw commercial credit in addition to this performance data. And it turned out that combination of data had the signal that was extremely accurate predicting way before these big companies would ever know it that this supplier might be at risk of ceasing operations. And so we became sort of standard best practice in the aerospace industry and some other heavy manufacturing industries because they were just really reliant on their suppliers.

So started my journey with AI, if you will. And when I went to the insurance space, I saw another market that had lots of data, but not always the best data, not always perfect data, not always the same format of data, but lots of potential information and insights, if you could just bring it together. And that was my thesis going in, was if I could turn that into actionable insights and very precise predictions, there's a lot of value I could deliver to companies in the insurance space.

And that was the genesis of of creating AI and starting there.

James BenhamIs it fair to say that in early days of pattern recognition that you were looking for correlation and causality? Is that what you mean by pattern recognition, or are you just looking for correlation?

Stan SmithBoth. Both.

James BenhamBoth. Yeah. Because you're looking for, like, what cause you know, what are what what contributing factors are around when this good or bad thing happens, and then Yep. Which ones cause it to happen. Right? Yep. Yep. So,

Stan Smithyou know, good good example. In the supply chain, in small suppliers, it was different by size of suppliers, but you could see a combination of if their credit score had a short but steep increase improvement, paying bills, you know, doing things to kind of get their credit score up, but their actual performance either was flat or somewhat declined. You can see a company that might be stressed trying to get a new loan. Whatever was going on, we realized there was a combination here. We could see performance combinations that would indicate for a short period of time maybe they're getting better.

But if that same supplier had a dip in their credit score, in other words, they tried to improve their finances, get a loan, didn't get it, whatever, and then they had to turn around and start paying bills slowly, if they saw a dip in their financial performance, again, based on other factors as well, not just this alone, but that was within the next ninety days, we could pick up that a lot of these companies would file. And that was something that nobody we didn't know that pattern existed.

We didn't know that combination of things existed, but that was something that came out of the data, became very predictive and something we could monitor across tens of thousand suppliers at any one time. Yeah. That's fair. I mean, we think we know what causes things just because of common sense. Right? Yeah. Like, we look at it and we go, well, that causes that. Yes. And I I can tell you what. Getting involved and you utilizing machine learning models

James Benhamhas taught me is that about half the time I'm right and half the time I'm flat wrong.

Stan SmithCan I you can I give you that that an example in work comp for that one, Sure? So my first client, actually, on our first engagement for work comp claims and trying to identify what they call creeping catastrophic claims, they look benign. Some people call them jumper claims, claims that just didn't look that bad, but would be in top 10% of highest cost claims later. The goal was, could we identify those claims early? But one of the things they told us that they already knew, knew, I'll use air quotes for that, is that if the claimant themselves was a frequent flyer claimant and their definition of that was five or more previous claims by that individual, that that claim would always be very expensive.

So they thought there was a direct correlation causation between five or more previous claims and high cost complex claim. And so when our models returned a series of claims that we thought were high risk and the vast majority were low risk, they went and looked at the low risk claims and said, You've got 40 claims in that low risk group that are all frequent flyer claimants. We think your model's incorrect. And so we did some data science work to see if the models had missed something or whatever, but it took us a week or so to go through that.

When we delivered the results, we showed them that in those cases, approximately seventy percent of the claims that had frequent flyer claimants, the model was pretty confident that those claims were not gonna be high cost claims, and they were not believers of that. But I had one fact to to use for them. And this we were only had been scoring claims for thirty days in total. Day one. So we're only thirty days in to scoring all their claims. And of those 40 claims that are freaking flyer claimants, half had already closed.

Half. So I had objective proof that those that half of those claims were not high cost claims. They closed in a few days for a few $100, and that was it. And they were floored. And I was excited because I thought this was we'd missed something. And it turned out the models were correct, and they're and and they're overfitting on that one variable, how many previous claims you had, wasn't as accurate as they thought it was. So previous claims wasn't the indicator for high cost claims in the future that they thought it was?

Yes. As and on a standalone basis. Is it an input? Sure. Does it play a role in some cases? Yes. But on a standalone basis, is that enough to tell you it's gonna be a high cost claim or not? Absolutely not. In fact, 70% of the claims turned out when we finally did some because this was such a a surprise to them, we did some work and turned out of seven 70% of the claims that had a freaking flyer claimant actually closed within the first I can't remember now exactly, but thirty, forty five days.

They it it wasn't actually anywhere near as important as they thought it was. Whereas they thought it was a key driver to

James Benhamhigh claim high claim cost. Yep. Exactly. So high frequency is not necessarily a driver to high severity? Nope. That was the core assumption they made that Yep. The model really proved out that it was wrong. Correct. It's interesting.

Stan SmithBecause you you would just in in your logical brain, you would think that it was. And their logic was when they looked at some of their worst claims, a lot of those claims had frequent flyer claimants. So it was they that's where they thought it was, you know, causation. They said, okay. We see it there. That must mean that's the end. And again, I think the way the human brains, we can handle, you know, some people say up to seven variables. I don't know about that.

My wife would tell you I can't handle Maybe two. Not even one. But the point being is, you know, they looked at, you know, high cost claims and said, what do those claims look like? And they tried to find the pattern of those claims to use to identify them before they got there. And that's the way humans can do it. This math, this AI can do it at a much, much higher scale than we Yeah. Well, it can look at every data element everywhere all at once.

And and never forgets never gets tired.

James BenhamYeah. Continue to learn. We tend to we tend to identify it's like the example that I used in our in our talk at WCI where, you know, in Star Trek, the red shirt guy always dies here for wear to red shirts cause death, but they don't yeah. Same thing. Or if they don't use their name, that guy's gone. Right? Yeah. He's gone. He's yeah. He's gone. That's right. He's crew member number two. Yeah. That's right. That's right. Rob, what you got? Yeah. Stan, it's great to have you on.

I I cannot believe it's been so many episodes,

Rob Galbraithfrom the the time you're on as one of the very first guests to to to now. So one of the early guests. I appreciate that. Yeah. And I encourage those, particularly those that are big fans of the show, to go go back, kinda listen to those that that early episode. I think it'll be really instructive. So, for those that don't necessarily wanna do that homework though, Stan, maybe you can just kinda give a a broader overview. What was gradient back then and what is gradient today?

How have you evolved? Particularly, think we've had maybe, like, a pandemic that's happened in then. We've had, like, you know Yeah. A little few things. Yeah.

Stan SmithSo, again, not it's not too much of a commercial, but as James said, we start this business started in Milliman. My thesis there was knowing I wanted to build a contributory business a contributory data based, you know, business model, meaning that clients will give me their data because AI needs data to train on. The Milliman brand would help me. And it turned out to be luckily, it worked out. And people trusted me because I was part of Milliman because they trusted giving their data to Milliman actuaries.

They trusted me with their data. But we were able to establish that relationship. I was the first person at Milliman to put all data in the cloud. In fact, my clients insisted that their data never touch a Milliman server because they didn't want anybody at Milliman, somebody that didn't know at Milliman, an actuary at Milliman that might be working for a competitor to ever see their data. So we were basically walled off and we were in the cloud, so nobody at Milliman was authorized to view it.

But it helped me build a very large dataset, tens of millions of work comp claims and work comp policies in that database. And so when we spun this business out in 2018, that was a key asset that I brought with me. So we left none of that data behind. We left no copies of it behind, and that came with me. But so since we rolled out, I was bootstrapped. James has a little bit of knowledge about what bootstraps means, and I know that he's bootstrapped a lot.

Bootstrapping can be effective for a lot of things, but it also can make you know, it's challenging in some respects. And we'd done okay, I'd say, as part of Milliman, but I wanted to grow faster. So as an independent company that started out in 2018, we've raised now three rounds of capital. We've raised just over, I think, dollars 88,000,000. We've grown a lot. We've gone from a handful of clients to well over 200. And we've also entered the group health market. So we were on the P and C when we started, and we've added group health.

And what's really exciting right now is we're finding ways to bring insights from group health into the P and C market in ways that nobody's done before and just find new ways to add value to our clients, actually bringing in insights that we think can help our P and C clients do more, do it sooner, do it better and and improve their results. So it's it's I think we're I'd like to say we're just getting started even though we've got we've had some good results. I think we're just getting started.

Yeah. I mean, AI, though,

James Benhamin the broader sense, large language models have turned the entire world's attention to a very specific form of AI. Yep. You know, LLMs. Now the the challenge is that the LLM providers are putting their stacking other AI models behind their large language model. So they're confusing people because they're using it as a user interface to access other types of artificial intelligence. Right? So now it can perform mathematical functions that aren't really the function of a large language model or using it as a user interface for it to be able to, you know, perform image recognition, image generation, mathematical prediction, etcetera.

Stan SmithHas has all the attention on AI been good or challenging for y'all? I think if I net it out, James, I would say it's good because I again, it just it's lowered the barrier to people saying, I first of all, if I don't understand what this thing's doing, I won't touch it. It. So they always assumed AI was a black box. Now I think what ChatGPT and these LLMs have done is understand people now understand better that it's trying to predict a series of words. That's I mean, at some level, that's pretty simplistic.

When it gets it right, it's pretty impressive. It can, you know, can edit a note you want to send. It can do some things I find to be very useful and valuable. At the same time, it's not perfect. And now, of course, the news loves to report all the things it does wrong, and some of those things are actually, you know, pretty you know, not very good, so you gotta be careful with that. But I think, overall, it's it's lower the the concern as long as people understand, where's my data?

What kind of models are you using? Can I see the backtest to understand how accurate this thing is? Then, ultimately, is it useful in what it produces? And I think that's always been the case, but I think now it's a lot more I think we're just having more pragmatic discussions with people who have, you know, chat GBT on their phone and and think about this stuff as more part of their daily lives. And I think and and, James, I think you and I have talked about this.

As they start seeing more uses of it around their livelihood, like taking large amounts of medical documents and summarizing those so they can quickly understand what's going on in a case. If this claim has been diary for thirty days, what's happened in the last thirty days, they don't have to go read all the documents that summarizes, okay, they've gone for two more PT sessions or, you know, they've had these new drugs dispensed or, in fact, they've completed a prescription and they're not asking for renewals. And there's lots of things that they can pick up on instantly and say, wow, That's better than I it's faster for me, more concise for me.

And I I think they start to look for more of those value added helps that can make their job simpler, hopefully more enjoyable because they get to do the the most important stuff, which is still make a decision on what I should do now that I have this information.

James BenhamYeah. You know, human in the loop allowing adjuster licensed adjusters and underwriters Yep. You know, to continue to to make the final call, I think, is really important in particular in a highly regulated state regulated industry like insurance. We've had to put a lot of stop checks in. You know, we have have been putting in document and text summarization quite rapidly over the last three months. And even there, we we pinned it as a pending note and then let a license adjuster review it. Right. And then they decide if it goes in with modifications or not.

We just don't insert summaries without into a note record. It's things like that that are that that are that are super important, and and certainly, I've implemented gradient before, and we use it as a total incurred prediction tool. We use it as a recommendation tool. But at the end of the day, the licensed human, they make the final call. Right? Correct. Which is an an important hallmark of of how, you know, you and I both believe this this type of technology should be used. So Correct.

You know, it's it's, it's important, I think. We recently announced a partnership between Terra and and Gradient. Certainly for us, building a core system was a daunting task, as you can imagine, you know, doing a ground up cloud native core policy and claim software. Yep. We we did we did it with 95% less money what a lot of very, very well funded startups have done. Yeah. Because we're bootstrapped, we're very capital efficient. We've been really working hard to deploy a great core system. But having the data and the capital to also build a really a really intelligent recommendation engine for treatments and and predicted rev you know, reserves and total incurred predictions and, well, contributing factor determination, that's another really big thing that's that's really important that that your system can do.

That was, you know, very daunting for us, and so we decided to partner. We recently announced it. Why is being partnered with the core systems I I just explained why it's important to us. Why is that important to you?

Stan SmithWell, so we think we think a couple of things. One, I think the core system folks, you're working with a client, a direct client. That client is is responsible for or is is reliant on your core claim system to help them get their job done. And, again, highly regulated environment. I I think I can't emphasize enough how many clients I've talked to that bought from a vendor that didn't really understand WorkCop, and the client did not understand how little the vendor understood WorkCop until their 50% or more through the implementation realized they didn't understand WorkCop.

That's a little late to the party. So that's one thing. I think cloud native is another thing. I think at the top of the show, you mentioned that we're truly AI native, really using AI. I think when you talk about being cloud native, there's a lot of folks in the industry that say they have a cloud instance of their product. What it is, is their on prem software have been ported to a version of some cloud instance. And it's not cloud native. It wasn't built to be multi tenant or a lot of things you can do in the cloud that makes things more scalable and more cost effective.

So to us, finding cloud native a partnership that's cloud native like we are, that's helpful. I also will say, and I've said this directly to you, James, I think the most important feature I think about what's going on with Terra claims and Terra as a company, and I have used the word nuts as a description of this, is that you won't charge your customer till your product is live. And as a company, some of our biggest challenges with some of our clients is we're waiting for them to get their core system turned on because our deployment is reliant on that being turned on.

And when they're one year behind or two years behind or more than that behind, it's challenging. So I think we are excited to work with somebody who's got the ability to get something turned on quickly and get the client live so we can actually roll out. And both of us, of course, start to helping that client. And I've got clients that are on their third core system that haven't gotten the first two live and may not get the third one live, but they keep trying because of this effort to get through that.

So bottom line is I wanna serve clients as as fast as much as possible. I think that's the Tara's view as well. And I think that's what that attracted us that value prop attracted us to this partnership. Yeah. And then, you know, it's it that's also what's challenging in the scenario you just laid out with some of these, you know, funds, groups,

James Benhamand and companies that have tried three systems as they've paid millions of dollars to, implement systems that never went live. And so it's very, very challenging because eventually, their the respective boards and executives, you know, grow weary and then go wary of all technology adoption. And so it it creates a it creates a lot of challenges for adopt we talk a lot about technology adoption on this show, a lot about adoption. And it is really hard to get adoption when other vendors burn people out, financially burn them out Yeah.

And emotionally burn them out by saying they're ready when they're not, knowing that they're gonna float their work on the implementation fees, and then literally are rolling the dice that they can finish core system functionality in time for the client to be live. Like, it's it's a really, really big challenge. So Yep. I, definitely understand what you're saying and and certainly agree with it. That's also why it's you know, as a bootstrap company, it's my neck on the line and my business partner's neck on the line financially and we're we're certainly willing to put put our put our word behind it.

So, I think that's you know, you can innovate all day long in technology but have a crappy business model. We'd like to have both. You know what I mean? A great business model and a great technology play. Rob, what you got?

Rob GalbraithYes. Stan, so following up, and James, obviously, you know, you're talking about the partnership. So feel free to weigh in as well. But I'm just curious, Stan, maybe you can discuss a little bit about what a, I hate to use the word, typical integration looks like, but maybe discuss some of the ways that you have partnered with clients. Is it via API? Is it through a core system? And then maybe talk about the uniqueness about your integration with Terra and what makes it maybe different or especially unique from your perspective than some of the other deployments that you've done historically.

Stan SmithSo Rob, great question. So in some of the cases with the legacy systems we are involved with, where they were built a long time ago In fact, we have clients still to this day, once in a while on a conference call, will say, Hey, do you know any COBOL programmers? Because that's what their stuff is built on. And we we've been telling them for years that we don't know anymore. They've all retired. We just but they're they're so desperate to find them. I I I think I don't know.

James, you mentioned that you can actually now use some of these these tools actually program in COBOL. So now you might that that problem may go away. But but the bottom line is it's been sort of challenging. So in some cases, we've had to just put a hyperlink in the system. So if they wanna use our information, they have to they have to link out, which is not what we wanna it's not the way we really wanna build this. We wanna actually have a a tight integration and as native as possible.

And I think to flip to the other side of how we see this working with Terra is that to have information that's important to the adjuster at the time they need to know about it, present it to them in the page they're looking at. So they're actually looking at a claim, looking at information about the claim, what's going on in this claim, what's happened on this claim. And there in that exact same panel is information as you may want to consider this next best action on this claim.

You may want to consider assigning, you know, it's an early claim, tell them a nurse you know, a telephone case management interaction with a nurse could be valuable. Tell them that they should consider other things with this claim or that they might. They should. Some indication that this might be useful, but maybe not even recommended yet. But have it there so they're not going somewhere else. They don't have to look someplace else because guess what? If they have to look someplace else, they're not gonna do it.

Our the data show that if they have to go hunting for information, they're they're they're not looking for new information that might make their job harder or just take more time because they don't have extra time. They just deal with what they gotta deal with and they move on. So as close as we can put it in the cockpit of where that adjuster's living and making decisions, that way they make that that they have that information right at their fingertips, Can look at it, agree with it, disagree with it, ignore it, whatever they want, but it's there.

It it gives them information. And I also refer to this industry right now. It's a perfect storm. A lot of folks are retiring. With all the remote work, there's a lot of turnover. So training and supporting these folks that you're bringing on is challenging. So this also reinforces training, best practices, things like that for our clients that can help that. But I think that's all available through this partnership with Terra because, again, Terra's got a newer chassis and and and, you know, has a a ability to deliver this kind of inter interaction,

James Benhamwhereas a lot of the core systems don't. Let's talk about next best best action because that's something that I don't think we talk about. We did a lot of talks at conferences about AI. Yep. But very few conversations around nest best action because, historically, software was designed to have an input screen, storage and processing on the back end, and an output screen. Yep. And most software that most people in insurance use is just data grids. Right? It's just grids of software made to look as much like Excel as humanly possible.

But, I mean right? This is the that's just the the god's honest truth. Yep. Well, the problem with that is that we're not actually telling the user the next thing they need to work on. Like, we need the human to be a human and use their brain. Software needs to do more than just feed lists to users. It really needs to tell you not just the nest act next action, the nest best action. On a claim, what's the nest best action? Okay. It's time. In ninety percent of cases, you should be bringing a nurse case manager on right now.

In ninety three percent cases, you should already be bringing a physical therapist in because there's a super high likelihood. Instead of just saying, here's your results, say, here's the results, and here's our recommended nest best action for you. And then most importantly, Stan, like you said, why? So talk talk to me about this because this is a major pivot in the way software works.

Stan SmithGreat, James. So in in the beginning, we didn't we were we we knew we wanted to figure this out. We just hadn't figured out. We started building models that had, you know, high, medium, low risk. And so what we're telling somebody, an adjuster, this claim is gonna be much more expensive than you think. Now we're telling you that at day five, day six, whatever. And our clients came back and said, Okay. We've seen the backtest. This model is correct. This claim is gonna be much more expensive than we expect.

But I don't have the tools at this stage. I don't have the information. Don't have the medical bills. I don't have all the information I normally have. By the time I realize a claim might be medically complex and other things, that tend to be thirty, sixty, ninety days old. And I had a lot more information for me to understand why. Comorbidity. Turns out they they smoke and they whatever. They had a lot more information to deal with. And so my early clients used this term, you gotta give me the now what answer.

I said, what do you mean? Okay. You've told me this. It's gonna be more expensive than I expect. Now what am I supposed to do about it? And we really didn't have we just didn't have the capabilities of bringing that to the surface back then. Today, what we do is we'll identify that high cost claim. We'll tell them what we think they should do about it, like whether it's PT, nurse case management, IMEs, other things like that. And then the why. Why? Because in a claim like this, if you apply that service, that intervention, whatever you wanna call that, here's the benefit that you'll see on this particular claim if you do it at this particular time.

And we're able to see, because we have all this data on lots of claims, that there's so many opportunities missed by putting these services on either later than optimum or not at all. In fact, I think you and I have worked with somebody who saw an opportunity to drive more value for the clients, it's a TPA, by using nurses more effectively, more consistently. And when they did their own tests of historical claims, they found they had a lot of complex claims that were medically complex and a nurse had never been involved at all.

And they just felt like that if a nurse had been involved early, that would have made a difference. Well, it turns out the data supports that. If a nurse had been involved, many of those cases would have been lower cost and lower duration. So the essence here is to be decision support, but in a really specific way that that service at this time would have this kind of value. So the what you should do and the why you should do it. And again, it still can be accepted, rejected, or ignored, but track those interactions and then use outcomes to assess, was the model correct?

If so, maybe that client starts to reinforce the use of that recommendation. If the model is incorrect, retrain the model. There's a lot of different ways to use that. But I think it is bringing an assistant to the side of the folks trying to make these decisions in a timely, consistent manner.

James BenhamYeah. That's fair. Rob, last question.

Rob GalbraithSo, I gotta ask you to kinda you know, you've been at the forefront of a lot of this stand for the last, you know Right? Couple I won't put a number on it. But so looking ahead, right, what's your vision for how AI is going to continue to transform insurance? And where do you hope Gradient makes the biggest impact in the years to come?

Stan SmithGreat. Appreciate that. Two part question. So one is the money that's being spent, you can see how many billions and billions are being spent for infrastructure and NVIDIA chips and all this stuff, and these guys are all gonna try and outspend each other. But what's always happened in these worlds where the market goes crazy and the money goes in, technology just moves faster. So I think what we're gonna see is rapid, rapid evolution and some big leaps along the way of all this technology. I think the LM was a big leap in an interesting way.

But the billions that are being fought over from a revenue perspective, market share, market cap, that's going to just create more drive. So LLMs are going to become small language models, be easier and faster to train them. It won't cost as much. Think there'll be some leaps in capabilities there, auto labeling of data. There's just all sorts of things that's going to create more and better use cases of AI for all of us. I think we all benefit from this large amount of deployment of capital.

And ultimately, that helps us. The more things that are mundane that we can just have done by an LLM or have done by a core system that are good, necessary, but sort of steps that are regular, let the core system do that. Let the LLMs do that. We want to be taking the information that most likely they don't have enough of in house or they have some, but they want more. They don't want to use just what they've got. Things like benchmarking, total incurred predictions. What does a claim like this typically cost?

Things that bring information from outside the organization, they still can use it, but it really helps them drive better outcomes for themselves. So that's where we see our play. The smarter our clients get, I think the more they can use our technology. And I think that's a good thing. Again, partnering with Terra thinks that's gonna put an AI native claim system in front of the users. I think it's a great thing for the user, it's a great thing for Terrence. It's a great thing for Grady.

James BenhamYeah. Awesome. Great comments. Good good discussion today. Obviously, unfortunately, we're out of time, but always a a great chat around AI, around insurance. Insurance, a data heavy industry that can benefit greatly from the study of correlation and causality. I think that's a that that almost describes the quest forever in insurance from the the coffee shop days of Lloyd's of London, the sixteen hundreds, back to the bottomary contracts, the Babylonians struck thirty five hundred years ago that there were the first insurance contracts. Is everybody's always tried to price risk by looking at correlation and causality.

Unfortunately, for a long time, we didn't have the tools to look at all the data and now we do. Right. You know, part parting thoughts, Stan, And I I know I said I was done, but this really this is hard for me to ignore. The large LLM providers act like the LLMs are gonna eat the world on software and AI. You know, they've said if you're building anything on top of our platform, beware. Mhmm. And, obviously, you haven't. You built your own your own machine learning models that you've trained.

It's a it's a different subset of AI. You, of course, are using LLMs to interpret free text and to to determine, you know, sentiment and summarization. You're doing a lot with those LMs. Just in general, do you think the large LLM model the, like, the chat GBT and Anthropics of the world are are gonna eat software, or do you think that they might be, you know, more complementary than than replacement?

Stan SmithSo so, James, I think when you see these these technology waves, the way I I've seen them in the past and I've seen the exact same thing now, the first major wave are what I call horizontal. They're general use cases that can be used, generally speaking, across multiple industries. And then what starts to crop up is the verticalization. And especially in AI, you need verticalization of training data. So if you look at LLMs being used by law firms, that's not you know, again, I think ChatGPT probably does some of that right out of the box, but there are to be companies that train on legal whether it's defense side only or the other side.

You know, there are going be companies that are building vertical solutions that leverage things that are horizontal, as big as these guys are, as much as they spend. So I think what I've seen in the past and the reason I focus on vertical software startups is because verticalization allows you to specialize and build things that are uniquely positioned, not just because of the technology, because of your knowledge of the customer's needs and better than the horizontal players can ever do it. So I think there's always gonna be a place for both in this world.

If we're doing something that a generalist can can replicate and do as well as we can, then we shouldn't be doing it. That's the way we look at our own business. If LLM could do this out of the box,

James Benhamthen that's not something we should build. That's the way we're actually using it to help pick the our targets as well. It's a good answer. It's a good it's a good summary. It's a it's a brave new world that we're always trying to the the lines are shifting quickly in software. It's never seemed like this before, but, you know, technology moves faster every year. So I think that's just normal. It it it literally does. There's a lot to, there's a lot to consider in all of that.

Well, thank you for your discussion, Stan. It's always good to talk to you. Always good to see you. And, Rob, as always, appreciate your guidance there, brother.

Rob GalbraithYep. Great conversation. So glad we were having to catch up, Stan, and congrats to both of you on the partnership.

Stan SmithK. Thanks. Thank you, guys. And

James Benhamthank you for tuning in today to geek out our interview with Stan Smith from Gradient AI.

Browse all episodes

Never miss an episode

Sign up for the InsurTech Geek newsletter and get new episodes, insurtech trends, and articles straight to your inbox.

The showAll EpisodesTopicsGuest DirectoryBooksAboutContactPrivacyTermsSponsor the showPitch a guest
ListenApple PodcastsSpotifyAmazon MusiciHeartRadioYouTube
FollowLinkedInYouTube