Episode: 558 |
Astrid Malval-Beharry:
AI Project Case Study:


Astrid Malval-Beharry

AI Project Case Study

Show Notes

In this episode of Unleashed, Astrid Malval-Beharry discusses an AI case study with a top 50 homeowners insurance carrier in the US. Astrid was approached by their underwriting and innovation teams to digitally transform their underwriting workflow.  Astrid shares an overview of the industry at present. The industry is facing challenges due to an increase in natural catastrophes, inflation, disruptions in the supply chains, and policyholders who prefer to have an Amazon or Uber experience with their insurance carrier. The client had three goals for the digital transformation project: increasing the level of straight-through processes, improving risk assessment, and realizing greater investment in inspection. Astrid explains what  straight-through processing is and how it works using data analytics and AI-based and technology solutions. The second goal was to improve risk assessment by analyzing the location of the property, the condition of the property, and the policyholders themselves. The client wanted to know how AI solutions could help enhance risk assessment, reduce premium leakage, and charge the right price for coverage.The third goal was to improve the inspection process, which currently costs carriers a lot of money but only yields a few actionable insights. To achieve this, Astrid’s team shadowed underwriters across both regions and senior IDI to understand how consistently underwriting guidelines are being applied. The team also interviewed and benchmarked against competing carriers, InsurTech carriers, and carriers that look at the underwriting workflow with a different lens. This allowed them to see the art of the possible and make informed decisions about their underwriting practices without disrupting the workflow.


Employing AI Solutions for Insurance Companies

Astrid talks about what follows the research and benchmarking exercise and how they mapped the workflow and the ideal future state.  Premium leakage occurs when insurance companies charge less for a policy than the actual premium should be to reduce losses and charge the right price for the coverage. The inspection process is often done by agents or license inspectors, leading to a lack of actionable insights. To address this issue, a preferred digital transformation engagement was conducted by shadowing underwriters across both regions and senior IDI. This allowed the team to understand the consistency of underwriting guidelines and the impact of different levels of underwriters on the process. Competitive intelligence benchmarking was conducted against carriers with similar profiles and InsurTech carriers. This allowed the team to map the workflow as the ideal future state from an underwriting workflow perspective. However, the change should not be too abrupt, as the procurement process in the insurance industry is notoriously long. A middle ground was identified by analyzing claims activities on the book of business NIS to identify the biggest losses and how implementing AI solutions would give the highest return on investment. Change management is also important, as it involves both technology and people and processes. The organization’s readiness to implement new digital tech-driven solutions is also crucial. Astrid also touches on the convergence of people and processes when implementing technological solutions in change management.


Questions to Ask an AI Vendor

Astrid shares a list of questions to ask an AI vendor, including accuracy, model explainability, model bias and fairness, and scalability. She has experience working with insurance carriers, analytics, technology vendors, and private equity firms, giving her a deep understanding of what solutions work and don’t work. When selecting an AI vendor, it is important to understand a series of fundamentals about the solution.

The first question is about the accuracy and performance of the AI model. It’s crucial to understand how the vendor measures accuracy and how they handle situations where the model may not perform as expected. 

The second question is about model explainability, which is crucial in the highly regulated insurance industry. 

The third question is about model bias and fairness, and how the vendor addresses and mitigates biases in their AI models.

The fourth question is about scalability. While some solutions are considered vaporware, and Astrid explains what vaporware is, there are legitimate, enterprise-grade solutions that have legitimate AI technology. By asking these questions, clients can better engage with the right AI vendor and ensure the right decision-making process. She states that licensing data from a vendor is the right path due to the ongoing maintenance required. AI vendors are now incorporating large language models, such as chat GPT, into their AI models. However, this is not the core competency of an insurance carrier, which is to assess risk.

Astrid stresses that results should not be expected too quickly. However, she does mention that they are already seeing results. She mentions a project that has been 16 months in development, and it is not expected that a solution will immediately bring new business or reduce expenses. However, the results have been significant, with a client seeing a 75% increase in straight-through processing and reduced manual injury interventions. Operational efficiency has also soared, and better risk assessment has been achieved.



01:02 Digitally transforming underwriting workflow for a top 50 US homeowners insurance carrier

03:08 AI solutions for insurance industry digital transformation

07:14 AI implementation in insurance industry

13:42 AI model accuracy, explainability, bias, and scalability in insurance industry

17:54 Evaluating AI vendors for insurance industry use cases



Website: https://www.stratmaven.com/

LinkedIn: https://www.linkedin.com/in/astridmb/


One weekly email with bonus materials and summaries of each new episode:


  1. Astrid Malval-Beharry

Will Bachman, Astrid Malval-Beharry


Will Bachman  00:02

Hello and welcome to Unleashed. I’m your host will Bachman. Today is an episode in our AI case study series. And I’m so happy to welcome my friend and one of the founding members of Umbrex Astrid Melville Buhari. Astrid, welcome to the show.


Astrid Malval-Beharry  00:21

Thank you so much. Well, it’s a pleasure to talk to you. And I’m so grateful to be part of the Umbrex. Community. This is exciting to be talking to you today.


Will Bachman  00:30

Thank you. You’ve been you’ve been part of the community since the very beginning 70 years ago. Astrid, it’s great to have you on the show. I know you are an insurer tech consultant focused on insure tech industry. Walk us through your case study.


Astrid Malval-Beharry  00:46

Absolutely, it will be my pleasure. So this particular case study will involves a top 50 homeowners insurance carrier in the US. And I was approached by their underwriting and innovation teams to help them digitally transform their underwriting workflow. Maybe for context, we’ll for your for our audience, I’ll give kind of a quick high level overview of the homeowners insurance industry in the US. And I think it will help to have that context for what my client was looking to achieve. But it’s about a $134 billion in premiums industry, there are 1000s of insurers, but it’s actually a fairly concentrated market. So the top 25 carriers account for 70% of premium, the top 50 carriers for 80% of premium. And my client, as I said, was a top 50 insurance carrier. This is an industry that’s going through a number of challenges, right, so we hear all the time on the news, you know, the wildfires in California, flooding in both California and the Southeast, the northeast, obviously, I’m based in Miami, and lots of hurricane there, you know, heal in the Midwest and whatnot. So all all those natural catastrophes have been increasing in severity and frequency, and will only continue to do so right with climate change. Additionally, over the past couple of years, you know, there’s been obviously inflation disruption in the supply chain that has made it more costly to reconstruct a company, I mean, sorry, a property after a disaster or, you know, any type of of claim. And so those are some of the challenges facing insurance carriers. Additionally, you have policyholders, across frankly, all ages, it’s not just the Gen Z, that are, you know, use, obviously, to the Amazon experience, the Uber experience and want to have something similar with your insurance carrier. And so I would say, you know, a lot of insurance carriers are grappling with those issues, and my client was as well. So when they came to me, they had three goals in mind in for this digital transformation project. The first one was they wanted to increase the level of policies that were straight through process, and I will explain a little bit, you know, what this means. Second, they wanted to increase, you know, improve the risk assessment of policies they were about to put on their book. And then third, they wanted to realize a greater investment on their inspection. So maybe let me start with the first the first goal, which was around Street to processing, and basically straight through processing is the process of essentially able to review and approve an insurance policy smoothly without any delays or extra requirements and Frankie, without really manual intervention. And typically, you achieve that using data analytics, you know, AI based in technology solutions. So that was the first goal. The second goal, as I said, was around the improve risk assessment. So when you underwrite a policy, in the insurance industry, specifically in homeowners, you are looking at specific risk insights. So one would be around the location of the property itself. Is this a property that’s located in an area that’s prone to flooding or hurricane or he’ll? Is it an area that’s you know, more exposed to crime and then other areas? So this will have an impact in terms of how you underwrite and price that policy The second is around insights into the property itself. So how old is the roof? What condition is this property? Is there a pool a trampoline, you know, are they are outdated appliances that can cause you know, all types of hazards inside the home. And then finally, insights into the policyholders themselves, you know, are these people that tend to file, you know, claims, you know, at the drop of a hat, and that tends to indicate, you know, a greater propensity to file claims, for example. And so, my client wanted to figure out what were some of the AI solutions that could help them, you know, enhance risk assessment so that there would be less premium leakage. Premium leakage is when an insurance company charges less for a policy than the actual, you know, premium should be. But they also wanted to reduce losses, right, ultimately, so charge the right price for the coverage that’s being provided. And then finally, as I said, the inspection process, so right now, carriers spend a ton of money on inspection, and very few of them, you know, will lead to actionable insights. So very often, it will be your agent at the end of the day, you know, on his way, or way home, stopping by taking a few pictures, or an actual license inspector will come on site. But you know, we’ll just do an exterior, a quick exterior inspection. And like I said, a few of these will lead actionable insights, and carriers still end up spending quite a bit of money. So the first hour, I guess, I’ll walk you through sort of the process we took in, you know, this preferred digital transformation engagement, the first thing we did was to actually shadow underwriters. And that’s one of the favorite parts of any project that I do is the primary research component of it. And in particular, I love you know, shadowing the actual users of a solution. So in this case, we shadowed underwriters across both region and senior IDI. So it was important for me to understand and my team, you know, how consistently are the underwriting guidelines of the scarier, you know, being applied? Is there a difference between how you know, a more junior underwriter is approaching this versus, you know, a more senior underwriter. And that has implications when we talk about change management when we’re introducing new innovative solutions, particularly tech solutions, right into a process. We also interviewed and benchmark if you will, against not only competing carriers, so carriers that were direct competitors had a similar profile book of business, but also InsurTech, carriers and mgas, that, you know, look at this underwriting workflow with a different lens, if you will. And that is, that is important because it can show what I call the art of the possible very often, you know, these are, you know, these can be 100 year old organizations, right, that are stuck in their ways. And it’s important to understand what is the art of the possible but also, what is the path to getting there, you don’t want to disrupt, you know, the workflow to such an extent that, you know, you lose your agents, they recommend other competitors versus you because the change has been too abrupt, too disruptive and same thing for your policyholders. And so, having shadowed the underwriters done the competitive intelligence benchmarking exercise, we were able to map you know, the workflow as is the ideal future future state from an underwriting workflow perspective. But as I said, you don’t want the change to be too abrupt. And it does take a long time right. procurement process in the insurance industry is notorious to be very long, as many of those AI vendors will attest. And so we came up with sort of a step in between between the as is state and the future ideal state. And for that middle ground, the intermediary step. We did sort of an analysis to prioritize, you know, what were the solutions that would we believe would give us the highest return on investment to begin with? And one way we did that was honestly to analyze you know, the claims activities on the book of business NIS to realize, you know, what, where were the biggest losses coming from? And if we had had a solution in place back then, when we first took this policy on our book of business, would we have been able to avoid those losses either by, you know, first of all, not even letting that policy in the door. But working with but letting the policy in the door and working hand in hand with the policyholder to mitigate against that risk, and be able to offer the proper coverage. And so we were guided by that analysis, to understand, you know, what were the AI solutions that would have priority from an implementation point of view, for that intermediary step. I think one thing that I will add there, we’ll that I think is important. I’m obviously if you cannot tell, I’m actually really passionate about technology. But I also have found over 30, plus engagement, around deploying AI solutions, and around digital transformation, that change management is also equally important, if not more important. And so there’s obviously one side of change management that’s around the technology itself. But there is another side to change management that’s around people and processes, and how ready the organization is to implement, you know, these, the these types of new digital tech driven solutions. And so we weren’t guided by all of that, you know, really to inform our approach. So let me stop there, we’ll and see, you know, if you have any, any questions, and then I’ll be happy to tell you, then, you know, how we went about, you know, bringing in those AI solutions?


Will Bachman  11:57

Yeah, let’s, let’s get to that. And I know, you put together a helpful list of questions to ask an AI vendor. And I’d love to have you walk through that as well. Absolutely,


Astrid Malval-Beharry  12:08

yes. So I have the benefit, because I work at stripe, even with insurance carriers, good analytics, technology vendors and private equity firms who tend to invest in those vendors, it gives me sort of a really deep understanding of, you know, what solutions work, don’t work. And, you know, some of those vendors, I’ll readily admit are or customers of mine. And so I had a fairly good understanding of the type of vendors that I wanted my client to engage with. In some, for some of them, I facilitated the introduction. For others, my client already, you know, was aware of the vendor had spoken to them in the past, but not fully engage. And, exactly to your point, one of the things that I did was really to guide my client in asking the right questions. And I think so often, you know, we can be blinded, if you will, a little bit by the next shiny objects, nothing wrong with that, I think that’s actually required, again, to think about the art of the possible. But when it comes down to selecting a vendor and AI vendor and implementing that solution, it’s really important to understand, you know, a series of fundamentals about that solutions. And so, you know, there are a series of questions that I that I typically like to ask. So the first one is around the accuracy and the performance of the AI model. So what is the accuracy rate? How does the vendor Miss measure that accuracy rate, and along those side, right, we want to see how they handle situations where the AI model may not perform as expected, that’s actually very important to understand the type of feedback loop that those vendors have in place to incorporate, you know, the false positive or the false, you know, negatives, and learn from that have their AI algorithm learn from that. The second one is around model explainability. So what I what I call the black box phenomenon, and because the insurance industry is highly regulated, and it’s regulated at the state level, so 50 different state regulations. And you know, a lot of our insurance regulators, you know, are not necessarily right data scientist and modelers. And so, you have to really watch out for AI model opacity, if you will. And so, I the the real test for me is Kennedy eyes vendor explained the decision making process of their models in a way that is understandable to non experts. And so that’s one of the tests that I use. The third one is around the model bias and fairness. And, you know, how does the vendor address and mitigate biases in their AI models, and I’m particularly so happy to see, you know, regulatory actions that are being taken, you know, even in the United States, in Europe, around ethical ai, ai safety. And even in the insurance industry, I see many of the newer AI solution vendors really incorporate that, you know, into their decision making process of their models. And I only think this will increase, you know, as more innovations are happening in AI, and that’s, that’s really critical. The first fourth one is around I would say, scalability. So we’ll there is quite a bit of solutions that I would say are vaporware. And so obviously, you don’t want those. But there are solutions that are not vaporware that have legitimate, awesome, beautiful AI technology. But they are not what I call enterprise grade. So


Will Bachman  16:22

sorry, what do you mean by the term vaporware,


Astrid Malval-Beharry  16:25

vaporware meaning, you know, it’s, it’s it’s not real, like it’s a, it’s more of a can demo that the vendor will present to you. But it’s not really, it’s not really proven? The AI models so a lot of people will throw the word AI, you know, in their brochure, their marketing materials, because they want to create that buzz. But deep down when you look into it, it’s not it’s not, it’s not the real thing. On the flip side, as I was saying, you do have solutions that are the real thing. But they are not what I call enterprise grade. So they are not, you cannot really deploy them at scale. You know, with an increase in data and user load, the model falls apart. And and that is actually critical for any insurance company, right? Especially when you think of new business, you have tons of quotes coming in, not all those quotes will obviously end up in being a buying policy or sold policy. But you do have to account for that for that volume. Another area that’s important is, you know, nowadays, a lot of those AI solutions, you know, have to be able to be integrated through API into a carriers stack. And you know, that can be applied to other industry as well. Right. You could imagine the same thing in banking and in ag tech and whatnot. And, and so it’s helpful, right, when an AI solution is already pre integrated with other technology stacks, that a carrier leverages already right, it speeds up the deployment. And it’s also, you know, it’s one more way to, to verify, right, that the solution has been successfully deployed in leverage by others in the industry. There are other things that I other questions that I asked that, but the last one, you’re just in the interest of time that I think is important is the maintenance and support of those AI models. You know, what kind of ongoing maintenance and support does the vendor provide? You know, how do they handle updates and upgrades to their AI model? And as I said, how do you incorporate that feedback, right from the users to continuously improve your AI models? And that is, very often, you know, I’ve worked with some carriers that will say, you know, given the price of implement of, of implementation or licensing the solution, Isn’t it worth our while we have enough data to build the solution in house and unless you’re a State Farm or Allstate, you know, some of the you know, top five carriers where you may have enough data to build the solutions in house and even then, I still don’t think, I think depending on the use case, yes, it might make sense for you to build it in house. But generally speaking, I believe that licensing data from a vendor is the right path precisely because of that ongoing maintenance that’s required. And we’re just seeing it now. Right? We’re just seeing AI vendors that you know, we’re leveraging computer vision for example, right, which is a form Over artificial intelligence or, you know, natural language processing NLP, they are now starting to incorporate those large language models, right the chat GPT type of AI solutions. And so these things will only continue to evolve more AI techniques, if you will, will emerge. And, you know, that’s not the core competency of an insurance carrier, the core competency of an insurance carrier is to assess risk. And so, you know, I believe that, you know, leaving that to the vendor is is important. But But likewise, when you’re selecting a vendor, it’s really important to understand those various elements that we’ve discussed.


Will Bachman  20:47

Well, that is a great that what you just gave with by itself is a fantastic many resource, walk that questions to ask an AI vendor.


Astrid Malval-Beharry  20:59

Oh, I’m glad. I’m glad that’s helpful. Well, and yes. And in your audience, if they want to reach out and ask any questions, I would be happy to do that. I’ll just quickly again, just at least, to give a sense of what the results were, you know, this is a fairly new project. So we are 1416 months in. And one thing that I will see as well, and I think this is across any industry, but it’s not to think that you’re going to turn on a solution and day one, and the magic will happen, you know, a bunch of new bliss business will flooding in or will you will reduce your expense line overnight through automation, you know, these things take time, there’s a sequence to deploying those solutions. And but we are already seeing quite a bit, you know, very impactful results, I mentioned the straight through processing. So my client has already been able to go from 45%, straight through process to 75%. That is absolutely remarkable results. So that means, right, manual injury interventions being drastically reduced. operational efficiency is soaring. And, and, and honestly better risk assessment in the process, because what I observe when I was shadowing those underwriters, even though they were following the underwriting guidelines of my client, but two underwriters could actually come up with two different solutions. And I think when you’re introducing AI solutions, there is definitely more consistency and more accuracy in how a particular risk is being assessed. And so, you know, and by the way, we’ve so we’ve definitely seen increased new business. And I believe they will see an improvement in their loss ratio, and, and so forth. And from an inspection point of view, which was a third point that I had mentioned, we’ve also been able they are in the process right now of testing solutions, that are basically phone Bayes solutions, where the policyholder can, you know, basically scan the exterior and interior of their property, send that to the carrier. And so that reduces costs speed up the process, but it also Frankie provide more actionable results as well. And so we have, that solution has not yet been deployed, but I think will be very promising for my client.


Will Bachman  23:48

Astrid, thank you for this case study. And you made my job easy on this one you had, you know, thought thought through it. so deeply. This was really amazing overview. By itself, your list of questions around you know, how to judge an AI vendor is super valuable. And a will include if it’s okay with you, those questions in the shownotes? Absolutely. Where can listeners find more about your firm?


Astrid Malval-Beharry  24:19

Absolutely. So they can go to our website, WW dot strat, maven.com. And also feel free to connect with me on LinkedIn. So Astrid, Madhavi, Harry.


Will Bachman  24:31

Great. We will include your LinkedIn URL as well as your firm’s website in the show notes. Astrid, thank you so much for joining today.


Astrid Malval-Beharry  24:40

Thank you. Well, it was a pleasure. Thanks, everyone.

Related Episodes


Automating Tax Accounting for Solopreneurs

Ran Harpaz


Integrating AI into a 100-year-old Media Business

Salah Zalatimo


Author of Second Act, on The Secrets of Late Bloomers

Henry Oliver


Third Party Risk Management and Cyber Security

Craig Callé