AI in Insurance: How Carriers Are Using AI to Improve Underwriting, Claims, and Operations

Introduction: meet Chaz Perera, CEO of Bevaya

Will Jones, Alliance InsightsWelcome to Alliance Insights. Our guest today is someone who just didn't wander into insurance. He actually understands it, which already puts him ahead of almost all the so-called insurtech experts. Chaz Perera is the co-founder and CEO of Bevaya, an AI agent platform purposely built for insurance operations.

And when I say insurance, I mean the real work, underwriting, claims, policy servicing, the workflows that actually drive combined ratios and not just conference panel discussions. What Chaz and his team have built isn't theoretical AI. These are production AI agents that can read submissions, analyze claims data, and make recommendations across the core insurance processes. So instead of just adding another layer of technology, he's focused on actually increasing throughput and decision quality inside the workflows that already matter to insurance carriers.

Chaz has been working across the intersection of insurance and automation long enough to understand the nuances. He knows where straight through processing works, where human judgment still matters, and how to bridge the two without breaking your existing systems or your underwriting discipline. He's also a member of the Forbes Technology Council. So, in addition to building this stuff, he's been shaping how people think about AI adoption across the industry, ideally in a way that goes beyond just adding AI to a strategic plan slide.

So, if you're thinking about scaling operations, improving loss cost management, or just trying to figure out what AI actually looks like inside insurance carrier or broker environment, this is the guy who's already doing it. Chaz, welcome to Alliance Insights.

Chaz Perera, BevayaThanks, Will. I appreciate the opportunity to be able to speak with your audience and share a little bit of of sort of my perspectives on the industry.

The biggest operational bottlenecks in insurance

Will Jones, Alliance InsightsAbsolutely. So Chaz, you spent over a decade in senior roles at AIG across claims, operations, transformation. What were the biggest friction points you saw that led to build the current organization?

Chaz Perera, Bevayayou know, well, the the bottleneck in insurance has never been the decision. it's everything feeding that decision. It's the documents, the data hunting, constantly having to hop between systems. You know, frankly, it's a lot of smart people doing really low judgment work.

and and every transformation was just an exercising adding another tool that nobody really owned. Imagine more screens, not fewer steps. So for us the insight for me the insight for for the team has been carriers brokers agencies TPA they don't need more software they need the work done and what we're really focused on is document to decision end to end

AI agents versus traditional insurance automation

Will Jones, Alliance Insightsso when you describe Bevaya today what makes it fundamentally different from the automation tools carriers have already tried

Chaz Perera, Bevayaso I suppose to level that well old automation was rules. Brittle breaks the moment reality starts to vary even a little bit. where Bevaya runs on judgment not a series of scripts. our product is vendor-neutral orchestration.

So it sits on top instead of sort of locking into a single vendor road map. and you know, ultimately what are you trying to succeed in? You're, you know, you're not buying a workflow. You're trying to own one.

You're trying to control the outcomes and ensure that your best people are able to coach it. And that's really what our platform, the Bevaya platform, is about. It's the human at the center being able to make those high value judgment calls, but taking everything else off their plate. So that's where they can focus.

How insurance carriers should approach AI adoption

Will Jones, Alliance InsightsThat's awesome. Absolutely. So when you when you sit down with a client, you know, what are kind of the first steps before any technology is implemented that carriers really need to get right.

Chaz Perera, BevayaYeah. start with the decision you're trying to make. And I don't mean should you use AI, for example. I I I mean more specifically how fast are you trying to accomplish the task that you're doing today?

How completely or how accurately are you trying to to achieve that? what type of customer experience are you trying to create? Understanding those the decisions there not the document is really what your where your focus should be initially. It's important here essentially to name the judgment that you want to have made and who actually owns the outcome.

It's not as simple as just it's not as simple as saying this is what we want to do, but who's accountable for making sure that actually happens. I suppose people have a tendency to also focus on trying to create this very utopic or per perfect series of outcomes. And I would argue just define what good looks like. And as part of that, where should a human be part of the decision-making process when you start to automate things and start to leverage artificial intelligence?

because governance is a design input. It's not a a bolt-on after the fact. And I think this is probably where like these two things become really important, right? Pick the right thing, put the right owner around it so that you get the right decisions at the end of it.

Last thing I would say, it's automation. It's something that you're looking to have happen early, often at high volume. So, picking a a use case with real volume, with a clear set of key performance indicators that can be proven to create value in weeks. That's really the the the starting point.

Why many insurance AI projects fail to scale

Will Jones, Alliance InsightsAnd do you find it Chaz that that organizations or or people in those organizations have a hard time laying that out before you know engaging with you all or any other automation?

Chaz Perera, BevayaI I think that most customers have a half perspective at a high level on what they want to solve. But I think the tendency is let's try to do one big bang. Yeah. And my suggestion here is pick a single product line within a within a use case.

Even then maybe pick a a geography where you want to start that because so much of this is about ensuring that the human that is the recipient of all of this information and ultimately is making that decision, making that judgment call has high degrees of confidence in the information available to that individual. and and and that's I think where people tend to fall down. They they they they don't see it as a baby step that can then sort of roll down a hill and and and you know gather steam. They think of it as like I've got to do it all because I only have one shot at it.

And that's really not the case.

AI governance, compliance, and auditability

Will Jones, Alliance InsightsYes, that's a great point. So governance around AI has becoming more and more important. And so how do how do you see the balance of kind of like the development getting everything we want out of this tool or this this workflow but with the governance the auditability especially in a high regulatory environment like insurance.

Chaz Perera, BevayaYeah. I I suppose first well you know I think there's a lot of like people tend to think of these things as trade-offs and I I don't think I I think the I don't think they're trade-offs like auditability is a feature. It shouldn't be considered friction in your in your value chain. for example, every judgment are our AI agents render is traceable.

It's explainable. we have human in the loop checkpoints that are built into these workflows. and it's the end user and the customer that's deciding where where autonomy of these agents stops. and a human signs off. It's our customers that are determining how you know broad or how narrow the guard rails in which these agents are operating.

And as that as that comfort and confidence in the performance of these agents and the agents and humans grow, you know, the wider those guard rails can can become. The the only thing I would say here that I think is important for people to understand a lot of agentic solutions in the market today born out of things like Claude Code and you know OpenAI's Codex everyone thinks everyone thinks the idea here is give the solution a goal give the agent a goal and then allow it to get to there on its own. that works really well in the context of a one-off problem. For example, I'm I want to build a system to do X or I want to build a feature to do Y. that works really poorly in an environment one that's highly regulated but two where you're trying to do the same thing over and over and over again where the human on the end of it that's making the judgment calls and the decisions is looking for predictability consistency a you know a pattern so to speak and I think what's important here is to understand that deciding what the workflow will look like and what the human and the agent how the human and the agent will interact. That starts off as this probabilistic exercise, right?

You give the agent a goal and have it come out with a plan. But when it comes to the normal day-to-day execution, that has to be deterministic. Meaning the, you know, the the same inputs produce the same path because that's what your regulators expect, that's what your leaders expect. You need that level of auditability.

How AI improves underwriting decisions and risk selection

Will Jones, Alliance InsightsThat's a great explanation of that. Absolutely. And so in, you know, in that sort of same vein, like when we're talking about underwriting, you know, where do you really see AI, you know, potentially having the to create a true competitive edge and and obviously I think everyone's sold on the efficiency, right? But yeah, how does AI actually help the processor or even help the human within that not just be more efficient, but actually make maybe better or faster decisions?

Chaz Perera, BevayaYeah. Yeah. Look, I I think I I would argue here it's not even just faster, it's better. Just plain old better, right?

You're you're the AI is going to afford you the opportunity to see more risk and see it more consistently and see it frankly across every submission instead of just the ones the underwriter gets to. Yeah. Then you start to think about well I'm going to apply triage at the top of the funnel so the right risks reach the right underwriter sooner. So you're you know and then your speed to quote becomes a win rate lever.

that's the exercise and freeing up especially the senior people for the hard judgment calls where the relationship and the engagement may sway the decision in your favor. Right? That's I think where this really starts to change. And so so then you're then you're starting to evaluate risks on the basis of profitability instead of intake on the basis of the grunt work that needs to get done.

AI-powered submission intake and risk analysis

Will Jones, Alliance InsightsAbsolutely. And I've yet to meet an underwriter that wants to have more losses on their book. So any more information they can get or other risks that have a similar profile or any kind of correlation that they can find that they've already written or another underwriter a huge hu just a helpful for their own consciousness and are they making the best judgment so 100%. And so, you know, in that, you know, how does an AI agent really approach things like the submission intake or the clearance and and even the risk analysts you the risk analysis differently than maybe a traditional workflow or just a human could do on its own.

Chaz Perera, BevayaYeah. So, I I suppose tra traditional flows are typically sequential and very manual, right? You got to re key stuff, you got to look stuff up, you got to route things places. imagine the AI agents doing all of that in parallel.

So that's fundamentally changing the way the the world works. imagine sort of the AI agent who's responsible for this clearance step, right? It's not it's no longer a search problem. You're not trying to figure out like, you know, where where is this in the in in in the mix.

the agent is actively the AI agent is actively reconciling all of these decisions. across multiple systems, multiple decision points, and chasing things if necessary, right? Go back to the broker. I actually need three years of three additional years of loss runs. Yeah.

Right. There's no way I can put this in front of my underwriter and get a good answer otherwise. So, suddenly the underwriter is not in the business of chasing things. They're just there ready to make the the best evolved decision they can make with the information available to them at the right moments in time.

I think the other thing that you also end up having then and this is again where I think the world changes and this becomes less traditional. you can surface risk signals to your earlier point will about you know there was another underwriter that maybe was in an office in another city that we're seeing something similar. Yeah. Right.

There's a trend here potentially. How do we how do we evaluate that? maybe that becomes a product decision and a change in either you know manuscript maybe it becomes a manuscript thing you take on and that then eventually evolves to something different from the product perspective right that's the sort of thing I think that you're starting to you'll shift to because you can find those opportunities faster and the AI will keep a sort of you know its brain on it versus what an underwriter might do is look I've got five other things to work on I'm going to move on to the next thing.

Using AI to transform claims management

Will Jones, Alliance InsightsYeah that's awesome man. so let's shift a little bit to claims real fast. So claims is always sort of, you know, it obviously it all impacts the customer, but claims seems to be the thing that obviously customers look at directly that impacts them the most, right? At a loss, how is this carrier going to respond to this claim?

So, how are carriers using AI in that claims process to maybe improve speed or accuracy without compromising trust and obviously having a fair claim settlement at the end of the day?

Chaz Perera, BevayaYeah, look, I I think the reality for most insurance carriers is your brand is made at the claim and trust comes through consistency. Yeah. Right. So imagine using the AI and people say they use the AI to ensure that the same standard is applied across every claim every time.

That predictability is highly accretive not just to the policyholder but to the agent or to the broker that's that's your partner in this exercise. Speed on its own is valuable, but speed where it's safe, where scrutiny matters is an is an important exercise here, right? There are certain subsets of claims, for example, where it makes sense to just pay it. So do it.

Why create complexity when it's unnecessary? And then there's clearly a larger subset of claims where judgment, perspective, understanding of the relationship like there's so many other factors that matter in the decision and those are the ones that you know you don't fasttrack. You take your time on and again you do it in a way that is consistent because that creates trust and transparency. and then I I I would just say probably really important to your customer.

They don't necessarily have to agree with your decision and sometimes they won't, but if you can show your work, that, you know, bridges the the unknown for the for the for the policyholder and the claimant. And I think that's one of the areas where AI can be really valuable. And we've seen customers do this where they're using the input, the inferencing, the decisioning from the AI to help the adjuster shape what the work what the the show your work look like.

Why human-in-the-loop AI matters

Will Jones, Alliance InsightsAbsolutely. That's a great point 100%. And so you mentioned to you know AI not just standing alone. Obviously you guys look to involve a human in the loop.

So, where does kind of the human in the loop oversight become critical in the claims workflow?

Chaz Perera, BevayaYeah, what what I'm about to say will I think you could apply to underwriting to claims to policy servicing or or frankly any workflow or process where humans and AI are working together. But you know where does the rubber meet the road? I suppose in claims it's anywhere a decision touches a customer's money or your money where you know a coverage decision is is is is either favorable or not to your policy holder or or to the claimant. I think that's really where the human in the loop really has to to sit in because if the AI were to make the decision on its own, I think it creates ambiguity for the person that has to communicate the decision to another human and removing that ambiguity goes back to my earlier point of, you know, transparency and consistency builds trust.

I think this is sort of where the human in the loop really comes to fruition. So it's not simply about making sure the AI is correct. Over time the AI will be correct because the right humans are participating but it's also about giving the human who is the human in the loop the confidence that they have the right information shaped in the right way for them to be able to turn around and have a fulfilling conversation with the customer with the claim policy holder who whomever it might be. and in claims, unfortunately, it's often when something bad has happened, people want to get to their semblance of normal as quickly as possible, and you might not be able to afford them that.

Integrating AI into insurance workflows

Will Jones, Alliance InsightsAbsolutely. Great point. So, you've seen carriers deploy AI in a lot of different ways. Some fully connected to their core system, some running alongside them.

In your opinion, Chaz, what actually determines whether AI changes how decisions get made versus just adding another tool as you mentioned before that no one needs and and no one cares about?

Chaz Perera, BevayaYeah, I think integration in integration into a decision, not just a screen, really matters here. So really thinking long and hard about you know where can AI be best leveraged because it's not it's not the entirety of the process first and foremost you know and frankly if it sits behind a workflow or isn't well sort of connected to the workflow then I I think you're going to have like it's shelfware like so many other technologies have have historically been u I think fundamentally well that's why co-pilots that concept of co-pilot really doesn't work in the ether and why people don't get value from it because it's not well integrated into the way people work. It's an option that you can use and therefore it becomes a behavior you choose to use very sporadically. I I also think ownership really matters here and you have to enable the people at the ground level who are the humans in that loop who are responsible for the decision or the outcomes of the decisions rather like they have to feel like that agent is theirs and that they can coach it and that they can shape it and that they can guide it. and that the only way that happens by the way is if they feel value from it and the only way to get value from it is for it to be highly integrated into the workflow.

So I so I would say you know it's important for these AI agents certainly to connect to core systems to connect to the ancillary systems that help you shape a decision because as I said earlier a policy decision an underwriting decision a claims decision doesn't happen in the silo of one one piece of software like Guidewire there's many things coming together to make that decision and you know an agent that can act on all those things and suggest across all those things will be a lightly used agent if used at all.

Insurance-specific AI versus general-purpose models

Will Jones, Alliance InsightsAbsolutely. That's that's great. And and to segue from that Chaz is like so you've built on an insurance specific models trained on proprietary databases you know not only as to your point are they baked into the workflows is it sort of non-optional behavior to use it but how important is it to have like domain specific AI versus just general purpose you touched maybe a little bit on the co-pilot and and how it's optional but what about the actual data that the model's trained on and to your point earlier the pred predictability of those results. Does that only come through a domain specific AI?

Chaz Perera, BevayaYeah, it's by the way if you were to go to our website you know bevaya.ai and look and go to benchmarks or /benchmarks you could see how we perform relative to these frontier general knowledge models like Gemini 3.1 and OpenAI you know 5 etc. GPT-5 etc. I I think the thing about you have to appreciate about these frontier models these they're they're general knowledge models they understand language English French German they understand language where our product differs is it also understands insurance it understands the nomenclature the nuance in language from one geography to another one products to another it understands codes and forms essentially it it understands insurance the way you know we do well because we've been working in the industry for a really long time and we've gone through accreditations and you know we've learned by being on the job so to speak. That means it's trained on real submissions, real claims, not just sort of what people have made available on the public internet, right?

Things that have historically been closed off from the worldwide web. And what we find typically is the the data when run on a general or flows when run on a general knowledge model might only get you to a level of completeness and accuracy that's about 80%. But that uncertainty that 20% is a big big big gap to overcome. and what it tends to do is it forces the humans on the in these processes to review everything and then you start to lose the speed.

You start to lose the quality. You start to lose why you wanted the AI to begin with in the first place. And so that's why we would argue it's really important to have proprietary data sets to to to to be using technology that is, you know, trained on your tooling, right? in much the same way that there's a there's an there's an obvious and immediate lift to using Guidewire as your core system versus trying to build something off of Salesforce, right? Like there's just a difference in what shows up on day one.

And I think that's how people have to think about it.

Lessons from 120+ insurance AI deployments

Will Jones, Alliance InsightsIt's a great analogy. Yes, 100%. So Chaz, you've seen now over 120 or so live deployments. What patterns do you feel like separates carriers that successfully scale AI versus those that maybe stall out or don't get the full value or potential out of what they're implementing?

Chaz Perera, BevayaYeah, we're pretty proud of this by the way that we're well that we we've done 120 plus you know deployments of our AI in production at scale. but and the pattern is pretty pretty common right the scalers as I said earlier start with what's the decision what's the KPI that makes a difference stallers love to just talk about the technology scalers understand inherently here's the owner here's the coach if I get to production quickly I can shape this to the outcomes I want stallers love to run pilots. In fact, they'll run like a 100 pilots and they're trying to optimize at the pilot level and eventually they're like, "Oh, it's optimized. They declare victory." But they never operationalize it because they misunderstood operationalizing something is very different than running something in this sort of highly sanitized environment that a pilot often is. and then I think scalers have a have a notion that they're going that before they even start their objective is not to get the first thing into production, it's to expand into the next workflow and the next workflow and the next work.

So they're thinking about the architecting that they have to do organizationally and systematically to be able to do that. And I think stallers typically are thinking about well I'm just going to build this. it's going to work and then I'm going to go to the next thing and I'm gonna start from scratch and build it again instead of thinking about the scaffolding, right? And and we would argue what our platform is, it's the scaffolding, right? It's it's the harness that lets you go fast and build, you know, and and and and build into the next thing and the next thing and next thing without ever starting from scratch.

Common AI misconceptions leaders must unlearn

Will Jones, Alliance InsightsThat's great, man. Absolutely. and and you've touched really like on how the workflows and and the and the human in the loop works, but what do you feel like the one misconception about AI and insurance you think leaders still need to unlearn would be?

Chaz Perera, BevayaI I don't think and I appreciate that, you know, we live in a world where particularly some of the frontier labs have talked about the jobs are going away, right? and maybe one day in the far distant future jobs will go away. But guess what? New jobs will come up.

It's just it's just the cycle of life. Yeah. so I I would what I would what I would advise here is this isn't about replacing people. Yeah.

It is very much about moving people up the value chain to the judgment work to the decision-making and critically in insurance owning and maintaining and enhancing the relationship you have with the policyholder you have with their risk managers you have with their brokers and agents right because that's that's where the work gets done that's where the relation the relationship gets made no matter how bad something is you can solve and sort based on the relationship you built so I would that a couple others really quickly. the idea that the bigger general models will win. I just think in in in like in insurance, in banking, in healthcare, you know, where domain expertise and understanding and ownership is really important. Those things will beat raw horsepower because what you get with those things is raw horsepower, but you don't get you know, the right shape to what you're trying to solve.

you know, last thing I'd say is, you know, you don't you don't have to bet the farm. start narrow, prove it, expand quickly into the next thing, and build that virtuous circle. Like that's that's the way you win here.

Building an AI strategy inside a large carrier

Will Jones, Alliance InsightsThat's perfect, Chaz. And really kind of our last question. And so with your experience in a carrier like you know AIG for example you know if you had to build an AI strategy again or currently inside a large carrier like that knowing what you know now what what would you do differently or what would be sort of day one for you leading up to maybe like an implementation or or a pilot as you mentioned

Chaz Perera, Bevayafigure out the judgment the decisions rather that move the loss ratio or improve your win rate and start with those because that has a direct impact on your P&L. Right? So, I'm not talking about tech. I'm not talking about a tech road map.

I'm just saying what drives the behaviors and the outcomes in your business. Start there. Land a workflow. Prove it cold.

Then let demand pull the next one in. because you can't think of this as like I'm going to do this massive transformation, right? Like all these boards and CEOs have turned to their tech leaders and been like let's go do AI and so they're just purchasing AI products with no true balance for what they're trying to achieve. And then the last thing I'd say is think think about how you do this in a in the most neutral way possible because the tech is changing so quickly which is why I make the argument about think about the scaffolding and the harness to be able to plug in and plug out of whatever you need because there is a future right some people may not like to hear this but I I believe there is a future where even the idea of a core system doesn't exist because a core system fundamentally exists to record a decision, but the decision is happening above it.

Will Jones, Alliance InsightsThat that is Yeah, that would be amazing. Well, Chaz, thank you so much for your time. Is there anything that we didn't cover or anything I didn't ask that that you would like to mention?

Chaz Perera, Bevayaonly that, you know, we are we're insurance people through and through. We love the industry. come check us out at bevaya.ai and on there by the way you can also try our software get an interactive demo and you know you don't need to put your email address to try that. So you know we encourage people to go give it a go.

Closing

Will Jones, Alliance InsightsWell Chaz, thank you so much for being on Alliance Insights. We truly value your time and thank you for letting us know what the current environment of AI is like.

Chaz Perera, BevayaYeah, thanks Will. I appreciate the opportunity.

Will Jones, Alliance InsightsThat's it for this episode of Alliance Insights. For more information and to hear previous episodes, check out our website at riskeducation.org. Be sure to join us for the next one. And until then, thanks for listening.