Where to Start With AI in Insurance (2026)

Where insurers should start with AI

Jill Zakarian, BevayaI am thrilled to be joined by our co-founder and CEO, Chaz Perera. Chaz is going to walk us through a practical approach for figuring out just where to actually start with AI. So, without further ado, I'm happy to hand things over to you, Chaz. Take it away.

Chaz Perera, BevayaAwesome. Thanks, Jill. Okay, we talked to a lot of insurance teams. Many have started testing AI, but there are still plenty who are trying to figure out where to begin.

We actually quantified this in our "State of AI Adoption in Insurance" report which we released last month. As Jill mentioned, you can grab a copy of it in the related content widget. What you'll find when you read it is that 93% of insurers are engaged with AI in some way, but only 27% anything in production. That means about two-thirds of the industry is exactly where many of you probably are right now.

Knowing you need to do something is easy. Knowing where to begin isn't. That's exactly what today is about. It walks through the thought process before behind adding AI, mapping workflows, examples of AI in insurance and picking a starting point. So with that, let's get into it.

Why buying an AI tool first backfires

Chaz Perera, BevayaThe instinct most teams have is to go find a tool. It's a pattern we certainly see play out many, many times across the industry. a carrier decides one day that they need an AI strategy. They check out a variety of vendors.

They evaluate the the underlying technology and then they go buy something. Six months later, they're not sure they're not they're not sure anything has actually changed in their business. See, it's it's not the technology that was wrong. It's that they bought before they understood their own operations.

they bought before they understood what problem they were actually trying to solve. AI isn't a product you just plug and play. It's a fit problem. It's it's knowing where it belongs in your workflow before you look for the solution.

So what what are we really talking about? You have to start with the workflow. The technology will follow. The next few slides are going to walk us through how that architecture plays out.

Three steps to find where AI fits

Chaz Perera, BevayaSo, we're going to cover three steps. Mapping mapping your workflows, identifying the steps AI can own, and seeing what that looks like in practice.

Step 1: Map your real workflows

Chaz Perera, BevayaLet's begin with step one. Before you evaluate anything or see any demos of any products, you need to understand what your team actually does. Not what an org chart says, certainly not what process documentation alone say. What actually happens day-to-day.

There is almost always, in fact, forget the almost, there's always a gap between the documented process and the real one that people are doing every single day. Real processes include workarounds, handoffs via email that no one ever wrote down. A senior person in your organization who quietly fixes things before anyone notices that something's gone wrong. AI doesn't see that gap.

It doesn't know that gap. So, it will automate the documented process and miss the real one. So, three things to look for. First, separate high volume repetitive work from judgment heavy decisions.

They are very different categories. If someone on your team is pulling the same data, filling in the same fields or forms, and routing documentation a hundred times a day to the same kind of places, that's worth paying attention to. If they're making a very nuanced coverage decision informed by years and years of experience, that's a different conversation. Second, identify where time is actually being lost.

In most insurance operations, it's not the decision-making that slows things down. It's everything that comes before that decision. The data gathering, the manual entry, the document review. That's where the hours and hours go and where AI tends to have the most immediate impact.

Third, define what good looks like before you start. If you don't have a baseline, you can't measure whether anything has actually improved. How long does it take to do this work today? How often does it produce errors?

What do those errors actually cost you to? You need that anchor before you introduce any technology. So map the work, identify where the time is lost, set the baseline. That is step one.

Step 2: Four questions to test AI fit

Chaz Perera, BevayaStep two is about taking everything you just mapped and running it through a filter. Not every high volume or time consuming task is a good fit for AI. So how do you tell the difference? We tend to use four questions.

The first question is the input sufficiently consistent for AI to process. AI is really good at handling structured data. It also is pretty good at handling unstructured data that sits in all these different documents you might have or emails if it's been trained on insurance. ACORD forms, loss runs, SOVs, medical bills come in a variety of formats, but the data inside them is consistent.

That's the fit. Question two is the output verifiable by a human. The strongest AI deployments in insurance keep a keep a keen interest in having humans in the loop especially in the early stages. Not because the AI is wrong but because many decisions we make in the insurance industry are reg have a regulated aspect to them and so that requires accountability and accountability has to rest with a person.

Think about submission intake. AI extracts the key data fields from the application and populates your underwriting system. An underwriter glances at it, confirms it looks correct, and moves it along. That's the loop you want.

AI handles the extraction, a human verifies the output, and the work moves faster than it ever had before. If the output goes somewhere no one reviews, it's a it's a risk calculation and a risk conversation that you need to be willing to have. Question three, does volume or speed matter here? AI value compounds when work is high volume or time-sensitive?

A task you do 12 times a year at your own pace probably doesn't need AI. A task you do 1,200 times a week where every hour of delay costs you a renewal, that's where AI changes the math. Question four, what's the cost of a mistake? This doesn't mean avoiding AI where wherever errors are possible.

It means understanding the risk before you deploy. A misclassified document caught in review is very different from an error that makes it into a policy or a payment. The cost of a mistake tells you how much human oversight to keep in the loop and how often to audit and how confidently you can scale it. AI is a strong fit if volume and speed are high.

The input being consistent and a human having the capacity to verify the output means mistakes are caught. That's the filter. Run every task through these four questions and you'll you'll know very quickly where to focus.

Step 3: Where AI works in underwriting

Chaz Perera, BevayaStep three is where this gets very concrete. We've discussed mapping workflows and filtering for fit. Now, let's see what that looks like in underwriting. On your screen, you see you see six areas where AI is doing real work today.

I'm not going to walk through every one, but I want to highlight a few that come up most often in our conversations. I said six, but I really meant five. submission intake is usually the first place teams look and for good reason. It's high volume, the inputs are predictable and the manual effort is substantial.

AI handles extraction, passes the structured data directly into the underwriting platform and can be extended to the full clearance triage and rating process. policy renewals are one of the are one people don't always think about first, but it's high volume, time-sensitive workflows that fit the filter in a really neat way. and and importantly here, monitoring expirations, extracting updated information, and flagging what needs attention. That's a strong candidate for AI.

Where AI works in claims

Chaz Perera, BevayaWhat about claims? This is where insurers typically feel the pressure first. Volume is very high. Cycle time expectations with your customers and your and and your claimants are intense.

And the documentation flow through a claims operation is enormous. Claims indexing is often the first task teams tackle. Every day, documents arrive across thousands of claim files, including demand letters, medical records, and police reports. Someone has to read them, determine what they are, and route them to the right place.

AI has this magical ability to do that well, and do that automatically thousands and thousands of times a day. One insurer, for example, that we work with saved over 60,000 hours of manual work last year doing exactly this. Legal demands are one area that doesn't get enough attention. By the way, a demand package often arrives buried in a multi-page document and someone has to find it, flag it, and ensure that the right person acts within the required time frame.

Missing that window has real consequences. AI tracks those deadlines, surfaces what it what needs attention before it becomes a problem, and can follow up to make sure that actions have been taken. Claims summarization or claim summary are a newer capability that resonates immediately as well. Instead of reading through pages of notes to get up to speed on a file that was just assigned to you as a claims adjuster, AI can generate a concise summary so that you as a claims adjuster can focus exclusively on the decisions that need to be taken rather than the underlying review.

Where AI works in policy servicing

Chaz Perera, BevayaFinally, in policy servicing, you know, it this is an area that doesn't often get the spotlight in AI conversations. People love to talk about underwriting, love to talk about claims. But let's talk about policy services. Premium audit is a strong example of where AI can accurately extract data from questionnaires, payroll reports, t tax returns, and 1099s at scale.

We're seeing 98% accuracy in data extraction. And that matters because errors here directly affect earned premium. This so so this isn't just an efficiency story. It's a revenue story.

Endorsement processing follows that same pattern. Requests come in classifies the type, extracts the data, organizes the data, updates that information in the system ready for an underwriter to make a decision if the flag is premium bearing.

Example: AI for submission intake

Chaz Perera, BevayaAll right, we've just covered a lot of ground on where AI is used today. Now, I want to make it tangible with a few workflow examples. Let's take submission. Let's look at submission intake.

On the left side is what this looks like today for most underwriting teams. A submission comes in, someone manually retrieves the documents. It gets routed. then there's a request for missing information.

The key key data gets ex gets reviewed, extracted and entered into the platform. By the time the underwriter can begin the review, it might be a few days. On the right is what this looks like. When AI handles the intake, the submission automatically comes in from any source.

data is extracted quickly and and validated and ensured that missing fields are flagged that struct that the data is passed into the rating system pass into the workbench or the underwriting platform in a highly structured way. Clearance and triage are completed automatically. and then the underwriter opens a file that's ready to work. So same outcome just a completely different path to get there.

And importantly here the underwriter is just focusing on judgment not chasing data or having to make new new edits and data entry changes. And so suddenly that quote time can be compressed from days to hours. That's the shift that can happen with AI.

Example: AI for first notice of loss

Chaz Perera, BevayaLet's let's look at another example in first notice of loss. a typically high volume process with a lot of times sensitive moments in a claims claim operation. most carriers do a lot of this work manually especially in prop in commercial property and casualty. So on the left side a claim comes in staff manually logs it gets entered in the system a number gets assigned requests for documents go out maybe one by one a physical file might be set up and then it's assigned to the adjuster to start working on the right claim comes in AI creates a record identifies what documents are needed sends that request out immediately starts to build the file and pre-populates the file with all the information that's available as it's coming in.

So, not just once, but multiple times throughout the day, weeks and years as that claim file develops. So, in other words, claims are set up proactively from day one and then because a claim file is organic, continuously maintained. And adjusters, they're not starting with an empty shell waiting on documents. They're working things as the information comes in, just a lot faster and with a lot less manual intervention.

Example: AI for COI creation

Chaz Perera, BevayaAll right, last example. Policy servicing. Believe it or not, COI creation is is is still one of those workflows that looks simple from the outside, but can quickly when you're doing it hundreds of times a week. it follows the same p before and after pattern.

An insured submits a a request. Staff at an agency or broker has to locate the policy, verify coverage, draft the certificate, review it for accuracy, and then email it back. Hours can pass. With AI, the policy data and the coverage details are verified quickly.

A certificate is generated very quickly and delivered near instantly. The policy holder gets what they need faster. Your servicing team, importantly here, is not buried in in a in repetitive repetitive requests or tasks. And importantly here, it happens whether it was at 10 a.m. in the morning or 10 p.m. at night. That's what AI owning a workflow actually looks like.

The pattern: AI takes the intake, people take the judgment

Chaz Perera, BevayaSo we've just walked through three very different functions underwriting claims and policy servicing each with different teams workflows and pressures. But look at the pattern. In every case AI handles the intake, the extraction and the setup. And in every case your team is handling the judgment.

That's not by coincidence. That's the fit. Repetitive high volume work belongs to AI. Work that requires accountability belongs to your people.

Once you see that pattern, appreciate that pattern, it becomes a lot easier to know where to start.

How to start: one workflow, one team, one win

Chaz Perera, BevayaSo, how do you actually get started? Very common question. the answer is relatively simple and certainly simpler than most people expect, right? Which is don't start with a massive transformation.

Start with a test. Pick one workflow. Don't pick your most complex workflow, by the way. Not the one or or the one with the most political baggage.

Pick one workflow with high volume and clear steps. Something where the problem is obvious, the inputs are clear. Pick one team. Ideally, a team of people that are curious and open to trying something new and willing, importantly here, willing to provide brutally honest feedback.

Then pick one pain point with a measurable baseline. This step makes everything else defensible. You need to know what the process looks like. you you can present a view on what it might look like afterwards.

Understand things like time per transaction, error rates, cost of of correction, volume, you know, whatever the right metrics metric is for that workflow, but importantly established at the very beginning. Then run a focused test and expand based on those results, not hype. If it works, you have a great proof point. You have an internal story that makes the next use case easier to fund and staff.

The goal is one win, something real, measurable, and scalable. That's how insurers who see results actually got there. Not with a big bang, with one workflow done really, really, really, really well. So that's the approach.

Map your workflows, identify the steps. AI can own aspects of a process, see where it's already working across the industry, and start small with one workflow, one team, and one measurable win. If there's one thing I want you to take away from today, it's this. You don't need a perfect strategy to get started.

You do need a starting point. Thanks everyone for joining us today. With that, I'll hand it back to Jill.

Jill Zakarian, BevayaThanks, Chaz. That was great. Hopefully, everyone's walking away feeling like the path forward is a lot less intimidating than you may have initially thought. And before we move into our Q&A, I'd like you to explore the many other sessions available on demand.

And you'll also find a link to our webinar library in a copy of today's presentation slides.

Q&A: How long workflow mapping takes

Jill Zakarian, BevayaAll right. So, it looks like we do have a few questions coming in. give me a moment as I pop these up in my queue. Okay, Chaz, the first question I'm seeing, on the workflow mapping step, how long does that exercise actually take and what do you need in place to do it well?

Chaz Perera, BevayaSo, you know, this is probably a step that most insurers underestimate and insurers in this context, sure, carriers, but I I think this is also true for for brokers and agencies, it's true for TPA, it's true for reinsurers. and you know, it's because because we have a tendency to underestimate it. It's the one that determines whether it determines whether the rest of the work can really really sort of create value. but the good news is it doesn't have to be a long effort with the right focus.

We're talking a week to a few weeks. We're not talking months. The key is getting the right people in the room. And importantly here, I don't mean people who manage the process.

I'm speaking specifically about the operators who actually do the work every single day. Like I said in, you know, in one of the earlier slides of the webinar, that 20-year professional who just knows what the workarounds and has never put it down on paper, but probably has spoken to it many times. That's the that's the friction in the process. That's the workarounds.

That's where the real change lives. And so it's really important to bring those people into the room, pair them with the managers who can speak to volumes and SLAs and other expectations around the process. That will quickly paint the clear picture of what the of what the the process needs to be and what the work workflow needs to look like. I don't I'd also say by the way the other the other thing that slows teams down is trying to get really really perfect.

You don't need perfect. You need defensible. Some some some examples are volumes, times per transaction, error rates. It doesn't need to be spot-on.

You have a good inclination of what those numbers are. and and and by the way, as you start to run the process, you'll start to get more detail and more data that will get you a better number to shore up.

Q&A: When a task is too judgment-heavy for AI

Jill Zakarian, BevayaFantastic. the next question that I'm seeing is how do you know when a task is too judgment heavy for AI versus just needing better AI?

Chaz Perera, BevayaYeah. so I suppose the clearest signal that something is genuinely judgment heavy is when the right answer depends on a lot of context. And that context isn't captured in the data or the documents immediately in front of you. An underwriter, for example, deciding whether to write a risk based on a conversation they had with a broker last year or an adjuster weighing a coverage decision based on how likely citiz how likely a situation might develop.

To me, those things are that's human judgment, right? That that's where experience and understanding comes into play. Those two inputs I just gave you examples of aren't fully visible to the the average person, right? It requires time and experience and so why would it suddenly be visible to AI?

So, I would say that's how you should think about, you know, what's a genuine judgment heavy decision. I think where teams get this wrong is when they label something as judgment heavy just because it's a complex process or because it has many exceptions to a process. And and I just want to make this the distinction here that complex isn't the same as judgment heavy. A lot of what looks like judgment is actually just pattern recognition at scale.

And that's exactly what AI is very very good at. So if someone on your team can explain the rules behind a decision, even if those rules are nuance, AI can probably learn them and learn them very quickly. So the better question maybe is to ask isn't isn't is this judgment? It's what's happening when AI gets it wrong.

If a mistake gets caught in the review before it has consequences, that's a workflow you can bring AI into with the right human oversight. if mistakes create a compliance issue, a coverage dispute, or potentially a customer problem that might be hard to unwind, you know, I think it's really important then to think about how you keep a human in that decision matrix.

Q&A: What AI implementation feels like for the team

Jill Zakarian, BevayaAwesome. Next question. I think we might have time for one more question. the before and after looks clean on a slide. What does implementation actually feel like for the team going through it?

Chaz Perera, BevayaWell, I suppose nothing ever is as clean as a slide makes it look. the technical side I would suggest is usually the easiest part. The people side is the hard part and it's the long pole in the tent. When you introduce AI into a workflow, you're changing what's how someone perceives their own job and their own day-to-day.

Even when that change is actually positive, for example, you're just taking a bunch of bunch of grunt work off somebody's plate. people still need time to build trust that that work will be done correctly. So the first few weeks teams tend to override everything. and and in some sense they're like they're looking for aha and gotcha moments.

but you should you should accept that for what it is and consider it to be a healthy exercise because ultimately it's the thing that builds trust between the human and and the AI. and you know, I think it also means that it c it'll catch the teams off guard, right? You let's let's remember here if you've done this well, you've mapped the workflows, you've sat down with the experts and really understood what the process really looks like. but there will always be edge cases and those edge cases will always cause some amount of friction.

so what's important there is that it's clear who owns those edge cases in the organization and who's and who who will work with the AI to educate it or or take over from the AI when appropriate. la last thing I'd say if you do this well it the conversations internally will move from does this work because that's what a lot of conversations start with to hey where else would we do this right so once a team starts to trust the output and the outcome and the quote-unquote decision making of the AI in the context of this the you know the the the confines you've given it then they're going to come to you with lots and lots of other ideas where they where they would be excited, frankly, to to apply things.

Jill Zakarian, BevayaAwesome. And we're going to stop there. We appreciate you making time to log on and join us today. For any questions we didn't have time to cover, we'll reach out to you after the session.

As a reminder, we'll share a recording of today's session by tomorrow afternoon. And we hope to see you at a future session. And we hope you have a great rest of your day.