Top 3 Hurdles of AI Deployment in Insurance (2026)

Intro: top hurdles of AI deployment in insurance

Chaz Perera, BevayaOkay. So, wherever you are with AI right now, this session should be useful. Some of you could still be scoping your first use case. I suspect most of you are mid pilot and hopefully some of you already have something live and are trying to push it further and further.

The challenges I'm going to walk through show up at every one of these stages. So none of this depends on being far along. If anything, the earlier you see these coming, the less they will ultimately cost you. We'll cover the three hurdles that slow AI work down most in insurance, why they show up in this industry specifically, and then go one at a time through each hurdle in depth.

We'll close with what it actually costs you to leave these hurdles standing. So with that, let's get into it.

What's slowing insurance AI down

Before I name these three hurdles, think through what challenges you have been facing lately. Go ahead and drop it in the Q&A widget. What's the thing that's really slowing your AI work down right now? Not the version you put in a board deck per se, the real thing, the blocker that you keep running into.

I'm asking because I've had this conversation with a lot of insurance leaders across underwriting, claims, policy servicing, IT, AI, and the honest answer in the room is almost never the model isn't good enough. The technology is rarely the thing. It's something more practical, something operational, and it tends to be one of a small handful of things that come up again and again. So, take a second and again in the Q&A widget, if you're willing, tell me what's in your way and let's see how closely it matches to what we've found when we asked around the industry.

All right, so these are coming in. It looks like some of you are dealing with figuring out what the human does versus the AI. How does this fit in? How does AI fit into legacy tools?

Some of this is around being aware of regulations. So, it looks like some of you are thinking about governance. And I'm seeing a mention of measurement. So I'm assuming most of that falls under the broad umbrella of ROI.

That's the first hurdle of the three in our list. So let's get to that list.

Top 3 hurdles to AI adoption in insurance

All right. So we asked insurance leaders about the biggest challenges to AI adoption in our "State of AI Adoption in Insurance" report that we released earlier this year and the answers were consistent. It didn't matter company size. It didn't matter which product lines people were focused on or even how far along people were in the AI journey.

So the first is an unclear return on investment or business case. The second is data quality and data availability. And the third is security and data privacy.

Why insurance faces these AI hurdles

There's a reason these hurdles cluster the way they do and it's not that insurance teams are behind. It's the shape of the industry we happen to work in. Insurance is regulated. It's data heavy.

And because it's data heavy, decades of that data live across policy systems and claim systems, spreadsheets and PDFs. Some things are sitting in your emails. And the truth is like these formats, they don't necessarily line up. And because it's regulated, you can't just point a tool at the data and hope.

Every step has to hold up to some compliance standard and eventually to an examiner. It's also an industry where decisions carry real financial and legal weight. If you get a claim or a policy decision wrong, and it's not just an internal miss. It can cost a customer their coverage.

It could theoretically land you in front of a regulator and outcomes take a long time to validate. You often don't know if a decision was the right one for a few months, maybe even a few years in some instances, which leaves less room to move fast and to fix later. So AI doesn't get a free pass here. It has to earn its way in the way a person, a new person on your team earns trust by being right on real work, by handling sensitive material responsibly and proving their value, right?

You wouldn't hand a brand new hire the hardest claim on day one without at least watching their work. I suppose AI is no different. That's not a knock on AI. That's just the bar in this industry.

And honestly, it's the right bar. Every hurdle we're about to walk through is a version of AI earning that trust. So, let's take each of them one at a time and we'll start with ROI.

Hurdle 1: Unclear ROI and business case

Almost everyone believes AI can help. Far fewer can put a defensible number on it or on how fast it pays back. And without that number, it's hard to get the go-ahead to start and then later to scale beyond a single use case. So the enthusiasm is real, but it stalls right at the point where you commit budget and other resources.

Leaders rightly need justification before approving a big project like bringing a new tool especially an AI tool since it's relatively new territory. What it does cost, what it does save and what it's going to take are these like core questions. A CFO isn't going to fund a feeling. Belief doesn't really get you the budget, right?

So a business case does and a business case has to survive someone poking holes at it or in it. So how do you move past it? Start where the proof is easiest is our best advice. Pick a repetitive high-cost workflow.

The kind of high volume manual work that's expensive precisely because it eats so many hours of your day. Those workflows have two things going for you. The current cost is easy to measure because you already know what it takes today and the improvement is dramatic enough that nobody argues with it. That's where savings are easiest to prove and hardest to sort of wave away.

And once you can point to a clear number on one workflow, you've done more than save money on that one process. You've built internal credibility to earn and fund the next one and the one after that and the one after that. So the first number is the one that unlocks everything else. Once you know your use case, define the number you're going to track before you start.

Let me give you a couple real examples of what I'm talking about. A US P&C carrier doubled its submission volume by cutting document processing time by 87%. A top insurance brokerage cut manual COI processing effort by 80% and saved more than $2 million versus outsourcing that work to a third party. And a regional workers comp carrier brought claims indexing from five days down to hours, saving more than 22,000 hours in the mail room.

None of these, by the way, started as a grand transformation. Each one started as a single well-chosen workflow where the number was very easy to prove.

Hurdle 2: Data quality and availability

So hurdle two, data quality and availability. Here's what it looks like day-to-day. The data you need already exists, but it's scattered across sources. It likely comes in inconsistent formats and it sits in systems that were never built to hand it to AI.

Commonly, that means it's locked inside a legacy system. The same field is named three different ways in three different systems for most of you. Some of it might be clean structured records, but in our experience, most of it is locked in PDFs and scanned documents and emails and attachments to emails in a data lake, stored, not sorted and sourced. So, you know, it's all there somewhere.

Getting it into one clean usable shape is the hard part and unfortunately it's a part nobody really enjoys. So that's the underlying condition. But here's the part that catches teams off guard. This hurdle usually doesn't show up during the pilot.

In a pilot, you handpick a few very clean examples. You point the tool at them and everything looks phenomenal. It becomes visible the moment AI moves from pilot to production. However, that's when it meets the full mess of real everyday data.

The edge cases that are really in people's heads, the odd formats that you weren't expecting because you don't work with particular carrier, particular broker on a regular basis. You know, the one carrier who sends everything as a photo, for example, or the one customer who sends everything as a photo. Those are the sorts of things that don't come up in a pilot. So teams ultimately get a false read early and they feel really good, but then they get to production and they immediately hit a wall.

So, the shortcut to moving past this one is using data models built for insurance right from the beginning. When the system already understands what a policy is, what a claim is, what a submission looks like, it cuts the cleanup and mapping work dramatically because it isn't learning the industry from scratch. That's what lets you move without stopping to run a separate months-long data cleanup project before you can actually get to value. You make the data usable as you go on the workflow in front of your eyes.

And once you know the use case, map the data behind it so that you can start to move really quickly.

Hurdle 3: Security and data privacy

The third hurdle is security and data privacy. Since you're working with some of the most sensitive, most regulated data there probably is. Putting AI anywhere near it raises real questions about access, vendor risk, compliance. So questions like who can see this data?

Where does it go? What is a vendor going to do with it? Can we prove all of that to a regulator months from now on a specific transaction? Those are the questions you really need to be able to answer for yourself.

And every one of these questions is a fair question. They're going to land on either your desk or someone else within your organizations. Usually somebody in a security, compliance or AI risk or enterprise risk role, right? And they need to get comfortable before you can start to move things forward.

Because the data is sensitive and tightly regulated, the stakes become very real. Let's remind ourselves this is policyholder information. This is claims detail. This is health and financial records, right?

Really sensitive information. And like the last hurdle we talked about, it intensifies as you go into production. A small pilot with a handful of files feels contained. And nobody's going to really worry about that.

The moment the AI is touching real world volume, the security and privacy questions get louder and louder. A lot more people are going to start asking where's this information going, what's the vendor going to do with it, etc., etc., etc., including, by the way, people that probably weren't in the pilot to begin with. So, how do you move past this one? You got to treat governance as the thing that lets you move faster, not the thing that slows you down.

It's a mindset shift. When you have strong governance and clear accountability in place from the start, security stops being a gate you're stuck at and becomes the enabler that lets you move forward with confidence. Concretely, that means three things. Access controls that limit who sees what.

Audit trails that show exactly what the AI did and why on every transaction. And human oversight built into the process rather than bolted on after the fact. Put those in place and the questions flip. It stops being is it safe enough to expand and becomes where do we expand next?

And it means knowing upfront what your legal and compliance teams will need to see to make not just a pilot but production successful.

The cost of standing still on AI

All right. So, these three hurdles, don't hold still, by the way, while you wait. Leave the ROI question unanswered and another initiative gets this quarter's budget instead of yours. Leave the data messy and it doesn't clean itself up.

It gets messier the longer nobody touches it. Run a pilot without governance in place and what could have been approved upfront becomes something you have to go back later and redo. None of this fundamentally gets easier by waiting. If anything, it just gets more and more expensive.

So, you don't have to solve all three at once. Pick the one that's costing you the most right now and start there. That's usually the one most in your way and clearing it will earn you the room and allow you to move to the next opportunity. Okay, running into these three hurdles isn't a sign you're doing AI wrong, by the way.

It's a sign you're tackling the right areas. And a quick note on the Bevaya side before I hand it back to Jill. Earlier I mentioned insurance-specific data models. InsurGPT is ours, trained on real insurance documents instead of the open web.

It's what's powered our platform across 120 deployments in production at scale. Not pilots in production at scale. That's where the governance and the provable ROI come from. So if one of these three is the hurdle in front of you, there's a good chance we've seen something close to it before.

If you want to talk through it, click the Request a Demo option. And with that, thanks everyone for spending this time with us. Jill, I'm going to hand it back to you because I know that we've got a bunch of questions from folks out there.

Q&A

Jill Zakarian, BevayaYes, we definitely do, Chaz. And before we hop into our Q&A, I'd like to invite you all to explore the many other sessions we have available on demand. You'll find a link to the Bevaya webinar library in your copy of today's presentation slides. All right, Chaz, let's hop into our Q&A and we do have quite a few questions, so hopefully we can get through a few of them.

I'm just going to go in order as I see them. Okay, so first question you said start with a repetitive high-cost workflow. What if we don't have one that's obviously high cost? How do we find it?

Chaz Perera, BevayaI would suggest it's highly unusual and maybe uncommon to not have a high cost workflow. I suspect it's that somebody just hasn't done the math to realize that workflow X is an expensive workflow. So maybe the questions you can ask yourself to get to that answer without having to do a lot of the work is where does your team spend most of its time doing repetitive things? It's not the most visible work.

It's just stuff that people are heads down doing all the time. And examples of this would be like generating certificates of insurance. Classifying claims documents as they come in via email or snail mail. Submission ingestion is a good example of this kind of work.

A lot of people doing it, a lot of time to do it. Sometimes it's the wrong type of people doing it like claims adjusters and underwriters that you want focused on other things. So that's like how I would sort of shape what are those expensive processes? The other tell: what annoys your teams like what's the work?

What's the task that they're just like, "Oh god, I got to do this thing again." It's going to be the things that are high volume and manual and the things you want to focus on. So once you have those candidates, you're going to go through an exercise to size the cost. Most of that is going to be associated with people doing the work and the time associated with them doing that work and where they could otherwise spend that time.

Jill Zakarian, BevayaAll right. Awesome. So next question. You said the data problem doesn't show up until production.

Is there a way to catch it during the pilot instead of getting surprised later?

Chaz Perera, BevayaYes, but it takes doing something I suppose that feels counterintuitive, which is don't pick your best examples for the pilot. Don't pick the pristine and the clean, right? The instinct in a pilot is typically to take your cleanest use cases, your cleanest data because you want it to succeed. But that's often what ends up hiding the problem.

So if you want the pilot to tell the truth, you have to feed it the ugly stuff and you have to do it on purpose. So, you know, include the scanned PDF that, you know, when you visually look at it, it looks messy. Include the carrier who formats things differently every single time. Include the record with three different names in the same fields.

Right? If the pilot can hold up to that, you've learned something and you've done something substantive and then when you get to production, you're less likely to face the same volume of challenge. There'll still be challenges, but you're less likely to face the same volume of challenges because you've gone through some of the dirty work, so to speak, right up front.

Jill Zakarian, BevayaOkay, awesome. So, next question. If all three hurdles are blocking us at once, which do we tackle first?

Chaz Perera, BevayaIf I had to pick one, I'd start with the one that's actually stopping you from starting. I know that sounds obvious. It doesn't necessarily mean it's the biggest problem. It's just the one that's blocking you.

You know, for example, security and compliance conversation can happen in parallel with almost everything you're doing in as you go from idea to pilot, pilot to production. And you can be having that conversation with legal while you're also proving out the workflow. You know, data quality sort of falls in that same bucket, right? Which is you can start narrow.

Pick some hard ones but start narrow and then you know start to expand it. And you know if you've chosen the use case well you can clean the data as you're going right but you know in my answer the obvious thing here is I left out ROI and ROI is the one that usually is load-bearing for the other two. Without a number to point to, you're not going to get budget. You're not going to get the attention of the powers whether they be at the executive level or at the governance committee level to generate the conversations you need to start the data conversations, to start the legal conversations.

So, you know, look, if you're stuck in all three, start with ROI. It's gonna unlock the others is really what I'm saying.

Jill Zakarian, BevayaAwesome. Okay. So, Chaz, the next question I have, it's kind of long, so if I do need to repeat it, no problem. Okay.

So, the whole case on the data hurdle rested on starting with models already built for insurance. We have data scientists. Can we just fine-tune one of the big models on our own? Or sorry, can we just fine-tune one of the big models on our own documents and get that same insurance-specific edge?

Chaz Perera, BevayaJill, just so I'm following. So the question was data is a hurdle. Can internal data scientists for example fine-tune using other models?

Jill Zakarian, BevayaYeah. Yep.

Chaz Perera, BevayaOkay. So look, it's a good instinct. It's an appropriate instinct to have. You know, you can't really fine-tune a general frontier model, right?

They're closed weight models. What you can do is tweak API parameters, but that's not really fine-tuning. You know when you talk about fine-tuning what you're really talking about is leveraging open-source, open- weight models and then applying a variety of different techniques to be able to you know design something for your very specific use cases. I just would suggest to you it's not simply doing the model.

The model is one part of it, but there's more complexity to these insurance documents than we sometimes appreciate. For example, insurance documents are among the most complex from an unstructured data point, right? Hundreds of different formats, inconsistent layouts. It's a mix of like nice typed text and then handwriting.

Sometimes it's multiple documents actually contained in one document and how do you separate those? Sometimes it's tables and text and images and sometimes it's information nested within other information. All of that is, you know, data science. It's engineering.

And that stuff could take months and months, if not years, to get to a good place before you can start to take advantage of the, you know, applying AI to the use case you're interested in. Then you start to think about okay well certain types of use cases require more than just what's the answer right you want to validate that the answer makes sense in the context of the work you're doing so now you need to start to think about some type of verification layer or like a champion challenger sort of mentality in the way these models work and work against each other. So again lots of complexity around engineering. So I don't think it's as easy as just being like hey let's can a data scientist get there.

I think they can but the question you have to ask yourself is was that time to get them there to something that's production grade and ready really worth it when you could be partnering with folks that will bring you something that's more rich and more complete so you can move. Right? That the exercise here is time to value. We live in a world where it takes a long time typically for insurance companies to get to production.

And if you can't accelerate that time to value, you're going to lose people in the exercise. Not just the CFO who's like, hey, we committed to a dollar, but the people who are, you know, you're trying to engage with to get the AI to the right place and are excited about getting work off their plate. That's not going to happen fast. And that's the change management exercise that you risk by just trying to do it all on your own.

Sorry, Jill, that was a really long-winded answer.

Jill Zakarian, BevayaNo, that was fantastic. All right, everyone. We do have to wrap up here. We appreciate you making time to join us today.

And so for any questions, and I know there's quite a few that we didn't have time to cover. We'll reach out to you after the session. Also feel free to email us any questions you may think of later at resources@bevaya.ai. And as a reminder, we'll share a recording of today's session by tomorrow afternoon.

We can't wait to see you all again soon, and we hope you have a great rest of your day.