How to Measure AI ROI in Insurance

Why ROI is now the second most important criterion

Chaz Perera, BevayaSo according to our "State of AI Adoption in Insurance" 2026 report, when insurers evaluate an AI solution, ROI is now the second most important criterion. Uncertainty around the business case and ROI is also a top barrier to adoption. So, the question isn't is AI interesting or did we run some pilots. It's did it move a number that matters.

That's what today is about.

What this session covers

Here's where we're headed. We'll walk through a simple two-track framework for thinking about AI ROI. We'll talk about two different lenses to apply, what to do before you even launch anything, when you should actually expect to see results, what to track while you're waiting on those results, how to translate what AI is doing into the language finance already uses, and then we'll close with what to do next.

What teams say when you ask about AI ROI

So, here's the thing. When you ask teams about their AI ROI, this is what you get back. We completed 12 AI pilots this year, or AI was in our CEO's annual letter, or we logged 2,000 hours of AI training. All true, all great, but none of them answer the question that was actually asked.

None of those are the same as this question. What did your AI program actually return? Not effort, not activity. What did it return?

And if you can't answer that clearly right now, you're not alone. But you also can't stay there much longer because eventually someone with budget authority is going to ask.

Two tracks: automation and augmentation

So let's fix that. Here's the framework I want you to leave with today and it's genuinely this simple. There are two tracks and almost every AI initiative falls into one or both of them. When an initiative does both, you measure the two effects separately rather than picking one label for the whole thing.

Track one: automation, or cost out

Track one is automation or cost out. This is AI doing work that a person used to do. hours get redeployed. Contracts you don't need anymore. More of your volume going through without a human touching it.

Less rework. This hits your expense ratio and your loss adjustment expense line. The question you're asking here is are we doing this cheaper or faster?

Track two: augmentation, or value in

Track two is augmentation or value in this is AI making your team's judgment better not replacing it less claims leakage because the data is complete lower bad faith exposure better quote-to-bind ratios sharper risk selection this hits your loss ratio your combined ratio your growth line if you're not property and casualty translate that as your loss ratio equivalent medical loss ratio if you're in health mortality or morbidity experience if you're in life combined ratio specifically as a P&C member but the logic is the same everywhere augmentation hits value not just cost the question here is are we creating value a person alone couldn't if you blend your automation and augmentation results together into one number and report it up you can't tell which type of AI is doing what that matters since they hit differently in the financial line. So if you say our AI saved us $2 million without separating which track it came from, nobody can verify it, nobody can build on it. And when leadership asks a follow-up question, you might not be able to answer it.

What automation looks like in practice

So, let's stay on automation for a second because this is usually the easier and faster one to prove. Think about straight through processing. Claims, submissions, policy documents handled start to finish with no one touching them, hands-free essentially or with very light touch. Think about cycle time and handling costs coming down because your adjusters and your underwriters spend less time per file.

Think about document intake, mail, first notice of loss, invoices getting read and routed automatically. The common thread in each of those is that AI is taking on the volume. The work that's document heavy, rules-based, and doesn't necessarily need a human's judgment. Here's what that looks like in real numbers.

A regional property casualty carrier processes 14,000 claims invoices a year. The average handling cost is $15 for each invoice. Automate 90% of that and you've taken out about $204,000 a year out of your loss adjustment expense, without having to hire anyone, or frankly lose anyone. We've seen this directly.

Eastern Alliance took claims indexing from 5 days down to an down to under an hour and saved more than 22,000 hours doing it. A Fortune 500 carrier recently hit 99% straight through on claims acknowledgements and saw a 246% return in ROI in six months.

What augmentation looks like in practice

Now the second track augmentation this one's less about speed and much more about quality of decision substance. Take claims leakage reduction. When your data is complete and structured, you catch payment errors and coverage gaps that a manual review likely will miss. Things that would otherwise slip through.

Or think about bad faith exposure. Accurate timestamped documentation lowers your litigation risk and it lowers your reserve exposure on the complicated claims, the ones where a dispute could otherwise drag on. And then there's risk selection and quote to bind. Cleaner submission data means sharper underwriting calls and faster turnaround means you're winning more of the business you actually want instead of losing it to whoever got the quote out fastest.

The idea here isn't that AI is doing the thinking of your team. It's actually sharpening the judgment of your team. The judgment that they're already applying. They're the ones making the call.

AI is just making sure they're making it with the best information possible. Let's look at an example. A commercial lines carrier recently ran claim to policy comparison against 40 million in annual payments. AI surfaced 2.4% in billing errors and coverage gaps.

That's roughly $960,000 a year in reduced leakage. Another example, an enterprise carrier avoided 1.3 million in cost through better document classification and indexing, not by processing faster, but by catching what it would have otherwise gone through incorrectly. So, augmentation is the harder story to tell in a boardroom, but it's often the bigger number.

ROI versus total cost of ownership

All right. Now, one distinction people mix up is ROI and total cost of ownership or TCO. They're not the same. they're not the same question. They're not the same that but but you do need both.

ROI answers what this specific initiative is actually worth. Was it worth it? That's a per use case question. If you're if your go or no go decision, it's it's kind of like your no or go no-go decision on the deployment.

TCO or total cost of ownership answers something bigger. What does running AI at scale actually cost? This factors in things like implementation, integration, training, ongoing operations, and by the way, the piece that people typically forget, technical debt, that adds up across every initiative you have, not just the one you're looking at in the moment. This distinction matters most the moment someone asks, why don't we just build this ourselves?

Because an internal build carries engineering cost, annotation cost, compliance costs, retraining costs. None of which show up if you're only looking at a single initiative's ROI. You have to zoom out to really answer the question around total cost of ownership. Honestly,

Three things to do before you launch anything

all right, here's a part that can trip up a team and it's not technical. It's all about sequencing. You cannot prove what you didn't measure first. So, three things before you launch anything.

Name the financial statement line first. Expense ratio, loss ratio, growth, retention, whatever it is. Decide it before deployment, not after the results roll in. Capture a baseline before you launch, a hold out period, a comparison period.

It has to exist before go live. You cannot reconstruct it afterwards. That data is just gone. And get finance, by the way, to sign off and on how you're going to measure this.

Do it now while nobody's incentives are on the line so the number isn't challenged down the road.

Timeline: go-live day is not results day

All right, let's talk timeline because expectation setting here saves you a lot of pain later. This matters more than people give it credit for in that same adoption report I mentioned earlier at the at the start of the session insurers named how quickly AI delivers ROI as a critical success factor timeline isn't a footnote to the ROI story it's very much part of it the industry average from project start to go live is 12 to 18 months that covers vendor evaluation data prep, model training, integration. Most of that time is happening before AI is ever live. What purpose-built AI with purpose-built AI, meaning AI that's already trained for insurance workflows, that go live window compresses from 8 12 to 18 months to 8 to 12 weeks.

But here's the part leadership needs to hear before you before launch, not three weeks after. Results don't show up on go live day. AI going live just means it's processing volume and the workflows are running. If you're using a purpose-built AI tool, then you can start to look at numbers around the 90-day mark after go live.

Set that expectation early. It's a much easier conversation before launch than it is when somebody's already asking why the number hasn't moved.

What to watch during the first 90 days

So, what do you actually watch during the 90-day window? You don't need to wait for the P&L to move to know whether this is working. You just need metrics that are measurable from day one and tied to the financial line you already named. Here are some real world examples.

How long is it actually taking from submission to quote to a quote decision? The faster that moves, the more business you win. Then are the data fields getting pulled correctly. That's your extraction accuracy.

And bad data here causes rework. It undermines everything else downstream for the files that still need a human to intervene. Is that person moving faster, too? That average handling time should ideally be dropping right alongside everything else.

Are your time-sensitive items landing inside their window? That's SLA compliance, and it connects directly to your bad faith exposure. How often is a human review actually getting triggered? That's your manual intervention rate and expect it to fall as the AI learns on your data.

And lastly, how long from an endorsement request to issuance? with the right setup, same day is very much achievable here. And I'll just plainly say it, Bevaya's commitment is over 95% accuracy and over 70% straight through processing within 90 days of go live. It's not a projection. It's not a hope.

It's in the agreement.

Translating it into the language your CFO already speaks

So, okay, let's connect all this let's let's connect this all the way back to the language your CFO already speaks. Think of it like an equation. Expense ratio is your automation side, loss ratio is your augmentation side, and them together, you get your combined ratio, the number leadership is actually watching. That's the property and casualty version.

If you're in life or you're in health insurance, the same math applies to your numbers, just under different names. Here's a table putting everything we've covered together into one handy view for you. This is how you can map out what AI is delivering. You have your operational metrics, the financial line your finance team tracks, and whether it's automation or augmentation, it's consistent.

By the way, we've included a couple examples here for life and health, too, since the same logic holds regardless of your line of business.

What to do next

So, as you leave this webinar today, what do you do with this information? What are your next steps? Six things you should think about. One, keep the two tracks apart in how you report on them.

Separate by outcome, not by initiative. Keep automation and augmentation results apart in how you report them since they show up on different lines of the financials. And blurring them together muddies both stories. That said, if a single initiative touches both tracks, claims indexing is a great example of this, which can cut manual hours and also catch documents that used to get missed at the same time.

That's fine. Just measure each effect on its own against its own against its own baseline so that it lands on the right financial line it's supposed to be moving. So again, the initiative can be mixed, the measurement cannot. Two, before you go live, know exactly which number you're chasing.

Is this about your expense ratio or your loss ratio? Pick one going in because if you didn't set a baseline up front, there's going to be no going back to create one later. Three, tell leadership now before anything launches that the day AI goes live and the day you'll have real results to share are two very different days. Get that expectation set early so nobody gets caught off guard.

Four, loop finance in before you have a single result to point to. Hash out how you'll actually measure this while it's still a hypothetical. That conversation and that convention goes a lot smoother before there's a real number people will start to argue about. Five.

From the moment you go live, know what you're watching. Straight through processing rate, extraction accur extraction accuracy, handling time, SLA compliance. Those are the early signals that will tell you it's working and frankly will surface themselves well before the financial results catch up and confirm it. Six, take whatever results you've already got and run them through the translation table we covered on the last slide.

So every single one lands on the on the financial line it's actually supposed to move. Nothing should be left dangling or unconnected to a number that matters. All right, thanks so much for spending this time with us. I know everyone's calendars are packed, so I appreciate it.

Thanks again everyone. Jill, I'll hand it back to you for any questions we might have.

Q&A opens

Jill Zakarian, BevayaThanks, Chaz. That was great. We hope this helps you get to the answer of what your AI program returns and how to show those numbers with confidence. All right, so let's move in to our Q&A.

And Chaz, we do have a few questions that came in.

Q&A: Two baselines for one initiative without doubling the work

Jill Zakarian, Bevayathe first question that I see in my queue is practically speaking if one initiative is producing two separate numbers say cost savings and leakage reduction how do you set up two baselines for it without doubling your measurement workload?

Chaz Perera, BevayaLet let me start by saying it's probably less work than it sounds. Like certainly the way I'm going to explain it, I promise it's it's less work than that because you're not really running two separate measurement projects. You're just trying to capture two different things during that baselining period, you know, before you launch. a real world example with our customers take claims indexing. before you go live, you track two numbers over the same window, average handling time per claim or per document. which is your automation baseline and then you know your current mistake rate or error rate you know how much you're missing or misclassifying which is your augmentation baseline. Again, same time period, same file sample, same effort to pull the data.

So, you're not really doubling the work. You're just not stopping at the first number you find. Now, once you're live, you compare each one against its own baseline. If handling time is dropping, say from, I don't know, 12 minutes to four minutes, that's your automation number. and you can see how it would improve your expenses. if leak if the error rates are dropping from say 3% to a half a percent.

That's your augmentation number and you can see how it's improving your loss ratio for example. So two results one initiative and each one is defensible on its own because you know exactly what is being me what it's being measured against. the only real discipline required is deciding before you launch which effects you expect an initiative to produce. and if you only are thinking about cost savings then only capture the cost savings baseline. If if if improving leakage or you know improving some other metric that's more focused on augmentation then prove that fact out. there are clearly opportunities to affect both and both upfront. So capture both baselines frankly because it's not unusual for there to be quite a few use cases that fall into both spaces.

Q&A: Already live and never captured a baseline. Is it gone?

Jill Zakarian, BevayaAwesome. All right, Chaz, the next question I see is what if you're already live and never captured a baseline before launch? Is that number just gone or is there a way to recover it?

Chaz Perera, Bevayaso this isn't unusual and something we try to help our customers stay ahead of meaning get to before go live. look, the honest answer is you can't get back the number you didn't capture, right? There's no reconstructing what your handling time or leakage rate looked like six months ago if nobody was tracking it to begin with. But that doesn't mean you're stuck with nothing. It means you shift from a before and after comparison to a right now comparison. and you know here's what that would look like. instead of comparing today against a baseline you don't have you build a control group today.

So pick a subset of your work that is still being handled the old way. Maybe it's a region or maybe it's a particular line of business or maybe it's just one team that has yet to be onboarded with the new way of working. run those in parallel with the AI against the same volume or similar volume within a predefined window of time, 30 days, 60 days, 90 days, for example, because that will give you a real comparison that you can start to measure against instead of trying to recreate something that may never have existed to begin with. Now, if you genuinely can't run something in parallel, it's not uncommon for people to turn the AI onto everything all at once, then I think the best opportunity there is to fall back to external benchmarks. Industry averages for handling time or leakage rates aren't as precise as your own, of course. but they're not wildly off that you couldn't use them as a defensible reference point instead of just asserting a number with nothing behind it. the bigger point I suppose here is don't let this be the reason your next initiative repeats the mistake.

Whatever you're rolling out next, capture the baselines before it launches. Even if it's imperfect, even if it's just for two weeks of data, it just has to represent what's happening in your real environment. An imperfect baseline still beats no baseline every time. And it's a lot cheaper to grab now than to try to manufacture it later.

Q&A: Getting finance aligned on how you measure

Jill Zakarian, BevayaAwesome. I think we have time, Chaz, for one more question. okay. So, a lot of this depends on finance agreeing with how we measure. Any advice on getting them aligned.

Chaz Perera, Bevayatiming matters, but so does what you actually bring them. So, I would say don't walk in asking finance to bless them bless a measurement philosophy in the abstract. bring them a specific proposal, something like we're measuring this against handling time per claim baselined over a 60-day period before we went live and here's the sample. A a concrete proposal is something finance can approve or just or push back on. A vague how should we think about measuring this is a conversation that's just going to stall. it also helps to bring the right person.

So not just any finance contact. You want whoever owns the number downstream, the person who actually will be asked to defend the ROI, defend the measurement in a lead in a leadership setting. and it's important, I suppose, in that sense, for it to be someone who's not one or two steps removed from the work, right? So, if it's a I don't know, if it's an underwriting initiative, it's not uncommon for that underwriting line of business, if not that team, to have a finance counterpart. And it's really that individual that you want to engage with in this exercise.

And look, if you get any push back on your proposed baseline or methodology, it's a great outcome. Believe it or not, it's not a setback. It's good to have the disagreement while it's hypothetical. It's good to evaluate the right balance because it has to be measurable.

It has to be it has to be seen in the numbers. and frankly, your reputation, their reputation is attached to it. So you want to get it right and getting your finance team to pressure test your measurement approach and methodology up front, they'll become your biggest defenders. and sooner or later that's going to be critical in affording you the opportunity to go after the next thing and the next thing and the next thing.

Jill Zakarian, BevayaFantastic. All right, we are going to wrap up here. We really appreciate you making time to join us today. Also feel free to email us any questions you think of later at resources@bevaya.ai.

And we can't wait to see you again soon. Have a great rest of your day.