Get a demo
Select the type of demo you'd like to experience with Indico Data.
Select the type of demo you'd like to experience with Indico Data.
In episode 23 of the Unstructured Unlocked podcast, Tom Wilde, CEO at Indico Data and Laura Drabik, Chief Evangelist at Guidewire Software, discuss how AI is changing the way insurers handle commercial lines underwriting. With growing pressure to reduce risks and costs, they explore how turning unstructured data into structured information can improve underwriting efficiency and help insurers make better decisions faster. Tom shares how AI, including generative and agentic models, can streamline underwriting, improve pricing, and boost profitability, all while ensuring that human oversight remains key to the process. Tune in for a clear look at how AI is shaping the future of underwriting.
Tom Wilde: Welcome to Unstructured Unlocked, a podcast where listeners discover how insurers are entering the decision era, utilizing artificial intelligence to refine their decision-making processes, boost underwriting profitability, and achieve premium growth. Hi, Laura. Great to be here.
Laura Drabik: Well, it’s great to have you. So commercial lines, underwriting performance continues to deliver strong results for p and c insurers with revenues up 8% annually over the last five years, and a combined ratio of 91%, which is actually very strong, but the sector faces mounting pressure to reduce exposure to risk and lower costs while driving profitable growth. Yet according to some recent Capgemini research, 77% of insurers, they struggle to achieve pricing precision due to an inability to leverage relevant underwriting data. Tom, the availability of data isn’t necessarily the issue. I mean, we got lots of data. It’s that so much of the data is unstructured creating what you’ve called a Tower of Bible problem. Can you help listeners understand what you mean by that?
Tom Wilde: Yes, absolutely. I think especially in commercial and specialty, you see this where you have many, many parties involved in any kind of risk assessment, be it the insured, the broker, the carrier. As you point out, there’s lots and lots of data, but generally it comes in all sorts of different formats. Nobody has the ability to sort of dictate a standard. There are some quasi standards like Accord, et cetera. But generally the art here is taking all of this unstructured data and making it high fidelity structured data because at the end of the day, structured data is what drives all the decision systems that surround it, be it pricing, rating, policy, admin, et cetera.
Laura Drabik: So in commercial lines, a lot rides on personal relationships and deep underwriting proficiency. As a result, many carriers manually process new broker submissions, and today those submissions are coming in faster than sometimes what they can manage. By some estimates, only 25% of broker submissions ever become written policies and as much as 60% of all submissions they never even get reviewed. So Tom, what are the challenges this creates for carriers in the marketplace, and in your opinion, what else might keep them up at night? Yeah,
Tom Wilde: It’s interesting. I think as you pointed out in your intro, we’ve been in a hard market for several years now, and in a hard market, the carriers were able to make their numbers without having to worry too much about how well they were triaging and getting at that 75, 80% of the submissions that they weren’t able to look at. There’s some signs that we’re in a transitioning market Now, I don’t think it’s a soft market yet, but what that means is that as it becomes more of a buyer’s market, the carriers are going to have to work harder to make sure that they’re being efficient in how they select that risk. They’re going to need to look at more of those submissions to achieve the same gross written premium targets as it’s going to be more of a competitive market. Obviously in the insurance space, they can’t just go out and add 25% more underwriting capacity. It doesn’t work that way. Underwriting is a specialized skill art, and so now you need to be able to focus on making the resources you have more efficient, more effective in selecting the right risk.
Laura Drabik: Yeah, that makes a lot of sense. And you’re spot on. You just can’t throw people at this problem. And you mentioned why there’s a lot of tribal knowledge associated to being a commercial lines underwriter. I actually read recently that the average age of a commercial lines underwriter is in fact 54 years old in the us. So what can we do then to utilize that mature underwriter in order to improve efficiency? What are some of the things that you think carriers can do?
Tom Wilde: Yeah, I think it’s going to continue to be a relationship business for a long time to come. That’s an important component of the way business gets done. So the art here is making the underwriter more efficient by creating submissions that are ready for them to assess much more rapidly, eliminate a lot of the data entry that currently gets done today. I mean, I think what a lot of people don’t realize is a lot of these submissions start as an email. An email comes in from a broker either to a centralized carrier inbox or perhaps to an underwriter directly. There’s a lot of PDFs and Excel files attached to that. The body of the email probably has important information in it as well. So the key is how do you start with that unstructured payload and turn it into structured data that is complete? IE has the broker submitted all the information needed to properly assess the risk and accurate.
And in addition to that, from when I talk to customers, they’ll tell me that only about 40 or 50% of the data they need to assess that risk comes from the insured or the broker. The other 50% has to come from external third party data sources or internal first party data sources information they already know about this particular risk. So really the task is to combine that submitted data with external data and normalize it in a way that allows the underwriter to be very efficient in assessing that risk. Because as we are in a transitioning market, the speed of response becomes absolutely critical in helping them win that business.
Laura Drabik: Yeah, it makes sense and it now explains why we read that underwriters touch anywhere from five to 15 systems just to perform their daily duties. So thanks for sharing. When integrated with a modern insurance platform like Guidewire’s, for example, AI automation solutions like those from Indico data, they help to streamline underwriting processes to really help carriers quickly and accurately price risk and achieve premium growth of up to 58%. But it’s not just about making technology investments to modernize underwriting. As McKinsey recently pointed out, top performers formulate clear strategies for leveraging them to optimal effect. Tom, you’ve said this requires looking at decision-making holistically across the entire decisioning supply chain. What does that mean and why is it so critical?
Tom Wilde: Yeah, this is something we’ve begun to articulate to customers is this notion that key decisions in the company actually have supply chain like attributes to them. If we think about the physical world like manufacturing an iPhone, the supply chain is absolutely critical, right? You need to understand the provenance of all the parts that go into it, the labor that’s being used, is it being done legally and to the company’s policies, et cetera. So in a physical supply chain world, there’s a well-established understanding of being able to monitor all aspects, all steps of that manufacturing process. We think that decisioning has very similar characteristics. If we think about underwriting as ultimately a decision, right? You’re going to decide if I should underwrite this risk and how to price this risk increasingly. Now, what you need to understand is work backwards to where did that data come from originally?
Can I trace it all the way through the system, including what did I do to this data as it progressed through my system to get it to a point where it was decision ready, right? So if you think about the origins of the data, it’s probably coming in as I mentioned to an inbox. Then you have various AI and heuristics applied to that data to go from purely unstructured to high fidelity structured data. You now with gen ai, have fascinating opportunities to add context to that data from existing information you have at the company such as underwriting guidelines, underwriting authority, policy documents, et cetera. All of that needs to be trapped and auditable so that as you reach the end of these decisions at any time, you can always trace backwards and improve the system and understand how you’re making these decisions.
Tom Wilde: So generative ai, for instance, is of course gaining traction with its ability to review incoming applicant input, accord, forms, photos, spreadsheets, and so on against underwriting guidelines. It can then summarize that information and act as a copilot for underwriters, but for that to work, underwriters must trust the accuracy of gen AI output, and this could be a tall order. Given the complex criteria involved with risk decisioning in the sector. Tom, what strategies must be in place to build that trust?
Tom Wilde: Yeah, it’s a great question. The promise of gen AI is really exciting because of its very deep interpretive capabilities. Things like, as you point out, summarization, understanding context from unstructured data like policy guidelines is a good example. These are great opportunities for applying things like gen ai. However, gen AI is not really designed for deterministic type outcomes where there is one answer and that answer has to be right. So when you deploy ai, you have to think about various strategies that you use or various flavors of AI to get to the result you’re thinking about. A lot of times people don’t realize there are more than one flavor of ai, so generative is just one flavor. Think about these as different programming languages as a good metaphor here. So things like discriminative AI is very useful where you have highly deterministic outputs. I need to extract a particular date or dollar amount or loss from a document, and there’s really only one answer to that.
Combining that with generative AI means I can then use that to do things like summarize and make recommendations as to risk or score risk and prioritize those submissions. So your ability to combine those things together is very powerful. One thing though that I see insurers struggle with when beginning AI projects is being able to establish what’s known as a ground truth. It’s really important to know ahead of time, what does ground truth look like when you’re handling these submissions? Meaning, what is the data that you’re going to use to improve the process? And I would say that in the majority of cases when we show up, the insurers don’t really have that information at the ready, they have to assemble it. This may seem kind of surprising because you would think that these are processes that have been underway for decades, but I think that we as an industry have assumed that if these are people-driven processes, then of course they’re consistent and accurate, and oftentimes they’ve never measured their existing systems. AI really forces hard thinking about that and the establishment of standards and ground truth to know whether AI is performing properly and effectively or not.
Laura Drabik: So something I’m personally excited about is the ability for AI and automation technologies to also streamline initial risk assessment and pre-clearance checks to ensure that only viable submissions get to the quoting stage. Tom, why should the earliest stages of underwriting be a specific focus for underwriters that are seeking to leverage AI automation?
Tom Wilde: I think most carriers have pretty robust heuristics in place that allow them to triage, prioritize, and select risk. The challenge they face is those systems assume you’re going to feed those systems high quality schematic structured data. So the absence of structured data means those systems aren’t very effective. So if you trace backwards, let’s go back to our supply chain metaphor, trace backwards and figure out, okay, what is the raw material I need to feed these decisioning systems? It turns out that your ability to classify and extract the submitted data is the key to making a lot of the systems you already have simply work better, whether they’re triage algorithms, risk selection algorithms, policy admin systems, et cetera. So that’s one of the keys here is can you do that? Can you assemble enough quality structured data to feed these systems?
Laura Drabik: In recent months, the conversation about generative AI and human in the middle models has given way to a new wave of agen AI solutions. And for our listeners, Egen AI is a type of artificial intelligence that can act autonomously, make decisions, and take actions to achieve specific goals with as much or as little human supervision is desired. So I can see AG agentic AI as a compelling option for the two early stage underwriting processes. We just talked about triage and of course pre-clearance. Tom, what are some especially dangerous pitfalls to avoid when thinking about leveraging ag agentic AI in commercial lines? Underwriting?
Tom Wilde: Yeah, this is a great question. Kind of harking back a little bit to what I mentioned earlier. Most of these processes are ultimately deterministic, right? There is a set of data that has to be accurate for the underwriter to do their job effectively. This is where you hear the term human in the loop come in, and human in the loop is very similar to a lot of the ways that AI’s been described. Microsoft calls it a copilot. We’ve often referred to it as a bionic arm, but what both of those have in common is the notion that it’s human at the center that these decisions have to be supervised. Is there a point in the future where more and more of these can be straight through processed? Probably, but right now we have to walk before we run, which means that all of these systems have to be designed with human the loop as a core part of the decision system.
We need to make sure that at every step of the process, we can intervene and supervise the decisions that a agentic AI may be making. What you’re hearing this described as often is next best action. What this really means is the agentic AI is surfacing its recommendation for the next best action, which can dramatically accelerate a lot of these decisions, but at its core still assumes there’s human supervision. As we collect more and more data with these systems in place in terms of their accuracy and efficacy, we can begin to get more confidence in automating parts of the process. Oftentimes, I’ll describe this as lowercase a automation as opposed to capital a automation meaning that parts of the process can be configured to be straight through, but overall, the process still has a heavy human in the loop supervision to it.
Laura Drabik: Tom, if you look at Gartner’s hype cycle for artificial intelligence, generative AI is about to drop out of the peak of inflated expectations and into the trough of disillusionment. A agentic AI isn’t even on the cycle yet. Tom, how should carriers cut through the AI hype and identify what matters to their bottom line results?
Tom Wilde: The concern I have is that customers, in our case, carriers, MGAs, et cetera, are a little bit too prescriptive in requiring that the solution they want has to be generative AI or whatever else it may be, and I think that’s a bit dangerous. It’s much more effective to define a desired future outcome that you’re trying to drive towards. For example, I want to reduce my turnaround time on submissions by 40%. I want to increase the number of submissions that we’re able to evaluate from 25% to 35%. These are business outcomes and can be linked directly to gross written premiums and combined ratios and things that then I think it’s important to work backwards and figure out, okay, what is preventing us from achieving that outcome? And what technology tool or product should we apply to solve those various bottlenecks or pinch points in my existing process?
I think if you do that, you’re likely to avoid a lot of the frustration where you more generically just try to apply a technology to all of the steps in the process and then find out that either in the case of Gen ai, it may be a very expensive tool to apply to the problem. It may be a tool that can’t give you the deterministic reliable outcomes that you need it to. So I think that there’s a little bit of a rush to simply apply this technology blindly across the problem, and I think ultimately that leads to disappointment because obviously that’s not the right way to attack a business process.
Laura Drabik: Yeah, I couldn’t agree more. I would always recommend starting with the business opportunity or challenge, and then as you always say, work backwards or try to understand then what the best fit for achieving that is. So looking at the year ahead, what do you think the most surprising development in AI relative to commercial lines underwriting will be?
Tom Wilde: I certainly think that a agentic AI is real, and it’s going to be very interesting. I think it’s very early days. The promise of it is well ahead of what it can actually do, but that doesn’t mean you shouldn’t start working on it today. The way I think about age agentic is maybe slightly different, which is age agentic is really an orchestration layer that if used effectively can coordinate a number of different technologies and capabilities to achieve and drive this business outcome. That’s a very powerful concept and I think has a ton of promise in it. But again, going back to the sort of human supervision, human in the loop trust factor, I think that’s going to be really important as we begin to deploy these kinds of technologies. Unquestionably, claims and underwriting are very ripe for business transformation, resulting from the proper implementation of these technologies. But understanding the business outcome, number one, and then as I mentioned earlier, understanding your ground truth that exists today is vital so that you can measure the impact. Is it working, is it not working? And what do I need to do to tune or fix it as opposed to hoping that it’s going to take effect or not being able to measure whether it’s had the desired impact.
Laura Drabik: Tom, thank you so much for your time today and for your incredible insights. You’ve shown us p and c. Innovation isn’t just about ideas. It’s about making ideas happen.
Tom Wilde: Great to be here, Laura. Thank you.
Thank you for joining us for this episode of Unstructured Unlocked. You can find all of our episodes wherever you listen to podcasts today, Spotify, apple Podcasts, anywhere. Be sure to write a review if you like what you hear.
