Hello, and welcome to today's webinar, The Future of Underwriting in the Small Business Market. Our presenter today is Jesse Lowe, Principal Product Manager for Guidewire Analytics and Data Services. All lines have been muted. However, let's take a moment to familiarize ourselves with the webinar console. On the webinar console, you will find an area where you can submit questions at any time, including any technical issues, and we will address those at the conclusion or as appropriate during the webinar. You'll also find an area where additional resources have been made available, including information on how to contact us if you have even deeper questions and details about Jesse, and with that, I will turn it over to Jesse to get started.
Thanks so much for that. And hey, everyone, thanks so much for taking the time today. My name is Jesse Lowe, and I'm the Product Manager for Cyence for Small Business, one of the products within our Analytics and Data Services business unit at Guidewire. So what I'd love to do today is to share a little bit about our views on the small business market, what we see as some of the opportunities and challenges, and why we decided to further explore this space and provide more value to our customers by creating this analytics tool to basically differentiate small business risks, risks that we feel might look similar on paper. So I'll talk through a little bit of that and the data behind it all, and then talk through some of the validation and the proving of value that we've gone through with our partners and our customers.
At the end, we'll open it up for Q&A. In the meantime, I think if you look on your console, there's a Q&A section tab. Feel free to punch in any questions at any point during the presentation. At the end, we'll go through those and talk through any of the questions that the group has. I wanted to start off with getting the lay of the land here and talk through what we see as some of the market opportunities and challenges. In thinking through the small business market, I think it's immediately obvious that to a lot of carriers and players in the space, it's a definite opportunity. It's a growing opportunity. More businesses are being started every year. I think a non-trivial amount of employees actually work at small businesses.
But as you think about the changing evolution of these small businesses, the business owners are behaving more and more like consumers as we think about all these other technologies and these other services that exist to us today as consumers. And so when you think about and you look at some of these polls that are out there about what these small business owners are looking for, a lot of these small business owners aren't that familiar with insurance. And so they definitely need some guidance and some hand-holding in that process. But at the same time, they're also asking for a very streamlined, easy-to-use, self-service kind of approach when they are buying insurance. So they're looking to make that process as easy as possible and as little bit of a headache as possible. And so I think that really leads to growing opportunities in the small business space.
I think metrics and figures will vary, but I think by some estimates, there's over $100 billion of EWP in the U.S. written for small businesses across the lines, and so I think that's something that we saw as really, really exciting and an opportunity to explore, but at the same time, when we talk to our partner carriers and our customer base, we quickly realized that across the board, the way that small business insurance is being underwritten right now is just not efficient, especially from a cost-efficiency perspective when you think about how little small businesses pay in terms of premium and the amount of time and effort it takes to actually go and do the due diligence and process it, etc.
Not to mention the fact that these business owners are looking for a faster turnaround, low friction, and then compounded by the fact that if you actually wanted to automate this process or structure some kind of approach to streamline this overall underwriting process, it really comes down to having the underlying capabilities to process this quickly en masse and kind of this volume or high-throughput approach. But what you really need is actually the data underneath to be able to make faster decisions and to empower your underwriters to either triage risks that might require more attention and time, and then to also surface the ones that don't require any further attention and time that can get passed through. So when we look at the market, we saw this opportunity.
We wanted to be able to bring value by building some kind of tool that can really quickly determine and assess the risk of a small business and then either triage it and serve it to an underwriter so that they can then further explore or enable carriers to take a more low-touch or no-touch underwriting approach, and so I think that's what really takes us to this maybe not so distant view of how this low-touch underwriting world could look like. Because on one hand, you can automatically quote and bind a lot of risks that are more straightforward, and at the same time, then triage and review other risks that essentially get floated up to the top using risk assessment tools that can make these decisions very, very quickly and then overlaid on top of that with business rules and decisions within your policy framework.
And so when we think about really kind of the value that we want to bring to the market, it is how do we help accelerate the adoption of technology and data across the P&C industry, insurance industry? And this is kind of where we're starting off with small business as where we feel like a really interesting opportunity to do that. And so if we can then double-click into one of those accounts that you want to triage, then this is where we can then augment the ability for underwriters to really dig a little bit deeper and to be more focused in their exploration.
If you were to have some of these different risk factors surface to say, "Hey, maybe we should take a look at these two things out of 10 or 20 different potential areas," we want to be able to support the underwriters in maximizing the efficiency of their time, again, through the use of data and analytics. So what that looks like, kind of one layer behind the scenes, if you imagine looking at the far left-hand side of the screen there, a business owner might be using an iPad to buy insurance, and when they do, they're submitting their information there, and what gets ported over to the Science Data Listening Engine or the Science Risk Assessment Tool, at a minimum, is really just the name and address.
And then we can take that information and run a lot of these different data collection and these sophisticated risk assessments behind the scenes, serve that information back to the carrier in the form of a risk rating and a set of risk factors that I'll go into later, and have that flow through the underwriting guidelines to then ultimately to the eyes of that policyholder, that person applying for insurance. They can see a really streamlined end-to-end process that really only takes minutes.
Now, when we think about these kind of future state views of what low-touch or no-touch underwriting could look like, especially when enabled with data, I think it's not so controversial that there is a place for where data can help augment and streamline a lot of opportunities and avoid having underwriters really peeing up on their time across all these different accounts where only a small subset really deserves their time and attention and focus, and so I think the idea of having data be an integral part of that underwriting process really across the full insurance value chain. I don't think it's controversial, but I think it's also non-trivial to actually take that idea of integrating and implementing data and have it come to life.
And so that's what I want to call out here and talk through this kind of thought process and these steps that we've seen in our conversations with carriers that we've had to go through. And so first, there's this idea of capturing data, all this data that's out there. And when we think about all this data that's out there through a lot of these different platforms and technologies, there's a lot of digital exhaust, so to speak, that's being created and generated on a daily basis. And so as all this data is created out there, I think a lot of carriers recognize that while the data is there, it's actually hard to systematize it, capture it, distill the value of that data, and have it flow through to the underwriting process.
And so the first step of that is really to have some sort of tool to go out and to know where to look and to collect that data in some kind of automated, productionized way so that you can then even set up the processes to distill the value and extract those insights from that data. And so you might want to take an AI or a machine learning approach, apply natural language processing to look at the words that are being said, look at historical analyses and look at how things are trending, look for outliers, etc., and to really figure out what are the insights that you can glean from any given set of data, and then to productionize that and turn that into a pipeline so you can collect similar types of data on a continuous basis.
And then comes the insurance tooling where you want to take that data and you want to bring it into these different tools that actually drive decision-making in these different use cases. And so up to this point, we've been mainly talking about underwriting, but it's really one component of that overall value chain. And so whether it's underwriting, pricing, prospecting, or claims, we want to be able to have insurance tools that take all this data, distill the insights, surface that data in some kind of readable or ingestible format, whether it's human or machine as the audience, and then ultimately have that flow into the workflow.
And so you have all this data, you have the tooling behind it, and then ultimately, I think where rubber really meets the road and where there's truly value that can be appreciated is if you're looking at the screen that you normally would as an underwriter, and now you have more information or now you're moved to different workflows because of information that's gathered by these tools, I think that's where the value creation really happens. And that's where you really can start seeing either higher throughput because the different folks along that value chain are more efficient, they can focus their time better, and I think that's where you really see the cost savings impact the bottom line. And so this is where I think I'd like to call out the Guidewire.
I think it really brings that full-stack solution from the very beginning of the data ingestion phase through to the data synthesis, tooling, and ultimately the workflow. So what I'd like to do right now is really quickly open it up to the audience here and to the group and learn a little bit more about what some of your roadblocks are when it comes to taking in third-party data to streamline underwriting and probably other processes. I have only a couple of buckets here just to see if we can get some interesting insights, whether it's more internally driven or whether some of these roadblocks are more externally driven. So we'll take a minute or two for you guys to punch in your response here. All right. I'll give them another 10 seconds. All right. Yeah. So I think this is really interesting.
Looking at where existing solutions are not meeting needs as being the biggest one, followed by a close second around internal prioritization. And I think that echoes some of the conversations that we've had with our partners where either it's hard to really understand and dig underneath a surface level of explanation of what the capabilities are, or it's hard to see kind of how these different tools provide business value without either going through some kind of more involved pilot process or seeing kind of how this data, while it might seem really interesting or valuable to other carriers, ultimately, I think some of the challenges that we've heard are, "How does this actually impact me? And how is it going to be relevant for me?" So I think that's interesting to see here. And then for the others, we don't have an option here.
I can have you type in the answers here, but I think later we'll share some contact information to our team, and we'd really appreciate it, be interested in for the folks who said other to share more about their views. So hopefully, that was helpful as a high-level overview of kind of the why and a little bit of the what around what we're doing with Science. So I'll dive into this next layer of detail around what is the data, what are some of the approaches that we take, and some of the thinking behind the scenes. So it really comes down to this underlying approach of what we call Data Listening.
And it's this approach to collect data at what we call internet scale, where we want to look at data that's somewhat obvious to collect a lot of this information that comes in on applications, but also data that is less obvious. And a lot of this is also a legacy of the work that we did within cyber insurance when we first started out a few years ago and now expanding to all lines of business. We're finding that a lot of the characteristics around looking at a company's digital footprint actually translates well to us understanding a little bit more behind the scenes around what the behavior looks like, what the management looks like, etc.
We then curate that data, synthesize it with our AI machine learning techniques, and then that's where we can then distill those insights into something that's more usable for these different use cases, whether it's underwriting, pricing, or marketing, and I think the actual output of this data also differs and is modulated based on whether it's a human or a machine audience on the other side and what kind of needs they have in terms of doing their job.
And so, as we think about these different use cases, we've talked a little bit about underwriting, pricing to an extent, but I think what's also really interesting is around this marketing and prospecting use case, where I think the ability to actually make decisions and learn more about a business before they even come to your door is this ability to collect data from the internet and from the world that's out there on demand and get it back in very few seconds using very little pieces of information, just like a business's name and address.
And so with that, it unlocks and opens a lot of doors for marketing, for example, where now you can start to prioritize how you go about that process of reaching out to customers, and you can really fine-tune that segmentation to really target the set of audiences that you want to go to. And that can really help cut down and focus your acquisition spend, as an example. So I also have another quick poll to learn a little bit more about for each of you. I'd love to learn a little bit more about how are you guys thinking about using data and analytics. And of these, I'm sure you're focusing on many of these, but if there is a biggest focus for data and analytics across these different use cases, really appreciate your insight here. All right. I'll give another 10 seconds. All right.
I think it's just really cool to see everyone's feedback here. I know that this webinar was set up as more of an underwriting-focused underwriting webinar. I think it makes sense that the bulk of the responses are towards streamlining underwriting processes. Also interesting to see that that actuarial, I think, is also very tied into the underwriting process, but also claims and marketing as less of a priority. It could be just based on the audience that we have here, but very helpful. Thank you. I'll dive in now. It's kind of this peeling back one more layer and going behind the scenes in terms of how we think about our data collection. We'll start off with a little bit of the approach.
I'll go into some of the examples of the data that we collect, and then we'll go into looking at the validation of that afterward, so when we think about the data that we want to collect, we definitely combine all the different resources that we have in-house as well as leverage the expertise that our partners have, and so it's very much a collaborative process where we're actually going out and building out these sets of data and our products and our tools, and starting off, really, as you see on the left-hand side there with these data-driven hypotheses, what I'm showing here is just an example of some of the causes of injury that you can get from OSHA.
And when we looked at it, we thought that, "Hey, maybe this might be a good place to start to think about how we're measuring workers' comp risk, for example, and what types of data we should really collect to try to quantify that and build that into a risk factor." And so in working with our partner carriers and coming up with what some of the types of causes are more top of mind, we took one that's related. We took a couple that are related to motor vehicles, machinery, and boats.
And how we then turn that into insight is we go through all these different websites, and for businesses with websites, we'll basically crawl through every single one of their pages, look at all the words, click on the links, look at the pictures, and distill that into, as an example, what you see here in that word cloud. And we're going to be doing that for millions and millions of different companies. We then go back and assess for the companies that turn out to be related or work with boats and motor vehicles what are the words that show up more on their websites related to more of their services, related to their customers' feedback, etc.
That bar chart right there in the top right-hand corner was a kind of completely algorithmically driven chart where words like boat, freight, marine actually surface up to the top and automatically telling us, "Hey, for these types of businesses that work with boats and motor vehicles, the words, when they show up on the website such as these, actually indicate that these businesses work with boats and motor vehicles." And now this is one example, again. But if you imagine you can take this kind of one-sliver approach and attach that to every single one of those causes of injury on the left-hand side bar chart, in addition to really any question that you might have when it comes to, "Does a company do delivery?" And if so, is it fulfilled through a Grubhub or their actual employees that go and do the delivery, among many other questions.
And so that, I think, is one of the pillars of how our technology works to go out and try to prove or disprove hypotheses and then turn that into insight. And so one of those applications could be things like industry classification, where we feel like there's a lot of attention on this, both from the application intake process where the industry is either not filled out correctly, is not comprehensive enough. And so if there's a way that we can actually really pick out what those different class codes are, what those industries are, we feel like that would be a really interesting way to bring value. And so if you take a company like this example, it's privacy, and you were to go to some company database and look it up, it might provide pretty high-level, SIC 2-level categories.
If you were to go to the website, then as kind of another click down, you might see boats on that front page. So it kind of looks a little Mediterranean, pretty fancy, but I think it has to do with boats or cruises. But then at the very low level, if you were to read every single word on their website, as the robot there is, you actually start to identify words like boating, marine, and support. And through not just this website and your knowledge of websites and companies, but across a blend of different companies, we can then start to build models that really help refine these classifications down to the NAICS 6 level, for example, around other support activities for water transportation. And so it turns out that this company does kind of support and maintenance for boats.
I think this is an example where you can then enumerate multiple ones so that an underwriter can still leverage their expertise to say, "Hey, this NAICS was that one," or, "I didn't realize that this company provided these services, so I'll also need to tack on this other industry code or this other class code on top of that." That's one example of how these data points can be applied to different use cases. Broadly, we want to expand that to kind of all things data, both data that at the surface level is more straightforward, but also expanding to things that are more non-obvious. So this is where then some of the other data points come into play, like web sophistication or consumer sentiment, that I'll go into a little bit more later, where it's really interesting to think about.
And that's where our minds really go to when we talk about, "Oh, there's so much digital exhaust," or, "There's so much data that's out there." Now, if you can only go and enumerate through all of a company's Yelp reviews, that would be a great way to be able to get a sense of how people feel about the company, to, again, learn a little bit more about the management, etc. Now, it's not enough to have this data that's collected from the outside in. We also want to be able to look at data that really tells a story around whether or not these data points have relevance. So that's where the proprietary data comes in, where we're partnering with third-party data providers who are our carriers to look at the historical data.
It's this having claims, I think, that makes this whole process that much more credible, especially when you layer on then the historical data, which is not only do we have all this data at some point in time. If you are able to pull this data now back historically, you can then start to do backtesting and validation, especially against that proprietary data like claims to say, "Hey, for these certain types of data points, if you now had them 10 years ago through today, how would you have priced these policies differently?" Or, "How would you have made different decisions?" I think that's what really enables our ability to bring value. As a double-click, like I mentioned, some data might be more obvious or more top of mind, but a lot of this data is actually really hard to parameterize.
What I mean by that is it's easy for each of us to say, and sometimes not even, to look at a Yelp rating or a Yelp review and say, "Okay, this person is pretty negative," and we can pick out some of those challenges. But then for a machine, especially when you have double negatives or the Yelp rating saying that this food tastes really bad, but at least service is great, or the opposite, I think a lot of those different components of a review, for example, can actually throw off more traditional machine learning approaches or automated approaches to looking at this.
That's where we have then taken on some of these different approaches to categorize the data, think about what some of these high-level keywords are, apply a lot of the different open-source type approaches to synthesizing and distilling this data to then try to look at how we can featurize some of these different word phrases and overall sentiments and turn that into some quantifiable measure that can use the model. On the flip side, there are other types of data that we think is really interesting, but still requires a lot of validation. One example is that when you look at websites, you can look at the environmental technologies that a website uses and also whether or not they're adhering to not just the best web standards, but also ADA standards around disability.
And so as someone who's a little colorblind myself, it's actually really meaningful for websites to take on some of these responsibilities to adjust their different colors, for example. And so we have a hypothesis around whether or not websites that adhere to best practices, both from a technical and from an ADA perspective, are more reputable or more credible or better risk. And so it's through this process of coming up with these hypothesis-driven approaches, distilling the data, having some kind of technology to pull down that information and then turn it into some kind of usable unit of measure that's been really interesting. And that's where I think just this volume of different types of data and the ability to freely source some of the different types of data that I think is really one of our strengths.
And so this last section, which I think is where the crux of it all is, how do I know that you guys talk about this data? How do I know that it really works? And it's through this idea of historical backtesting, ultimately, to determine the kind of efficacy that we're talking about. So on one hand, you can take an approach where you can do backtesting, whether it's 10, 15, 20 years of data to either build or improve your models, or actually look at the underlying risk factors like ADA compliance to determine whether or not those types of data actually have relevance. And then if they do have relevance, how much relevance and how much lift.
And so in a nutshell, what this is really doing is saying if you had a time machine to go back 10 years in the past, now armed with the knowledge that these types of data have lift, you can then append all the historical data to those historical policies, monitor that performance year by year, and to look at where along that curve you're doing better or you're doing worse, and then how you can improve and how you can use these data points to, even if you have an in-house model today, to augment how you're thinking about evaluating and assessing risk. Again, for an underwriting approach, I think this is pretty tactful in terms of how you might make different decisions. But I think that also extends to all kinds of other different use cases, whether it's prospecting all the way to claims.
And so this is a process that we take whenever we're coming up with new data points, for example, and working with carriers. Or for new carriers that are interested in our product, we oftentimes take this approach where we'll go and backtest their data and basically take a test drive or take the carriers on for a test drive to say, "If you had this data now 10, 15 years back, how would that have looked? How would you have done things differently?" You can pull in your model as well, and we can compare what it would look like if you had your insight, if you had your application data to make your decisions versus our totally outside-in approach, just given, again, name and address. So I'll call out a couple of things here, which I thought were pretty cool.
But basically, this is one way to look at backtesting and validation and efficacy. And so I've just called out a couple of these different risk factors that we have within our product. And what you'll see here, and I'll just choose one of the couple on the bottom, the gray lines that you see are each individual year that we were doing in the backtesting, whereas that darker line is the average of all the performance. And so if you look at some of these charts, the bottom one on the far left-hand side is the number of nearby medical services.
On the X-axis, if you go further to the right, you'll see that as the number of medical facilities goes up around your area, we actually see that the likelihood of an event actually either goes down compared to the commercial vehicle size or some of these other prior OSHA violations where, as you look to the right, going from left to right, the larger the commercial fleet size or the number of previous OSHA violations actually increases risk. You can also look at this from a loss ratio perspective. Here, I'll turn your attention to the bottom right-hand corner around Web Sophistication.
We've seen that if you look at the websites of different businesses, that if you look at the technologies and generally try to get a sense of what the hygiene looks like on these businesses, we saw that, interestingly enough, there was negative correlation around the more sophisticated website, the lower the loss, and so this is something that we thought was interesting and needed to prove and test out, and so we started looking at other metrics related to web sophistication to try to tease out what it was related to. And ultimately, I think we just came to the hypothesis that it might just have to do with the hygiene and internal practices, whether or not a business runs a tighter ship, and so those are some of the interesting things that we've seen coming through storytelling of the data.
Ultimately, when we think about now the hundreds of different data points that we have, we want to expand this approach to have as robust and as comprehensive of a way to assess these businesses as possible. So in addition to just looking at lift and the directionality of whether or not a data point makes sense, we also look at coverage. We want to make sure that the data point is relevant to as many businesses as possible, but also correlation so that we're not double-counting and having a lot of redundancy around these different types of data points. Not that redundancy is bad either, because we also then want to have a way to triangulate data points to make sure that there's multiple data sources that all tell the same story.
And so, again, happy to discuss more offline if you're interested, but there is a lot of work that goes on behind the scenes to make sure that the data is both valuable and has lift, but also is as robust as possible. So ultimately, what we can do is actually try to provide value to carriers and then cover a lot of this business value, which is if you had now this outside-in data to then append to your own internal models, especially if you have an internal model, you're really only looking for a couple of other data points that add incremental marginal lift. And so from tests that we've seen with carriers, they've seen 2%-3% improvements on loss ratio, especially when compared to their in-house models.
And so that's something that we've seen where there is value that's shared across different customers and something that we're really excited about, especially in sharing with you. And that is, again, in addition to these other opportunities like operational efficiencies, looking at making a lot of these different components around employee training, model development, and pricing compliance, making these processes easier when you have a more systematic approach and a more auditable approach to how you're underwriting. And so with that, I know it was a lot of material, and I think we'll be able to share a recording of this later on. But with that, I wanted to turn it over to some of the questions that you guys might have and have a bit of a discussion around some of the topics that are more interesting to you.
Thanks, Jesse.
So I'm sorry. Yeah. So I think there's a couple of questions that have come in, and I can read them out. One of them was, at what point, based on the size of the risk, does the automated risk selection and pricing begin to lose credibility? So thank you for that question. So I'd say that when we think about, and this is kind of like, how do we define really small business? I think we take the high-level approach of at the point where the economics really don't make sense for you to spend a lot of time underwriting that account. I think that's kind of how we would define small business. Internally, in terms of how we've been building the model, we've been more so looking at single-location businesses. And that, I think, in terms of employee size or revenue, somewhat does vary.
But I think that's the kind of the main street small business, single locations where we've seen our model have more relevance. So there's a question around how many data points are required to produce a quote or policy. And I'll interpret that as, so I think there's a couple of things here. In terms of how many data points are required to produce a quote or policy, I think that's ultimately up to the carrier in terms of what their business rules are and where their appetite is. You can make it as in-depth or as not in-depth as possible, depending on the appetite. But then from our end, out of the tool, we have 101 different data points that we call risk factors, along with a risk rating that really synthesizes the learnings and the signal from each of those different risk factors.
Hopefully, it answered that question. If I didn't, feel free to follow up again. Then there's a question to go back to the previous slide, so I'm happy to do that. There's another question about how are state insurance bureaus accepting the outside-in rating factors as valid determinants for rate? I think that's a great question, especially from more regulated lines like workers' comp for workers' compensation. Candidly, this is something that we are exploring with our partner carriers as we go along this process of building out our product. For the time being, how we've approached it is that a lot of these carriers are using these data points at pricing that's coming out of it as just price guidance.
A lot of it's going into the scheduled debit and credits to help underwriters augment how they're thinking about among the different categories as part of their filed debit and credits, how to come to a better decision around what their assessment is for each of those categories. There's a question about what sort of explanations are provided with the risk indicators. So there's a few levels of explanation. We have a model documentation that basically goes into a lot of detail around how the data is collected, what it means, etc. But then generally, there's also, not to get too in the weeds here, but the API documentation has kind of an explanation of what the risk factor is, what the different risk factor buckets and values signify.
So there's different levels of documentation, as well as kind of being able to pull in data scientists and our product team as needed to help better explain what's happening behind the scenes. There's another question about how accurate are models for risks that have no web presence. And I think that's a great question. Thank you for bringing that up. So what we've seen is that using kind of our approach, there's certainly, in terms of companies with a website as a starting point, there's certainly companies without a website. And we've seen that oftentimes, whether it's contractors, etc., they'll have a Facebook page or a Google page or a Yelp page, etc., or a more kind of industry-focused page. And so with that, we recognize that not all companies have a website.
In those cases, the website risk factors aren't going to be tailored exactly for that company. It's that idiosyncratic risk of that company. We do have models and fairly sophisticated imputation models to try to predict, well, if this company had a website, what roughly would that website's kind of risk look like? Given that, if a company doesn't have a website, they oftentimes will still have a web presence. I think if you look at even companies who are fairly new, they'll often have some kind of profile on Google. There's a Google Places that talks a little bit about kind of what their open hours are or what people have said about them or their customer review rating, again, on Yelp, etc. We have found that a lot of companies have some kind of web presence.
Now, even if a company didn't have a web presence, a lot of the other risk factors that we do have for a product are related to very kind of physical-based things. And so as long as you provide a name and address, for example, we'll actually be able to look at certain risk factors like how far away is the closest fire department or the police station, etc. And that was something that we have also found, kind of going back to one of those slides earlier around risk, to show that that actually does correlate with risk. And that's something you don't need a web presence for either. And so we try to take a balanced approach and try to have as wide coverage as possible. There's a question, I think, related to this page about a line here that says we have petabytes of external data.
And the question is, is it already built into Science? So I'll say that we have petabytes and petabytes and petabytes of data because we've been collecting this for years now. And so a lot of that data is in Science across our really wide kind of data infrastructure and our databases. And then the process of collecting that petabyte of data on a monthly basis or a periodic basis is just part of our overall data collection process. And when you think about all the different companies and millions of different companies that are out there and the digital exhaust that they're creating among other types of basically sensors or data collectors on the web, it adds up very quickly to just a ton of data that we're able to sort through. So if I'm not answering that question correctly, please follow up.
There's a question around how much transparency there is to the underlying data. So I think that this depends. On one hand, we don't provide the actual vendors or the data sources that we use. And so there's not so much transparency there. In some cases, that's like a kind of data licensing issue. In other cases, we do feel like in our exploration of finding really high-quality data sources, that it is kind of part of our IP to protect those sources. But then beyond that, if it's a question of what was the process to get that data or what are the, I guess, the nuances and the variances around that data, that's something that we can definitely share. Depending on the data source standpoint, the sets of metrics historically, etc.
And so that's something that we've done often for carriers to show them more of what's happening behind the scenes.
This question around how do you weight the different data points or risk factors to get to your target price for a given risk? And does that vary by geography, industry, etc.?
So I'll take this in kind of two parts. So the first part is, how do you weight the different data points and risk factors to get to your target price? So two things here. One is the output of our product isn't actually a price. It's a risk rating, and it's an indication of price. But in terms of how we weight these different factors, we actually run through a ton of these different machine learning models, and then happy to pull in our team to go into more detail here.
So a lot of that is kind of first produced using our machine learning models. And then we actually go and take kind of a fine-toothed comb and magnifying glass to that to look at what actually makes sense from an intuitive perspective. And if something doesn't make intuitive sense but has a lot of signal, we'll actually take it back through and try to do more analysis on that and also surface that to our partner carriers to try to figure out what happens. And so we're not manually adjusting the weights for different data points. That oftentimes comes out of the modeling where, yeah, we can go into more details there if you're interested. And then the other question is, does it vary by geography, industry, etc.? And it definitely does.
Geography and industry are definitely some risk factors and considerations when it comes to thinking about the model and the model output. So we do also break out companies differentiated by geography and industry. So I think we're close on time. And so there's a couple of questions around, does it only work with Guidewire systems? The answer is no. It's actually underlying system agnostic. And then kind of how often we review the data sources around the modeling. We are looking at, and we review the kind of stability of this data on a monthly basis. And then we also do kind of periodic model refreshes just based on underlying behavior of the data as that's needed. And so thank you so much for your time. If you have additional questions, I've also included Brett Schneider's email here. And he's really the lead of all this client engagement.
And he can also put you in touch with our strategic advisory teams as well as product teams if you're interested in learning more. And so with that, thank you so much for spending the time with us today. And like I said, I'll turn it back over to you.
Thank you, Jesse, so much. And thank you all for joining us.