Getting Data Ready for Work

Storing the data in a usable database. Right now, most brokers and carriers store all of their data on their corporate shared drives, spreadsheets, and PDFs, and there’s no way to access that in an organized way. So that’s probably the single biggest challenge across all of insurance, not just property. Property is particularly detailed, and getting that data into a database so it’s usable, reportable, and able to be analyzed is very challenging.

The vast majority of small business data comes on PDFs, and the vast majority of large business [data] comes on spreadsheets. The challenge is there is no structure to those spreadsheets. A human can look at them and usually understand what the user was trying to convey, but it’s challenging. One example is an address that says it’s across the street from Tesco. That’s not helpful. Or an $850 million factory and the address is, “about 6 miles south of town,” and it’s in Florida. These are not edge cases.

The fact that all this data is sitting in spreadsheets and PDFs on corporate shared drives, when a broker is looking to put together a better submission means the process is extremely manual. They have to cut and paste, they have to retype, and if you think about doing this for many, many fields across tens of millions or hundreds of millions of buildings around the world, it’s monumental. Some brokers just send off the documents and hope for the best.

The better brokers are focused on advanced techniques to extract that data and get it into a database. Once they have it in a database, they can run it through algorithms to look at the exposure and put it in a structured form, so they can better advise their clients, and so they can provide better data to the carriers. The brokers are in this very unique position where if they have bad data at the start, it’s very expensive to fix it. However, if they do, they start to change the [insurance placement] process from the very beginning.

That’s where Ping comes in. We say, if you have all this unstructured data in PDFs and spreadsheets, we’ll help you put it in a structured format and we’ll validate it. We use independent sources and our proprietary data to help you validate it and get that in better shape to put it in databases, get it into analytical tools, put it on maps, and get it out to the carriers in the way that makes sense.

There’s now a fairly large batch of companies that are using computer vision with aerial photography, either based on high-resolution satellite images or even better images taken with drones and civilian aircraft. Certainly, there’s wide availability of such data across countries including the United States, Australia, Canada, and the U.K. And that imagery is being taken more frequently. It’s much higher resolution. So, using computer vision algorithms, they’re able to determine things like estimated roof age, condition, shape, and there are many new fields of data that are highly desirable for underwriting. This is obviously used by the carriers, and so brokers looking to gain an edge will also look to purchase that data if they can. It’s expensive, however.

After the recent wildfires in California, you could see what the damage to a building was using this type of imagery. It can tell you if that house or building is still there, or if it’s got any damage, minor or significant.

Then it’s really just understanding your exposure at that point. Post-event, these are very valuable tools. And then pre-event there’s lots of use cases, such as showing that your client has put a new roof on their house. How would the underwriter know that? Being able to get these kinds of tools that provide much more recent imagery, and to be able to provide the underwriter and the broker the comfort that indeed the party had replaced the roof on their building or house.

There are estimates in the billions of dollars’ worth of time and resources spent by the industry to cut and paste this data, putting it into submissions and so forth, and often that doesn’t end up in the databases.

The large language models, as one branch of AI, are very powerful, although they need to be used cautiously, they are not a panacea. They make mistakes, and when they do, they don’t tell you they make mistakes. But they are certainly leading the charge with extracting data out of PDF documents, or Word documents, or text-based documents.

Other types of machine-learning technologies are also available. And so, this whole world is blurring. You might have heard the term multimodal in AI; that means it can take any kind of image, it can take sound, it can take text, and it can put all those things together and make sense of it.

So, for instance, you could see an inspection report with pictures, and AI could in theory read the entire document, including the pictures, and put that into a database and summarize it for an underwriter. Where historically that may take an hour or more depending on how long that document is, you can do it in seconds. Just on the vision side, you can take an image of a roof, and you can say, “How old do you think it is? Do you think there’s damage? Give me the shape. How big is that roof? How much is it going to cost to replace it?” They can estimate now extremely quickly, all using AI-based tools. So, on the claims side, it’s profound as well.

We recently rolled out a new visualization platform [Ping.Maps]. And the old adage of the picture is worth a thousand words jumps off the page. When you see the building is in the Florida Keys, you [an underwriter] might say, “I’m not really interested in that risk.” But when you see that there are a few buildings in the Florida Keys, but most of the risk is in a different place, you can be more interested because there’s more exposure in another place.

As humans, we are trained to make decisions with our eyes. We evolved that way, and I think that’s why we find pictures so compelling. And so as long as the pictures are presented appropriately and quickly, they’re invaluable. You use it in terms of claims planning; you use it in terms of measuring impact. For the recent wildfires, it was invaluable. And getting that very accurate data in the hands of professionals has been a [big] change. It wasn’t available 20 years ago, and much of it wasn’t available 10 years ago.

Ping is focused on property in general, but largely commercial property. We read the entire submission, which includes information in the body of the email, in PDFs, and spreadsheets. That includes Acord apps, inspection reports, loss runs, or other documents. But the main focus, especially for commercial property, is statements of value—Excel spreadsheets where the data resides. The data is very unstructured and very difficult to analyze. My example about the addresses is just one. There’s often missing data, it’s often disorganized. A human can understand it most of the time, but a machine can’t.

We ingest statements of value and other PDF documents for submissions. We put it into a database; from there, we can integrate it into clients’ analytical workflows and directly into their models, like Verisk’s and Moody’s catastrophe models, mapping frameworks for spatial analytics. We can introduce things like high-res imagery. All of that can be done in minutes instead of days. One of our clients coined a tagline for us, which we use all the time now, which is that what used to take them 24 to 48 hours now takes them 24 to 48 minutes.

I think that really summarizes it. There’s a lot of detail underneath there, but if you think about how many man-hours it takes to process a single submission or 100 submissions, we reduce that by five times or 10 times at a minimum. In certain operations, there is even more lift than that.

There’s one other component, which is that we store all of the original data. When you cut and paste data, you lose things. They don’t keep it in the original insured’s language, if you ever want to go back and look at it. You’re losing some of the original data, which is very valuable.

We put all that in a database, and it’s there to mine on the fly, but it’s also there for future mining as well. The part of Ping that you don’t see very much of is that it is building a big data infrastructure for the entire property industry. And as evidence of that, just this week, we crossed over processing 875,000 statements of value in the past three years. That’s a large number.

All of the other tools and platforms that have existed, and even the new ones coming out, they’re self-contained environments. Let’s take something like a catastrophe model. You put the data in there, and it just stays there. Then you have to get your data into a policy admin system, and it stays there. And you put your data into a mapping environment, and it stays there. Everybody’s got a different data format. It’s very difficult to keep it in sync. You basically can’t. And this is for companies that process thousands of policies. And, by the way, you’ve got to get all that organized and get it to your reinsurer once a year as well.

Ping takes a fundamentally different approach. It’s a single platform that carriers and brokers can access to manage their data through time.

It’s really important that you understand property in general and store it in a master taxonomy that allows you to speak any [data] language, like a master translation language. All of the other platforms have proprietary formats and aren’t interchangeable. We store it in the master taxonomy so that we can always give it back to them [carriers and brokers] in the correct format if they ever want it.