Leveraging Machine Learning to Transform Insurance Underwriting at VGM

There are many use cases for machine learning (ML) in the insurance industry, from helping underwriters classify risks to identifying fraudulent claims, which results in an overall reduction in losses, improved profits, and more satisfied customers. For this reason, there are increasingly more insurance companies looking to leverage ML in their daily operations.  

VGM Insurance happens to be one of those companies. With more than 30 years of experience providing specialty insurance programs to clients and partners across the country, VGM has built a reputation as an innovative forever partner that provides protection beyond policies. 

In effort to streamline and automate processes, and make more informed decisions around claims applications, VGM engaged the team at SPR to create ML models that would support them in gaining valuable insights based on historical application and claim data. Of course, this wouldn’t be without its challenges. But we were up for it! 

It is estimated that 85% of ML projects fail according to McKinsey. Factors like inadequate data, poor data quality, lack of skilled personnel, unrealistic expectations, and integration challenges all play into the success rate. For this reason, our machine learning mindset is to ‘fast track and fail fast,’ allowing us to learn quickly and pivot as needed so we can work towards a successful outcome. In our original proposal for VGM, we included a technological solution that made sense in theory, but once we learned more details, we realized we needed to evolve our solution. We did this and it pointed us in a new, better direction. 

The speed in which our team got ‘up to speed’ was noticed by VGM. “It was almost like jumping onto a team of people that I'd been working with for years,” said Scott Hagberg, application and automation architect for VGM and key stakeholder on the project. “We were already talking the same language. We were in lockstep the entire way, understanding our environment, requirements, and establishing services in compliance with our security standards. We were able to complete that [initial] work in about two less sprints than intended once we were up and running. It’s so critical to highlight how fast we started.” 

With years of historical application data stored in various different application formats and versions—including images and PDFs of non-digital applications as well as digital applications—the question became: could we leverage VGM’s non-digital historical data to make predictions? 

VGM had admittedly attempted to “read” and extract data from non-digital applications in the past without success. With many different changes to applications over time, it proved difficult to account for inconsistencies. Since then, however, optical character recognition (OCR) technology has been refined. Utilizing the latest technology—Microsoft Azure’s Document Intelligence Studio running on Azure AI for OCR services—we were able to extract this historical data into a usable, consumable format. Now, we had a working extraction model. 

The fact our team was up-to-speed on the latest tech also proved meaningful to the VGM team. “I’ve worked on OCR projects for years…SPR had a level of understanding of the new tools, services that were released by Microsoft in July of 2023, and they had us running on them within months of their release,” said Scott. “It’s that expertise in these newer technologies that really upskills a team.” 

Because VGM met challenges with OCR in the past, we were very intentional about how we moved through this phase of the project. We held ‘go, no-go’ conversations on a regular basis—if we were unable to make progress, we would stop the project to avoid wasting resources. 

“I don't think we've ever gone five hours without meaningful interaction,” said Scott. “It’s that level of commitment that has made this so successful for us. I hope I'm articulating that well because it's just so beyond normal, it’s impressive. In my entire career, I’ve never been a part of something so high-performing as this group.”

Fortunately, OCR was a win! We were able to successfully ingest data from 31 different application versions, including images and PDFs.  

Once we were able to ingest the non-digital application data, we realized that far more files than just applications were stored together. An important component to any cloud project is considering near and long-term costs in the cloud, FinOps. To avoid incurring additional, unnecessary Azure spend and skip the need for a large data clean-up effort to separate out the application data, we fine-tuned by adding a classification model before the extraction model, helping us to weed out the non-application data; we hooked it all together with LogicApps to manage the orchestration. 

Now, with full access to all historical data, we’ve begun to create and train ML models to utilize the extracted data sets to predict the likelihood of claims. This phase has consisted of a lot of back-and-forth experimenting as we define and train the model. Because we finished this first phase of the project under budget and ahead of schedule, our team is now determining how we can leverage the remaining budget to make these models operational and get the predictive data into the hands of VGM’s underwriters. 

Since we are now more familiar with the data at hand, alongside operationalizing the prediction data, we are also considering other ways we can leverage these data sets with additional ML models to empower VGM underwriters and other team members in making more informed decisions.  

When asked about SPR’s ability to deliver “beyond the build,” Scott had nothing but good things to say. “Not only did they deliver a technology solution, but they understood our needs, and adjusted their delivery to those needs, ultimately arriving at something that is going to become meaningful and impactful to our business.” 

“It takes courage and it takes amazing, high-performing people who ultimately are willing to jump into scenarios like this, and arrive at what we've built. So, to me, it's more than just a saying, it's a mindset. It's a personality. It's part of who you are. It's part of your DNA. SPR demonstrated that this entire time.”