Speedy recovery: How AI research is helping insurers boost claims response

The calculus of catastrophe is getting harder to factor, and 2021 was one of the hardest tests yet for the insurance industry. 

A February blizzard shattered pipes and power grids across Texas and Mexico. Summer heat waves from Kazakhstan to British Columbia were so intense, forests combusted. Another heavy season of fires in Australia, California, Greece. And then came the rains. A ten-day storm inundated Europe from London to Croatia, with particular devastation in Germany and Belgium. Hurricane Ida, one of 20 to make landfall, gouged an unusual path up the middle of the U.S., flooding heartland cities as much as New Orleans and New York. 

Insurers have always approached the duty of sorting out disasters with a careful balance of diligence and expediency. Yet as climate change makes the destruction more frequent and widespread, carriers are being tested as never before. Especially for those affected by natural disasters, a timely response remains paramount, even as insurers must manage a growing number of complex claims. 

Practical tools that allow insurers to marshal people and resources where they’re needed most are becoming increasingly essential to deliver on the policyholder promise. 

One area where carriers are looking is a better view of the current destruction at hand. This can be invaluable not only for recovery but also in terms of prevention and mitigation of future damage. One of the best ways to gain such perspective is from above, as satellite and aerial technology continue to advance in both affordability and access. At the same time, this availability, coupled with the marked rise in disasters, leads to more ground to cover, figuratively and literally. 

Insurers are increasingly turning to AI to help sort through claims and destruction, even putting experimental research to the challenge, as Munich Re and Google Cloud have recently undertaken. 

“If your house gets destroyed by a hurricane, for example, every day you can return to normal sooner counts—this includes quick reimbursement, as well as every day less in the hotel,” Johannes Kuhn, a senior technical data specialist at Munich Re, said in a recent interview. “Given the increasing frequency of storms and other weather-related natural catastrophes, it's important to manage the bottlenecks in the best possible way: There’s only so many loss adjustors, they’re increasingly stressed and stretched thin, especially during major events.” 

And so, Kuhn and Thilo Horner, a senior venture building specialist, are hard at work looking for technology to make their lives easier. “Any way to support adjustors to perform assessments easier and faster is hugely valuable for effectiveness,” Horner said. “And early and concrete information helps both the injured party and the insurer make the best decisions. We’re trying to address all of this with our technology solutions.”

Horner leads a team at Munich Re, one of the leading reinsurers in the world, that has been very active in building AI models to help understand the scope of damage following a disaster even faster and better, notably through Munich Re’s CatAI program. Eager to improve on CatAI, Munich Re began working with Google researchers to further adapt AI tools from across industries—including visual inspection programs first deployed on assembly lines, and cutting-edge AI-specialized processors. Their goal was to further enhance and accelerate the development and deployment of Cat AI models to support claims adjustors, who in turn help policyholders bounce back from storms quicker. 

The specific aim has been to optimize and accelerate the use of computer vision as a means to analyze post-storm aerial imagery, which in turn helps identify the most damaged areas following storms. The company not only offers such services to its own reinsurance clients, either. In the spirit of every company becoming a tech company, they are developing a suite of services available to any insurance carrier, regardless of affiliation. 

“Digital services are definitely having a growing impact on the solutions we offer to our customers and our customers’ customers,” Kuhn said. 

With AI, a new view on the world

In the case of its aerial imagery offerings, Munich Re is helping to solve a problem that is both logistical and financial. Overhead imagery is nothing new for insurance carriers. For decades, insurers have been surveying the aftermath of storms from above, first using airplanes, then satellites, and now drones have entered the mix. Policyholders also regularly submit their own photos and videos to help their carriers sort through, and sort out, their claims, using new services like Munich Re’s Remote Inspection solution (which itself has AI built in to help streamline data intake and tagging). 

The growing stream of information is hugely valuable, and the more there is, the more insights Munich Re and its partners can glean. At the same time, they must be able to assess it all, and quickly, which is where AI comes in. 

“This is one of the few use cases where saving money ranks behind saving time,” said Dr. Ulrich Nöbauer, the senior technical data specialist at Munich Re helping lead the project. “If we can enable a model with only 10% of the label data as before, we gain a great advantage. This not only applies to the claims adjuster and the insurer. Homeowners also benefit directly, and, by extension, communities win by getting rebuilt faster. Everybody wins.”

Insurance carriers have a history of dealing with huge amounts of data, and they often look to the latest technology to help. The work on CatAI is a natural extension of this tradition, and the significant growth in its use is a testament to its effectiveness. At the same time, given the particulars of each storm and each location, at least an initial phase of human review is still required. 

Insurers often turn to labeling companies to help sort through an initial batch of aerial imagery, and the selection is then used to train AI models. A review for each image can cost upwards of $3. With potentially hundreds of thousands of images after a major disaster, the cost of such analyses can be substantial, as well as consuming valuable time, during which damaged properties often remain unprotected and exposed to the elements, and adjusters may be tediously fanning out over an area, instead of focusing their efforts. 

Munich Re and Google Cloud had already begun working together in the area of data-driven cyber insurance in the spring of 2021. Seeking other areas of cooperation, and with their shared experience in machine learning, improving the work already done on Cat AI provided a fruitful opportunity to collaborate. 

“Our mission is to help the insurance industry provide the gold standards for data, analytics, AI, and ML,” Dr. Henna Karna, Google Cloud’s managing director for global insurance and risk management, said. “Working together with Munich Re, we’ve been able to discover and use insights across multiple use cases.” 

Semi-supervised learning, fully capable in the field 

One of the first places the teams looked for opportunity was Google Cloud’s Visual Inspection AI. Originally conceived as a tool for manufacturing, to spot defects within the consistent output of an assembly line, Visual Inspection AI quickly proved itself similarly adept in the highly unique environment of property damage, as well.

Even if drawn from the same blueprints, rarely are two homes exactly alike. A garage or deck, new paint, an herb garden, oak tree or bed of roses can make each appear different, as can the seasons. All that variability can make differentiating damage from other features a challenge for computer vision.

Thanks to the advanced, off-the-rack capabilities of a cloud-based program like Visual Inspection AI, Munich Re found it could more quickly deploy new models to CatAI and the customers using it without having to develop custom programming for each new disaster, as had been the case. While results were comparable to their current approach, the speed of deployment and scalability proved useful.

“Tools like AutoML are so impressive for the way they can help customers create their own models on the fly, without having to build from scratch,” Nöbauer said. Even bigger breakthroughs came from applying new AI techniques, known as Semi-Supervised Semantic Segmentation with Cross-Consistency Training, or more generally semi-supervised learning. 

Rather than taking a decent-sized volume of labeled data to build a model, semisupervised learning only requires a small amount of labeled data—about one-tenth as much, in the Munich Re and Google Cloud work—and pairs that with a larger volume of unlabeled data.

That larger quantity of data must be available, of course, though in the case of a natural disaster event, that is rarely an issue. And while requiring an overall greater volume of data, the new models can ultimately achieve results in about a third the time as traditional models, given how much less labeling must be done. Over time, this efficiency could likely increase, too, as the models replicated and learned from themselves, building greater accuracy. 

Thus fewer human-labeled images, and less time, were required to build a more accurate understanding of the most affected areas. 

Saving time, money, and communities with AI 

Hoping to advance the results even further, the team ran some of their semi-supervised routines on mainframes built with Tensor Processing Units, chips specially designed by Google to accelerate AI workloads.

Raw training times were reduced by a remarkable 92%. Running TPUs is currently more Aerial costly than traditional mainframes, so there were some tradeoffs, but the overall savings was one-third the cost of traditional methods.

So with this trio of tools, a post-disaster model in CatAI could be achieved at a fraction of both the time and cost. At present, the combined performance and total cost of ownership are nearly 19-times greater than traditional methods when Visual Inspection AI, semi-supervised learning, and TPUs are used in concert. The savings and performance can be expected to grow as the technology becomes more available and the models better trained by more data inputs.

Nor are the possibilities limited to insurance. 

“This research really showed us the potential to get Visual Inspection AI out of a controlled setting like a factory and out into the real world,” said Anant Nawalgaria, a senior machine learning specialist at Google who led the research on that side. “And when you add in TPUs, the ability to work not just in the real world but in real time only grows. Anything with segmentation or classification problems, like retail or construction, could really benefit.” 

For now, Munich Re is devising how best to integrate this research into its own real-world products, Kuhn said. 

“Thinking into the future and with many specialties that we insure or reinsure, the scope could be expanded beyond homes and businesses,” Kuhn said. “There’s so many use cases out there for this technology, the more we can scale it, the more we can achieve.”