Next-Gen Insurers Are Going to Need (Way) More AI Horsepower

While involving AI/ML in the complex process of insurance claims now might be piecemeal, the future is bright for insurers to speed time to claim resolution by using image-based data and machine learning models to understand the scope of damage to vehicles or eventually, entire geographic regions.

Tractable, a UK-based startup, is one of several on the front end of this evolution in insurance claims. While their business started with using machine learning models to pinpoint the extent of damage on cars, they say they are working with twenty of the largest global insurers, to branch into home and property insurance.

The timing couldn’t be better. With climate-driven catastrophes increasing in frequency, leaving areas difficult to access for humans to understand the extent of insurance claims, bringing rich models with satellite and other data to bear makes sense. But as with other areas of the economy that have long since had many humans in the loop, making AI the trusted source for claims might take some time.

There are a few things that might happen over the five years. First, we might expect large insurers to kick the tires with AI using standalone startups (like Tractable) as a proof of concept. If the time to claim completion improves dramatically, we could see acquisitions of those teams and talents to bring that capability in-house as a strategic asset, or to see insurance companies hiring their own legions of computer vision experts to streamline the claims flow.

Either way, expect this industry, whether its through startup intermediaries or eventually, the companies themselves, to need a lot of AI training and inference, especially if the “boots on the ground” approach to assessing claims becomes less important if only for speed’s sake.

Tractable’s CTO, Razvan Ranca, tells us while there are still humans involved, the claims process, at least for vehicles can be shifted from close to two months to a few days. “This will be gradual, it’s a change management process,” he adds. “But we are demonstrating it works and shifting the status quo in the market.” The company is currently working with an insurance company in Japan to apply the technology to typhoon damage claims and expects the home and insurance areas to keep growing.

For our purposes here at The Next Platform, of course, the real question is what an AI-driven insurance industry might look like from a systems and software standpoint. Given the uniqueness of claims in any arena, vehicle or land, having a go-to training set that can be retrained seems a difficult task. For Tractable’s car insurance claim arm, retraining on models of a wide range of vehicles, components, and damage types is one thing. For specific geographies, it might get quite a bit more complicated, pushing the need for ever-increasing training investments.

In the case of automotive, as the company enters new markets, it encounters cars with different design features, even among vehicles of the same make and model. There are other variations that also have to be accounted for, including different methods used to fix vehicles in different countries. All of this must be learned by the AI for each territory that Tractable takes its service to – making training and model refinement an everyday activity at Tractable.

Right now, for its car claims business, the company is using AWS for production AI training and inference runs. “We’re using something like 60-70 GPUs for model and training and research, then another 40-50 in production,” Ranca says.

Tractable began with TitanX gaming graphics cards from Nvidia and has moved through the GPU product timeline, using K80s and now, latest generation A100 GPUs on AWS for production with a mix for their on-prem hardware.

The company now has a new addition to their model training and research arsenal—a Graphcore IPU POD system. He says they went through a successful proof of concept before deciding on the system, which so far has proven a 5X speedup over equivalent hardware at less cost, Ranca says.

The software framework “took some investment” he adds. Teams at Tractable had to spend a fair bit of time wrapping their heads around the Poplar stack “but for 5X, which means we can run our models that much faster, it’s a worthwhile investment.”

The insurance industry is no different than others where the value of field employees making decisions is less efficient and in some cases, accurate, than what a well-trained model can deliver. However, what makes insurance different is that there is no one-size-fits-all model that can be optimized or added to for nuance when branching into the property domain. AI will have to get better, but so too will the sources of data for an ever-expanding cadre of calamities beyond mere auto claims. This also means hardware will have to be more affordable and efficient. Larger, more complex, varied models mean big training times. We’ll keep an eye on the ROI, as always.

Sign up to our Newsletter

Featuring highlights, analysis, and stories from the week directly from us to your inbox with nothing in between.
Subscribe now

Be the first to comment

Leave a Reply

Your email address will not be published.


This site uses Akismet to reduce spam. Learn how your comment data is processed.