Transforming Auto Claim Review with Machine-Learning: Q&A with Dr. Ranjini Vaidyanathan
with Dr. Ranjini Vaidyanathan
Every day, hundreds of CCC’s AI models process billions of points of data from thousands of auto claims. These inputs continually sharpen CCC’s AI models and constantly adapt to the ever-changing industry. AI is enhancing and accelerating the auto claims process to achieve improved outcomes across the ecosystem – for insurers, repairers, and their customers. And CCC’s data scientists are continually monitoring, refining and creating the models to improve and expand the pace of decisions and results.
Dr. Ranjini Vaidyanathan, one of CCC’s senior data scientists, shared insights on this process and the CCC Smart Audit product, a solution that leverages AI to automatically assess auto claims estimates, at the annual Insurance Data Science Conference earlier this year. Below are additional insights from Dr. Vaidyanathan on the product, the future of AI, CCC’s approach to innovation and more.
Tell us how long you’ve worked at CCC and what your primary role is.
R: I joined CCC in 2016 as a data scientist to work on machine learning models for photo estimating. That period coincided with an inflection point in machine learning and deep learning research, with significant advances in computing, algorithms and new modeling frameworks. It was extremely interesting to leverage these and apply them to yield breakthrough results for CCC’s use cases. In my current role, I manage the data science research division in the Innovation Foundry, leading a team of data scientists with full-stack expertise in data science and making it work in practice. We build machine learning and deep learning features for CCC’s product suite – this includes prototyping, scoping, research and development.
Tell us about your experience with data science and AI – what drew you to this field and what excites you about it?
R: I love how multi-disciplinary the field is – so much of data science work involves connecting the dots between different areas. In one of my first months at CCC, I recall using ideas from graph theory to model damage on the car surface, while also implementing an optimized form of the algorithm in Python. More recently, we are exploring ideas from GAN (generative adversarial network) architectures for a completely different potential use-case: returning customized predictions for carriers. (For the record, this doesn’t involve generating synthesized data!)
Tell us what attracted you to CCC.
R: I am an applied mathematician by training, and before CCC, I worked on curated/clean datasets in mostly academic settings. The opportunity to apply the latest technology on real data in a fascinating domain that I had never known about was really exciting to me.
How is CCC using machine learning to enhance auto claims review? What problems is it trying to solve, and what makes CCC’s approach unique?
R: CCC has processed over a trillion dollars in historical claims. With an increasing volume of data, we build more and more intelligent software with predictive capabilities. Predictive models help appraisers write estimates faster, help insurance carriers make more informed decisions on claims sooner, etc. These algorithms, running in production today, have already been applied to over 5 million claims.
One example is a product called CCC Smart Audit, a workflow solution that prioritizes claims audits for insurance re-inspectors using AI models. Our models do this prioritization based on ranking metrics that we’ve identified in collaboration with carriers. When I joined CCC in 2016, the product was good but there was still potential to leverage the latest advances in data science. These brought significant gains to how estimates were ranked, leading to a 20 percent lift in the types of claims selected. Even better, the team was able to improve the product to help insurers identify which claims to audit but which lines in an estimate may need additional review. Once we added that feature, re-inspectors were auditing twice the number of claims within the same amount of time.
When you consider that a major insurance carrier can receive over 3,000 estimates in a single day, a product like Smart Audit can offer significant value to the insurer. A carrier can cover a lot more ground just by targeting the top 20 – 30 percent of estimates that are predicted to need supplements.
What were some of the challenges to building effective machine learning algorithms to optimize auto claims review?
R: Only with complete damage context can an estimate be thoroughly reviewed, and one of the main sources the re-inspector has are photos submitted with the claim. An algorithm needs to be able to look at photos of the vehicle and aggregate information from each of these into meaningful info about damage. We’ve used 3D deep-learning and computer vision algorithms to identify not just where damage is in the photo but where the damage is on the car’s surface.
Another piece of the puzzle was incorporating carrier-specific practices when desired. Every insurance carrier has unique practices and rules. Carriers were able to provide historical data to allow the AI models to learn these correlations, and we also provide carriers the ability to further modulate prediction outcomes using tunable knobs.
Putting these improvements together has created a more comprehensive photo-based smart audit tool to identify claims that may need more scrutiny.
What’s next for Smart Audit?
R: We’re always improving our AI models with more data and the latest advances in machine learning and deep learning. Currently the product identifies lines in the estimate so that the adjuster knows it may wish to review the lines further.
In your experience, tell us what are the key components to building a successful machine-learning model?
There are so many important components, but two things that are top of mind are:
- Data representing the use-case: A clear use-case with data true to the actual real-world process (as much as possible) is ideal. Once these are in place, a lot of the data science process is focused on designing metrics that reflect a successful outcome, followed by developing algorithms that capture and represent signal from different data modalities. For example, the Smart Audit algorithms learn correlations between different lines in an estimate, conditioned on the damage itself.
- Domain understanding is really important. We have ongoing discussions with insurance and estimation domain experts and account for their input right from the time of model design, as well as for consistency checks with results. In one of the first models I developed at CCC, we had been working with the CCC field consultant group, specifically, David Garson (Regional Field Services Manager) who helped us learn more about the repair space. It was very satisfying to see the hypotheses he had also show up in the prediction model’s behavior: for example, damage on edges between car panels impacted the outcome significantly.
What’s next on the horizon for your team and machine-learning?
R: Our main focus right now is straight-through processing with a product like CCC® Estimating – STP. This product will be designed to create a seamless, straight-through experience for qualified repairable claims, leveraging our advanced AI models in collaboration with insurer rules and network connections to guide decision making. It offers a high level of configurability to insurers, and it should accelerate time to resolution for their customers.
We’re expanding our prediction capabilities to new targets that are more challenging. To overcome some of these challenges, we use a few shot learning and other semi-supervised techniques.
We’re also working on finding new ways to leverage high-resolution photos and videos. We have algorithms that automatically capture optimal frames from videos of a damaged car, and we’re applying other approaches in our research like new transformer architectures, differentiable rendering, among others.