We are excited to bring Transform 2022 back in-person July 19 and virtually July 20 - 28. Join AI and data leaders for insightful talks and exciting networking opportunities. Register today!
The immediate aftermath of a vehicular crash is difficult for all parties involved. The drivers have to process the trauma, address injuries with medical attention, exchange information, and figure out how to get their damaged cars repaired or replaced. Insurance companies have to conduct damage assessments and figure out the most efficient and least expensive route to get the claims resolved. Claim adjusters, body shop workers, and various other parties all have a say in the process, which can often take weeks to wrap up.
In addition to the time and trauma involved, crashes generate a lot of data, whether they’re pictures of damaged parts or associated documentation from police reports. In addition, the frequency of crashes — 2019 saw nearly 6.8 million vehicle crashes in the United States alone — means a large volume of data to be processed constantly. Auto insurance claims result not just from crashes, but also from other kinds of damage, such as floods and trees falling on bumpers.
AI ramps up
These collective factors make for a particularly compelling argument for implementation of artificial intelligence in claims processing, says John Goodson, chief technology officer at CCC Intelligent Solutions, a technology solutions provider for the automotive and insurance industries. (CCC is itself not an insurance company.)
The use of AI in insurance claims processing has been steadily accelerating. CCC reported a 50% year-over-year increase in the application of advanced AI for claims processing in 2021. The company reports that more than 9 million unique claims have routed through its deep learning AI solution – a number that grew more than 80% in 2021.
When a crash claim comes through, the insurance company has to dispatch claim adjusters who attend to a laundry list of questions: is the car completely damaged or can it be fixed? How much will it cost? What’s the best way to fix the car? Where should replacement parts be sourced? Will the parties need a rental?. The same questions need to be asked every time, which makes them particularly suited to a deep learning model: understand the damage and solutions from previous crashes and apply that learned knowledge to future ones.
CCC processes about 16 million auto crash claims annually, which gives it a rich base of data on which to base AI models. CCC’s deep learning model is built on billions of photos of vehicular damage, incident reports, and line items from claim forms. Deep learning algorithms and computer vision begin to detect patterns — a dent that looks a particular way will need a certain kind of intervention and will cost a certain number of dollars — and deliver recommendations for next steps.
“We have built an AI model that determines line by line the things that are going to be needed to be done: what parts are going to be needed? How much time is it going to take to be repaired? There are many questions that we answer,” Goodson says. CCC’s algorithms also sort data by car model, so the algorithm can continue learning as new claims are filed.
If a customer files a report on the scene with pictures from the crash, those are matched against the database to find pictures from the similar model along with learned information about what various repair processes involve. “We can immediately say, ‘this car is going to be repairable, here’s the shop to take it to if drivable, if not, call this towing agency,’” Goodson says. “It makes the entire process much faster and much less traumatic for the consumer and much less labor-intensive for the companies involved.”
To make its AI-delivered recommendations more understandable, CCC delivers its estimates with “heat maps” that highlight the damaged spots and make them easier to visualize.
Building robust ML models
To decrease bias, CCC scrubs its models free of identifying information such as vehicle identification numbers, street address, and town or city names. License plates are also obfuscated. “It’s a very strenuous process to make sure the data is really ready to be trained on,” Goodson says, estimating that nearly 35% of their time is spent on data readiness.
Natural language processing (NLP) comes into play for documents that might not be easily digestible in digital formats.
Insurance claims processing is a particularly good fit for AI applications because of a large bank of data and the possibility for inference-based recommendations to apply. Similar mechanisms can translate to other industries with documentation-heavy tasks and a large repository of information. Goodson cautions against leaning on AI to gain time efficiencies without basing it on robust data. “Most companies want an AI practice but they don’t have enough data or they don’t have ethics principles in place to ensure that bias doesn’t creep in,” Goodson points out.
“You have to train and retrain your model if biases do surface, you really can’t take shortcuts, you have to pay a lot of attention to data cleaning and readiness,” Goodson says. What does “enough” data look like? “It’s subjective to the industry and it’s hard to answer, but it’s definitely not in the hundreds,” Goodson says.
CCC’s own AI ventures will move toward processing information at the edge. In the future, expect that consumers can simply livestream a video of the damage to the insurance company’s portal and receive instructions about next steps in minutes. “We’re using advancements in AI to advance not only our back-office techniques, but to leverage that technical capability to advance our [front-end] solutions as well,” Goodson says.
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Learn more about membership.