Artificial intelligence drives car-insurance claims estimates before the tow truck is called
September 21, 2020
On a typical day, about 80,000 U.S. drivers have accidents serious enough to warrant calling their insurers. After the initial shock comes a predictable sequence of worries: Was anyone hurt? Am I at fault?
The driver’s first call is often to the insurance company, which leads to the next questions: How long will it take to get an estimate, get my car into the shop and then get it back on the road?
The time it takes to settle auto insurance claims is being shortened, and the accuracy of initial estimates is improving, because U.S. insurers now use artificial intelligence to generate repair estimates.
The latest technology powered by AI is much different from the “virtual claim” you might have filed after your last fender-bender. About five years ago, photo-based estimates became increasingly common. Insurance companies sometimes had customers download an app that helped them provide consistent photos, but some insurers just told customers to attach pictures to an email.
Insurance companies liked photo-based estimates because appraisers, who could average only four in-person estimates a day, could complete as many as 15 virtual ones by staying in the office and scrolling through customer-supplied photos on a computer monitor. However, once damaged cars got into body shops, those estimates proved far less accurate than the ones done in person. Insurance companies were bedeviled by costs that surpassed estimates — called claim supplements — sometimes running as much as 50% higher. Customers were frustrated by unexpected delays. And body shops hated being caught in the middle.
That was then. Now, customers can download phone apps through their insurers to guide them through the process of taking and uploading photos that can be evaluated by AI, producing a near-instantaneous damage estimate. The apps are not yet in wide U.S. use, but their time is coming.
The algorithms are trained in image classification, and they identify damage and hand off the claims to companies like Mitchell International, based in San Diego, that price out parts and calculate labor costs. The best algorithms already provide estimates in a few seconds that are as accurate as those produced by experienced human estimators. The pandemic has made AI-powered estimating even more attractive because the technology reduces or even eliminates the need for face-to-face interaction between drivers and insurance adjusters.
By eliminating the need to make appointments with appraisers or make a separate trip to the body shop for an initial estimate, these apps take days off the “cycle time” — how long it takes to get customers back into their cars.
Algorithms also learn and adapt more quickly than human experts. A simple bumper replacement is not necessarily simple anymore, because new bumpers often have expensive integrated sensors, like the ones that warn drivers if they’re backing up too close to another car when parallel parking. As a result, those claim supplements are increasing.
One of the leaders in this “insuretech” market is Tractable, a company based in London that was founded in 2014 by entrepreneur Adrien Cohen and two computer vision experts, Alex Dalyac and Razvan Ranca. Since then, Tractable has received more than $50 million in venture capital funding and grown to over 100 employees in London, New York and Tokyo. Major insurers in Europe and Asia have used Tractable’s AI to settle more than $1 billion in claims.
Dalyac joined Tractable from Imperial College London, where he had led the computing department’s first industrial application of “deep learning.” That’s the approach the company used to train an algorithm to interpret auto damage — a task that had previously been performed only by skilled humans.
“These algorithms are very different to how people used to do computer vision, because you actually get the algorithm to figure out the right patterns in the object,” Dalyac said. “Instead of telling the AI, ‘This is what a front bumper looks like; look for a corner like this and pixels like that,’ you feed the algorithm millions of images. Some contain a front bumper and some don’t. On a rainy day, a dark day, or a sunny one; an undamaged bumper; or one that needs three hours of repair. And the algorithm itself figures out the best combinations of pixel patterns that give it the most accuracy. It’s kind of magical, but it’s very data hungry.”
To date, Tractable has fed its algorithm about 10 million photos of damaged cars, most of which were taken in body shops and submitted to insurers along with repair estimates.
As insurance companies have pulled their employees out of the field, the use of virtual estimates has jumped. CCC Information Services, a Chicago company that markets its own AI-enhanced Quick Estimate app to insurers, recently reported a 125% increase in app use since March — even though traffic levels and accident numbers plunged when states locked down.
Even before the pandemic, major U.S. carriers were exploring the use of AI to speed claims settlements. Liberty Mutual’s in-house technology incubator, Solaria Labs, began work on an AI estimating algorithm in 2018. The company now uses it to give appraisers a head start on estimates.
USAA took a different approach. Rather than develop its own algorithm, it teamed with Google. Customers can upload photos of their damaged cars for analysis by Google Cloud’s Vision API. That damage assessment is then handed off to another partner, Mitchell International, which also uses AI to prepare a parts and labor estimate.
“Today, we send those estimates back to our appraisers because we’re still training the system,” said the company’s chief claims officer, Sean Burgess. “But in the near future, you won’t need that step. We’re going to take the process from days or weeks to minutes.”
That’s the approach that Tractable has taken, too.
“As comfort with the AI’s results is gained,” Dalyac said, “this human quality check is gradually reduced and removed, and so the proportion of AI touchless cases increases.”
Drivers insured by Admiral Seguro, a major Spanish auto insurer that uses Tractable’s tech, can already upload photos and completely resolve some claims — right down to receiving an offer of payment — in minutes on the first phone call.
How soon will U.S. drivers have access to nearly instant claim settlement? Every insurer has to make its own decision about when it is ready to drop that last human quality check, but the day will come. Tractable is confident that it will soon be operating in the United States.
“We are getting pretty close,” Dalyac said. “In the next few quarters, there’s going to be an announcement of a very big American carrier — a household name — that’s going to be doing this.”
Although Burgess said USAA customers would always have the option of a human estimate, it recently filed a trademark on the phrase “Flash Estimate” and expects to bring its own AI claim settlement technology to market in 2021 or 2022.
The rise of AI could be bad news for thousands of people working at insurance companies, but Dalyac bristles at the suggestion that Tractable will necessarily put those people out of work.
“The goal of our technology is to take care of the repetitive, straightforward cases so they can focus on the complex ones, or on providing better customer service,” he said. “Because sometimes when you’ve had an accident, you’re pretty shaken and want additional touch.”
This story was originally published at nytimes.com. Read it here.