Canadian Underwriter

Humans required to train AI to assess images for damage losses

August 22, 2018   by Greg Meckbach

Print this page Share

Claims adjusters could soon have part of their work replaced by software that looks at photos and decides whether property is damaged, but there is still a need for humans in the process, one researcher predicts.

When properties are damaged in a major catastrophe by a wind storm or flood, “typically it takes the insurance company a few weeks to send someone out to assess the damage to property,”  Aditya Kaul, lead analyst for research firm Tractica, said in an interview Tuesday.

To expedite the process, some P&C insurers are using unmanned aerial vehicles (or drones) to survey damage and take pictures. In some cases, they use computer software to assess the damage in the pictures automatically. This is also true for auto damage, explained Kaul, who works out of London, England for Boulder, Colo.-based Tractica.

Using artificial intelligence, computer software can be programmed to flag certain things in images. Based on a photo, the software is designed to decide whether or not a piece of property is damaged. Some claims organizations “automate the process to a considerable degree, so rather than waiting a few weeks, you typically would get the decision on the claim in a few days,” Kaul said. “That would not be possible without artificial intelligence.”

But there is still a manual aspect to the process right now. “Someone has to sit and tag a lot of these images, train these models and a lot of effort needs to be put in [by humans],” Kaul said.

In a few years’ time, AI models will come pre-trained to identify roof damage or flood damage. So it is possible that data service vendors or business software vendors would sell software that is “trained” to decide whether or not a photo shows flood or wind damage. In those cases, carriers and adjusters do not “have to go and train these models,” Kaul said. “These would be pre-trained with existing images. Someone can just pick these off the shelf.”

Kaul does not know of any vendor providing this specific type of service, although he said there are thousands of “one-man garage shops” in the technology industry. “There is a gap in the market.” he said, but has no doubt there will soon be companies providing these pre-trained software algorithms.

In the meantime, Kaul said, “I would say over the next three, four or five years, as this gets rolled out, there will definitely be humans checking a lot of those decisions because the trust in AI is not that high. Up until now, it has been a human making these decisions, and it is a business-critical decision, so it is going to be hard for an insurance company to completely trust AI.”

Tractica announced Monday the release of a report, Artificial Intelligence Market Forecasts, which forecasts that global revenue from AI software implementations will increase from $8.1 billion in 2018 to $105.8 billion by 2025. All figures are in U.S. dollars and are for AI software spending by all kinds of organizations – including but not limited to insurance.