Suncorp deploys Watson to determine claims liability

Suncorp has put a Watson-powered accident liability determination system into production for its AAMI, GIO, APIA, Bingle and Suncorp insurance brands.

The company first revealed it was testing IBM Watson back in June. The system has ingested 15,000 anonymised claim files, which it uses as the basis to determine who is at fault in car accidents.

Suncorp insurance CEO Gary Dransfield told iTnews the digital team started looking at Watson last year to determine liability initially just for claims lodged with AAMI.

Of Suncorp’s insurance brands, AAMI has “historically quite a high level of self-service take-up”, and Suncorp wanted to extend that capability even further.

This is presented as an option available to customers when they interact with the insurer, allowing them to perform more of a transaction online if they wish. However, they can still “drop out of the digital process” and talk to a consultant at any point.

“It’s part of a broader process we’ve built called zero-touch claims beginning with the lodgment online, Watson liability determination, and then once the liability has been determined the customer can choose to book their repairs as part of a seamless digital process, or they can drop out and talk to a consultant,” Dransfield said.

“We didn’t just want to pressgang our customers into being forced down a digital process end-to-end.”

The Watson-powered liability determination engine has now been integrated into the claims system shared by Suncorp’s major retail brands.

Keywords, rules and confidence

Like many enterprise AI projects, Suncorp’s journey began as an informal “passion project” of a member of the financial company’s digital team.

“Really, to a degree, it started off as a skunkworks for one of our smaller online only brands to try to prove out the concept,” Dransfield said.

The first trial was months in the making, given it occurred “in parallel with the day jobs” of the digital team.

“It’s fascinating that it’s very quickly become quite core to our claims processing,” Dransfield noted.

“As people started to talk about it a bit more [internally], you could see that there was potentially quite a strong uplift in customer experience that it could deliver.”

When a car collision claim is lodged, there is typically a “slight lag” involved while the insurer tries to determine who was liable for the accident.

This was considered a “pain point” for customers because it delayed the assignment of a repairer, which in turn kept them off the road longer.

Liability determination, in and of itself, is also a challenge.

“If you’ve got a two-vehicle motor accident, neither party are typically in a rush to have liability determined against them because that then affects no claim bonuses or discounts, claims histories and the like,” Dransfield said.

Insurers look at a number of data points to determine liability. Much of this data is collected directly from the customer, as they are asked to share their version of events as part of the claims process.

IBM’s Watson natural language classifier is used to analyse the text submitted by the claimants. It then makes a determination by applying business rules and recognising keywords, and attaches a “confidence level” to the decision, based on what the system knows of similar decisions that have occurred in the past.

“When a customer lodges [a claim] online, the classifier capability of Watson is analysing that text description to try and determine who was at fault in the accident,” Dransfield said.

“Put simply, it’s looking at the free text and applying business rules to the text that will be indicative of liability. For example, it looks at the direction of travel, the driver’s sequence in a string of vehicles, and the status of the car when a collision occurs in terms of whether it was moving or not.”

If Watson is unsure - that is, if its confidence assessment is below a certain threshold - then the claim is referred to a human overseer for vetting.

“When Watson returns an assessment and it doesn’t have a high confidence rating - it may still be accurate but the system’s not sure - that will get referred to a human consultant to review and look at in-depth,” Dransfield said.

Watson is able to accurately determine liability in “around 90 percent” of the cases it is referred.

Just how many cases Watson is now working on is hard to determine.

Dransfield suggested Suncorp saw “hundreds of thousands of car claims each year” from its “many millions of motor policy” holders.

The majority of those claims are still not lodged digitally but rather over the phone.

“The maximum population [of claims that Watson] can deal with are those claims lodged digitally,” Dransfield said.

“That’s only a subset of motor claims. [Specifically] it’s the subset of those ones lodged online where the customer doesn’t want to drop out and engage a human around the [liability] determination.

“Well less than 50 percent of claims are lodged digitally, and then it’s a subset of those that the customer wants to stay in an online process for [that are sent to Watson].”

However, the level of digital engagement varies by retail brand.

Some brands - such as AAMI - saw customers demanding a greater degree of self-service capability due to the demographic they appealed to.

AI evolution

Dransfield said Suncorp is looking more broadly, both inside and outside of insurance, for further opportunities to streamline processes with either AI or robotic process automation (RPA).

“We’re scanning right across the business for opportunities to improve customer experience with some form of automation,” he said.

On the RPA front, Suncorp is not backing a single software or services provider but is instead looking to its business process outsourcing (BPO) partners, domestic services providers, and the major software players in the space for guidance on how to proceed.

This project has been named a finalist in the finance category of the iTnews Benchmark Awards 2017/18. The full list of finalists can be found here.