Gathering Intelligence

Computer programs are nowhere close to replacing humans in providing financial services functions, especially with regard to dispensing advice to clients. Still, basic forms of artificial intelligence (AI) and machine learning are already up and running and, in fact, may be more common in Canada’s property and casualty insurance space than some might think.

AI “deals with using advanced technology to mimic human cognition and activities,” Mark Breading, a partner with Strategy Meets Action (SMA), notes in the April 2017 paper, AI/Machine Learning in Insurance: A Force to be Reckoned With. “These activities may include identifying patterns, deriving insights, learning from experience, making decisions, and taking actions either autonomously or in collaboration with humans,” Breading writes.

Machine learning, for its part, “is a branch of artificial intelligence that focuses on getting computers to act without being explicitly programmed,” SAS Institute Inc. reports in Statistics and Machine Learning at Scale: New Technologies Apply Machine Learning to Big Data. Central to machine learning is “the idea that with each iteration, the algorithm will learn from the data,” the paper explains.

Some in Canada’s p&c industry are already employing basic forms of machine learning – and the number of insurers dipping their toes into the pool will, no doubt, continue to grow.

With the marked increase in computer processing speed in just the past few years, some Canadian insurance providers say they are using machine learning for, among other applications, generating quotes, determining rates and even deciding which properties to inspect.

Those applications, too, are expected to grow. With regional pockets like the United Kingdom and Silicon Valley already delving deeper into the uses of AI and machine learning, risk management and fraud detection are likely two areas of interest to financial services and insurance players in Canada and elsewhere.

What follows, though, is expected to be so much more.

ADVANCING LEARNING

“Insurers have implemented early forms of AI for decades,” Breading writes in the SMA paper, reporting that “case-based reasoning and rules engines” have been used in underwriting and claims since the 1980s.

London-based Craig Beattie, senior analyst with the insurance practice of Celent, part of Marsh & McLennan Companies Inc., observes that “for an insurer of any size, I would be surprised if they managed to use, across all their applications, no machine learning.”

Jeffrey Baer, manager of advanced analytics at Economical Insurance, says that people “interact with machine learning every day.” Examples include everything from receiving content in news feeds on Facebook to getting recommendations for content to watch on Netflix and email servers that flag messages as spam, Baer points out.

Commenting on predictive modelling in general, “most medium and large insurance carriers in Canada are using predictive modelling in rating and risk selection, and many are also starting to apply predictive modelling in marketing, underwriting, claims handling, et cetera,” he reports.

“Many of these predictive modelling applications rely on traditional statistical techniques like regression, although machine learning does appear to be growing in popularity,” Baer says.

In the past, “someone had to tell a computer exactly what to do and given a set of inputs, you could predict what the outputs would be,” Beattie explains. “Machine learning allows a computer to understand something from experience, so it adapts to a set of data or a set of inputs,” he notes.

“In traditional predictive analytics, every question will need to be defined and tested independently,” notes Jean-Francois Lessard, chief data officer for Intact Financial Corporation (IFC).

“In machine learning, this manual trial-and-error process disappears. With the increase of computing power, the machine tests millions of combinations and comes up only with those that are meaningful, those that have real predictive value. And naturally, the larger the data sets, the more meaningful interaction can be found,” Lessard reports.

LEARNING FROM EXPERIENCE

Machine learning is “an advanced form of AI in which machines can ingest massive amounts of information, detect patterns and analyze outcomes in an iterative manner that continually improves the accuracy of the results,” Breading writes. “The machine is, thus, learning from experience in an automated fashion rather than relying on human intervention or reprogramming.”

At some point, though, human intervention is needed, Baer makes clear.

“Most of our machine-learning applications at Economical are what we would call ‘supervised learning,’ which means that we are trying to determine the relationship between various inputs and a defined output, and that output is known as the response,” he says.

“It requires careful and thoughtful preparation of data by highly skilled humans to feed this data into the machine-learning algorithm. Once the algorithm has identified relationships in the data, we then need to validate that the algorithm has done its job well, and that we can confidently use its predictions to make business decisions,” Baer adds.

Breading has a somewhat different take. “Up until relatively recently, all AI systems had to rely on extracting and codifying expertise from human subject matter experts,” he writes.

“Now, machine-learning capabilities enable AI systems to learn on their own via analysis of massive amounts of data and automated iterations to improve results. The science of machine learning has advanced at the same time that computing power has increased significantly, enabling big data analytics and very frequent iterations,” Breading points out in the paper.

Foteini Agrafioti would likely agree. “I think it is because of the advances in computational power that we are able to even speak about AI today,” says Agrafioti, head of RBC research at the Royal Bank of Canada.

“With the availability of more and more data, as well as with the evolution in computational power, it became possible to finally do simulations that would traditionally have taken a very, very long time,” she points out. “Some things that would have taken us a week to simulate a couple of years ago, we can do that within a couple of minutes today,” Agrafioti reports. “We can fix our algorithms very quickly and that is how deep learning came to be.”

Machine learning is “really taking all those data points to inform logic that your systems and platforms can use,” says Sam Natur, president and chief executive officer of Bullfrog Insurance Ltd., an Ontario-based brokerage that sells commercial coverage over the Internet.

MAKING A BUSINESS CASE

Machine learning can be used for “anything from better user experience to the actual planning and development of products and services,” suggests Beattie.

Baer notes that his company has “used machine learning in a variety of applications, generally focused on improving our customer experience and ensuring that we are operating efficiently.”

Citing as an example, the insurer’s direct writer, Sonnet, it uses data – some of which is provided by the customer – to offer coverage that is “really tailored and customized” to meet the needs of the client, he reports.

To be effective, machine learning requires “business knowledge and context,” Baer emphasizes. “Machine learning requires high-quality data to be effective,” he notes. “Otherwise, it’s what we call GIGO – garbage in, garbage out – and that is never going to work well.”

IFC, for example, is using machine learning to rate its automobile product. “With telematics, we collect 20 data points every second the car is driven,” Lessard reports. Noting that “this is generating terabytes of data every year,” the data gleaned from telematics gives the insurer “insight into individual driving patterns,” he points out.

“Couple that with our policy and claims transactional database, it had a significant lift on our rating formula,” Lessard says. “The data is 30% more powerful than what had previously been our most predictive variable.”

Martha Schrader, vice president of business intelligence and analytics for Northbridge Financial Corporation, reports the insurer is “actively exploring opportunities to use machine learning.” The long-term goal “is to apply machine learning and create a differentiated customer experience,” Schrader says.

In financial services, Agrafioti notes that fraud detection and risk modelling are “two of the more immediate applications” of machine learning. In risk analysis and fraud detection, sector providers “would traditionally employ some statistical models, usually regression models, which have been around for a long time,” she notes.

Any time someone swipes a card, “we are using machine learning to determine whether that is a fraudulent transaction or not,” Agrafioti says. “We have an algorithm in the background learning a particular user as they go because transactions are very specific to clients and a certain pattern of behaviour can be unique to one person,” she explains. Fraud-detection algorithms “learn that uniqueness automatically and adjust to that, so if you do something out of the ordinary, we run something called outlier detection,” she says.

Beattie points out that “most fraud solutions have some sort of machine learning or at least adaptive computing in there as well.”

Machine-learning algorithms “perform very well especially when you have very large amounts of data… where there is a lot of variation,” Agrafioti says, adding, however, that the technology has its limits today.

FOSTERING HUMAN INTERACTION

An area “that has drawn a lot of attention from people in this industry is financial advisors – in particular chatbots for financial advisors,” Agrafioti says. “It’s not something where we feel the technology is there today that we can have AI replace a human in interacting with a client and offering the same quality of service that a human would, but that’s of interest,” she suggests.

That said, Agrafioti is quick to add there are “a lot of interesting advantages should we ever be able to interact to have meaningful communications between humans and machines.”

Baer says each algorithm has its own strengths and weaknesses. “We consider the type of problem we are being asked to solve and select an algorithm that is best suited to the use case,” he notes.

Consider how machine learning can be employed to decide which insured properties should be inspected, Bear says.

“We took a business problem, which is essentially the idea that we have a limited budget to be able to perform property inspections, and we said, ‘This is a great application of machine learning, because we have a lot of historical data that tells us which properties we benefit from inspecting the most and that our property owners benefit from when we provide them with recommendations to improve the conditions of the property,” he reports.

“As a result, we are able to take that data and focus our inspections by using machine learning to identify the properties that would benefit most from that inspection,” says Baer.

Beattie cautions, however, that machine learning can sometimes be “a problem both for the insurer and for the regulator” if a computer program adapts to a set of data or input.

“Over here in Europe, you can’t use gender to drive the price of insurance and what an insurer found when they were using the machine-learning algorithm is that it started to use the height of an individual in order to determine the price,” he relays. “Of course, the average height of a female versus a male is quite different, so there was some interest in that from both lawyers and the regulator in terms of, ‘You are actually using a proxy for gender in order to determine the price,'” he reports.

Although it was not the intent of the insurer to use height as a proxy for gender, “it appeared to be what the machine-learning algorithm was doing,” Beattie says.

“So there is a challenge here, I think, both for the regulator and for the insurers and for chief data officers who are trying to communicate what these algorithms are trying to do. Is somebody watching what these machines are doing in order to make the decision about whether it is appropriate or not?” he asks.

SHARING NATURAL LANGUAGE

Beattie reports that some applications of AI include optical character recognition, autonomous driving and natural-language processing.

Ontario-based Mitchell & Whale Insurance Brokers Ltd. is one brokerage that is currently using natural-language processing, via a chatbot that provides quotes, accepts claims, sets up calls and deals with change requests, among other services.

Natural-language processing “is just the ability to interpret words for their intent, so understanding auto to mean car or vehicle and things like that… without having to fully program out all of those,” says Mitchell & Whale president Adam Little. “The machine learning part of it would be progressively picking up answers. So you can do it a couple of different ways – either by watching the answers that are being given and then correcting it, or you can have it watch a human and then it will start to pick up the patterns,” Little notes.

The brokerage is also beta testing an application that would read incoming email, he reports. Noting that “there is a changing trend in demographic,” Little suggests “most Millennials don’t want to talk to people at all.”

Natur cites how responses can be made more attractive to certain segments by simulating texting. “One usage of AI is emulating the experience of texting with a friend in the context of a claims submission. I have seen some of our competitors doing that already. During the entire claims submission, it feels as if you are texting back and forth with a human and you are not,” he reports.

Beyond providing service in line with how a particular customer chooses is offering that information with a view to helping prevent or mitigate damage and losses.

In the United States, for example, Natur says that some companies are using machine learning to alert property owners that there is a hailstorm approaching. Customers are warned either by email, text message or phone, depending on their preferred method of communications, he reports.

Citing the “venture capital interest in insurance technology” in many cities in the world – including Singapore, London, New York and San Francisco – Beattie’s advice to Canadian insurance providers is to “try and keep a finger on the pulse of what is happening in the various insurtech hubs around the world.”

What is happening elsewhere, coupled with developments unfolding at home, may provide the kind of lessons learned that will allow insurance providers to identify how best to employ AI and machine learning not just today, but in the future.