November 2017    |    14 minute read   |    By David Gambrill

— Hi Hal.
— Hi Dave.
— I need insurance for a new sports car I just bought. Here’s a picture of it.
— Nice, Dave. Will Maureen be driving it?
— No, she’s going to take the other car.
— Okay then. Is your profile with me up to date?
— Yes.
— In that case, I’d say Company XYZ’s policy should do the trick. Here are the details. Just click the button and I’ll email you the bound coverage.
— Thanks Hal.
— Happy driving, Dave.
Dave exits the chat with Hal, a robo-advisor on Broker ABC’s app, and puts away his smartphone.

Can machines with artificial intelligence (AI) replace humans as insurance professionals?

Five years ago, the question seemed absurd, a mere detail in a science fiction movie. But with rapid advances in collecting data via the Internet of Things (IoT), as well as new powerful and cost-effective tools for analyzing Big Data, AI is quickly becoming fact rather than fiction.

“Despite skepticism of AI as just another technology buzzword, its momentum is very real,” Accenture stated in its Technology Vision 2017 report, which includes the results of a global survey of more than 5,400 IT and business executives. “Eighty-five per cent of executives we surveyed report they will invest extensively in AI-related technologies over the next three years.”

Most (79%) of the executives in Accenture’s survey believe AI investment will help accelerate technology adoption throughout their organizations. Out of the $1.7 billion invested in the Insurtech sector last year, more than 40% was aimed at AI, according to Norman Black, insurance industry principal consultant at the tech company SAS.

So, what is AI, exactly? And to what extent is it being used in the Canadian p&c industry?

What is Artificial Intelligence?

“AI is a rather loose, amorphous term,” says Black. “Essentially, AI is the capability of a computer to imitate intelligent human behaviour.”

AI is made up of a lot of different technologies; since these technologies are commonly used in conjunction with one another, overarching definitions of AI tend to be elusive. For simplicity’s sake, AI can be subdivided into technologies that mimic specific types of intelligent human behaviour — learning, speech and language, vision and perception, and analysis.

Learning
Machine learning, a subcomponent of AI, enables a computer program to analyze data while continuously adjusting future behaviour and predictions based on previous analyses.

Machine learning can happen with or without human assistance, notes Jamie McDougall, vice president of business intelligence and analytics at Gore Mutual Insurance. He gives an example of a computer learning to identify cat images with human assistance. Humans supply the computer with pre-labelled cat photos; the computer compares random images it receives with the pre-labelled photos.

“When processing any given photo, the computer will seek to determine, ‘Is this a cat?’” McDougall says. “Based on whether the answer is yes or no, the machine will be able to figure out the differences between which edges in the photo define a cat. It will figure out the elements of cat images, such as fur (softer edges), eyes, or legs, until it can finally put all of the detailed segments together into an overall determination of a cat.”

Deep learning is a much more complex offshoot of machine learning, characterized by sophisticated neural networks. The computing system essentially mimics the biological neural networks that allow animals to learn.

AXA, for example, used an open-source deep-learning framework called ‘TensorFlow’ to predict large-loss traffic accidents with an accuracy of 78%. Analogous to McDougall’s cat example, AXA used large-loss data in Google Compute Engine to train the TensorFlow framework to recognize “large losses” of more than $10,000. The company then fed 70 auto insurance data points – including age of the driver, age of the car, record of previous accidents, annual insurance premium range, region of the driver’s address, etc. – into the neural network. Using this data, the framework would identify the likelihood of a large loss, allowing AXA to optimize its premium pricing model.

Speech and Language
Machines use Natural Language Processing (NLP) to understand the meanings of words in conversations, both in audio and written text. NLP uses statistical modelling to help recognize topic areas and patterns of words occurring in a collection of documents [i.e. claims investigation notes or medical reports]. It can mine online/offline interaction data [i.e. mainstream media or public records] to help build predictive models; it can also analyze sentiments expressed in unstructured data such as social media posts.

Chatbots are a common example of machines using NLP. Insurify, an MIT-founded insurtech, used NLP to analyze more than 20,000 car insurance chat conversations to build its virtual assistant on Facebook Messenger. “The virtual assistant acts as a broker, interacting with the consumer via Facebook Messenger to assess their cost and policy benefit preferences,” as Max Kraus of life and reinsurance consulting company LOGiQ3 explains in a blog piece. In a separate example, Lemonade, a New York-based insurtech, allows an insured to purchase renter’s or home insurance in under two minutes, exclusively acting with Lemonade’s AI chatbot, ‘Maya.’

Vision and Perception
Image analysis identifies objects, locations, events and individuals in photographs and video, Infinilytics, a technology-based startup in Silicon Valley, writes in its blog. The company provides AI and machine learning services to the insurance industry.

“Imagine the possibilities of receiving a live video stream of an area just destroyed in a catastrophic event or firestorm,” Infinilytics writes. “What about photographs of vehicles used over and over in multiple false claims (each time masking the true identification of the vehicle both VIN and license plate)? A robust image analysis solution can help validate genuine claims, and stop the fraudulent ones before a payment is issued.”

No longer the stuff of imagination, live stream video captured by drone flights helped adjust catastrophic home insurance claims in response to the wildfire damage in Fort McMurray, Alberta last year. Calgary-based Ventus Geospatial received emergency clearance in 2016 to operate an unmanned aircraft over the areas affected by the wildfire.

The initial risk assessment was completed by Opta Information Intelligence (Opta), a technology and analytics company serving the Canadian p&c industry, using AI imaging. “The images allowed insurers to see what percentage of the homes were damaged by fire,” says Greg McCutcheon, president of Opta. “There is a lot of interpretation involved in assessing damage based on images. A lot of technology was used to identify all these features.”

Image analysis is also part of a claims process innovation announced in October 2017 by Mitchell, a technology provider to the collision repair segment of the Canadian p&c industry. Mitchell WorkCenter Assisted Review uses visual computing to analyze photos and help identify incorrect replace or repair decisions; as a result, insurance companies can review more estimates in less time while refining estimating guidelines and consistency.

Analysis
Advanced analytics, a discipline related to AI, offers a way to sift through countless gigabytes of data, extract meaningful insights, and make business decisions supported by timely and accurate predictions. This is especially important in a world living under Moore’s Law, which states that computer technology will double in power and halve in cost every 18 months.

IBM predicts insurance data will increase 94% by the end of 2018. A separate IBM study shows the average industry IT department is wrestling with annual data storage compound growth rates of 60%.

The Internet of Things (IoT) will increase the amount of data exponentially, Black observes. “We will have much more data than any human being can possibly analyse, so you have got to apply machines to that data if you are going to use IoT data for anything.”

IoT refers to sensory data collected through devices connected to Internet. Examples include smart phones, vehicle parts, security monitors, biometric wearables, household appliances, etc. It is estimated that by 2020, each person will have several devices connected with the IoT. “This is a significant amount of data that can be gleaned and analyzed in the business of insurance and risk management,” Infinilytics observes.

One obvious use of advanced analytics is in the field of telematics — also known as usage-based or pay-as-you-drive insurance. In a telematics program, a small wireless device is installed into a car’s diagnostic port (typically under the steering wheel). The technology assesses driving habits, including distance driven, time of day, and how the driver accelerates and brakes.

Intact Insurance uses advanced analytics to sift through telematics data; the results help the company to rate its auto insurance product. “We collect 20 data points every second the car is driven,” Intact says in an emailed statement about its AI initiatives. “This is generating terabytes of data every year. This is also a very good [risk] selection tool for us…The data gives us insight into individual driving patterns and, couple that with our policy and claims transactional database, it has a significant lift on our rating formula. The data [obtained in the telematics program] is 30% more powerful than what had previously been our most predictive [auto insurance risk] variable.”

Who’s using?
AI in Canada’s p&c industry
Ironically, it is difficult to find data on AI use in the Canadian p&c industry. One complicating factor is that Canadian p&c insurance organizations compete on technological advantage; they don’t want to negate this advantage by revealing their forthcoming innovations. In the absence of any seminal research study on the topic, AI’s footprint in Canada is largely a matter for conjecture.

“AI is still at an early stage in the Canadian p&c sector,” says Black. “There is much more progress in sectors like the U.K. where competitive pressures are very intense.”

Barriers to AI use in Canada include high investment costs, which could be prohibitive to small insurers and brokerages, as well as a general lack of trust in allowing machines to handle personal interactions with humans. “Trust levels have not evolved to the extent of throwing away what we do now,” says Greg Purdy, co-founder of getClarity, an analytics company serving the Canadian p&c industry. “The regulatory and legal infrastructure around AI just isn’t there yet.”

Another barrier is what McDougall calls “the velocity of data.” Machine learning is dependant on having large enough data sets against which to compare. Considering claims frequency and policy renewal periods, p&c data is generally less voluminous than in other financial industries. “That means there is less data available to train machines based on data observation,” McDougall says.

Despite the barriers, Canadian p&c insurance organizations are investing in AI.

Insurtechs are one route into AI investment. Five of the largest 15 insurtech investments deals in 2017 went to AI startups, Kraus notes. Investment firm Mill Street & Co. in early 2017 bought a stake in the Toronto-based p&c insurtech Tuque, which, when launched, will offer digital solutions for buying home, auto and business insurance online. In October 2017, Montreal-based Covera Technologies Inc. raised $1 million in seed money for its AI work in helping consumers to re-shop their insurance automatically.

Alternatively, PwC suggests a type of “sandbox” approach for p&c organizations interested in experimenting with AI. For example, an organization might build an AI pilot project internally using existing vendor solutions or open source tools; then they conduct “parallel runs,” comparing and improving the results of their AI solution against the decisions of human beings. Organizations could also create their own proprietary databases – “crash images” from claims data, for example – and then measure the accuracy of AI algorithms against them. The Cookhouse Lab is an open, collaborative space where p&c professionals, finance service professionals, entrepreneurs and academics can discuss the development of AI and other insurtech projects.

Across all sectors, Canada is often described as positioned to become a “global hub” of AI research, with approximately $400 million committed to AI research initiatives in 2017. This includes a total of $170 million committed to Vector Institute, a University of Toronto affiliate, by Intact Insurance, life insurers, Google, Uber, and the federal and provincial governments.

Intact created its Intact Data Lab in 2016. The Data Lab has 30 resources, including actuaries, data scientists, geomaticians, software engineers, and meteorologists dedicated to analyzing data with machine learning. “Our goal is to have our 200 actuaries doing data analytics with these new [AI] techniques,” the company says.

McDougall thinks p&c claims organizations would benefit from using the same AI technologies banks use. “I’m confident that banks are using AI-based tools to identify fraud in credit card transactions,” he says. “You could deploy that same AI to identify claims fraud if you had access to enough data.”

AI advocates in the blogging sphere predict chatbots will evolve into “robo-advisors.” Risk Genius, which analyzes language in claims policy documents, and PolicyGenius, which compares insurance quotes, are cited as current examples of AI moving in this direction.

In the broker space, PwC predicts robo-advisors will soon offer more “recommender” systems and “someone like you” statistical matching. When fully mature, robo-advisors will understand individual and household balance sheets and income statements, PwC predicts. They will be able to recommend, monitor and change financial goals and portfolios for their p&c policyholders.

This begs the question of how AI will affect the workplace demographics of the Canadian p&c industry in the future.

Will machines replace humans?
There is a wide spectrum of opinion on whether this is possible, much less advisable.

Brenda Rose, vice president and partner at FCA Insurance Brokers, wonders whether machines could ever evaluate or make judgments like humans do. To illustrate the point, she cites the development of autonomous, self-driving cars.

Three years ago, MIT observed that autonomous vehicle technology could not easily detect “humans that invariably pop up alongside (or in the middle of) the road.” Rose questions whether machines could make trustworthy decisions if they confronted ethical choices that were not part of their programming or learning process. What would a machine do, for example, if it had to choose between swerving into a child or a traffic cop if it wanted to avoid a dog that suddenly appeared in its path?

Brokers must make value judgments and recommendations that take them beyond information processing and making predictions, Rose points out. In the example of Hal above, an AI attempting to replicate broker services must do more than identify a policy to serve Dave’s purposes, Rose says. “It would also need to identify options, reviewing with him their various merits, some of which might relate to his personal preferences or comfort levels, and continue that discussion until Dave was satisfied.”

Hypothetically, McDougall believes machines could learn to make such human-like evaluations and judgments. “The question is, what are humans teaching machines to learn? To offer the best coverage to make a sale? Or to offer the best coverage to meet the client’s needs?”

In the near term, McDougall and Black think humans and machines are much more likely to work in tandem to improve insurance business processes.

Several commentators see machines automating and streamlining less complex tasks, leaving more time for human insurance professionals to work on sophisticated tasks. That could mean underwriting more complex risks, selling more customized coverages to clients, or adjusting multi-faceted claims. “AI will take over roles,” says Purdy. “We’re seeing it happen right now.”

But McDougall sees AI more as a tool for human beings to improve insurance processes. “AI is not like people walking around in machine bodies,” he says. “These technologies are tools used to improve a predictive action.”

In the future, McDougall predicts p&c professionals will likely apply available AI technologies to accomplish specifically defined insurance goals or tasks.

AI technologies can be applied to solve insurance problems even if they are not designed with a specific insurance purpose in mind, McDougall says. He cites the example of Alpha Go, a machine-learning technology used to play the strategy game Go.

“No one programmed the computer to play Go,” he observes. “The tool was built to learn, and then it developed strategies and tactics [for Go] that were not part of its initial programming. Humans applied this AI tool to gain insights in how to play the game.”

Implications for the Insurance Industry in Canada

A Changing Workforce

A new 2017-2018 demographic research study is currently underway. Stay tuned! We anticipate the new research to provide a wealth of valuable information that will help employers address challenges and implement initiatives to meet current and future talent needs.