The End of Insurance as We Know It

by Rob Galbraith

  An area where unsupervised AI can be applied to existing data infrastructure that is already tuned for analysis purposes is in data mining. This data mining approach can lead to four possible outcomes:

  1.Identification of trends that were already known to business experts such as claims adjusters, underwriters, etc.

  2.Identification of non-meaningful patterns that show correlations but do not have power to explain causality

  3.Identification of dynamics deep within the business that are important and non-trivial which were not apparent through traditional directly analytical approaches

  4.Invalidating commonly-held business assumptions or conventional wisdom

  The third and fourth outcomes are the ones that add tremendous value for both entities within the insurance ecosystem as well as offer business opportunities for those seeking to disrupt the industry.

  A second major benefit of AI is the ability to make use of data that is commonly acquired but is not being fully utilized today. Textual descriptions, voice recordings, images and videos are all examples of non-fielded or unstructured data that is rich with value and insight if used properly. Too often what is captured is simply a few dozen data points in fielded data for analysts to understand the cause of the claim and for it to be investigated and resolved. The ability to use AI in combination with technologies such as Optical Character Recognition (OCR), Natural Language Processing (NLP), image pixels and geolocation from still and video images is quite powerful. These technologies when brought together as AI can show trends and patterns that may exist in a way that humans would not be able to easily decipher on their own. Potential use cases include fraud detection, faster claims resolution, underwriting for a problematic cause of loss, greater insights on loss trends and even the potential to pursue better loss control and prevention strategies. Some initial examples of the use of AI in these areas are performing an initial estimate of auto claims based on pictures of the damage and triaging renters claims to speed up payment rather than routing each one to a human adjuster.[104]

  What are some potential sources of unstructured data in insurance? Here are a few examples:

  •recorded conversations with customer service representatives, claims adjusters, etc.

  •notes taken by agents, underwriters, claims handlers, billing specialists, etc.

  •photos of exposures and damages from a loss event

  •inspection reports

  •chat sessions

  •videos

  •social media posts

  •any other non-fielded data that could be used in an insurance context

  To the extent that these forms of data have previously been captured and stored, their use was generally limited to an individual review. Phone calls could be retrieved and listened to again for quality purposes, images could be reviewed by claims adjusters, agent notes could be reviewed by underwriters, etc. What was previously unattainable was to perform the same level of summarizing, aggregation, trending and in-depth analysis using this data because the tools to do so simply did not exist. AI makes full analysis of these types of data sets at an aggregated level possible. For example, a machine learning algorithm can be trained to identify all homes that have a pool in the backyard.[105] Additionally, an algorithm could be trained to determine whether each pool has a childproof fence around the pool, which could make a difference in the rating and underwriting for a policy due to the liability concerns of a child unknowingly getting into a pool and drowning. These possibilities were dreamed of previously but were impossible to execute until now.

  A third benefit of AI is making sense of the explosion of sensor data from telematics, smart home and other IoT sensors. The value of these devices is that they allow insurance carriers to directly observe behaviors and “see” losses in a way that previously had to be indirectly observed through correlations with other variables that could be observed.[106] These sensors are a constant presence capturing and recording data, 99.9999% of which is likely of limited value or relevance on its own except to distinguish the 0.000001% or so of data that is meaningful because it is the loss event or a “near miss”. These near miss scenarios are extremely valuable data points for insurers because they provide insights into losses that could have occurred but were remediated through some action - or damage that was caused but not claimed (yet).[107]

  A fourth benefit that AI can unlock in the insurance ecosystem is the ability to learn from itself through unsupervised learning. As mentioned before, much of analysis in the insurance space in the past was deterministic and supervised by business experts. This has created some impressive outcomes and advances in the industry, but also leads to a lot of groupthink and conventional wisdom that is hard to overcome. Deploying AI to continually train, test and learn using a host of new algorithms outside the traditional suite of linear regression approaches that have dominated in insurance is bound to uncover previously hidden patterns. By implementing unsupervised AI to find patterns without prejudging what the outcomes “should be”, both traditional incumbents and insurtech startups alike have the ability to gain new insights that competitors do not have and to seize upon these new insights to their competitive advantage. The possibilities are similar to those found in technologies such as Google Translate where the machine is not “taught” by experts but rather uses the distinct advantages that machines possess to test millions of potential combinations or outcomes in seconds and recognize the most relevant patterns.[108]

  This ability of AI to learn and improve over time as more data is fed into it opens up a wide range of new potential use cases including streamlining processes and improving service quality. One popular use case receiving a lot of attention is in the world of chatbots. Chatbots are an AI technology that can mimic a human to consumer interaction as well as between two or more parties handling back office work. According to VentureBeat, some of the benefits of using chatbots in insurance include:

  •Reduced customer confusion

  •24/7 availability

  •Streamlining of tedious processes[109]

  The biggest criticism of chatbots today is that they lack human empathy and can be easily exploited, but developers will undoubtedly continue to make improvements. For an industry that 72% of consumers say uses jargon that is too confusing, the opportunity for chatbots to improve service quality is real.

  The ability of AI to replace humans is a familiar topic in the media recently, and insurance is no exception. From the ability to:

  •capture information from a myriad of data sources to

  •identifying hidden patterns in data to

  •making decisions based on data and learn over time to

  •24/7/365 availability

  the potential disruption to the human workforce in insurance is quite large. AI technologist Francisco Corea says that incumbents in the insurance sector “should be ready to engage intelligently with new types of data and adapting their models and infrastructures to fully embrace the potential of AI”.[110]

  THE REVOLUTION WILL NOT BE TELEVISED - IT WILL BE STREAMED

  AI holds enormous potential to disrupt the insurance ecosystem by leveraging existing data in new ways and understanding new data being generated by a proliferation of cheap sensors in our world. The streaming amount of data generated by these sensors not only hold promise for improving upon the existing insurance business model - which is highly reactive to losses after they occur - to one that is more proactive and detects losses before they occur. The potential value for customers of preventing claims before they occur is a game changer. The benefits from consumers and businesses appears compelling, both in terms of direct financial losses avoided as well as large indirect benefits such as time saved and disruption avoided.

  Loss avoidance has been challenging to quantify in these early stages, but that is not stopping a full scale deployment of sensors. According to Alex Sun, President and CEO of Mitchell International, new vehicles typically have 60 to 100 sensors in th
em and that number is expected to grow up to 200 in the next few years. It is estimated that there will be 22 billion sensors embedded in vehicles by the year 2020[111], all part of what is known as the advanced driver assistance systems (ADAS). Not only to these sensors provide data on how vehicle systems are performing and the driver’s behavior, but they also enable capabilities such as forward collision detection, lane departure warnings, adaptive cruise control and other features. Grand View Research estimates that the global market for ADAS is rapidly growing at a CAGR of 19% and expected to reach $67 billion by 2025.[112]

  The use of sensors continually streaming data combined with AI to directly monitor the behavior of systems is a fundamentally different paradigm that will challenge the traditional insurance ecosystem. If enough losses are prevented (lower frequency) and those that occur cause less damage (lower severity), the premiums that carriers can justify charging must be reduced as well. While this is a boon for insureds, the loss of revenue can negatively impact not just the carriers themselves but agents, brokers, contractors and other third parties whose success is directly dependent on commissions or other income that flows from insurance carriers.

  The journey to implement AI solutions is also fraught with peril and potential missteps. The technology is still relatively new and untested for most insurance applications. In addition, finding reputable companies and talent to execute on an AI strategy is exceedingly challenging. At InsureTech Connect 2018 in Las Vegas, there were over 6,000 attendees in just the third year of the conference with 180 insurtech vendors, up from 83 in 2016 and 112 in 2017. Many of them new startups have formed in the last 3-5 years,[113] and a large fraction of these startups offer AI-based solutions for a host of pain points that traditional incumbents have. Are these real solutions or merely vaporware? How easy are these solutions to integrate and deploy at scale? Which use cases should be prioritized? The stakes in terms of risk and reward for pursuing an AI strategy have never been higher for both incumbents and startups alike.

  1

  CHAPTER 16 - STOP AND GO: THE CASE FOR AND AGAINST TELEMATICS

  TELEMATICS: A SNAPSHOT

  Telematics is the insurtech of the 2000s: a powerful new technology that holds great promise but also presents challenges. Since auto manufacturers (commonly referred to as original equipment manufacturers or OEMs) began to include event data recorders (EDR) or “black boxes” in all manufactured vehicles starting in the 1990s,[114] the ability to revolutionize the insurance industry has been technically feasible. For the first time, it was possible for insurance carriers to gather data directly on the driving habits of each individual. Prior to this point, insurers have had to rely on proxy variables such as age, gender, marital status, and miles driven (among many) that have strong statistical correlation with losses. Technical hurdles remained however: each black box was originally custom to the auto manufacturer and required expensive software to read the data. The amount of data that was captured originally was also quite limited as compared with the amount that is captured today by EDRs in conjunction with the proliferation of sensors on vehicles today.[115]

  For a myriad of reasons, telematics has not revolutionized the auto insurance industry in the way that many predicted 10-15 years ago. Telling the full story of why could be a book by itself. The short version is that significant hurdles have delayed adoption: from the technical knowledge to access, store and interpret the data; to the analytical ability to change pricing and underwriting based on this data; to creating a value proposition that is compelling for consumers to share their data.

  Each insurer evaluates and makes decisions on one or more telematics vendors to partner with. This meant historically each insurer had to negotiate the challenge of distributing the special devices used to read the EDR data, known as a dongle, to insureds and incentivize them to install them in the vehicle. These incentives usually took the form of discounts to premium, although some focused on the training and education benefit for drivers, especially parents of teenagers who wanted more awareness of their child’s driving behaviors when they were not riding along.

  Bottom line: Developing a telematics program for insurers has historically has been an expensive and challenging proposition:

  •to acquire the dongles and distribute them

  •to collect the data, store it, format it and analyze it

  •to develop new rating and underwriting algorithms

  •to persuade reluctant consumers to install a dongle and share their driving data

  •on top of it all, the loss of premium in the form of discounts for insureds who choose to participate in the insurer’s telematics program

  BACK TO THE FUTURE

  Given the mixed success (at best) of telematics, why is it still a hot technology in insurtech? A November 2018 report titled “The Societal Benefits of Telematics” from LexisNexis Risk highlights the promise of telematics to reduce road casualties based on data from the United Kingdom,[116] where telematics insurance adoption has reached a critical mass among young drivers. According to the report, the UK has experienced a large rise in telematics policies from 100,000 in 2011 to 975,000 in 2017. Over 4 in 5 young drivers are now covered by a telematics policy in the UK, and the resulting impact is a 35% decrease in road casualties for young drivers ages 17-19 over this time period compared to 16% for the overall population of drivers.[117] Government estimates are that collisions cost the UK economy an estimated £16.3 billion ($20.9 billion) annually and the average claim for younger drivers ages 18-25 is double that for older drivers 51-70. In 2017, a total of 1,793 people will killed on British roadways, an average of 5 per day.[118]

  Back in the United States, LexisNexis reports that the majority of auto insurance carriers are actively pursuing telematics programs or expanding them.[119] In addition to traditional players, new startups such as Metromile looking to compete in the auto insurance space are providing product offerings known as usage-based insurance (UBI).[120] UBI efforts rely on telematics to collect the information needed to go beyond the standard time-bound policy term. What do all of these auto insurers, large and small, new and old, see that justifies continuing down this path?

  Advances in telematics technology have gone a long way to decrease barriers to adoption and reduce the overall cost of pursuing such a telematics strategy. In particular, the ability to go beyond a dongle and instead rely on the ubiquity of smartphones to read and record the vehicle statistics is a game changer. The use of smartphones also enables more consumer-friendly services similar to OnStar. One example is sensing when a vehicle has been in an accident and calling the occupants to check if they are injured. By providing value-added services (VAS) that can be enable through the combination of telematics and smart phones, insurers can overcome some of the resistance to consumer adoption that were present in earlier efforts.

  An increase in consumer receptivity to having telematics used in their vehicles unlocks a powerful new capability: gamification. Gamification has a lot of potential for influencing drivers’ behaviors and can add immediate value to insurers adopting a telematics strategy without the need to wait until a brand-new rating algorithm and underwriting program is in place. By providing feedback on driving habits to consumers on their smartphone and “scoring” how well or poorly they drove in relation to other drivers, telematics can leverage social psychology to unlock the competitive juices for some drivers and encourage them to improve their driving habits to earn a better score. In doing so at scale, insurers may be able to do the unthinkable: shape the loss performance of their auto insurance book en masse.

  The provision of a driving score along with other diagnostic statistics can, in and of itself, provide an incentive for some drivers to be safer on the roadway. For another subset of drivers, the provision of a score may not prove enough incentive by itself, but if tied to potential tangible rewards and incentives, it could persuade them to drive safer. An example of this is the San Antonio’s Safest Driver contest held over the summer
of 2018 which provided cash prizes exceeding $60,000 to winners.[121] The ability to change driving behavior with the goal of reducing accidents and losses without the need to radically alter pricing algorithms or underwriting practices through telematics and gamification is a game changer - pun intended.

  SHOULD I STAY OR SHOULD I GO?

  With all of these technological advances and reduction of barriers to adoption, are carriers right to invest heavily in telematics? The roughly two decades of the telematics era have brought mixed results to date and the technology overall have not proven to be as revolutionary to the auto insurance industry as originally anticipated by evangelists - yet. On the other hand, the number of carriers actively pursuing telematics strategies remains high - arguably the highest level of interest since the technology came into being. In addition, more companies are seeking to make telematics an integral part of their product offering and even core business strategy, rather than a side project. So should carriers aggressively pursue a strategy to adopt telematics? Reasonable debate exists, but I would advise caution and restraint for a number of reasons.

  A major hurdle that has been consistently underestimated is the time and expense in transitioning to a fundamentally new way to price and underwrite risk. To fully leverage telematics, carriers have to upend the traditional approach of using many “proxy” variables such as age, gender, marital status. Insurers need to start capturing, storing, processing and analyzing raw sensor data on excessive speeds, hard braking events and other driver behaviors. This data is vastly different in the size, type and patterns from traditional fielded, relational data. Storing the data for starters is already a huge challenge - it cannot be stored in standard relational databases due to the sheer volume of data. Once you solve the storage issue, how do you analyze this much data? There is an enormous amount of data and little of it is of genuine interest. As an insurer, you really only care about the moments before an accident (or near accident). The remainder of the driving data is important context - but just that, nothing more. The difficulty in converting from the traditional approach to a telematics-based one should in no way be minimized. New pricing algorithms must be developed that radically depart from the traditional approach. The same goes for tiering and underwriting - these must be developed from analysis and then filed and approved by state regulators.