Actuarial science, and insurance more generally, is a crucible for this concern. At a 2017 Insurtech Connect Conference in Las Vegas, Chief Executive Officer Daniel Schreiber of the startup Lemonade suggested that AI will supplant the traditional actuarial role, declaring that “the next insurance leaders will “use bots, not brokers, and AI, not actuaries.” Increasingly actuaries have been incorporating machine learning techniques into mathematic models.
Gort. The Terminator. Wall-E. HAL 9000 … and AutoML?
Popular images of artificial intelligence (AI) can terrify, inspire, and amuse us – but the real thing, AutoML (automated machine learning or AML), often just confounds us. AML doesn’t walk. It doesn’t talk (necessarily). But this form of software is so smart that it can teach itself to code, and it could enable insurers to automate tasks in fields as diverse as fraud investigation, valuation, and lapse detection. For some, this outcome sounds terrifying. For others, it’s exciting. But for many, it simply confuses.
Many past attempts to apply AI have foundered against the natural limits of data availability, storage capacity, and computing power, but those barriers are now falling. As advances in chip design have delivered exponential increases in processing power and information storage, the results are visible across multiple branches of this discipline from rapid advances in the most basic forms of AI voice, speech, and object recognition to more advanced forms of machine learning.
AutoML sits near the top of this technological food chain. Several years ago, Google’s AI researchers invented AI software to generate more AI software – to effectively regenerate itself independently. To its creators’ surprise, AutoML was not only capable of creating neural network architectures without human input, but the software actually created models that were more powerful and efficient than those designed by humans. The researchers coded themselves out of a job.
Naturally, the idea that a computer could be capable of such feats of ingenuity and cognition stirs deep fears – particularly among knowledge workers. In fact, if one types “will automation” into the Google search bar, the first four AI-generated, auto-complete sentences are all job-related: “will automation kill jobs”, “will automation create more jobs?”, “will automation replace jobs?”, and “will automation take Our jobs.”
Actuarial science, and insurance more generally, is a crucible for this concern. At a 2017 Insurtech Connect Conference in Las Vegas, Chief Executive Officer Daniel Schreiber of the startup Lemonade suggested that AI will supplant the traditional actuarial role, declaring that “the next insurance leaders will “use bots, not brokers, and AI, not actuaries.” Increasingly actuaries have been incorporating machine learning techniques into mathematic models. Will the advanced is machine learning eventually make Actuaries redundant as Daniel Schreiber suggests?
We don’t think so. While AML is a tool that can augment or assist, it cannot fully replace human reasoning. The most likely outcome is that it will fuel a transformation in the roles and responsibilities assigned to actuaries, automating more of the routine actuarial tasks, only to create new opportunities for actuaries to develop.
Automation is not binary – it’s a continuum, a complex combination of human reasoning and machine processing. For example, carmakers have been on a journey to achieve self-driving cars since the invention of anti-lock brakes, but few of us today would be willing to place a child alone inside an autonomous vehicle. That’s because complete automaton of complex tasks is enormously difficult, and even after years of research and development, self-driving cars that can be 85% automated still require a human driver to take over the wheel in an emergency. Indeed, achieving 100% automation could take some time.
Similarly, AML requires an actuary or data scientist to drive the modeling process, but the AI can automate many analytics tasks across a vast array of information sources. And there has never been more data for insurers to work with. Actuaries are swimming in a sea of alternative data opportunities, from social and streaming media to wearable devices on our wrists to an array of Internet-connected devices in our homes (the Internet of Things). All of these connected systems have the potential to provide insight into risk in virtual real-time, but only if the data can be analyzed efficiently. AML automates several of the most challenging steps in the data pipeline, enabling insurers to accelerate the use of alternative data sources for risk assessment, and it opens up the potential for new actuarial rating factors and more finely tailored, bespoke product designs.
AML makes sense for carriers seeking to supercharge the modeling process and free actuaries from mundane tasks. It also makes sense for smaller carriers with limited access to data science talent. And yet, it is critical to approach AI with caution. In insurance, it is not uncommon to face sparse or gap-filled data sets, such as information on drug prescriptions or on younger policyholders. Faced with these gaps, AML interpolates or extrapolates mathematically – and at times incorrectly. Under these circumstances, significant actuarial input and assumption setting are still required.
Insurers also need to bear in mind the heavily regulated nature of the industry. While the production of data is growing, so too, is concern over data privacy and regulatory protections provided to consumers. Most notably, the European Union’s General Data Protection Regulation (GDPR) provides consumers with the right to opt-out of data sharing, and the willingness of the insured to take a leap of faith and provide this information will directly influence the amount of data available for risk analysis.
It is important to continually prove the relevance of data to risk prediction and engage in a broader societal debate on the value and future of insurance. This will require carriers to address concerns about purely correlational relationships or the potential for certain proxy variables to unintentionally be used to make rating decisions based on protected or socially unacceptable factors such as ethnicity, sexual orientation, or educational attainment. AutoML may be smart, but it is also utterly naïve. It can use any variable provided to achieve predictive conclusions, and it will be up to humans to manage its use ethically. The industry must constantly ask itself, “Just because we can, does that mean we should?”
While no one can precisely predict the future, actuaries come close, and it seems clear that the profession will need to evolve to co-exist alongside AutoML and other forms of AI. The actuarial syllabus seems likely to adopt more data science training and techniques in the years ahead. Similarly, demand for purely technical actuarial functions seems destined to decline as automation spreads, just as emphasis on creative and relationship-building skills seems likely to grow. These trends are not limited to actuaries, or insurers. A recent Oxford University study found any professional who can demonstrate these three skills in his or her role will be able to better manage the upcoming wave of automation:
The actuarial role will remain at the beating heart of the insurance business, but to remain relevant, the profession and the industry will need to learn, adapt and evolve. As the great physicist William Pollard once said, “The arrogance of success is to think that what you did yesterday will be sufficient for tomorrow.” For those who are willing to use machine learning to the advantage of the profession and the end consumer, great opportunities await.
