Artificial intelligence is the new electricity. We hear it will fundamentally shift the balance of power between labor and capital, mostly by rendering labor obsolete. It will enable and empower transformative technologies that will rearrange the sociopolitical landscape and may lead to humanity’s transcendence (or extinction) within our lifetimes. As it changes the world, it will necessarily rewrite the rules of insurance. That’s the myth, and the nature of the headlines. Interestingly, insurance is heavy on intellectual property (think of proprietary underwriting models), technology and data. And AI is hungry; hungry for data, of course, but also hungry for systems that can be automated and for proprietary classification problems that can be improved. That places insurance right in the appetite of artificial intelligence and its promise of transformation. If we want to act on artificial intelligence’s transformational potential, we need to understand what it actually is, separate the technologies from the hype and develop a practical understanding of what is required to implement AI-powered solutions in the insurance sector. This article will highlight these three steps and offers a realistic approach for carriers to take advantage of the opportunities. Defining Artificial Intelligence Unfortunately, our first step is also our hardest, as a working definition of artificial intelligence is difficult. The scope of the term AI is broad, and it requires careful consideration to avoid becoming hopelessly confounded with its own hype. It is also challenging to come to a clear definition of natural intelligence, which leaves us struggling for a definition of artificial intelligence because the latter is so often compared to the former. AI tends to be discussed in two flavors. The first is general artificial intelligence (also, artificial general intelligence and strong AI). GAI is machinery capable of human-level cognition, including a general problem-solving capability that is potentially self-directed and broadly applicable to many kinds of problems. GAI references are accessible through fictional works, such as C-3PO in Star Wars or Disney’s eponymous WALL-E. The most important feature of GAI is that it does not currently exist, and there is deep debate about its potential to ever exist. The second is usually referred to as narrow AI. Narrow AI is task-specific and non-generalizable. Examples include facial recognition on Apple’s iPhone X and speech-to-text transliteration by Amazon’s Alexa. Narrow AI looks and feels a lot like software or, perhaps, predictive models. Narrow AI can be described as a class of modeling techniques that fall under the category of machine learning. See also: Seriously? Artificial Intelligence? What is machine learning? Imagine a set of input data; this data has one or more potential features of interest. Machine learning is a technique for mapping the features of input data to a useful output. It is characterized by statistical inference, as advanced techniques often underlie machine learning predictive models. Through statistical modeling, software can infer a likely output given a set of input features. The predictive accuracy of machine learning methods increase as their training data sets increase in size. As the machine ingests more data, it is said to learn from that data. Hence, machine learning. Perhaps most important of all, machine learning (as an implementation of narrow AI) is real and here today; for the remainder of our discussion when we say "AI," we mean narrow AI or machine learning. Beyond the Hype The hype around AI and its potential is extensive. Silicon Valley billionaires opine on the potential implications of the technology, including comparing its power to nuclear weapons. Articles endlessly debate if and how quickly AI will structurally unemploy vast swaths of white collar workers. MIT’s Technology Review provides a nice summary of the literature, stating that up to half of all jobs worldwide could be eliminated in the next few decades. AI may well have this kind of impact. And the social, political and economic implications of that impact, especially around questions of potential large-scale unemployment, deserve careful long-term consideration. However, executives and business owners need to evaluate technology investments today to improve their current competitive position. From that perspective, we find it more practical to focus on examining which existing tasks could be automated by AI today. Enter Pigeons In 2012, researchers trained pigeons to recognize people based only on their faces as part of a study on cognition. Suppose you had millions of face-recognizing pigeons; this force of labor could be deployed in a comprehensive facial recognition system -- a system remarkably similar in function to the facial recognition AI of devices like modern smart phones. It turns out pigeons have also been trained to recognize voices, spot cancers on X-rays and count, among a host of other tasks related to headline-grabbing AI achievements. The metaphor is admittedly silly. Instead of pigeons, imagine an army of virtual robots capable of classifying information from the real world to produce a machine-readable data set. In machine learning language, these robots take unstructured data and make it structured. Said robots resemble the automation machinery of a factory; like spot welders tirelessly joining steel members to form automobile frames, our virtual robots tirelessly recognize if a face is featured in a photograph. In contemplating the question, what could be automated with AI, a useful starting place is the army of robots (or pigeons!). For example:
- What existing analyses could be improved or optimized? Could pricing or underwriting be improved using better classifiers or non-linear modeling approaches?
- What data currently exist at the firm that could be made available for new types of analysis? Claims adjusters’ notes can be processed by natural language algorithms and cross-referenced with photos of physical damage or prior inspections.
- What data would you analyze if it could be made available? What if you could listen to all the policyholder calls received by your customer service department and annotate which questions stumped the customer service representatives? Or which responses lead to irritation in the policyholders’ voices?