This paper is the first in a series that will explore strategic areas where life insurers may deploy artificial intelligence (AI): in underwriting, customer experience, risk assessment, fraud detection, and operational efficiency. We will also publish an installment dedicated to regulation and compliance considerations in the use of AI.

Introduction

Life insurers play an essential role in protecting the financial well-being of people by providing products that mitigate financial risks and aid in retirement planning. Traditionally, insurance companies have been “second movers,” slower in adopting technology due to their long-term focus and desire to clearly understand the risks and rewards of change. New capital investments in technology must meet internal benchmarks for opportunity cost and overcome the high barriers to adopting unproven innovations. Many insurers interact with customers through third parties or digital platforms, providing another barrier to decisiveness. This cautious approach has led some to view the life insurance industry as ripe for disruption, especially with the rise of fintech and insurtech in the late 2000s. There have been numerous historical examples of insurers whose early adoption of transformative technologies such as advances in computing (MetLife, Franklin Life, and PacLife’s UNIVAC mainframes, 1954), the internet (Progressive’s auto-insurance.com, 1995), and development of derivative financial instruments (Keyport Life Insurance Company’s KeyIndex FIA, 1995) gave them a competitive edge.

Artificial intelligence (AI), and in particular generative AI, is just such a new technology that will drive growth in embedded value for those companies that can successfully harness it.

This paper is the first in a series that will explore strategic areas where life insurers may deploy AI: in underwriting, customer experience, risk assessment, fraud detection, and operational efficiency. We will also publish an installment dedicated to regulation and compliance considerations in the use of AI.

The term artificial intelligence is a catch-all that encompasses multiple other terms, including both “machine learning” and “generative AI.” While all the flavors of artificial intelligence are interesting, we believe the value of machine learning to insurers is already well-understood. The rest of this paper (unless we specify otherwise) will focus on how insurers will create value from generative AI, such as large language models (LLMs) and similar models and capabilities. For simplicity, we refer to this as ”AI” throughout.

Generative AI capabilities for insurers

The following features and uses of generative AI are key to insurers:

• Generalist capabilities: AI has broad capabilities and potential reasoning abilities across a diverse array of tasks, with the ability to become further specialized in particular domains or task areas through additional fine-tuning. This flexibility allows insurers to deploy AI across multiple functions, reducing the need for multiple specialized systems.
• Human-computer interface: AI provides a new way to interact with computers. Spoken or written colloquial language can ”prompt” a unique response from a machine. This more stochastic approach is a departure from the traditional interface humans have with computers: entering commands to receive outputs determined through deterministic algorithms. These more intuitive and natural interactions can lead to improved experience, efficiency, and satisfaction.
• Multimodal content generation: AI can produce substantial written, programmatic, pictographic, video, and audio responses based on simple requests. This core capability drives higher productivity for knowledge workers and can allow insurers to scale this generation with fewer resources.
• Personalization and customer experience: AI can correlate general information (from its training data) with specific information from end-user input and customer data to create personalized writing and responses. This stands to elevate the customer experience when compared with deterministic chatbots or static knowledge bases, as people typically respond better to more personal interactions.
• Automation and improving efficiency: AI can automate routine and repetitive tasks, which reduces the scope for human error and increases efficiency. Even in instances where the accuracy is not high enough to completely entrust to AI, it can expedite tasks by performing a first pass that a human can improve upon.
• Facilitating innovation and creative thinking: AI can accelerate innovation in companies by assisting with scenario analysis, suggesting solutions to complex problems, and providing predictions and challenges for the feasibility of new ventures.

Optimizing and expanding AI in insurance

We examine the role that AI will have in the insurance space across the following dimensions:

Optimizing and improving current capabilities

Generative AI stands to provide widespread efficiencies across current operations. The mechanisms for these improvements rely on taking tasks that are currently done by people (usually working with computers) and inserting generative AI into the workflow. One example of this is using generative AI to summarize existing documents, or to take notes on a call. Microsoft has incorporated this feature (”Copilot”) into the Office Suite already. Insurers can leverage these capabilities by creating generative AI chatbots (for example, through a “CompanyGPT” interface) and search assistants that reference company document repositories to support querying of complex documents such as contracts, policy forms, and product specifications. Envision a new company employee in the marketing division asking an AI bot detailed questions about a specific annuity policy. While this may not fully replace detailed reading of the source documents and conversations with knowledgeable colleagues, using an AI chatbot can reduce the amount of human time spent searching for the right document or chunk of text embedded within a document.

By improving the efficiency of current capabilities, AI assistants will lower the cost for activities such as searching for information or providing additional teaching or training to employees. As the costs for these activities drop, we expect to see insurers develop employees who are functionally more knowledgeable because they have access to more information through efficient search assistants.

Expanding abilities available from generative AI

We recognize that generative AI will also expand the opportunities available to insurers. The mechanism of this expansion is using new AI capabilities that did not exist before LLMs and generative AI: machines producing content, collating ideas and concepts with less time and effort than people do. AI will serve as an assistant to the marketer (Holloway, 2024), to produce copy and images in the company’s own style, at first as a way for marketers to deliver their visual ideas to professional artists, but eventually producing content that is ready for publication.

Specific to insurer operations, generative AI may help the insurer fraud division detect signs of fraud activities in real-time. Currently, call center employees need to balance providing good customer service while also watching for red flags but, with an AI assistant, these duties could be bifurcated between the human and the machine. For example, an AI assistant could listen to voice calls between customers and administrators and be on the lookout for social engineering tactics, unusual requests, or suspicious communication, while the person on the line focuses on the customer’s needs. With access to a repository of call transcripts, the AI could also detect repeated scripts used in organized fraud attempts or similar schemes. The AI could keep the service representative in the loop regarding any potential fraud issues and suggest countermeasures, including alerting an organization’s fraud specialists for intervention. Of course, fraudsters already employ generative AI in creative ways, such as voice cloning (Bethea, 2024), in this “arms race.”

Upcoming articles in this series will examine the optimization and expansion of insurer capabilities across key areas that we believe are ready for disruption: underwriting, customer experience, risk assessment, fraud detection, and operational efficiency.

Challenges to AI adoption in insurance

Insurers must be strategic in deploying AI in their organizations: adopting these technologies comes at a financial and resource cost. Despite that, we believe that failing to adopt AI in various ways produces an opportunity cost that will eventually put most insurers at a strategic and competitive disadvantage, as widespread adoption of AI will lead to—for example—better customer service, higher worker productivity, and lower fraud.

Envision learning of the first spreadsheet technology (say in the late ‘70s or early ‘80s) and believing: this will put company actuaries and accountants out of a job, it will cost a ton of money, and the company puts itself at risk by relying too much on machines to do the work we have done successfully as humans thus far.

Business—and insurers—are at the next major technology inflection point, and (despite what actuaries say) it’s bigger than the spreadsheet.

Insurers must devise a clear strategy for AI, considering opportunities that make sense for their core business, and consider taking calculated risks on others.

Building and buying

Innovating at the cutting edge of any new technology is a high-risk, high-reward gambit. Early innovators that succeed can reap outsized returns and durable advantages. However, the unpredictable nature of technological innovation also means that many early bets will be misplaced and thus result in wasted resources.

The prospect of building new AI tools and processes within an insurance company will be tempting. One approach is to build the new technologies in-house: to hire new experts and integrate new workflows. After all, if there are major new opportunities available, then why not invest and earn the returns of that investment?

Another strategy is to be a more passive fast follower. Let other companies take the early bets and risk their capital and only be reactive, seeing which bets pay off and only implementing those proven ideas, but necessarily with a time lag. With this approach there is less overall risk of overinvesting in unproven technology, but it will reduce the value of any competitive advantage and, indeed, if the first movers are truly successful, it could put your company at a significant disadvantage playing catch-up.

Another uncertainty looming over both options is that it is still unclear to what degree in-house customized solutions will be more useful than off-the-shelf third-party tools, if at all. It is possible that the products of generalist providers such as ChatGPT will suffice, or that successful specialist vendors will emerge in the insurance space. Going back to the spreadsheet example, no insurance companies would have benefited by creating their own customized spreadsheet software, as major software developers such as Apple, IBM, and Microsoft offered superior, general solutions. Moreover, few companies were significantly harmed by adopting spreadsheets at a slower pace than their competitors. Yet while both those things are true, any insurer that resisted the technology entirely would have found itself eventually defunct. Similarly, as major business technology players continue to pour billions of dollars into developing AI solutions, keeping an in-house solution current with external developments may prove to be a Sisyphean task.

The decision will likely come down to company culture. Those companies for which such disruption is more welcome, and that are willing to test ”tried-and-true” methods against new models, will find themselves faster adopters of this new technology. They will be the companies most willing to hire external guides for their AI build, and to hire a new team themselves. If the company or CEO has been rewarded in the past for quick adoption of new technology, they will be ready for this next paradigm shift.

Regulation and compliance

In addition to our focus on key strategic areas, this Milliman series will also dedicate an installment to compliance and regulatory considerations. The National Association of Insurance Commissioners (NAIC) has issued the 2023 model bulletin “The Use of Artificial Intelligence Systems in Insurance” (NAIC, 2023). Insurers, primarily through their legal and compliance teams, have adopted internal policies on the use of AI at work, and have begun to articulate their positions regarding the use of AI for business operations. This area of practice is constantly evolving with the new tools and techniques that generative AI brings us and it will continue to evolve as the cost of new intelligence continues to decline.

Overreliance on AI

Even the most techno-optimistic insurance companies will realize that placing too much reliance on AI presents new risks:

• Relying too heavily on machines for tasks traditionally done by people will atrophy our human connections to the business. The more we cede tasks and decisions to models, machines, and electronic agents, the less we are in control of the core operations of the business, and we risk losing culture and the raison d’etre of the venture. Employees might not take kindly to the prospect of AI-generated performance reviews, for example.
• There is value to the institutional memory that we generate through our daily interactions, tied together over decades to produce quality products and experiences for our customers. Allowing machines to do more tasks may bring efficiencies to our customers—through better experiences and lower prices—but this could come at the cost of not developing sufficient human intelligence about the customer experience.

In the end, the balance between reliance on people and machines will be a company decision, dependent on culture and economics and leadership, as well as external factors such as competition and larger business trends.

Reliability

Knowing when and how to rely on the output of AI will become a harder task as AI becomes more prevalent. AI in its current form is prone to hallucination. However, though often less discussed in the same terms, humans also are prone to such “hallucinations”—our own human frailty can manifest in diverse ways such as misunderstandings and overconfidence.

Underwriters will try to override decisions where the AI suggests accepting profitable business. Claims might get flagged as fraudulent when the model is noticing rare but unconcerning behavior.

The future of work will involve the complementary hybridization of work between people and AI. The employees best able to navigate this future will have the openness to let the technology automate things when it can, and then further use it augment their judgment when it can’t. We must retain the healthy skepticism to understand that, just because AI determined an answer, it is not automatically correct.

Work force transformation

Any adoption of AI to perform tasks currently done by people will transform the way a company thinks about doing work. To start with, we envision AI completing or assisting with certain tasks done by managers in much less time. These tasks include rudimentary work, such as taking notes and writing memos, but also include more cerebral tasks such as assigning work (or suggesting assignments), brainstorming ideas on new ventures, and outlining project workflows.

This requires an ”upskilling” for current employees to adopt AI in their own roles—in a compliant manner—to improve their own efficiency. A large adoption of AI will no doubt lead to additional business gains: with extra time, skilled and thoughtful employees have more opportunity to contribute to other business ventures.

Most roles will be transformed not only by ”upskilling” to incorporate AI into their current processes, but also by changing the nature of those roles altogether. We envision actuarial and underwriting roles, for instance, being rewritten over time to incorporate new parameters of risk from new data sources and newly available analyses. Just as the advent of spreadsheets within analytical professions revolutionized what could be done in those positions (leading to new positions, new products, and more consumer choice), the new AI capabilities could produce new roles that were previously unfathomable.

The imperative of trust in the age of AI

The insurance industry’s greatest asset is trust: consumer trust that the company will be there to pay claims when it promised it would, to alleviate financial distress in a time of need. With lifetime contracts, this trust must be held ”forever,” for as long as customers live. However we move forward with adopting AI, we must ensure that we retain and enhance consumers’ trust.

In noninsurance settings we already see the online landscape muddied with ”AI slop,” or the vast array of AI-created content that proliferates on social media and in email and more.

In insurance, we must present the case for AI use to our customers: how we reach comfort with “black box" algorithms, or models that seem to lack explainability. We must be vigilant to recognize the potential for biased outcomes to emerge from our models, which are inevitably based on data and methods that can never be completely free of bias. Finally, customers will rightly have concerns about the use of their personal data in our AI models. We must address these concerns with transparency and understanding as an industry.

But in addition to defensive trust fortification, insurers should consider the variety of ways that AI allows for a better customer experience, thereby expanding trust. As an example, insurers that use AI to successfully anticipate their customer needs and behaviors may create a bond that is hard to generate when seeing every policyholder as simply ”part of a pool” of risks. Insurance companies can foster longer-term relationships with consumers through proactive engagement, sending personal, relevant communications when people will most welcome them (e.g., at the time of life events like purchase of a new house).

In that regard, some companies will use AI to their competitive advantage, to strengthen trust among their customers.

Forthcoming research

This is the initial entry in a Milliman series of research and commentary on the value that AI will add to life insurers in the United States. Our subsequent papers will delve into key life insurer functionality, and explore how AI stands to change the company value proposition in the following areas:

• Underwriting
• Customer experience
• Risk assessment
• Fraud detection
• Operational efficiency

In addition, we will explore the regulatory and compliance environment. We will also demonstrate the impact that AI will have on the customer value proposition of life insurers, and outline key strategic milestones that will guide leaders on the path.

References

Ajay Agrawal, J. G. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Boston, Massachusetts: Harvard Business Review Press.

Bethea, C. (March 7, 2024). The terrifying AI scam that uses your loved one's voice. New Yorker. Retrieved November 10, 2024, from https://www.newyorker.com/science/annals-of-artificial-intelligence/the-terrifying-ai-scam-that-uses-your-loved-ones-voice.

NAIC. (2023, July 17). Use of Algorithms, Predictive Models, and Artificial Intelligence Systems by Insurers. NAIC Model Bulletin. Retrieved November 10, 2024, from https://content.naic.org/sites/default/files/national_meeting/07.17.23%20Exposure%20Draft%20AI%20Model%20Bulletin_0.pdf.