GenAI in Insurance
How are AI advances redefining coverage and actuarial practices?
August 2024Photo: Shutterstock/PeopleImages.com - Yuri A
The general sentiment around generative artificial intelligence (GenAI) is that it came into public awareness with astonishing speed and a tremendous amount of good and bad press. There have been articles about the early iterations of large language models (LLMs) being abused to enable harmful behaviors and make eerie predictions, which is no longer easy to accomplish with the controls since put in place.1
We’ve seen a new arms race between tech companies and chipmakers, and thousands of startups connected to AI.2 Yet, if we focus our attention on the actuarial profession, we observe that it’s not yet obvious how quickly this technology will alter our daily work.
Regarding GenAI in the insurance industry, underwriting and claims management are the two areas that most readily come to mind when we think of the potential impacts. Knowledge management is another area, along with life underwriting in particular because the process inherently collects a large amount of unstructured medical data.
This brings us to the actuarial profession and what we have seen progressing daily over the last year. Looking at the end-to-end valuation process, we do not currently see an obvious place for actuaries to use GenAI in insurance. It appears to us that actuaries aren’t looking for creativity in their financial results, nor poetry. GenAI is not intended as a mathematical tool, either. Based on our experience, it’s plausible that GenAI tools could operate at the end of the process—explaining results—relying on generative capabilities, but GenAI use currently is limited due to a lack of mathematical ability.
While GenAI may, as of now, not be ready to integrate into the actuarial production process, it demonstrates significant potential to enhance productivity. Actuarial roles increasingly require programming skills akin to some technology functions.3 This evolution reflects actuaries’ deep product knowledge, mathematical prowess and data proficiency, positioning us as primary developers instead of just serving as a business unit that dictates requirements. Actuaries are involved in coding, testing and deploying sophisticated models—activities once solely attributed to software developers.
Given this, we believe GenAI in insurance has the potential to be a transformative tool in typical day-to-day actuarial operations.
Potential Applications of GenAI in Insurance
So, what are impactful, potential actuarial use cases for GenAI tools?
Code Generation and Documentation
Model development and enhancements potentially could be accelerated with GenAI by having these tasks completed automatically with a few simple instructions to an LLM. While these models mainly generate text, it is possible to enhance their capabilities by allowing them to execute code, make direct edits to existing codebases, run test cases and debug code to essentially act as your own dedicated codeveloper.
Alongside GenAI’s ability to generate code, a much more straightforward application is code documentation. The classic undocumented Visual Basic, Python or R program, originally intended to be an ad hoc process, can now be summarized and explained to a new user in simple text. Taking this a step further, GenAI could be used to generate code documentation and save a documented version of the program for the user, greatly reducing the effort of doing this manually, especially for someone with limited prior knowledge.
Additionally, GenAI’s ability to generate code unlocks a variety of other use cases. It can generate usable functions and daisy-chain them together with other libraries to perform more complex tasks.
Process Orchestration and Results Explanation
Aside from helping with coding-related tasks, GenAI could be used to orchestrate current processes, such as existing actuarial models and results databases. By providing the LLM with instructions on running a model, we could trigger model runs using natural language, develop and kick off ad hoc sensitivities, and retrieve results after the runs are complete. This can be done with any open-source model or via application programming interface (API) calls to existing closed-source software. Taking this concept one step further, feeding the model results back into the LLM allows it to generate draft explanations of the results.
It is important to note that these use cases require engineering to be implemented. As of July 2024, there are limitations that might impact the scalability of some of these use cases. For example, limits on tokens, which are units of text that LLMs process and can represent common character groupings of text,4 mean that summarized information would need to be input for the model to perform results analysis. Hallucinations, which are fictitious and inaccurate GenAI output,5 also introduce limitations around the extent to which we can rely on GenAI models without supervision. But again, we expect these to improve as LLMs and monitoring techniques become more sophisticated.
GenAI Traction Within the Actuarial Space
Based on our market knowledge and talking with our clients, there have been many actuarial proof-of-concept projects to evaluate using GenAI in insurance, but the actual creation of tools beyond the proof stage has been scarce. Three reasons for this are:
- New skill sets, such as the ability to safely maintain and monitor GenAI tools, are necessary. Yet, we believe they are not currently widespread in the insurance industry.
- There are learning curves involved in engineering solutions that overcome the various limitations previously mentioned, including token limits and hallucinations.
- Given the risks that the widespread use of GenAI in insurance could introduce, companywide governance policies need to be updated accordingly. This takes time and additional support from subject-matter experts.
We also find that approaches to GenAI tool adoption range from data- and IT-led use case identification and development efforts to individual research initiatives and small-scale proofs of concept. There are pros and cons to either approach; however, under the primary approach, actuarial use cases are often lower priority in initial rounds relative to the more obvious targets (e.g., underwriting, claims) for which the costs and benefits are likely clearer and would be visible sooner. This is for a few reasons, not least because quantitative benefit tracking can be difficult when GenAI is used as an accelerator of change activities.
For those of you starting on your GenAI journey, three things are worth considering:
- Pick the right use cases. If available, those that are likely to get broader company support could help you move past the proof-of-concept stage. We also suggest highlighting the qualitative benefits.
- Remember the fundamental principles of risk management, governance and controls. These are just as crucial to your GenAI proof of concept as they are to other areas, and we suggest incorporating them into the design.
- Give teams the liberty to experiment within a safe framework. While not all use cases may succeed, this could foster engagement and enthusiasm among your team members.
What Does This Mean for Actuaries Going Forward?
We expect that GenAI tools will affect actuarial work in the long term, but it’s uncertain how quickly this will happen. We believe some of the everyday use cases will materialize quicker, and capabilities requiring heavy engineering may take years to implement as the tools mature and actuaries learn the technology.
From a business perspective, we believe the potential efficiencies and qualitative benefits make using GenAI in insurance worth considering. And, with the additional capacity for actuaries to focus on deeper analysis, strategic business decisions and addressing complex challenges, there is potential for net-new business value and a clear appeal for early adopters who may see a competitive advantage as a result.
This also will require adding multiple skill sets to the actuarial field. Actuaries will need to understand and validate the outputs of GenAI (and AI) models. Programming, data and GenAI-specific skills, and knowledge of the responsible use of AI could become even more critical than they are today.
Lastly, we assert the need for actuaries to focus on judgment-based analysis and strategic questions will not go away. GenAI is just the latest technology in a long history of advancements that push actuaries away from rote computation and toward higher-level analysis. We look forward to embracing this change.
Statements of fact and opinions expressed herein are those of the individual authors and are not necessarily those of the Society of Actuaries or the respective authors’ employers.
References:
- 1. Zacharakos, Alexis. How Hackers Can Abuse ChatGPT to Create Malware. TechTarget, February 22, 2023. ↩
- 2. David, Emilia. Chip Race: Microsoft, Meta, Google, and Nvidia Battle It Out for AI Chip Supremacy. The Verge, July 19, 2024. ↩
- 3. Purushotham, Marianne, and Dararith Ly. Top Actuarial Technologies of 2022-2023. SOA Research Institute, September 2023. ↩
- 4. Stricklen, Michael, Khalid Khan, and Jeff Vogel. Four Steps to Implement a Large Language Model Successfully. EY Parthenon, November 15, 2023. ↩
- 5. Wray, Jeff, Dan Diasio, and Gautam Jaggi. Five Generative AI Initiatives Leaders Should Pursue Now. EY, October 5, 2023. ↩
Copyright © 2024 by the Society of Actuaries, Chicago, Illinois.