This essay is one of five essays that were submitted in response to the CAS Publications Department’s call for essays on the “Intersection of Artificial Intelligence and Actuarial Science.” See the CAS 2024 Summer E-Forum for three other essays, including the prize-winning essay by Ronald Richman.
Maybe you too have attended a meeting to work out with your colleagues the best way to structure the team’s folders. Surely there must be a naming convention and structure that will finally result in things not getting lost. In that meeting, did someone point out that Microsoft and all other current file management systems allow you to use labels? You can categorize your files simultaneously in various hierarchies, rather than in just a one-folder structure. Digital files don’t need to be put in a single bin or folder because they aren’t physical objects. If you take the time to create meaningful labels, you’d just need to name your file and apply labels, rather than stress over exactly where to put it. That works well, but again, it relies on you and your colleagues to design and apply labels. The point of failure is, of course, the “you and your colleagues” part. So, with our advanced technology, what behavior have we adopted? Super-fast searching. Our computers index everything, and most of us use various search functions when needed, rather than putting much effort into organizing anything very well beyond the project level.
Invention, innovation and a desire to make life incrementally better seem to be a basic drive of humanity. In ancient times, the innovations stayed localized and only moved between groups when those groups met up to trade, mingle or fight — very much like dolphins and other animals have localized foraging and hunting skills. Sometimes knowledge was lost when the person knowing the information died. Then came writing. Knowledge could be preserved across generations. Detailed instructions for making beer, recipes for tasty dishes and accounting of taxes and debts owed are some of the oldest recorded documents. Then came the printing press and innovations in all areas of life spread like wildfire. You’ve heard the story.
AI is turning out to be as transformative as the printing press and the internet were. As far as innovations go, it’s not even in the same ballpark as labeling folders.
Hopefully, you’ve read not just about other great inventions in history but about how well they were received. People whose lives were expected to be made better by big inventions were ecstatic and talked of progress. Those whose jobs were being automated or politicians facing angry populations were anxious. As each society-changing invention was poked, prodded, improved upon, and finally used widely by society, the inventors, or usually the refiners, made huge profits, while those whose work was automated either mastered the new tools or were left to find other work. Were the inventions forced upon society? Mostly no, we chose to embrace them. Where’s my proof? When was the last time you bought a hand-sewn dress or suit jacket? You still can; nobody is stopping you. There was a time before sewing machines, when fabrics were sewn together by hand to produce clothing. To buy clothes made that way today would cost you 25-100 times14 the cost of the item made with the help of machines. Embracing innovation is what changes the way we live collectively. It frees up the laborer’s time and reduces costs for the consumer. Economists sometimes have predicted that people will spend less time working because of all the inventions. But that hasn’t ever proven true. People find other employment, work just as much, and the benefit is passed to society in the form of new stuff to buy or the same stuff to buy at lower cost. Laborers of all sorts spend less time working because of social standards and regulations, not because their jobs were automated. People who don’t like the new inventions have the option to live in societies that don’t have those inventions. Where do you draw the line though? Even those living off the grid use Mason jars to store their herbs and copper wires to connect their solar cells to their LED light bulbs.
But now I’m rambling. Let me get back to the subject of what AI will mean for the actuarial community.
Let’s start by defining some terms. Well, just one: AI. I won’t add a footnote linking to the controversies around the words “Intelligence” and “Artificial.” If you haven’t already, please Google “What is AI” and immerse yourself in the ongoing effort to define human intelligence and what an artificial version would be. I’ll define the term by distinguishing AI models from the models we’ve been building for 30+ years.
Actuaries have always designed and used models of various forms. When people talk about models in modern actuarial contexts, they are talking about models that require the use of computers — what is often called machine learning outside actuarial groups. Among actuaries, generalized linear models (GLMs) have been implemented for pricing since at least 1989. The pioneering work was done in unregulated personal motor lines in the U.K. and Europe.
GLMs, decision trees and other machine learning approaches deliver to the analyst a model which has inputs and functions structured in such a way that the builder of the model can see and document exactly how a given output is calculated. In the context of pricing, you can state which characteristics of the insured are determining the price that the model produces. A broad definition of AI would include those models. But that’s not what people mean when they say “AI.”
Whenever I hear people use the term “AI,” they are referring to a computer program that does calculations resulting in capabilities that they previously thought only people had — things like identifying objects in pictures, correcting phrases, translating languages, making up stories and pictures, driving a car, carrying on a conversation or writing computer programs.
Most models that people refer to as AI are instances of what practitioners would call very large artificial neural networks, or just neural networks (NNs). The weird thing about NNs is that the engineers don’t know quite what the model is using to make its inferences. The engineers know how many layers and nodes there are, they know the weights assigned to the connections that came from the training process, they know what the utility bills are for running the servers and graphics cards, and they know what information was fed into the model. But there are billions of nodes and weights. The model created by the training process is something that seems to reproduce some of what our own brains do, that is trained in kind of a similar way, and in the end, that seems to be just as hard to understand. That is what I’m going to call AI: any model that is built in such a way that its creators don’t know exactly which inputs and functions lead to a given output.
The creators know how to set up the environment. They know how to kick off the training process. They know how to evaluate the accuracy of the model. To track exactly how the model comes up with any one of its inferences would probably require an even bigger and less scrutable model. All you can really state is what all the inputs are, then look at the output and decide whether you like it or not. That’s what people mean when they say AI.
Using NNs researchers can build headbands today that convert electroencephalogram (EEG) waves to text with 40% accuracy.15 Today! With those headbands, we the analysts will be writing code and responding to messages by what will feel like tomorrow.
Over meals, I’ve heard people a generation (or two) younger than me balk at the idea of brainwave reading devices and laugh at Neuralink’s ambitions. But they are wrong. Society will accept these communication devices with open arms. Just like we did with printed words, cars, light bulbs, machine guns, nuclear power, computers, the internet, YouTube, smartphones and Amazon. What makes “now” any different? People aren’t different, and there are better safety protocols in place. I bet that as a percentage of the population, far fewer people will ever die from Elon Musk’s Neuralink implants than died in automobile accidents during the first 10 years after the Model T hit the roads.
But how can actuaries use such tools? I’ll address the three areas of work I have experience with: predictive analytics, capital modeling and reserving for non-life insurance lines.
Let’s start with reserving work. There are two main branches of reserving work: claims (or case) reserving and actuarial reserving. These functions are performed somewhat separately by the claims departments and actuarial departments. And there’s no need for me to try to summarize them — here’s the ChatGPT summary:
Actuarial reserving refers to the work typically done by actuaries, which involves estimating the required reserves for the entire portfolio of insurance contracts to ensure the company can meet its future liabilities. This process often uses statistical models and historical data to predict future claim payouts.
Claims reserving, or case reserving, refers to the estimates set by claims adjusters for individual claims. These reserves are set based on the adjuster’s assessment of the amount the insurance company will need to pay to settle each specific claim.
These generalized labels distinguish between the macro-level, statistical and model-based approach of actuaries (actuarial reserving) and the micro-level, individual claim-focused approach of claims adjusters (claims reserving or case reserving).
Claims adjusters are exposed to many individual cases and develop a detailed understanding of expenses and payouts associated with claims. Case reserve departments follow their own norms like any other corporate function. For example, on first notice, an adjuster will create a record in the claims system. The team may have a rule that you set a minimum case reserve of $1,000 to indicate the claim has been opened. Another rule may be to set the case reserve to $50,000 to indicate you expect legal action rather than a simple payout. Many companies already have models that estimate the ultimate payouts, models that update the initial estimates and models that prioritize claims likely to go to court. The ability of AI to read documents and understand images is already drastically changing workflows in this area. For example, some case reserves are set automatically by an AI based on images uploaded by the policyholder; that is, if the fraud detection AI doesn’t flag the policyholder as a risk.
Case reserving workflow changes will continue as more and more models are trained on the various stages of the work. They will ingest and utilize internet data, legal information and cost trends, and they will interact more and more with vendors that insurers use for settling claims.
I think the areas of non-life actuarial reserving most likely to be impacted by AI have to do with data collection, workflow automation and document generation. I doubt that AI will be used to dramatically change the calculations performed for three main reasons.
1.Lack of incentive. Improving ultimate loss estimates by 1% does not have the same impact on profitability as improving pricing accuracy by 1%.
2.Strong auditing and regulatory oversight. Altering reserving methods increases audit and regulatory burdens, which is difficult to justify given the first issue.
3.Executive involvement. Automating reserve calculations reduces expert judgment and management flexibility.
Capital modeling is in the same boat as reserving. In capital modeling, you have exposure information at various levels of granularity and risk measures for nearly all categories of risks a company faces. Investment portfolios, natural catastrophe risk, cyber risk, price and wage inflation — it’s a long list. It is the area I spent most of my career working in, and maybe that is why I struggle so much coming up with ways AI will improve capital modeling-specific tasks.
Soon we’ll have copilots helping us with everything we look at digitally. And soon after that, everything we look at will be converted into a digital representation so that our copilots can help us with the real world, too. Documentation, designing presentations, coding workflows, research and many other tasks that are part of capital modeling will be improved. But what about fitting probabilistic distributions to historical data, selecting correlation coefficients and designing risk tolerance statements? Maybe the copilot will be there to help remind actuaries that describing risk in terms of percentiles is far more effective than talking about standard deviations.
I’m sure there are ways we could use AI to improve Montecarlo sampling, though my imagination is hitting its limits on that one. A market constraint exists: How much money is in it? How much improvement can really be made? Those are the things that draw investment in innovation. If a new killer app for capital modeling shows up, it will be from a boutique startup run by an enterprising team of actuaries. And if it is successful, it will certainly attract the attention of larger firms.
That brings us to the last branch of actuarial work I’m going to write about in this essay: predictive analytics. As I pointed out earlier, the actuarial community has been applying computer-based predictive analytics in pricing since at least the late 1980s and early 1990s. The datasets have grown larger, and the complexity of the models has increased. One thing that’s remained constant, though, is that the input variables used by the model to produce predictions are well-known and subject to regulatory oversight. There is typically a logical link between the predictors and the target variable. With the advent of telematics, the predictors are more arguably cause-and-effect as opposed to correlative.
Why use a neural network to do something a GLM or decision tree is doing very well? In competitive environments, we don’t introduce complexity unnecessarily. I suspect that pricing and underwriting models will continue to be mostly based on GLMs and decision trees.
But everyone knows the modeling part is the easy, fun part of the job. The big lifting happens in building our datasets. And that is where the Large Language Models (LLM) and NNs will really shine. These tools are already giving us access to vast stores of features in documents that were very difficult to extract with older Natural Language Processing (NLP) strategies. Companies can see the value in structuring the storage of these documents and making them accessible to the modeling teams. Over the last 20 years, predictive analytics and data science skillsets have turned into full career paths. The same is happening with feature engineering, and that space will be strongly influenced by AI.
All those impacts I’ve described didn’t address how AI will change the nature of the risks we insure. The world has steadily grown safer for people and more connected in trade and governance. At the individual insured level, the amount of information garnered through AI interpretations of video, audio, GPS and various other sensors will continue to improve safety and reduce risk. This information will impact the work of predictive analytics and reserving. However, the interconnectedness of things will create opportunities for contagion that didn’t exist previously. For example, at present my car’s ability to navigate an intersection does not depend on the same cloud computing resources that my bank transfers do. Without some time of planning, they soon will. That tendency will impact the capital modeling teams and create more topics for them to research and model, just like it did with cyber risk and cyber insurance over the past 10 years.
These are my views on the intersection of AI and my corners of actuarial science – in the near term anyhow. Looking a bit further down the road, I’m certain we’ll be hiring people who grew up using virtual reality headsets to play games and to interact with their friends. Those new recruits will be perfectly comfortable using immersive devices. In fact, they’ll find it silly to use just a screen or even a keyboard. I’m sure I don’t know all the ways AI will change actuarial work, but I’ll love watching it develop and, hopefully, being part of it. ●
Mario DiCaro, FCAS, CERA, works for Tokio Marine HCC.
