Jason Alleyne
Cofounder, Reimagine Risk
A new workflow is emerging in the insurance world. It will remove the staff who code actuarial models, displace traditional actuarial software and the checking work it creates, and automate how results are interrogated and analysed.
In this new revolution, some will simply use an “AI-equivalent” version of an existing calculation engine by making a like-for-like replacement swap of an existing actuarial workflow.
That would be akin to the early factory owners who replaced the steam engine with an electric motor but kept the line shafts, and the factory floor built around them, exactly as they were. The true winners in the Industrial Revolution were the industrialists who fundamentally revamped their manufacturing workflows to optimise the benefits of the new power source.
Actuaries now face the same choice with AI.
“Generative AI will break the current model apart.”
Jason Alleyne
Cofounder, Reimagine Risk
Within the current actuarial function, the workflow runs as a production line of task-based roles across five areas: product development, pricing, ALM, valuation, and risk and capital management. Each role has its matching tool: product illustrations, Excel spreadsheets for pricing, Bloomberg terminals for market data, actuarial valuation software, and stochastic generators for capital.
Despite heavy regulation and a high barrier to entry in insurance, these workflows and the SaaS tools built around them proved efficient enough to win wide adoption. Generative AI will break the current model apart.
An era of AI-led workflow should start with the recognition that AI will be the most efficient coder and developer for any computational tasks.
Users will work through prompts that call up the right “actuarial thinking unit” (ATU): a small finance and actuarial model, similar to an Excel worksheet, each one built around a specific actuarial concept, that carries actuarial judgement into what-if analysis.
That judgement is what sits behind the answer actuaries are known for: “it depends”. A reply that has long given the profession an air of authority.
Our new AI, relying on its ATUs, will traverse the “it depends” questions with ease.
The first leg: building the ATU
The work an ATU does is to weigh related but competing constraints against each other at the same time. A good ATU will cover its subject in enough detail to represent the real world usefully, while staying simple enough to remain intuitive.
One way to picture ATU development is through the actuarial rules of thumb everyone already knows.
Take duration mismatch in life insurance, which correlates strongly with an insurer’s economic risk capital. The size of the interest rate shock in a jurisdiction’s RBC regime, set against a rough estimate of the duration mismatch gap in the portfolio, gives a useful proxy for that capital position.
That rule of thumb is directionally right for what-if work on investment duration, and an ATU built on it captures the concept of duration mismatch in a form that can train an LLM.
Duration mismatch is a single-factor rule of thumb, but a working ATU usually has to balance several constraints at once. One such case, drawn from a real “it depends” situation, is the ATU for an indexed universal life (IUL) product.
The IUL design rests on an ALM strategy built around a fixed-income portfolio that generates a coupon yield. That yield sets how many call options the insurer can buy to gain exposure to an equity market index. The actuary must also weigh the cost of insurance (COI) charge for death benefits on top of the client funds deposited. Then comes the capital that the regulations demand in the market where the product sells, and the return required on that capital, namely the profit margin. Finally, there is the sales commission.
These five factors—the fixed income choice, the option budget, COI, cost of capital, and commission—drive the IUL, and the simplest what-if question is a yes or no on whether a design works.
“ATUs could well become the main source of actuarial skill and job satisfaction in AI-system development in this new era.”
Jason Alleyne
Cofounder, Reimagine Risk
In this ATU, the Excel version shows the investment choice driving economic risk capital along two dimensions: the regulatory risk capital charge for the asset class within the SAA, and the duration mismatch gap.
The first comes straight from the available assets, cross-referenced to the regulatory requirements. The second leans on the rule of thumb set out earlier, which links the duration mismatch gap to a proxy for economic risk capital, typically a shock scenario.
The model then compares the weighted average yield from the assets in the SAA with an estimate of the yield the product needs to cover its components: option budget, COI, cost of capital, and commission amortisation. From that comparison, it generates a language-based result: the yes or no.
More sophisticated language-based results would examine calculated metrics such as ratios or averages or percentiles for the results relative to multiple instances of the input parameters.
This ATU shows how the rules of thumb of actuarial insight can be captured in a reasonably realistic construct, which can then generate language-based results for millions or billions of combinations to train an LLM.
ATUs could well become the main source of actuarial skill and job satisfaction in AI-system development in this new era.
The second leg: the user side
On the user side, the tool takes a query framed in actuarial terms.
In our example, the user describes an IUL product that will offer brokers a certain commission, or whose market posture demands a certain option budget.
The prompt engineering guides the user towards questions the ATU can answer, and the richness of the exchange depends on how well the ATU and its language-based results are designed.
Users with sufficient knowledge of the subject matter and intuition for the ATU would be more capable in designing their prompts and extracting more from the computed insights embedded in the language-based results.
The system also has to remember. That means updating the ATU’s library of language-based results as new market data and user interactions arrive. Human reinforcement learning workflows could be set up to improve the language-based results with subject matter expertise and richer context.
“Those who study human history may prove to be the ones best placed to master our future.”
Jason Alleyne
Cofounder, Reimagine Risk
Actuarial career implications
Nothing described here sits far beyond what the individual contributors can already do: asset managers, subject-matter expert actuaries, AI specialists, and system developers each have a role to play.
Coordinating that expertise into customisable AI tools will take time. Actuaries will either embrace the new workflow or miss the shift in mindset it demands and fade into irrelevance.
AI will disrupt our work, but it will also make our careers more rewarding. Embracing change will mean different things to different people. But change inevitably means different industrial complex workflows that optimise for the new AI capabilities.
Equally likely will be the emergence of new roles that displace the current delineation of tasks and disrupt the job descriptions and scopes of the current actuarial departments.
Those who study human history may prove to be the ones best placed to master our future.