AI isn’t your father’s reserving software: Why that matters more than you think

By Frank Huang, CEO of Insurance Solutions North America

One of the interesting things about AI is the ever present question of “Will it take my job”. So far to my knowledge I am still gainfully employed but it’s clear that the actuarial profession will not be immune to the impact of AI. Today we’re already seeing clients and friends of the firm investing in AI for the purposes of bolstering internal actuarial capabilities similar to how investments in reserving software occurred 10-20 years ago.

But as the hopefully catchy tagline notes, we are noticing interesting parallels between these two generations, such that we can learn from the past to realize the upside of AI and minimize the risk.

From calculations to decisions

If you think about traditional actuarial software – ResQ, ReservePro, Arius, and to a lesser extent Excel – the benefits have always been clear: faster calculations, more consistent processes, and scalable workflows for known methodologies. At the end of the day, these tools helped us do what we were already doing, just better and faster.

AI is different.

It doesn’t just take inputs and run with it, it participates in the thought process at each step of the way. For example, it can suggest methodologies, make assumption selections, explain reserve movements, draft narratives for stakeholders, and run all sorts of scenario testing to the user’s delight.

The difference is meaningful as it impacts how we humans use and validate the results. We can no longer just validate the tool itself, we must validate how people think with these tools every step of the way.

Where did the risk go?

With traditional reserving software, most of the risk lived in the tool itself. “Was the data mapped correctly?” “Were the assumptions reasonable?” “Were the parameters set up properly?” Once you validated those inputs, you had a level of comfort that should you put in the same inputs, you would get the exact same outputs. If something was wrong, it was usually consistently wrong – and that made it easier to identify and fix the issue.

AI flips that dynamic. Now instead of the risk lying primarily in the inputs, the risk has shifted to the human using AI. You could have two people sitting side by side use the same AI and seemingly immaterial differences in prompts can lead to materially different answers.

In other words, every prompt is effectively a new model run of potentially a different model.

This introduces a new kind of variability that traditional validation and governance were not designed to handle.

Where did the risk go?

One of the things I’ve always appreciated about reserving and actuarial software is that when it’s wrong, it’s usually pretty easy to understand where it went wrong. Perhaps it was an incorrect LDF selection, or a bit outlier trend assumption, or certain years were given more/less weight than ideal.

AI errors are different. Because traditional software has codified calculations and processes, only the things around it like inputs are where things go awry. But with AI, any step along the way could be erroneous. For example, AI may have miscalculated steps or skipped steps along the way, it may have incorporated hallucinated data points, or even utilized a twist in the methodology that the user fails to fully comprehend.

These are such new types of risk that expand not only the types of errors that can manifest itself in usage but consequently expand the depth and breadth of validation and governance necessary to ensure the AI models are appropriate for the use case.

The worst situation is that a human is in at the beginning and end of the loop but there is not one ensuring the veracity of the loop itself, and then management/board/client makes decisions based off those results.

Next steps

AI is not all bad, so apologies if that has been the tone thus far. The point to highlight here is that AI has very real benefits to scale, efficiency, and insight but the process has to be overseen, validated, and governed more fully than software of the past.

Years ago, I stole from a mentor the explanation of what actuaries do – “We get paid to do other people’s math homework”. In this new generation, AI can now do the math homework, but we need to make sure it shows its work and gets the answer right.

If you would like to continue the conversation, get in touch with CEO of Insurance Solutions North America, Frank Huang at frank.huang@us.davies-group.com