Actuary — tasks, one by one

Preparing the data and the experience study

This is the part of the work that is structured, repeated every period and checkable against the source, which is the profile that automates first across this site. Code and tools have taken on much of the assembly for years, and generative tools now write the queries and first drafts of the analysis as well.

No source on this page measures how much of this is automated at any insurer. Faster preparation does not make the assumptions behind it any less a person's responsibility, and a data problem found late is still found by someone.

Building the pricing model

Machine learning is used in pricing and underwriting, and regulators have responded by governing it rather than banning it: the European framework treats AI used for risk assessment and pricing in life and health insurance as high-risk, and New York expects insurers to show that the external data in a pricing model is supported by accepted actuarial standards. The model gets more powerful; the obligation to justify it grows with it.

These are expectations about how pricing must be governed, not measurements of how pricing work is now divided between people and software. They also apply to regulated insurers in two jurisdictions; elsewhere the obligations differ.

Setting the reserves

The calculation is heavily tooled and the tooling keeps improving, but the European supervisor names the actuarial function as responsible for the controls on AI systems within its remit, and gives coordinating the calculation of technical provisions as an example. So the arithmetic is assisted and the accountability for the number is placed, by name, on a function.

Naming who is responsible for the controls does not say how much of the calculation is done by software, and the supervisor's opinion is addressed to national authorities rather than directly to firms.

Testing a model for unfair discrimination

Work created by the arrival of AI in pricing rather than by anything that existed before it: New York expects this testing before a system goes into production and on a regular cadence, and lists quantitative methods such as adverse impact ratios, denial odds ratios and marginal effects. The European supervisor lists fairness metrics too, and warns that some of them can contradict actuarial fairness — charging the same price for the same risk.

An expectation that testing happens is not evidence of who does it; at some insurers it sits with data science or compliance rather than actuaries. And the tension between group fairness and actuarial fairness is named by the regulator, not resolved by it.

Signing off on the numbers

The European framework gives the actuarial function the opinion on the overall underwriting policy, and responsibility for controls on the AI within its remit; responsibility can be supported by software but not transferred to it. This is the part of the job that exists because someone must be accountable, and more automation beneath it makes the sign-off more demanding, not less.

A duty that stays with a named function says nothing about how many people that function employs, and a smaller team signing off on more automated work is entirely consistent with it.

Explaining the number to the people who decide

Both regulators expect boards and senior management to understand and oversee the AI used in their insurers, which places a translation job between the model and the people accountable for it. That job gets larger as the models get harder to read.

Nothing here says actuaries in particular are the ones asked to explain, rather than data scientists or risk managers, and no source on this page measures it.