Senior Actuaries Tell SOA Study AI Use Is Wide, Shallow and Unguided

Thirteen senior actuarial leaders interviewed for the Society of Actuaries' Foresight Project describe a profession using AI widely but shallowly, under pressure from executives to adopt it and without a roadmap for doing so, according to a report published by the SOA Research Institute on October 13.

The report, written by researchers at the MIT AI Risk Initiative at MIT FutureTech and funded by the SOA, is the first phase of a multi-phase project on AI risks to the profession. It is built on one-hour interviews with chief actuaries, consulting partners, founders and academics, about half of them in the United States, and its findings will be tested in a survey of SOA members in the next phase.

Governance failure, transparency and interpretability, fraud and scams, privacy and model failures including hallucinations were the risks the leaders raised most often, the report says. Those match the consensus of 272 AI risk experts in earlier MIT research. The actuaries parted from the experts on reputational risk and on the human cost of overreliance, lost judgment and displacement, which they ranked far higher.

Key Takeaways

The SOA's Foresight Project interviewed 13 senior actuarial leaders, a 72% response rate, and found AI adoption "experimental and fragmented," with no standard the profession can follow.
One leader put weekly chatbot use among 5,000 consultants at about 95%, use inside a process at 25%, and frontier use at "me and three other people."
Governance failure came up in almost every interview: frameworks built for repeatable quarterly output do not fit models whose output is non-deterministic and continually updated.
The leaders ranked reputational risk and labor displacement well above the 272-expert MIT consensus, with near-unanimous concern for the junior talent pipeline.
Proposals for the SOA include exam reform away from rote memorization and a parallel "actuarial technologist" credential track.

Wide use, shallow use

The report's picture of adoption is consistent with the SOA Member AI Survey of more than 500 actuaries in 2025, which found most respondents spending under five hours a week on AI tools and using them mainly for peripheral tasks. A year on, the leaders saw no qualitative shift. "What I see is just little experiments here and there," one said. Another said companies are "very much still in the exploratory phase."

The sharpest description came from a leader describing a firm of about 5,000 consultants. Asked how many had used a general-purpose model in a given week, the answer was "like, 95%." Asked how many had built an AI tool into a process: "I'd bring it down to 25%." Using a coding agent to produce presentations: "me and three other people."

The report says the common day-to-day uses were writing, research and administration, with coding assistance rising fastest: rewriting spreadsheets in Python, converting legacy models and building calculators.

Several leaders questioned the economics. One expected companies to find they "can't throw agents at everything because it's actually becoming way too expensive." Another said "90% of the organizations are yet to see any significant returns." A third, after prototyping a solvency tool with generative AI, found reading the regulation directly "a lot less frustrating" than checking the output.

The leaders also said the attention on generative AI is crowding out work on the predictive models the profession has historically built for itself. "It shouldn't be at the cost of surrendering one of the core disciplines of actuarial science, which is building good predictive models," one said, pointing to actuarially aware neural network designs and tabular foundation models as the neglected frontier.

Where the pressure comes from

The report sorts the forces behind adoption into drivers and barriers: cost and market pressure push it; IT security policy, missing guidance and lack of bandwidth hold it back. "Possibly my CEO is saying, use more AI in your work, but, you know, she or he doesn't have any kind of firm advice for how to do it," one leader said.

Security rules can be self-defeating, another said: staff without access to approved tools "are doing stuff on their own laptops, and obviously that introduces data security risks." A third described investment flowing past the actuarial function altogether: "if they have a few million dollars to invest, they are not going to the actuarial team, they're going to the tech team."

Influence on adoption	Direction	Leaders' view
Internal pressure to reduce cost	Driver	Efficiency and "cost arbitrage" are the dominant motive; one leader doubted the headline savings would materialize
External market pressure	Driver	Signaling AI capability has market consequences for insurers
Personal appetite for technology	Driver	A stronger predictor of use than practice area, type of work or seniority
Security and IT infrastructure	Barrier	Blocked tools push staff to riskier workarounds on personal devices
Lack of implementation guidance	Barrier	Adoption confined to isolated pockets; "no widespread standardized implementation"
Lack of bandwidth	Barrier	"If you are already 90% loaded, that 10% doesn't help you to clear your mind and be creative"
Professional standards and sign-off	Both	A barrier to rapid adoption and the profession's main means of managing AI risk
Experience and seniority	Disputed	Some saw juniors as more cautious after early experiments; others saw seniors using the tools more critically

Source: Table 1 of the report, summarized. Direction follows the report's own markers for agreement among the leaders.

Governance built for repeatable work

Governance failure, defined in the report as AI capability and adoption outpacing an organization's ability to manage the risk, came up in almost every interview. The leaders said traditional actuarial governance and model risk frameworks were designed for routine, repeatable output and are "not merely lagging but are inadequate" for models whose output changes from run to run and which are updated continuously.

"If things are mechanical and every quarter you produce the same thing again and again, then I guess the governance and risk management frameworks that are in place work," one leader said. "I get very nervous when we get into the non-recurring or the non-regular work ... when is it really in production?"

Interpretability drew the same concern, framed in terms of professional sign-off. "Our professional standards says that our results should be replicable. And they should be explainable, right? And if I don't do that, I'm at risk of losing my license," one leader said. The report notes two further problems for model risk management: as models become more specialized, the small pool of people able to validate them stops being independent, and once agentic systems are used "like a person," even maintaining a model inventory becomes difficult.

The governance in use is layered rather than designed, the report says: contractual guardrails, internal "do's and don'ts," SOC 2, the NIST Generative AI Profile, ISO 42001, NAIC initiatives and OSFI's Guideline E-23, with professional accountability underneath. "AI will not sign a report, will not be sued, will not be asked to ... be disciplined by a judiciary body," one leader said. The informal practices the leaders described include breaking AI work into verifiable pieces, validating on the deterministic parts of a model first, running several models and comparing them, and using coding agents to build deterministic pipelines that then run without AI.

Views on specific risks were shifting. Most leaders still named privacy, security and hallucinations as primary risks, but a minority said hallucinations are "not the sort of problem that it was a couple of years ago," and one argued that cascading errors across networked agents should now rank higher. Discrimination, which the MIT experts rated among the sector's most vulnerable areas, drew mixed views; some leaders said AI merely mirrors existing bias and that "insurance as a business is discriminatory."

The pipeline problem

Concern for the junior talent pipeline was near-unanimous. The leaders described two compounding effects: the loss of entry-level work, and the loss of the hands-on tasks through which junior actuaries build the judgment later needed for oversight. One asked how anyone plans for "career advancement where the machine is doing all the core learning." Another described staff asked to document their thinking who then wonder "Am I being replaced by AI eventually because they just want to download my brain?"

Talent is also leaving. One leader said that at the largest university in their country, where actuarial science "is extremely prestigious," it had for the first time been overtaken by computer science. Another said they were "losing my best people to tech teams" that focus on actuarial problems from outside the actuarial function. The report ties this to a longer-running move of data science and finance into actuarial work: "tech [is] cannibalizing risk areas."

The leaders agreed on which skills to preserve, with first-principles understanding first. "If you've done the math and done the spreadsheets enough times, you start to understand the drivers of things ... I think people have lost that a little bit over time as the models have done more and more of that work," one said. Coding ability is being commoditized, they said, while the ability to read and evaluate code is not. Some expected the learning curve to compress, with skills that took five to seven years developed in two to three.

They disagreed on what deeper expertise the profession should own. One camp wants actuaries back to building their own models on current machine learning methods, calling the loss of that discipline "without a doubt the biggest gap." One leader proposed a dual track within the SOA: the traditional path to appointed actuary alongside an "actuarial technologist" route combining actuarial training with external technology certification and SOA-supervised projects. Another camp argued for judgment AI cannot supply, and for an education that adds economics, geopolitics and systems thinking, because "the past is not predictive of the future" for climate, AI and geopolitical risk.

The report lists four opportunities for professional bodies: structured guidance and standards for AI use, exam reform that reduces rote memorization while keeping first principles, a broader curriculum with different career tracks, and public position statements on safe and ethical AI. The next phase will survey SOA members to test where the leaders agree and disagree, including differences by seniority and which skills are being eroded.