Actuarial Judgment Elevated

As artificial intelligence (AI) advances, it is worth pausing to note that the actuarial profession has successfully navigated technological disruption before. Using primary tools such as manual calculators and logarithmic tables, the actuary of yesteryear checked and rechecked mortality tables, premiums and reserves estimates, albeit via laborious computation. When the room-sized mainframes arrived—with processing speeds that seem glacial by today’s standards—they did not render the actuary obsolete; rather, they did something more interesting: they freed the profession to ask harder questions. Before computers, running stochastic models or complex scenario testing at scale seemed unfathomable; computational prowess enabled the actuary to evolve and take on work that was simply too unwieldy to achieve previously.

I would like to gracefully assert that we find ourselves at an equivalent inflection point today: AI’s predictive capabilities will catapult the profession into a new order of analytical possibility. The actuary, yet again, finds themself at the brink of change. Will the profession lead that change or be led by it?

At its core, the celebrated superpower of AI, specifically as it pertains to our profession, is its ability to advance predictive analytics to the next level. The gamut of data at our fingertips, such as claims histories, wearable device data, behavioral signals and satellite imagery, can be processed by AI with unfathomable speed, and the inherent underlying patterns discerned with insights that can arguably exceed what can be achieved by the human mind alone. This is possible primarily due to advances in quantum computing and neural networks, which have unlocked unbridled computational power to process high-dimensional data and detect complex patterns that are difficult for humans to replicate at scale.

The economic logic of this shift becomes explicit in the book “Power and Prediction: The Disruptive Economics of Artificial Intelligence,” where authors Ajay Agrawal, Joshua Gans and Avi Goldfarb argue that as the cost of prediction falls, its relative value declines, increasing the importance of complementary inputs, particularly in human judgment. Correspondingly, what rises in value is judgment: the interpretational ability of the human in the loop, who ultimately decides what to do with the prediction and is accountable for the decision. This is a fundamental expression of how the expertise of the human professional is leveraged to create value.¹

This reframing is worth noting. For actuaries, the profession is a combination of prediction and judgment: estimating mortality but also setting defensible assumptions, establishing prudent margins, crafting documentation for regulators and policyholders—the list goes on. AI can be leveraged to sharpen the prediction end of the equation considerably, whereas the judgment end is where the practical opportunity for actuaries becomes more concrete, making the intellectual framework of “Power and Prediction” most instructive for me.

The authors draw a critical distinction between task-level thinking and system-level thinking. Task-level thinking asks whether AI can perform specific functions better than a human. In many actuarial contexts, the answer is an emphatic yes when it comes to computational analyses. For example, gradient boosting models, in certain contexts, can often outperform generalized linear models (GLMs). On pricing tasks, natural language processing (NLP) can easily extract structured data from unstructured claim narratives. And fraud patterns can be aptly detected using anomaly detection algorithms. However, the authors caution that task-level thinking is currently the dominant approach to planning for the introduction of AI into all sectors of the economy, which may not be enough.²

Realizing AI’s transformational opportunities may require examining entire systems and understanding how AI can improve them. This raises a more consequential question: Given that AI potentially can perform certain tasks with greater speed and accuracy, how should the entire decision-making system evolve to harness that capability? Answering this requires someone who fully understands the problem’s architecture, including the regulatory environment, institutional or enterprise incentives, downstream implications of model outputs, and feedback loops between decisions and behaviors—in short, an actuary. This perspective is most powerful when applied alongside technologists and other domain experts, reflecting the inherently multidisciplinary nature of AI-enabled decision systems.

I believe where we stand today with AI mirrors the moment the profession faced when hand calculations gave way to computers: The actuaries who thrived then were not those who automated what they had previously done manually, but rather, the ones who reimagined what was possible and asked what new problems could now be solved. The same shift can be envisioned today from task-based to systems-based thinking. Just as the advent of computing power did not replace the actuary but redefined the actuary’s role, similarly I foresee AI handling the computationally cumbersome predictive tasks, giving way to the actuary to serve as the architect of the system in which those predictions are used responsibly.

THE CASE FOR THE HUMAN IN THE LOOP—AND WHY CREDENTIALING MATTERS

In his 2019 book “Human Compatible: Artificial Intelligence and the Problem of Control,” Stuart Russell argues that human oversight of AI systems is not just preferable, but rather structurally necessary. Russell outlines the Gorilla Problem: the problem of whether humans can maintain supremacy and autonomy in a world that includes machines with greater intelligence. The gorilla analogy is pointed: gorillas are not endangered because humans are malicious toward them, but because humans now control the economic, environmental and agricultural systems that were never designed with the welfare of gorillas in mind. The gorillas simply fell out of the loop, and Russell’s provocation begs the question that humans may be inadvertently engineering themselves into the same position.³

At a more subtle level, Russell argues that the greater risk is not that AI systems will turn hostile, but that they will become misaligned with what humans value. As he states in the book, “A technical way of saying this is that we may suffer from a failure of value alignment—we may, perhaps inadvertently, imbue machines with objectives that are imperfectly aligned with our own.” I believe that for actuaries, this is not an abstract philosophical concern but rather a description of model risk, a risk the profession is uniquely equipped to manage.

I see actuarial judgment in AI-augmented work playing out across two levels.

The first level is what I call deep technical judgment; the accumulated expertise that comes from years of working with models, understanding their failure modes and developing an instinct for when an output warrants skepticism. A model can produce a narrow confidence interval around a result that is fundamentally wrong, because the world has changed in ways the training data did not capture. Recognizing when to trust the model and when to override it is a skill that takes years to develop and, in my experience, cannot be extracted from a dataset, nor can it be delegated to AI to determine.

The credentialed actuary in the loop is certainly not there to second-guess the model’s arithmetic; rather, they are there to ask questions the model cannot answer about itself: Is this result reasonable given what we know about the underlying risk? Does this output create obligations we have not accounted for? Would we be comfortable explaining this to a regulator?

The second level is ethical judgment, addressed by Brian Christian in his 2020 book, “The Alignment Problem: Machine Learning and Human Values.” Christian explains that classical reinforcement learning assumes a fixed reward structure, whereas the alignment problem asks the reverse: how to design rewards that produce outcomes aligned with the objectives defined by organizations, regulators and key stakeholders.⁴ This connects to Goodhart’s Law, which states that when a measure becomes a target, it ceases to be a good measure. AI systems optimize proxies, and in insurance, the gap between the proxy and the underlying objective is precisely where unintended consequences can emerge.

The stakes of getting this wrong were anticipated before the age of machine learning (ML). In 1960, MIT’s Norbert Wiener, writing in “Some Moral and Technical Consequences of Automation,” issued a warning that today, reads as prophetic: “If we use, to achieve our purposes, a mechanical agency with whose operation we cannot efficiently interfere once we have started it, then we had better be quite sure that the purpose put into the machine is the purpose which we really desire and not merely a colorful imitation of it.”⁵

The “colorful imitation” of sound actuarial judgment—a model that optimizes a proxy metric while remaining blind to the values it was meant to serve—is precisely the failure mode the profession may want to be on guard against.

A model optimized to minimize claims costs may systematically underprice risks in communities with low credibility or biased experience data, not out of malice but because it infers risk from incomplete signals or proxies that fail to capture underlying exposure. A model optimized for retention may keep policyholders in products that are no longer appropriate for their circumstances. In each case, the model is doing exactly what it was designed to do, and the problem is inherent in the design—catching it requires someone with both the technical fluency to understand what the model is optimizing and the ethical formation to recognize why that objective is insufficient. In practice, actuarial work operates within established frameworks of model validation, peer review, and regulatory oversight designed to identify and mitigate these risks before they affect outcomes.

Alongside model risk sits AI governance more broadly. Regulators in jurisdictions including the European Union, the United Kingdom and Canada are actively developing frameworks for the oversight of algorithmic decision-making in financial services. The questions they are asking—around fairness, transparency, explainability, accountability and systemic risk— closely align with many principles familiar to actuarial practice.

Actuaries are bound by professional standards to act with integrity, exercise sound judgment, and consider the interests of policyholders and the public. In an AI-augmented environment, these obligations extend to how the profession engages with the technology itself. Where the model supplies the prediction, the actuary supplies the oversight needed to ensure those predictions are applied responsibly and in accordance with professional standards.

The professional designation or credential matters here precisely because it carries accountability. An algorithm cannot be sanctioned by a professional actuarial association, but the actuary who signs off on the work it informs can be. That accountability structure is not a simple bureaucratic formality; it is the mechanism by which public trust is maintained, and it is one that cannot be replicated by AI.

Credentials such as the CERA are particularly well-positioned for this moment. Enterprise risk management, by design, requires holding the whole system in view, from understanding how risks interact to how models can mislead, and how organizations should be structured to make sound decisions under uncertainty. These are precisely the capabilities an AI-augmented risk environment demands.

THE ‘AI’CTUARY

I believe the actuaries who will excel in the AI era are those who thrive in three areas that complement their technical credentials:

1. ML. A working understanding of ML is essential, not at the level of writing production code, but sufficient to interrogate a model’s assumptions, interpret its diagnostics and identify when its outputs warrant skepticism.
2. Communication. The ability to translate complex model behavior into language that resonates with executives, boards and regulators, making consequential decisions based on that output.
3. Intellectual curiosity. Intellectual curiosity about interrelated adjacent domains, such as data science, behavioral economics, climate science and legal and regulatory frameworks, because the risks of the next decade do not neatly fit within any single discipline.

AI is not arriving in a vacuum—it is entering a profession with over two centuries of accumulated expertise in measuring uncertainty, modeling complex systems and making defensible decisions under conditions of incomplete information. The alignment problem that philosophers and computer scientists are wrestling with—how to keep a powerful optimization system pointed at the right objective—has clear parallels in actuarial practice. We call it model validation. We call it peer review. We call it professional judgment.

If AI handles the prediction, the actuary’s energy may be best utilized by shifting toward what encapsulates it: the framing, the validation, the interpretation and the governance. This is not a diminishment of the role of the actuary; rather, it is an expansion into territory that is genuinely underserved.

Statements of fact and opinions expressed herein are those of the individual authors and are not necessarily those of the Society of Actuaries or the respective authors’ employers.

Copyright © 2026 by the Society of Actuaries, Chicago, Illinois.