Artificial intelligence is beginning to influence decisions that can shape an insurer's financial results for years: which risks are accepted; how they are priced; the claims that receive scrutiny; the assumptions that feed reserves; and the deployment of capital.
However, despite its capabilities, an AI system can be wrong. If that error is not identified, it can be repeated across thousands of decisions before anyone can see the pattern. When the problem finally emerges, the firm may not be able to reconstruct exactly how and where the AI system influenced each outcome.
Tracing that decision-making stream is a new kind of governance challenge. Approval systems are important, but they only confirm that the approved process was followed, still leaving the cumulative steps opaque. What leaders increasingly need is transparent evidence of what actually happened to produce the outcome.
We call the capability to identify the steps taken to produce the outcome — and to reproduce them so the result is reliably the same each time — legibility. For AI, legibility is the ability to reconstruct how AI influenced a consequential decision, including the model and data involved, the system’s recommendations, the human or control intervention that occurred, and the ultimate outcome. Legibility turns an opaque AI-enabled process into evidence that can be challenged, measured, and actioned.
As AI moves deeper into underwriting, pricing, reserving, claims, credit, investment, compliance, financial crime, and capital decisions, legibility is likely to become an increasingly important part of risk management. It is also an area in which the actuarial profession has an opportunity to contribute more deliberately.
Actuaries do not currently lead AI governance in most organizations, nor should they replace technology, risk, legal, compliance, data science, cybersecurity, internal audit, or business leaders in that work. But actuaries already work where models, uncertainty, controls, data, and consequential financial decisions intersect. As AI governance increasingly needs to support measurement as well as control, that expertise becomes highly relevant.
AI governance must go beyond policies to AI legibility
Most companies have some form of AI governance. They maintain policies, review use cases, validate models, establish committees, and assign senior accountability. These measures are necessary, but they do not always create enough evidence to understand the exposure that remains after the controls are applied.
Consider an external model used to identify claims for further investigation. The formal claims decision may still be made by an employee. However, if the model determines the files that receive attention and the reviewer has limited information with which to assess the recommendation, the AI system is already shaping outcomes across the claims process.
Now imagine that the model changes over time and a pattern of disputed decisions later emerges. The firm may know the model was approved and that employees made the final decisions. However, can it identify which model version influenced each claim? What information did the model see? How often did reviewers disagree with it? Which other claims may have been affected by the same issue?
Those questions determine whether the organization has an isolated problem or an accumulating exposure. Governance tells the firm that controls existed. Legibility helps determine what happened, how far the problem may have spread, and what needs to be remediated.
Once the questions become “how often might it happen?” and “how large could this be?,” the problem begins to look distinctly actuarial.
Why governance must evolve as AI scales
The urgency is partly driven by the pace of adoption. The Oliver Wyman Forum's 2026 CEO Agenda survey, conducted with the New York Stock Exchange, found that organizations further along in scaling AI were roughly three times as likely to report that their investment was meeting or exceeding expectations as organizations still at an early stage of adoption. Nearly a quarter of respondents said AI had not yet affected revenue.
The relationship between deployment scale and ROI creates strong pressure to move from experimentation to full adoption. However, governance models designed for pilots can become strained when AI expands into consequential functions, connects to third-party systems, or becomes embedded in complex workflows.
At that point, exposure is no longer limited to whether a model produces an incorrect output that can be corrected quickly. The more important questions are how far the output can propagate, how long an error can remain undetected, how many decisions it can affect, and whether multiple business units — or even multiple firms — depend on the same model, data source, vendor, or assumption.
These are questions of frequency, severity, accumulation, dependency, discovery lag, and tail risk. They are questions actuaries have long been trained to address in other forms of financial risk.
Regulators are also wrestling with how AI should be governed. Approaches vary around the globe, but the direction is increasingly clear: Firms need to understand how AI is being used, who is accountable for it, how its performance is monitored, and how it affects consequential decisions. When firms cannot do that, problems can persist longer, remediation becomes harder, and uncertainty around losses, capital, and risk transfer grows.
The actuarial opportunity in AI governance
The case for actuarial involvement does not rest on the claim that actuaries already own AI governance. They generally do not. The stronger case is that the next phase of AI governance will require capabilities that are familiar to the profession.
Actuaries routinely assess the reasonableness of reserves, develop and review insurance rates, evaluate models and assumptions, assess data quality, quantify uncertainty, stress test scenarios, and communicate conclusions that boards, regulators, investors, and management teams rely on. They are accustomed to working with imperfect information while still forming a disciplined view of financial exposure.
Existing actuarial standards also point in this direction. ASOP No. 56, Modeling, addresses actuarial work involving models, including circumstances involving models developed by others. ASOP No. 23, Data Quality, addresses the selection, review, and use of data and reliance on data supplied by others. ASOP No. 41, Actuarial Communications, addresses the communication of actuarial findings and their bases.
AI complicates the application of those disciplines because an actuary may increasingly sit downstream of an AI-enabled decision. Data reaching an actuarial model may already have been selected, classified, augmented, or filtered by AI. An underwriting or claims process may produce experience data that already reflects AI recommendations. If that influence cannot be reconstructed, understanding the basis for the information — and the uncertainty around it — becomes harder.
Better legibility therefore supports existing actuarial disciplines while also creating a new opportunity: Actuaries can help translate AI governance evidence into a view of exposure, control effectiveness, uncertainty, and financial consequences.
Five dimensions are critical to AI legibility
Rather than standardizing rapidly changing AI technology itself, firms can standardize the minimum evidence needed to understand how consequential AI-enabled decisions are made. Five dimensions are especially important to this effort.
- Attributability: Can the organization establish what was decided, on what basis, using which model and data, and by whom or what? Without attribution, credible exposure and loss experience are difficult to develop.
- Detectability: How quickly would the organization know that a model had changed, drifted, degraded, or been replaced? The answer shapes discovery lag and the amount of exposure that can accumulate before intervention.
- Bounded severity: How bad could a failure become before controls stop it? Scenario analysis and stress testing can help estimate plausible financial consequences and establish tolerances around them.
- Independent challenge: What checks are genuinely capable of identifying and overturning a flawed recommendation? Human review is meaningful only if the reviewer has enough information and authority to challenge the system.
- Demonstrability: Can the organization produce the evidence supporting an actual consequential decision on request? A governance process is more credible when it can demonstrate how a real decision was reached, not merely describe the policy that was supposed to govern it.
Together, these dimensions make AI risk more containable, measurable, and defensible. They can help firms trace affected decisions, estimate the scope of an exposure, target remediation, set limits, assess control effectiveness, inform capital decisions, and evaluate whether a risk can ultimately be transferred.
This is where actuarial participation can add value. The contribution is not to own the technology or replace existing governance functions. Rather, it is to help convert the evidence produced by governance into questions management can answer: What is the exposure unit? Where could losses accumulate? What assumptions matter most? How quickly would a failure be detected? What is the plausible range of financial outcomes? What controls materially reduce that range?
Leaders can apply a practical test for AI governance
A useful starting point is not another policy review. Select a small number of consequential AI-influenced decisions and try to reconstruct them end to end.
Can the firm identify the model and version involved, the relevant data, the recommendation, the human intervention, subsequent model changes, and other decisions affected by the same issue? Can it estimate how many decisions might be affected if the model were wrong? Can it determine how quickly the problem would be detected and what evidence would be available for remediation?
If the organization cannot answer these questions, it may have governance documentation without sufficient risk evidence.
Once the answers to these questions are understood, boards and management teams can decide where stronger controls, monitoring, evidence retention, or independent challenge are needed. As these questions become more quantitative and more closely tied to financial consequences, actuarial expertise can become a useful complement to the functions already responsible for AI governance.
The actuarial profession can help organizations move from asking only whether AI is governed to asking whether the resulting risk is understood. Actuaries should not lead every aspect of AI governance. However, the profession's existing disciplines are uniquely positioned to help firms build the measurement layer that can help them quantify, understand, and remediate risks arising from AI use. This capability is increasingly critical as AI is used even more heavily to inform consequential decisions.
The time to make AI-related exposure legible is before loss experience makes the need unavoidable. When was the last time you had a conversation with your actuary?