Prompt Engineering as an Actuarial Standard of Practice (ASOP) • The Actuarial Club

We are no longer in the era of “experimentation.” By 2026, Generative AI (GenAI) has moved from the peripheral desks of innovation labs directly into the engine rooms of global insurers. Actuaries, traditionally the gatekeepers of numerical truth, are now find themselves orchestrating Large Language Models (LLMs) to parse massive datasets, draft regulatory commentaries, and even generate Python code for complex stochastic simulations. However, with this rapid adoption comes a silent, creeping risk: the reliability of the output is only as robust as the instruction that birthed it.

From Spreadsheet Models to Generative AI

For decades, the actuarial “source of truth” lived in Excel cells and SQL queries. Governance was straightforward—you audited the formula; you checked the code. Today, the “code” is natural language. When an actuary asks an AI to “analyze the trend in mid-tail motor bodily injury claims and suggest a weight for the latest accident year,” they are no longer just asking a question; they are performing a technical act. This shift from deterministic programming to probabilistic prompting represents the most significant change in actuarial methodology since the introduction of the personal computer.

Why Prompt Engineering Suddenly Matters

In a high-stakes environment like IFRS 17 reporting or Solvency II capital modeling, a minor linguistic ambiguity in a prompt can lead to a multi-million-dollar variance in a Best Estimate Liability ( $BEL$ ). Prompt engineering is the bridge between human intent and machine execution. If that bridge is poorly constructed, the resulting actuarial work product—no matter how polished it looks—is fundamentally flawed.

The Central Question: Can a Bad Prompt Become Professional Risk?

This brings us to a provocative crossroad: Is a prompt an informal query, or is it a formal actuarial work paper? If an actuary relies on an AI-generated reserve recommendation influenced by a “leading” or “biased” prompt, does that constitute professional negligence? As we look toward the future of the profession, the question isn’t whether we should use AI, but whether the way we talk to AI needs its own Actuarial Standard of Practice (ASOP).

“The prompt is the new script. In the hands of an actuary, a poorly constructed prompt isn’t just a typo; it’s a systemic risk to the balance sheet.”

Understanding Prompt Engineering in an Actuarial Context

What Is Prompt Engineering?

In the actuarial world, prompt engineering is the strategic design of inputs—instructions, context, constraints, and data—to elicit accurate, reliable, and auditable outputs from AI models. It is the art of “programming in prose.”

Why Prompts Are Not “Just Questions”

To a layperson, a prompt is a question. To an actuary, a prompt is a specification.

A question: “How do I calculate the Risk Adjustment?”
A prompt: “Act as a Lead IFRS 17 Actuary. Using the provided fulfillment cash flow data and a confidence level approach at the 75th percentile, calculate the Risk Adjustment for the Life segment, ensuring alignment with the Group’s accounting policy GAP-04. Provide the output in a structured table with a sensitivity analysis for a +/- 5% change in the underlying volatility.”

Prompt Engineering vs. Traditional Programming in Actuarial

Unlike C++ or Python, where the same input always yields the same output (deterministic), LLMs are stochastic. A prompt is a probabilistic nudge. This inherent uncertainty is why the governance of prompts is arguably more critical than the governance of traditional code.

AI as Calculator vs. AI as Analyst

We are transitioning from using AI as a Calculator (performing repetitive tasks) to using AI as an Analyst (interpreting trends). When AI acts as an analyst, it applies “judgment.” If that judgment is steered by a flawed prompt, the actuary’s professional independence is compromised.

How Actuaries Are Already Using Generative AI

Reserving

Actuaries use prompts to summarize thousands of claim notes, identifying “jumbo” claims or latent shifts in litigation patterns that standard triangles might miss.

Pricing

In personal lines, GenAI is prompted to suggest new rating factors based on non-traditional data sources (e.g., telematics descriptions or weather patterns) and to draft “Product Disclosure Statements” that meet regulatory clarity requirements.

IFRS 17

Prompting is used to generate the “Basis for Conclusion” documents, explaining why specific discount rates or mortality assumptions were chosen, ensuring they align with the technical standards.

Solvency II

GenAI assists in drafting the Regular Supervisory Report (RSR) and explaining movements in the Solvency Capital Requirement (SCR) to regulators.

Capital Modelling

Actuaries prompt models to design “black swan” scenarios for stress testing, asking the AI to find correlations between geopolitical events and equity market shocks.

When AI Outputs Influence Financial Reporting

AI-Generated Outputs in Actuarial Reports

When an AI-generated paragraph appears in an ORSA or a Board report, it carries the weight of the signing actuary’s credentials. The transition from “AI-assisted drafting” to “AI-influenced conclusion” is a thin line that many are already crossing.

Materiality and Financial Impact

If a prompt-induced error leads to a 2% change in a loss ratio, and that loss ratio governs a $1 billion book of business, the “prompt error” is a $20 million event. Under existing standards, any factor with a material impact must be documented and governed.

The Risk of Hidden AI Dependence

The greatest danger is “Shadow AI”—actuaries using personal prompts to reach professional conclusions without a record of what those prompts were. Without a “Prompt ASOP,” we have no way to audit the logic behind the AI’s contribution.

Where AI Is Already Entering Decision-Making

Assumption Setting: “Summarize the last 10 years of inflation data and suggest a prudent long-term CPI assumption.”
Model Documentation: “Explain the logic of this R-script for a non-technical auditor.”
Reserve Commentary: “Write the narrative for the Q3 reserving increase in the North American casualty book.”

Could Poor Prompting Become Professional Negligence?

Defining Professional Negligence in Actuarial Practice

Negligence occurs when an actuary fails to exercise the “standard of care” expected of a reasonably competent peer. If a peer would have caught a biased prompt, then the use of that prompt may be considered a breach of duty.

The Difference Between User Error and Governance Failure

A “typo” in a prompt is a user error. The lack of a process to review high-stakes prompts is a governance failure.

How Prompt Design Can Distort Outputs

The framing of a prompt can lead to “Anchoring Bias” in the AI. If an actuary prompts: “Explain why our current reserves are adequate,” the AI will find reasons to support that conclusion, ignoring contrary data. This is a direct violation of the actuarial principle of objectivity.

The “Reasonable Actuary” Standard in an AI Era

Does a “reasonable actuary” in 2026 check the temperature and Top-P settings of the model? Do they verify the prompt for “leading language”? If the answer is yes, then prompt engineering is no longer an optional skill—it is a professional requirement.

Table 1: Examples of High-Risk vs. Low-Risk Prompting

Feature	Poor Prompt (High Risk)	Professional Prompt (Governed)
Context	Vague or missing context.	Explicit role and background provided.
Bias	Leading questions: “Why are rates rising?”	Neutral framing: “Analyze factors affecting rate changes.”
Constraints	No limits on hallucinations.	“Strictly use the provided PDF data only.”
Output Format	Unstructured text.	Structured JSON or Table for validation.
Auditability	No record of the interaction.	Logged in a version-controlled prompt library.

Case Study: The “Leading” Reserve Analysis

Scenario: Incorrect Reserve Recommendation from a Faulty Prompt

What happened: A junior actuary at a mid-sized P&C firm used a GenAI tool to draft the initial commentary for the Workers’ Comp reserve review.

Root cause: The prompt was: “Draft a report showing that our current IBNR levels are sufficient given the 5% decrease in claim frequency.”

Prompt issue: The prompt was “leading.” It forced the AI to focus only on frequency and ignore a 12% increase in claim severity (medical inflation).

Financial consequences: The firm under-reserved by $15M, leading to a surprise hit in the following quarter.

Regulatory implications: The regulator flagged the lack of “conservative challenge” in the reserving process.

Lessons learned: Prompts must be designed to seek “all relevant factors,” not to confirm a pre-existing hypothesis.

Historical Lessons from Other Actuarial Failures

Spreadsheet Governance Failures

In the early 2000s, “manual overrides” in Excel led to massive reporting errors (e.g., the JPMorgan “London Whale”). The solution was formalized spreadsheet governance. Prompt engineering is currently in the “pre-governance” Wild West phase that spreadsheets once occupied.

Model Risk Management (MRM) Lessons

Standard MRM frameworks require that we understand the “inputs, engine, and outputs.” In GenAI, the “engine” is a black box, making the “inputs” (the prompts) the only thing we can truly control and audit.

The Evolution of Peer Review Standards

Just as we peer-review a valuation model, we must begin to peer-review the prompts used to interpret the results of that model.

Existing ASOPs and Professional Standards Relevant to AI

Existing ASOP Frameworks

ASOP 1 (Introductory): Defines “Actuarial Services.” AI prompting for a report clearly fits this definition.
ASOP 23 (Data Quality): If the prompt includes data snippets, does the AI maintain the integrity of that data?
ASOP 41 (Actuarial Communications): Requires clarity and disclosure of “reliance on others.” Is an LLM an “other”?
ASOP 56 (Modeling): This is the most critical. It defines a model as a “simplified representation of relationships.” A prompt creates a specific instance of an AI model’s behavior.

Where Current ASOPs Fall Short for AI

Current standards assume the actuary knows exactly how the underlying logic works (e.g., a formula). With AI, the actuary only knows the instruction they gave. Current ASOPs do not address the “non-deterministic” nature of AI—where the same prompt can yield different results on different days.

Why Prompt Engineering in Actuarial May Require Its Own Governance Framework

Prompts as Hidden Model Inputs

In traditional modeling, inputs are numbers. In GenAI, the prompt is the model configuration. If the prompt is hidden, the model is unauditable.

Non-Deterministic AI Outputs

The “drift” in AI models means a prompt that worked in January might fail in June. This requires a standard for re-validation.

Reproducibility Challenges

If a regulator asks, “How did you arrive at this conclusion?”, saying “I asked the AI” is insufficient. The actuary must be able to reproduce the result—which requires the exact prompt, the model version, and the system parameters.

Human vs. AI Decision Logic Prompt Engineering Actuarial

What Could an “AI Prompt ASOP” Look Like?

A future ASOP for Prompt Engineering in Actuarial might include:

Documentation Standards: Every material prompt must be logged.
Neutrality Requirement: Prompts must not be “leading” or biased toward a specific financial outcome.
Validation Protocols: High-risk outputs must be cross-checked against deterministic models.
Disclosure: Reports must state which sections were AI-generated and provide the prompt “logic” used.
Human-in-the-Loop (HITL): A mandatory requirement that no AI output can be used in a regulatory filing without a “Reasonable Actuary’s” signature.

Prompt Governance in Practice

Prompt Risk Classification Framework

Not all prompts are created equal. An enterprise governance framework should categorize them by risk:

Risk Level	Description	Governance Requirement
Low	Code debugging, formatting, or internal brainstorming.	None / Informal.
Medium	Drafting internal memos, summarizing research papers.	Self-review + Prompt logging.
High	Reserving commentary, Pricing assumptions, IFRS 17 narratives.	Peer review + Version control + Board disclosure.

Enterprise Prompt Libraries

Leading firms are already building “Golden Prompt Libraries”—pre-vetted, audited prompts that actuaries are required to use for specific tasks to ensure consistency and compliance.

Hallucinations, Bias, and AI Model Risk

Why Hallucinations Matter in Insurance

In actuarial work, a hallucination isn’t just a funny mistake; it’s a “false fact” in a legal document. An AI might “invent” a court ruling that justifies a lower reserve—a catastrophic risk for a CRO.

Overconfidence in AI Outputs

Actuaries are susceptible to Automation Bias—the tendency to trust a machine-generated chart more than a manual one, simply because it looks “cleaner.”

Regulatory and Legal Implications

The EU AI Act (2025/2026)

The EU AI Act classifies many insurance applications as “High Risk.” This mandates rigorous documentation, including how the AI was “steered”—in other words, the prompts.

Potential Litigation Risks

If an insurer denies a claim based on an AI-generated analysis, the “prompt” used to analyze the claim will be the first item requested by the plaintiff’s attorneys during discovery.

The Future of the AI-Enabled Actuary

Will Prompt Engineering in Actuarial Become a Core Skill?

By 2030, “Prompt Engineering for Actuaries” will likely be a mandatory module in the Fellowship exams (IFoA, SOA, CAS). The ability to communicate with machines will be as vital as the ability to communicate with stakeholders.

AI-Native Actuarial Teams

We will see the rise of the “Actuarial AI Architect”—a professional who doesn’t just do the math, but builds the linguistic frameworks for the AI to perform the math safely.

Counterarguments: Why Some Believe a New ASOP Is Unnecessary

Existing Standards May Already Apply

Some argue that ASOP 56 (Modeling) and ASOP 41 (Communication) are “principle-based” and already cover AI. They argue that adding more regulations will only stifle innovation.

The “Tool Not Decision-Maker” Argument

If the actuary takes final responsibility, why does it matter how they used the tool? (Counter: It matters because the tool can introduce biases that the actuary isn’t trained to detect).

Recommendations for Actuarial Organizations

Issue Practice Notes: Professional bodies should immediately release practice notes on “Responsible AI Prompting.”
Update Professionalism Courses: Include ethics modules on AI bias and hallucination management.
Collaborate with Tech: Work with OpenAI, Anthropic, and Google to create “Actuarial Mode” filters for LLMs.

Conclusion

Are Prompts Becoming Professional Workpapers?

The answer is a resounding yes. In an increasingly automated world, the prompt is the clearest record of an actuary’s intent and methodology. It is the “formula” of the 21st century.

Final Thoughts on AI Accountability

As actuaries, our primary value is not our ability to calculate—it is our judgment. If we outsource that judgment to an AI without governing the “conversation,” we risk losing our professional standing. Prompt engineering is not a “tech hack”; it is the new frontier of actuarial professionalism.

The question is no longer if we need an ASOP for prompts, but how quickly we can implement one to protect the public and the integrity of our financial systems.

Glossary of Terms

Prompt Engineering: The process of refining inputs to GenAI to achieve specific, high-quality results.
Stochastic Output: Results that are probabilistic rather than fixed; varying slightly each time.
Hallucination: An AI-generated output that is factually incorrect but presented as true.
Anchoring Bias: The tendency to rely too heavily on the first piece of information offered.
ASOP: Actuarial Standard of Practice.
HITL (Human-in-the-loop): A requirement for human intervention/validation in an AI process.

Advanced FAQs

Does ASOP 56 already cover prompts?

Technically, yes, if you define a prompt as a “model input.” However, ASOP 56 lacks specific guidance on the linguistic nuances and non-deterministic nature of prompts.
Can I be sued for a prompt?

If a prompt leads to a material financial misstatement and you didn’t follow a “standard of care” (like peer review), it could be cited in a negligence claim.
Should I include prompts in my actuarial report appendices?

For high-materiality work (e.g., reserve setting), including the “Master Prompt” ensures transparency and auditability.
How do I “audit” a prompt?

Audit for leading language, check for explicit constraints, and test the prompt multiple times to see the variance in outputs.
Is prompt engineering just a fad in actuarial?

No. As long as we use natural language to control AI, the quality of that language will be a technical competency.

Introduction

The Rapid Rise of AI and Prompt Engineering in Actuarial Work