The International Actuarial Association announced two artificial intelligence papers on November 8, a risk framework and a prompting guide, and said they complete its AI Task Force's series of six.

A Framework for Considering AI Risk and Prompting for Actuaries are both dated November 2026 and appear on the association's papers page under November 7. The announcement said both are meant to be read alongside the IAA's Artificial Intelligence Governance Framework, which the same page lists on December 1, 2025.

They arrive as the NAIC's Big Data and Artificial Intelligence (H) Working Group has adoption of version 7.0 of its AI Risk Evaluation Supplement on its agenda for November 16.

Key Takeaways

  • Hidden proxies and hallucinations are among 10 third-level risks, each with an example such as a motor model inferring gender from braking intensity.
  • A home quoting tool whose reward mechanism favours sales could drift into "increasingly generous pricing" over time, one of 18 worked examples in the risk paper.
  • The sample pricing prompt names five triggers for human review and tells the model to label every figure as observed, modelled, assumed, forecast or AI-written.
  • The same prompt can return different answers at different times, the guide says, so critical prompts should be version-controlled, validated and peer reviewed.

The three levels of the risk framework

Most AI risks are familiar to actuaries, the risk paper says. It uses the OECD definition of an AI model, which it calls "very broad and captures many traditional actuarial models." AI makes the risks harder to manage because the models tend to be more complex, adapt to their operating environment, react sharply to inputs, or cannot reproduce results.

Table 1 of the paper sets out six level 1 categories, nine level 2 risks and 10 level 3 risks. Model and data risk holds five of the nine second-level risks and all 10 at the third level.

Level 1 categoryLevel 2 risksLevel 3 risks
StrategicNoneNone
Compliance and regulatoryNoneNone
Model and dataSecurity; privacy; transparency; bias and discrimination; errorTechnological integrity; consent; confidentiality; aggregated data; model bias introduced by data; model bias introduced by training process; hidden proxies; correlation and causation; inability to react to changing data or external conditions; hallucinations
ImplementationComprehensibility; model suitabilityNone
MonitoringRobustness; model evolutionNone
AccountabilityNoneNone

Source: IAA AI Task Force, A Framework for Considering AI Risk, Table 1, November 2026.

Table 2 then gives an example for each risk at its lowest level, 18 in all, which the paper calls "both real and hypothetical."

Several come from pricing. Under bias from the training process, a regional insurer trains a car insurance model on its own 20 years of claims history and uses it to price a broader market, where the model misprices demographics the data never held. Under hidden proxies, a model trained on motor data infers gender from braking intensity or the number of claim-free years.

Both monitoring examples use one system: a home insurance quoting tool that writes new pricing rules from continuously updated claims data. A misaligned reward mechanism that favours sales could lead it to offer "increasingly generous pricing to prospective customers over time," the paper says. Monthly reviews of quoted prices may then be too slow to catch discriminatory rules that develop in between.

Under confidentiality, an actuary pastes client information into an off-the-shelf large language model, in breach of company policy or client terms. Some tools train on user inputs, the paper notes, so the information could surface in answers to other users.

Three uses are suggested: a basis for a risk taxonomy within enterprise risk management, "a checklist to assess the robustness of existing model risk management frameworks," and a starting guide for member associations' professional standards. Mitigation is outside the paper's scope; for that it points readers to the task force's case study repository on GitHub.

What the prompting guide tells actuaries to do

Written for "the actuary new to prompt engineering," the second paper breaks a prompt into four components: instruction, context, examples and technical constraints. It then separates six prompt objectives, such as extraction and analysis, from 10 techniques, running from zero-shot and few-shot prompting through chain-of-thought and prompt chaining to meta prompting and tree-of-thought.

Appendix A demonstrates the techniques on a synthetic auto insurance claims dataset from Kaggle. The prompt chain runs to four steps: extract the losses, average them excluding zero records, replace missing values with that average, then recalculate. The role-based example reads: "You are an actuary, calculate the loss ratio using the losses and premiums in the data."

Arithmetic gets a blunt warning. Language models "should not be used for calculations without extensive checking of results," it says, in a list of eight pitfalls that also covers chat history changing how a prompt is read, hidden memory files personalising outputs over time, and models ignoring a requested output format.

Reproducibility is treated as a professional issue. "Due to the stochastic nature of LLMs, the same prompt can produce different responses at different times," the paper says, and it asks actuaries to test and document that a prompt gives consistent responses before using it in professional work. Its documentation protocol is version control and validation of prompt templates, with peer review of critical prompts.

Appendix B is a block of text to append to a prompt, written for a report on insurance profitability and premium rates by policyholder type. It has four parts: 10 "do" instructions, seven "do not" instructions, five triggers for human review and a housekeeping line.

Among the instructions: reconcile reported totals to the approved control total; state which measure "profitability" or "premium" means, including written or earned and gross or net; and name the possibility where a rating variable could stand in for a protected characteristic.

On confidentiality the block says: "Do not send policy-level or special-category data to any gen AI environment not approved for that data." Its housekeeping line asks the user to record the block version, the model version and the date against any output used downstream.

Such wording "should not be relied on to ensure an ethical outcome," the appendix says, particularly where the data has problems. Responsibility stays where it was, in the guide's words: "the actuary remains professionally responsible for the work product even if it is produced by (or involving) an AI model."

The six-paper series and the NAIC Supplement

Neither paper binds anyone. Each carries the IAA's statement that it "is not an International Standard of Actuarial Practice (ISAP), nor does it set standards or requirements." Both were written by drafting groups of the AI Task Force, which exists to deliver the statement of intent on AI that the IAA Council adopted on March 8, 2024.

The IAA papers page lists four earlier task force papers. The Artificial Intelligence Governance Framework, Testing of Artificial Intelligence Models or Systems and Documentation of Artificial Intelligence Models or Systems are listed on December 1, 2025. Professional Considerations for Actuaries Interacting with Artificial Intelligence followed on October 1, 2026.

The Governance Framework is organised around the life of a system: design, development, approval, implementation and ongoing monitoring. Two of the risk paper's six categories, implementation and monitoring, carry the names of those stages. For bias and fairness the risk paper sends readers back to section 3.2.1 of the framework.

Some of the same items are in the Supplement the NAIC working group takes up on Monday, where agenda item 4 is "Consider Adoption of the AI Risk Evaluation Supplement v7.0." Exhibit C of the clean text asks a company to discuss "testing for model drift, accuracy, unfair trade practices, unfair discrimination, performance degradation" for each high-risk model. The IAA's model evolution risk covers drift and also a system that "learns new rules over time that, while not necessarily wrong, were not intended." Hidden proxies are a named risk in the IAA paper; version 7.0 dropped the previous draft's question on proxy testing, the site reported last week.

One piece of the IAA's own standard-setting was open when the series began. The Governance Framework said the association was considering amendments to ISAP 1, its model standard on general actuarial practice, and that the parts relevant to AI governance "may change in the near future."

Further Reading

Sources

  1. International Actuarial Association, "IAA Releases Two Educational Papers on Artificial Intelligence," news release, 8 November 2026
  2. IAA AI Task Force, "A Framework for Considering AI Risk," IAA Paper, November 2026
  3. IAA AI Task Force, "Prompting for Actuaries," IAA Paper, November 2026
  4. International Actuarial Association, Papers (listing with publication dates)
  5. IAA AI Task Force, "Artificial Intelligence Governance Framework," IAA Paper, November 2025
  6. International Actuarial Association, news release on "Professional Considerations for Actuaries Interacting with Artificial Intelligence," 6 October 2026
  7. IAA AI Task Force, Actuarial AI Case Studies repository
  8. NAIC, Big Data and Artificial Intelligence (H) Working Group, agenda for 16 November 2026, Fall National Meeting, Dallas (draft dated 4 November 2026)
  9. NAIC, AI Risk Evaluation Supplement v7.0 (clean)