1.1 Business Objective Definition
1.2 Use Case Assessment
1.3 Regulatory & Ethical Considerations
1.1 Business Objective Definition
Align AI to clearly defined business needs and outcomes
→ Start by aligning AI system capabilities with business needs and clearly defined intended outcomes; actuaries are encouraged to collaborate with development teams early to define limitations and intended use across the lifecycle.
References
G. 2 & 3; D. 3; T. 2
Calibrate governance to model risk rating (high/medium/low)
→ Use the model risk rating methodology (high/medium/low) based on criteria such as adverse financial impact, complexity, impact on reserving, solvency, pricing, underwriting, claims processes and fair treatment of beneficiaries to determine the intensity of governance and oversight.
References
G.1, 2, 3; D. 1; T. 4
Design for fairness, robustness, transparency from the outset
→ Recognize that the governance framework covers design, development, implementation, ongoing monitoring and retirement of AI systems; objectives should be framed so that fairness, safety, robustness and regulatory compliance can be monitored over time;
References
G. 2, 3; D. 1, 2, 3, 4; T. 1, 2, 3, 4, 5, 6, 7
→ Consider early how transparency and explainability requirements will influence choice of model type (transparent vs explainable vs opaque such as LLMs), recognising that this determines the feasible testing, validation and monitoring approaches.
References
G. 2, 3; d. 3, 5; T. 2, 4;
→ Document the business purpose, target population, intended use and high‑level limitations as part of a model overview, forming the front of the model card.
References
G. 2, 3; D.1; T. 3, 4
→ Apply the principle of proportionality when setting expectations for documentation and governance – depth should scale with significance, risk and complexity of the AI system.
References
G. 3; T. 2; D. 3, 5;
1.2 Use Case Assessment
Assess feasibility, data, and limitations before committing
→ In designing an AI system, first test whether capabilities actually match the actuarial business need; actuaries are encouraged to engage with development teams early to identify limitations and define intended outcomes across the model lifecycle.
References
G. 3; D. 3; T. 2, 4, 6
Explicitly consider fairness and discrimination risks for each use case
→ Evaluate data suitability – robustness (accuracy, diversity, completeness, resilience to noise, timeliness), privacy and security – as these are pre‑conditions for a viable AI use case.
References
G. 2, 3; D. 3, 6; T. 2, 3
Decide early if AI (vs traditional models) is appropriate for the context
→ Consider the transparency needs of stakeholders: transparent models (e.g. GLM) may be more appropriate than opaque models (e.g. LLMs) in contexts needing strong explainability or customer‑facing decisions.
References
G. 3; D. 3; T. 2, 4, 6
→ Assess potential for bias, fairness and discrimination early, including whether protected characteristics could be directly or indirectly embedded in data or features and how dynamic feedback loops might change model behaviour over time.
References
G. 2, 3; D. 2, 3; T. 3, 4, 5, 6
→ Reflect on whether AI is necessary versus traditional modelling; the Governance paper notes that AI brings particular risks (opacity, dynamic learning, data sensitivity) and may require more frequent validation and monitoring than traditional actuarial models.
References
G. 2, 3; D. 3; T 2, 4
→ Record assumptions about scope (e.g. line of business, geography, product), noting that models developed for a specific product or demographic may not generalize to other products or markets.
References
G. 2, 3; D 3; 6 T. 2, 6
1.3 Regulatory & Ethical Considerations
Embed regulatory expectations (AI, data, insurance) into design
→ Build AI governance framework by leveraging existing governance frameworks to meet emerging regulatory requirements in AI around data governance, model governance, testing and validation in their region and sector.
References
G. 2, 3
Address bias, discrimination, and fairness as core design constraints
→ Follow applicable regulatory requirements or guidance (e.g. EU AI Act, Monetary Authority of Singapore’s FEAT principles (Fairness, Ethics, Accountability and Transparency)
References
G. 2, 3; D. 2, 3; T. 2, 5
Connect with external principles such as FEAT and OECD AI guidance
→ Be mindful of bias, fairness and discrimination, underlining that harmful bias can arise from data, model design and system context
References
G. 3; D. 2; T. 4, 5, 6
→ Be mindful that fairness is context‑dependent, influenced by legal norms and societal expectations (e.g., fairness in insurance pricing), so one must interpret fairness in light of local market practice and regulation.
References
G. 3 & Appendix; D. 3, 6; T. 3, 4, 6
→ Document compliance with relevant data laws, actuarial standards and entity policies, and explicitly address ethical implications of data usage, bias and fairness in both data and model documentation.
References
G. 2, 3, 4; D. 2, 3, 6; T. 3, 5, 7