IRDAI's Principles for AI/ML in Insurance: What Underwriting Models Must Prove

What Are IRDAI's Core Principles for AI and ML Use in Insurance?

IRDAI released its AI/ML guidelines as part of the broader global convergence of insurance regulators on responsible AI. The Indian framework draws from:

• IAIS (International Association of Insurance Supervisors) principles on AI
• EU Solvency II model governance requirements
• GDPR-derived explainability obligations
• India-specific consumer protection priorities under IRDAI's mandate

Scope of application: The guidelines apply to any AI/ML model that:

• Influences premium calculation for individual policyholders
• Makes or contributes to claims decisions
• Determines coverage eligibility or exclusions
• Communicates personalised risk information to policyholders

What is NOT in scope:

• Pure actuarial tables (traditional statistical models are exempt from the AI principles, though the boundary is contested)
• Internal operational AI (fraud detection for internal processes, not policyholder decisions)
• Market-level pricing models that do not differentiate between individual policyholders

Enforcement timeline: IRDAI is conducting supervisory reviews of AI governance at major insurers in FY2026. Companies that cannot demonstrate compliance with the five principles are subject to remediation orders and, for repeat non-compliance, product approval delays.

How Do Explainability Requirements Apply to Underwriting Decisions?

The premium differential problem:

Life insurance AI models commonly produce premiums that vary by 20–40% based on AI-assessed risk factors. When a customer asks "why is my premium ₹18,000/year when my colleague pays ₹12,000?", the insurer must be able to answer.

Acceptable answer: "Your premium is higher because of three factors: (1) your BMI places you in a higher metabolic risk category (+₹3,200), (2) your occupation involves regular travel to high-risk geographies (+₹2,100), and (3) your family medical history includes two first-degree relatives with cardiac conditions (+₹700)."

Unacceptable answer: "Your premium reflects your individual risk assessment" — this is the current practice at most insurers and will not satisfy IRDAI's explainability requirements.

The counterfactual obligation: For life and health insurance, IRDAI guidance suggests that explanations should include counterfactual guidance where possible — "Your premium would be X if you achieved Y." This is most applicable to health insurance where behavioural changes can influence risk.

Technical requirements:

• Feature attribution for every AI-driven premium calculation, stored for 5 years
• Plain-language translation of feature attributions (not raw SHAP values)
• Customer-facing explanation portal accessible within 48 hours of request
• Audit trail confirming that explanations were generated and stored at the time of the decision (retroactive explanation generation is not sufficient)

Claims decisions: AI-assisted claims rejections face even higher explainability scrutiny. A claims rejection must explain: which specific policy exclusion applies, what evidence (from the AI analysis) triggered that exclusion, and what the policyholder could do to challenge the decision.

What Does a Compliant AI Governance Stack Look Like for Indian Insurers?

Component 1: Model inventory

IRDAI requires that insurers maintain an inventory of all AI models affecting policyholders, including:

• Model purpose and the decisions it influences
• Input features and data sources
• Validation date and next scheduled revalidation
• Named model owner and accountable committee
• Known limitations and monitoring in place

Component 2: Explainability infrastructure

For underwriting AI, the infrastructure must:

• Generate SHAP or equivalent feature attributions for every decision
• Map attributions to plain-language explanations using insurer-approved templates
• Store explanations with decision records for 5 years
• Provide an API endpoint that can serve explanations to customer-facing portals within 48 hours

For claims AI, higher standards apply — explanations must reference specific policy language, not just feature attributions.

Component 3: Fairness monitoring

IRDAI specifically prohibits using the following as AI features or proxies for features: religion, caste, gender (except where actuarially justified), and geography (except where directly relevant to risk — e.g., flood zone pricing). Monitoring must demonstrate that outcomes are equitable across these groups.

Component 4: Drift monitoring

Insurance models face a specific drift challenge: the underlying risk environment changes with regulatory policy, climate patterns, and health epidemics. Models must be revalidated when drift is detected, with a maximum revalidation cycle of 24 months even without detected drift.

Component 5: Human review escalation

High-value or disputed decisions must route to human review. The AI governance framework must define the thresholds (premium above X, claim value above Y, complaint received) and document the human review process.

Claris's Glass Box Suite covers all five components through a single integration, with pre-built IRDAI compliance templates for each reporting requirement.