Insurance is the most AI-intensive sector in BFSI after banking. Actuarial models have incorporated statistical methods for decades. AI augments these with machine learning on behavioral and sensor data, LLM-assisted claims processing, and real-time underwriting engines that combine hundreds of variables from multiple data sources. The governance frameworks for each of these AI use cases have distinct requirements — and insurers who treat them as one problem consistently under-govern at least one of them.
The governance gap in insurance AI is structural: insurers often apply their actuarial governance framework to AI (actuarial certification plus internal review) without applying model risk management standards, or they apply banking-derived model risk standards without addressing the insurance-specific dimensions — actuarial justification standards, IRDAI regulation, health data obligations under the DPDP Act, and the tension between permitted actuarial differentiation and proxy discrimination in rating factors. We have seen both failure modes in the same organization, on different product lines, at the same time.
Six Insurance AI Use Cases and What Governs Each
Actuarial AI (mortality, morbidity, claims frequency modeling). AI-augmented actuarial models blend traditional actuarial tables with machine learning on behavioral, device, and third-party data. Governance requires: actuarial professional standards compliance, where the model is reviewed by a qualified actuary (typically a Fellow of the Institute of Actuaries of India) who attests to its actuarial soundness; model risk management — independent validation, documentation, monitoring; and proxy variable analysis, because behavioral features that predict mortality or morbidity may correlate with religion, caste, or socioeconomic status in India. When actuarial justification and proxy discrimination analysis are in tension, governance must document the resolution and get oversight committee approval — not quietly let the actuarial sign-off settle the fairness question too.
Underwriting AI (eligibility and pricing). AI underwriting engines score applications for eligibility and premium pricing. Governance requires: permitted rating factor compliance, where features must be actuarially justified and filed under IRDAI's product approval requirements; proxy variable analysis on geographic rating features, since pincode and district-level features may capture genuine actuarial risk (claims frequency geography) or caste/religion demographic concentration; explainability for customer refusal or loading, so customers can get an explanation of underwriting decisions that affect them; and fairness monitoring of differential approval rates and premium loading across India proxy dimensions, with escalation triggers when differentials exceed appetite thresholds.
Claims processing AI (validity assessment and amount estimation). AI claims systems assess validity and estimate payable amounts. Governance requires accuracy monitoring by claim type and customer segment — aggregate accuracy masking subgroup errors is a real governance gap we see repeatedly; fairness monitoring for claims fraud AI, since a model that flags claims from certain communities at systematically higher rates creates discriminatory claims-handling exposure; explainability adequate for grievance resolution when a customer disputes a decision; and a defined human review pathway for materially contested claims, not just an AI reconsideration loop.
LLM customer service and claims guidance. LLM chatbots handle policy queries, claims guidance, and agent-assisted processing. Governance requires an insurance-advice boundary — analogous to banking's financial-advice boundary — so the LLM does not provide individualized coverage recommendations or claims-strategy advice beyond factual policy information; claims-process guidance that accurately reflects the insurer's current process, since incorrect guidance that causes a customer to miss a claims deadline creates real customer harm; oversight extended to any intermediary or agent channel through which the LLM is deployed; and protection for health information processed in claims conversations.
Telematics and behavioral insurance AI. Usage-based insurance (UBI) and telematics AI use device-collected behavioral data — driving patterns, fitness tracker data, smartphone usage — for underwriting and pricing. Governance requires a documented consent basis for using this data in underwriting, consistent with DPDP Act standards (specific, informed, freely given, and revocable); data minimization, so the AI uses only the behavioral data necessary for the underwriting task; proxy variable analysis, because device usage patterns may proxy socioeconomic status or disability status; and transparency so customers understand how their behavioral data affects their pricing.
Third-party data AI (social media, credit bureau, public records). Insurers increasingly use third-party data in underwriting AI. Governance requires purpose-limitation discipline — data obtained for one purpose, such as credit assessment, cannot be repurposed for insurance underwriting without a fresh lawful basis; due diligence on data source accuracy, recency, and discriminatory-proxy potential; proxy discrimination analysis, since social media and public records data are particularly likely to carry India-specific demographic correlations; and customer transparency about which data sources fed an underwriting decision that affected them.
Banking AI Governance vs. Insurance AI Governance: Where They Diverge
| Governance dimension | Banking AI | Insurance AI | Governance implication |
|---|---|---|---|
| Fairness — permitted differentials | Demographic differentials in credit AI are treated as suspect by default; an adverse impact ratio below 0.75 (the US-origin "four-fifths rule" threshold many Indian banks have adopted as an internal benchmark, not an RBI-mandated figure) typically triggers remediation | Actuarially justified risk differentials are permitted; geographic and health-based pricing differentials are allowed where an actuarial basis is documented | Insurance governance must assess whether a differential is actuarially justified or a proxy for discrimination; the banking fairness framework cannot be applied directly |
| Explainability standard | Per-decision local explainability expected for credit AI; plain-language explanation for credit rejection under the Fair Practices Code | Underwriting decision explanation required; claims decision explanation required for disputes; actuarial models need actuarial-standard documentation alongside any customer-facing explanation | Insurance explainability blends actuarial documentation requirements with customer-facing explanation obligations |
| Regulatory anchor | RBI's FREE-AI recommendations (2025); SR 11-7-style model risk practice | IRDAI product filing and conduct regulations; actuarial professional standards (Institute of Actuaries of India); DPDP Act; Insurance Act, 1938 obligations | Insurance governance maps to IRDAI, not RBI; actuarial certification requirements coexist with, rather than substitute for, model risk validation |
| Data governance | DPDP Act for customer financial data; RBI's 2018 circular restricts storage of payment system data to India | DPDP Act applies the same single consent/purpose-limitation framework to health data as to other personal data — there is no separate "sensitive data" statutory tier for insurers to rely on; consent requirements for behavioral/telematics data; purpose limitation for third-party data |
India-Specific Insurance AI Governance
India proxy variable analysis in insurance rating. The proxy variable challenge in India insurance AI is more complex than in banking AI. In banking, all demographic proxies are suspect in credit decisions. In insurance, some geographic and demographic correlations are actuarially justified — claims frequency genuinely varies by geography, and age is a legitimate health insurance rating factor. Governance must distinguish geographic variables capturing genuine actuarial risk (flood zone, hospital density, local claims frequency) from geographic variables that primarily capture caste or religion demographic concentration. The hard case — and the one most insurers skip — is when the correlation with claims frequency is real but itself reflects social discrimination in healthcare access rather than independent actuarial risk. Governance documentation has to address this at the feature level, not assert blanket actuarial justification for the whole geographic variable class.
Pradhan Mantri Fasal Bima Yojana (PMFBY) agricultural insurance AI. Agricultural insurance under PMFBY and allied schemes uses AI for crop loss assessment — satellite imagery analysis, weather data modeling, claims verification. The fairness question here is different from retail insurance fairness: the concern is accuracy and timeliness of AI-based crop loss assessment across regions, crops, and smallholder farmer profiles, not demographic fairness in underwriting. An AI system that systematically underestimates crop losses in certain geographies creates direct financial harm to farmers who paid premiums and received an inadequate settlement. Governance of agricultural insurance AI should include accuracy monitoring by geography, crop type, and landholding category — and insurers running PMFBY books should expect this to be the first thing a government audit asks for.
DPDP Act obligations for insurance AI processing health data. Health insurance AI processes health data as personal data under the DPDP Act, 2023. Insurers using health data in AI underwriting models need a consent architecture that tracks consent at the data-type and use-case level, not just at the policy level — an AI underwriting model trained on health claims data without a documented basis for that specific use has a real compliance gap, with or without a formal "sensitive data" label attached to it.
Insurance AI Governance Action Checklist
- Stop citing "sensitive personal data" protections for health data under DPDP Act in policy documents — rewrite that language to reflect the Act's actual consent and purpose-limitation obligations, which apply uniformly.
- For every rating factor with a geographic or demographic footprint, document separately whether it is actuarially justified, a proxy risk, or both — do not let actuarial sign-off double as the fairness sign-off.
- Build a consent-tracking architecture at the data-type and use-case level for any model trained on health or behavioral data; policy-level consent is not sufficient evidence.
- For agricultural insurance AI, add accuracy monitoring broken out by geography, crop type, and landholding category before the next claims season, not after a dispute.
- Extend LLM claims-guidance systems' advice-boundary controls to any intermediary or agent channel through which they are deployed — oversight that stops at the insurer's own app misses where most claims conversations actually happen.
Related: BFSI AI governance beyond banking
Insurance AI governance — actuarial model documentation, IRDAI-mapped regulatory evidence, India proxy variable analysis with actuarial-justification annotation, DPDP Act health data consent tracking
Zytra supports insurance-specific AI governance requirements alongside the core BFSI governance framework: actuarial AI documentation, fairness testing with permitted-vs.-proxy annotation, IRDAI regulatory evidence, and DPDP Act health data consent management. India proxy variable analysis includes geographic rating factor proxy discrimination assessment.