How AI Is Improving Underwriting Accuracy in Commercial Lines

Commercial lines underwriting in India has historically relied on manual assessment of proposal forms, past claims data, and surveyor reports. This process, while thorough, introduces inconsistency. A 2025 IRDAI annual report highlighted that Indian non-life insurers collectively reported a net incurred claims ratio of approximately 85% across commercial segments, suggesting systemic mispricing.

The root cause is not negligence but cognitive limitation. A senior underwriter evaluating a manufacturing risk must synthesise fire survey reports, financial statements, loss histories, and industry benchmarks simultaneously. AI augments this process by identifying patterns across thousands of similar risks that no individual underwriter can retain.

Modern ML models for commercial underwriting ingest structured data — sum insured, industry code, location, claims history — alongside unstructured inputs like surveyor narratives and financial filings. Gradient-boosted decision trees and neural networks can identify non-obvious correlations: for instance, that textile units in certain Surat industrial estates with specific boiler configurations have 3x the fire loss frequency.

These models output risk scores that complement, rather than replace, underwriter judgement. The underwriter retains authority over pricing and terms, but the AI flags risks that warrant deeper scrutiny or merit preferential rates.

Early adopters among Indian insurers report measurable gains. Portfolios where AI-assisted risk scoring is deployed have shown 8-12% improvement in loss ratios within 18 months. The improvement comes from two sources: better risk selection (declining or re-pricing adverse risks) and more competitive pricing on good risks that previously received unnecessarily conservative quotes.

For SME commercial packages — a segment with historically thin underwriting data — AI models trained on aggregated industry data provide a baseline assessment that would otherwise require extensive manual investigation.

The most successful AI deployments in Indian insurance do not replace existing systems. Instead, they operate as a decision-support layer within the underwriting workbench. When a proposal arrives, the AI model generates a preliminary risk score, flags anomalies in the submission, and suggests comparable risks from the portfolio.

This approach minimises disruption. Underwriters continue using familiar tools — their policy administration system, rating engines, and reinsurance treaty structures — while gaining an additional analytical input. Training typically requires 2-4 weeks for underwriting teams to calibrate their trust in model outputs.

AI model accuracy is directly proportional to data quality. Indian insurers face particular challenges: inconsistent coding of industry types across branches, incomplete digitisation of legacy claims records, and variable quality of surveyor reports. Before deploying ML models, insurers must invest in data cleansing and standardisation.

Leading carriers are addressing this by implementing structured data capture at the point of proposal submission, mandating standardised surveyor report templates, and backfilling historical data from scanned documents using OCR and NLP pipelines.

IRDAI's 2024 guidelines on AI and ML in insurance require that algorithmic underwriting decisions remain explainable and auditable. Insurers must be able to demonstrate why a particular risk was declined or surcharged. Black-box models that cannot provide interpretable outputs face regulatory pushback.

This has driven adoption of interpretable ML techniques such as SHAP (SHapley Additive exPlanations) values, which quantify each feature's contribution to a risk score. Indian insurers deploying AI underwriting must maintain audit trails that satisfy both IRDAI examination requirements and internal compliance standards.