Indian health insurer achieves 50% faster underwriting decisions with AI-powered risk classification

For a pan-India health insurance provider operating a network of 7,000+ hospitals and serving millions of customers across retail, group, and overseas medical policies, underwriting volume is substantial and time-sensitive. Non-STP (Straight-Through Processing) proposals — those requiring manual underwriter review due to medical complexity — represented a significant operational bottleneck. Underwriters were manually parsing unstructured medical documents to assess risk, a process prone to inconsistency and human error. The absence of automation produced inaccurate risk categorization, elevated loss ratios, delayed policy issuance, and capped the insurer's capacity to scale. High manual processing costs compounded the problem, limiting the business's ability to expand without proportionally growing headcount.

Turing built an AI-driven underwriting engine that automated the non-STP proposal workflow end to end. At its core, the system used Azure OpenAI's large language model to perform medical entity extraction — parsing unstructured clinical documents to identify and categorize diagnoses, conditions, and treatment history. An OCR pipeline digitized incoming paper and PDF records, feeding structured data into a risk classification engine that sorted proposals into Red, Amber, and Green tiers using codified underwriting heuristics. AI-assisted pricing models reduced reliance on manual actuarial judgment for standard cases. A Streamlit-based interface gave underwriters a single screen for document uploads, real-time risk summaries, and decision support — integrating into existing workflows without requiring a full system replacement.

The deployment cut underwriting processing time by 50%, directly accelerating policy issuance across the insurer's high-volume retail and group health books. The most significant operational shift was in proposal throughput: 70% of non-STP proposals are now handled automatically, redirecting underwriter capacity toward genuinely complex, high-risk cases that warrant human judgment. Risk classification accuracy improved by 60%, reducing the discrepancies that previously inflated loss ratios. Operational costs fell by 30% as manual document handling and review hours contracted sharply.

• 50% faster underwriting processing speed
• 70% of non-STP proposals fully automated
• 60% improvement in risk classification accuracy
• 30% reduction in operational costs

• LLM-based medical entity extraction is production-ready for insurance workflows — it handles the variability of unstructured clinical documents well enough to replace manual review for the majority of standard cases.
• Red/Amber/Green tiering provides a practical automation boundary — automating Green and most Amber cases while routing Red cases to underwriters preserves human oversight where it matters most.
• OCR and AI risk classification deliver compounding returns: speed, cost reduction, and accuracy improve simultaneously rather than trading off against each other.
• A lightweight UI matters for adoption — embedding AI outputs into a purpose-built underwriter interface, rather than forcing workflow changes, accelerates rollout and reduces resistance.