Cytora uses generative AI to automate underwriting risk digitization for insurers

Commercial insurers process thousands of risk submissions annually, yet the intake pipeline remains largely manual. Submissions arrive as PDFs, spreadsheets, emails, and broker API feeds — each with different schemas, field naming conventions, and attachment structures. Underwriters must search, parse, and re-key data across multiple systems before a single underwriting decision can be made. Compounding this, brokers asking semantically identical questions phrase them in structurally different ways, making automated matching to internal risk appetite criteria unreliable. The result: broker response times measured in days, inconsistent underwriting decisions, and underwriter capacity consumed by data wrangling rather than risk judgment.

Cytora built a SaaS platform on Google Cloud's Vertex AI — incorporating Gemini, PaLM 2, and Gecko text embeddings — to automate commercial insurance risk digitization end-to-end. The platform applies zero-shot predictions and Retrieval Augmented Generation (RAG) to begin making accurate extractions without pre-existing training data. A proprietary synthetic data generation engine addresses the cold-start problem, enabling fine-tuned, customer-specific private models from day one. Cytora's three-pillar architecture combines adaptable natural language understanding (via PaLM 2 fine-tuning on each insurer's private data), a risk taxonomy layer that maps extracted fields to standardized commercial lines schemas, and chain-of-thought prompting for multi-step reasoning across complex submission documents. A human-in-the-loop console captures underwriter corrections as continuous training signal. Vertex AI's enterprise privacy guarantees — where customer data remains the customer's property — were a prerequisite for deploying in this regulated environment.

Cytora's platform reduces broker response times from days to hours or minutes, a step-change that allows insurers to compete on responsiveness as well as price. Underwriters receive 'decision-ready risks' — submissions already parsed, classified, and mapped to internal risk appetite criteria — enabling consistent, auditable decisions at scale. Key outcomes include:

• Broker response time: reduced from days to hours or minutes
• Model accuracy: fine-tuning on PaLM 2 surpassed previous risk digitization performance benchmarks with significantly fewer training examples
• Time-to-production: Vertex AI's enterprise-ready infrastructure allowed Cytora to move from prototype to production generative AI workloads rapidly

Underwriter capacity is redeployed from data entry to judgment-intensive tasks, improving both throughput and portfolio quality.

• Chain-of-thought prompting is essential for insurance reasoning: multi-step logic in risk assessment — where one answer conditions the next — benefits materially from structured chain-of-thought techniques that also produce auditable outputs.
• Synthetic training data eliminates the cold-start barrier: insurers don't need large labeled datasets to begin fine-tuning; proprietary synthetic data generation can stand in until real corrections accumulate.
• Human-in-the-loop feedback loops compound over time: capturing underwriter corrections as training data turns routine quality control into a continuous accuracy improvement engine.
• Enterprise data ownership is non-negotiable in regulated industries: privacy and security guarantees must be contractually established before generative AI can move from prototype to production in insurance.