Cytora

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.