Zurich Insurance

In Property & Casualty insurance, flood risk represents one of the most costly and unpredictable perils — yet traditional actuarial models assess risk at a portfolio level, not at the individual property level where intervention is most valuable. Zurich Insurance, serving millions of customers across the UK and globally, needed a way to predict flood claims weeks in advance so it could proactively engage at-risk policyholders rather than simply process claims after damage occurs. Manual risk assessment methods could not scale across the volume of individual properties in Zurich's book, and data science teams had no governed, repeatable path to move ML models into production — leaving AI initiatives siloed, slow to deploy, and difficult to audit.

Zurich built an end-to-end MLOps platform on Amazon SageMaker AI, delivered in just 8 weeks from prototype to production deployment. The platform ingests data from Snowflake alongside external geographic and weather datasets, with each property matched to a UK Unique Property Reference Number (UPRN) to enable precise, location-level risk scoring. Critically, personal identifiers — names, addresses, and UPRNs — are stripped before model training to ensure privacy compliance and reduce bias. The architecture spans three segregated AWS accounts: Tooling, PreProduction, and Production, each enforcing standardized workflows for model training, evaluation, approval, and continuous monitoring. This structure gives data scientists worldwide a consistent, auditable path to production, replacing ad hoc deployment practices with governed, reusable patterns applicable beyond a single use case.

The MLOps platform cut model deployment time from approximately 26 weeks to 8 weeks — a 69% reduction — by standardizing the infrastructure data scientists need to move from experimentation to production. Beyond speed, Zurich gained capabilities that were previously out of reach:

• Continuous property-level monitoring: Thousands of customer properties now have active flood risk scores updated as conditions change, something manual methods could not achieve at scale.
• Reusable infrastructure: The same SageMaker patterns built for flood prediction are now being extended to generative AI tools and customer interaction simplification across the global organization.
• Auditability and governance: Every model deployment follows a traceable approval workflow, meeting the compliance standards required in regulated financial services markets.

• Invest in MLOps infrastructure before scaling models: Zurich's 69% deployment speedup came from standardizing the path to production — not from the models themselves. The platform is the product.
• Privacy-by-design is a prerequisite in regulated industries: Stripping identifiers before training enables both compliance and broader data use — build this into the pipeline from day one.
• Geographic and third-party data unlock property-level precision: Integrating external weather and location datasets transformed a portfolio-level problem into an addressable, individual-property risk model.
• Reusable patterns compound value over time: Infrastructure built for one use case (flood) now accelerates every subsequent AI initiative across the organization.