Cloud Services in Munich

Insurance / Reinsurance

A Munich-based global reinsurer (EUR 48B GWP, operating in 80 countries, 42,000 employees — providing reinsurance, primary insurance, and risk management services) needed cloud infrastructure to support actuarial analytics, catastrophe modelling, and AI-driven underwriting across its global operations. (1) Actuarial compute constraints: quarterly Solvency II internal model calculations required 200,000+ CPU-core-hours — running on an on-premises cluster of 8,000 cores that took 5 days per calculation. The actuarial team could only run the full model 4 times per year — meaning decisions between quarters relied on approximations. More frequent modelling would improve risk understanding and capital allocation — but on-premises compute could not support it. (2) Catastrophe modelling: the reinsurer used multiple catastrophe models (AIR, RMS, CoreLogic) to assess natural catastrophe exposure across its global portfolio. Full portfolio re-analysis required massive compute — each model provider running took 3-4 days on dedicated hardware. The reinsurer wanted to run catastrophe models after every significant event to understand its exposure in real-time — but on-premises infrastructure limited this to quarterly runs with ad-hoc runs for major events. (3) Data fragmentation: operational data was spread across 28 systems in 15 countries — policy data, claims data, financial data, and market data each in separate systems with different data models. Analytics required manual data assembly that consumed weeks of analyst time and produced results that were often outdated by the time they were ready. (4) AI/ML adoption: the reinsurer wanted to deploy AI-driven underwriting, claims analytics, and risk assessment — but lacked the cloud infrastructure to train and serve ML models at scale. On-premises GPU capacity was limited to 2 research machines — insufficient for production ML deployment across global operations. (5) BaFin compliance: as a German-regulated insurer, cloud adoption required VAIT compliance — access management, encryption, change management, and incident response meeting BaFin's requirements. The reinsurer's previous cloud initiative had stalled due to VAIT compliance concerns — the security team could not demonstrate that cloud met or exceeded on-premises security controls.

We designed and implemented a reinsurance analytics cloud platform over 16 weeks. (1) Actuarial compute cloud: elastic compute on AWS eu-central-1 — scaling from 2,000 baseline cores to 250,000+ cores for quarterly Solvency II calculations. Parallel execution: internal model calculations distributed across thousands of instances using Apache Spark — reducing runtime from 5 days to 8 hours. Frequency: monthly Solvency II model runs (from quarterly) — providing actuaries with updated risk metrics and capital figures every month. Cost optimisation: spot instances for interruptible calculation workloads — reducing compute costs by 65% compared to on-demand pricing. (2) Catastrophe modelling: cloud-hosted catastrophe model environments — AIR, RMS, and CoreLogic models running on cloud compute with licensed software deployed through approved vendor arrangements. Event-triggered analysis: automated catastrophe model runs triggered by significant natural events — full portfolio re-analysis completed within 24 hours of an event (from 3-4 days on-premises, if resources were available). What-if scenarios: unlimited capacity for scenario analysis — underwriters could model the impact of specific risks or portfolio changes without competing for scarce compute resources. (3) Enterprise data platform: Snowflake-based data warehouse consolidating data from all 28 operational systems across 15 countries. Automated data integration: event-driven data pipeline collecting and harmonising policy, claims, financial, and market data. Data governance: classification, lineage, and access control per DSGVO requirements — personal data identified and protected throughout the analytics pipeline. Analytics-ready: business users accessed harmonised, current data through SQL and BI tools — eliminating the weeks of manual data assembly that previously preceded every analysis. (4) ML platform: GPU compute for model training — NVIDIA A100 clusters for training underwriting, claims, and risk models. Model serving: dedicated inference infrastructure for production ML models — underwriting risk scoring, claims fraud detection, and portfolio optimisation models serving predictions in real-time. MLOps: automated model training, validation, deployment, and monitoring — ensuring models were current, accurate, and compliant with BaFin expectations for model risk management. Explainability: all ML models included explanations per BaFin MaRisk requirements — underwriters understood why the model recommended specific risk assessments. (5) VAIT compliance: comprehensive security architecture — encryption (AES-256 with customer-managed keys), MFA (hardware tokens for administrative access, software MFA for general access), network isolation (private subnets with no direct internet access for reinsurance data), logging (comprehensive audit trails for all data access and processing — maintained per BaFin retention requirements), incident response (automated detection and response procedures — meeting VAIT incident notification requirements), and change management (documented change procedures for cloud infrastructure changes — satisfying VAIT change management requirements). Betriebsrat: cloud adoption presented to and approved by the works council — addressing data protection concerns and change management.

Actuarial efficiency: Solvency II calculation from 5 days to 8 hours — enabling monthly modelling (from quarterly). Actuaries gained 4x more frequent risk metrics — improving capital allocation decisions and regulatory reporting accuracy. Capital optimisation: more frequent and accurate modelling identified EUR 850M in capital that could be released from conservative buffers while maintaining regulatory solvency — generating EUR 42M in investment income annually on the released capital. Catastrophe response: full portfolio re-analysis within 24 hours of a natural catastrophe event (from 3-4 days, if resources were available). During a major Mediterranean flooding event, the reinsurer provided clients with exposure estimates within 18 hours — significantly faster than competitors, strengthening client relationships. ML adoption: 12 ML models deployed in production within 18 months — underwriting risk scoring (improving combined ratio by 0.4 points on EUR 48B GWP = EUR 192M), claims fraud detection (identifying EUR 85M in previously undetected fraudulent claims), and portfolio optimisation (improving diversification benefit by EUR 120M in capital terms). Data platform: data assembly time from 3 weeks to same-day — analysts accessed harmonised, current data through self-service tools. 350 business users adopted the analytics platform — from 45 who previously had access to consolidated data through IT-prepared reports. BaFin compliance: VAIT examination passed with zero findings related to cloud — the examiner specifically noted the comprehensive security architecture and change management procedures. Betriebsrat approval: cloud adoption approved by the works council without objection — proactive engagement and transparent data protection measures built trust. Annual value: EUR 192M (underwriting improvement — 0.4 combined ratio points from ML-assisted risk selection), EUR 120M (capital optimisation — improved diversification and modelling accuracy releasing conservative capital buffers), EUR 85M (fraud detection — ML models identifying previously undetected fraudulent claims), EUR 42M (investment income — return on released capital), EUR 18M (operational efficiency — faster analytics, reduced infrastructure costs, and automated data processing) = EUR 457M. Development cost: EUR 1.2M. Payback: under 1 day.