Zurich Insurance UK

In Property & Casualty insurance, claims settlement speed is a critical differentiator — yet manual policy checking remained a significant bottleneck for Zurich Insurance UK. Experienced claims handlers were required to read and interpret dense policy wordings, cross-reference coverage terms, and assess claim eligibility by hand. This process routinely stretched settlement decisions across multiple days, creating friction at the moment customers are most vulnerable. At enterprise scale, the inconsistency inherent in manual review also introduced accuracy risk and regulatory exposure, while customer expectations for rapid, transparent outcomes continued to rise. The status quo was unsustainable: slow resolution damaged satisfaction, and manual throughput capped the volume of claims that could be processed without proportional headcount growth.

Zurich Insurance UK partnered with Sprout.ai to develop an Automated Policy Checking tool built on Natural Language Processing (NLP) and Knowledge Graph technology. The system was trained on over 20,000 historical claims, enabling it to replicate the judgment depth of a handler with an estimated 100 years of combined experience. The engine processes policy wordings at 10,000 words per microsecond, extracting coverage intent and applying it to incoming claims in real time. A key differentiator is its integration with the Financial Ombudsman Service database, allowing the AI to surface precedent-based comparisons alongside each recommendation — providing handlers with explainable, context-rich outputs rather than opaque decisions. The solution was piloted between December 2020 and February 2021 before broader rollout, with the AI operating as an augmentation layer integrated into the existing claims handling workflow.

The pilot demonstrated measurable improvement across speed, accuracy, and transparency. Claims that previously took multiple days to settle were triaged and processed within hours, with settlement times falling to under 24 hours. Policy checking accuracy exceeded 98%, surpassing the rate achieved through manual review. Qualitative outcomes were equally significant: handlers received structured, explainable recommendations rather than raw policy text, reducing cognitive load and improving consistency across the team. Customers also benefited from greater transparency, as Ombudsman comparisons were surfaced during the claims process.

• Settlement time: Multiple days → under 24 hours
• Policy checking accuracy: Above 98% (exceeding human baseline)
• Training corpus: 20,000+ historical claims

• Historical data volume matters: Training on 20,000+ claims gave the model sufficient breadth to handle diverse policy types and edge cases — smaller corpora risk coverage gaps.
• Explainability drives adoption: Surfacing Ombudsman precedents gave handlers and customers confidence in AI recommendations, which is essential in regulated industries.
• Augmentation over automation: Keeping humans in the decision loop preserved accountability and regulatory compliance while still delivering speed gains.
• Pilot before scale: A controlled three-month pilot allowed Zurich to validate accuracy and refine integration before committing to full deployment.
• Structured knowledge graphs extend NLP: Pairing language models with structured knowledge sources improved the system's ability to reason over complex, interconnected policy terms.