The Challenge
The insurer had developed a hybrid machine learning model designed to identify claims where earlier intervention at First Notice of Loss (FNOL) could help prevent unnecessary escalation and higher settlement costs. However, the model remained at prototype stage and could not yet deliver scoring quickly enough to influence decisions while a claim was being logged.
The challenge was to move the model into a reliable, governed production environment capable of scoring claims in real time, giving claims handlers actionable information at the point of FNOL rather than later in the claims journey.
Approach
Advancing Analytics productionised the hybrid model within the Databricks Data Intelligence Platform, transforming the prototype into a scalable real-time service.
The prototype feature transformations were re-engineered into a reliable and repeatable production pipeline, with multiple models brought together behind a single low-latency serving endpoint. Lifecycle management, versioning and governance were also embedded into the production environment to create a robust foundation for ongoing model management.
Architecture
The architecture creates an end-to-end journey from the moment a claim is logged through to an intervention recommendation for the claims handler: FNOL Event → Feature Pipeline → Hybrid Model → Real-Time Model Serving → Claims Handler
- FNOL event: The process begins as soon as a claim is logged.
- Feature pipeline: Claims data is transformed on Delta Lake through the production feature pipeline.
- Hybrid model: Multiple models are combined to generate the appropriate claim score.
- Model serving: The model is exposed through a real-time, low-latency serving endpoint.
- Handler intervention: The resulting recommendation is available while the claim is live, supporting earlier intervention.
- Unity Catalog: Provides governance across the data, features and served model, including lineage and access control.
- MLflow: Supports model lifecycle management and versioning from prototype through to production.
Outcomes
The project successfully moved the capability from prototype into production, enabling claims to be scored in real time at FNOL rather than hours or days later. This gives claims handlers the opportunity to identify relevant cases and intervene earlier in the claims journey, with the objective of preventing unnecessary escalation and reducing avoidable claims handling and settlement costs.
Headline outcomes
Real-time service
Claims can be scored at the moment they are logged, bringing machine learning directly into the operational claims process.
Lower handling costs
Earlier identification and intervention are designed to help prevent costly claims escalations and reduce avoidable downstream costs.
$80–160bn industry opportunity
An estimated $80–160 billion in industry savings by 2032, attributed to Deloitte, highlighting the wider potential for AI and automation across insurance claims.
This is an industry-level statistic rather than a measured saving achieved by this individual insurer.
Production-ready AI
The insurer now has a governed production architecture supporting real-time machine learning, model versioning, lineage and access control.
Sustainable in-house capability
The engagement was delivered using a blended team approach, with the build transferred into the insurer's internal team to support longer-term ownership.
Move beyond the AI prototype
See how Advancing Analytics can help you productionise AI, embed it into real-world processes and turn experimentation into measurable business outcomes.
