USAA

Auto insurance carriers have long sought to eliminate manual appraiser involvement in straightforward claims — a process known as touchless claims processing. For USAA, which serves millions of military members and their families, the volume of auto claims makes manual review a significant operational bottleneck. While an earlier collaboration with Google Cloud had produced a computer vision API capable of identifying damaged vehicle parts from photos, that system stopped short of actionable decisions: it could not determine whether a part should be repaired or replaced, nor estimate the labor hours required. Without those downstream predictions, fully automated claims settlement remained out of reach, leaving appraisers as the necessary final step and limiting cost efficiency and claims cycle time.

Over 16 months, USAA and Google Cloud's AI Industry Solutions Services team built a multi-model ML system designed to extend the existing computer vision pipeline into full repair/replace and labor hour predictions. Millions of vehicle damage photos were scored through the existing REST-based computer vision API using Google Cloud Dataflow for parallel processing, with results stored in BigQuery alongside structured claim data — vehicle make, model, year, point of impact, zip code, and drivability status. The team engineered 20 features from these combined sources, then evaluated three modeling approaches: AutoML on Vertex AI, K-Nearest Neighbors, and a custom TensorFlow/TFX pipeline. Production infrastructure was built on Vertex AI Pipelines, supporting both real-time and batch prediction APIs with model explainability and a human-in-the-loop model promotion process designed to meet insurance regulatory requirements.

The project achieved a 28% peak ML performance improvement over baseline models across the combined system. AutoML on Vertex AI was selected for production deployment based on both performance and maintainability:

• 6.78% improvement in the repair/replace classification model (AUC ROC)
• 16.7% improvement in the repair labor hours estimation model (RMSE)
• 120+ models trained across 4GB of processed data during development
• Real-time and batch prediction APIs deployed with full explainability support

Beyond the metrics, the project delivered a complete automated retraining pipeline, a single consolidated BigQuery training table for regulatory auditability, and an MLOps infrastructure positioned to support ongoing model improvement without rebuilding the system.

• Feature engineering drives more lift than model architecture. Domain expertise from insurance specialists — not algorithmic complexity — was the primary source of performance gains; investing in 'golden' features pays higher dividends than tuning model hyperparameters.
• AutoML can beat custom models in production. On common vehicle body styles, AutoML on Vertex AI matched or exceeded the custom TFX pipeline while dramatically reducing ongoing maintenance burden — a meaningful trade-off in regulated industries.
• Regulatory compliance shapes MLOps design. Human-in-the-loop model promotion and a single auditable training table were not optional enhancements; they were prerequisites for deploying ML in an insurance context.
• Parallel processing infrastructure is a prerequisite, not an afterthought. Scoring millions of images through an existing API required serverless orchestration via Dataflow — plan for this before model development begins.