What machine learning techniques are used in the assistant?

The system uses decision tree ensembles, specifically XGBoost, trained on historical policy applications, underwriting decisions, and claims data. Features are selected using recursive feature elimination (RFECV) to ensure optimal predictive performance.

How are decision thresholds configured?

Administrators can configure score thresholds for applications: green zone for auto-approval, yellow for manual review, and orange for high probability of decline, allowing flexible risk management and control over the auto-decision process.

What metrics are used to assess algorithm performance?

The system uses ROC AUC (Receiver Operating Characteristic Area Under Curve) to evaluate performance, ensuring high recall for declined applications while minimizing false positives, which is critical given the dataset imbalance.

What results did the client achieve with the assistant?

The client achieved a 45% increase in straight-through processing, processed 2.5 times more applications in the same period, and enabled underwriters to focus on high-value or complex cases, improving overall operational efficiency.

Policy Application Decision Assistant for Underwriters

Azati designed an automated policy application decision assistant for insurance underwriters. With machine learning tools and improving computing capacities, this has now become a reality. The assistant processes policy applications and provides recommendations to approve, decline, or manually review them, helping underwriters make faster, more informed decisions. The system utilizes historical data for predictive modeling, significantly enhancing the underwriting process.

Discuss an idea

45%

increase in straight-through processing

2.5x

increase in application processing capacity

0.75-0.85

Auto-Approval Efficiency Score (AAES)

Motivation

To address manual decision bottlenecks, inconsistent underwriting outcomes, and time-consuming processing by leveraging historical data and machine learning, enabling faster, more accurate, and consistent policy application decisions while freeing underwriters to focus on complex or high-value cases.

Main Challenges

Manual decision-making caused delays and bottlenecks in the underwriting process, preventing efficient handling of policy applications. Azati addressed this by automating the decision-making process, reducing the need for manual intervention and speeding up policy application processing.

The underwriting process was time-consuming due to repetitive tasks, making it difficult to quickly evaluate and process applications. Azati developed an automated system that analyzes policy applications and provides recommendations, significantly reducing processing time.

Human factors led to inconsistent decisions, affecting the quality of underwriting decisions and increasing the risk of errors. Azati solved this by using machine learning to analyze historical data and provide data-driven decision recommendations, ensuring consistency and accuracy in the process.

There was limited use of historical data, which hindered the potential to make data-driven decisions and improve the accuracy of policy assessments. Azati leveraged historical underwriting data to train the machine learning model, enhancing the decision-making process with predictive insights from past applications.

Our Approach

Machine Learning-Based Predictive Model

We developed a machine learning-based predictive model that processes policy applications, assigns them scores, and provides recommendations for approval, rejection, or manual review.

Historical Data Training

The model was trained using historical data, including previous underwriting decisions, policy submissions, and claims, allowing the system to learn from past patterns and make informed decisions.

Architecture Planning

Identified limitations of GraphQL as a standalone solution and proposed a hybrid approach combining GraphQL and standard JSON API to ensure flexibility and comprehensive data handling.

Automation and Efficiency

By automating the decision-making process, the system reduces manual intervention and significantly increases the speed and accuracy of underwriting decisions.

Want a similar solution?

Just tell us about your project and we'll get back to you with a free consultation.

Schedule a call

Solution

Automated Decision Engine

Processes incoming policy applications using a machine learning-based predictive model that scores each application and provides recommendations to approve, decline, or manually review. Eliminates repetitive manual assessment and speeds up underwriting while maintaining high accuracy.

Key capabilities:

Automatic classification of applications into green, yellow, and orange zones
Prediction of approval or decline probability
High recall for declined applications to minimize risk
Stratified 10-fold cross-validation for model validation

Historical Data Integration

Utilizes extensive historical data from past applications, decisions, policies, and claims to train the predictive model. Ensures that the system learns from past patterns, improving consistency and reliability of recommendations over time.

Key capabilities:

Training dataset with over 140k applications and 1000+ features
Recursive feature elimination to optimize model performance
Integration of prior policy outcomes and claims data
Enhanced predictive scoring for new applications

Configurable Decision Thresholds

Allows administrators to adjust thresholds for approval, review, and decline zones, balancing risk and application throughput. Enables the business to control the level of automation and fine-tune system performance based on risk appetite.

Key capabilities:

Green zone for auto-approval
Yellow zone for manual review
Orange zone for high probability of decline
Thresholds adjustable by line of business or risk tolerance

Statistical Performance Analysis

Provides underwriters with statistical analysis of similar past decisions, showing historical effectiveness of approvals and declines. Supports informed decision-making and continuous improvement of underwriting strategies.

Key capabilities:

Panel showing historical accuracy of similar decisions
Measurement of algorithm performance using ROC AUC
Supports continuous improvement of predictive models
Facilitates informed operational adjustments

Hybrid API and Architecture

Implements a hybrid API architecture combining GraphQL and JSON endpoints to handle complex data needs of the underwriting system, ensuring flexible integration with internal and external platforms.

Key capabilities: