Faster End-of-Life Insurance Claims Payouts with Automated Fraud Checks

Handling the end-of-life expenses of a loved one can present a significant and often unexpected cost burden to the survivors. Purchasing an insurance policy with a death benefit (often called funeral insurance) can help cover the costs associated with fulfilling the last wishes of the deceased, including burial, travel expenses to attend a memorial service, and more. This financial benefit makes it an attractive target for bad actors who submit fraudulent claims against these policies. Spotting which claims are legitimate and can be paid out, and which ones are suspicious and should be reviewed carefully for evidence of fraud, is crucial to continuing to offer these benefits to customers.

To learn more about how machine learning (ML) and artificial intelligence (AI) services are helping spot fraud as part of automated insurance claims processing, I sat down to talk with Ashia Bowers, Head of Automation at Standard Bank Insurance, and Tiaan Taljaard, Head of Cognitive Automation at Standard Bank Group.

AWS: Who is Standard Bank?

Ashia: Standard Bank is a major South African financial services group that been in business for over 160 years and is Africa’s largest lender by assets. The organization offers a range of products and services to its clients, including investment solutions, home loans, vehicle and asset finance, along with insurance to name a few.

AWS: What is your fraud challenge?

Ashia: Funeral claims are an area prone to fraud due to the nature of the product which requires faster processing of claims. According to the Daily Maverick, South African life insurers detected 4,287 fraudulent and dishonest claims worth R787.6-million across all lines of risk business in 2021, significantly higher than the 3,186 claims valued at R587.3-million in 2020. The Association for Savings and Investment South Africa (ASISA) revealed that funeral insurance attracted the highest incidence of fraud and dishonesty, followed by death cover, disability cover, hospital cash plans, and retrenchment benefit cover.

The automated traditional process used by our Funeral Claims team still has many manual checks in place to make sure that claims are legitimate prior to claims finalization. These checks may slow down the claim process, which in turn could delay the release of funds for legitimate claims at a time when they are needed to help pay the beneficiaries for the costs of end-of-life expenses.

Tiaan: The Automation team at Standard Bank Insurance came to us with a challenge: could we use ML to help enable a better customer experience with a faster, frictionless, and efficient payout of valid funeral claims? The goal was to see if we could meet these objectives and improve the claim process to reduce the time to process claims from 48 hours to 1 hour without introducing more risk to the business. Some of the data points used to assess these fraudulent claims were based on existing business rules, which claims-assessors use to manually assess the validity of a claim.

AWS: Why did Standard Bank Insurance choose Amazon Fraud Detector?

Tiaan: The business area required a managed solution that is scalable without the need to increase the technical skills required to develop and deploy a solution. While looking at options to address the issue, we had to carefully explore technology that is available in our toolbox. As a team, we have a firm understanding of AI technology and the phases to develop AI solutions. Moreover, not being insurance domain experts, we needed a platform where we could iteratively develop and engage with insurance subject matter experts to review progress and guide to us to success.

The ease of use, flexibility, and low code nature of Amazon Fraud Detector were the primary drivers in deciding to use the solution. It has been trained and tested with real fraud use cases which reduced the need for huge amounts of data labeling effort on our specific requirements. Amazon Fraud Detector enabled us to easily test and present findings from a pre-trained AI service, allowing us to focus on the model output and alignment to business objectives. With the assistance and guidance of our AWS team, we were able to run a successful proof of concept within a few weeks.

We also firmly believe that any successful AI deployment requires frequent checks to make sure that it’s performing accurately and still aligning to the business objectives. As data is generated daily, we must enable the domain experts to produce new and accurate data for model retraining.

To continuously improve the accuracy of the Amazon Fraud Detector model, we have paired this service with Amazon Augmented AI (A2I). This enables our claims experts to feed the model with any inputs for model re-training. Amazon A2I enabled us to bring these domain experts closer to the AI solutions, and allowed them to review the predictions made by the models. Differences between model predictions and expert reviews are valuable data points to measure model performance and retrain the model should there be any drift.

AWS: What results have you seen so far from using Amazon Fraud Detector?

Ashia: Since we launched Amazon Fraud Detector into production, our results have been great. The model that was developed by the service is accurately identifying high-risk claims at the time of claim submission which helps reduce risk.

Approximately 94% of our claims are typically rated as low risk and for these claims, the turnaround time has already reduced from 48 hours in February with our traditional manual process to less than 6 hours by the end of August. This has resulted in improved customer experience. We have also seen a 36% increase in our Net Promoter Survey (NPS) scores between February 2022 to August 2022 since going live in production. We attribute this to the faster payouts due to safely automating approvals for low-risk claims.

For the 6% of claims rated as high risk, we now have more capacity to interrogate these better than before. As a result, we are able to thoroughly investigate suspected cases and stop more claims with actual fraud. Overall our confirmed fraudulent cases that we have been able to identify before paying out a claim has increased over 100%, which has greatly reduced the business’ exposure to risk.

Tiaan: Amazon Fraud Detector has not only enabled Standard Bank Insurance to solve a challenge by streamlining the process of identifying claims that are ready for direct processing, but also the platform has helped us identify rules that we were using through the manual process and illustrate which rules were making the biggest difference. Amazon Fraud Detector calls this “feature importance”, and it has added immense value. We believe that as we plan to enhance and further scale the solution, we can make better decisions because we understand the way that the data drives risk detection and identifies valid claims.

AWS: What future plans do you have to expand your use of Amazon Fraud Detector?

Ashia: Reusability is a key feature to help enable and accelerate the use of AI-based initiatives throughout the business. We see enormous possibilities to use Amazon Fraud Detector and other AI services from AWS to benefit the organization’s other insurance initiatives as well as other business areas. We believe that AI will help us strengthen our value proposition and greatly enhance the customer experience by delivering fast, efficient, and reliable solutions to our customers.

Tiaan: Adding to these comments, from a technical standpoint, we believe that the success of Amazon Fraud Detector has shown value. As a result, we plan to use the solution within the organization to drive decision support solutions. We believe Amazon Fraud Detector can be packaged within our Cognitive Automation team to easily be reused by making full use of the AWS ecosystem whereby other teams within the bank can benefit.

If you have questions or other feedback for Ashia, Tiaan, Standard Bank, or AWS, please leave it in the Comments section. For more information about Amazon Fraud Detector, see our Catching Fraud Faster by Building a Proof of Concept in Amazon Fraud Detector post.