Leveraging AI in Fraud Detection: Four Critical Components

By Jim Hulett


Insurers have made significant investments in streamlining the claims process to adapt to changing customer expectations. But fast and simple claims processing comes with a risk—increased fraud.

One out of every ten claims is estimated to be fraudulent,1 costing the (non-health-related) insurance industry approximately $40 billion annually.2 And as insurers look to remove friction from the claims process, unscrupulous actors—whether individuals, third-party providers, or organized rings—have more opportunity to submit questionable claims that can slip through the system.

Most carriers are cognizant of this problem, which is why they’re looking to leverage big data and artificial intelligence to detect and deter fraudulent activity. But to incorporate those technologies in fraud analytics, it’s important to understand the four critical components of operationalizing them: variety, value, volume, and velocity.

  1. Variety: Do you have enough data?

Developing AI algorithms and predictive models for fraud detection requires data—a lot of data. The more data that’s analyzed, the more effective the system will be at recognizing patterns and characteristics indicative of fraud. While an insurer’s historical claims data is a good foundation for analytics, incorporating third-party data sources provides a more holistic picture because it taps broader industry data for analysis.

But integrating third-party data has its challenges. Mapping data from a third-party source to an internal model can take a significant amount of time (a year or more in some cases) and resources. But having an established infrastructure built into analytics can create more seamless data integration, accelerate implementation, and save money.

Another type of variety is unstructured data, such as images, audio, video, text, and social media. Analyzing this kind of data can enhance the scope of fraud detection. For example, image analysis can determine if loss photos submitted by claimants have been tampered with or if the image time stamp is before the date of loss. In addition, text mining technology can read and analyze claim notes and identify trends and patterns associated with certain keywords.

Analyzing a wide variety of data along with internal data sets is essential to recognizing new and emerging patterns and identifying fraud risks.

  1. Value: Are you getting the right insights?

Across industries, businesses struggle to derive actionable insights from data. Nearly 60 percent of organizations suffer from inaccurate business intelligence because of slow or poor access to the right data.3 For insurers, part of the issue can be attributed to disparate data within their organizations. McKinsey reported that insurer data is typically “incomplete or miscoded and substantial effort is required to bring the data into working condition.”4

Organizing data often requires investing in data science resources, and insurers are starting to answer the call. According to an industry study, 35 percent of large property/casualty insurers and 24 percent of midsize carriers are expanding their data science efforts.5 But as data science teams apply analytics, insurers need business analysts to apply the proper business logic to models to get actionable insights for fraud detection.

For example, predictive models that not only score claims for fraud potential but also provide reason codes that detail suspicious behavior and attributes within a claim can help direct SIU on how to intervene and investigate.

  1. Volume: Is your data deep enough?

Deriving actionable insights not only requires the right data but also the necessary depth of data. While many people think of data volume only in terms of breadth, it’s equally important that data is deep and rich enough to create effective analytic models.

Fortunately for insurers, claims often include depth of information. For example, an auto bodily injury claim can include extensive details: involved parties, injury details, medical provider information, drugs prescribed, auto body shop, vehicle damage photos, police report, weather report, and more. Analyzing that depth of information—including historical claims data from involved parties and providers—offers a more complete picture of a claim.

While this extent of detail usually isn’t available at first notice of loss (FNOL), it can unfold throughout the life of a claim. And as adjusters gather reports, interviews, and specifics on the event and entities, adding that information is vital for accurate analysis.

  1. Velocity: Do you have processing speed?

A fraud analytics system that uses AI is only as good as its processing power. Each claim can have hundreds of data points that need to be analyzed and weighed against each other to determine its risk score. And it needs to be done fast!

Crunching a large variety of structured and unstructured data requires significant speed and capacity that most internal servers can’t handle. A carrier needs the proper tech stack to enable the modeling and analyzing processes. Cloud computing makes it possible. Novarica reports that more than 70 percent of insurers are using cloud computing in some fashion, while 10 percent run most of their infrastructure on the cloud.6

Fast processing also enables a critical aspect of fraud analytics—real-time insights. As new information becomes available, it needs to be analyzed to update the claim as quickly as possible to equip claims handlers with the right insights either to process a claim or triage it for further investigation.

Keeping pace in the fight against fraud

Technology such as AI, machine learning, and predictive analytics is becoming more common in the insurance industry, and it’s transforming all aspects of the business. There’s potential for a strong return on investment in fraud analytics because these technologies can directly impact fraud leakage in the claims area. Whether you’re building in-house analytics or partnering with a vendor, it all starts with the four key components necessary for effective fraud detection.

jim hulett.jpg

Jim Hulett is vice president of the ISO ClaimSearch® Antifraud Solutions Group.

1. Insurance Information Institute. “Background on: Insurance Fraud”

2. FBI Reports and Publications

3. Digital Insurance. “Poor data access putting many analytics projects at risk”

4. McKinsey & Company. “Transforming into an analytics-driven insurance carrier”

5. Insurance Innovation Reporter. “Insurers Are Prioritizing Data Science More in 2019.”

6. Novarica. “70% of Insurers Now Using Cloud Computing, With Additional Growth Planned, Says Novarica”