It’s not true that insurance carriers are averse to paying claims, it’s just that many insurers hate paying claimants that don’t deserve the money—particularly those who file fraudulent claims that end up costing honest policyholders in the long run.

Tim Wolfe heads the special investigations unit (SIU) for CNA where his team’s responsibility is to detect and investigate claims fraud. To assist in that job, CNA began using the Fraud Framework tool from SAS in 2011 after taking several years to vet suppliers of similar tools and go through pilot programs to select the best tool for the carrier’s business.


CNA is a commercial lines carrier with multiple lines of business. SAS built the predictive models for CNA’s four major lines: workers’ comp, general liability, commercial property, and commercial auto, based on the carrier’s own historical data.

“All the claims we have successfully investigated for fraud were sent to SAS—our financial data, medical bill data, and our case management system. We sent them several years’ worth of data and they built models based on factors that are common to known fraud claims,” says Wolfe.

With the models in place, CNA sends SAS all of its open claims data for those four lines of business in batch-form on a weekly basis. So, when the SIU team arrives at the office on a Monday morning, they enter the SAS user interface to look for possible fraud.

“The first area we look at is the claim alerts—individually scored claims with fraud potential,” says Wolfe. “Our investigators go through these claims, which have not been referred by adjusters. They start with the highest scoring claims and determine if it is appropriate to launch an SIU investigation.”

Wolfe explains that CNA uses a sophisticated model so there are dozens of factors that contribute to the fraud score. With workers’ comp, those elements might include the type of injury that would affect the fraud score—lower back, soft tissue, or subjective kinds of injuries—which could bring about a high score. Also, if an injury occurs on a Monday morning, that is a red flag for investigators.

“Sometimes, people get injured over the weekend and you find they come in to work on Monday morning and 30 minutes later they allege that they hurt their back or something like that,” says Wolfe.

CNA has been using predictive analytics to fight fraud since 2008, but Wolfe explains the tools the carrier uses today are more sophisticated and effective. CNA built a basic model in 2008 just for the workers’ comp line of business.

“The hit rate was low—in the single digits,” says Wolfe. “It wasn’t worth the time we spent weeding out the false positives.”

The following year, CNA began using a pair of models built by another company which Wolfe describes as marginally better. The insurer worked with those models for eight months, but determined the models still weren’t as effective as what CNA felt it needed.

That led the carrier to the SAS model, which Wolfe explains has three major components that appeal to CNA:

  • The ability to apply the carrier’s own business rules
  • To opportunity to build predictive models
  • The social networking aspect.

“By social networking I’m talking about the ability of the tool to identify connections between multiple claims,” says Wolfe. “When we go to the user interface we have the ability to look at provider networks—people who provide service to the claimant such as doctors of all specialties, legal providers, physical therapists, transportation companies, and durable medical equipment providers. All those people live off the claims process. Where the big bucks go in our business is to those providers rather than the individual claims.”

The SAS tool allows CNA to see connections between claims. For example, the same doctor and attorney may be representing multiple claims. Those claims might be handled by individual adjusters who might not be able to connect the dots. The tool connects those dots and looks for patterns of questionable activity.

CNA applies its own business rules, which are available in the data fields. Wolfe explains an example of this might involve a property claim where the carrier can list some of the factors that the tool needs to consider.

“A red flag would be a new insured that files a claim within 30 days of policy inception,” says Wolfe. “That would be suspicious.”

With predictive modeling, another facet of the tool is anomaly detection. Anything out of the norm with a claim would trigger a higher fraud score and merit a second look by members of the SIU. An example would be stress claims.

“Stress claims are not accepted as a compensatory injury in many states, but in states like California and Hawaii they are,” says Wolfe. “If a claim for a stress injury is filed by an injured worker, it might trigger a score under the anomaly detection feature.”

The tool is able to mine the unstructured text such as the adjuster notes, which CNA adjusters enter into the claims system. When they interview a claimant or a witness, the adjusters put that information in the claim notes.

“The tool has the ability to detect concepts of fraud—a phrase or a word that are common to known fraud claims,” says Wolfe. “What the adjusters write in the text can trigger a fraud score. With stress claims, we typically find people are having performance issues or maybe they’ve had non-industrial injuries before. The tool can pick up on phrases in the text.

In addition, investigators can go to the medical provider side, the medical billings that are provided out of the norm, such as doctors billing the maximum amount of face time with a patient, which is known as upcoding. So if certain doctors continually bill the highest amount for face time with a patient—not every patient needs a complex examination, explains Wolfe—that will generate a fraud score.

With the abundance of data available for insurers, Wolfe believes the job of the SIU is in some ways easier because the system points investigators in the right direction, but it also compounds matters.

“You are going to get false positives, but before it would take many man-hours to compare multiple claims files; now [the software] does it in seconds,” he says. “The problem becomes there is so much information that one network can involve 100 claims. How do you drill down to the specific fraud? That continues to be a challenge for us.”