Insurance scams cost the U.S. economy tens of billions of dollars each year. Now, insurers are adding artificial intelligence to their fraud-fighting toolkit. My company uses AI to stop e-commerce payment fraud by detecting unusual behavior and stolen data. Insurance claims are less standardized and more complex than a typical online purchase, but the industry is finding ways to use AI to reduce fraud by policyholders, providers and criminal gangs.
The FBI estimates that insurance fraud costs more than $40 billion annually. That means the average household pays $400 to $700 more in premiums each year. Those numbers don't include healthcare insurance fraud, which is rampant and harder to quantify. Estimates of the scope of healthcare fraud range from 3% to 10% of annual healthcare spending. That translates to anywhere from $68 billion to $230 billion per year, driving up premiums for businesses and families.
Who commits insurance fraud?
When it comes to identifying fraudsters, the insurance industry has a very different challenge than the e-commerce industry. E-commerce fraudsters are typically unethical shoppers committing return fraud or organized criminals buying items for resale with stolen card data. But insurers face a different fraud landscape with more players.
Individuals can commit fraud in several ways. One is by providing false information that gets them a lower premium. For example, 10% of auto insurance policyholders falsely state that their vehicles are garaged at night. Policyholders can also exaggerate the amount of damage caused by an accident or injury – or fake an accident or injury.
A small number of unscrupulous agents embezzle premiums instead of forwarding them to the insurance company. Scammers posing as agents can also steal premiums and damage the insurer's brand in the process.
Claims adjusters are responsible for validating claims and spotting fraud before insurers must pay for it. Unfortunately, some adjusters go bad and exploit their knowledge of fraud prevention to enrich themselves. One high-profile case in Florida involved an adjuster who took bribes from businesses and consumers. He got caught and then worked as a police informant without his employer's knowledge. When the news broke in 2018, the company had to review its internal processes and do public damage control.
Providers, like claims adjusters, are largely honest and conscientious, but a few engage in fraud. In healthcare, provider fraud includes false or exaggerated claims, medically unnecessary procedures, kickbacks for referrals and other schemes. Other types of provider fraud include overbilling for auto and home repairs. Auto shops and contractors may also charge for new parts and materials but use cheaper used items to make the repairs.
Organized criminal groups commit the most healthcare fraud, according to the National Health Care Anti-Fraud Association. Crime rings have also been busted for scams against federal crop insurance programs, homeowner insurers and commercial trucking insurers.
Faced with so many types of fraud from so many sources, insurers are always looking for better ways to screen applications, evaluate claims and identify bad actors.
How AI can help detect different types of insurance fraud
The same capabilities that make AI so powerful against e-commerce fraud also help fight insurance fraud. AI algorithms can analyze large amounts of data quickly to find patterns. Then they can spot anomalies that don't fit the patterns. For example, by comparing new claims to existing data, AI can help detect claims values that are unusually high.
Visual analytics can assess auto or property damage based on images and videos and determine if the damage amounts claimed are accurate or not. Is that damaged rear bumper really a $10,000 claim, based on the visual evidence and other data from similar claims? AI can help sort it out quickly.
AI can also take a higher-level view of a company's claims, employees and policyholders to look for patterns that could indicate large-scale fraud. This kind of analysis can include behavioral data and transaction histories. For example, if employees' online behavior indicates that they're struggling with debt, and they also have unusual new patterns of financial transactions, the algorithm might flag them for more review to see if they are receiving kickbacks in exchange for claims approvals or inflations.
Or, to take an example from healthcare, if a clinic or hospital has an unusually high number of expensive procedures compared to similar facilities, or procedures that cost substantially more than the industry average, investigators may want to take a closer look. The U.S. Department of Health and Human Services reported to Congress in early 2019 that it's exploring AI and machine learning to help it root out organized healthcare fraud.
Federal adoption of AI could change the way large-scale healthcare fraud is identified. Already, the HHS Office of Inspector General and its state and local law enforcement partners lead regular takedowns of national and regional healthcare fraud rings. One bust in 2018 named more than 600 defendants in $2 billion worth of Medicare and Medicaid fraud.
The investigation required the resources of more than 1,000 law enforcement agents across the country. AI could assist in cases like this by analyzing multiple large sets of data quickly and identifying anomalies faster than humans can. That could give agents the information they need to target their investigations and stop fraud faster.
In the case of healthcare fraud, the stakes are more than money. The 2018 scheme uncovered by the HHS included illegal distribution of addictive opioid medications which have caused a public health crisis. So, using powerful analytics and machine learning may not only help root out fraud and save taxpayer money, but it may also save lives.
The limits of AI in fraud detection
It's important to keep in mind that AI can't, and shouldn't, replace human expertise when it comes to evaluating fraud, for two important reasons. First, the algorithm may not always get it right, especially when it's new and has a relatively small dataset to learn from. Sometimes what looks like criminal behavior may have a valid and benign explanation. And it's important to keep in mind that AI can learn biases that skew outcomes if the databases contain biased information or were built based on biased practices.
In e-commerce fraud prevention, it's best practice to have a fraud analyst review any orders that AI flags as possible fraud – a process that often includes talking with the customer to validate their information. Rejecting a good customer will often drive them away for good. Rejecting a valid claim can damage the insurer-policyholder relationship and may lead to legal action against the insurance company.
The second reason it's critical for AI-flagged claims to go through an expert review is that human expertise helps the algorithm get smarter. When analysts find flagged claims or anomalous patterns that aren't fraud, they can feed that data into the algorithm so it can learn to identify similar cases and evaluate them more accurately. Over time, the AI system gets better at telling the difference between fraud and legitimate activity.
Right now, fraud is a costly fact of life in many industries. It wastes money and draws resources away from core business functions. As we've seen, health insurance fraud can also endanger people's wellbeing and lives. But as insurers and law enforcement agencies start to work with AI to spot potential problems and evaluate claims more accurately, insurance fraud may become much harder to get away with.