Using Artificial Intelligence In Expense Reporting While Performing Fraud Detection

Financial frauds have been prevalent for a long time. Banks and all other industries have been trying to control frauds for quite some time. Credit card and debit card, payroll, andexpense frauds form the largest types of frauds. Artificial Intelligence can be used to reduce auditing and fraud analysis cost while reducing fraudulent transactions in companies. Below are some of the most common types of frauds and examples of how AI can lend itself to detection of these frauds.

Credit Card and Debit Card Frauds: In 2015, credit, debit and prepaid card issued worldwide reached around $21. 84 billion according to Bloomberg Report and is expected to increase by 45% by 2020.

Some of the ways in which credit card and debit card frauds can be prevented are listed below: Device Intelligence can be used to determine whether a device profile from which transaction is occurring is legitimate within a span of milliseconds, allowing banks to stop the frauds before the occur. Information can be collected from devices based on cookies and web beacons the moment a customer visits a bank’s website (e. g. login, checkout, account creation, account registration). The solution finds whether a device, which the user has utilized, has an abnormally high amount of online activity and uses it, in conjunction with other factors, to determine fraud patterns and recommend whether a transaction needs to be denied, approved or reviewed. All this can happen in less than a second. Such a system can be trained using a list of previous frauds (both internal and external) that have occurred and been detected by fraud analysts.

Continuous Authentication: Assesses a user’s behavior to authenticate the identity of the logged in user in cases wherein logged in users forget to logout. Can use cognitive factors such as eye hand coordination, applicative behavioral patterns, usage preferences, and device interaction patterns, physiological factors such as left, right handedness, press size, hand tremors, etc. and contextual information such as transaction, navigation, and device patterns. Fraud Prevention: Can perform real time detection of criminal behavior, malware, Remote Access Tools (RAT), aggregators, robotic activity, social engineering attacks not recognized by traditional fraud prevention methods. Based on aforementioned criteria, machine learning can analyze a user’s behavior and provide him a risk score. Decision Intelligence can be used to reduce credit card frauds. As a user makes a purchase, Deep learning models can be used to determine whether the type of purchase, time, location, purchase cost along with a ray of other data points such as IP address, device ID, email phone number are in line with client’s previous transactions. Such an application can also check the merchant’s system to confirm whether the customer or system is assigned a risk score.

In absence of a pattern that is consistent with fraud, the system can approve the transaction.

Payment Gateways: Machine learning can help merchants, financial service consultancies, and payment service providers differentiate between fraudsters and genuine customers. A customer’s digital identity can be determined by using data points such as email, phone, location, IP address, device ID, passport number, etc. These details are defined when an individual transacts online and are updated as customer evolves. Based on these inputs from the system, merchants and banks can improvise the security to authenticate or assess the risk process if additional checks are required. Machine learning algorithms can study a customer’s recent online activity such as payment behavior, social media, social security, IP location, device activity, and billing address. The more the data points are available for a customer, the lower the risk score. For example, if the same fraudster lists different name variations when opening an account, say Kris Jefferson, Kris Jeff, Kris Jesse, etc. the algorithm will analyse data points (IP addresses, devices, bank accounts, payment behavior)from these logins to determine a risk score associated with such transactions. These data can also be used to update the customer’s profile and determine the trustworthiness of the customer. This would allow merchants to be aware of fraudulent transactions such as chargebacks, fake account, spams, account takeover, etc. The risk score as well as the data points used to arrive at the risk score can be shown to a human analyst who can intervene as and when required. This human intervention can be fed into the system to increase its future accuracy and improve its algorithms.

Fraudulent T&E expenses:According to the Association of Certified Fraud Examiners, the approximate median loss due to reimbursement fraud approximates to $40,000. Some of the ways in which fraudulent claims are made include fictitious business receipts, inflated expenses or illicit upgrades, false merchant codes, multiple claims for the same bills or codes, etc. An example, when we purchase pillowcasesat Walmart, we expect to see the retailer spelled out clearly in our bank statements. However, when it comes to gentlemen’s club industry, expenses are not so obvious. Irrespective of whether these claims are a part of generic mistakes or intentional frauds, they lead to heavy financial impact for all types of organizations. Conventional methods of fraud detection perform random sampling which generally covers only 1 to 15% of the total reimbursements. Real time fraudulent detection can be performed through the usage of AI. Below are some key steps that any AI engine would need to adopt to classify whether an expense is a legitimate or fraudulent expense.

Extracting information: Machine learning, computer vision, deep learning and NLP can be used to understand the context of the reimbursements by scanning receipt images, boarding passes, travel documents, etc.

Data Augmentation:These data can be searched real time against thousands of external and social data sources to establish the validity of business merchants, their pricing, and background information to confirm the submitted expense reports. This can be also used to detect whether the business that issued the receipt was a club, casino, etc. Based on this background research, reimbursement requests can be approved or rejected. In order to ensure that a specific alcohol is not being billed into the restaurant bill, application can also match each item in the bill against a dictionary of thousands of brands of alcohol.

Additionally, a guest claimed on the T&E can be searched against multiple news and government sites to reduce the risk of providing company expense to a government employee or a high risk individual. Pattern Recognition: Using Machine Learning, AI engine can detect patterns to detect employeeswho are repeat offenders or make accidental or opportunistic claims.

Company wide analytics: This sort of analytics can also allow your company to determine company wide spend and audit trends providing real time alert when an expense is flagged as high risk. Manager can sort and filter expenses between policies, cost centers, departments, and drill down to the high risk individuals/employees and top expense areas. There is a misnomer that predictive analytics can be used in isolation for fraud detection.

A detail to note is that predictive analytics cannot be used in isolation for fraud detection. For example, in the case of PDS2 (Payments Services Directive) across EU member states, lack of historical data would starve predictive analytics of training data rendering it ineffective in the short term. In such cases, the risk can be mitigated through the use of a hybrid detection methodology, involving the use of business scenarios and detection of anomalies through the use of experienced peer groups.