Machine learning platform focused on finance can spot a strip club alias a mile away

AppZen uses computer vision, data augmentation, and natural language processing to spot fraud and compare contract terms with actual bills.

Cracking the code: Why more companies are focusing on AI projects Karen Roby interviews a Gartner analyst about why some companies are doubling down on artificial intelligence and machine learning projects.

After seven years of analyzing expense reports, AppZen's algorithm has learned enough to expand its expertise beyond travel and entertainment expenditures.

AppZen's expanded platform combines natural language processing, computer vision, and a data augmentation engine to give the machine the ability to build context around all financial information.

SEE: Travel and business expense policy (TechRepublic Premium)

The updated platform can now analyze the contents of contracts and use online searches to find relevant information about expenses. This allows the platform to spot fraud, unapproved expenditures, and potential conflicts of interest. 

Co-founder and CEO Anant Kale said that the platform's expanded capabilities will allow companies to audit all expenses and invoices, instead of a fraction of these transactions.

"One hundred percent of all transactions that go through the platform are reviewed by the engine in real-time, and the result is that we are able to find things that are impossible for humans to find," he said.

Kale explained that the engine adds information to an expense report line item by searching for Yelp reviews and considering the location of a restaurant and the number of diners as well as the employee's role in the company. The algorithm uses this information to make the decision to approve or deny an expense.

"The platform is smart enough to make a decision that with this number of people at this restaurant, this spend makes sense," he said. "This information has to be gathered from multiple outlets inside and outside the company." 

Kale said his team used the question, "How would a human optimize decision making with unlimited time and unlimited resources?" to build the platform's capabilities. This highlighted the need to use data augmentation, which allows the algorithm to develop domain expertise about a company's rules, policies, and even business culture.
 
 "We had to port into a machine all these things we make decisions on when we make it manually," Kale said. 

The artificial intelligence (AI) engine also can identify unstructured data from contracts, invoices, and expenses to make sure companies are not paying for goods and services that fall outside company policies. 

Kale said the process is the same for a $10 expense report item and a $1 million dollar contract.
If the platform finds a questionable expense, say travel expenses for a companion, it flags the transaction for review by a human. The reviewer makes a decision and provides an explanation for the decision. The platform uses this information to adjust decision-making parameters.

The AppZen platform analyzes data to look for risk, such as duplicate payments or potential collusion, such as a vendor being linked to employees in a company.

"The platform can detect these things in real time, which means you can stop it from getting executed in the first place," he said.  

Spotting unapproved expenses, fraud, and conflicts of interest

The AppZen platform analyzed millions of items on expense reports to train the algorithm to tell the difference between approved expenses and expenses that fall outside company guidelines. The algorithm found small-dollar problems, like Starbucks gift cards added on to a legitimate morning coffee run. A $25 receipt from Starbucks maypass the meal limit threshold set by company policy, but the AppZen platform can catch "gift card" on a receipt and flag the item for an auditor's review. The platform also can spot doing-business-as names that strip clubs use to stay under the auditor's radar. 

The platform also looks for employees who are not just bending the rules, but breaking them by committing fraud. In one instance, two friends worked together to make money from an inflated AirBnB listing. One person posted his home at an inflated price. The other friend stayed there on business travel and submitted the expense for reimbursement. The two split the "profit" from the expense report. The AppZen algorithm noticed the higher-than-average room rate for the region and flagged the expense.  

The platform also can catch potentially illegal expenses. Some industries have strict conflict-of-interest policies that consider lunches, gifts, or services as bribes, including the Foreign Corrupt Practices Act and UK Anti-Bribery Act. The AppZen platform can scan names on meal receipts to track who employees take out to lunch and avoid the fines associated with violating conflict-of-interest laws.

Building on seven years of data

When working with a new client Kale said that his team personalizes the engine with the company size, industry, and anything unique to the firm, such as additional regulations or compliance requirements.

The engine then reviews a few week's worth of data to look for risk and non-compliance to set thresholds. AppZen presents initial findings to the client and then based on feedback from the client, the system adjusts its analysis.

Kale said that because the system has already been trained on seven years worth of data, it knows what problems to look for, which improves time-to-value for customers.

"We identify non-compliance issues within a few weeks of a customer implementing AppZen," he said.

AppZen has more than 1,800 customers in 40 countries, including banks, media companies, pharmaceutical manufacturers, aerospace companies, and software providers. Kale and Kunal Verma founded AppZen in 2012.

Also see

screen-shot-2020-01-21-at-1-18-19-pm.png

AppZen has built a machine-learning platform for finance departments that allows customers to review all transactions from invoices to expense reports.

Image: AppZen