Crushing Fraud and Protecting Public Funding with AI

In early 2020, the nation was grappling with the unprecedented economic effects of the Coronavirus Disease 2019 (COVID-19) pandemic. The Small Business Association (SBA) disbursed approximately $1.2 trillion of COVID-19 Economic Injury Disaster Loan (EIDL) and Paycheck Protection Program (PPP) funds to help eligible small business owners and entrepreneurs affected by the crisis.

To expedite this aid, key internal controls at the SBA were reduced or eliminated in favor of self-certification for eligibility. The SBA Office of the Inspector General (OIG), tasked with independent, objective oversight to improve the integrity, accountability, and performance of the SBA, needed a way to handle a surge in potential fraud, waste, and abuse (FWA), tracking down bad actors who were taking advantage of the relaxed security controls intended for American businesses in immediate need.

eSimplicity partnered with the OIG to create a solution powered by AI to define the evolving fraud landscape and identify actionable leads for investigators.

eSimplicity employed a variety of supervised and unsupervised AI/ML methods, in tandem with big data and cloud computing techniques, to uncover trends and patterns in data which we were ultimately able to correlate and link to better understood examples of fraud. Working closely with auditors and investigators, these trends and patterns allowed us to identify a set of 11 unique indicators that flagged suspicious activity in EIDL and PPP financial data and define a set of rules that auditors could trust, ultimately allowing investigators to move towards prosecutions.

Leveraging these indicators as guideposts throughout our analyses, we increased the size and scope of our search for potential fraud across the agency’s diverse COVID lending programs. We then augmented the most likely leads by linking additional records (both from within- and cross-program) by matching application and loan metadata. We performed link analysis on these potential schemes to identify additional loans associated with a source loan suspected of fraud, to create a list of potential fraud clusters. These fraud clusters, combined with actual investigative casework, OIG prior experience, and careful review to eliminate false positives and bias, served as the basis for future modeling.

Indicators on lending programs, lenders, and loan characteristics were identified through unsupervised ML methods (k-means, Guassian Mixture Models, and Isolation Forests, etc.). Potential candidate leads, with similar characteristics to traditional fraud, were flagged utilizing semi-supervised ML methods with SME oversight (autoencoders, data mining, frequent item sets, etc.). Supervised ML methods included convolutional neural networks for object detection and labeling of images.

Recently, eSimplicity headed a study to incorporate more sophisticated Natural Language Processing (NLP) methods to triage complaints received through a public hotline system used by the public to report fraud. Sifting through submission data manually is not feasible, NLP is leveraged, from simple regex/keyword, to unsupervised topic modeling, to generative AI summarization using LLMs. Additionally, we are leveraging AI capabilities to automate classification of emails for nuanced investigation tasks.

Presently, we leverage Azure AI and related services for vectorizing data to simplify complex tasks in investigative case management, summarizing large sets of data. We also have chatbots deployed to support individuals outside of audits and investigations, for corporate-level tasks.

eSimplicity’s automated FWA identification resulted in over $200B identified as potential fraud and more than $30B in seized or reclaimed funds.

Our team has identified more than 95,000 actionable leads, representing more than 100 years of manual investigative case work. These leads were passed onto investigation teams who used this data to bring wrongdoers to justice, resulting in 1632 indictments, 1,213arrests, and 1,045 convictions related to COVID-EIDL and PPP fraud.