SUMMARY:   AI and data analytics have become indispensable in SBA’s fraud prevention toolkit. They allow SBA to scrutinize 100% of applications (something not feasible manually) and do so in milliseconds, effectively stopping many fraudulent loans in their tracks while letting legitimate ones proceed unhindered. SBA’s ongoing investment in this area – building data infrastructure, obtaining new data sources, refining models – will only increase the precision of fraud detection and thereby protect program integrity without encumbering the application process for genuine borrowers.

SBA’s Overarching Strategies for Fraud Risk Management

In this section, we delve into the concrete initiatives and tools the SBA has deployed to combat systemic fraud in its loan programs, while streamlining processing for honest borrowers. Over the past few years – especially accelerated by lessons from the pandemic – SBA has significantly modernized and strengthened its fraud-fighting arsenal. These initiatives range from cutting-edge AI-driven fraud detection systems, to old-fashioned verification of identities and financial information, to revamped oversight of lenders and third parties. We will examine each in detail, providing examples and case studies illustrating their impact. We’ll also look at how SBA has adapted policies in response to past fraud cases (what went wrong and what changed thereafter), and how SBA works with partners (other agencies, auditors, financial institutions) as force-multipliers in fraud prevention. The tone here is explanatory, aimed at assurance professionals who want to understand not just what measures are in place, but how they work and why they matter.

AI-Driven Fraud Detection and Data Analytics

One of the most transformative developments in SBA’s fraud prevention strategy has been the adoption of artificial intelligence (AI) and advanced data analytics to detect fraud. Historically, fraud detection in loan programs relied on manual audits or basic rule checks, which could not scale to millions of applications. SBA’s recent experience forced a leap into automated analytics.

By 2021, SBA deployed first-of-its-kind AI tools to scan loan applications for suspicious patterns or anomalies, effectively acting as a digital fraud filter (U.S. Small Business Administration Releases Report on Anti-Fraud Control Measures in Pandemic Relief Programs | U.S. Small Business Administration) (U.S. Small Business Administration Releases Report on Anti-Fraud Control Measures in Pandemic Relief Programs | U.S. Small Business Administration).

These AI-driven checks operate at high speed and volume, flagging potentially fraudulent applications in real-time. According to SBA, using these tools it was able to block over 21.3 million loan applications during the pandemic relief efforts, preventing an estimated $511 billion in potentially fraudulent or ineligible disbursements (U.S. Small Business Administration Releases Report on Anti-Fraud Control Measures in Pandemic Relief Programs | U.S. Small Business Administration) (U.S. Small Business Administration Releases Report on Anti-Fraud Control Measures in Pandemic Relief Programs | U.S. Small Business Administration).

This is a staggering figure that highlights how indispensable AI has become – no team of human reviewers could have intercepted that magnitude of bogus claims in short order.

How do these AI and analytics systems work? While specific algorithms are not public, we know they look for fraud indicators and patterns gleaned from known fraud cases. For instance, AI might cross-check application data against external databases to catch fake identities (e.g., flagging if an SSN or EIN doesn’t match the applicant’s name or if the business was not active as claimed).

Machine learning models can be trained on confirmed fraud cases to recognize subtle signals – such as multiple applications originating from the same IP address or physical address (a sign of organized fraud rings), or inconsistencies between a business’s stated payroll and industry norms.

In PPP, data analytics helped identify that some borrowers overstated employee counts or payroll and thus got larger loans than allowed. Those anomalies led SBA to deny forgiveness on about $4.7 billion in loans – essentially catching the misuse through data-driven review (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO) (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO).

Another example: SBA’s analysis found many duplicative or cross-program applications (entities trying to get both PPP and EIDL loans improperly or applying multiple times). AI models were tuned to spot these duplicates and reject them automatically, sparing SBA from paying out multiple times to the same fraudulent actor (U.S. Small Business Administration Releases Report on Anti-Fraud Control Measures in Pandemic Relief Programs | U.S. Small Business Administration) (U.S. Small Business Administration Releases Report on Anti-Fraud Control Measures in Pandemic Relief Programs | U.S. Small Business Administration).

A case study illustrating AI impact is the fraud indictors analysis performed by GAO and SBA. GAO, using its own analytic algorithms, identified over 3.7 million unique recipients of pandemic loans with fraud indicators (such as nonexistent businesses or inflated employee numbers) and referred them to SBA OIG (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO) (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO).

SBA’s internal analytics similarly churned through loan data and flagged hundreds of thousands of suspicious loans. In fact, SBA reported referring over 669,000 potentially fraudulent PPP and EIDL loans to OIG after using data analytics and manual review to confirm red flags (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO). These referrals often came from patterns AI picked up. For example, if an address was used by dozens of different loan applicants, the system would tag those for closer inspection – many turned out to be fraud rings using drop boxes or fake locations.

Importantly, AI has enabled SBA to implement “continuous monitoring”. Rather than one-time checks, the agency can re-run analytical models on evolving data (e.g., as new data sources are added or as loan forgiveness applications come in) to spot fraud that might have evaded initial detection. SBA’s analytics capabilities were enhanced during the pandemic and the agency has recognized the need to further develop this into a mature program (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO) (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO).

Going forward, SBA has opportunities to better integrate external data into its analytics. Currently, limitations exist – for instance, SBA did not initially have access to certain federal wage data (the National Directory of New Hires) that could have verified if businesses truly had employees on payroll (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO) (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO). GAO recommended SBA pursue access to such databases (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO) (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs | U.S. GAO).

In response, SBA has been exploring data-sharing agreements (for example, with Treasury or IRS) to feed its AI with richer data, which will improve detection accuracy without slowing applications (the checks happen behind the scenes at machine speed).

One tangible initiative is the creation of a dedicated Data Analytics Team or Project within SBA. The agency started a Data Analytics Strategy Project to map out current and future fraud analytic needs (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs) (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs).

This includes performing a gap analysis and developing a roadmap for a “mature data analytics program” within SBA (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs) (COVID Relief: Fraud Schemes and Indicators in SBA Pandemic Programs). By institutionalizing AI/analytics, SBA is moving away from ad-hoc tools toward a sustained capability. The Fraud Risk Management Board, mentioned earlier, likely oversees this effort, ensuring that data analytics is applied enterprise-wide, not just in pandemic programs.

For example, even in core 7(a) lending, SBA can use analytics to spot outlier loans (say, a lender whose default rates spike or an industry category showing anomalies in loan performance, which could hint at fraud or abuse).

From an assurance perspective, AI-driven detection dramatically improves SBA’s fraud risk coverage but is not without challenges. Models must be well-calibrated to avoid excessive false positives that burden honest borrowers. SBA has to continually train and update its algorithms as fraudsters adapt.

There is also a need for human analysis to backstop AI – flagged cases often require an investigator or loan specialist to review and confirm if it’s truly fraud or just an unusual but legitimate case. SBA’s approach thus far merges the two: automated mass-screening combined with “fusion teams” of analysts and investigators who dive into the AI hits.

This hybrid approach paid dividends: SBA’s enhanced analytics helped produce numerous leads that led to enforcement – by May 2023, OIG’s work (fueled in part by SBA data) had resulted in over 1,000 indictments and 529 convictions for COVID-19 loan fraud (COVID-19 Pandemic EIDL and PPP Loan Fraud Landscape) (COVID-19 Pandemic EIDL and PPP Loan Fraud Landscape).