On September 8, 2023, the IRS announced that it is leveraging artificial intelligence (AI) to target large partnerships. https://www.irs.gov/newsroom/
And not only has the IRS been unsuccessful in generating revenue from these audits, but large partnership audits are time consuming. IRS revenue agents spend on average 345 hours on each large partnership audit, and the audits last on average 1.7 years. The average audit closed 2.99 years after the original return was filed. This was just inside of the three-year statute of limitations on assessment. And less than one percent of these audits had a statute extension signed – surely due to the seasoned practitioners representing the taxpayers in these audits. This indicates that many of these audits may be closed because the IRS revenue agents assigned were unable to identify issues within the three-year statute of limitations, and the taxpayers wisely decided not to provide the agents with additional time.
Because of its poor selection of large partnership returns for audit, and the lack of revenue generated by such audits, by 2019 the IRS had largely abandoned large partnership audits. The IRS audited 54 large partnerships out of more than 20,000 that filed tax returns in 2019 (an audit rate of 0.3 percent). But despite these difficulties, the Government was unwilling to give up on enforcement of large partnerships. Congress granted the IRS more expansive powers when auditing large partnerships and the IRS reworked its audit selection tools.
With the enactment of the Bipartisan Budget Act, Congress centralized the partnership audit regime, effective beginning in the 2018 tax year. That regime made it easier for the IRS to audit large partnerships by (1) allowing tax adjustments and assessments to be made at the partnership level and (2) establishing a partnership representative with sole authority to act for the partnership in the audit. These changes to the previous (TEFRA) partnership audit regime made it easier for the IRS to resolve audits, assess adjustments, and collect the tax due.
Even with this increased statutory authority, the IRS still needed to figure out how to better select large partnerships for audit. The IRS has various methods for selecting returns for audit. But the most prevalent is statistical modelling. Generally, the IRS will run filed returns through its statistical models to identify returns that have a higher risk of noncompliance. Partnership returns will be run through these models twice a year. And additional runs are made whenever the IRS needs to increase its audit inventory. Classifiers then manually review returns identified as higher risk. Those classifiers will identify specific issues to audit on a return and then forward the case to an audit manager to assign to a revenue agent.
The IRS uses two statistical models to review partnership returns – the Partnership Model and the Large Partnership Compliance Model (LPCM). The Partnership Model has been used since 2018. It reviews all partnership returns for which the IRS’s Large Business & International (LB&I) Division is responsible (those being all partnership returns filed showing $10 million or more of assets). For the 2020 tax year, the Partnership Model reviewed 249,464 partnership returns with $10 million or more in assets. As of April 2023, 400 of those returns had been selected for audit.
The LPCM is the IRS’s newest statistical model aimed at partnerships, and is currently, as its name implies, being used exclusively to identify audit potential for large, complex partnerships. The new model uses multiple indicators of risk for noncompliance that are based on accounting rules, tax law, and a machine learning algorithm. IRS subject matter experts assign importance to different metrics, which the LPCM then uses to classify the risk of noncompliance for each return as very low, low, medium, high, or very high. The IRS is quite impressed by this statistical model, and has classified it as an AI system. Federal agencies are required to publish lists of AI use cases, see Executive Order 13960, Promoting the Use of Trustworthy Artificial Intelligence in the Federal Government, and the IRS includes the LPCM on its list. (Interestingly, the Treasury Department has included a redesign of the National Research Program on its latest list of AI use cases.)
But putting a name to something does not make it so. The Government Accountability Office (GAO) found deficiencies in the development and testing of the LPCM. See GAO-23-106020, IRS Audit Processes Can Be Strengthened to Address a Growing Number of Large, Complex Partnerships.
First, IRS subject matter experts tested returns that the LPCM flagged as high risk to determine whether it was correctly identifying issues with audit potential. But checks were not done of returns identified as lower risk to determine if the LPCM was failing to identify issues with audit potential. Therefore, the LPCM was developed with the use of unrepresentative data. The IRS conceded that this was due to budget and time constraints. Second, the GAO also found that the LPCM used untested assumptions to select assign audit risk to returns. As one example, assumptions used to develop the business risk indicators on partnership returns and the weighting factors used to develop aggregate risk scores were drawn from interviews with IRS subject matter experts. These individuals’ views were not tested against outcomes from audited returns or validated research. The GAO notes that these shortcomings violate best practices of statistical modeling and AI development.
Ultimately, the issues identified by the GAO can be fixed. And the GAO provides guidance to the IRS on the steps that it can take to improve the LPCM. But the IRS used the LPCM in April 2023 to identify 150 large, complex partnership returns from the 2021 tax year as having high audit potential. The IRS has assigned 75 of those returns for audit. The September 8, 2023 public announcement of those audits touts the LPCM, and specifically ties its development and deployment to the Inflation Reduction Act funding.
Based on the GAO’s findings, the IRS’s public relations efforts seem like a gamble. There may be deficiencies in the LPCM that tarnish its audit selection. And the realities of these large partnership audits, including the relative resource and information disparities between these wealthy partnerships and IRS field agents, make it difficult for the IRS to sustain successful audit adjustments. If the IRS’s large partnership enforcement shortcomings continue with these 75 large partnership audits, it may diminish support the IRS’s needed technological development and additional funding. But the IRS has apparently taken a calculated gamble that either (1) its efforts will be successful or (2) the announcement will garner positive press, and the potential negative results of these audits will not garner any press.