IRS Audit Selection Criteria and Discriminant Function Scoring

IRS audit selection criteria, including Discriminant Information Function scoring, form a layered system for identifying possible tax shortfalls. A return may be selected because of statistical suspicion, a documentary mismatch, a random examination, or a referral from a connected return. Understanding those paths gives practitioners a sounder basis for preparing documentation and responding to an examination than trying to predict a hidden score.

Why the IRS audits at all: the tax gap mission that drives everything

The IRS audit program exists to close the gap between taxes owed and taxes actually paid. Its selection systems help by finding returns that may need examination and gathering evidence to improve how returns are selected in the future. DIF and UIDIF estimate potential gaps, while information-return matching identifies discrepancies in reported income. Random examinations supply the compliance data needed to keep statistical comparisons useful. This mission explains why a flagged return represents a possible shortfall and why the practitioner’s response should rest on supporting the return.

How DIF scoring works: comparing every return to its statistical peers

The Discriminant Information Function, or DIF, turns that mission into a numerical assessment of every filed return. Its proprietary formula is classified, but its central method is peer comparison. A return is compared with others that are like it, including those with the same income range, filing status, and industry. DIF measures how far its reported characteristics depart from the norms of that group. A high score signals a statistical basis for examining the departure.

A large charitable deduction is assessed against the giving patterns of similarly situated filers. The amount therefore matters in relation to the taxpayer’s circumstances, rather than simply because it looks large on the page. Statistical rarity and substantive error are different conclusions, and DIF supplies only the first signal.

For business returns, NAICS industry codes make the peer group more precise. A restaurant reporting unusually low food costs relative to revenue departs from the pattern of other restaurants. The industry classification gives the comparison its meaning because normal cost structures differ across businesses. Practitioners therefore need to understand the relationship between reported figures and the relevant business model. Looking at an expense amount in isolation misses the relationship the scoring process is designed to assess.

A high-scoring return enters a selection queue. A human IRS agent reviews flagged returns and decides whether to open an audit. The IRS does not disclose DIF scores to taxpayers or representatives, and no specific score threshold is publicly established. Practitioners must therefore assess exposure through known return characteristics. If an examination follows, the appropriate posture is narrow documentation of the flagged items, because unnecessary disclosure can expand the scope.

What actually moves a DIF score: the return characteristics that create statistical outliers

The known DIF risk dimensions translate peer comparison into features practitioners regularly encounter. Schedule C expense ratios outside industry norms matter because self-employed filers control both the income they report and the expense deductions they claim. Cash-intensive businesses, including restaurants, salons, and construction trades, present another verification problem because cash receipts are harder to trace through third-party reporting. The algorithm weights those returns differently for that reason. Neither characteristic proves understatement, but both make the relationship between reported receipts and expenses consequential.

Charitable contributions that are large relative to AGI can be outliers within the giving distribution at that income level. A home office deduction can also attract attention when it is atypical for the reported occupation. Round numbers across expense categories suggest that the filer used cost estimates. These are patterns the system weighs.

Income can create a further inconsistency when it appears insufficient to support reported expenses or evident living costs. A business loss with no other reported income source illustrates that mismatch. The practitioner’s task is to establish the factual support for the unusual relationship, rather than make the return resemble an average taxpayer’s return. Documentation should explain why the figures are correct in the client’s circumstances. The classified formula does not support more precise claims about the weight of any individual feature.

UIDIF: the parallel system scoring the income side of the return

The Unreported Income DIF, or UIDIF, asks whether reported income is likely complete. It draws on bank deposit data, lifestyle indicators, and information returns the IRS already holds to assess possible understatement. Where DIF can identify deductions that depart from peer norms, UIDIF identifies income that appears insufficient relative to other available data. The two scores are generated simultaneously, and a return can score high on one, both, or neither. Each score can lead to selection on its own, so support for deductions alone leaves other statistical reasons for scrutiny unaddressed.

Automated Underreporter matching: when the IRS already has the income data

The Automated Underreporter program, or AUR, moves from statistical inference to documentary comparison. W-2 and 1099 information filed by employers, banks, and brokers is cross-referenced against the taxpayer’s return. Income appearing on a third-party document but not on the return triggers an automated notice. The IRS is therefore identifying a specific mismatch rather than estimating that a return looks unusual. That evidential distinction is the most useful starting point when explaining an AUR notice to a client.

The response must consequently reconcile the identified discrepancy, rather than merely defend the plausibility of the return as a whole. AUR contacts often take the form of correspondence notices rather than field audits, but that form does not lessen their legal weight or documentation burden. Practitioners should treat them with the same seriousness. Form 1099-DA reporting for digital assets, effective for the 2025 tax year, extends this matching process to cryptocurrency transactions covered by the new reporting. The practical change is that the IRS receives third-party transaction information against which it can compare the filed return.

NRP random selection: how audit data feeds back into the scoring models

The National Research Program, or NRP, provides a selection path that does not depend on a suspected defect in the return. The IRS established it in 2000, replacing the older Taxpayer Compliance Measurement Program. It examines a limited sample of individual returns in each active cycle to collect compliance data. For a practitioner, every line requires organized documentation, even when nothing about the item looks unusual.

The structural value of these examinations lies in the data they return to the scoring process. Filing patterns change as income, industry mixes, and deduction behavior shift. NRP findings bring the statistical norms used by DIF up to date, helping keep peer comparisons in line with current returns. DIF then identifies departures from those norms, while examination results further inform model updates.

Related-party referrals and the whipsaw problem

A related-party referral can bring a return into examination without an independently significant score. An examiner reviewing one taxpayer may find a transaction suggesting that another party should have reported differently. These referrals are common with partnerships, S corporations, and family transactions, and an examination of a large partnership can lead to referrals to partners’ individual returns. Selection therefore extends beyond the characteristics visible on any one filing.

The clearest example of this exposure is whipsaw, in which two taxpayers take positions on the same transaction that cannot both be correct, and the IRS examines both sides to prevent inconsistent treatment from producing a tax benefit. Practitioners representing clients in multi-party structures should treat consistency across related returns as a compliance requirement. A well-supported position on one return does not resolve a contradiction with the other party’s treatment. The shared transaction is the relevant unit of review.

How the IRS is extending its selection models with machine learning and new data sources

Machine learning extends the established selection framework in targeted areas. The Large Partnership Compliance model applies machine learning to large partnership returns and incorporates expert review by tax specialists. Its population includes hedge funds, private equity, and real estate investment groups. The model selected 82 high-risk returns for tax year 2021, an increase over the program’s initial round before the AI model was applied. Certain late-filed returns were excluded from the initial run, so its coverage was not complete.

Other developments show why deployment status matters as much as the existence of a model. The Small Business/Self-Employed Division developed a return-selection model for self-employed filers, but it remained in pilot testing as of mid-2025 without a final deployment decision. A March 2026 GAO report identified 126 AI use cases in the IRS inventory as of June 2025, including many still in development. Those uses spanned audit selection, fraud detection, and taxpayer services.

IRM 10.24.1 establishes IRS-wide AI governance covering audit selection, examination support, generative AI, and other functions. It was originally effective February 10, 2026, and superseded August 10, 2026. Its rules require human oversight of final audit rulings.

How staffing reductions interact with automated selection: the flag-but-cannot-follow dynamic

The IRS lost nearly a third of its revenue agents in the first three months of the current administration, while broader auditing and collections staffing also fell substantially over fiscal 2025. Examinations of high-income taxpayers declined, and staffing in the Global High Wealth program fell sharply. Complex audits are particularly exposed to these reductions because they require substantial examiner time. Automated systems can continue flagging returns even when human capacity to pursue them is constrained.

Technology has nevertheless offset some of that pressure. In April 2026 Senate Finance Committee testimony, IRS CEO Frank Bisignano said technology had helped staff get more done, leading to higher enforcement revenue in the first five months of fiscal year 2026. The GAO identified a return-prioritization AI model that might not be deployed because the program could lack staff to conduct the resulting audits. Fewer examinations can also reduce the data available to develop and refine models, weakening the feedback process described earlier.

For practitioners, the immediate distinction is between a flag and the capacity to act on it. A return may remain in queue for an extended period, and staffing reductions do not themselves erase the flag. The statute of limitations still governs exposure, while documentation obligations remain unchanged. A delay in examination reflects an operational constraint on the IRS.

What practitioners cannot learn from the selection process, and why that shapes the advisory posture

The selection architecture is understandable at a level that the individual selection decision is not. The DIF formula is classified, practitioners can see only some of the data UIDIF uses, and the IRS does not share the taxpayer’s score. The score typically remains unknown to the assigned examiner as well, and discovery of which AI tools or scoring methods led to selection falls outside the available formal procedures. Those limits make reverse-engineering a client’s score a poor foundation for advice.

The more defensible approach is to match preparation and response to the likely selection path. DIF calls for focused support of questioned positions, AUR for reconciliation, and NRP for documentation of every line. Related-party exposure calls for checking that connected returns are consistent with each other. For high-volume practices, especially those serving Schedule C, partnership, or digital-asset-heavy clients, maintaining that support is a substantial workflow burden. Automating routine intake and document review preserves time for the risk-calibrated analysis that makes the framework useful. The practitioner’s advantage is not access to the hidden score, but the ability to prepare accurate returns whose unusual features can be explained and defended.