Fewer Agents

Why you should read this. This week the AICPA and former IRS Commissioner Danny Werfel announced the Council on AI Risk in Tax (CART), a standing forum to test and refine the 20-risk framework Werfel published in The Tax Adviser on August 31. Martha Waggoner’s report in the Journal of Accountancy gives the details. The news arrives as the IRS widens its use of AI with far fewer people to run it.¹ Three things follow for funds and family offices. A correct return can still be costly to defend: when a screen or a newer examiner cannot connect two reported amounts, the taxpayer pays to connect them. Make the reporting trail easy to follow: complete the required statements, cross-reference them, and keep the calculation where someone can find it in a day. Know who answers: name the partnership representative, set an internal response process, and calendar each notice’s own deadline.

A partnership return can be right and still be expensive to defend. (In tax, being right is necessary; it is rarely sufficient, and it is never free.) A basis adjustment, a special allocation, or a timing difference may explain why two reported amounts do not match. When the supporting records are hard to connect, a routine question can become weeks of reconstruction.

Automation may help the IRS work more efficiently. It also makes a clear reporting trail valuable when a system or a reviewer does not connect separately reported information. The risk for clients is a higher cost of answering questions, even on correct returns; the oversight reports document the staffing and systems changes but do not measure that cost. The practical response is to complete required disclosures, preserve the support, and decide who will answer an inquiry.

Executive summary

The money and the people are down. About $26 billion of the $79.4 billion provided by the Inflation Reduction Act survives four rescissions; the IRS had spent $16.5 billion of it by March 31, 2026, including 97.7% of the enforcement account. Revenue agents fell from 10,736 to 8,609 in fiscal 2025.

AI use is growing faster than its oversight. GAO counted 126 IRS AI use cases in June 2025, most still in development, while the AI governance team lost more than three-quarters of its staff. IRS policy already treats AI that influences audit selection as presumptively high-impact; the open question is capacity.

Two initiatives show different limits. In an 82-return large-partnership initiative, 33 of the 36 exams closed by year-end 2025 ended with no change. In a separate letter campaign on partnership balance-sheet discrepancies, the IRS rejected 182 of 320 responses, often for inadequate support, and then chose not to examine any of the returns.

The practical response is a clear reporting trail. Our worked example shows how a correctly reported §743(b) adjustment can still prompt a question, and how a one-page cross-reference can turn reconstruction into retrieval.

01 What is established, and what is not

Start with the money. The Inflation Reduction Act gave the IRS $79.4 billion in August 2022, by TIGTA’s count.² Four laws have since rescinded about $53.4 billion: $1.4 billion in the Fiscal Responsibility Act of 2023; $20.2 billion in the FY2024 appropriations act; a separate $20.2 billion in the FY2025 full-year continuing resolution; and $11.7 billion in the FY2026 appropriations act signed February 3, 2026. By account, TIGTA attributes about $41.8 billion of the cuts to enforcement and $11.7 billion to operations support. Its September 1 snapshot (Rpt. 2026-IE-R014) shows about $26 billion of authority retained, of which the IRS had spent $16.5 billion by March 31, 2026. Only about $85 million remained in the enforcement account.

The IRA money is a multi-year supplement. The annual budget is separate, and it is also smaller. Congress enacted $11.2 billion for FY2026 (about $5.0 billion for enforcement), down from $12.3 billion in FY2024. The President requests $9.8 billion for FY2027; the House Appropriations Committee approved $10.2 billion in April. A continuing resolution signed September 2 carries the FY2026 rate into FY2027 through December 11.

The workforce shrank faster than the budget. TIGTA counts 31,273 separations between January 2025 and January 2026, about 30% of the workforce, most through deferred-resignation programs; about 2,000 people were hired. The IRS Data Book shows revenue agents, the examiners who handle partnership and complex individual cases, falling from 10,736 on September 30, 2024 to 8,609 a year later.

Partnership examination starts fell from 6,709 in FY2023 to 2,285 in FY2024 and 1,589 in FY2025. Much of that decline predates the 2025 departures. TIGTA also cites the transfer of staff into a new pass-through unit, training sequences that kept new agents out of partnership courses until FY2026, and onboarding demands on frontline managers.

Swipe sideways for the other columns.

Exhibit 2. What the record shows, and what it does not. Sources: TIGTA Rpts. 2026-IE-R014, 2026-IE-R009 (June 9, 2026), 2026-3S0-045 (Aug. 26, 2026), and 2025-308-022 (May 19, 2025); IRS Data Book 2025, Table 6-3; Treasury, IRS FY2027 Budget in Brief; House Appropriations Committee (Apr. 22, 2026); GAO-26-107522 (Mar. 24, 2026); IRM 10.24.1 (transmittal Aug. 10, 2026).

The best-known use of AI in audit selection deserves a careful reading. For the 2023 large-partnership initiative, a model screened 282,884 partnership returns for tax year 2021 and identified 1,617 for consideration. Staff in the Large Business and International Division (LB&I) narrowed those to 150, chose 78, and expanded the group to 82 for related transactions. The model assisted; people chose. TIGTA reported in March that 33 of the 36 examinations closed by December 31, 2025 ended with no change; 43 remained open and three had not been assigned. No-change examinations often close sooner, so the rate may move. The result may indicate that selection needs work. It does not show that sophisticated partnerships pose little compliance risk, or that smaller funds will be next.

The same TIGTA report describes a campaign closer to our subject, with limits of its own. In October 2023 the IRS sent Letter 6585 to 483 partnerships whose balance sheets did not reconcile and lacked the required explanatory statements. It received 320 responses and rejected 182 of them, for reasons including insufficient evidence and inadequate documentation; 163 partnerships did not respond. These were letters, not examinations, and in April 2024 LB&I decided not to examine any of the returns because it needed the resources for higher-priority work. The episode shows the value of required reconciliations and adequate support. It does not show that AI falsely flagged a correct, fully reported return.

Clients also ask how often the IRS examines returns like theirs. The IRS Data Book groups individuals by total positive income (TPI): the sum of the positive amounts from each income source, with losses left out. That matters for fund investors. A filer with $1.5 million of positive income and $1.4 million of losses may net about $100,000, yet sits in a much higher TPI band: rich on paper, for once in the least useful way. Exhibit 3 uses tax year 2021, the most recent year the Data Book treats as largely mature. Daniel Hood’s September 9 report in Accounting Today, drawing on J. David Tax Law’s analysis of IRS data, highlights the same pattern at the top (66 and 39 audits per thousand), and Janka Reinhardt and Robert Faltings, writing in Tax Notes in 2025, trace the long decline in individual coverage since 2010. Partnerships with $10 million or more of assets tell the steepest story: TIGTA puts their examination rate at 2.7% for tax year 2011 and under 0.1% for tax year 2023.

Exhibit 3. Historical examination coverage, tax year 2021. Source: IRS Data Book 2025, Table 3-1 and notes. Coverage counts closed and in-process examinations as of September 30, 2025, as a share of returns filed. These are rounded historical rates, not forecasts or any client’s probability. They concern examinations; they do not measure the chance of a matching notice, a refund-verification request, or other IRS correspondence.

02 Harbor’s reporting trail: a worked example

Section 743(b) gives the buyer of a partnership interest its own adjustment to the tax basis of the partnership’s assets when the partnership has a §754 election in effect (or a mandatory adjustment applies). The adjustment belongs to that buyer alone, which is why it can make one partner’s numbers look out of line with everyone else’s. (Think of the dinner guest who brought his own wine: perfectly proper, and confusing to whoever reconciles the bill.)

The facts are hypothetical. Harbor Fund IV, L.P. (“Harbor”), a buyout fund sponsored by Harbor Partners, bought a 60% interest in Northline Holdings, L.P. (“Northline”), an operating partnership, from a prior investor in 2023. Northline has a valid §754 election and is not publicly traded. In 2025 Northline sold a business line at a $50 million gain. Oakridge Family Holdings LLC (“Oakridge”), an LLC classified as a partnership, held 5% of Harbor throughout 2025.

Swipe sideways for the other columns.

Exhibit 4. Harbor’s bridge, from Northline’s gain to Oakridge’s share. Assumptions: the §754 election is valid; the $6 million is the adjustment remaining after prior recovery; the §755 allocation supports its use against the assets sold; character is preserved; the 60% and 5% shares apply to this item, with no carry or special allocation changing them. The table shows net economics, not the actual K-1 presentation.

The required reporting already tells the story. Northline’s Schedule B discloses its §754 election. Harbor’s Schedule K-1 from Northline reports the gain and, separately, the §743(b) effect: the 2025 Form 1065 instructions use box 13, code V, for negative §743(b) income adjustments, with a box 20, code U statement for partnerships that are not publicly traded. Northline also filed a statement in the 2023 transfer year. Those records show the reporting reconciliation; the underlying basis and valuation still require support.

Suppose, though, that a screening process captures Harbor’s $30 million gain component but fails to incorporate the separately reported $6 million §743(b) adjustment when reconciling the net taxable amount allocated through Harbor. That incomplete reconciliation could prompt a question even though the required reporting is correct. A model may flag a difference without connecting the explanation; so may a newly trained examiner working a long queue.

The difference is large enough to draw attention. If the full $6 million were a sustained, net positive adjustment, an unmodified imputed underpayment under §6225, computed at the 37% top rate in §1, would be $2.22 million. On the stated facts Harbor owes none of it, and modifications, character, and partner status may change any real computation. The number illustrates the potential amount in dispute; it does not establish how an IRS model would score the return.

The cost lies in the response. If Harbor’s file holds a one-page cross-reference (the source K-1 and year, the box 13 and box 20 entries, the transfer-year statement, and the basis workpaper), answering an information document request (IDR) is largely retrieval. Without it, Harbor’s partnership representative may spend weeks pulling the 2023 purchase agreement, the §755 allocation, and the valuation behind it. On purely illustrative assumptions, the cross-reference takes three hours of adviser time at filing ($2,250 at $750 an hour), and rebuilding the trail under an IDR takes forty ($30,000), most of them spent on archaeology in a 2023 data room. These are assumed costs, not typical savings. A good cross-reference does not establish the valuation or guarantee a quick resolution, but it can reduce the reconstruction work.

Oakridge cannot fix Harbor’s return, but it can inherit Harbor’s audit. If the IRS sustained an adjustment and Harbor elected to push it out under §6226, Oakridge would take into account its 5% share: $300,000 of the adjustment, an income item rather than tax. As a partnership itself, Oakridge would then push that amount out to its own members or pay the imputed amount under the pass-through procedures in the Form 8986 instructions. Oakridge must act by the applicable pass-through deadline, filing Form 8985, with Forms 8986 when it pushes the adjustments further out. A partner that is not a pass-through would instead compute additional tax in its reporting year using prior-year effects, as the Form 8978 instructions describe.

Timing matters. If Harbor filed its 2025 return on September 15, 2026, the ordinary three-year period in §6235(a) would run to September 15, 2029. Administrative adjustment requests, notices, agreed extensions, and special rules can change that date.

A PRACTICAL RULE

For each material tiered item, make the reporting trail easy to follow: identify the source K-1 and year, show the adjustment and its tax character, cross-reference the applicable statement, and retain the calculation and supporting documents. Prepare every required disclosure regardless of size. Add explanation where it helps; for returns already filed, organize the workpaper rather than assuming a new filing is needed.

03 What clients should do now

A word on the calendar. The extended deadline for calendar-year 2025 partnership returns passed on September 15; individual returns on extension are due October 15, subject to any disaster relief. For returns already filed, organize the supporting workpapers. An optional narrative is not, by itself, a reason to amend or to file an administrative adjustment request.

Exhibit 5. Actions by seat. La Presa Partners analysis; IRS guidance on partnership representatives and notices offering appeal rights; Treas. Reg. §301.7216-2; Form 8986 and Form 8978 instructions.

04 Oversight, and what to watch

What has the IRS built, and what might it build next? Exhibit 6 gives our map. Each entry carries a status, because absence from a public inventory does not prove a tool is absent internally, and a listed tool may since have been withdrawn. The risk column adapts Werfel’s labels; the mapping of risks to uses is La Presa’s analysis.

Swipe sideways for the other columns.

Exhibit 6. IRS AI uses by status: documented, reported, retired, or hypothetical. Sources: TIGTA Rpts. 2026-308-011 and 2026-400-019 (May 13, 2026); GAO-26-107522; Axios (Nov. 21, 2025); FedScoop review of the Treasury AI inventory; Werfel, The Tax Adviser (Aug. 31, 2026). The fraud-filter figures cover the broader identity-theft process, not the effect of AI alone. Use-case mappings are La Presa’s analysis.

The IRS has already written the right rule. IRM 10.24.1.4.2, dated February 10, 2026, lists AI that “informs or influences whether a taxpayer will be subject to audit, or what aspects of a return will be subject to audit” among presumptively high-impact uses. Such systems call for pre-deployment testing, an impact assessment, monitoring, human oversight, and remedies or appeals, subject to documented determinations and to waivers that only Treasury’s Chief AI Officer may grant. These internal rules do not, by themselves, establish a taxpayer right to obtain a model or invalidate an audit. The practical question is whether the IRS can apply them consistently after its AI governance team lost over three-quarters of its staff, including contractor support, during 2025 (GAO).

There is a related gap. In May 2025 TIGTA found that the IRS was not feeding examination results back into its four AI case-selection models, and the IRS agreed to do so subject to staffing constraints (Rpt. 2025-308-022). We would ask for more than a no-change rate, which alone can reward easy, low-value adjustments. For mature case groups, the IRS should publish dollars sustained and collected per staff hour, elapsed time, no-change rates, and taxpayer burden.

Taxpayers can contest any adjustment on the merits and can seek much of their examination file. The selection formula is different: §6103(b)(2) lets the IRS withhold return-selection standards when disclosure would seriously impair enforcement, and courts have long upheld the withholding of DIF scores on that basis (Small v. IRS, 820 F. Supp. 163 (D.N.J. 1992)).

EXISTING DUTIES, NEW TOOLS

OPR’s June 24, 2026 guidelines apply existing Circular 230 obligations to AI: due diligence (§10.22), fees (§10.27), competence (§10.35), firm procedures (§10.36), and written advice (§10.37). The contested piece is OPR’s reading of §10.27 that AI savings should be credited to clients. The AICPA’s Eva Simpson called that “an overly simplistic view that ignores the full economics of AI adoption,” and the AICPA is seeking clarifying FAQs. The Tax Court has made the citation point for everyone: in Clinco, T.C. Memo. 2026-16, it warned that “submitting a brief with fictitious caselaw is a recipe for sanctions.” We could not improve on that sentence, and we checked that the case exists.

Werfel’s framework is useful because it names the failure modes in plain terms (hallucination, drift, audit targeting, loss of workforce expertise, control dilution) and shows that fixing one can worsen another. As he put it this week, the profession cannot assume “that emerging AI tools will police themselves.” CART gives it a standing home.

What we would watch next:

1. December 11, 2026. The continuing resolution expires; watch whether a further IRS rescission returns.
2. The FY2027 appropriations bill. The House committee mark ($10.2 billion) sits above the request ($9.8 billion).
3. CART’s first work product, and whether IRS officials take a formal part.
4. TIGTA’s next look at the 43 open large-partnership exams. Their outcomes will say more than the first 36.
5. Master-file modernization. GAO’s September 2025 report described an FY2028 target in earlier plans, followed by pauses and reprioritization; watch for a current schedule.

Our take

The IRS will work with more software and fewer people for some time. GAO documents expanded AI use alongside substantial losses in governance capacity. We do not know which returns those tools will favor, and we would be wary of anyone who claims to. (Forecasting IRS priorities has humbled better prophets than us.) We do know who bears the first cost of an incorrect flag: the taxpayer who has to answer it.

Better records will not remove that risk, but they can make the answer clearer and less disruptive. No one, so far as we know, has ever been penalized for a workpaper that was too easy to follow. The task for funds and families is straightforward. Make the return accurate, and make its explanation easy to find.

Selected research and primary authorities

1. Waggoner, Martha. “AICPA, former IRS commissioner to lead initiative on AI in tax.” Journal of Accountancy, Sept. 23, 2026.
2. Werfel, Danny. “A risk framework for AI use in tax administration and preparation.” The Tax Adviser, Aug. 31, 2026.
3. TIGTA. Rpt. 2026-IE-R014, IRA spending through March 31, 2026 (Sept. 1, 2026).
4. TIGTA. Rpt. 2026-308-011, large partnership compliance (Mar. 18, 2026).
5. TIGTA. Rpt. 2026-3S0-045, compliance trends through FY2025 (Aug. 26, 2026).
6. TIGTA. Rpt. 2026-IE-R009, workforce (June 9, 2026).
7. TIGTA. Rpt. 2025-308-022, AI examination case selection models (May 19, 2025).
8. TIGTA. Rpt. 2026-400-019, identity-theft filters (May 13, 2026).
9. GAO. GAO-26-107522, IRS artificial intelligence (Mar. 24, 2026); GAO-25-107611, IRS modernization (Sept. 16, 2025).
10. IRS. IRM 10.24.1, IRS Policy for AI Governance (transmittal Aug. 10, 2026).
11. IRS. Data Book 2025 (Pub. 55-B), Tables 3-1 and 6-3.
12. IRS. Instructions for Form 1065 (2025).
13. IRS OPR. Introductory Guidelines for Responsible AI Use in Federal Tax Practice (June 24, 2026).
14. Waggoner, Martha. “AICPA seeks IRS clarity on AI guidelines, CPA fees.” Journal of Accountancy, Sept. 8, 2026.
15. Treasury. IRS FY2027 Budget in Brief; House Appropriations Committee, FY2027 FSGG release (Apr. 22, 2026).
16. Hood, Daniel. “IRS tax audits hit the richest — and the poorest.” Accounting Today, Sept. 9, 2026.
17. Axios. “Exclusive: IRS deploys AI agents.” Nov. 21, 2025.
18. Small v. IRS, 820 F. Supp. 163 (D.N.J. 1992).
19. IRS. Instructions for Form 8986 (pass-through partners; Forms 8985 and 8986).
20. IRS. Instructions for Form 8978 (reporting-year tax on pushed-out adjustments).
21. Treas. Reg. §301.7216-2 (disclosures and uses permitted without consent).
22. Newkirk, Kaitlin, and Sarah Webber. “A recipe for sanctions: AI-hallucinated citations in tax.” Journal of Accountancy, Sept. 2026 (discussing Clinco v. Commissioner, T.C. Memo. 2026-16).
23. Reinhardt, Janka, and Robert Faltings. “U.S. IRS Audits 2025 — Quo Vadis?” Tax Notes (2025).
24. Code §§1, 743(b), 754, 755, 6103, 6223, 6225, 6226, 6235, 6330, 7216; Circular 230, 31 C.F.R. Part 10.

Important notes

This alert addresses U.S. federal tax administration only. The Harbor example is hypothetical, with rounded figures and illustrative professional rates; it is a template for discussion, not a description of any client. Funding and budget figures are stated as of the dates shown and may change as Congress acts.

This client alert is for general information only. It is not tax, legal, investment, or accounting advice. Consult your own advisers about your particular circumstances before acting on any matter discussed here. Questions: Carlos Schmidt, CPA, La Presa Partners LLC, carlos@lapresallc.com, (917) 558-6393.

Download the alert (PDF, 318 KB)

This alert is for general information only and is not tax, legal, investment, or accounting advice. Consult your own advisers before acting. Carlos A. Schmidt, MBT, MBA, CPA · Managing Member · carlos@lapresallc.com · (917) 558-6393.