Industries / Finance and business services / Legal and professional services / Accounting / Tax preparation
Audits and notices
Audits and notices are what can follow filing: the tax office's computer checks and letters, its examinations of returns (the US term for audits), and the preparer's work answering them.
4
Milestones achieved, from 1964 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1964
Random audits aim the rest
The IRS's Taxpayer Compliance Measurement Program (TCMP) began in 1964 with tax year 1963 returns: about every three years it audited a random sample of 50,000 to 55,000 individual returns line by line, and used the results to update the formulas that pick returns for audit. TCMP findings also led to new reports from third parties, of mortgage interest and of proceeds from sales of shares and homes. The last study covered tax year 1988; after a 1995 House hearing on the burden it put on randomly chosen taxpayers, the IRS was pressed to stop, and a successor, the National Research Program, began in April 2000 with lighter methods.
LessonWithout fresh random audits the IRS made decisions on data "more than 20 years old", and the Taxpayer Advocate writes that older formulas likely meant more unproductive audits of compliant taxpayers.
Credit
IRS, as described by its researchers Bob Brown and Drew Johns and by the National Taxpayer Advocate
2021
Nationality in fraud checks
Between 2004 and 2019 the Dutch Tax and Customs Administration stopped childcare benefits for thousands of families and ordered them to repay; a parliamentary commission called it an "unprecedented injustice", and Mark Rutte's cabinet resigned over it in January 2021. In December 2021 the Dutch Data Protection Authority fined the tax office €2.7 million for keeping applicants' second nationality on record and using it in deciding whether to investigate families for fraud. About 15,000 people had received €30,000 each in compensation by then.
LessonA fact with no bearing on a claim, here a second nationality, became a reason to investigate once it was in the data the checks used.
Credit
Dutch Data Protection Authority (Autoriteit Persoonsgegevens); the parliamentary commission
2023
Audit picks skewed by race
Researchers from Stanford, the University of Michigan and the University of Chicago, working with economists in the US Treasury's Office of Tax Analysis on more than 148 million returns and about 780,000 audits for tax year 2014, found that Black taxpayers were audited at 2.9 to 4.7 times the rate of other taxpayers, though the IRS does not record race. Most of the gap lay in mail audits of people claiming the Earned Income Tax Credit (EITC), a refundable credit for low earners: Black taxpayers made 21% of EITC claims but were the focus of 43% of EITC audits. On 15 May 2023 IRS Commissioner Danny Werfel wrote that the IRS's "initial findings support the conclusion" and called review of its EITC audit-selection algorithms the "topmost priority".
LessonA selection rule that never sees race still produced a racial gap; the researchers found that a focus on over-claimed refundable credits would give rates like those observed.
Credit
Hadi Elzayn, Evelyn Smith, Arun Ramesh, Jacob Goldin, Daniel E. Ho and Treasury economists
2025
AI audit picks go unmeasured
TIGTA reported in May 2025 that the IRS's models for choosing returns to audit, including the Line Anomaly Recommender, used since March 2021 on corporate returns with assets of $10 million to $250 million, and a machine-learning model for large partnerships, do not always use the results of past audits; that a high share of examinations ended with no change to the tax; and that the IRS had no specific results for the corporate model, while the partnership audits had not yet closed. The IRS agreed with all three recommendations, two of them subject to staffing.
LessonThe IRS Commissioner had told Congress in February 2024 that AI could cut such "no-change" audits, but the agency had not set up the feedback that would show it.
Credit
TIGTA
Watch for
AI to offset fewer auditors
Treasury Secretary Scott Bessent, also acting IRS Commissioner, has said that fewer revenue agents need not mean less revenue because the IRS can rely on "smarter IT" and "this AI boom", according to a letter of 26 January 2026 from ten senators led by Ron Wyden, ranking member of the Senate Finance Committee, which asks him for evidence that AI tools improve detection or yield. GAO reported in March 2026 that the IRS's research unit had lost 63 staff who worked on AI and had not identified the skills it needs, and FedScoop quotes an official who said an audit-selection model could be shelved because "the program may no longer have staff to conduct the audits".
Achieved whenTIGTA or GAO reports audit and collection results holding up as staff fall, with the difference traced to AI tools.
Claimed by
Scott Bessent, 2025, as quoted by the senators in January 2026
Open
01
Read the letter right
A public test of whether an AI agent can read a tax office letter and state correctly what kind of letter it is, what the office proposes, the amount, the reply date and the ways to reply. Make 500 made-up notices, each marked SPECIMEN, in the layouts of common IRS letters such as the CP2000 (a proposed change after third-party reports did not match the return) and the 5071C (an identity check), with an answer key for every field.
The briefCredit · unclaimed02
Checking a CP2000
A test of whether an agent can check a CP2000 against the taxpayer's papers, say whether the taxpayer should agree in full, in part or not at all, and recompute the tax. Build 300 made-up cases: some where the IRS is right (a forgotten Form 1099), some where it is wrong (a share sale reported without its cost, an income counted twice, a form issued in error).
The briefCredit · unclaimed
Open 01 · the brief
Read the letter right
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of whether an AI agent can read a tax office letter and state correctly what kind of letter it is, what the office proposes, the amount, the reply date and the ways to reply. Make 500 made-up notices, each marked SPECIMEN, in the layouts of common IRS letters such as the CP2000 (a proposed change after third-party reports did not match the return) and the 5071C (an identity check), with an answer key for every field.
Why it is still open
The IRS sends about 170 million notices a year to individuals and over 40 million to businesses, and in January 2024 began redesigning the most common ones to make them simpler and clearer; in its own trial, a two-page version of letter 5071C led to 16% fewer people calling as their first step. The published AI tax tests, such as TaxCalcBench and TaxLab, cover preparing returns, not reading letters.
Built with what exists
the IRS's "Understanding your notice" pages, which explain each notice and its reply options (upload, fax or mail), to write the answer key; Python and ReportLab to draw the notices; a fixed-answer model such as Gemini Decision 1 for the yes-or-no and choice fields (see the Note "A thousand questions, one pass").
First step
Copy the layout of one CP2000 page into a template, make 20 made-up versions, and score one model.
Done looks like
500 notices across at least 15 notice types; for each system, the share of notices with every field right and, reported apart, the share with a wrong reply date, the costliest mistake.
Open 02 · the brief
Checking a CP2000
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test of whether an agent can check a CP2000 against the taxpayer's papers, say whether the taxpayer should agree in full, in part or not at all, and recompute the tax. Build 300 made-up cases: some where the IRS is right (a forgotten Form 1099), some where it is wrong (a share sale reported without its cost, an income counted twice, a form issued in error).
Why it is still open
The IRS closed 987,460 such underreporter cases in fiscal 2025, assessing $5.9 billion, and tells recipients that a CP2000 "isn't a bill and your response may be required". I found no published test of checking one.
Built with what exists
OpenTaxSolver or PolicyEngine's TAXSIM emulator, both open source, to recompute each case; the notice templates from the item above; agents built on current models.
First step
Write ten cases by hand, five where the notice is right and five where it is wrong, and recompute each with OpenTaxSolver.
Done looks like
300 cases; for at least three systems, the share where the agent's verdict matches the key and the share where its recomputed tax is within $1.