Industries / Finance and business services / Legal and professional services / Accounting / Tax preparation
Individual returns
Individual returns are worked out from a person's income, deductions and credits, by the taxpayer with software or by a paid preparer, to find the tax owed or the refund due.
4
Milestones achieved, from 1955 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1955
Preparation sold to the public
The Bloch brothers had prepared returns free of charge as a sideline of United Business, the Kansas City bookkeeping firm they formed in 1946. When the IRS stopped offering such help to the public in 1955, a client who sold advertising for the Kansas City Star talked them into two newspaper advertisements, and on the first day the office was "flooded with customers". They sold the bookkeeping business and reincorporated as H&R Block, which took in $20,000 in its first year and by 1967 had nearly 1,700 offices in 1,000 cities in 44 states.
Credit
Henry and Richard Bloch
1996
UK taxpayers work it out
Under Self Assessment, which has applied in the UK since the tax year 1996-97 under legislation in the Finance Act 1994, the taxpayer files a return instead of waiting for the tax office to assess the tax, and a paper return sent after 31 October must include the taxpayer's own calculation of the tax due. Online filing, which does the calculation automatically, opened on 3 July 2000, and 38,000 people had used it by 31 January 2001.
Credit
UK Parliament; the UK tax office, now HM Revenue & Customs (HMRC)
2012
Refund loans pushed out
A refund anticipation loan (RAL) is a short loan, sold through a preparer, against the refund a client expects. Lenders relied on the IRS's "debt indicator", which said whether any part of a refund would be taken to pay other debts; in August 2010 the IRS said it would stop providing it from the 2011 season. Regulators followed: the Office of the Comptroller of the Currency (OCC) ordered HSBC, H&R Block's bank, to stop making RALs, and in December 2011 Republic, the last RAL bank, settled with the Federal Deposit Insurance Corporation (FDIC), paid a $900,000 penalty and agreed to stop after 30 April 2012. The Consumer Federation of America and the National Consumer Law Center counted 7.2 million RALs in 2009, with $606 million in fees; a typical $1,500 Republic loan cost $61.22, plus $29.95 for a refund anticipation check, a temporary account the refund passes through, which the groups put at an annual rate of 149%.
LessonWithdrawing one piece of IRS data that lenders' risk checks depended on was enough to end a product sold to millions of people a year.
Credit
IRS; OCC and FDIC; the two consumer groups, which counted
2026
Models finish most test returns
TaxCalcBench, a test published by the tax software company Column Tax, gives a model a made-up taxpayer's papers and scores whether the whole return comes out exactly right. Its tax-year-2025 version, released in June 2026, has 50 cases whose inputs are "realistic PDFs" such as W-2s and 1099s, and covers state returns, among them California, New York and Virginia. The best result, Anthropic's Claude Fable 5.2 with web search, got 86% of returns exactly right; Meta's Muse Wonder and Google's Gemini 4 Hydrogen, both with web search, got 84%. Each model ran once on each case.
LessonColumn Tax scored 155 web-search runs as wrong because a model had searched for the test case itself, so the answers to a public test can leak into the runs it scores.
Credit
Column Tax, which tests other companies' models
Watch for
Agents that ask for facts
Columbia University's DAPLab describes TaxLab, dated July 2026: 3,142 rules over 63 issues in US federal individual income tax, and 55 made-up tax-year-2025 cases in which some facts are held back by a simulated taxpayer who "answers only the specific question asked", so an agent must notice what is missing and ask for it. The lab reports that Claude Opus 5.5 got every key field labelled by a certified public accountant (CPA) right in 64.0% of cases, and that Fable 5.1 got the final Form 1040 right in 81.8%. The benchmark has not been released.
Achieved whenthe lab releases the benchmark and its test environment, as it says it will, and someone outside the lab reproduces the scores.
Claimed by
Columbia DAPLab, July 2026
Repeat runs disagree
Rivet, which calls itself an AI-enabled accounting firm, tests models on more than 500 prompts drawn from its own client work, with outputs checked by more than 40 accountants. Its best model, GPT 5.5 Pro, scored 77.7 on a first try (pass@1) but 29.3 on pass^5, the share of prompts answered right on all five of five tries. The prompts are not public.
Achieved whenRivet releases the synthetic dataset it plans, or another tester finds a similar gap on public data.
Claimed by
Rivet, 2026 (the page is undated)
Open
01
Find the planted mistakes
A public test of AI review, the step in which a senior preparer checks a finished return against its papers before it goes out. Take 200 made-up taxpayers with correct federal returns, plant one to four mistakes in each of 150 of them (a figure copied wrong, an income statement left out, a deduction on the wrong line, a credit claimed by someone who does not qualify), leave 50 clean, and score each reviewer on mistakes caught and false alarms.
The briefCredit · unclaimed02
Same return, five runs
Rerun TaxCalcBench's 50 tax-year-2025 cases five times per model and publish pass^5, the share of cases a model gets exactly right on all five runs, beside the one-run score, with the list of cases that flip between right and wrong.
The briefCredit · unclaimed
Open 01 · the brief
Find the planted mistakes
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI review, the step in which a senior preparer checks a finished return against its papers before it goes out. Take 200 made-up taxpayers with correct federal returns, plant one to four mistakes in each of 150 of them (a figure copied wrong, an income statement left out, a deduction on the wrong line, a credit claimed by someone who does not qualify), leave 50 clean, and score each reviewer on mistakes caught and false alarms.
Why it is still open
TaxGPT says its review agent flags transcription, omission, classification and mismatch errors and automates "up to 90%" of review, but gives no catch rate. TaxCalcBench is "solely focused on" calculation and does not test review.
Built with what exists
TaxCalcBench's tax-year-2025 cases (MIT licence) as the clean returns; OpenTaxSolver (federal Form 1040 with its main schedules, and 11 states) or PolicyEngine's TAXSIM emulator, to recompute the effect of each planted mistake; reviewer models such as Claude and Gemini (see the Note "Seven more flagged").
First step
Take five TaxCalcBench cases, plant ten mistakes by hand, and count how many one model finds.
Done looks like
200 returns with about 400 planted mistakes and a scoring script, published; for at least three systems, the share of mistakes caught by kind and the false alarms per clean return, over three runs.
Open 02 · the brief
Same return, five runs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Rerun TaxCalcBench's 50 tax-year-2025 cases five times per model and publish pass^5, the share of cases a model gets exactly right on all five runs, beside the one-run score, with the list of cases that flip between right and wrong.
Why it is still open
The tax-year-2025 leaderboard ran each model "once per case", although the TaxCalcBench tool has a --num-runs option and can print pass^k. Rivet's private test shows how far such numbers can fall: 77.7 on one try against 29.3 on five out of five.
Built with what exists
the TaxCalcBench code and cases (MIT licence); the models on its leaderboard, through their makers' programming interfaces; the repeat-run method described in the Note "All twenty times".
First step
Run one model five times on ten cases with --num-runs 5 and count the cases whose result changes.
Done looks like
a public table for at least five models giving one-run accuracy and pass^5 over all 50 cases, the cost of the runs, and the cases that change between runs.