Industries / Finance and business services / Legal and professional services / Accounting
Tax preparation
Tax preparation is the work of gathering the records behind a tax return, working out what a person or business owes or is owed, filing the return with the tax office and answering the office's letters afterwards, whether the taxpayer does it with software or pays a preparer.
1955
Henry and Richard Bloch, who ran a bookkeeping firm in Kansas City, Missouri, reincorporated it as H&R Block to prepare people's tax returns full time, after the US Internal Revenue Service (IRS) stopped offering that help to the public, according to the Gale company history on Encyclopedia.com.
27
Milestones achieved across 6 domains.
7
To watch for: announced, not yet confirmed outside.
12
Open deliverables. Nobody has shipped them.
In one minute
Where it started
Paying someone to fill in a tax return became a mass trade in 1955, when the Bloch brothers of Kansas City turned a free sideline of their bookkeeping firm into H&R Block; by 1967 it had nearly 1,700 offices in 44 US states, by the Gale company history. Machines then took over the sums and the post. The IRS tested electronic filing in 1986 with 5 preparers and about 25,000 returns, and the UK tax office opened online filing in July 2000.
Where it stands
Most returns now travel as data, and more of the data starts with the tax office. HMRC says 97.25% of the UK Self Assessment returns filed by 31 January 2026 went online, and Brazil's tax office says 59.8% of its 2026 returns at the deadline began from a form it had filled in itself. Paper is the costly remainder: by IRS estimates that its inspector general reports, paper was 6% of US individual returns in 2025 but 72% of the cost of processing them. Given every document, AI models now calculate most test returns: the best got 86% of 50 made-up 2025 returns exactly right in Column Tax's TaxCalcBench.
What is open
The published tests ask whether a model can work out a return. I found no published measure of how often AI reviewers catch mistakes planted in a return, how well models read the longest documents, such as partnership K-1s and brokerage statements, whether they read a tax office's letter correctly, or whether they give the same return twice. Each can be tested on made-up taxpayers with open tax calculators such as PolicyEngine and OpenTaxSolver, without advising or filing for anyone real.
Domains
AchievedWatch forOpen
Domain
Progress
Last achieved
Watch for
Open
Gathering documents
Gathering documents is collecting the records a return is built from: the wage and income statements that employers, banks and brokers send to both the taxpayer and the tax office, receipts, and, more and more, data the tax office already holds.
4 · 1 · 2
Last achievedHalf of Brazil starts pre-filled2025
Watch forAI reads the K-1s
2
openIndividual 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 · 2 · 2
Last achievedModels finish most test returns2026
Watch forAgents that ask for facts
2
openBusiness returns
Business returns are the returns that companies and partnerships file: income tax returns, and the value-added tax (VAT) and sales tax returns that report the tax charged on sales.
5 · 1 · 2
Last achievedFirst global minimum tax returns2026
Watch forAgents for business returns
2
openFiling
Filing is getting the finished return to the tax office and accepted, now mostly as data sent over the internet, and getting any refund paid.
5 · 1 · 2
Last achievedNearly all UK returns online2026
Watch forAll paper, scanned
2
openAudits 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 · 1 · 2
Last achievedAI audit picks go unmeasured2025
Watch forAI to offset fewer auditors
2
openPlanning
Planning is arranging a person's or a company's affairs, within the law, so that less tax is due or it is due later; it is bounded by rules that set aside steps taken only to save tax.
5 · 1 · 2
Last achievedConfidential plans shared2023
Watch forEU trims scheme reporting
2
open
Advances to build with
2 research notes

Muse Ember
The jobs agents repeat
Meta's Muse Ember matched Claude Haiku 5.5 on an outside firm's tests at a third of the cost per task, though both list the same prices for text in and out.

Microsoft ThinkingBox
All twenty times
By Microsoft's own count, 50 steps of training took an AI model from passing 38 of 507 test tasks on all 20 tries to passing 107.
Open in tax preparation
3 of 12 open deliverables
Open · Gathering documents
Read the hardest statements
A public test of AI reading the two documents that run longest and vary most between issuers: Schedule K-1s with their footnote statements, and the consolidated Form 1099 that a broker sends, which bundles interest, dividends and every sale of shares (Form 1099-B) over many pages. Make 300 of each for made-up investors, every page marked SPECIMEN, in several fund-style and broker-style layouts, as PDFs and as scans, with an answer key for every box, footnote item and sale lot.
Built with what exists
Python with Faker for made-up investors and ReportLab to draw statements; the IRS's fillable Schedule K-1 and Form 1099-B PDFs, filled in with pypdf; Augraphy to make pages look printed and scanned; models that read long documents, such as Claude, Gemini and Meta's Muse Ember (see the Note "The jobs agents repeat").
Open · Individual returns
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.
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").
Open · Business returns
VAT return from e-invoices
A test of whether an AI agent, or a program it writes, can turn a month of a made-up German company's electronic invoices into the boxes of its VAT return: sales and purchases by tax rate, input tax and the balance due. Make 100 made-up companies with 50 to 500 invoices each in XRechnung and ZUGFeRD, the German formats of the EU e-invoice standard EN 16931, including credit notes, reverse-charge purchases (where the buyer, not the seller, accounts for the tax) and invoices that fail validation.
Built with what exists
the KoSIT validator and its XRechnung configuration (Apache-2.0 licence), to check each invoice; Python to generate the invoice files; agents built on current models.
Where it sits
Broader
In outline