Industries / Finance and business services / Legal and professional services / Accounting / Tax preparation
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
Milestones achieved, from 2011 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2011
A phone reads the W-2
The W-2 is the yearly statement of pay and tax withheld that a US employer sends to each worker and to the government. Intuit's SnapTax app let people photograph their W-2 with a phone; optical character recognition (OCR), software that turns a picture of text into text, moved the figures onto Form 1040EZ, the short federal return, and the return could be filed from the phone for $14.99. It was for simple returns only: no dependents, the standard deduction and income under $100,000, by USA Today's report. After a California-only trial on the iPhone in 2010, it opened in all states on iPhone and Android in January 2011.
Credit
Intuit (TurboTax)
2021
India shows taxpayers their data
On 1 November 2021 India's Central Board of Direct Taxes (CBDT) put the Annual Information Statement (AIS) on its compliance portal. It lists what banks, fund houses and others report about each taxpayer, such as interest, dividends, share and mutual fund trades and money sent abroad, and can be downloaded as PDF, JSON or CSV files. Taxpayers can mark an item as incorrect, duplicated, someone else's or from another year, and the corrected summary is used to pre-fill the return, "in a phased manner".
Credit
CBDT, Income Tax Department of India
2023
Free online filing of 1099s
Forms 1099 are information returns: statements of payments such as contract pay, interest and dividends, which the payer sends to both the payee and the IRS. On 25 January 2023 the IRS opened the Information Returns Intake System (IRIS), a free portal where businesses can key in or upload Forms 1099 without special software; the older FIRE system stayed open for bulk files. From 2024 a business filing 10 or more information returns in a year, counting all types together, must file them electronically, down from 250 of each type, under final rules of 21 February 2023. By 2026 the IRS had also changed IRIS to take in data scanned from paper information returns, the Treasury Inspector General for Tax Administration (TIGTA) reports.
Credit
IRS and US Treasury
2025
Half of Brazil starts pre-filled
A pre-filled return is one the tax office fills in from data that employers, banks and others have reported, so the taxpayer checks figures instead of typing them. Brazil's Receita Federal offers one to holders of a gold or silver Gov.br account, a verified online government identity, which 84% of filers had in 2025. Its share was 23.9% of income tax returns in 2023 and 41% in 2024, and in 2025 50.3% of the 43,344,108 returns filed on time used it, against the 57% the Receita had forecast. On the afternoon of the last day of the 2026 season the share stood at 59.8%, and the head of the Receita, Robinson Barreirinhas, said the country was "very close" to the finance ministry's goal of a return 100% pre-filled.
Credit
Receita Federal
Watch for
AI reads the K-1s
A Schedule K-1 is the statement in which a partnership tells each partner its share of income, deductions and credits, often with pages of footnotes. K1x says its Aggregator Plus, generally available from July 2025, uses "patented AI-powered data extraction" for K-1s and Forms 1099 and gives "100% digitization of key tax elements"; it gives no accuracy figure from an outside test.
Achieved whenan independent test on a published set of K-1s with footnotes reports K1x's field-level accuracy.
Claimed by
K1x, July 2025
Open
01
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.
The briefCredit · unclaimed02
Check the pre-filled data
A test of whether an AI agent can compare a tax office's pre-filled statement with the taxpayer's own papers and flag every wrong line, using the four faults India's AIS lets taxpayers report: incorrect, duplicated, someone else's, or from another year. Build 200 made-up taxpayers, each with an AIS-style statement in JSON and the bank, broker and employer papers behind it, and plant two to five faults in each.
The briefCredit · unclaimed
Open 01 · the brief
Read the hardest statements
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 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.
Why it is still open
K1x publishes no outside accuracy test, and Byron's "more than 97%" figure for K-3 and footnote extraction is its own. TaxCalcBench starts from PDFs but is "solely focused on" calculation and scores the finished return, not the reading of each document.
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").
First step
Draw one made-up K-1 with two pages of footnotes and one 12-page consolidated 1099, write their answer keys by hand, and score one model on them.
Done looks like
600 documents with at least 30,000 labelled fields and sale lots, released with a scoring script; at least three systems scored, with field accuracy (amounts within $1) given separately for clean PDFs and scans and for footnote items and box items.
Open 02 · the brief
Check the pre-filled data
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 AI agent can compare a tax office's pre-filled statement with the taxpayer's own papers and flag every wrong line, using the four faults India's AIS lets taxpayers report: incorrect, duplicated, someone else's, or from another year. Build 200 made-up taxpayers, each with an AIS-style statement in JSON and the bank, broker and employer papers behind it, and plant two to five faults in each.
Why it is still open
Pre-filling is spreading, to 59.8% of Brazil's returns at the 2026 deadline, and India's AIS asks taxpayers to correct it, but the AI tax tests I found start from the taxpayer's documents and score the return. I found no test of checking pre-filled data against source papers.
Built with what exists
the generator from the item above; the AIS categories and download formats described by the Income Tax Department; models that compare long documents side by side (see the Note "Seven more flagged").
First step
Write ten made-up AIS statements with their source papers, plant 30 faults, and see how many one model finds.
Done looks like
200 cases with about 700 planted faults; for each system, the share of faults found, the number of false flags per case, and both figures over three runs.