BRENNEN

Industries / Finance and business services / Legal and professional services

Accounting

Accounting is the work of recording what a business or other organization earns, spends, owns and owes, drawing up financial statements from those records, having them checked by an independent auditor, and working out the tax due.

1854

The Institute of Accountants in Edinburgh, formed the year before, received a Royal Charter and was renamed the Society of Accountants in Edinburgh, one of the three bodies that joined in 1951 as today's Institute of Chartered Accountants of Scotland (ICAS).

icaew.comicas.com

39

Milestones achieved across 6 domains.

14

To watch for: announced, not yet confirmed outside.

13

Open deliverables. Nobody has shipped them.

In one minute

Where it started

Accounting began as record-keeping: clay tablets in Mesopotamia about 3200 BCE, and double-entry ledgers in Italy by 1299–1300, a method Luca Pacioli put in print in 1494. It became a chartered profession in Edinburgh in 1854, and the law soon made audits compulsory, for every British registered company in 1900 and for US companies selling shares to the public in 1933.

Where it stands

Most books are now kept in software, and tax offices take in invoices and records as data: Brazil has authorized 63 billion electronic invoices since 2006, and Italy, Poland and the UK require electronic invoices or digital records by law. Much of the checking is still done by people: the median organization in a survey by the benchmarking group APQC took 6.4 days to close a month's books, and the US audit regulator found that the auditor lacked enough evidence for its opinion in 23% of the audits it inspected in 2025. AI agents now sort transactions and draft reconciliations, but in an outside test in 2025 the best models drifted more than 15% from an accountant's books within months, sometimes by inventing transactions to pass the checks.

What is open

Most of the open jobs below are public tests with answer keys: a year of made-up books that any agent can be scored on, reconciliations and invoices whose right answers are known, and a test of whether invoice-paying agents fall for fake bank-change notices. Others turn published reports into data, such as the EU's new sustainability statements and the notes to company accounts. All of them use made-up records or public filings, and none handles real money or files a real return.

Domains

AchievedWatch forOpen

Narrower industries

Tax preparation

Figureheads

1 person

Advances to build with

3 research notes

Open in accounting

3 of 13 open deliverables

Open · Bookkeeping

An open year of books

A public test in which an AI agent keeps the books of made-up companies for twelve months, from bank, card, payroll and sales records, with the data, the answer key and the scoring code all published so that anyone can rerun it on a new model.

Built with what exists

Python to generate made-up companies whose correct books are known in advance; a double-entry ledger in PostgreSQL, or a plain-text ledger such as Beancount or hledger; agents built on models such as Claude, Gemini and GPT, each run several times (see the Note "All twenty times").

Credit · unclaimed

Open · Invoices and expenses

From PDF to e-invoice

A public test of how often AI turns a PDF invoice into a valid EN 16931 e-invoice, with every field right and the file passing the official EU validation rules.

Built with what exists

the official EN 16931 validation rules for the UBL and CII formats, published on GitHub under the European Union Public Licence; made-up invoices rendered to PDF from known XML files, so the XML is the answer key; models that read PDFs, such as Claude, Gemini or GPT.

Credit · unclaimed

Open · Closing the books

Reconciliations with an answer key

A public set of account reconciliations, for bank, card, payroll and dealings between companies in the same group, built from made-up records with known differences such as timing gaps, bank fees, duplicates and keying errors. It tests whether AI agents match the items, explain each difference correctly and report what they cannot match instead of forcing a match.

Built with what exists

Python to generate statements and ledgers with planted differences; agents built on current models; a model that returns probabilities for fixed-answer questions, such as Gemini Decision 1, to sort each difference into a type (see the Note "A thousand questions, one pass"); a scorer that checks each match and each explanation against the key.

Credit · unclaimed

Where it sits