BRENNEN

Industries / Finance and business services / Legal and professional services / Accounting

Bookkeeping

Bookkeeping is recording every sale, purchase, payment and receipt in the books (the ledger), so that the owner, the auditor and the tax office can see where money came from and where it went.

6

Milestones achieved, from c. 3200 BCE to 2019.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. c. 3200 BCE

    Clay tablets keep the accounts

    In Sumer and Elam, in today's Iraq and western Iran, people counted goods with small clay tokens from about 8000 BCE, one token for each unit. About 3200 BCE they began pressing the token shapes into flat clay tablets instead. The archaeologist Denise Schmandt-Besserat finds that these tablets were used only for accounting at first, and that writing grew out of this record-keeping.

    sites.utexas.edu

    Credit

    Sumerian and Elamite record-keepers; the research is Denise Schmandt-Besserat's, University of Texas at Austin

  2. 1299

    Double entry appears in Italy

    In double-entry bookkeeping every amount is written twice, once as a debit in one account and once as a credit in another, so the two sides must balance and many mistakes show. Sources disagree on the oldest surviving example. The historians Mikhail Kuter and Marina Gurskaya call the 1299–1300 ledger of Giovanni Farolfi & Company, a Florentine firm trading in Provence, "the earliest accounting ledger kept according to the rules of double-entry bookkeeping". The city museums of Genoa give Genoa "the oldest known use in public accounting", in the treasurers' ledger of 1340.

    link.springer.commuseidigenova.it

    Credit

    the Farolfi company of Florence; the treasurers (massari) of the commune of Genoa

  3. 1494

    Double entry in print

    Luca Pacioli's Summa de arithmetica, printed in Venice in 1494, included a treatise on double-entry bookkeeping, which spread the method beyond the Italian merchants who already used it. By the account of the MacTutor history archive, Pacioli borrowed freely from earlier writers.

    mathshistory.st-andrews.ac.uk

    Credit

    Luca Pacioli

  4. 1951

    A computer runs office work

    J. Lyons & Co., a British catering company, built its own computer, LEO I. In November 1951 it ran "Bakery Valuations", a program that costed the ingredients of the bread and cakes made at the company's London factory, and soon afterwards it ran the staff payroll. London's Science Museum calls the bakery job "the world's first routine, real-time office application".

    sciencemuseum.org.uk

    Credit

    J. Lyons & Co. and its LEO team, led by John Pinkerton

  5. 1992

    Bookkeeping software for small firms

    Intuit, maker of the Quicken home-finance program, released QuickBooks, a full bookkeeping program meant to be easier and cheaper than the accounting software of the time. A company history published by Gale reports that by about 2000 QuickBooks held 85% of the US retail market for small-business accounting software.

    referenceforbusiness.com

    Credit

    Intuit, founded in 1983 by Scott Cook and Tom Proulx

  6. 2019

    Digital records required by law

    Under Making Tax Digital, the UK tax authority, HM Revenue & Customs (HMRC), began requiring larger VAT-registered businesses, which pay 99% of the VAT collected, to keep their VAT records in software and file VAT returns from it. Smaller traders followed in 2022, three years later than planned. The National Audit Office (NAO) reports HMRC's estimate of £185 million to £195 million of extra VAT in 2019-20, against a cost of about £295 million by March 2023.

    nao.org.uk

    Credit

    HMRC

Watch for

  1. QuickBooks gets an accounting agent

    Intuit says the Accounting Agent in QuickBooks Online sorts transactions into categories, does the bookkeeping and helps reconcile accounts. It says its set of AI agents saves businesses "up to 12 hours a month"; that figure comes from a survey Intuit paid for, in which 45% of customers using its AI bank feed said they saved 12 hours a month.

    Achieved whena study that Intuit did not run or pay for measures the time saved and checks the agent's books against an accountant's over a full year.

    investors.intuit.com

    Claimed by

    Intuit, July 2025

  2. An agent beats hired bookkeepers

    Digits says its Bookkeeping Agent sorted 2,000 real small-business transactions into accounts with 97.8% accuracy, against 79.1% for the 12 outsourced accountants it hired for the same test, and worked 8,500 times faster.

    Achieved whenan outside group repeats the test on its own data, or the agent holds that accuracy over many months of real books, the long run in which the models in Penrose's study failed.

    digits.com

    Claimed by

    Digits, June 2025

Open

  1. 01

    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.

    The brief
    Credit · unclaimed
  2. 02

    Clean up inherited books

    A test of whether an AI agent can take over a ledger that already holds mistakes, such as duplicated payments, costs in the wrong account, personal spending and entries in the wrong month, and find and fix them without adding new ones. Bookkeepers call this clean-up work, and it is common when a new client arrives.

    The brief
    Credit · unclaimed

Open 01 · the brief

An open year of books

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

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.

Why it is still open

The long-run test that exists, Penrose's AccountingBench, uses one real company's private data. Digits tested its agent on its own 2,000 transactions, and DualEntry's benchmark of 101 short cases names no download on its page.

accounting.penrose.comdualentry.com

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").

  1. First step

    Generate one made-up company's first three months with its answer key and run one agent on month one.

  2. Done looks like

    Ten made-up companies, twelve months each, three runs per model, with every month's account balances scored against the answer key and a count of the entries the agent made up to force the books to balance; code and data under an open licence.

Open 02 · the brief

Clean up inherited books

Credit · unclaimed

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 take over a ledger that already holds mistakes, such as duplicated payments, costs in the wrong account, personal spending and entries in the wrong month, and find and fix them without adding new ones. Bookkeepers call this clean-up work, and it is common when a new client arrives.

Why it is still open

In Penrose's study, models that recorded transactions twice noticed it and failed to undo the changes, and old discrepancies left them confused about the current month. Digits' test measures sorting new transactions, not repairing old ones.

accounting.penrose.comdigits.com

Built with what exists

made-up ledgers with planted errors of known kinds, in Beancount or a SQL ledger; agents built on current models; a scoring script that compares the repaired ledger with the clean original (see the Note "Checked without being asked").

  1. First step

    Plant ten errors of five kinds in one made-up company's quarter and try one agent on it.

  2. Done looks like

    100 ledgers with ten planted errors each; the share of errors each agent finds, the share it fixes correctly and the number of new errors it adds, over three runs.