AI bookkeeping changes who categorizes a transaction, not what the record has to show

AI bookkeeping changes who proposes the category. It does not change what has to sit behind that category once it is in your books. The IRS expects a supporting document that identifies the payee, the amount paid, proof of payment, the date incurred, and a description of what you actually bought. A bank feed carries three of those five.

That is the whole argument, and it is not the argument you have probably already read. The usual one is about accuracy: will the model get the category right, what share does it get wrong, at what point do you still need a person. Those are reasonable questions. They are not the ones that decide anything. The questions that decide things are what your ledger has to be able to show afterwards, and who is answerable for it.

What does AI bookkeeping actually do to a transaction in your books?

It proposes. In QuickBooks Online a downloaded bank transaction lands in the For review tab, and Intuit's own documentation is explicit that those transactions will not affect your books until you match or categorize them. The suggestion comes from a bank rule if you have one, and otherwise from your own past transactions. Until somebody accepts it, nothing has been recorded.

That gap between the suggestion and the acceptance is the most useful thing in your ledger and the least discussed. It is where the distinction lives.

A model's categorization is a proposal. The record is the entry, plus the document behind it, plus the trail of who accepted it. Three parts, and automation supplies one of them.

The last row is the one worth knowing about, because it is the case where the proposal step disappears. A QuickBooks bank rule has a setting called Automatically confirm transactions this rule applies to, and when it is on, matching transactions add themselves. Turn it off and the rule still categorizes them, but they wait in For review until you accept them. Both behaviors are legitimate. Only one of them puts a person between the suggestion and the ledger.

What does the record have to show once a model has proposed a category?

Everything it had to show before. IRS Publication 583, Starting a Business and Keeping Records, revised December 2024, says your books must show your gross income as well as your deductions and credits, and that you must keep the supporting documents behind them. The IRS page on what kind of records a business should keep adds the sentence that settles the question for anything automated: all requirements that apply to hard copy books and records also apply to electronic records. There is no automation clause. Who or what produced the entry is not a variable anywhere in it.

The specifics on that same page are worth reading once, because they are more demanding than most people assume. A supporting document for a business expense should identify the payee, the amount paid, proof of payment, the date incurred, and include a description of the item purchased or the service received showing that the amount was for a business expense. Five things.

So the categorization can be perfect and the record still incomplete. A $4,180 charge from a supplier you use every month will be categorized correctly by any competent system, because it looks exactly like the last eleven. If the invoice showing that $4,180 bought inventory rather than a piece of equipment is not kept with the transaction, the entry is a number in the right account with nothing proving it belongs there. This is the same discipline as tying every line on your balance sheet back to a document, and it does not get easier because software filled the line in.

Revenue Procedure 97-22, issued in 1997, adds the mechanics for the storage side. It requires an electronic storage system to index, store, preserve, retrieve and reproduce the stored books and records in a legible form, and it requires the taxpayer to maintain, and to give the IRS on request, a complete description of the system including the procedures for using it and the indexing system. "Our provider uses AI" is not a description of a system.

Where does your ledger record who or what made an entry?

In the audit log, and it already names the machine. QuickBooks Online creates user profiles of its own that appear in the log alongside the people, and Intuit documents four of them: System Administration, which marks changes made automatically by QuickBooks Online, including where a connected third-party app sends data or changes data that is already there; Online Banking Administration, which marks a change to your connected bank accounts made automatically by QuickBooks Online; Support Representative; and Import Administration. The log also cannot be turned off. Xero keeps the same trail under a different name: each time a transaction is created, edited or changed, an entry is added to that transaction's history with the user's name and the date and time.

This is the part of the question that has a concrete answer, and you can check it this afternoon rather than taking anybody’s word for it. Filter last month by user and see which entries have a person’s name against them and which do not.

There is one thing about that trail worth knowing before you rely on it. Intuit's documentation says events recorded in the QuickBooks Online audit log are available for two years. The IRS period-of-limitations table says to keep records supporting an item of income, deduction or credit for three years where no other rule applies, four years for employment tax records, six years where more than 25% of gross income was left off a return, and seven years for a worthless-securities or bad-debt claim. The trail explaining who made an entry therefore stops being available before the entry itself stops being required. That is not a flaw anyone is hiding; it is a retention window doing what retention windows do. It is a good reason to treat the attached document, rather than the log, as the thing that has to last.

Which transactions does automated categorization get wrong, and why?

Recurring vendor charges categorize reliably. What breaks is anything where the right answer lives outside the text of the bank line, and those cases fall into recognizable groups rather than happening at random.

None of that is an argument against automation. It is an argument for knowing which pile a transaction is in before you accept forty of them at once, and it is why the review step exists rather than being a formality somebody added for comfort.

Who is answerable for an entry a machine proposed?

Whoever accepted it, which in a business your size is usually you or the person you pay. That is not a moral position, it is what the audit log records, and it is the reason the acceptance step is the only part of this that cannot be handed to software.

There is field evidence that the acceptance step is where the difference actually shows up. Stanford Graduate School of Business reported in 2025 on research by Jung Ho Choi of Stanford and Chloe Xie of MIT Sloan, "Human + AI in Accounting: Early Evidence from the Field", dated 3 May 2025, which drew on survey data from 277 accountants and field data from 79 small and medium-sized businesses. Among their findings: senior accountants treated the system as a collaborator and stepped in when its confidence dropped, while junior staff were more likely to accept its output at face value even when that output was flagged as uncertain. The tool was the same in both cases. What differed was what happened at the moment of acceptance.

Which is why the discipline for checking automated books is the same discipline as checking a person’s. The questions in a monthly review of your bookkeeper's work do not change when the entries were proposed by a model: did the bank accounts reconcile, is anything sitting in uncategorized, does the balance sheet still tie out, is there a document behind the entries that matter. The same is true of what you should be getting handed every month. Automation changes who drafts the entry. It does not change the list, and if you are buying a service where an accountant reviews what the automation proposes, the list is what you are actually buying.

Will AI replace bookkeepers?

No, and the reason has more to do with the record than with the technology. Somebody has to accept an entry, and the IRS expects the supporting document behind it to exist and to be retrievable for the whole period of limitations. Software can propose the category and it cannot produce the receipt, decide what an ambiguous payment actually bought, or be the party answerable for the entry afterwards. What is changing is the shape of the work: less typing the category in, more deciding which proposals are wrong and making sure the document is attached.

Can AI do my bookkeeping?

It can do a large part of the data entry, and if you use QuickBooks Online or Xero it is already doing some of it. Downloaded bank transactions arrive with a suggested category or a suggested match, drawn from your bank rules or from your own history, and in QuickBooks Online they sit in the For review tab without affecting your books until somebody accepts them. What it cannot do is the two things that turn an accepted entry into a record: attach the document proving what the payment was for, and be the party answerable for the category chosen.

What does AI get wrong about bookkeeping?

The most expensive thing automated bookkeeping gets wrong is not a category, it is a missing document. A wrong category on a correct amount is visible on your profit and loss and takes a minute to fix. An entry with nothing behind it looks correct on every report you will ever run, and only fails at the moment somebody asks what it was for, which can be two years later. The categories that genuinely defeat a proposal are narrower and predictable: one-off payments with no vendor name in the descriptor, single payments covering two different things, vendors you buy more than one kind of thing from, personal charges on a business card, and loan payments or transfers between accounts you own.

What can AI bookkeeping do and not do yet?

As of September 2026, automation inside a mainstream ledger does three things well: it proposes a category for a bank-feed line, it proposes matches between downloaded payments and invoices or bills you already entered, and it flags transactions that look unlike the rest of your history. What it does not do is settle anything ambiguous, produce the supporting document the IRS expects behind an expense, or carry responsibility for the entry. The boundary is not drawn by how clever the model is. It is drawn at the point where the answer stops being derivable from the data the ledger holds.

When do I still need a human?

At the acceptance step, every time, and specifically on any transaction whose correct treatment depends on something the ledger cannot see. That means a payment covering two things, a vendor you buy several kinds of thing from, a one-off payment with no vendor name on it, anything mixing business and personal spending, and anything where the amount has to be split across accounts. It also means the month-end check itself: reconciling the bank accounts, clearing whatever landed in uncategorized, and confirming the supporting documents are attached to the entries that matter.

Is AI bookkeeping accurate enough for taxes?

Accuracy is the wrong test for a tax return, and retrievability is the right one. A return is prepared from a full year that reconciles, with categories mapped to the lines on the form and documents behind the entries, and whether a machine or a person put each category in is not something the return records. What matters is that the books show gross income, deductions and credits, that the supporting documents exist, and that you can produce them. Which expenses are deductible and how a particular item should be treated depends on your entity type and your circumstances, and it is worth ten minutes with whoever files your return before you change how you categorize anything.