Where AI belongs in preparing financial statements — and where it has to stop

This is a first-draft article, added to prove the structure of the articles section. The content is the owner's to edit or replace.

A lot of new accounting tools promise that "AI prepares your financial statements". The promise is attractive, and the problem with it is not accuracy. It is auditability.

Suggestion and computation are not the same job

There are two entirely different kinds of work inside preparing a set of financial statements:

• Classification. Which line item does "building maintenance expense" belong to? That is a judgement about text, language and context — precisely what language models are good at.
• Computation. What is total non-current assets? What is the transfer to the statutory reserve? That is deterministic arithmetic with one correct answer, and it needs no language model of any kind.

Conflating the two is the root of the problem. A language model that produces a figure may produce the right one nine times out of ten — and in financial statements that is not ninety per cent success. It is a defect that cannot be audited.

Why "cannot be audited"

The first question in reviewing a working paper is: where did this number come from? An acceptable answer is a chain — this balance, from this account, aggregated with those, under this rule. "The model suggested it" is not a link in that chain, because it is not reproducible, not explainable and not arguable.

The line we hold

In Qaimah the rule is stated and enforced in the architecture, not only in a written policy:

• AI is used to propose how accounts map to line items, and to read the column headers of an uploaded file.
• Every figure in the financial statements is computed by deterministic accounting code, from the trial-balance amounts you entered.
• Every model invocation is logged, every suggestion is labelled with its source on screen, and every suggestion is editable before you confirm.

What the preparer gets from that line

First, every figure traces back to an account in the trial balance. Second, regenerating from the same inputs produces byte-for-byte the same result — a property no tool that produces figures through a language model can offer. Third, responsibility stays legible: the platform prepared a draft, the professional reviewed and approved it.

That line is not a reservation about the technology. It is what makes the output usable professionally at all.