The accrual looked right. A model had estimated it, the number was plausible, it slotted neatly into the deck, and the meeting moved on. A quarter later an auditor asked the ordinary question — show me how you arrived at this figure — and no one could. The model had produced a number, not a calculation. There was no formula to walk, no inputs to re-run, nothing to reproduce. The figure was fluent and confident and completely unaccountable, which in accounting is another word for wrong.
We are not skeptics about AI in finance. It is genuinely useful, and we build with it. It drafts narratives, flags the anomaly hiding in ten thousand rows, proposes a mapping a person would take an hour to find, summarizes a contract's revenue terms in seconds. What it must never do is own the number. The dollar amount is not the model's to produce, because money is deterministic and a model is not.
So we draw a hard line, and we draw it in the same place every time. AI drafts and checks. The ledger owns the dollars.
Fluency is not correctness
The thing that makes a language model so useful is also the thing that disqualifies it from owning a figure: it is built to produce the plausible. A plausible accrual and a correct accrual look identical on a slide and are completely different things to an auditor, a regulator, and a CFO who has to sign. Correct, in accounting, has a specific meaning — it can be reproduced. Run the same inputs through the same logic and you get the same dollar, today and at audit and a year from now. That property is called determinism, and it is the whole foundation the ledger stands on.
A model does not offer it. Ask the same question twice and you may get two answers, each reasonable, neither reproducible. That is fine for a draft and fatal for a posting. The dollars have to be calculated — in SQL, in code, in a formula someone can point at and re-run — never generated.
Where AI helps, and where it must not
The line is not "AI good" or "AI bad." It is a line down the middle of the workflow, and it sits in exactly the same place every time:
- Drafting the narrative — yes. The flux commentary, the variance explanation, the first pass at a footnote. A human edits and owns it, but the blank page is the model's to fill.
- Flagging the anomaly — yes. "These twelve entries don't fit the pattern." Surfacing what deserves a look is exactly what it is good at. Acting on it stays with a person.
- Suggesting a mapping — yes. Proposing which account, which match, which classification — as a suggestion a human confirms, not a decision it commits.
- Calculating the dollars — no. The amount that posts is computed by deterministic logic you can re-run and audit. The model never produces the figure itself.
- Owning the posting — no. What hits the ledger is owned by code and a person, traceable end to end. AI is upstream of the entry, never the entry.
Read the list and the principle falls out: AI belongs everywhere around the number and nowhere inside it. It accelerates the judgment and the prose. It does not get to be the calculation. Keep it on the right side of that line and it is one of the most useful tools in the close. Let it cross, and you have imported plausibility into a place that only accepts proof.
The cost of letting it own the number
A figure that cannot be reproduced does not just fail its own audit test. It poisons confidence in everything next to it. Once one number on the page turns out to be ungrounded, every other number inherits the doubt, because the reader no longer knows which ones were calculated and which were generated. The work of restoring that trust costs far more than the time the model saved, and it is paid at the worst possible moment — under audit, in front of the people you least want to tell that you are not sure where a number came from.
This is the same discipline that runs through the series, stated at a different layer. An ERP only accepts what reconciles because the posting must be provable; a good integration runs twice because the result must be reproducible. Determinism is the thread. AI is welcome on a deterministic system — as long as it stays out of the math.
It ties back to the ledger
The numbers hold up because every dollar on the page can be walked back to a calculation, not a generation. AI did real work to get there — it drafted the words, caught the outlier, proposed the match — but it never touched the figure. That is the arrangement we build toward and defend: the model drafts, the ledger owns, and the dollars are calculated, never guessed. Used that way, AI makes the close faster and the numbers no less certain. That is the only version worth deploying.