Can AI Draft Your Financial Statement Footnotes?

Public companies have a large new disclosure landing on them, and the timing is rough. Under FASB's ASU 2024-03 on expense disaggregation, public business entities have to break out purchases of inventory, employee compensation, depreciation and intangible amortization inside the relevant expense captions, for annual reporting periods beginning after December 15, 2026, with interim periods following a year later. So calendar-year filers are building this into the fiscal 2027 annual report, which means the tagging and mapping work starts long before that, and it lands on the same people who are already the last ones out of the building at year end.

Which makes it a fair time to ask the question a lot of controllers are quietly asking: can you point AI at a footnote and get something usable back? Yes, for the drafting. The judgment stays with you, and the gap between those two things is where every horror story lives.

What can AI actually do with a footnote today?

It's good at the first pass, and the first pass is most of the hours. Give a model last year's disclosure, this year's trial balance and the supporting schedules, and it will roll the numbers forward, rewrite the narrative in the same register you used before, flag the captions where the balance moved enough to need new language, and produce something that looks like a footnote in a couple of minutes instead of a couple of days. It's also very good at the boring compliance sweep - checking that every number in the note ties to the face of the statements, that the prior-year column matches what you actually filed, that a defined term is used consistently in all nineteen places it appears.

Where it earns the most is on the notes that are long, formulaic and highly repetitive year to year. The significant accounting policies note. Property and equipment. Debt maturity tables. Commitments. Fair value hierarchy tables. Anything where the structure is stable and the content is a roll-forward, which is the same reason it works so well in the close - we made that argument at length in rolling reconciliations and what AI actually changes about the grind, and disclosure drafting is the exact same shape of problem one step further down the calendar.

Where does it break?

It breaks on the part that makes a footnote a footnote, which is the judgment about what a reader needs to know. A model will happily roll forward a contingency note that stopped being accurate in March because the litigation changed, it will keep a materiality threshold you set three years ago and never revisit, and it will write a beautifully fluent paragraph describing an accounting policy you quietly changed in Q2. None of that shows up as an error. It shows up as confident, well-formatted prose that reads exactly like the note you signed last year, and that's genuinely dangerous, because your review instinct is calibrated to catch bad writing and this writing is not bad.

The other real failure mode is citation. Ask a general-purpose model for the ASC reference behind a disclosure requirement and you will sometimes get a paragraph number that does not exist, stated with total confidence. So treat every standard reference as unverified until you've opened the codification yourself. The AICPA's guidance on AI-powered tools puts it plainly: professional responsibility can't be delegated to the tool, and AI should support your judgment rather than stand in for it. That's not a disclaimer, it's the actual operating constraint, and building your workflow around it is what keeps this useful.

What does the workflow look like when it works?

Structure it so the model never has to guess. Four pieces, in order.

Ground it in your own documents. Feed it the prior-year filing, the current trial balance, the supporting schedules and your disclosure checklist as source material, and instruct it to draft only from what you gave it. A model working from your documents is doing extraction and rewriting, which it's excellent at. A model working from memory is doing recall, which is where it invents things.

Make it show its work. Require every figure in the draft to carry a reference back to the schedule and line it came from. This one change turns review from re-deriving the note yourself into spot-checking a set of tie-outs, and it cuts review time more than the drafting speed does.

Ask it what changed, before you ask it to write. Have it diff this year's balances and activity against last year's and produce a list of movements that likely need new or amended language, with the reason. Then you decide which ones matter. This is where the tool is genuinely additive, because it will surface the small movement in an obscure caption that a tired human skims past at 11pm.

Keep a human sign-off per note, in writing. Somebody with a name attaches to each disclosure and records what they checked. Your auditors are going to ask how the note was prepared, and "a person reviewed and approved this against these sources" is a very different answer from "the system generated it."

What the new expense disaggregation rules change about all this

DISE is a mapping problem before it's a drafting problem, and that's the part teams are underestimating. To disclose purchases of inventory, employee compensation, depreciation and amortization inside each relevant expense caption, you need your chart of accounts and your cost allocations to actually support that cut, and for a lot of companies they currently don't. The work is figuring out where compensation is buried inside cost of sales, whether your depreciation allocation is defensible at that granularity, and how you'll produce the same tabular disclosure again next quarter without rebuilding it by hand.

AI is useful on the mapping reconnaissance - pointing it at your chart of accounts and asking which accounts plausibly contain each of the required categories gets you a first-draft mapping in an afternoon that a human would spend a week assembling. Then you argue with it, correct it, and lock the mapping down as a rule your system executes going forward. That's the pattern that holds up generally: the machine drafts the structure, you own the decisions, and the whole thing gets cheaper every period after the first one. Same logic we walked through for revenue recognition under ASC 606 and for the monthly management reporting package.

What I'd do

Pick your dullest note and run it now, while nothing is on the line. Property and equipment, or the debt maturity table, something where you already know the right answer cold. Give the model last year's note plus this year's schedules, make it cite every figure, and time both the drafting and your review. You'll learn more from that hour than from any amount of vendor demo, and you'll come away with a real sense of where your own review instincts need to be sharper.

Then do the DISE mapping exercise before your auditors ask for it, because that's the one with a deadline attached and the one that gets expensive if you start it in the fourth quarter. If your current close and reporting stack can't produce the disaggregated cut without a manual rebuild every period, that's worth knowing in August rather than in January - our accounting and finance software directory is a decent place to see who's actually built for it. And if you're the person who ends up owning this, it's a genuinely good line on a resume, because technical accounting and AI fluency together is a rare combination right now and the postings hiring for it pay like it.

Frequently asked questions

Can AI write financial statement footnotes?

AI can produce a solid first draft of a footnote when you give it the prior-year disclosure, the current trial balance and the supporting schedules, and it's especially strong on repetitive roll-forward notes. The judgment about materiality, changed facts and what a reader needs to know stays with the accountant who signs off.

Is it safe to use AI for financial reporting disclosures?

It's workable with controls: ground the model in your own source documents, require a citation for every figure, review each note against the schedules, and record a named human sign-off. The AICPA's position is that professional responsibility cannot be delegated to the tool, so build the workflow around human review rather than around automation.

When do the new FASB expense disaggregation disclosures take effect?

ASU 2024-03 applies to public business entities for annual reporting periods beginning after December 15, 2026, and interim reporting periods beginning after December 15, 2027. Early adoption is permitted. For calendar-year filers that means the fiscal 2027 annual report.

What does DISE require companies to disclose?

It requires disaggregation of specified categories - purchases of inventory, employee compensation, depreciation, and intangible asset amortization - within the relevant expense captions on the face of the income statement, in a tabular note.

Will AI hallucinate accounting standard references?

It can, and general-purpose models sometimes produce ASC paragraph numbers that don't exist while sounding completely confident. Verify every codification reference against the source before it goes into a note.

The teams that get ahead here are the ones who run the experiment on a boring note this month rather than debating the technology for another year, so pick the note, run it, and see for yourself what it does.