Drafting the 10-K With AI, and the Line You Can't Cross

Disclosure drafting is a strange kind of work — enormously high-stakes and, at the same time, deeply repetitive. Much of a quarterly filing is last period's language with the numbers rolled forward and a few events described. That combination, high stakes plus repetition, is exactly where GenAI is both tempting and dangerous.

Let me be clear up front about where it helps. Grounded on your prior filings, a GenAI model can draft the MD&A narrative, roll forward footnotes, update boilerplate, and produce a genuinely usable first pass in a fraction of the time. It knows how you phrased the revenue recognition policy last year and can update it. It can flag where this quarter's numbers moved enough that the narrative probably needs new language. For a reporting team staring down a tight filing calendar, a strong first draft is worth a lot.

Grounding is the whole game

The word "grounded" is doing heavy lifting. A GenAI model left to write disclosures from general knowledge will produce fluent, professional-sounding text that is confidently wrong about your company. That's a nightmare in a document that gets filed with the SEC. So you constrain it hard: it drafts only from your actual filings, your actual trial balance, your actual approved numbers. It retrieves and assembles; it does not invent.

Even then — and I can't stress this enough — every number and every assertion gets human-verified. The model can transpose a figure or misattribute a variance in prose that reads perfectly. The review doesn't go away. It changes shape: your team spends less time drafting and more time verifying, which is honestly the better use of a technical accountant's time anyway.

What actually improves

Cycle time on the drafting stage comes down. Consistency improves — the same defined term is used the same way throughout, cross-references line up, prior-period language stays intact where it should. Tie-out is easier when the draft is generated from the source numbers rather than retyped. Those are real gains and they're worth pursuing.

The risks are equally real. Hallucination in a filing is a control failure, full stop, so the verification gate is non-negotiable and should be documented as a control. Materiality judgments stay with humans — the model can suggest that a change looks significant, but whether something's material is a call for qualified people who understand the business. And be thoughtful about confidentiality: this is your most sensitive pre-release financial data, so it belongs in an environment that treats it accordingly.

Would I let AI file a document? Never. Would I let it hand my reporting team a solid draft so they can spend their nights checking rather than typing? Absolutely. The distinction between those two things is the entire discipline. Get it right and you get a faster close and a sharper review; blur it and you get a fluent, filed mistake.