Copilots draft the memo. You still defend the model.

The fastest-adopted AI use case in finance isn't forecasting or reconciliation — it's drafting. Board memos, variance write-ups, investor updates, internal one-pagers: the blank page has quietly become a solved problem for a lot of finance teams. A model pulls the numbers, drafts a first pass, and a human edits it down in a fraction of the time it used to take to write from scratch.

That's a genuine win, and it's the reason drafting keeps showing up as the first AI habit finance teams actually stick with. It's also where the risk quietly moves — not away from the team, but into a different part of the process, and one that's easy to under-staff if nobody names it explicitly.

Drafting moved the risk, it didn't remove it

When a controller wrote the variance memo from scratch, the act of writing it was also the act of checking it — you couldn't explain a number you didn't understand. Drafting with a model breaks that link. The first version can read fluently, sound confident, and still contain a subtly wrong causal claim, because the model is optimizing for a plausible-sounding explanation, not a verified one.

That means the review step isn't optional polish anymore — it's the entire control. If your team is spending the same amount of time reviewing an AI-drafted memo as it used to spend reading over its own first draft, that's a sign the review is happening too casually, on the assumption the draft is "basically right" because it looks fluent.

What actually changes about the workflow

The teams handling this well haven't added a review step so much as they've relocated where scrutiny happens:

• Every claim needs a traceable source. A draft that says "marketing overspend drove the variance" needs a line item the reviewer can point to — not just a sentence that sounds right.
• The reviewer signs, not the drafter. Whoever approves the memo for distribution owns it fully, the same as if they'd written it by hand. That accountability doesn't transfer to the tool.
• Templates constrain the model's room to wander. A structured prompt that requires "driver, magnitude, source" for every claim produces a far more checkable draft than an open-ended "summarize the variance" prompt.

Where this goes wrong in practice

The failure pattern is consistent across the teams who've had a bad experience with AI-drafted memos: the tool performs well enough, for long enough, that the review gets lighter over time. Nobody decides to stop checking — it just erodes a little each cycle, because the last twenty drafts were fine. Then one draft contains a plausible-sounding but wrong explanation, it goes to the board, and the conversation afterward is never really about the tool. It's about why nobody caught it.

The model can draft the explanation. It can't be the reason the explanation was right.

A workable standard

A reasonable line, and the one most finance leaders land on once they've been through a near-miss: AI-drafted content is fine for anything internal and anything where the reviewer is genuinely re-deriving the claims, not skimming for tone. It's not fine for anything leaving the building — board decks, investor updates, regulatory filings — without a named human who traced every number back to source and is willing to defend it in the room.

That's not a step back from using AI in drafting. It's the same standard finance already applies to junior staff, applied consistently to a new kind of first draft.