Most finance teams can close the books faster than they can explain them. The numbers come out of the ledger and the reporting model within days; the words come out of email threads, calls with business partners and the controller’s memory, and they take longer. Variance commentary, the sentences that say why revenue, margin or costs moved against plan, is usually the most time-consuming part of the monthly pack and the part most likely to hold up its release.
Generative AI changes the economics of that work. A language model can read a variance bridge and write a clear first draft in seconds. The question for finance leaders is not whether to use it, but which parts of the commentary it should write, which parts it should leave to people, and how the pack stays accurate when a machine writes the first draft.
What AI drafts well: movements the bridge already explains
Much of the commentary in a monthly pack restates the bridge in words. Revenue was ahead of plan because volume in the Northeast was higher and prices held; cost of goods was over plan because freight rates rose. When the bridge is built on certified data, with volume, price, mix, currency and cost drivers already calculated, the explanation is arithmetic plus grammar. That is exactly what a language model does well. The comparison matters as much as the movement: against plan, against last year or against the latest forecast. Settle which comparison each section of the pack uses, and the model will apply it the same way every month.
The draft should be generated from the governed numbers, never typed into a chat window. The reporting model computes the variances and their drivers, a simple rule selects the lines above the materiality threshold, and the language model writes one comment per line: the movement, the driver, the amount and the comparison. If the bridge is not yet part of the monthly routine, build it first; everything else rests on it.
What still needs the controller
Three kinds of commentary cannot be drafted from the bridge alone.
- Judgment. Whether a variance matters, whether it reflects timing or a trend, and what the business should do about it. The model can say that the Midwest is behind plan; only the people close to the business can say whether it will recover.
- One-off items. A legal settlement, a reclassification, the reversal of last quarter’s accrual. These rarely leave a trace the model can read, and a confident but wrong explanation of a one-off does more damage than no explanation at all.
- The view forward. What the month means for the quarter and the year, which risks and opportunities have changed, and what leaders should expect next. Executives read this part most closely, and it should carry a name.
A simple test separates the two kinds of work. If an explanation can be traced to a number in the reporting model, the AI drafts it. If it depends on something that happened outside the systems, a person writes it.
Design the review, not just the draft
The risk in AI-written commentary is not clumsy prose. It is a plausible sentence attached to the wrong number. The review workflow is what keeps the pack accurate, and it has four parts.
- Numbers come from the model. Every figure in a comment is inserted from the reporting model rather than generated by the language model, so the reviewer checks the reasoning, not the arithmetic.
- Known exceptions are flagged first. Lines with one-off items, manual entries or late adjustments are marked for the controller before drafting starts.
- Every edit is recorded. The controller accepts, edits or rewrites each comment, and the system keeps the history, so the team learns where the drafts fall short.
- Sign-off does not change. The pack goes out under the controller’s name, with the same approval as before. AI changes who writes the first draft, not who is accountable.
Review time should go where the risk is. A comment that restates a large, clean volume variance needs a glance; a comment on a line with manual adjustments needs a careful read.
Keep one version of the numbers
AI commentary loses trust quickly when the words and the tables come from different places. If the draft is written from an extract taken on Tuesday and the pack is refreshed on Wednesday after a late adjustment, the comment and the table disagree, and readers stop believing both. Generate the commentary from the same certified model that produces the pack, in the same refresh, and regenerate it whenever the numbers change.
Definitions follow the same rule. If gross margin includes freight in the tables, it includes freight in the words, and the language model is told so in its instructions rather than left to guess.
Write the house style down once
Good commentary has a recognizable shape: lead with the movement, name the driver, quantify it, and stop. Write that style down once, including sentence length, sign conventions, favourable and unfavourable wording, rounding and the names of business units, and give it to the model as standing instructions, with a few comments written by your best controller as examples. Consistency is one of the quiet benefits of AI drafting. Every business unit’s commentary reads the same way, which makes the pack faster to read and real differences easier to spot.
What changes for the finance team
The controller’s month shifts from writing to reviewing. Analysts spend less time chasing explanations for small movements and more time on the few that matter. Business partners see a first draft as soon as the close is done, so their questions arrive earlier. Because every edit is recorded, the team can see which kinds of commentary the model handles well and tighten the rules where it does not.
Start with one section of one pack, usually revenue and margin, where the bridge is most mature. Run the AI draft alongside the manual one for a few closes, compare them line by line, and switch when the controller is satisfied. The pack that results is not written by a machine. It is written faster, reviewed more carefully and signed by the same person as before.