Every month, finance writes the same document. Revenue came in 4% under plan, gross margin held, marketing overspent by $80,000, and someone has to explain why in language the board will read. Assembling the numbers takes an afternoon. The commentary takes the evening. This is repetitive, language-heavy work, so plenty of teams have started pointing a model at it.
Good instinct. Where it usually goes wrong is the wiring.
Why variance analysis fits AI
Most finance leaders are cautious about AI, and they should be. A model that guesses at next quarter's revenue is a liability. But variance analysis is not prediction. The period has closed and the numbers are final. The only open question is why they moved, and that is a writing task with a known answer sitting underneath it.
So the model is never asked to be right about the future. It only has to describe the past clearly. Done quickly, that is a job it handles well, as long as you build it so it cannot wander off the facts.
The failure mode: a plausible lie
Here is how it breaks. You paste the trial balance into a general-purpose model and ask why marketing ran over. It tells you the team front-loaded a campaign and pulled agency fees forward. The answer reads well. It is also made up. The model never saw the invoices, so it filled the gap with a story that fit the shape of the data.
For finance, that is a real problem. A variance report earns its keep because every line traces back to something. An explanation no one can verify is worse than none, because it carries the authority of the report while resting on nothing.
The governed pattern
The fix is to stop asking the model to know the numbers. Split the work into two layers.
The first layer is deterministic. Compute every figure (actuals, budget, variance, percentage) directly from the ledger with ordinary code. No model touches the arithmetic. These numbers are locked before the AI sees anything.
The second layer is the narrative. Hand the model the locked figures and the detail behind them, the transactions driving the largest movements, and give it one job: explain and summarize what is already there. It does not calculate. It does not speculate. It writes around numbers it cannot change.
Then require citations. Every driver the model names has to point to a source: an invoice, a journal entry, a department report. No citation, no claim. A reviewer should be able to click any figure in the commentary and land on the entry that produced it.
Keep the human sign-off
The last step is not technical. Before the commentary reaches the board, a controller reads it and signs off. The AI drafts; an accountable person owns the result. That approval is what makes the speed safe. It is also what auditors, regulators, and your board already expect, and building it in from day one costs nothing.
This is the pattern for governed AI in finance generally. The machine drafts, the human decides, and every claim traces back to source. Variance analysis is a clean place to start because the answer already lives in the ledger. The only thing you are automating is the writing.
Getting started
You do not need a platform. You need a clean chart of accounts, a dependable export from your accounting system, and the discipline to run one month in parallel before you trust the output. Build the figure calculation first and confirm it ties to the close. Add the narrative layer second. Keep a person approving every report until the commentary is reliably right.
Done this way, the month-end write-up stops eating the evening. And you give up nothing that made the report worth writing, because you can still prove every word of it.