Stop handing AI your forecast. Hand it your variance commentary.
The judgment stays human. The two days spent writing it up don't have to.
Every finance team I talk to wants to point an AI at next quarter's forecast. That's the wrong first job. Forecasting is mostly judgment: which assumptions to trust, which account is about to churn, whether the sales VP's pipeline is real this time. Give that to a model and you get confident numbers with nothing underneath them.
The job worth automating is the one nobody wants. Variance commentary.
Think about what actually eats your close week. The numbers are locked by day three. Then someone, usually the most expensive analyst on the team, spends two days turning a variance table into sentences. Gross margin fell 180 basis points against plan. Why? Volume was up, but mix shifted toward the low-margin SKUs, and input costs ran ahead of standard. Writing that paragraph for fourteen cost centers is tedious, repetitive, and pattern-heavy. It's also exactly what language models are good at.
The model doesn't need to understand your business to draft the commentary. It needs the bridge.
Hand it a clean price/volume/mix decomposition, the same numbers you'd drop into a waterfall, and it can write "the $2.3M unfavorable gross-margin variance was driven by a $3.1M adverse mix shift, partly offset by $0.8M of favorable volume" in the house style, every line, every month. The math is yours. The prose is the part you were never paid to enjoy.
Draw the line at causation
The trap is letting it explain causes it can't see. A model will cheerfully tell you mix shifted "due to changing customer preferences." It has no idea why mix shifted. So you draw a hard line: the agent describes what moved, sourced from your bridge; a human supplies why, sourced from actually knowing the business. Cross that line and you're back to confident nonsense, which is the fastest way to lose a CFO's trust.
Get the brief
One sharp piece on FP&A and AI, once or twice a week.
Start smaller than feels worth it
One report, one month. Take last month's actuals, feed the agent the bridge you already built, and have it draft the commentary. Then read it next to what your analyst wrote. The first pass will be 70% there and wrong in instructive ways: too hedgy, inventing causes, burying the number that matters. That gap is the work, and you close it by writing down the rules your analyst keeps in their head. Lead with the dollar variance. Name the driver. Never speculate on cause. Three sentences a line, maximum.
Do that and you've automated two days of someone's month without betting the forecast on a machine. You also walk away with something better than the time back: a written record of how your team explains its own numbers. Most FP&A teams have never put that on paper.
That's the right first project. Not the forecast. The paragraph under it.