When the Board Asks Why the Forecast Moved

A finance lead I work with at a $14M SaaS company rolled out an AI forecasting tool last fall. It pulled from the CRM, the billing system, and two years of actuals, then produced a rolling 18-month revenue forecast that refreshed every night. Clean interface. Faster planning cycle. The board liked that the numbers updated on their own.

Then the March forecast came in 9% below February's. The lead director asked a simple question. What changed. She didn't have an answer. The model had reweighted pipeline velocity against historical churn and landed on a lower number, and which driver moved by how much wasn't visible anywhere in the tool. She spent the rest of the meeting defending a figure she couldn't take apart.

What the tools optimize for

FP&A software crossed a line this year. The newer platforms sit inside Excel and Google Sheets, read your billing and CRM data, and select the forecasting method on their own. Planful's Signals watches metrics in real time and flags variances against adaptive thresholds. Cube's agents answer plain-language questions and regenerate forecasts without anyone rebuilding a model. More than half of finance leaders now name AI agents a top transformation priority for 2026.

That's real progress. None of it produces an explanation. The tool hands you a number and a confidence band. It doesn't hand you the sentence you need when someone with a board seat asks why the number moved.

The model doesn't take the hit

When an AI flags a variance and you can't explain how it got there, you lose the room. The model doesn't take the credibility hit. You do.

For a growing company the exposure is sharper than at a large enterprise. A $14M company has one person who owns the forecast and the board relationship at once. There's no FP&A bench to reconstruct the logic overnight. When the model trades pipeline against churn and nobody can say so out loud, the credibility that gets spent is personal, and it doesn't come back with the next accurate number.

What to require before it reaches the board

Driver attribution comes first. The tool has to show which inputs moved the output and by how much. If it can't break a 9% drop into its parts, you're signing your name under a black box.

Second, write the assumptions down every cycle. Four sentences on what the model leaned on this month, in your words, before the deck goes out. If you can't write them, you can't defend them, and you've found the problem at your desk instead of in the room.

Keep a manual baseline running for another quarter or two. Not because the AI is wrong. Because a number you built by hand is a number you can dismantle on demand, and that's the actual skill a board buys when it hires a finance leader.

The forecast was never the hard part. Defending it is. A model that produces the first and skips the second hasn't saved you the work. It moved the work to the worst possible room.