The AI Forecast Won't Flag Your Bad Data

A founder asked me last month why his shiny new AI cash forecast kept missing by six figures. He'd bought the tool to stop babysitting a spreadsheet. The forecast looked beautiful. It was also wrong, every month, in the same direction, and nobody on his team could say why.

The reason turned out to be boring. Two of his bank feeds double-counted transfers between accounts. A revenue category had been miscoded for a quarter. His old spreadsheet had surfaced both problems the ugly way, as numbers that didn't tie. The model just absorbed them and produced a clean line.

The bottleneck was never the tool

A Crisil survey of more than 100 treasury teams this year found that fewer than 10% use AI for core work like forecasting or fraud detection. Half hadn't started at all. What's more useful is why. Roughly 59% named their own data quality, not the technology, as the top barrier to an accurate forecast.

That number should stop you. These are teams with real systems and real staff. They didn't say the AI was too dumb or too expensive. They said the inputs were too dirty to trust the output. The smartest forecasting engine on the market can't tell the difference between a transfer and a sale if your ledger doesn't.

A model hides the holes a person would catch

Here's the part founders miss when they buy the tool to save time. A human building a forecast in a spreadsheet trips over the bad data. A cell goes red. A total doesn't foot. The mess is annoying, and the annoyance is doing real work, because it forces someone to look.

A model removes that friction. Feed it a gap and it interpolates. Give it a miscoded quarter and it learns the wrong pattern, then projects it forward with a smooth confidence band. You don't get a fuzzy answer you know to distrust. You get a precise answer you have no reason to question, which is worse.

AI forecasting doesn't lower the bar for your books. It raises it. Bad data used to produce an obviously broken forecast. Now it produces a polished one, and the error only shows up in your bank balance.

Fix the feeds before you buy the engine

The move isn't to wait until your data is perfect. It never will be. The move is to close the loop that catches errors while they're small. Reconcile every bank and card feed monthly. Lock your revenue categories so a coding mistake gets flagged, not buried. Tie the forecast back to actuals every month and hunt down any miss over a set threshold until you know its name.

Do that, and an AI forecast becomes a genuine multiplier on a clean base. Skip it, and you've paid to automate the manufacture of confident nonsense.

The 59% are telling you where the work is. It isn't in the model.