Analytics teams keep producing accurate forecasts that managers never use, and the usual cause is that nobody built a model of the choice itself: how much to make, which customers to keep, which loans to approve. Decision science exists to close that gap. The gap is about to open in every small business that asks an AI for a forecast.
A forecast gives you the middle, and the decision is at the edges
Take a baker who sells croissants. Each one costs $1 to make and sells for $3, and demand runs anywhere from 80 to 120 a day, spread evenly. An assistant asked to forecast tomorrow will say 100. If she bakes exactly 100, she sells out on about half of all days.
But a lost sale costs her $2 in profit, and a leftover croissant costs her $1. Missing a sale hurts twice as much as having one left over, so she should bake enough to sell out on only one day in three. That is about 107, by a standard rule from inventory planning that divides the cost of running short by the cost of running short plus the cost of having extra.
Nothing in the forecast changed. What changed is that she gave the model what each kind of mistake costs. A forecast that is scored on accuracy treats five croissants too many and five too few as the same error, and her bakery does not.
The same gap shows up wherever a prediction feeds a choice. A phone company can predict which customers will leave, but that does not say which ones are worth keeping or which offer would make them stay. A bank can estimate the chance that a borrower defaults, but approving a bad loan and turning away a good one cost very different amounts, and the right cut-off depends on those amounts.
Cheap prediction raises the value of the question
Getting a forecast used to mean hiring someone who could build one, so most of the effort and attention went into the forecast itself. When prediction gets cheap, the judgment about what to do with it is what is left to pay for.
Microsoft already combines the two for the people who plan its cloud hardware supply, with a tool called OptiGuide that pairs a language model with an optimization program. A planner asks a what-if question in plain words, the optimization program does the math, and the language model explains the result.
A baker or a warehouse shift manager does not need that machinery, because a general assistant can write the small calculation above in a few seconds. What it cannot supply is the $2 and the $1. Those numbers live in the owner's head, in the margin on the invoice, in how upset a customer gets about an empty shelf. I expect the people who get the most from AI forecasts in the next few years to be the ones who can say what they are choosing between and what each mistake costs.
Four lines to write before asking for a forecast
A hospital finance director who wants to plan beds for winter could write four lines before opening an assistant: who decides, what the choices are, what limits apply, and what it costs to guess too high and too low. An empty bed costs staff time. A patient turned away costs far more, in money and in other ways. Those two numbers decide how much spare capacity is right, and the forecast only tells her where the middle is.
If you cannot fill in the last line, the forecast was not what was missing, and finding those numbers is the first job.