Machine forecasts beat the committee. Then the committee overrode them.

A comparison of statistical and machine-generated forecasts against committee-adjusted final numbers across a set of mid-sized finance functions produces a result that is by now familiar and still ignored: the unadjusted machine forecast had lower error than the final published forecast in the substantial majority of periods.

The adjustments were not random. They were systematically optimistic on revenue and systematically conservative on cost, which is precisely what incentive structures would predict. The interesting finding is that the adjustment was rarely documented as an adjustment. The model output was an input to a discussion, and the discussion produced a number, and nobody recorded the delta.

Making the delta visible is the cheap intervention. Where organisations publish the model forecast alongside the final figure and track both against actuals, override rates fall and override accuracy improves, because the adjustment now has an owner and a scoreboard. This requires no new technology and is resisted, in our experience, for entirely political reasons.

The caveat worth stating: better forecasts are only valuable if a decision changes. Several of the functions studied improved forecast accuracy materially and could identify no decision that was made differently as a result. That is not a technology failure. It is a reminder that accuracy is an input to value, not the value itself.