The two AI mistakes that cost the most
Both mistakes are invisible at the level of a tool and obvious at the level of a step.
Mistake one: a model where a rule would do
If a step is genuinely deterministic — if the invoice is over a threshold, route it; copy these fields into that system — then it should be automated, but it does not need a model.
Putting one there costs more per run, behaves less predictably, and introduces a failure mode that did not previously exist. The work still gets automated; it just gets automated with the wrong instrument.
Mistake two: a machine where a human must stay accountable
The second error is the expensive one. A step carrying judgment, consequence, relationship or sign-off stays with a person — not because a machine could not produce something plausible, but because someone has to own the outcome.
The tell is that the surrounding language sounds automatable. Drafting a decision letter looks like generation right up to the moment you notice a person has to stand behind what it says.
Why decomposition catches both
Neither mistake is visible when you evaluate tools, because a tool is not the unit at which the error occurs. Both are obvious once the workflow is broken into steps and each step is sorted by who should carry it.
That is the whole reason the method decomposes before it evaluates anything. The sorting is where the money is, and steps that contain both — preparation a machine can do and a call only a person can make — are where the largest gains sit, because splitting them is what makes both halves work properly.