The number that is usually missing
Enterprise AI business cases tend to compare a fully loaded human process against a partially loaded automated one. Licence and compute are counted; review, exception handling, and rework are not.
The correction is arithmetic, not philosophy. Compute cost per unit, licence amortised per unit, review minutes at the reviewer's rate, exception handling at the rate of whoever handles exceptions, and the rework rate observed rather than assumed.
Why the reviewer's rate matters more than the model's
Reviewers are usually more expensive than the people who did the original task, because review is a seniority function. A workflow that moves work from an analyst to a manager can raise cost per unit while appearing to save time.
This is also why quality bars have to be set by the reviewer who will defend them. A bar set elsewhere produces review volume nobody budgeted.
The crossover point
Variable cost grows with volume. Human cost is usually stepwise. There is a volume at which the two lines cross, and after it the workflow you automated stops being cheaper than the one you replaced.
Calculate it before you scale, monitor it after. [Expand](/native-framework/expand) is where most organisations discover this by surprise, at renewal.
A defensible comparison
Publish the comparison in this shape and it will survive challenge.
- Cost per unit of work before, fully loaded, measured over a stated window.
- Cost per unit of work after, including review and exception handling.
- The volume assumption behind both.
- The crossover point and the monitoring that will catch it.
Next step
Read the Validate stage, which specifies the pilot charter and the fully loaded cost comparison in detail.