The cost of an AI investment arrives as an invoice. Its return arrives, if it arrives at all, as a change in how work is done, and that change is rarely measured with the same care.
Spending is easy to count and return is not.
A licence line and a consumption bill are unambiguous. What the spend produced is a question about workflows, and many organizations do not measure their processes closely enough to answer it. So the cost side of the case gets built precisely and the benefit side gets estimated, and the estimate is often made by the team that proposed the investment.
Where the missing return may sit
In many cases the benefit exists but has not been attributed. A tool saves 40 minutes at a step nobody tracks, and the 40 minutes never appear in a report because the report measures the end of the process rather than the middle of it.
The reverse is also common: a saving that is real at the task level and disappears at the organizational level, because the recovered hour was absorbed and never reallocated. A recovered hour only becomes recovered cost if something changes about how the team is staffed or what it takes on.
What published survey figures cannot tell you
Survey figures on AI spending and return describe a population. They do not tell you which side of it you are on, and the answer depends less on how much you spent than on what you agreed to measure. Without that agreement there is nothing against which a return can be measured.
The practical lesson is to decide how success will be measured before the money is committed, and to write that decision down where the people approving the budget can see it.