The CFO's AI Allocation Problem

AI spending can look deceptively simple from a finance seat. There are subscriptions, implementation budgets, vendor contracts, cloud charges, and internal experimentation costs.

But the real allocation problem is not whether AI belongs in the software budget. The question is whether AI changes the firm’s cost of producing useful work.

If a claims process, credit memo, procurement review, customer support response, or software change can be completed with less human time and acceptable quality, then the accounting category matters less than the economics of the workflow.

The useful comparison

A CFO should eventually be able to compare several production choices:

• human execution;
• human execution with AI assistance;
• AI first draft with human review;
• autonomous AI with exception handling;
• external vendor or software product;
• no action.

Each option has a different cost, quality, risk, speed, and organizational burden.

The challenge is that most AI reporting does not measure these options in comparable units. It reports licenses, adoption, tokens, productivity claims, or anecdotal time savings. Those are inputs to the decision, not the decision itself.

What should be measured

The minimum useful metric is total cost per accepted output. That requires more than API logs. It requires knowing whether the output was good enough, how much human review it required, how often it failed, and what a failure costs.

This is the beginning of a more useful AI finance discipline: not AI spend as a percentage of revenue, but AI expenditure per unit of economic output created.