CFOs are asking how to get more value from AI spend, but the old software metrics—seats purchased, active users—don’t capture what matters.

The article argues that the real measure should be work accomplished versus cost to produce it, coining the term Useful Intelligence per Dollar.

A lower-cost model may have cheap tokens but require more retries and human review, while a more capable model might finish the same task in one pass at a higher token price.

The core tension is that cost per token is a misleading proxy for the full cost of a successful outcome, and the market needs a scorecard that tracks value, not just volume.