The cost of intelligence got a scoreboard. Make sure you own the denominator.

BCG proposed return on intelligence; KPMG found cost-visible leaders win five to one. The number only moves if you control the cost.

BCG gave the field a number to compete on this week. Sylvain Duranton proposed return on intelligence, a metric that divides the value of the output by the combined cost of the labor and the tokens that produced it, and framed the whole moment as an era of token-based competition. It is a useful idea, and it quietly names the problem underneath it. If you cannot see the bottom of the ratio, you cannot manage the ratio.

KPMG put hard numbers on exactly that. In its Q2 Global AI Pulse, 49% of leaders had already scaled back agent deployments on cost, and the leaders who can actually see their AI spend were roughly five times more likely to show an established return. The firm also tied ownership to outcomes: where the chief executive is explicitly accountable, returns run close to four times higher, 14% against 4%. Cost visibility stopped being a finance chore and became the dividing line between a program that pays back and one that does not.

Here is where the metric bites. Priced by the metered token, the denominator of return on intelligence is a moving target that nobody at the table controls. The labs reprice, usage spikes, and the number a client optimized last quarter is stale this quarter. A firm that bills by the hour has no answer to a rising meter except to pass it through and hope the client absorbs it.

The answer that holds is to fix the cost side before the work runs. Price the analysis, not the tokens: a fixed, named price per task, known and auditable before a model spends anything. That turns the denominator from a guess into a constant, so the client spends its attention on the value of the output rather than chasing a meter that can reprice overnight.