The new economics of AI: 88% deployed it, 39% can prove it paid off.

The gap finance actually has.

88% of organizations have deployed AI in at least one function. Only 39% can trace that investment to enterprise-level EBIT impact.That’s not a technology gap — it’s a measurement gap, and it’s squarely finance’s to close. Usage-based pricing turned intelligence into a variablecost that’s hard to predict, and you can’t govern what you can’t see.

Why cutting cost won’t flatten your bill (Jevons).

Here’s the part most cost pitches won’t tell a CFO. When AI gets cheaper or better, usage rises with capability rather than falling — the Jevons paradox. Cursor saw workers send 44% more agent messages per week after a model upgrade, with high-complexity work up 68%. Cheaper tokens don’t shrink the invoice; they get spent on more.

So the honest truth from a cost advisor: you will not optimize your way to a flat AI bill. That’s not a reason to skip optimization — a 7× cheaper tokenis still 7× cheaper. It’s the reason optimization alone isn’t a strategy. The bill is going up; the only question is whether every dollar of the increase is buying something.

The variance is where the money hides.

The same task, run different ways, costs wildly different amounts. Cursor measured cost per agent request varying nearly 9× across model families, and cost per accepted line of code varying roughly 7× — and 84% of power users already run multiple models each week. Different models are better for different work: planning, frontend, debugging, low-cost execution.

That spread is not noise — it’s the single biggest controllable line item you have. Matching the right work to the right level of intelligence is the finance lever, and it’s exactly what our Cost Engineering Playbook and model guide are for.

Spend isn’t the enemy — unmeasured spend is.

The counterintuitive part for a cost hawk: heavier AI usage correlates with growth.Companies in the highest token-usage quintile saw 16.5% median year-over-year revenue growth versus 5.1% for the lowest. And the benefit is wildly concentrated — top-percentile developers produced 46× more AI-assisted output per day than the median, a spread more unequal than income distribution in any country on earth.

Read those two facts together and the strategy writes itself. The floor is cutting waste — real, and soon table stakes. The ceilingis steering spend toward the people and tasks where it compounds into revenue. A CFO’s job here isn’t to minimize the AI bill; it’s to make sure every rising dollar is a bet you can see.

What to instrument on Monday.

• Visibility first. If you can’t see cost per team / per workflow / per outcome in near-real-time, that’s the whole problem — fix it before you negotiate a single contract.
• Measure per-outcome, not per-token. Cost per accepted line, per shipped PR, per closed ticket — tie spend to a unit of value, or you’re flying blind at 9× variance.
• Route, then re-check. Match work to the cheapest model that clears the bar; re-check quarterly, because the “best” model is a moving target.
• Budget for growth, not flatness. Plan for the bill to rise with usage — and make each increment defensible.