AI spend is shifting from experimental pilots into a major recurring operating expense that reached $1.5 trillion globally in 2025.
The potential returns are greater than ever, but so is the gap between investment and impact. According to a recent McKinsey study, 88% of organizations have deployed AI in at least one business function, but only 39% can trace that investment to enterprise-level EBIT impact.
There is no established discipline for making these investments measurable, predictable, and efficient. That is why we are launching the Cursor CFO Council, a working group of finance leaders focused on answering a single question: How do you keep AI spend tied to value?
The council will meet quarterly in rotating cities around the world, giving members a standing forum to compare what they are seeing and develop a shared framework for AI economics.
Intelligence is showing up in revenue
A recent BCG analysis using Cursor data found that companies in the highest quintile of token usage saw 16.5% median year-over-year revenue growth compared to 5.1% for companies in the lowest quintile.
A separate study of Cursor usage found that following major model improvements in late 2025, workers sent 44% more agent messages per week. The largest increase came from high-complexity work, where messages rose 68%.
Better models are expanding the set of work teams are willing to attempt, pointing to a Jevons-style dynamic where usage tends to rise with capability rather than fall. But it's also clear that the benefits of AI adoption are not showing up everywhere in equal measure.
The returns on intelligence are unevenly distributed
Our recently released Developer Habits Report found that p99 developers produced 46x more AI-assisted lines per day than the median active user and merged 15x more pull requests per week than the median active pull request author.
In other words, a small number of people are getting an enormous amount of leverage, while most others are not.
We observed similar concentration around spend, token consumption, and AI-generated code. Measured by Gini coefficient, these distributions are more unequal than income distribution in any country in the world.
Cost per unit of work varies widely
Even where AI is clearly working, cost varies a lot. In the Developer Habits Report, cost per agent request varied by nearly 9 times across model families, while cost per accepted line varied by roughly 7 times.
This cost gap shows why it helps to have access to multiple models and providers. Different models are better for different kinds of work — planning, frontend development, debugging, lower-cost execution — and in Cursor, 84% of power users already use multiple models each week.
That optionality is becoming more important as AI providers move toward usage-based pricing, which turns intelligence into a variable cost that is harder to predict.
Matching the right work to the right level of intelligence has significant cost-saving advantages which are only growing over time.
A forum for thinking about AI economics
The productivity value of AI is growing with each major model release, but adoption is uneven, usage is concentrated, and costs vary widely depending on how work is routed.
The CFO Council will give finance leaders a place to work through these questions together. It will work on developing shared benchmarks for AI productivity, frameworks for measuring returns on intelligence, and practical approaches to model allocation and cost management.
The first meeting of the CFO Council will take place in August. We look forward to announcing participants as we track toward that meeting, and beyond. We also plan to publish updates on the group's work, so the broader community can benefit from what we learn.