AI spending is shifting from experimental pilots to large recurring operating expenses, reaching $1.5 trillion worldwide in 2025.
The potential returns are greater than ever, but the gap between investment and impact is also greater. According to a recent McKinsey study, 88% of companies have deployed AI in at least one business function, but only 39% can attribute this investment to its impact on EBIT at the enterprise level.
There is no established discipline to make these investments measurable, predictable and efficient. That’s why we’re launching the Cursor CFO Council, a working group of finance leaders focused on answering a single question: How do you tie AI spending to value?
The Council will meet quarterly in rotating cities around the world, providing members with an ongoing forum to compare their observations and develop a common framework for the AI economics.
Table of Contents
Intelligence is reflected in sales
A recent BCG analysis using Cursor data found that companies in the highest quintile of token usage experienced average year-over-year revenue growth of 16.5%, compared to 5.1% for companies in the lowest quintile.
A separate study of cursor usage found that after major model improvements, employees sent 44% more agent messages per week in late 2025. The biggest increase came in highly complex tasks, where messages increased by 68%.
Better models expand the range of work teams willing to engage in them, suggesting a Jevons-like dynamic where usage increases rather than decreases with performance. But it is also clear that the benefits of AI introduction are not felt to the same extent everywhere.
The returns from intelligence are unevenly distributed
Our recently published Developer Habits Report found that p99 developers produced 46x more AI-powered rows per day than the average active user and merged 15x more pull requests per week than the average active pull request author.
In other words, a small number of people get huge leverage while most others don’t.
We observed a similar concentration in spending, token consumption, and AI-generated code. Measured by the Gini coefficient, these distributions are more unequal than the income distribution in any other country in the world.
The cost per unit of work varies greatly
Even where AI clearly works, costs vary widely. In the Developer Habits Report, the cost per agent request varied by nearly nine times between model families, while the cost per line accepted varied by about seven times.
This cost gap shows why it’s helpful to have access to multiple models and providers. Different models are better suited for different types of work – planning, front-end development, debugging, lower-cost execution – and in Cursor, 84% of power users already use multiple models each week.
This option is becoming increasingly important as AI vendors move to usage-based pricing, making intelligence a variable cost that is harder to predict.
Matching the right work with the right level of intelligence offers significant cost-saving benefits that only grow over time.
A forum for thinking about AI economics
The productivity value of AI increases with each major model release, but adoption is uneven, usage is concentrated, and costs vary widely depending on how the work is routed.
The CFO Council provides a place for finance leaders to address these questions together. It will work to develop common benchmarks for AI productivity, frameworks for measuring returns to intelligence, and practical approaches for 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 move into this meeting and beyond. We also plan to post updates on the group’s work so that the broader community can benefit from what we learn.
https://cursor.com/blog/cfo-council
