AI Budgets 2027: Extensions Denied Without Proof of Value

Financial departments are tightening the screws on AI spending. Without usage metrics linked to tangible economic gains, license extensions are being denied. Low productivity effects are being observed at this stage, while expectations for the next three years are higher.

Without proof of value, budget extensions are denied

In a major French real estate group, an extension to expand AI licenses was not granted due to the inability to demonstrate the returns from previous purchases. Management had explicitly requested these elements before making a decision. This case illustrates the question now posed by financial departments: why double a batch of 50 licenses without quantified evidence of returns from the first batch? The initial enthusiasm has waned, and the size of budgets is evolving. Scaling up means having agents in production and tools accessible to the entire company, but this fall, many CIOs will need to defend these ambitions with more solid elements than mere enthusiasm.

What surveyed executives report on productivity

Nearly 6,000 executives, primarily CEOs, CFOs, and senior executives from the United States, the United Kingdom, Germany, and Australia, were surveyed between November 2025 and January 2026 as part of a study led by the National Bureau of Economic Research in collaboration with the Federal Reserve Bank of Atlanta, the Bank of England, and the Bundesbank. According to their responses, 69% of companies are using AI. The effect on productivity is estimated at +0.29% over three years, while 89% of executives report seeing no change. However, for the next three years, these executives anticipate +1.4%.

Measuring usage and converting time into economic value

A connection rate or a few favorable testimonials are not enough to defend the ROI of an agent. The challenge is to establish what the initial euros have actually changed in the work. This requires tracking, for each license and each agent, who is using the tool, for what tasks, how often, and what volume of time is actually saved. This time only counts in a budget if it allows for absorbing more workload without hiring, avoiding a cost, or redeploying a team to another activity. If time savings are dispersed without visible effect, they are worthless in a trade-off, and coming with just a connection rate amounts to defending a promise.

Where the frictions come from: an easy initial budget, uneven usage

Most companies introduced AI one to two years ago using easily accessible innovation budgets: 50 Copilot licenses, an internal ChatGPT, a few Dust licenses, and about ten Claude licenses intended for developers. By 2023, it had become common to fund a test, and executive committees preferred to avoid having to explain the lack of experimentation. The second budget is of a different nature: usage is heterogeneous, with some saving a lot of time while others rarely open the tool; some employees request licenses while the cap remains at 50; advanced teams ask for an agent for repetitive tasks and time to properly configure Claude on the development side.