AI: Spending Exceeds Forecasts by 24% on Average, According to a Capgemini Study
According to a study by the Capgemini Research Institute (Capgemini's internal think tank), published on November 12, 2026, AI spending exceeded forecasts by 24% on average in the companies concerned. The survey shows that companies respond first by tightening controls, and less often by redesigning their tools.
AI Spending Beyond Budget for 59% of Surveyed Executives
The study, described as a “flash” by its authors, is based on a survey conducted in September 2026 among 500 senior executives of companies generating more than one billion dollars in annual revenue, across 10 countries and 13 sectors. It addresses the economics of AI in large companies in general, and does not specifically focus on Capgemini's own activity.
AI costs are a concern for 53% of the executives surveyed. Over the past year, 59% of them report that their company's spending in this area was significantly higher than planned or budgeted, and nearly one company in five (21%) exceeded its AI budget by more than 30%.
According to the institute, only 56% of executives are confident in their ability to fully anticipate the evolution of their costs, and half say AI spending is approved without sufficient visibility into the total cost. They express particular doubt about estimating indirect costs: human supervision, rework of produced content, and governance.
Two-thirds of companies that exceeded their AI budget suffered negative financial repercussions, such as a decline in free cash flow or a reduction in discretionary spending. The study notes that chief financial officers now play a greater role in validating major AI-related investments.
Tighter Controls, Multi-Model Strategies and New Indicators
Faced with these overruns, 53% of companies have strengthened control over their spending, by tightening budgets, introducing company-wide controls, or reducing the scope of projects. Only one company in five has redesigned its tools or processes, for example by changing models or configuration.
More than three executives in five plan to adopt a multi-model strategy within a year, in order to select models according to cost, capabilities, latency, or level of risk. Nearly three companies in five are also considering expanding their use of open-weight models, whose parameters are made publicly available.
On measuring value, 57% of executives find it difficult to demonstrate AI's contribution to the company, and barely one company in five calculates the “AI Yield,” a return indicator that incorporates all the costs of a secure deployment. Only 32% of respondents cite revenue growth among the benefits obtained.
Today, 60% of companies mainly compare their AI spending against the budget, but only 44% intend to keep this as a priority over the next 6 to 12 months. According to the study, the first-run success rate should become the priority indicator, ahead of the cost per completed task.
Volker Darius, Global Head of Capgemini Invent, stated: “Cost control alone will not make AI profitable, and value creation without cost control will not be sustainable either.” The study follows the first edition of The AI Compass, a Capgemini publication dedicated to the economics of AI.
