Six in ten big companies spent well over their AI budgets in the past year, by 24% on average
Most large companies are spending more on AI than they planned, and many cannot say exactly where the money goes. That is the core finding of a survey of 500 senior executives that Capgemini, the French IT services group, published on November 12. Trade outlets picked it up through the weekend.
Nearly six in ten executives (59%) said their AI spending over the past 12 months was significantly above what they expected or budgeted. Among those that overshot, spending ran 24% over on average. About one in five organizations overshot by more than 30%.
The overruns were not just accounting noise. Two-thirds of the companies that went over budget said it had negative financial effects, such as lower free cash flow or cuts to other discretionary spending. All respondents work at companies with more than $1 billion in annual revenue, across 10 countries and 13 industries, and were surveyed in September.
Why the bills are surprising buyers
Executives and investors have mostly heard that AI is getting cheaper, as the price per unit of AI output keeps falling. Capgemini's numbers show why total bills are rising anyway: usage is growing far faster than prices fall.
The report cites data from OpenRouter, a service that sends requests to many AI models, first reported by the tech news site The Decoder in August. It showed that usage by AI agents grew 14-fold in six months. Usage by people typing requests grew 2.8 times over the same period. Agents are AI programs that carry out multi-step tasks on their own, and agent coding tasks can use roughly 1,000 times more AI processing than chat-style coding help, according to research from Stanford University and others that the report cites.
Buyers are feeling it. The report quotes Uber's chief technology officer, Praveen Neppalli Naga, saying the budget he expected to need is "blown away already," according to diginomica's account of the study. Concern is rising: 53% of executives now call AI costs a worry, up from 45% a year earlier.
The bigger problem may be visibility. Six in ten organizations lack a company-wide view of what AI costs them. Half of executives said AI spending gets approved without a clear picture of its full cost. Only about a third can confidently estimate indirect costs, such as staff time spent checking AI output, fixing mistakes and running oversight.
Even where spending is tracked, the link to results is thin. Capgemini says only 61% of AI spending can be systematically tracked, and only 45% of that tracked spending can be tied to verified business value. Just one in five companies measures what Capgemini calls "AI Yield," its term for verified value after counting the full cost of delivering it.
Finance chiefs step in, and cheaper models gain
Companies are reacting mainly by tightening controls rather than rethinking how they use AI. Of those that overshot, 53% added spending controls. Only 20% reconsidered the tools or processes involved, and 5% paused or stopped projects. Chief financial officers now help approve major AI investments at 52% of companies, up from 42% a year earlier.
The other response is shopping around. More than three in five executives expect to use several AI models from different providers within a year, and nearly three in five plan to expand their use of open-weight models. These are models whose underlying files are published, so companies can run them on their own hardware or through cheaper providers. Capgemini estimates their output costs about 87% less on average than that of closed models from the leading labs.
AT&T shows what this looks like in practice. The telecom carrier now runs nearly 500 AI agent workflows in production and processes about 45 billion units of AI text a day, an executive told an Amazon Web Services event in Boston, SiliconANGLE reported on November 14. AT&T keeps humans reviewing each workflow until accuracy reaches the mid-to-high 90% range, according to the same report. Separately, an AT&T executive told The New York Times in September that about 40% of employee AI requests now go to cheaper open models, according to AI Weekly's summary of that article.
What the survey does not prove
The findings deserve some caution. Capgemini sells AI strategy, implementation and cost-management services, so a report concluding that companies need better AI cost governance also describes demand for its own business. "AI Yield" is Capgemini's own measure, not an industry standard.
The data are also self-reported. Respondents came mainly from roles tied to AI spending and strategy, such as AI leadership, engineering, technology finance and AI cost management. The report's example of an unnamed company running up a $500 million AI bill in one month comes with no further detail and cannot be checked.
Still, the direction matches other recent surveys. Deloitte reported in late October that only 44% of companies track the cost of each unit of AI work, and only 28% report revenue gains from AI. Neither survey suggests companies plan to spend less. They suggest the next year of AI buying will be shaped more by finance teams than by technology teams.
For AI vendors, that is a mixed signal. Demand is clearly strong. But buyers with tighter controls and more pressure to prove returns are more likely to move routine work to cheaper models, which puts pressure on the premium prices charged by the leading labs.