A survey of 400 companies conducted through May 2026 reveals that the vast majority of businesses have exceeded their initial AI implementation and operational budgets. Only a small minority managed to stay within planned spending limits, while 2% of respondents admitted they had made no financial forecasts at all.
The scale of budget overruns varies significantly among companies. 11% of participants reported exceeding their budget by up to 11%, while 39% saw costs rise by 11–25%. Another 35% of firms faced overruns of 26–50%, and 13% of companies spent more than 50% above their original estimates.
Research involving universities such as Stanford and Berkeley suggests that opting for cheaper AI models does not always reduce expenses. In a test comparing Google's Gemini 3.1 Pro and Gemini 3 Flash, the more expensive model completed a task for $1, while the cheaper version failed after costing $14 in token usage.
The unpredictable nature of AI costs is further complicated by the fact that the same model can charge different amounts for identical tasks over time. Experts note that AI services function differently than traditional software, often requiring more steps to process requests, which leads to inconsistent and often higher operational costs.
