Corporate budgeting for AI implementation has become increasingly difficult due to unpredictable costs. According to a survey of 400 businesses conducted between April and May 2026, only 11% of organizations managed to stay within their planned AI budgets, even when allowing for a 10% margin of error.
The survey revealed that 39% of respondents exceeded their budgets by 11–25%, while 35% saw overruns between 26% and 50%. Additionally, 13% of companies reported budget overruns exceeding 50%, and 2% admitted they did not forecast AI expenses at all.
Research by Microsoft and US scientists covering 6,800 tasks indicates that token consumption is highly variable. Using cheaper models for tasks they are not optimized for can lead to higher total costs compared to more expensive, efficient models. This inefficiency occurs in 32% of cases.
A case study involving Google's Gemini models illustrates this discrepancy. While Gemini 3.1 Pro completed a task for $1 in 85 steps, Gemini 3 Flash failed to complete the same task after 1,000 steps, resulting in costs of $14. Token usage can vary even when the same model performs identical tasks repeatedly.
Google representatives noted that cost variability is inherent to AI systems. To manage these fluctuations, the company recommends implementing usage limits, offering diverse pricing plans, and providing training for employees to improve the efficiency of their interactions with AI models.
