Survey: only 11% of ~400 businesses could forecast AI spending; a study finds lower-priced models cost more than higher-priced models on 32% of 6,800+ tasks

Your AI tool's operational cost can be higher than advertised, and forecasting budgets for AI is now materially harder.

A recent survey indicates that a mere 11% of approximately 400 businesses have been able to accurately forecast their spending on artificial intelligence. Concurrently, a separate study analyzing over 6,800 tasks revealed that lower-priced AI models were more expensive to operate than higher-priced models in 32% of those instances. The core issue identified is that while AI usage is metered by tokens, the actual token consumption for specific tasks by various models remains unpredictable. This unpredictability complicates cost management and forecasting for organizations integrating AI technologies.

The dual findings highlight a significant challenge in the AI adoption lifecycle: cost predictability. Businesses struggle to forecast AI expenditure, not just due to the volume of usage, but because the underlying economics of different models are not transparent. The study's revelation that cheaper models can become more expensive in practice suggests that current pricing structures may not adequately account for diverse real-world task complexities or that optimized usage is not guaranteed. This creates a gap between expected and actual operational costs, impacting budget allocations and ROI calculations for AI initiatives across industries.

This lack of predictability affects both AI providers and consumers, potentially slowing down widespread adoption and investment. Companies may become hesitant to scale AI deployments if they cannot reliably budget for them. Consequently, AI developers and platform providers might need to rethink their pricing strategies, perhaps moving towards more outcome-based pricing or offering better tools for cost monitoring and optimization. End-users, in turn, must develop more sophisticated methods for evaluating and managing their AI toolchain, considering performance-per-dollar across a range of potential use cases rather than just upfront model costs.