Survey: only 11% of 396 businesses could forecast AI spending; Microsoft finds lower-priced models cost more than higher-priced ones on 32% of 6,800+ tasks

AI task costs now vary significantly by model, retry count, and failure rate, making budget forecasts unreliable.

A recent Wall Street Journal survey revealed that only 11% of 396 businesses could accurately forecast their AI spending within a 10% margin. The article highlights that the cost of AI tasks is highly variable, with lower-priced models sometimes proving more expensive than higher-priced ones due to factors like the number of steps required, task complexity, retries, and failure rates. For instance, a "frontier-ish" model like Google's Gemini 3.1 Pro completed a task in 85 steps for approximately $1, while a cheaper model, Gemini 3 Flash, took 952 steps, failed, and cost $14 in tokens. This variability means businesses cannot reliably estimate task costs based on model pricing alone, complicating budgeting and planning for AI adoption.

The difficulty in forecasting AI spending implies that businesses may need to adopt new budgeting strategies that account for inherent variability, rather than relying on fixed pricing models common in traditional software. This situation could inadvertently benefit companies offering more advanced, potentially more predictable (though higher-priced) frontier models, as their costs might appear more stable compared to cheaper, but less reliable, alternatives for complex tasks.

The analysis suggests that the current AI market is experiencing a "carrying water" effect for frontier model providers, where the unpredictable costs associated with cheaper models might push businesses towards more established, albeit more expensive, solutions for critical operations. As AI adoption scales, the focus will likely shift from mere model pricing to the total cost of ownership, including development time, debugging, and the overhead of managing unpredictable task executions.