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
Context & Ripple Effects
Corporate AI budgets were already hard to govern: a June KPMG survey found only 26% of companies had a comprehensive view of AI costs. The new results put a sharper operational measure behind that gap in cost visibility: model price alone does not reliably predict the cost of completing work.
Buyers had begun shifting workloads toward cheaper models as spending rose, creating pricing pressure on leading model providers. The finding that lower-priced models can cost more on a meaningful share of tasks makes workload selection, rather than list price, the central procurement decision.
First-order effects
- The surveyed businesses face a budgeting problem: only 11% could forecast AI spending, while Microsoft’s task analysis means procurement teams must evaluate total cost per completed task rather than token or model pricing alone.
- Model buyers using low-priced models for cost control may need to retest those deployments where retries, longer reasoning, or weaker task performance raise the effective bill.
Second-order effects
- OpenAI, Anthropic, and lower-cost rivals face more buyer scrutiny over measured task economics, not simply headline price, weakening a procurement pitch built solely on cheaper access.
- Finance and engineering teams gain a shared incentive to instrument usage and outcomes, because incomplete cost visibility prevents reliable AI budget forecasts.
Third-order effects
- If enterprises standardize purchasing around useful-task cost, AI procurement shifts from broad model commitments toward workload-level routing and performance measurement.
- The resulting buyer discipline would make model pricing less transparent as a competitive signal, with operational efficiency and reliability carrying more weight in vendor selection.
The trend: Enterprise AI is entering a procurement phase in which the relevant unit of economics is the cost of useful work, not the advertised price of a model.