Why Trimming AI Tokens Often Backfires on Corporate Budgets

The report, titled "Token Reduction Is Not Cost Reduction," challenges the industry standard of minimizing input volume as a primary cost-saving measure. Data from the study reveals that roughly 80% of total AI expenditure is consumed by repeated prompt-cache traffic—instructions and tool definitions—rather than the actual generation of new content. In these instances, only 1.3% of the billed data represented novel information.

PointFive CEO Alon Arvatz suggests that engineering leaders have been making financial decisions without empirical evidence. When developers strip away context to save on tokens, AI agents frequently enter additional processing steps to recover the discarded data, ultimately driving up the final bill. The research highlights that true fiscal control stems from real-time visibility and governance of how sessions unfold, rather than blind compression. With 93% of companies currently exceeding their AI budgets, the findings advocate for a shift toward tracking actual usage patterns and retries, which remain largely invisible under current cost-management practices.