Corporate AI safety rules cost 25–35% extra, study finds

Enterprise AI spending is typically tracked at the token level, but a hidden tax is embedded in every commercial API call: the computational cost of enforcing safety guardrails. The researcher's measurement reveals that system-level safety mechanisms—including refusal behavior instructions, safety classifier injections, and mandatory output hedging—are not merely passive filters but active consumers of tokens, and therefore dollars. Organizations budgeting for large-scale AI workloads generally do not segregate this cost, and vendors do not break it out in their pricing documentation, making it difficult to audit or challenge. The scale becomes material: an organization running a million queries annually, each burdened with roughly 1,500 tokens of safety overhead, is paying for millions of tokens of processing that produces no direct business output. This opacity is the core of the story—the overhead exists and is measurable, but it remains invisible to finance and procurement teams.