Fraunhofer Institute for Software and Systems Engineering
Content
Many companies measure their AI success by the number of tokens consumed. This paper shows why that metric is misleading: high usage says little about actual value, while AI agents are driving costs up sharply. Blanket cost-cutting, however, only shifts the expense elsewhere. Targeted levers such as caching, model routing and deliberately avoiding AI where it isn't needed are far more effective. What matters is measuring outcomes rather than consumption and actively managing the system built around the model.