Agent workflows waste costs by overusing expensive AI models for simple tasks, study finds.

A developer observed an agent workflow costing hundreds of dollars because it used a top-tier, expensive AI model for every step, including simple formatting tasks. Research indicates enterprise AI systems waste 60-80% of their budget by routing most steps to such frontier models when smaller models suffice. A 2026 study found only about 14% of steps genuinely require the advanced models, yet those consume nearly half the total cost. The solution involves a tiered routing system, or model cascade, where steps first attempt a cheaper model and only escalate if a rigorous quality gate fails, dramatically cutting expenses.