The Efficiency Paradox in Agentic AI Workflows: When Optimization Increases the Electricity Bill

Problem definition: Firms deploying agentic AI workflows at scale face fundamental trade-offs between cost, speed, and quality of Large Language Model (LLM) calls. While individual queries are inexpensive, the cumulative cost of automated workflows quickly becomes substantial. Organizations can optimize LLM usage through two approaches: internal optimization (aggregating tasks and caching instructions to reduce computational costs) or external optimization (providing time flexibility to providers who offer substantial price discounts in exchange for processing delays). Methodology/results: We characterize the optimal strategy based on workload intensity and delay sensitivity, revealing clear decision rules for when each approach dominates. Including model selection uncovers an efficiency paradox: when firms can jointly optimize workflow strategies and model capability, cost and energy savings are offset-and can be more than offset-by selection of more capable, energy-intensive models, so that increased operational flexibility leads to higher energy consumption. Managerial implications: Operational optimization alone cannot deliver environmental benefits when model choice is endogenous. Efficiency improvements are neutralized by capability upgrades that consume the same total resources or more, so efficiency strategies may not reduce the electricity bill when firms can select model capability.