Marginal Value of Tokens in Business LLM Applications: Measuring Cost, Latency, and Utility of Prompt Components

Large language model (LLM) applications used in business settings are often optimized by informal prompt editing: shortening instructions, adding examples, increasing retrieved context, or imposing output constraints. Such edits are usually evaluated anecdotally, even though each prompt component affects quality, cost, latency, and operational risk. This paper introduces Marginal Value of Tokens (MVT), a component-level framework for measuring the incremental business utility of prompt segments relative to their token and cost footprint. The framework treats a prompt as a structured composition of functional components, including system instructions, task rules, business policy, retrieved context, chat history, few-shot examples, tool definitions, and output schemas. We define cost- and latency-adjusted utility, propose paired ablation and coalition-based estimators for component attribution, and give operational rules for classifying prompt components as high-value, low-value, negative-value, reusable, model-dependent, or workflow-dependent. The methodology is designed for common business workloads, including customer support and policy question answering, document summarization, and structured information extraction. The central argument is that business LLM systems should not minimize tokens blindly. They should maximize useful business output per token by preserving necessary context, pruning harmful context, compressing redundant history, caching reusable prefixes, and measuring prompt changes under non-inferiority constraints. The contribution is a practical measurement framework and experimental protocol for cost-efficient LLM adoption in business environments.