As the CEO of Arango, Shekhar Iyer leads the company’s mission to make enterprise AI contextual, scalable and trusted.gettySome AI agents are designed to retrieve and interpret customer information from structured and unstructured data to support teams such as sales, marketing, customer service and finance. They use that information to recommend and explain the best next actions based on a 360-degree view of the customer. That view doesn’t merely draw from a collection of customer records, but also from an understanding of customers’ histories, relationships, needs and current situations.That work often begins with an attempt to understand the business. Who is the customer? Which products are involved? What policies apply? Can the information be trusted? Has anything changed since the last interaction?Many companies already have these answers, but the answers are scattered across CRM and ERP systems, product catalogs, knowledge bases and operational applications. That disorganization can lead to redundant work for agents who need to reason, decide and act reliably. Redundant agent work drives up token consumption and costs over time. The Problem Before The PromptIn my conversations with CIOs and other business leaders about enterprise AI economics, one question comes up frequently: “How can we bring down our rapidly growing token costs?”It's a reasonable question. However, high token consumption is often a symptom of asking agents to repeatedly reconstruct from scratch the business understanding the enterprise already has. Specifically, leaders ask an agent that’s beginning a task to reconstruct that understanding from scratch. To me, that's like blaming your electricity bill when your house has no insulation. You can negotiate a better utility rate or replace every lightbulb with LEDs, but neither stops energy from escaping through the walls.From my observations, token costs have become one of the most visible measures of AI economics. But AI costs often reflect architectural decisions made long before a model receives a prompt.With agents, optimizing prompts, routing requests to less expensive models and reducing unnecessary output can help, but those steps come after the architecture has already determined how much work agents need to do.Increasingly, I’ve found that enterprise leaders are asking not just how to reduce token costs, but why agents consume so many tokens in the first place. Often, the answer is the business context agents must reconstruct. That reconstruction also has a human cost: Teams repeatedly spend engineering time integrating sources, resolving entities, encoding business rules and maintaining similar context pipelines for different applications and agents. That raises a better question: “How can we make that business context a reusable asset instead of rebuilding it every time?” Why Context Efficiency MattersContext efficiency is the ability to give an AI system the minimum trusted business context required to complete a task, without forcing agents to repeatedly discover, reconcile and translate that context at runtime. Agents shouldn't have to relearn the business every time they begin a task. When they start with the business context they need, agents can spend less effort reasoning, deciding and acting.Consider AI agents working on a customer account at a technology company. The context the agents need, and the best way to retrieve it, depend on the task. Resolving a billing dispute may require retrieving and relating the customer’s product configuration, contract terms, case history and recent interactions. Assessing renewal risk may instead draw on usage, support and engagement signals. The retrieval layer should select the right approach and supply the task-specific context. Rethinking Enterprise AI EconomicsI’ve observed that organizations progressing faster with AI are treating trusted, governed business context as a shared enterprise capability rather than recreating it for every request.But making business context reusable is only part of the equation. How it’s organized matters, too. Connecting records isn't enough. Business context should reflect the concepts and relationships people actually use to understand the customer, namely their history, needs, interactions, products, health and current situation.I’ve found that the closer the business context is to the language, concepts and relationships people actually use, the less reconstruction and translation the overall system may need to perform. When implemented well, that can mean fewer tokens, lower latency and less computational work to reach the same business outcome.That points to a broader principle for AI economics: Use agentic reasoning where it adds value, and use deterministic software where it can do the work more efficiently and predictably. Context efficiency is about giving agents the right business context, in a form they can readily use, while using standardized access patterns, appropriate permissions, persistent context and deterministic or hybrid retrieval to minimize unnecessary reconstruction, translation and reasoning.The benefits can extend beyond lower token consumption. Agents can respond faster and more consistently, decisions can become easier to explain and govern and more compute can be spent applying business understanding rather than recreating it.Context-efficient architectures don't just make business context reusable. They reduce the computational distance between a business question and a business outcome.Three Recommendations For LeadersLeaders preparing to build agents at scale should ask three questions. 1. “Are we measuring cost per inference or cost per business outcome?”Token consumption is an important operational metric, but it doesn't tell leaders whether agents are becoming more economically productive. Leaders should measure tokens, latency and compute against the business outcome being produced, not simply the cost of an individual inference.2. “Is our business context governed and reusable, or is every agent rebuilding its own version?”As organizations deploy more agents, allowing each to reconstruct its own understanding creates unnecessary work. Leaders should treat business context as shared infrastructure, with common definitions, access controls and provenance that applications and agents can build upon.3. “How much work are we asking AI to do unnecessarily?”Not every step requires reasoning. Leaders should organize context around the language, concepts and relationships people actually use, and leverage deterministic software where it can perform the work more efficiently and predictably. The goal is to reduce the computational distance between a business question and a trusted business outcome.The Crucial Role Of Context Efficiency In The Future Of Enterprise AI Architectures As organizations expand their use of agents, reducing token costs will remain important. But the larger opportunity is reducing the work they must perform before they can create business value.Organizations have spent decades building business knowledge. Agents shouldn't have to rediscover it for every new task. With trusted, relevant context, they can focus their reasoning on the business problem at hand, producing faster, more reliable and more economical outcomes.Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?