--- title: "MCP: A Standard for Action, Not Knowledge" author: "Daniel Gorld" author_role: "Consulting Director, cbs CX — The cbs Group Salesforce Consultancy" author_url: "https://cx-waves.com/about" publisher: "CX-Waves" canonical_url: "https://cx-waves.com/nodes/mcp-knowledge-vs-action" date_published: 2026-08-12 date_modified: 2026-09-20 language: en --- # MCP: A Standard for Action, Not Knowledge Source: Daniel Gorld, CX-Waves — https://cx-waves.com/nodes/mcp-knowledge-vs-action (published 2026-08-12, updated 2026-09-20) ## What is the proper use of the Model Context Protocol (MCP) in AI agent architecture? - **Targeted Actions**: MCP is optimally designed for connecting AI agents to systems to perform specific, targeted actions, such as changing a status or initiating an operation within a system. - **Standardized Integration**: It serves as an excellent integration standard, streamlining connections between language models and diverse systems by providing a uniform protocol, eliminating the need for custom interfaces. - **Encapsulated Write Access**: MCP is best utilized for write operations where a user's permissions can be passed along, ensuring that actions are executed in the user's name and respecting existing security protocols. - **Governed Operations**: When used for actions, MCP should be paired with a robust permissions layer. This ensures that the system validates user authorization before any model-initiated action proceeds, preventing security gaps. ## MCP in Perspective: What the Celebrated Standard Delivers, and What It Doesn’t The Model Context Protocol (MCP) is currently lauded as a universal connector for AI, often described as the "USB-C port for LLMs," promising to standardize AI interfaces and resolve the existing chaos of proprietary connections. While this aspect of MCP is true and represents significant progress in integration, the application of this standard is frequently misunderstood in architectural discussions. The issue lies not with the protocol itself, but with the misapplication of its capabilities. MCP brings substantial benefits by standardizing the connection between a language model and various source systems. Historically, each such connection required manual effort, including custom interfaces, unique logic, and ongoing maintenance. MCP eliminates this overhead by allowing models to interact with tools and data sources via a single, uniform protocol, rather than requiring a custom workaround for every system. This standardization marks real progress for integration work in AI architectures. ## The Critical Distinction: Knowledge vs. Action A crucial distinction often overlooked in the hype surrounding MCP is the difference between "knowledge" and "action," which profoundly impacts cost, performance, and operability. Knowledge refers to the process of reading data, such as querying "Which open quotes do we have for this customer?" This typically involves a single query yielding a specific answer. In contrast, action involves modifying something within the target system, exemplified by an operation like "Set the status of this quote to ‘presented’," which constitutes a targeted write operation. MCP is specifically engineered for actions. It excels in facilitating targeted, clearly defined operations within a system, leveraging its strengths in standardization, traceability, and low overhead for such tasks. However, when applied to pure mass data retrieval, MCP quickly becomes the most expensive and inefficient option available. Understanding this fundamental difference is key to leveraging MCP effectively and avoiding common pitfalls in AI system design. ## Where Misapplication of MCP Leads to High Costs and Latency Using MCP for mass data retrieval, such as answering "Which open quotes do we have for customer X?", introduces significant inefficiencies and costs. In a pure MCP approach, the language model is tasked with searching, retrieving raw data from a source, filtering it, reasoning about it, and potentially making further requests. This process leads to three reliably underestimated issues that impact project viability. First, a "context-window explosion" occurs. All data retrieved from the source system is loaded into the model’s working memory. When multiple systems are connected, the volume of data processed per interaction escalates rapidly, causing the model to become noticeably slower. This increased data handling directly impacts performance and resource utilization. Second, "latency from multiple steps" becomes a problem. If an AI agent requires several stages—like searching for data, requesting details, and then acting—this translates into multiple full processing cycles executed in sequence. For applications requiring real-time responses at the workplace, such multi-stage processing often results in unacceptably high latency. Third, and most economically perilous, "cost scales with the number of users." What a simple, deterministic filter can accomplish for a fraction of a cent, the model is tasked with re-evaluating at every reasoning step, and charges are incurred for every word processed. While this might go unnoticed with a small number of pilot users, scaling to hundreds or thousands of users during rollout will reveal substantial, unsustainable operational costs. ## The Underestimated Gap: Governance and Permissions A critical issue often ignored in the widespread enthusiasm for MCP, and one that proves to be a knockout criterion in industrial applications, is permissions management. MCP itself does not provide any integrated rights management system. This poses a fundamental question: if the model has access to a system via an MCP connection, how does the system verify whether the human user behind the screen is authorized to view or manipulate the data in question? In environments characterized by finely-grained roles, sales territories, and diverse confidentiality levels, this is not a trivial academic concern. Organizations that fail to adequately address this permissioning challenge face a dilemma: either they create significant security vulnerabilities or they incur enormous additional effort to implement external controls, typically leading to both. The lack of inherent governance within MCP necessitates careful architectural consideration to avoid critical security and compliance gaps. ## The Clean Cut in Practice: RAG for Knowledge, MCP for Action The most effective solution for integrating AI agents involves an unspectacular yet robust division of labor: using MCP for its intended purpose and employing established patterns for reading data. For all read access, termed "knowledge," the standard approach is a Retrieval-Augmented Generation (RAG) system. RAG operates by first semantically searching for and filtering relevant data, then presenting this compact, pre-processed context to the language model. An upstream cache can instantly answer identical questions, bypassing the need for any model call. This ensures the model receives precisely what it needs to answer, making the process cheap, fast, and traceable. Using MCP for this task would fundamentally misuse its strengths. For "action," or write access, MCP is correctly employed, but within an encapsulated framework and preceded by a permissions layer. When the model initiates an action, the user's token is passed along. The action then executes in the user's name, and if that user lacks write privileges in the target system, the backend system blocks the operation before the model can proceed. This strategically positions the governance solution where it belongs. In this refined model, questions about open quotes are efficiently and cheaply answered via the read path. MCP only becomes involved when a user explicitly wants to modify something, ensuring controlled, targeted, and singular operations. ## Conclusion: MCP as a Standard for Action, Not an Architectural Panacea MCP is undeniably an excellent integration standard, effectively ending interface chaos and serving as the correct pathway for action-based AI agent scenarios. However, it is crucial to recognize that MCP is not designed to be the operating system for a company's entire data landscape, and it should not be treated as such. The pervasive hype often suggests channeling all data interactions through this single, seemingly convenient universal connector. Organizations that blindly follow this advice risk encountering significant issues, including token explosions, unacceptable latency, and critical governance gaps. Conversely, those who correctly understand MCP as what it truly is—a powerful standard for targeted actions, rather than a solution for every data query—will unlock the genuine value the protocol offers. A standard defines connectivity, but it is not an architectural concept; conflating the two results in substantial costs, either in operational expense or through system failure.