Post 2 in the Series: Introducing MCP to Marketing Tech Leaders

By Dan Jaye, CEO

In my previous post about MCP as the Open API standard for AI, I introduced how Model Context Protocol is fundamentally reshaping AI architecture. Now let’s dive deeper into the specific concepts that make this transformation possible.

Two years ago, if you wanted to demonstrate AI sophistication at industry events, you’d casually drop terms like “prompt engineering” and “hallucination.” Today, I’m predicting the vocabulary that will distinguish the leaders from the followers: sampling, elicitation, and completions.

But here’s what’s curious – none of these terms means what you might initially think they mean.

Sampling: When AI Collaborates Mid-Stream

Now, let’s examine sampling, because this concept fundamentally warped my understanding of AI architecture when I first grasped it. In traditional AI workflows, the pattern is straightforward: client assembles information, constructs a prompt, calls the LLM, receives a completion. Think of it like ordering at a restaurant – you review the menu, place your order, and receive your meal.

Sampling completely inverts this dynamic.

In MCP, an AI server can actively call back to the client mid-inference and request: “Before I complete this task, could you run some analysis for me?” It’s like the kitchen calling you mid-prep to say, “We’re working on your dish, but could you come taste the sauce and guide our approach?”

This isn’t incremental improvement – it’s revolutionary. The server becomes an active collaborator in the inference process, not a passive responder. It transforms how we conceive collaboration between AI systems.

Elicitation: Building Intelligent Checkpoints

Elicitation takes this collaborative approach further. We already have “prompting” in AI terminology, so elicitation represents something more sophisticated. Elicitation occurs when an MCP server reaches back to the client – or directly to the user – for additional input or approval.

Imagine it as the AI equivalent of “Are you certain you want to delete this file?” but with nuanced intelligence. A server might communicate, “I’m preparing to execute this customer database query – does this approach align with your intentions? I’ve identified four different methodologies. Which would you prefer?”

This enables genuinely interactive AI workflows, where systems can pause, reflect, and proceed with enhanced information and human oversight.

Stop Thinking “Request-Response” – Start Thinking “Completions”

Let me start with a fundamental shift that reveals how differently we need to think about AI interactions. In traditional API architecture, we discuss “requests” and “responses.” But in MCP, we use the term completion – and this isn’t semantic posturing.

A completion represents the accomplished outcome of an MCP server interaction. Instead of merely returning data like a conventional API response, a completion emphasizes that something was achieved. It’s the distinction between asking “What’s the weather forecast?” (receiving data) and requesting “Prepare me for today’s weather conditions” (receiving actionable preparation).

This terminology shift reflects the agentic nature of MCP interactions – where AI systems don’t just retrieve information, they accomplish objectives. Start incorporating “completion” into your professional vocabulary. It signals you understand the paradigm shift happening beneath the surface.

Why This Can Transform Marketing Technology

These concepts aren’t abstract improvements – they’re practical capabilities that unlock entirely new approaches to marketing automation. Consider customer service systems that sample from recommendation engines, elicit preferences from users, and deliver adaptive interactions through completions – all orchestrated across specialized components.

This isn’t theoretical. Companies are already implementing these patterns. OpenAI officially adopted MCP in March 2025, and Google DeepMind confirmed MCP support in upcoming Gemini models, with CEO Demis Hassabis describing the protocol as “rapidly becoming an open standard for the AI agentic era.”

The organizations that master these new approaches via a new vocabulary – sampling, elicitation and completions – will architect the systems that define the next generation of AI-powered marketing technology. As we’ll explore in my next post tomorrow about chainable AI architectures, these interaction patterns become the building blocks for entirely new approaches to AI system design.

Move to Post 3 in the Series Here: The Architecture Revolution: Building Chainable AI Systems