MCP Explained for Marketing Leaders

MCP is quietly becoming the standard plug between AI models and business systems, the way APIs became the standard plug between software systems. If your team uses Claude or a similar assistant, the questions they type all day (how did Q2 acquisition perform, which campaigns drove repeat purchases) can now be answered from your actual company data instead of from the model's general knowledge. Marketing data is among the highest-value things to connect, because it is where the most expensive daily decisions get made on the least verified numbers. Leaders do not need to implement MCP, but they do need to understand it well enough to ask the right question about any tool that uses it. This piece covers that, in plain language.

What MCP is

The Model Context Protocol (MCP) is an open standard that lets AI models use tools and query data sources in a controlled, permissioned way. Anthropic, the company behind Claude, released it as an open protocol in late 2024, and it has since been adopted across the industry by the major model providers and thousands of software vendors. Before MCP, every connection between an AI assistant and a business system was a custom integration; with MCP, any compliant AI application can talk to any compliant data source, the same way any browser can load any website.

The word "controlled" is doing real work in that definition. MCP does not give a model open access to a system. The system's owner decides exactly which actions and which data are exposed, and the model can use those and nothing else.

It also helps to be clear about what MCP is not. It moves no data on its own; your warehouse stays exactly where it is. Nor is it a product you buy, any more than "API" is a product. The protocol simply carries requests and responses between a model and a system, and the quality of what comes back depends entirely on what sits behind the server.

Clients and servers, in one minute

MCP has two sides. The client is the AI application your team already uses: Claude, a chat interface, an agent runtime. The server is a small program that sits in front of one of your systems and offers the model a defined menu of capabilities. There are MCP servers for CRMs, ticketing systems, code repositories, and, most relevant here, data warehouses.

For a warehouse like Google BigQuery, a typical MCP server exposes a handful of capabilities: list the available tables or metrics, describe what they contain, and run a query. When someone asks Claude "what was our best channel last month," the model calls the server, the server executes against the warehouse, and the answer comes back grounded in your data, with the model narrating the result.

That is the whole mechanism. The strategic weight sits in one question: what exactly does the server let the model see and do?

Why this matters for a marketing leader

Until now, an assistant answering a marketing question was guessing from training data, and a dashboard answering it required someone to have anticipated the question. MCP closes that gap: ad hoc questions, asked in chat, answered from governed company data. Your Monday morning "why did CAC jump last week" stops being a ticket for the data team and becomes a two-minute conversation.

Play that out over a normal trading week. The Monday meeting surfaces a dip in returning-customer revenue. Today, someone opens a ticket, an analyst picks it up Wednesday, and the answer arrives after the moment to act has passed. With the warehouse behind MCP, the follow-up questions happen in the meeting: which cohorts, which channels, since when. The decision gets made while it still matters.

The compounding effect is on question volume. Teams with a working setup ask ten times the questions, because the cost of asking dropped to zero. Whether those ten answers are reliable depends entirely on the next section, and this is the part where a leader's attention pays off, because the protocol will be chosen by engineers and the guarantees will be inherited by everyone.

The server matters more than the protocol

MCP standardizes the plumbing. It says nothing about water quality. Two setups can both be "MCP" and produce entirely different business outcomes, and the difference lives in two choices.

Raw tables or a semantic layer. A server exposing raw warehouse tables makes the model guess what your metrics mean, which of your many revenue columns is the real one, which orders count. It guesses plausibly and is wrong often, in ways that are hard to spot; we break down the failure classes in Why AI Gets Your Numbers Wrong. A server exposing a semantic layer, a governed set of metric definitions, answers from the same definitions your reports use.

Unmanaged or managed. An unmanaged server, wired up by an engineer with a shared service account, typically means every user sees everything, prompts and results route to default US endpoints, and no log exists of who asked what. A managed server means per-user access control, regional inference, and a full audit trail. Your security team will ask about all three.

For the full technical account of why the middle column works in a demo and fails in production, hand your data lead Connecting Claude to BigQuery: What Works in the Demo and Breaks in Production.

Where Verity fits

Verity ships the right-hand column as a managed service: an MCP server over your semantic layer, running against your own BigQuery, with role-based access per prompt, AI routed through Google Vertex in the Netherlands, and a complete audit trail. The same definitions power Data Chat, so the answer in Claude and the answer in your dashboard are the same answer.

Frequently asked questions

Is MCP tied to Claude?

No. It originated at Anthropic but is an open standard, and the major AI providers have adopted it. A well-built MCP server is a durable investment rather than a bet on one vendor's model.

Do we need MCP if we already have dashboards?

Dashboards answer the questions someone predicted. MCP-connected chat answers the ones nobody predicted, which in practice is most of them. The two share the same foundation and work best over the same semantic layer.

What should I ask a vendor who says they support MCP?

Three things: does the model query raw tables or governed definitions, does access follow the individual user, and can you show me the audit log for a given question. The protocol itself is table stakes; those answers are the product.