Blog

Why MCP alone is not enough for AI analytics

Agentic Analytics · June 25, 2026 · 4 min read

MCP is how a lot of AI agents now connect to data. It is a standard, introduced by Anthropic, for plugging an agent into tools and data sources like a warehouse. It solves a real problem, which is access. What it does not solve is meaning, and for analytics that is the harder half.

Here is the short version. MCP gets your agent to the data. A semantic layer is what makes the answers it gives correct. You need both, and MCP without the second half is where a lot of agentic analytics projects are heading for trouble.

What MCP does

The Model Context Protocol is a standard way for an AI agent to discover and use external tools and data. Instead of a custom integration for every system, a tool exposes an MCP server, and any MCP-capable agent can call it. For data, that means an agent can connect to a warehouse, list tables, and run queries through a common interface.

This is genuinely useful. It is why MCP took off. Connecting agents to systems used to be bespoke work, and MCP makes it standard.

What MCP does not do

MCP is a transport. It moves requests and results between an agent and a data source. It says nothing about what the data means.

So an agent connected to your warehouse over MCP can see the tables, but it does not know that “revenue” is defined one way by finance and another by sales, or which of four user tables is canonical, or which filter everyone applies before counting active accounts. MCP hands the agent a connection and raw schema. The agent still has to guess the meaning, and on real data it guesses wrong. That is the same text-to-SQL failure as always, now reachable over a standard protocol.

What Gartner says

This is not a fringe concern. At the Gartner Data and Analytics Summit in 2026, analyst Andrés García-Rodeja predicted that 60 percent of agentic analytics projects relying solely on MCP will fail by 2028, for lack of a consistent semantic layer. The point was not that MCP is flawed. It was that MCP without a semantic and governance foundation has nothing stable to stand on.

The architecture that works

The fix is not to drop MCP. It is to put a semantic layer between the agent and the raw data.

MCP handles access. The agent connects to your systems through it. The semantic layer handles meaning. It defines metrics, dimensions, and rules once, so when the agent asks for revenue it gets the agreed definition instead of guessing at a column. Together they give the agent a governed path to a correct answer, meaning a standard way in and a defined set of things it is allowed to ask for.

This is a context engineering problem. The protocol is solved. The context, meaning the metrics, rules, and definitions the agent reasons over, is the part you build. Open-source agents like nao are built around exactly this, pairing standard connectivity with a governed context layer so the agent reaches for a defined metric rather than writing raw SQL and hoping.

The takeaway

If you are evaluating an agentic analytics setup, do not stop at “does it support MCP.” They mostly do, and connectivity is table stakes. Ask what sits between the protocol and the answer. An agent with MCP and no semantic layer can reach your data and still be confidently wrong. An agent with both can be trusted with a number.

Is MCP enough for AI analytics on its own?

No. MCP is a standard way to connect an agent to data and tools, but it does not define what the data means. Without a semantic layer to resolve metrics and rules, the agent guesses at definitions and returns confident wrong answers. Gartner predicts 60 percent of agentic analytics projects relying only on MCP will fail by 2028 for this reason.

What is the difference between MCP and a semantic layer?

MCP is about access. It is a protocol that lets an agent connect to and query a data source through a standard interface. A semantic layer is about meaning. It defines metrics, dimensions, and business rules so the agent uses agreed definitions instead of guessing. MCP gets the agent to the data, the semantic layer makes its answers correct, and analytics needs both.

Do you need a semantic layer if you use MCP?

For analytics that has to be right, yes. MCP alone gives the agent a connection and raw schema, so it has to infer what “revenue” or “active user” means and often gets it wrong. A semantic layer supplies those definitions, which is what keeps the agent’s answers consistent and correct across questions.