Connection Is Not Comprehension: Why AI Agents Need A Unified Semantic Layer

Felix Liao, Senior Director of Product Management at Denodo, specializes in AI-driven data architecture and enterprise data strategy.

​For years, the semantic layer was the behind-the-scenes part of the data stack. It sat beneath dashboards, making sure “revenue” meant the same thing in finance as in sales. When it was missing, a human analyst usually noticed and filled the gap.

Agentic AI has removed that safety net. Standards such as the Model Context Protocol (MCP) have made it remarkably easy to connect AI agents to enterprise data systems.

But connection is not comprehension. An agent can only reason reliably about data when the agent has been given the meaning of the data, and in most enterprises, that meaning is scattered, inconsistent or missing altogether.​

​Why Agents Need Meaning, Not Just Data

​​Data is what a system stores. If a field in a column called “rev_amt” holds the corresponding number 4,200,000, then semantics is everything needed to interpret that number correctly. Is it gross or net revenue? Which currency? Does it include returns? Calendar or fiscal quarter? Who is allowed to see it?

A semantic layer captures those answers in a form machines can use: business definitions, metric logic, the relationships between entities. In effect, it translates tables and columns into the language of the business.

Little of this meaning lives in the data itself. It sits in BI models, spreadsheets, documentation and the heads of experienced analysts. That worked fine as long as people were doing the interpreting.

When a dashboard shows a number, a person decides what to do with it. That person brings context the data doesn’t carry, such as knowing that last quarter’s figure excludes an acquisition.

In traditional BI, inconsistent definitions produced two versions of the truth and an awkward Monday meeting. With AI agents, they produce actions, and the errors compound at machine speed.​