Semantic layer: why AI guesses without agreed definitions

Ask three departments for last quarter’s revenue and you get three numbers. Finance counts by invoice date, sales by order date, and logistics treats returns differently. Colleagues know this and ask follow-up questions. An AI assistant does not. It picks a table, writes a query and returns one answer with complete confidence.

That is why the semantic layer is firmly back on the agenda since AI agents arrived. The concept is not new: BI tools have used it for decades. What is new is that the consumer of the data is increasingly a language model with no context, no colleague to call and no sense of when a number looks wrong.

What a semantic layer is

A semantic layer is an agreed translation layer between raw data (tables, columns, source systems) and whoever asks a question about it. It holds:

• Definitions. What is an active customer? What counts as revenue, and what does not?
• Calculations. How is gross margin calculated, at what level of detail, in which currency?
• Relationships. How do order, order line, product and customer connect, and on which keys?
• Synonyms. Debtor, account and customer mean the same thing in this organisation.
• Rules. Which filters always apply, such as excluding intercompany postings or counting completed orders only.

A dashboard, an analyst and an AI agent then all read the same definitions. The answer to “what was revenue in Q2?” no longer depends on who asks or which table happens to be found.

Why AI makes it urgent

In a classic BI environment the semantics often sit inside the report itself: a calculation in a dashboard, a filter a developer once built in. That works as long as people look at the data through that report.

AI applications bypass it. Text-to-SQL assistants, AI agents and integrations through the Model Context Protocol (MCP) access the data directly. Without recorded semantics, the model has to guess which column holds revenue, whether a customer who has not ordered in two years is still active, and how two tables join. The result looks convincing and is regularly wrong.

Vendors see this too. For questions on structured data, Qlik Answers uses the data model and business logic of Qlik apps. Snowflake added semantic views, so Cortex Analyst answers questions from defined metrics rather than raw tables. And in 2025 Snowflake, dbt Labs, Salesforce and others launched the Open Semantic Interchange initiative to make semantic definitions portable between tools. The direction is clear: semantics are moving out of the report into a shared layer that every consumer can read.

Where DAMA comes in

The technology of a semantic layer is the easy part. The hard part is that someone has to decide what revenue means, and that the agreement stays correct when processes change. That is not an IT question. The DAMA DMBoK framework offers three knowledge areas that apply directly:

Metadata Management. A semantic layer is active metadata: descriptions that are executed, not only read. A catalogue with a business glossary is where definitions live in plain language, with an owner and a version history.

Data Modeling and Design. A good conceptual and logical data model describes the terms and their relationships before a single table is built. If you have that model, most of the semantic layer is already on paper.

Data Governance. Every definition needs an owner with decision rights, and a process for changes. Without that, the semantic layer drifts away from reality over time, and nobody notices until an AI answer in a board meeting turns out to be wrong.

See the DAMA Wheel for how these knowledge areas connect to data quality and the other disciplines.

Where it goes wrong in practice

• Starting with the tool. Setting up a semantic layer in a new platform without agreement on the definitions gives you a tidy, technically correct layer with the same contradictions as before.
• Wanting everything at once. A complete semantic layer for the whole organisation is a multi-year project. It only pays off once it is finished, and it never is.
• Definitions without an owner. “The data team will handle it” does not work. The data team can build a calculation, but cannot decide whether returns are deducted from revenue.
• Several layers side by side. One semantic layer in the BI tool, a second in the data warehouse and a third in an AI application. The original problem has simply moved.
• No tests. A semantic layer is software. Changes should be tested against known answers, just like code.

Starting small

A workable start looks similar in most organisations:

1. Pick one domain with a concrete AI or analytics question. For example stock and lead times at a wholesaler, or order intake at a manufacturer.
2. Record ten to twenty core terms in a business glossary: definition, calculation, owner, source system.
3. Make them executable in the platform you already use, such as the Qlik data model or Snowflake semantic views.
4. Build a test set of questions with known correct answers, and check every change against it. This is also how you assess an AI assistant before it goes into production.
5. Expand domain by domain, with the same agreements on ownership and change.

The gain is reuse. A logistics company that defines delivery reliability properly once then uses that definition in dashboards, in customer reporting and in an AI assistant for the planning team. This ties directly into data products: a data product without defined semantics is only half usable for an AI consumer.

How we help

At cimt the semantic layer usually comes up as part of an AI Data Readiness assessment or an analytics engagement. We map which terms are critical for the intended AI application, where definitions currently differ, who decides on them, and which platform is best placed to carry the semantics. We anchor governance in the DAMA knowledge areas, so the layer is maintained after the first delivery.

Is an AI application on the roadmap and you are not sure the underlying definitions are in order? Then a short conversation is worthwhile. Book a meeting or read more about the AI Data Readiness assessment.