1. Data says nothing on its own
Take one row in a media group’s table: “Video ID 84721”. To the machine, it is a number. To the business, it is an episode, attached to a programme, in the sport genre, season 2. Everything that separates the number from what it means is semantics.
- DataVideo ID 84721
- ConceptEpisode
- ContextAttached to a programme
- Attributes and relationshipsGenre: sport · Season: 2
Semantics is a common language: it defines business concepts, the relationships between them and the rules for interpreting them. Its goal is easy to state and hard to hold: the same word keeps the same meaning across business teams, data, tools and AI models.
Data says what is stored. Semantics says what it means. An agent needs both.
2. What agents change
Until now, meaning mostly lived in people’s heads. The analyst knew that “revenue” on the dashboard excluded credit notes, that two tables could not be joined just any way, that an “active customer” was not defined the same way in marketing and in finance. That knowledge was written nowhere, and it was enough.
An agent does not have that knowledge. It reads column names, descriptions when there are any, and fills the gaps with whatever seems plausible. The result is a well-written answer, with figures, that is wrong.
| What is missing | What happens with an agent |
|---|---|
| A shared definition | “Active customer” is computed one way for marketing and another for finance: the agent picks one without saying so. |
| A reference calculation | Two similar questions return two different revenue figures, because the agent rebuilt the calculation each time. |
| Relationships between objects | The agent joins two tables on the wrong key and counts some rows twice. |
| Business rules | An exclusion everyone knows about (internal tests, divested subsidiaries, frozen periods) is not applied. |
| The sensitivity level | Confidential data is handled like any other, because nothing marks it as such. |
None of these problems is new. What is new is that a person who knows the context no longer catches them. The more autonomy you give agents, the more the meaning has to be written somewhere they can read it.
3. What a semantic foundation contains
- Concepts
- The objects the company talks about: customer, contract, episode, order, site, incident.
- Definitions
- What each concept covers exactly, and what it does not.
- Relationships
- How concepts hold together: an episode belongs to a programme, a contract ties a customer to an offer.
- Business rules
- The conditions and exclusions that apply: scope, effective dates, special cases.
- Metrics
- The reference calculations, written once: revenue, audience, resolution rate.
This foundation can take several forms, from the lightest to the richest. They are not exclusive: you usually start with the first and add the others where the need justifies it.
| Form | What it is | What it gives agents |
|---|---|---|
| Business glossaryBusiness glossaryThe list of terms the company uses, each with a definition approved by the teams that use them.See the glossary | The list of the company’s terms and their definitions, approved by business teams. | An unambiguous vocabulary to understand the question asked. |
| Semantic layer | Metrics and dimensions defined once, on top of the tables. | The same figures whichever tool or agent asks for them. |
| OntologyOntologyA structured description of the objects of a business, their relationships and their rules, written to be read by a machine.See the glossary and knowledge graphKnowledge graphA way of representing information as connected objects, so you can move from one to the next: from an episode to its programme, from a customer to their contracts.See the glossary | Concepts, their relationships and their rules, described in a machine-usable way. | The ability to reason step by step: from episode to programme, from programme to contract. |
4. Who does what: business, tech and governance
Semantics is not a project the data team can run alone. Business teams know what the words mean; tech teams know where the data is and how to make it reliable. The semantic foundation is where the two meet, and the agenticAgentic AIAI that does more than answer: it chains steps together and acts in tools to reach a goal.See the glossary platform is what makes it usable every day.
Data and AI governanceCommon rules and responsibilities · Data quality and lifecycle · Access and compliance
- Define the concepts
- Align metrics and vocabulary
- Express needs and business rules
Agentic platformMakes knowledge queryable and actionable, with rights and guardrails
- Make data reliable
- Structure the semantic repository
- Integrate, secure, trace
Shared semantic foundationConcepts · Definitions · Relationships · Business rules
- Business teams
- They define the concepts, align metrics and vocabulary, and express their needs and rules.
- Tech teams
- They make data reliable, structure the semantic repository, integrate it into tools, secure it and trace its use.
- The agentic platform
- It makes knowledge queryable and actionable by agents, with rights and guardrailsGuardrailsAutomatic checks that block risky use: sensitive data leaks, forbidden content, out-of-scope actions.See the glossary .
- Data and AI governance
- It sets the common rules and responsibilities, tracks data quality and lifecycle, and frames access and compliance.
A semantic foundation is first an agreement between people on what words mean. The tool comes second.
5. How agents use it
- Querying data: instead of writing a query by guessing the tables, the agent asks for a defined metric (“audience by programme, season 2”) and the semantic layer produces the reference calculation.
- Searching documents: concepts act as common labels, so every piece of content linked to a programme, a customer or a contract can be found, whatever wording was used.
- Acting in a process: business rules become checks the agent must pass before creating, changing or approving.
- Working with several agents: a common vocabulary stops two agents passing each other the same word with two meanings.
- Respecting rights: sensitivity is carried by the concept, not by each table; access is decided once and applies everywhere.
In practice, the semantic foundation is exposed to agents like any other system: through an MCP serverMCP serverThe small service that exposes a tool (an application, a document base) to agents through the MCP protocol.See the glossary , behind the gatewayGatewayA single checkpoint placed in front of services, controlling everything that goes in and out: identity, rights, quotas, logging.See the glossary , with the same rights and the same traceabilityAudit logThe record of who did what, when and with which data. It lets you check and explain every action.See the glossary as the restRESTThe most common style of web API: data is exchanged through addresses and simple verbs (read, create, update, delete).See the glossary of the platform.
MCP: the protocol, its limits, and what it needs around itHow agents plug into your systems, and what the protocol does not do for you.6. What you already have
Nobody starts from scratch. Part of the meaning is already written, scattered across tools that do not talk to each other. The work is less about creating than about gathering, arbitrating and making it readable by agents.
- The semantic models of your BI tools, where metrics are already defined.
- The metric layers of the data platform: dbt, Snowflake, Databricks and Microsoft FabricMicrosoft FabricMicrosoft’s data platform bringing together storage, transformation, analysis and Power BI dashboards.See the glossary each offer one.
- The data catalogue and its metadataMetadataThe information that describes a piece of data: its name, origin, owner, definition and sensitivity level.See the glossary , when it is kept up to date.
- The company’s reference dataReference dataThe reference list for one type of object in the company (customers, products, sites), which every system is meant to share.See the glossary : customers, products, organisation, contracts.
- The glossaries and scoping documents written by business teams, often forgotten in a shared drive.
On the standards side, the W3C languages (OWL for ontologies, SKOS for vocabularies) have been around for a long time. More recently, several data vendors launched the Open Semantic Interchange initiative so that one definition can travel from tool to tool. The question is no longer whether a semantic layer is needed, but how to avoid having five that contradict each other.
This article recommends no product. The right choice depends on the data platform already in place and on the maturity of the teams.
7. Where to start
- Pick one domain and one precise use case, not “the whole company”: audience questions, for example, or contract tracking.
- List the twenty or so concepts and metrics that use case relies on.
- Define them with the business teams involved, and settle the disagreements: this is the longest and most useful step.
- Write them in the tool you already have, rather than buying a new one.
- Expose them to an agent, then test it on a set of questions whose right answer is known.
- Measure, correct, then extend to the next domain.
Dasein helps put this foundation in place: we run the definition work with business teams, write it into your data platform and make it available to agents, under governance.
AI Platform: the shared foundation for agentsThe platform that makes this knowledge queryable and actionable, with rights and guardrails.Further reading
- W3C: Web Ontology Language (OWL) (opens in a new tab)
- W3C: Simple Knowledge Organization System (SKOS) (opens in a new tab)
- dbt: Semantic Layer (opens in a new tab)
- Snowflake: Cortex Analyst and semantic models (opens in a new tab)
- Databricks: Metric views (opens in a new tab)
- Microsoft: Semantic models in Power BI (opens in a new tab)
- Open Semantic Interchange: cross-vendor initiative launched in 2025
Share this article
