For decades, enterprises have invested in collecting, storing, and processing increasing amounts of data. Yet teams do not always agree on what that data means. Ask three teams to calculate "revenue" or "product adoption," and you may get three different answers. The answers may be defensible, but the definitions, relationships, and business rules behind them often live only in dashboards, SQL queries, or analysts' heads.

Semantic layers emerged to address this problem by translating technical data structures into a map of how the business defines and organizes data.  Historically, many organizations considered a dedicated semantic layer optional because BI tools and skilled analysts provided sufficient context to turn data into useful insights. Today, however, LLMs are increasingly interpreting questions and querying data on users' behalf.  Unlike experienced employees, they do not arrive knowing which metrics are trusted or which joins are valid.  As enterprises deploy AI agents and chatbots into production, the cost of leaving business meaning implicit has become too big to ignore, moving the semantic layer to the forefront of executives' minds.

An enterprise's ability to ground AI in its own data and institutional knowledge will increasingly determine whether AI becomes a competitive advantage or just another generic interface. Understanding why semantic layers are now so critical requires examining their evolution and the shift in their intended consumer.

Translating Data into Business Terms

Enterprise analytics requires both the technical ability to work with data and the business context to interpret it. Long before the rise of AI, businesses were already trying to bridge this divide. By the 1990s, relational databases had become common across the enterprise, yet using them effectively required knowledge of database structures and query languages.  Business experts often lacked the technical skills to retrieve data, while query writers often lacked the context needed to interpret it. 

Business intelligence software emerged to close that gap: an early Business Objects patent filed in 1991 for a "relational database access system using semantically dynamic objects" describes business-facing objects that map familiar concepts, such as customers and sales revenue, to underlying database structures. Over the ensuing decades, new tools and methods arose as the analytics function matured.  BI tools reduced technical barriers through graphical interfaces; dimensional modeling and curated data warehouses made the underlying data easier to navigate; data products packaged prepared data, documentation, and ownership into a well-defined, reusable asset.

But the technology was only ever part of the system: these tools emerged alongside dedicated BI teams that made enterprise analytics a reality.  An experienced analyst knew which revenue field finance trusted and which joins duplicated records.  When a request was ambiguous, the analyst could ask a follow-up question, consult a subject-matter expert, or recognize that a technically valid result did not make business sense.

The real "semantic layer," therefore, was distributed across the BI tool, the data model, and the tacit knowledge of the people doing the analysis.  This helps explain why dedicated, tool-independent semantic layers never became universal infrastructure during the BI era. While they improved consistency and reliability, the existing process was usually good enough.

A New Consumer of Enterprise Data

LLM-driven analytics changes that equation, as consumers of enterprise data are no longer limited to BI teams working within familiar tools.  Increasingly, a user asks a question in natural language and expects an LLM to interpret the request, identify the appropriate data, generate a query, and explain the result.

The model, however, lacks the BI team's institutional knowledge, and without explicit context, it must infer business meaning.  Because LLMs can produce plausible output even when wrong, missing context creates convincing errors rather than detectable failures.

A modern semantic model addresses this by providing a machine-readable representation of how structured business data should be interpreted and queried.  It encodes metrics, relationships, synonyms, instructions, and examples that connect business language to the underlying data.

Making context explicit is not the only change LLMs bring: as natural-language interfaces are embedded into more applications and tools, more employees are asking analytical questions in more places.  A definition siloed in a BI tool that only supported a bounded set of dashboards may now need to serve many systems operating across several data platforms, making semantic interoperability much harder to treat as optional.  The recent emergence of the Open Semantic Interchange (Apache Ossie), an open effort to standardize the exchange of semantic models across enterprise tooling, reflects the growing value of an interoperable semantic layer.

The enterprises that learn to ground their agents in institutional context, and do so at scale, will be best positioned to create differentiated value in the age of AI.  But enterprise context extends beyond databases: employees also rely on policies, documents, workflows, and knowledge accumulated over time.  If agents are expected to do more than retrieve data, they will need access to that broader context as well. Dataiku is focused on making that context usable across data, models, tools, and workflows so agents can reflect how the business actually operates. Companies that fail to capture and activate this institutional context risk deploying AI agents that are technically capable yet generic.

The rest of this series examines the new practices required to make enterprise context usable by AI. The next installment explores the rise of context engineering and how modern semantic models ground text-to-SQL systems. 

Semantic layers provide AI with a governed understanding of enterprise data. The larger opportunity is enabling AI to apply the knowledge and expertise that sets one organization apart from another.