Semantic layer for AI: why your business definitions decide where your AI runs

For most of the last twenty years, data platform vendors competed over where your data was stored and which engine processed it. That contest is winding down. Open table formats such as Delta Lake and Apache Iceberg let most modern engines read the same files, and every major platform can now read its rivals' data. The competition has moved up a layer, to the place where your organisation's numbers get their meaning.

That place is the semantic layer. It is where "revenue", "active customer" and "late order" are defined, where tables are joined the right way and where access rules are applied. Whoever holds it decides where your reports run and where your AI agents get their answers. Call it semantic layer supremacy. This article looks at who is competing for it, why a semantic layer for AI has become the prize, how to build one properly, and why Microsoft wants yours to sit in Fabric whichever platform holds the data underneath.

Why AI needs a semantic layer

Ask an AI agent "what was revenue last quarter?" and it has to turn the question into a query. Without a semantic layer it reads table and column names, guesses which ones matter and guesses how to join them. Then it returns a figure with complete confidence, whether the guess was right or not.

The research points one way. In a 2023 benchmark by data.world, GPT-4 answered business questions over an insurance database correctly 16% of the time when it queried the tables directly, and 54% of the time when it worked through a knowledge graph of the same data. dbt Labs reran a similar test in April 2026 with current models. Better models have narrowed the gap, but on a well-modelled project the semantic layer still lifted accuracy from 90% to 98% for one model and from 84% to 100% for another. dbt makes the more useful point itself: a semantic layer answers from a defined metric or says it cannot. It never quietly returns the wrong number.

Vendors run most of these benchmarks on small sets of questions, so read them as a direction of travel. The analysts have reached the same view. In March 2026 Gartner predicted that "by 2030, universal semantic layers will be treated as critical infrastructure, alongside data platforms and cybersecurity". In May it forecast that organisations which prioritise semantics will raise the accuracy of their AI agents by up to 80% by 2027.

Reports used to be the main consumer of business logic. AI agents are taking that role, and an agent that reads its definitions from one vendor's semantic layer runs on that vendor's platform. That is why every vendor wants to hold it.

An open semantic layer standard, with limits

In September 2025 Snowflake, Salesforce, dbt Labs and others launched the Open Semantic Interchange, an effort to agree one vendor-neutral format for semantic definitions. Databricks joined when the first specification was published in January 2026. In mid-2026 the project moved to the Apache Software Foundation as Apache Ossie. At FabCon Barcelona on 28 September 2026, Microsoft committed to it, and its engineers contributed a converter between Power BI semantic models and the Ossie format. Google is reported to be joining.

So the vendors competing for the semantic layer are also agreeing to make it portable. That looks like a ceasefire. In practice it changes the contest in two ways.

The first is that the standard is young. The current specification is a version 0.2 development draft. Tabular Editor's review of the Microsoft converter found that only simple metrics, such as a single sum or count of a column, translate into DAX. Row-level security, calculation groups and hierarchies have no equivalent in the standard yet. The parts of a semantic model that take the most work to get right are the parts that do not travel.

The second is that portable definitions move the contest to what sits on top. If any platform can read your metric definitions, vendors compete on what they do with them: the AI assistant, the reports, and the tools your staff already use every day. Microsoft is better placed than anyone to win that contest.

How to build a semantic layer for AI

For a mid-sized organisation, the practical lesson is this. You do not need to bet on which storage platform wins. You do need a semantic layer that is built properly, because the same model now serves your reports and your AI. A model that works for one dashboard can still mislead an agent.

Six practices make the difference:

• One model per business area, shared by many reports. Finance, sales and operations each get an endorsed model that everyone builds on. Dozens of near-identical models, each with its own version of revenue, defeat the purpose.
• Every definition written once, as a measure. Logic buried in report visuals, SQL views or spreadsheets is invisible to an AI agent and drifts from everything else.
• A clean star schema. Fact tables of events join to dimension tables of things, with the relationships set in the model. Agents get joins wrong when the structure is unclear, just as people do.
• Descriptions and synonyms. AI reads them. A measure called "Rev_Net_Adj" with no description gets guessed at. One described as "net revenue after credit notes, excluding VAT and intercompany sales" gets used correctly. Fabric's AI instructions and verified answers, both in preview, build on this.
• Security in the model. Row-level security set in the semantic model applies to every report and every agent, so an agent only answers with what the person asking is allowed to see.
• Version control. Keep the model definition in Git so that changes are reviewed, tested and reversible like any other code. Fabric stores models in a readable text format for exactly this.

If your data already lives in Databricks or Snowflake, you do not have to move it to benefit. Mirror it into OneLake, build the semantic model in Fabric and leave the engineering where it is. If your team models in dbt, keep dbt for transformation and use the semantic model for meaning. Simple metric definitions may move between the two through Apache Ossie in future, but do not design around that until the standard matures.

The organisations that get this right will find each new report and each new AI agent is a small addition on top of definitions they already trust. The ones that skip it will keep rebuilding the same logic, in more places, for more tools.