Why a semantic layer isn't enough for AI

Every data platform is shipping a semantic layer now, and most of them are free. Define a metric once, the pitch goes, and everyone downstream reads it the same way. It is a good idea, and an overdue one. It is also not the layer that AI needs, and the gap between the two is about to matter far more than it ever did for dashboards.

What a semantic layer does

A semantic layer is a dictionary for your dashboards. It takes the terms a business argues about, revenue, an active customer, gross margin, and pins each to one definition, so a query returns the same number wherever it is asked. That consistency is real and worth having. For the BI era it was often the missing piece.

The job ends at the definition. A semantic layer states what a metric means. It does not answer where the number came from, who stands behind it, when it was last agreed, or whether the answer ever reaches the person who has to act on it. Those were acceptable gaps when the output was a chart a human read and sense-checked. They stop being acceptable the moment something acts on the number without pausing to doubt it.

The same definition, a different job: governed, in the work, and yours.

Defining meaning is not governing it

There is a difference between a definition and a governed definition, and it is the difference between a number you can read and a number you can trust. A governed definition carries three things the metric alone does not: the source it is computed from, the owner who set it, and the date they set it. Lineage is first-class, not a diagram someone maintains on the side. When a figure is challenged, the answer is not a week of reconciliation. It is a trace, in seconds, to the record and the person behind it.

Definition, source, owner and date turn a number into a trace.

That is the line between a dictionary and a system of record for meaning. A dictionary tells you what a word should mean. It does not tell you who is accountable when the word is wrong.

Meaning has to reach the work

A definition that lives only in the query layer changes what a dashboard says. It does not change what happens on Tuesday morning. The value of a trusted number is realised where a decision is made: in the exception queue a controller works, the alert that fires before a cost lands, the fix in a process that was quietly broken. A semantic layer feeds a chart. Getting meaning into the operation, into the systems people already use and the way they already work, is a different discipline, and it is most of the distance between a defined metric and a result.

AI runs on meaning, not on dashboards

Here is why the gap is about to matter. A dashboard has a human in the loop who notices when a number looks wrong. An agent does not. Put a model on top of definitions that disagree with themselves, that carry no owner and no lineage, and it will produce answers faster and trust them less deservedly. It will be confidently wrong, at scale, in the language of certainty.

Intelligence on top is only as trustworthy as the meaning underneath it. A semantic layer was built to make BI consistent. The layer AI needs has a larger job: to be the one source of meaning that both people and agents run on, governed the whole way, so that everything built on top inherits the same definitions, the same owners, the same lineage. The phrase "under your data and AI" is doing real work there. A metrics layer was never asked to hold up an agent.

Meaning that compounds

There is one more property the dashboard era never needed. When meaning is governed, using it makes it sharper. Each definition settled, each owner named, each process wired in becomes a foundation the next one reuses instead of rebuilding. Ungoverned, reuse decays into a sprawl of conflicting spreadsheets that trust each other less over time. Governed, reuse compounds: the tenth answer stands on the nine before it, and the organisation gets steadily smarter without starting from zero again. A semantic layer is static by design. Meaning that learns as it is used is not.

Every settled term becomes the starting point of the next answer.

Owned, or rented

The last difference is where the layer lives. The semantic layers the platforms are shipping live inside their walls, and they are free for the same reason a first taste is free. The meaning of your business is the last thing you want to rent. A layer that holds it should be sovereign, run where you decide, and stay yours, contracts and definitions included, portable across whatever platforms you use now and later.

A system of intelligence

Put those together, governance in place of definition, meaning that reaches the work, one source that people and AI share, value that compounds, ownership that stays, and you are no longer describing a semantic layer. You are describing a different layer with a different job.

Datahub calls it a system of intelligence: the governed layer under your data and AI that gives an organisation one trusted, self-learning source of meaning, and keeps it theirs.

A semantic layer defines a metric. A system of intelligence governs meaning, puts it to work, and owns it. Definition is where the semantic layer stops. It is where the system of intelligence starts.