The Verbs Were Never Governed

An executive asks a routine question: “How is year-over-year sales growth in the Northeast?”

The semantic layer may define sales and Northeast. The word doing most of the analytical work is growth.

Gross or net? Same-store or all locations? Percentage change or absolute delta? Fiscal or calendar year? Currency adjusted? Several readings can be legitimate. A metric definition alone may not select one.

We governed many of the nouns of the business. The long tail of verbs still contains choices.

The nouns are necessary

Tables, columns, and keys describe the data model. Semantic models add business meaning: approved measures, dimensions, hierarchies, joins, and sometimes derived calculations. That work is valuable. An AI analyst should use it, not route around it.

But business questions rarely stop at a noun. They ask whether revenue is growing, a segment is churning, a product is underperforming, or customers are becoming less engaged.

Those verbs combine metrics, populations, comparisons, periods, thresholds, and judgment.

Some semantic systems can encode some of this logic. The problem is not that they govern nothing. It is that the combinations used in ordinary questions often extend beyond what the organization anticipated and authored.

Analysts have always supplied the missing step

A human analyst reads “losing regular buyers” and translates it into an analysis. They may reuse a known cohort, choose a period, compare against a baseline, and ask what “regular” means for this team.

That interpretation can be good or bad. The useful part is that another person can ask the analyst what they chose and why.

In strong teams, the answer points to a shared definition, local convention, prior analysis, or explicit clarification.

Metadata and lineage systems make much of the technical path inspectable. They do not, by themselves, record why a particular analytical composition was right for this question.

AI makes the hidden choice easier to overlook

A language model can interpret the same phrase and generate a polished result. If the system does not expose the interpretation, the choice is buried inside a fluent workflow.

Changes in phrasing, context, or model behavior can also change a material choice without making that change obvious to the reviewer.

The risk is not that every run will differ. It is that a difference in meaning can be silent.

The evaluation question, then, is not only whether the model found the right table. It is whether the system shows the analytical verbs it applied:

• What did “growth” mean here?
• Which population was compared?
• Which period and baseline were used?
• Which definitions came from governed context?
• Which remaining choice required clarification?

Govern the interpretation before the result

Our approach is to represent the intended analysis as a visible plan before SQL executes.

Available semantic definitions ground the nouns. The plan also exposes the operations being applied to them: comparison, ranking, segmentation, filtering, and time logic.

When an unresolved choice would materially change the analysis, the system should ask. When a governed definition resolves it, the system should cite that source.

Query generation can then be checked against the material bindings in the accepted plan.

This is not a claim that every analytical verb can be pre-modeled or that a visible plan makes an answer automatically correct.

It is a design principle: do not let the most consequential part of the question disappear between the prompt and the SQL.

The semantic layer gave us a strong foundation of governed nouns. AI analytics needs a reviewable way to compose the verbs around them.

See what an interpretation plan makes visible: how Spotonix works.