Governed knowledge gate for analysis Agents

Let the Agent remember the metric definition before it speaks.

The analysis bottleneck was not another chat window. Dashboards, metric definitions, permissions and source metadata had to line up before a natural-language answer could be trusted. Without that gate, an Agent can produce a fluent query with the wrong business meaning.

We delivered the gate inside that client’s analytics stack: semantic metrics, source metadata, permission-aware query context, SQL linting and model review. The Agent must ground, validate and execute before it explains or renders a chart.

In the work

A governed vocabulary
Metric definitions, dimensions and calculation semantics become grounding context instead of analyst folklore.
Evidence before language
Metadata and SQL checks sit before the response, so the Agent has to reach a real source before presenting an answer.
Permission-aware reuse
The same dashboard, workspace and query controls serve humans and Agents without creating a second truth system.

The problem

A self-service analytics estate carried 3,000+ dashboards and 20,000+ charts, but missing metric definitions made first-pass NL2SQL unreliable. The task was to make business meaning and source evidence available to the Agent before generation.

Measured result

The BI estate now carries 3,000+ dashboards and 20,000+ charts; semantic-cache hit rate rose from 50% to 90%+. Ad-hoc data delivery moved from days to the same day, and SQL governance intercepted 80%+ of erroneous SQL before release. DataClaw first-pass NL2SQL success is about 50%, with the semantic layer addressing the identified metric-definition bottleneck.

Scope