business-intelligence

# Business intelligence Answer business questions over the org's data **through a governed semantic layer** — define each metric once in versioned YAML, then route every "what was revenue last quarter by region" through that layer. Numbers come out consistent, auditable, and the same for everyone. This skill builds the layer and queries it in plain language. ## The one rule **Never free-hand SQL against raw tables to answer a governed business question. Go through the layer.** Why, with numbers: in dbt's April 2026 benchmark (ACME Insurance, 11 questions × 20 runs, ~15-table schema), an LLM grounded in a semantic layer scored **98.2% (Claude Sonnet 4.6) / 100% (GPT-5.3 Codex)** vs **90.0% / 84.1%** for raw text-to-SQL on the same schema; on the *unmodeled* schema it was 72.7% vs 64.5%, and a 2023 GPT-4 baseline managed 32.7%. The layer is not bureaucracy — it is the accuracy. The model writing SQL against undecorated tables is the failure mode you are eliminating. Your job is two motions: (1) **build** the metrics layer (entities, dimensions, measures, metrics) and (2) **query** it — translate a plain-language question into metric + dimensions + grain + filter, never into a hand-written query. ## The four primitives Every semantic layer (MetricFlow, Cube, warehouse-native) is built from the same four nouns. Learn these and the rest is syntax. - **Entities** — the join keys. `order_id` is the primary entity of `orders`; `customer_id` is a foreign entity that joins to `...