← Back to Writing
2026-11-12 • 4 min read
Why NL-to-SQL Fails Without a Semantic Layer (and How to Govern It)
AI-for-BISemantic LayerGovernanceH·AI·BI
Most natural-language-to-SQL demos look magical on a 3-table toy database, then fall apart the moment they hit a real enterprise warehouse.
When a business user asks “What was net margin by region last quarter?”, raw schema prompting runs into three immediate failure modes:
- Ambiguous business definitions: Three tables have columns named
amount,revenue, ornet_rev. Without an explicit semantic layer defining the canonical metric formula, the LLM guesses. - Silent permission leaks: A generic SQL generator doesn’t know which rows or fields a given user role is allowed to see unless Row-Level Security (RLS) and field sensitivity rules are enforced centrally, outside the prompt.
- Unverifiable answers: Executives don’t trust a chart if they cannot trace the exact SQL, tables, and filters behind it.
The Architecture Pattern That Holds Up
In H·AI·BI™, natural-language analytics is structured as a governed pipeline rather than a raw prompt wrapper:
- Semantic & Metric Layer first: Business entities, joins, and metric definitions live in a versioned semantic model with a
DEV → QA → PRODpromotion path. - Schema-aware guardrails: Generated queries are validated against allowed tables, columns, and read-only constraints before execution.
- Centralized Governance: RLS and field sensitivity policies apply uniformly to both Ask and Deep Dive queries.
- Evidence-linked Deep Dive (
[E#]): Multi-step analytical answers cite their underlying query outputs so analysts and leaders can audit every number.
AI-for-BI only earns adoption when the governance model is stronger than the prompt.