Beyond NL-to-SQLs A Semantic Architecture for Enterprise AI
Preserving Architecture for Enterprise AI over Structured Data
What if the SQL is correct—but the business answer is still wrong?
Enterprises are rapidly adopting natural-language interfaces to databases, warehouses, and semantic models.
The common architecture is simple:
Natural Language → SQL → Data → Natural-Language Answer
But business meaning can be lost at every transition.
- What does Revenue mean?
- Which Customer hierarchy applies?
- Does Last Quarter mean fiscal or calendar quarter?
- How should Margin be aggregated?
- Which exclusions, security rules, and business definitions should apply?
- An AI system can generate perfectly valid SQL and still produce the wrong business answer.
The Missing Problem: Semantic Transition Loss
We define Semantic Transition Loss as the loss or alteration of business meaning as a question moves through:
Natural Language → Interpretation → SQL → Results → AI Response
The solution is not simply better Text-to-SQL.
It is an architecture that resolves meaning once and preserves it throughout execution.
A Better Architecture
Natural Language
↓
Semantic Intent Contract
↓
Governed Enterprise Semantics
↓
Logical Query Plan
↓
Deterministic SQL / DAX / KQL Execution
↓
Typed Evidence + Provenance
↓
Semantic Verification
↓
Trusted Natural-Language Answer
What This White Paper Introduces
Semantic Intent Contract
A machine-readable representation of the user’s resolved business intent—including metric, entity, grain, time, filters, aggregation, rules, and policy.
Semantic Envelope
A way to keep meaning, units, lineage, and context attached to query results throughout execution.
Semantic Checksum
A mechanism for detecting whether important business meaning changed between question, query, result, and answer.
Round-Trip Verification
Validate generated SQL by reconstructing its semantic meaning and comparing it with the original intent.
Enterprise Answer Fidelity
A new way to evaluate enterprise AI—not simply by whether SQL executes, but by whether the final answer preserves the intended business meaning.
Why This Matters Now
Microsoft Fabric IQ, Databricks Unity Catalog Semantics, Snowflake Semantic Views, Looker, and emerging standards such as Apache Ossie all point toward the same shift:
AI needs governed business meaning—not just access to tables and columns.
The next challenge is to preserve and reconcile that meaning across heterogeneous enterprise systems.