How to build governed agentic analytics on Amazon Redshift with the dbt Semantic Layer
A revenue query that ran cleanly and looked completely plausible was still off by $3.5 million, and nothing in the query itself would have told you that.
Overview
AI agents can confidently generate incorrect metrics by misaligning tables or inventing definitions. Covers integrating the dbt MCP server with Amazon Redshift's MCP Server so Semantic Layer metrics compile into governed SQL via MetricFlow, with Redshift executing read-only, pre-governed queries. Includes a live demo and deployable AWS reference architecture.
What you'll take away
- 01
A query that looked plausible, $14.86 million in revenue, was actually $11.36 million, a 30.8% overstatement, because refunds, cancellations, and test accounts weren't accounted for in the raw SQL.
- 02
The fix: a two-layer governed architecture where the dbt MCP Server governs what gets queried against certified metric definitions, and the Redshift MCP Server governs how that SQL executes, with read-only, least-privilege access enforced at three layers.
- 03
Because an agent must select from an enumerated list of pre-defined metrics rather than generating its own, fabricated metrics become structurally impossible, and every answer comes with a receipt: metric name, compiled SQL, and a traceable query ID.
Session speakers
Neela Kulkarni
AmazonVenkatesh Aravamudan
Amazon