Data Quality / Metadata / AI Governance
AWS architects show how a semantic layer stopped an AI agent overstating revenue by 30.8%
At dbt Summit 2026, two AWS architects described routing agent queries through certified dbt metrics and read-only Redshift execution, rather than letting the model write its own SQL.
Two AWS solutions architects, Neela Kulkarni and Venkatesh Aravamudan, presented a governed design for AI agents querying production data on Amazon Redshift, according to a dbt Labs blog post on their dbt Summit 2026 session. Their example: an agent-style query summing gross order amounts returned $14.86 million, while actual revenue was $11.36 million, an overstatement of 30.8%. Four business rules on discounts, refunds, order status and test accounts existed only in the semantic layer, not the schema.
In their design, the dbt MCP Server limits agents to an enumerated list of certified metrics and MetricFlow compiles the SQL; the Amazon Redshift MCP Server enforces read-only, single-statement execution at the server, IAM and database-grant levels. Each answer carries a receipt with metric name, compiled SQL, source model, dbt test results and a Redshift query ID.
Editor’s note: This article relies solely on a blog post by dbt Labs, a vendor whose product features in the architecture described.
Why it matters
The example shows that business rules held only in people's heads or a semantic layer are invisible to an agent reading the schema, so metric definitions become a control, not a convenience. Data owners should check whether key metrics such as revenue are formally certified, with exclusions like refunds and test accounts documented in metadata. AI governance leads can use the 'receipt' pattern — metric, compiled SQL, test results, query ID — as a benchmark for auditability of agent answers.
Status: Published. Conference presentation reported by dbt Labs
Sources
Was this useful?
Thanks. What was the problem?
Thank you — this helps us choose what to cover. If you spotted an error, please tell us what it is.
Prepared with AI assistance and reviewed by the editor before publication. Spotted an error? Tell us — see our editorial policy and methodology.
Governance areas
Who should care
- CDO
- Data stewards
- AI governance lead
- Analytics engineering leads