AI Didn't Change Our dbt Project, Just Who Cared About It

In February 2026, an AI agent answered a question correctly on stage at Starburst’s company kickoff, and the room clapped. The next day, our own executives expected the same thing on real company data: correct, governed, repeatable answers to questions nobody had written down yet.

Two months later, AIDA was live for the executive team across five data products. We didn’t build a new warehouse, copy data, or rewrite pipelines. Years of dbt work (every model and column described, tested, and reviewed in pull requests) had already done the hard part. The agent didn’t need us to write documentation. It needed us to publish the documentation we already had.

In this webinar, Senior Product Manager Monica Miller walks through how Starburst’s data team turned an existing Iceberg lake, dbt project, and Trino engine into governed context for an AI agent, and what surprised them along the way. Every wrong answer traced back to metadata, never to the data. SME sign-off didn’t predict what executives would actually ask. And thin descriptions from excellent people turned out to be a template problem, not a people problem.

What you’ll learn

• Why a mature dbt project is the fastest path to a trustworthy AI agent
• How to structure data products (curated view, metadata contract, named owner) so an agent can actually use them
• How to keep dbt as the single source of truth for descriptions and access, synced through write APIs and Terraform
• Why you should test with the audience you’re shipping to, not just your SMEs