Shrink the surface: governed self-service for analyst-owned dbt pipelines

Most self-service dbt programs fail by giving analysts a full project and hoping governance catches up. The result: unreviewed SQL in production, engineers debugging code they didn't write.

We used dbt Mesh to give analysts a separate AI-native project, isolated from our production codebase. Analysts own business logic and metric grain. The platform owns everything else: microbatch config, schema isolation, CI, downstream registration. 10+ analysts have shipped production pipelines without touching deployment or backfill logic.

AI makes this boundary work. We built Claude skills: AI agents encoding dbt, microbatch, and git knowledge into repeatable steps. Analysts bring SQL; the skills scaffold models, generate tests, manage PRs, and enforce conventions.

This talk covers how we drew that boundary, the failures that reshaped it, and what we'd change.