Liftoff narrows analyst scope to make self-service dbt pipelines work

Giving analysts direct access to a production dbt project and training them on it wasn't enough to make self-service stick at Liftoff, according to analytics engineer Jinze Xin. Speaking at a dbt Summit 2026 breakout session, Xin said the fix was to shrink what analysts are responsible for rather than teach them more of the platform.

Liftoff, an AI-powered marketing and monetization platform for the mobile app economy, reported in public filings that its SDK reached more than 1.4 billion daily active users in the first quarter of 2026 across more than 167,000 integrated apps. Its roughly ten-person supply analytics team had faced weeks-long engineering queues for new production datasets, recurring questions answered by rescanning billions of rows of raw event data, and ad hoc requests pulling engineers off planned work.

Xin's team built dbt-pico, a separate dbt project connected to the main engineering-owned project through dbt Mesh cross-project references. Analysts own only business logic and metric grain, while the platform handles incremental strategy, CI, scheduling, review paths, and deployment. An AI agent built on Claude Code skills serves as the entry point, grounding requests in governed metadata and lineage and checking whether an existing dataset already answers the question.

After an early request touched 11 files and added roughly 620 lines, the team consolidated the pattern into a single shared macro, removing 564 lines from the codebase. The agent is limited to governed upstream references and a fixed model shape, and it can propose a model and open a pull request but cannot merge code, run production compute, change a schedule, or trigger a backfill. Deterministic CI checks validate naming, ownership, and metadata, with human-triggered validation paths comparing sample output before merge.

Xin shared early results: one production dataset that previously took around eight weeks to ship through the engineering queue went live in under a week when an analyst built it through dbt-pico. A recurring query that took roughly 30 minutes against raw event data now returns in under a minute against pre-aggregated datasets, and freed engineering time went back into planned project work between March and May.

Source: dbt Labs Blog