ETL pipelines you set up by asking: ad data to BigQuery without an engineer

Written by Uptimal, the team that builds and operates Upbuild and runs marketing analytics for multinational clients across the globe. Everything below reflects how the product actually works, including the parts that are deliberately boring.

Marketing ETL has been stuck in an ugly middle for a decade: real engineering tools are overkill for "Meta spend into a table every morning", and spreadssheet-era connectors buckle exactly when the account structure gets interesting. The 2026 answer is neither. You describe the pipeline to an assistant, and the pipeline exists.

What "ETL by asking" looks like

In Upbuild Enterprise: "every morning at 07:00, campaign-level spend, conversions and revenue from Meta and Google Ads into BigQuery, one table per platform." The assistant turns that into concrete extracts (platform tool, fields, window), a destination and a schedule, shows you exactly what will run, and the export starts on its cadence. Hourly pacing into a Google Sheet the client watches, a Monday cross-platform digest into Slack: same mechanism, different destination.

The extract layer is the same read-only MCP connector the assistants use: 55 tools across Google Ads, GA4, GTM, CM360, DV360, Meta, TikTok, Snapchat and Adobe, per-user OAuth, no write scopes anywhere. An export can only ever read.

Pipelines as versioned definitions, not chat residue

The failure mode of "AI did my ETL" is a pipeline nobody can explain three months later. Here every export is backed by a data stream: a stored, versioned query definition with its own history. Ask what a pipeline does and the answer is the definition, not an archaeology project through old conversations. Change it by asking, and the change is drafted as a new version, tested against live accounts, and published by a person, with a guard that refuses to silently drop a field a downstream table or dashboard reads. Rollback is one deliberate action.

Runs are recorded with provenance: what ran, when, against which account, delivering how many rows. When a platform API fails, the run reports it and the previous data stands; on a dead connection the schedule backs off instead of burning quota into a wall.

The honest boundary with real ELT platforms

If you are syncing Postgres, Salesforce and Zendesk into a governed warehouse with dbt on top, use the engineering stack; that is its job. The conversational pipeline wins the marketing-shaped slice: ad and analytics platforms, schema known to the connector, destinations your team already reads, latency of "this morning" rather than "this sprint". Teams that priced it out keep landing on the same split: warehouse ELT for the company, assistant-run exports for the marketing layer.

Where it goes next

Once the same definitions feed live dashboards and alerts, the pipeline stops being a separate discipline at all: one versioned description of the data, three consumers. That convergence, not chart generation, is the quiet win of the assistant-first stack.

Frequently asked questions