Webinar: The Seven Deadly Sins of AI-Generated Data Pipelines

On-demand webinar: seven ways AI-generated data pipelines break, each with its mechanism and its fix. Producing code got cheap. Verifying is where pipelines go to die.

Written by Chris Bergh on August 17, 2026

DataOpsData QualityDataOps TestGenOn-Demand WebinarAI with LLMs
Webinar: The Seven Deadly Sins of AI-Generated Data Pipelines

Key points

  • The meta theme, in Chris's words: when code generation is free, data verification becomes the constraint.
  • The seven sins are gluttony, pride, greed, envy, sloth, vainglory, and lust. Each is a specific failure mode of letting a model write the pipeline, not a general warning about AI.
  • Lineage is an artifact the pipeline produces. It does not help a model iterate toward the right answer, which is why an AI-written pipeline needs test coverage rather than more lineage.
  • Test coverage is measured against columns and tables, not rows. A couple of tests on every column and a couple on every table, so a table with billions of rows is no harder to cover than a small one.
  • Tests come in two buckets. The machine writes roughly 80% of them, and the organization-specific ones live in people's heads and need a UI to manage.
  • Catch the problem before it lands. Write-audit-publish does the write, audits it, and only then publishes, which keeps a bad append out of the table people read.

Producing pipeline code got cheap. Verifying it did not.

Point a coding agent at your warehouse and it writes SQL in seconds. It also writes SQL that runs clean, returns a number, and is quietly wrong, because it has no idea what your business means by “active customer” or which join silently doubles your rows. The code is free now. The checking is the bottleneck.

This session walks seven specific ways AI-generated pipelines break, framed as the seven deadly sins: gluttony, pride, greed, envy, sloth, vainglory, and lust. Each one gets its mechanism and its fix. Chris Bergh makes the argument, including the controversial part: data lineage is not enough for this problem, and test coverage is what an agent can actually iterate against. Then Chip Bloche runs the other half live, profiling a real dataset, generating tests from what the profile finds, and letting an agent propose fixes for the issues that surface.

You will see the coverage rule of thumb that makes a table with billions of rows no harder to cover than a small one, the 80/20 split between tests a machine can write and tests that need someone who knows the business, and the write-audit-publish pattern for keeping a bad append out of the table your reports are built on.

Fifty minutes, with the slides and the full transcript on the same page as the recording. If your team is shipping AI-written pipelines and you cannot tell which of them are right, start here.

The written version, with each sin spelled out, is in The Seven Deadly Sins of AI-Generated Data Pipelines.

Watch the webinar

Fill out the form and the recording, slides, and transcript open on the next page.

Install Open Source TestGen Free, no vendor lock-in Request a Demo See TestGen Enterprise in action
Chris Bergh

Chris Bergh

CEO and Head Chef at DataKitchen. He is a leader of the DataOps movement and is the co-author of the DataOps Cookbook and the DataOps Manifesto.

LinkedIn →

Don't want to give us your email address? Go directly to the webinar recording here.