The Seven Deadly Sins of AI-Generated Data Pipelines

Your team is shipping dbt models and DAGs faster than ever, and most of that code came from Claude Code or Cursor. Now count the hours going the other direction: reading it back, checking the grain, working out whether the load ran twice.

The output that hurts you was never the code that failed. It's the code that ran clean and quietly moved a number.

The seven

Seven ways AI-generated pipelines break, each with its mechanics and its fix.

• 

  Gluttony

  The failed run retries and loads your table twice.

• 

  Vainglory

  You ask whether it worked, and the person who wrote the code answers.

• 

  Pride

  It has your schema and no idea that ACTIVE_PENDING means canceled.

• 

  Greed

  4,000 new lines of SQL behind last quarter's 10 tests.

• 

  Sloth

  Writing the tests by hand takes longer than the pipeline did.

• 

  Envy

  You add 'APAC' to a CASE statement; the lineage graph doesn't move, and the commission report drops 30%.

• 

  Lust

  30 seconds from prompt to production, past CI/CD, peer review, and whoever was supposed to check.

Also on the agenda

• The coverage math: why a real warehouse needs more than 5,000 tests, and what happens to the teams that try to write them by hand.
• How to put a testing tool inside the agent's loop over MCP, so the agent finds its own errors before it opens a pull request.

Your presenter