AI now writes the code half of data engineering well (SQL, dbt models, pipeline boilerplate) and stays dangerous on the data half: a wrong transformation runs clean and returns plausible numbers, so every AI-written query needs a validation step the tools do not provide.
AI for data engineers was evaluated in August 2026 using Cursor, Claude Code, GitHub Copilot, and dbt Copilot across pipeline writing, SQL drafting, refactoring, and documentation tasks. The verdict: strong at generating code, weak at guaranteeing correctness.
- ✓Writes and refactors dbt models, Airflow DAGs, and ingestion boilerplate fast
- ✓Drafts SQL window functions, pivot logic, and cross-dialect translations reliably
- ✓Handles bulk refactors like column renames across dozens of models
- ✓Generates model descriptions, column docs, and YAML from existing SQL
- ×Produces wrong SQL that executes and returns plausible numbers undetected
- ×Rewrites queries for cost savings but actually scans more data
- ×Invents column names and hallucinate join paths on large schemas
Scorecard
SQL runs clean but semantic errors pass validation undetected
Terminal-native data engineers adopt agent workflows quickly
Models hallucinate columns and join paths on large warehouse schemas
Gains appear fast, especially for pipeline and SQL boilerplate
Wrong transformations produce plausible numbers with no warnings
A persistent semantic layer that stores schema meanings, approved join paths, and metric definitions where coding agents read them before writing a query and validate against them after. The pieces exist today (dbt semantic models, data contracts, catalog tools) but none sit in the agent loop.