ServicesAboutNotesContact Get in touch →
EN FR
Note

AI Limitations in Data Engineering

A reading path through the five core limitations of AI in data engineering — SQL failure modes, the context gap, architectural judgment, the production gap, and context engineering as the response.

Planted
dbtaidata engineering

This hub traces five core limitations of AI in data engineering — from specific SQL failure modes to the emerging discipline of context engineering that addresses them. AI adoption in analytics engineering has grown from 33% to 70% in two years (dbt Labs 2025); MIT Technology Review (December 2025) reported that 95% of enterprises found zero value from AI initiatives. The gap is structural: business meaning is not encoded in schemas, tests do not fully verify correctness, and column names mean different things across organizations.

Reading Order

  1. AI-Generated SQL Failure Modes — The concrete problem. 97% of incorrect AI-generated SQL runs without warnings. Research on temporal filter inconsistencies, join failures, and the confidence problem where AI modifies correct code into broken code.

  2. The Context Gap in AI Data Engineering — The root cause. Business context — what “Status” means, whether “Amount” is net or gross, tacit SAP knowledge — lives in people’s heads, not in any system the AI can query. This is what makes data engineering harder for AI than application development.

  3. Data Architecture as Human Judgment — The structural argument. DAG design, ownership models, temporal logic, and team boundaries are where human judgment is most durable. AI writes SQL; humans decide what SQL should exist.

  4. The AI Production Gap in Data Engineering — The practical barrier. AI gets you to 80% fast, but the remaining 20% — security, compliance, governance, temporal consistency — is where most of the real work lives.

  5. Context Engineering for Data Pipelines — The response. The emerging discipline of structuring what AI needs to know: semantic catalogs, instruction files, documentation as AI input, and the workforce pipeline needed to sustain it.