How AI is Changing the Data Modeling Process

For years, the analytics engineer's job has been strictly defined by the pipeline: extract here, transform there, join these tables, schedule that job. It was meticulous, sequential, and largely manual work. But a new generation of agentic AI workflows is reshaping that reality, and the role of the data engineer is shifting along with it. 

Xebia’s whitepaper Evolving Data Modelling for AI explores this change, exploring several areas where data modeling solves real business problems, from business alignment and warehouse migration to scalability and AI readiness. 

Let’s find out how this data modeling process is being affected by AI in a dramatic way. 

From step-by-step pipelines to reviewing AI-proposed models 

Rather than building pipelines step by step, as it was in the past, users can now interact with agents that analyze data patterns, propose transformations, and guide the sequence of steps from raw data to a usable dataset. The engineer's role is transformed, now being one of review, adjustment, and approval, it has become a fully iterative workflow. 

This shift makes self-service analytics more realistic than ever. Business users are now able to ask questions in plain language, and AI assistants can propose the joins, transformations, and metric definitions needed to answer them. The path from raw data to insight gets shorter.  

But this new process also raises one critical question: what happens when the AI confidently proposes a model that's just plain wrong? 

Why AI-ready data needs stronger governance

LLMs depend on well-documented semantic layers to navigate complex schemas. This is a clear advantage over the past but in the absence of clear boundaries, AI tools can hallucinate queries and they can do so with total confidence. As the whitepaper puts it, the same inconsistency that confuses humans now drives hallucinations at machine scale, propagated across every dashboard and downstream report that reads from the same model. 

This is why governance matters more than ever. As AI supports more decisions, it becomes more important to understand how those decisions are made, who is responsible, and how they can be traced. A solid data model is the layer that makes all of this possible. 

In the Evolving Data Modeling for AI whitepaper, we propose an AI-Readiness checklist that captures what a solid data model actually looks like in practice: 

A data model based on these standards will actually be an essential tool for your company’s data modeling and avoid future issues and potential AI hallucinations. 

The three context layers AI needs... and the one it's missing 

Xebia’s Data modeling whiepaper also describes the three layers in which data systems carry context.  They are: 

When AI fails at a data-modeling task, the missing piece is almost always the human layer. Joe Reis's 2026 State of Data Engineering survey outlines this issue quite concretely: 82% of practitioners use AI tools daily for data work, but only 11% say their data modeling is going well. The conceptual phase of data modeling, that is externalizing the human layer into a written and shared model, is what is needed to close that gap. 

Why agentic workflows depend on a semantic foundation 

It is necessary to accept that while a new generation of agentic workflows makes self-service more realistic and shortens the path from raw data to insight, Large Language Models (LLMs) still depend on well-documented semantic layers to navigate complex schemas. Without clear naming, well-defined metrics, and a single source of truth, AI assistants invent joins, misread grain, and produce outputs that look right until they're checked. 

We've seen this distinction be the deciding factor again and again. Scanning thousands of undocumented SQL objects with AI support produced a usable taxonomy in hours rather than weeks, but that is not “magic”. This process only works because the team committed to capturing the missing human context as it surfaced, rather than assuming the tool would supply it. 

The same logic applies to the "chat with your data" promise (text2SQL). These tools only work as well as the semantic layer underneath them. When a business user doesn't know how to read SQL, the model serves as the last line of defense against a wrong answer. If that model has already inherited inconsistencies and contradictions, wrong answers will abound. 

What next? 

This post only covers one part of the story: how AI is changing the way data models get built and why a solid semantic foundation matters more than ever. Xebia’s whitepaper Evolving Data Modeling for AI goes further, with: 

----

The whitepaper was developed by Ricardo Granados Lopez, with contributions from Jeziel Mateus de Abreu, Max Poirault, Camila Birocchi, and other Xebia experts.