Grounding AI Agents in Your Data Model
AI can already read and write your SQL and draft what a data model means. This talk is about what must be true for you to trust that draft enough to rely on it. We'll walk through a working pipeline, live, that grounds an AI agent's claims about your data in real evidence, and come away with a precise, testable answer for where that trust can be automatic and where it can't.
Every governance tool we already trust (ERDs, semantic layers, data catalogs, etc.) assumes someone already got the data model right before writing it down. That assumption is quietly breaking as AI starts drafting that first pass. This talk shows a citation-grounded pipeline, live: an agent drafts an ontology from a real dbt project, and a validator mechanically checks every claim it can, flagging rather than silently trusting the ones it can't.
Target audience: data engineers, analytics engineers, and ML/AI practitioners working with structured data pipelines. Basic SQL and data-modeling familiarity is assumed; no dbt or formal ontology/semantic-web background is needed. Especially relevant if you're already using AI coding agents in your own data work and have felt the "it forgot everything from last session" problem, or if you're evaluating how to bring AI agents into a data platform responsibly.
Audience takeaways:
A concrete, working pattern for citation-grounded AI-assisted documentation, adaptable to your own project
A precise mental model for what can be mechanically verified versus what always requires human judgment
A live demonstration of the pattern actually catching mistakes
An honest map of what this approach solves today, and what it deliberately doesn't
Scope: the approach, in four stages
Extract: pull structural facts from the project (e.g. a dbt manifest), zero AI involvement, deterministic and correct by construction
Draft: an AI agent proposes documentation/ontology claims, citing exactly where each one comes from
Validate: every citable claim is mechanically checked against the extracted facts; claims that can't be mechanically checked are flagged, never silently trusted
Signoff: a human explicitly confirms every flagged item before it counts as settled, tied to the exact claim text so a later edit invalidates a stale approval
Scope: what this does not solve
It doesn't automate grain or cardinality reasoning; those stay human judgment calls, permanently, by design, not a current limitation waiting to close
It doesn't check live data values at runtime (NULL drift, row-count anomalies, and similar); this operates at the documentation/metadata layer, not as a data-quality or observability tool
It's not a claim that AI can replace the architectural judgment good ontology design requires, only that AI can help draft and citation-check the structural layer faster, with a human still owning the design decisions
Outline (40 min content + 5 min Q&A):
0–5 min: The claims tools like ERDs and semantic layers make, and why AI complicates them
5–12 min: What it actually took to get an AI agent to draft a data model worth trusting
12–20 min: The mechanism: extract, draft, validate, signoff, live
20–28 min: Running the validator live, against a real project
28–35 min: Grounding a genuinely ambiguous table, live
35–40 min: Where this goes next, and what we're not claiming
40–45 min: Q&A
Ricardo Angel Granados Lopez is an analytics engineer specializing in data engineering and analysis. With a master's in IT management and a focus on data science, he is proficient in using various programming languages and tools. Ricardo is skilled in exploring efficient alternatives and has contributed to multicultural teams, creating business value with data products using modern data stack solutions. As an analytics engineer, he helps companies enhance data value through data modeling, best practices, task automation, and data quality improvement. His current focus is on grounding AI agents in real data models, making sure what they draft can actually be trusted, not just plausible.