Speaking of Data

The episode surfaces a handful of genuine practitioner lessons - don't let LLMs define your business glossary, build the semantic layer before you need it, use small SME groups - but they're buried under heavy conversational filler, repeated affirmations, and obvious advice about executive sponsorship and data quality. The insight-to-minute ratio is low.

“in one point in time we asked LLMs to define things for us. Yep, big mistake. Don't do that. You have internal knowledge. Right. Use it.”

“don't try to use the term data fabric just because this is. Oh, everyone is talking about that. Uh, uh, no, you need to understand exactly what is a, uh, data fabric.”

The content is almost entirely standard data governance and semantic layer orthodoxy - methodology first, single source of truth, executive sponsorship. The only mildly contrarian point is the warning against using LLMs to auto-generate business definitions, but even that is a known pitfall discussed widely in the community.

“in one point in time we asked LLMs to define things for us. Yep, big mistake.”

“don't think that this is not important. You must have a common language internally.”

Celso is a genuine practitioner - Global Director of Business Reporting Governance at McCain Foods with 25 years of analytics experience - who has actually implemented a data lakehouse with a semantic layer and business glossary. He is a credible operator, though not a high-profile industry figure or C-suite leader, and some answers remain vague about scale and impact.

“My most recent role is exactly, uh, connecting the dots between the two sides. So I'm very, I have very deep knowledge in terms of technology”

“we created agents to support people. Our, my team at that time, uh, to do the business analyst tasks, data engineering tasks, tests. Right. So the DBT tests.”

The guest references McCain Foods, a data lakehouse architecture, DBT tests, and a business glossary rollout process, but no metrics, timelines, team sizes, or tool names are provided. The one data point cited ('70% of people said the semantic layer was critical') comes from the host reading a TDWI report, not the guest. Most claims are anecdotal and vague.

“in a recent report it was like 70% of people said the semantic layer was critical. A unified semantic layer was critical to success with AI.”

“we had an opportunity to select one tool. Uh, I would say that it would be the 1 visualization tool, but that makes the semantic layer as well.”

The hosts repeatedly respond with 'great question' and 'that's really great' without genuine follow-up, never push back on vague claims, and Megan frequently digresses into her personal enthusiasm about AI agents rather than extracting deeper practitioner knowledge. Questions are predictable topic-transition prompts rather than incisive probes.

“Great, great question. Because you know that I would be lying if I said that we did everything correct.”

“I myself have used you know, natural uh, natural language interface to query like well structured data and have it deliver insights and at the most basic level.”