Overview
Large language models are changing how people interact with company data, and that shift puts new pressure on data governance. In this episode of The Dashboard Effect, Kate Eberle and Brick Thompson discuss what it takes to govern data responsibly when conversational analytics sit on top of a modern data lakehouse, from the committees that set direction to the permissions that control what an LLM can see.
Their conversation makes a clear point: AI does not reduce the need for governance, it raises the stakes. Organizations that want reliable answers from an LLM need curated data, rich business context, and security that is configured deliberately rather than assumed. See how Blue Margin’s Managed Data Platform helps organizations build a governed foundation for AI and conversational analytics.
What This Episode Covers
Two Lanes of Governance and the Role of Committees (2:55 – 4:40)
Kate and Brick frame governance as two parallel lanes: an organizational layer made up of committees and policies, and a technical layer made up of security and permissions. Both are necessary, and neither works well without the other. The organizational lane often takes shape as a data governance committee of engaged data users and managers who review reports and align on data strategy. This group keeps the data program connected to how the business actually operates.
Training for Self-Service (4:40 – 6:17)
Self-service analytics only works when users are trained well. The speakers explain that knowing how to click through Power BI is not enough. Users also need to understand the schema behind the data and how to validate the results they get. Without that foundation, self-service can produce confident conclusions built on misread data, a risk that grows when an LLM is generating the answers.
Context and Curation for Reliable LLM Answers (6:51 – 9:07)
This section covers the work a data team must do before an LLM can be trusted with business questions. That work starts with curating a platinum layer of data that is clean, modeled, and ready for consumption. It continues with giving the LLM rich context, such as markdown files that define business metrics, KPIs, and the rules for querying them. The quality of the answers depends heavily on the quality of this preparation.
Access Control in the Lakehouse (10:09 – 11:26)
In a single source system, permissions typically pass through to the user automatically. A lakehouse works differently, and the speakers explain why that matters for AI. Because an LLM can potentially reach any table a user is permitted to see, role based security has to be configured explicitly and thoughtfully. This is where the technical lane of governance becomes essential to protecting sensitive data.
Keeping Stakeholders Engaged (12:33 – 13:41)
Governance depends on participation from people who already have full schedules. Kate and Brick acknowledge this as a common challenge and offer a practical response: use adoption metrics to build a strong ROI case. When stakeholders can see how data informed decisions are improving outcomes, the time they invest in governance becomes easier to justify.
Who It’s For
This episode is worth your time if you are a data or analytics leader preparing to introduce LLMs into your reporting environment, an IT leader responsible for security in a lakehouse architecture, an executive sponsoring a data governance initiative, or anyone who wants AI generated answers they can trust.
Why It’s Worth a Listen
The most useful idea in this episode is the distinction between the organizational and technical lanes of governance. Many teams focus heavily on one and neglect the other, either building policies that are never enforced in the platform or locking down permissions without the committee alignment that gives them meaning. Seeing both lanes side by side makes it easier to spot where your own program has gaps.
The discussion of context and curation is equally practical. Kate and Brick make it clear that an LLM is only as reliable as the data and definitions it is given, and that preparing a platinum layer and documenting metrics is real work the data team has to own. Their point about lakehouse permissions adds an important caution: an LLM will use whatever access a user has, so security cannot be an afterthought.
For organizations moving toward conversational analytics, this episode offers a realistic picture of the groundwork involved. The teams that invest in governance now, on both the people side and the platform side, will be the ones whose AI tools earn lasting trust from the business.
