CUSTOMER Q&A: Databricks

How Databricks helps teams govern AI with speed and clarity

AI governance is often treated as a policy exercise. But as financial institutions move from experimentation to production, governance has to become part of how teams operate: who owns the data, how decisions get made, how models are monitored, and how companies prove what their AI systems are doing.

In this conversation from Plaid Effects, Suddu Seshadri, Head Data and AI at Plaid, sat down with Jennifer Miller, Global Head of Banking and Payments at Databricks, to discuss why governance is not a business detractor, but an accelerator—especially when organizations treat it as a product, codify it into the platform, and build clear ownership from the start.

This conversation has been edited for clarity and length.

Suddu: To start, tell us a bit about your role at Databricks and what the company does.

Miller: I lead banking and payments go-to-market globally. Before Databricks, I was a revenue line leader at a bank, so I find it funny that I’m doing a talk on governance! Five years ago, I thought governance was the business detractor; today, I view it as the accelerator.

Databricks is a data and AI company. We help customers unify data into a single view, which is critical for AI. That data doesn’t necessarily have to move into Databricks—we federate data from on-prem and other sources. We also govern data through Unity Catalog, which provides end-to-end governance for data and AI assets. Then we help customers democratize that data so business line users and functional leaders can make faster decisions, and build agentic AI and workflows on top of governed, unified data.

Suddu: AI governance can sound like a policy conversation. How do you build it so it gives teams clarity to execute?

Miller: We view governance through a five-pillar framework. The first is the operating model: How are you creating an AI- or governance-first culture? The second is regulatory requirements. Every industry has legal and compliance requirements, so how are you satisfying those? The third is responsible use: How do I make sure the data is being used responsibly and in service of consumers? The fourth is the technical tools that operationalize governance. And the last one is security: How do I keep the bad guys out and make sure data, AI, and agents are accessing data appropriately?

For us, that comes to life in Unity Catalog: a single pane of glass that governs both data and AI assets in one view.

Suddu: When you look across customers, who is getting governance right?

Miller: The patterns are clear. Companies getting governance right view it as a product. They invest in it, have a roadmap, clear owners, and a disciplined operating model bought in from the top down. They have one catalog that manages execution of the governance framework, and they measure governance publicly. They can describe how many data assets are appropriately cataloged and have a name attached. The companies getting it right view governance as an enabler, not a detractor of progress. When that cultural shift happens, and it’s clear who is responsible for each data domain, that’s when the magic happens.

Suddu: What does good governance look like in practice?

Miller: The most critical part is defining a clear operating model. Companies that are not doing this well have governance stashed in the back office. They’re the folks no one really wants to talk to. True governance means P&L owners, line-of-business executives, and functional owners are responsible for the data domain that supports their business. There is one human being attached to every domain who is responsible for governance.

Then you need the right catalog or toolset, and you need to codify governance within that tool. If your governance framework lives in a PDF, you’re already losing. Governance needs to be codified and applied automatically. Start with one domain and one or two use cases. Build a blueprint, execute it, and learn. The winners are picking one or two use cases and compounding the governance framework from there.

Suddu: What operating model helps teams scale governance without slowing everything down?

Miller: The fastest path to execution is a federated governance model. A central team defines the framework: How are we going to achieve governance across each pillar? But execution is federated to domain owners. In a bank, that might be the head of a consumer domain who sits in the line of business and is responsible for that data domain.

What slows progress is when governance is defined centrally and executed centrally. Then you work sequentially instead of in parallel. In a world where decisions are changing in hours, not weeks, you have to federate governance so agents or people have access to the right information as quickly as they need it.

Suddu: Once a governance framework is in place, how should companies think about evolving it as the business, technology, and use cases change?

Miller: The key is having a crisp framework that domain owners have to execute within. But you also need a thoughtful feedback loop. As the business or technology evolves, there has to be room to evolve the governance model. Without that feedback loop, governance feels rigid. You want domain owners to have high autonomy within the framework, and a culture where feedback is part of the equation so you can evolve and accelerate pace based on what you’re learning.

Suddu: What about when it comes to monitoring? Is a dashboard enough?

Miller: The feedback and monitoring framework can’t just be in a dashboard. If your monitoring framework is just a dashboard, you’re losing. Monitoring has to happen across levels. First is the data level: Who is monitoring data quality? Who is making sure pipelines are working as expected? How are you thinking about the data you need to feed and train your model? Second is the model level. Are you using the right model for the right use case? Not every model is right for every use case, so you need to monitor performance. Third is the business outcome. Is the model delivering what the business needs? Models drift. The data that feeds them drifts and changes. If the monitoring framework is not resulting in change, it’s failing.

Suddu: What should teams act on first?

Miller: To start, find one name to attach to your most important domain. Who is the one human being responsible for your most important data domain? If you can define that, you’re further ahead than most competitors. 

Next, ask whether your catalog covers both your data and your AI assets. In a world where agents are doing more on behalf of companies and people, it’s our responsibility to make sure those agents are reasoning over the right sets of data. It’s very difficult to do that if you’re governing agents in one catalog, models in another, and data in another. You need one single catalog that provides end-to-end lineage: Why did my agent make that decision, and what data did it use? If those assets are governed separately, you add risk and slow the AI agenda down.