Databricks Data + AI Summit 2026 | Bryan Clark, Databricks & Grant Veazey, Ensemble Health

Clips
More from Databricks Data + AI Summit 2026

Bryan Clark

Director of Product Management

Databricks

Grant Veazey

CTO

Ensemble Health

play_circle_outline LTAP vs HTAP: Single-Copy Architectures and Row-to-Column Translation for Unified Transactions, Analytics, Real-Time Query Engines

play_circle_outline From Custom Pipelines to Agent-Driven Workflows: Ensemble Health’s Clinical Appeal Agent Cuts Turnaround From Weeks to Days

play_circle_outline Observability, traceability, and human-in-the-loop in regulated healthcare agent deployments

play_circle_outline Branching in Lakehouse enables isolated experiments and safer agent development

Info
Transcript

Bryan Clark, Databricks & Grant Veazey, Ensemble Health

Bryan Clark

Director of Product Management Databricks

Grant Veazey

CTO Ensemble Health

In this interview from Databricks Data + AI Summit 2026, Bryan Clark, director of product management at Databricks, joins Grant Veazey, chief technology officer and enterprise performance manager at Ensemble Health Partners, to talk with theCUBE's John Furrier about how unified data architecture is enabling the shift from AI pilots to production-grade agentic applications in healthcare. Clark details Lakebase, Databricks' approach to consolidating transactional and analytical workloads into a single-copy architecture that eliminates brittle data pipelines. By... Read more

explore Keep Exploring

- What is the key to LTAP? - (Also asked in the excerpt) What are the trade-offs in consolidating OLAP and online transaction processing, and how should CIOs/CTOs think about that? add

What was the timeframe (ballpark) to build the clinical appeal agent and its data pipelines (how long did each pipeline and the overall process take)? add

What is the view on agents, and how are agent deployments being evaluated? add

At what point does the system become non-deterministic and generative? add

bolt Powered by CUBE AI

Bryan Clark, Databricks & Grant Veazey, Ensemble Health

search

John Furrier

>> Welcome back to theCUBE's live coverage here at the Moscone Center in San Francisco, Databricks' Data and AI Summit. I think theCUBE has covered every single event, Databricks, following it from inception through now mixturization and global expansion. AI intelligence is injecting into the business. The AI infrastructure is building out super-fast with super computing. Brings in higher performance and opportunities we haven't seen. We've got two great guests here to unpack what's going on at the data layer that unifies transactions, analytics, real-time systems into one, so that agents can perform their best and physical AI. Whatever comes next. We got Bryan Clark who's the director of product management at Databricks. Great to see you. Thanks for coming on.

Bryan Clark

>> Thank you for having me.

John Furrier

>> We've got the big product on stage, LTAP. And Grant Veazey, CTO of Ensemble Health. You guys are working hard to make sure everyone has what they need. Healthcare is huge. It's one of those use cases that pops with AI.

Grant Veazey

>> Absolutely.

John Furrier

>> So, looking forward to talking to you. Bryan, we'll start with you.

Bryan Clark

>> Yeah.

John Furrier

>> LTAP transactional center stage really is the final piece of the puzzle on that unification. There's probably some other things, but in terms of the big things, it's I think a really strong piece for Databricks to complete. Even Ollie's like, in the old days, we had Spark, data engineering, and you got warehouses in real time, now transactions come together. This is not an easy thing for an enterprise to grapple with. They were traditionally departments, technologies. What is the key to LTAP?

Bryan Clark

>> Well, it's the single copy. The difficulty before was managing multiple copies of databases, brittle pipelines, trying to understand how you're moving data back and forth. Those are the things that the LTAP really solves. The difficult part on our side is that we're now translating row-based format, this Postgres format directly into a columnar format. And so, there's a bit of work there, but it's deterministic. That work is solvable and then it creates this system where now you have an open format. Customers have all existing tools they can use to read that and do things that they've always done.

John Furrier

>> What was the trade-offs to make that consolidation happen? And how should people think about it? And how should CIOs and CTOs look at OLAP and online transaction processing? What is the trade-offs and how should people think about this?

Bryan Clark

>> Yeah. Well, instead of something like an HTAP, where you're trying to use the same query engine to do different jobs. And really, your trade-off is now it doesn't work well on certain things. You're using different query engines. So, customers have to say like, "I'm going to use my analytics query engine for analytics data and then I'm going to use the Postgres query engine for the operational data." And they have to understand that upfront, but you get the best of both worlds in that case.

John Furrier

>> Right. Talk about your environment, Ensemble Health. What do you oversee? What's your purview and how do you look at this? Because I mean, you have a lot of legacy, modern legacy, and current modern stuff happening because AI is a perfect storm of innovation for healthcare.

Grant Veazey

>> Yeah, absolutely. We were lucky enough to be partners at Databricks for over three years. But like a lot of people, we spent a lot of time building out a data lake platform to be able to do analytics. And now with the age of agents, we've gotten the place where we need to get that same level of data to those agents in a much faster way, to be able to start to do those things, like reasoning and actually take action, instead of that analytics. So, while the core foundation was great for us to be able to do analytics across the board, now we needed the ability to do something that was more real time, instead of making copies into transactional databases creating custom pipelines, every time we changed what the agent needed or how we wanted to reason about it. So, it really has been a huge use case for us to be able to take the huge data asset we have in healthcare and take that to action to drive care.

John Furrier

>> I want to pause for a second. Explain that, because you just laid out like, "Oh, my God, the old way." So, lay out the old way.

Grant Veazey

>> Yeah, absolutely.

John Furrier

>> And what's the new way to look like? And just scope the pain and suffering slog fest you got to take through and then what that turned into.

Grant Veazey

>> Right. So, a perfect use case for us was building a clinical appeal agent, in which we were going to be able to take information that we'd curated through our data lake to be able to reason on what was the clinical justification. You have an argument between a healthcare provider and insurance company on whether this was clinically necessary or not. So, we built a clinical reasoning agent to do that. What we would have to do is what are the data elements we needed to get to that agent to make that determination. Some of it was external, some of it was internal. We would have to figure out all those data elements. We would create a custom pipeline and create a transactional data store, in which that agent could consume that data. And then once we tweaked it and went through hill climbing and like, oh, it's doing pretty good, and now we think we need a couple more pieces of data. Let's go back to the lake. Let's create another custom pipeline and let's do that all over again. And went through five iterations.

John Furrier

>> phase. Okay. Just what's the timeframe on that? Because what was the duration of that? Just estimate, ballpark.

Grant Veazey

>> Probably each pipeline was a week or two, three or four iterations, a month or two.

John Furrier

>> And what's that now in terms of timing? Agents are doing it for you on the fly, coding agents happening.

Grant Veazey

>> Hours.

John Furrier

>> Hours?

Grant Veazey

>> Hours and days.

John Furrier

>> All right. So, weeks, maybe a month?

Grant Veazey

>> Yeah.

Bryan Clark

>> And that's just compounding the problem of like the brittle pipelines.

Grant Veazey

>> Absolutely.

Bryan Clark

>> If agents are moving different pieces within there, your web is starting to break really fast.

John Furrier

>> The MCP servers mentioned a lot up there. How does agents actually perform? Actually, with some really candid demos around, if you don't have the context, find those authoritative snippets. What's your view on agents? How are you evaluating your agent deployments? Because basically, Databricks was very straight up saying, "Hey, we're not ready for prime time in a lot of situations here."

Grant Veazey

>> Right.

John Furrier

>> I mean, that's honest.

Grant Veazey

>> Absolutely. That's one of the things we have to be very careful. We're in a very highly regulated arena and it's very important that we have that accuracy. So, in a lot of places, we're still leaving human in the loop from that perspective. But we're spending a lot of time on observability and traceability of those agents. And that's helping as well, is that we can take those transactions of what we're seeing, write those back to that same Lakebase architecture. Have that consumed right back into the lake structure, so we can build out that observability and traceability.

John Furrier

>> Right. I mean, I guess a little bit hyperbole saying it's not ready for prime time. If hype was the meter, it would totally be prime time, but it is a serious consideration for enterprise. The coding momentum from last year, going into this year and even halfway through this year, I think the enterprises go, I get it. This is not rag search and kind of good marketing copy. That's real productivity. What's your view on how the agents will perform now that you've got unification? What's the sequence of events that move it from enthusiasm to maybe confident workflows that you can bet on? When does it get non-deterministic and generative?

Bryan Clark

>> Yeah. That's a good question. I mean, our view has been Lakebase has branching built in. And the reason we built that was for developers to be able to work like they work and get... You create a branch, you have an isolated environment, you can try things. But now that agents are accelerating development, to us, that's the piece. I don't know that it's going to become less non-deterministic than it is now, as much as I'm going to run more agents. They're going to be more powerful. But you're still going to run them in an experimental fashion. You're going to say, "Hey, here's a copy of this data. Maybe I've removed all the PII and all the other information I wouldn't want an agent to be able to run through." But then once they run through that experiment on that isolated environment, you can then have a supervising agent review and get back to the human in the loop.

John Furrier

>> Grant, talk about the organizational challenges, because now you have agents coming, you got to look at the architecture. How do you view that change management piece? Because obviously, you got the lake house and the data lakes and all the good technology. What is the challenge as a CTO? It used to be called herding cast. Now you're going to be herding agents. How do you look at the personnel organizational impact as a leader?

Grant Veazey

>> Yeah. It's a paradigm shift completely of starting to look agent first. I think the technology allows us to do that. But what we're talking about is we have to stop looking at data as analytics, and looking at data as bits and pieces of behavioral data to drive action and reasoning. And so, it comes back to, again, making sure that you have good ontologies, which we're excited about some of the things that were announced today on that. Good ontologies, good observability, and really looking as an agentic first approach, but change is difficult.

John Furrier

>> Did you have a hand in the ontology product on the front end?

Grant Veazey

>> We did not, but we're excited to get our hands on them.

John Furrier

>> Okay. What are you going to do with it? What's going to be the action step there? Because ontology's open up a lot of doors for software to do things.

Grant Veazey

>> Absolutely. For healthcare, I think the ontology is paramount. We've worked on a lot of things internal to build out that ontology ourselves already to this day. But understanding that relationship between denial codes, clinical documentation, blocking cart codes from payers, understand how all that comes together, is the fuel for the AI that we see in the future.

John Furrier

>> Bryan, I have a joke of my kids when they were younger. "I used to walk to school, but bare feet in the snow." I mean, ontologies used to have to be hard coded by hand back in the old glory days, not to date myself. But now with automation and some of the intelligence, what does this open up for customers? And how would you explain ontology today, and what it does, what it is? What is an ontology today? What does it mean? How should people think about it?

Bryan Clark

>> Yeah. I mean, it's really I think about it like agent dreaming. I think that's the easiest way to think about it is like at night, in the background, the agent is kind of like pulling the different pieces and building this graph. And reproducing the memory of what exists in your data plane. Because like Ali said, either every time you talk to it, it's going to try and spider out its way and spend a bunch of time looking at different docs. Or in the background, it had figured this out and built a memory map of where things are. And then it can just go directly to where the information is instead of trying to reason about it from first principles every time.

John Furrier

>> And scaling is real?

Bryan Clark

>> Yeah.

John Furrier

>> So, ontologies have scale and does a lot of prep work. How do you see that envisioning for you guys on the agent side and the human in the loop piece?

Grant Veazey

>> We have a ton of unstructured data in healthcare as well. So, the first step of going from unstructured data to that structured data and then being able to build those relationships in that ontology is one of the things that's paramount for us to be able to understand what are those relationships. And healthcare is not as consistent as we'd all love it to be. And where we can find those patterns in ontology in a more dynamic way, and end up coding them all by hand, and turning that up to opinions. And specifically, we talk about opinions of doctors or versus insurance companies.

John Furrier

>> You can't recognize the handwriting.

Grant Veazey

>> Exactly. Exactly.

John Furrier

>> Computer vision doesn't work well on that. I mean, this just brings up a really good point about friction. In healthcare and all regulated industries I've covered, it's always been... but they've done a lot of data work. There's a lot of prep, tagging. In a way, they did all the right things for the wrong reasons. HIPAA was a disaster, but now you have all that data tagged and then combined with the unstructured. So, how do you see that friction being taken out, abstraction layers? How do you holistically look at leveraging the asset of the data estate, the legacy? And what's your innovation strategy to make the friction come out of that so that innovation can spawn?

Grant Veazey

>> I think there's still a bit of an interoperability problem completely across healthcare. But I think some of the advances we're seeing are going to allow us to possibly step over those or make those easy to bridge as we go forward. And again, a lot of what we're seeing from the data, if we don't have the interoperability, we're able to reverse engineer with all these pieces and parts. And basically, have the agents understand what is happening on the other side without that interoperability.

John Furrier

>> All right. Most important question I get asked all the time is I'm really struggling... This is the preamble. I'm really struggling with pilots to production. Recently, we've seen headway around accelerating production workloads with generative AI and now some agents that have structured workflows. What's your experience on your production strategy? How do you think about moving stuff from pilot, sandboxing, however you want to call it, into productions? Is it a narrow road? Is it getting bigger? What's your thoughts on getting stuff into production at scale?

Grant Veazey

>> Yeah. I think the more you do it, the faster it gets. But when we look at our use cases, again, we try to look at... Human in the loop's a big part of that. The ability to hill climb, starting with the standards of what we want to be able to do from a workflow and get out of that, is start small. And lots of times it may be don't try to eat the entire elephant at one time. Pick out a subsection or a sub process of that,. See if I can start to get real results in that and then start to build from there. It also builds the confidence. In our industry, human in the loop is huge. You get those doctors, those nurses, they want to have hands on making sure that that's important. And getting that fee loop back. And then once they start to trust it, it makes it a lot easier.

John Furrier

>> You must get a lot of laundry list items when, "Hey, what we can do with AI?" Innovate with AI. I mean, everyone I talk to at a CTO or CIO level, it's like, yeah, the list is miles long. So, prioritization becomes a big thing. So, I want to ask you your philosophy on this. We've heard different schools of thought on prioritizing AI projects because it's not an IT project.

Grant Veazey

>> Absolutely.

John Furrier

>> It's a business transformation model transition. So, people are saying, "Whoa, whoa, I know where the revenue is. I know where my problems were that I couldn't get to with non-super computing technology." So, the lens shifts with the new conditions. What's your thoughts on how you scope and look at and identify and prioritize?

Grant Veazey

>> Yeah. At Ensemble, we maybe take a little different approach than I hear a lot of people take it. It's a whole lot less about trying to get cost savings for us and automation than it is to be able to solve problems we weren't solving before. So, if you take a denial for an exempt, if our final denial rate is still 3%, how do we use this to shrink that down to 1% or closer to 0%, and drive more and more revenue for our clients and the providers? And so, we look through that lens a lot of how we prioritize what are the things that we're able to solve problems that we couldn't solve before.

John Furrier

>> Yeah. I mean, you know what the problems are.

Grant Veazey

>> Right. We know what the problems are.

John Furrier

>> And they have impact on people's lives. Everyone's having a story where they didn't get approved and surgery or this didn't happen. So, I really appreciate that and congratulations on that. Bryan, final question for you is, next 12 to 18 months, what are you optimizing for? What can we see from you and your team on the product side?

Bryan Clark

>> Yeah. Well, I don't know if you saw, but we have cross cloud disaster recovery. So, we've built out the platform in order to move data to different regions, different clouds. So, beyond just cross cloud disaster recovery, we'll be able to have replicas in different regions. You'll be able to switch over. And one of the things we were talking about earlier is how I now have customers who are actually using this to hunt down cheap GPU. Something that was built for disaster recovery to say, "Well, when a cloud is down, I want to move from this cloud to this cloud." But now instead of that, customers are saying, "I want to pick up my data and push it over here because GPUs are cheaper over there. So, can you move my data over to that thing?" So, that's the next, I mean...

John Furrier

>> I mean, I'm running out of time. We didn't even have a chance to talk about cyber resilience. Just in a nutshell, built in from day one, what's your thoughts on cyber resilience?

Grant Veazey

>> Yeah. For us in the regulated energy, security cyber is always a first class citizen, first nonfunctional requirement of everything we do.

John Furrier

>> And on your side, Panther?

Bryan Clark

>> Yeah. Same thing. I mean, integrating everything with that so that agents can be assisting in preventing intrusion.

John Furrier

>> And Panther is in your future-

Bryan Clark

>> Yeah. Yeah. Sure....

John Furrier

>> on the acquisition side? Thanks for coming out, guys. Thanks so much. I'm John Furrier, host of theCUBE, breaking it down. A lot of technology coming together. Again, transactional processing runs the business. Analytics tells you what it is. All's happening in real time around all the first principles of data hygiene and scale. Databricks is bringing that to the... And of course, the AI infrastructure is booming. More and more horsepower allowing you to solve problems that could not have been solved before, new opportunities for entrepreneurs and enterprises. Thanks for watching theCUBE. We'll be right back after this short break.