From Systems of Record to Systems of Action: Agentic Enterprise Architecture on Databricks

Summary

Six pillars: Data plane, cognition/control plane, memory, gateway and governance support architecture.
Agentic data plane: AI data fabric, semantic ontology, and hierarchical memory enable reasoning.
Curated catalogs: Agents query curated data products with freshness SLAs and quality assertions, not raw data.

From Systems of Record to Systems of Action: Agentic Enterprise Architecture on Databricks

Talk by Steve Lueck, Chief Data Technology Officer, Old National Bank
An enterprise agent control plane is the layer that routes, governs and observes AI agents across the business — and it is what separates a system of record from a system of action. This video lays out an agentic enterprise architecture in six pillars: an agentic data plane with an AI data fabric and a semantic ontology, hierarchical short- and long-term agent memory, a cognition and control plane for multi-agent systems, an agentic gateway with a semantic router, and governance expressed as declarative policy code.
The talk describes the agentic data plane with an AI data fabric, multi-model management, a semantic ontology (building on Genie Ontology), hierarchical short- and long-term agent memory, and observability. It covers watch, assist and autonomous modes to build trust, a gateway with a semantic router, governance and observability planes, and policy as declarative code with token cost controls. Steve Lueck of Old National Bank then shares how the bank moved on-premises SQL Servers to Databricks, why agents should target curated consumer catalogs instead of raw data, and why data products need freshness SLAs, data quality assertions and Unity Catalog tags.
Sponsored by Infosys.

Chapters

FAQs

What are the six pillars of an agentic enterprise architecture?

This video outlines agentic data plane with AI data fabric and semantic ontology, hierarchical short- and long-term agent memory, a cognition and control plane for multi-agent systems, an agentic gateway with semantic router, and governance expressed as declarative policy code. These pillars work together to separate systems of record from systems of action.

Why shouldn't AI agents query raw data directly?

This video features Old National Bank's CTO explaining that agents should target curated consumer catalogs instead of raw data. This ensures agents work with clean, trusted data while protecting raw sources and improving reliability.

What data product requirements support production agents?

This video emphasizes that data products need freshness SLAs, data quality assertions, and Unity Catalog tags. This video shows that Old National Bank moved on-premises SQL Servers to Databricks to provide agent-ready data products.

How does semantic routing improve multi-agent systems?

The agentic gateway includes a semantic router that directs requests intelligently across multiple agents. This video describes the gateway as a control plane for managing multi-agent systems and observing their interactions.

Full transcript

`00:00` Introduction

Hello everyone, this is Rajan. So we will be presenting you with the path towards the agentic enterprise. So I will have with me Steve, the CTO for data from Old National Bank as well. So we will start with the what is the path, what are the readiness and the pillars and then we will have Steve explain in terms of how from his practical experience how they are accelerating the journey towards being an agentic enterprise.
Today morning I hope most of you attended the keynote. We talked about the agentic system of record and you heard from Ali Gaudzi itself. But one of the key thing we would like to highlight here is while we talk about the systems of action there is a fundamental change from deterministic to more predictive to more more towards you need a multiple decision loops multiple decision loops and me more towards planning and executing which is going to be very critical what we are seeing is if you really look at the what is the a real bottleneck in the model is mostly it is not the primarily around data readiness you could see second thing is the biggest challenge would be about the the how the models are degrading over a period of time and then the second aspect is only 6% of the enterprises are really able to build and redefine an end-to-end business workflow if you really look at what are the six core pillars that are really required for to be successful in this agentic journey.

`01:43` Six pillars of the agentic enterprise

One is how do you define the agentic data plane right it is not about Lake House it is not only about Lake House or agentic database how you define the agentic data plane. The second part is what is the what is the pillar that is really required for process and integrations right and then how do you reimagine the agentic experience and then the necessity of agentic gateway what is the role that agentic gateway is going to present and the last and critical pillar is about agentic governance i will be walking you through what are the critical recipes for each pillar so that that would be a big takeaway for for us from a practical standpoint when we talk about the agentic data plane there are five key modules into it right one is your AI data fabric where you are building this as a product not as a simple metallian architecture building as a product and making it MCP five second part is the multi-model management this is where you are going to bring in even fabrics as well as the multi-model data images pdfs it could be all the embeddings with the real-time view the third critical component we all talked about genie we all saw about genie ontology what we here we are talking about is a comprehensive semantic ontology covering value chain process data measurements and the system so you need all the five layers of it and the last critical aspect is the agentic memory where it has to be comprehensive not only talking about episodic or semantic it has to have both the long-term working memory and the short-term memory which is pretty hierarchical in the nature and the last and key part is the observability how you are going to build the observability from a data system process overall when it comes to process and integration most of the integrations in the enterprises are so far it is an API where it is an intelligent integration but a dumb end point now it is going to pivot more towards a smart processing and a dumb end point right so from that standpoint you need first and foremost moving to a cognition and control plane where you need to decouple the reasoning separate the reasoning and execution.

`04:09` Cognition and control plane for multi-agent systems

Second part is an agentic mesh and you saw Omni agent for code harnessing but you need a broader agentic fabric that needs to be treated as a process integration across what is there in your data ecosystem plus all other your enterprise ecosystem with a hybrid bus architecture that is not only as I said a smart not a not less smart endpoint or a dumb endpoint but how do you really create that the data grid, that intelligence grid for really getting the real time information.
When it comes to the agentic experience, what we have seen, it is not about a chatbot type of an experience. You need to have three different modes, a watch mode where the agent shadows the workflows, an assist mode and an autonomous mode. The critical part in this whole aspect is when you are really thinking about agent experiences where agent has to be treated as a worker and human has to be treated as a supervisor.

`05:05` Watch, assist and autonomous modes to build trust

So you have to reimagine beyond a chat interface or beyond a prompt-based interface, more suited for your workflows where you can split into, where you should give that three steps of trust to establish that experience. One from understanding the intent. second from refining the plan, third from giving them the end-to-end audit drive. The other critical aspect is the agentic gateway. We talked about, we saw in the morning about unified AI gateway.

`05:59` Agentic gateway with a semantic router

How this agentic gateway is going to be different is because you are going to have agents that are going to build in the platform or the agents that you are going to register from a package applications or agents that is being built by your partners, right? So you need to have this gateway, which is much different from your typical LLM or MCP, but this is more content towards an agentic orchestration, agentic management, agentic traction.
It requires four key components. One is the semantic router. So that's where you kind of able to route based on what agents, discoverability of the agents comes into play. Second critical aspect is your governance plane. so where you can actually apply context-driven policies. And then the third critical aspect is your tools and adapters that are exposed to the gateway. And the last key part is your observability plane.

`06:55` Agentic governance and policy as declarative code

Then the last critical aspect is governance. While people are talking about unity catalog, unity governance, unified governance, and you've been hearing about agency governance, but what are four critical areas that are going to be very critical for an agentic governance please don't think it like a passive governance it has to be infused into the brain of the agents it has to be introduced into the the core brain of the agents so that the agents that can operate either in the enterprise either in the cloud or either in your desktops are completely governed and managed very well.
And that's where you need to have ontology layer where you bring business process governance ontology, system ontology, data ontology, which is going to be very critical combined with policy as a declarative code which you can directly infuse into the agentic memory even if the agents are not getting interacted or filed or able to recover. it is still available point in time into the agentic memory and the last part is the Applying the grain to fine grain policies and that's where you can cross leverage things like unity catalog and the Are at the last other critical part is your cost how it is able to consume the tokens and the cost with that What we see is when you are really thinking about the blueprint this is going to be a multi-layered blueprint, and each one can be evolved where the data plane will be the core layer for your context and the knowledge, which is going to be your mode, with control plane, cognition plane, and the experience, serving the humans with both governance and the guardrails.
With that, I would love to invite Steve, CTO for data from Old National Bank, and to share his experience. Thank you, Steve. Thank you. Thanks for having me. Thanks, guys. I just want to quick do a little bit of a recap here of who we are and what we're doing here at Old National and kind of bring a little bit of practical perspective to what we're talking about here. Old National, we're a regional bank headquartered in Evansville, Indiana.

`09:05` Old National Bank: moving to Databricks

We have headquarters in Chicago and Minneapolis-St. Paul markets as well. We're over $70 billion in asset size. we're starting to get into that real bank territory where we can do some pretty cool and meaningful things. We've been on about an 18-month journey to convert all of our old data lake systems that were on-premise, SQL servers, et cetera, over into Databricks. We finally turned our last SQL server off in October of last, maybe November of last year.
And we're able to get everything moved over. 50-plus sources, we have a lot more kind of in process and coming, but we have about five sub teams that we have working on our data infrastructure and our data product strategies as we move forward. Obvious challenges that we have are data and silos. You know, the analysts are spending too much time still figuring out how do we plumb data. The coding technologies, the capabilities that we have have grown tremendously.
But our ability to get that context, the business decisions into the data is what really is kind of still our bottlenecks and what's holding us back. obviously we have regulatory pressure that's sort of eased a little bit here recently, but definitely still always is at the forefront of what we do. So I wanted to talk a little bit about our architecture and approach and when it comes specifically on the data side and that data plane that was referred to a little bit ago.
And what we're looking at this as is it's not necessarily anything different than what we've been building or driving towards. It's actually an evolution of what we've been doing. So we are taking this approach to say data product architecture and mentality is still at the core of what we do. I think this may hopefully is obvious, but I'm going to say it out loud anyway, is that if you put agents and we expose all of this data on raw data in our systems, we will exponentially create more problems than what we have today, right?

`10:52` Why AI agents should not query raw data

We will we will introduce infinite scale of 10,000 times the problems we have today and our data. So all the things that we are talking about is probably the things we have all strived to do or should have been doing all along in some regard, even for how we were doing human interactions and working with data and application interactions. So some of the approach we're looking at is that system level catalog? How do we break this into components so that we continue to enrich data, get it to be curated into ways that make meaningful decisions and allow the agents to make meaningful decisions, not have to keep reinventing the wheel?
If we put agents on raw data every time, that's too much for them to do at scale. All the reasons why we still have done these practices over the years of we only want to do those transformations once, we only want to do et cetera, we want to make that more reusable, applies, again, tenfold to what we're doing now. So we're breaking it into the concepts around we have system catalogs, which is where we bring in our raw data.

`12:13` Domain and consumer catalogs for agents

We do our data standardization. We get it all together. We have the next tier, which is our domain catalogs, which is customer credit, loans, deposits, how we bring those business things together that are more system agnostic. And then lastly, we have our line of business or consumption-based catalogs. And Unity has allowed us to start to build this capability out. And this applies, again, even more importantly as we start to get into agentic development.
The agents will target, from a consumer perspective, will target that consumer catalog layer. A couple of kind of key components here, though, is that we have to make sure we're tagging everything to the nth degree, right? the metadata of what we're doing and the documentation of how we're doing it becomes that much more important for the agents to be able to use this and consume from these concepts as we go about it.
This kind of double clicks into some of the alignment of human analysts versus how it works natively for AI agents and data products we're considering are a big part of our AI infrastructure from an agent perspective. We're also looking at ways to say how do we now also have agents even build the data products themselves and bring the curation to bear and have that business context and knowledge in addition? So again, contract schema, freshness SLAs, data quality, delta sharing endpoints, unity catalog tagging, all those things are things we're doing to make, we have been doing to make human analyst functions better. that is that much more important as we go for the agentic approach.

`13:41` Data products with freshness SLAs and quality assertions

So having that reliability guarantee, having pre-validated inputs that go into those agentic decisions, having the governance and mechanisms to say, what are the agents doing? How are they acting on behalf of this? How do we kind of pinpoint them to make that as deterministic as possible so that they're getting trusted results on the end of this? kind of want to close a little bit here in the next couple minutes around what are some of the first moves that you guys should be thinking about or making or how we're at least thinking about approaching it you really want to focus on how do we get a product out there with with kind of show what good looks like how do we how do we get something out there and experiment what what the right approach is close the gap with whether it's real time or if it's still batch That's OK, too.
But have the SLA, have those DQ assertions, have the right metadata captured and tagged so that you have the ultimate component. The next thing is start to think about your internal marketplace and how you're defining your Unity catalogs with the agent consumption in mind. It's going to change the way you philosophically build your catalogs. Having a catalog called Bronze, a catalog called Silver, catalog called Goad is not going to make sense in an agentic world.

`15:03` Designing Unity Catalog layers for agent consumption

It needs to be closer to the business use cases to the concepts that you're putting forward and align those to the functions that they're doing and think about the security and how that rolls back. And then lastly, just run a scope to agentic curation pilot. Do these things, experiment, try it. We aren't going to get this right the first time we do it. Keep trying, keep exploring, and I think we'll start to see this movement and really make some big progress with what we've done.
Thanks.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.