Name: Shaun Clowes, IBM
Uploaded: 2026-10-09T17:00:00.000Z
Duration: 19 min 55 s

Clips
More from IBM's AI Operating Model Series

Shaun Clowes

GM, Confluent and Data

IBM

play_circle_outline AI shift: LLMs rely on continuous, contextual data compared to the cloud shift

play_circle_outline Streamhouse: Overcoming Data Gravity with Real-Time Unified Business State for LLMs

play_circle_outline System of intelligence/digital twin: harmonize people, places, processes, assets in real time

play_circle_outline Avoid islands of automation with federated state, guardrails, and enterprise operating model

18:45

22:47

Info
Transcript

Shaun Clowes, IBM

Shaun Clowes

GM, Confluent and Data IBM

This discussion examines artificial intelligence, including AI operating models, streamhouse architecture and enterprise agents for enterprise data management. It explores how emerging agent architectures reshape priorities for data delivery governance and workload design.

Shaun Clowes of IBM Data Products joins theCUBE Research hosts John Furrier and Dave Vellante to explain how large language model, LLM-driven agents change enterprise data priorities. Clowes draws on extensive industry experience to describe the streamhouse approach, addressing data ... Read more

explore Keep Exploring

How should one understand the "traditional enterprise" in the context of large language models and the changing role of data? add

How is data gravity creating a scarcity of "the right data in the right place" for LLM-based agents, why does that make an AI operating model necessary, and what architecture (e.g., a "streamhouse") addresses this need? add

How should an organization approach building a real-time "system of intelligence" or digital twin that provides context (what happened, why, and what will happen)—does this need to be centralized, and how do streaming, data enrichment/governance, and platforms like Streamhouse fit into that? add

How should a chief data officer or CIO prevent "islands of automation" when deploying AI (including autonomous agents) across an organization? add

bolt Powered by CUBE AI

Shaun Clowes, IBM

(INTRO)

John Furrier

>> Hello, I'm John Furrier, host of theCUBE, here with Dave Vellante for a special CUBE Conversation here at IBM's flagship One Madison Avenue, heart of New York City, part of the big Analyst Day. Shaun Clowes is here, GM of IBM Data Products. Long history, kind of like Dave and I, historian in the industry, seen all the data movements over the years. Shaun, great to see you again in person.

>> Fantastic, great to be here.

John Furrier

>> Yeah, you had a great talk. I think you stole the show. I love the Streamhouse 'cause it's consistent with our theme. That we've been kind of putting out there that we're in an always-on data world now. It's not stopping. And when I say always on, it's when I'm sleeping, there are agents working on my behalf.

Dave Vellante

>> 100%.

John Furrier

>> The headless system trend points to this new kind of abstraction/forming interface, personalization, and a slew of benefits on this shift. But you had a really interesting point. The shift has happened. How would you describe the AI shift and how does that compare to cloud shift? Traditional enterprise?

>> Yeah, one way to think about this is, before we always used to have effectively data was, people used to talk about data as the new oil, but what they really meant was data in some sense existed and was used by humans to make good decisions. And businesses were run by humans, and humans, that's how you win. But actually now, in a very literal sense, an LLM is like a smart brain in a computer, but it is limited entirely by what it can see about the world. In fact, data is quite literally the electrons that feed into the LLMs, and without them, nothing good happens. And those LLMs never rest. Like you just said, all night, every night, they're trying to make decisions, but they can't do anything unless they're given the information.

John Furrier

>> You know, we've covered— when we started 17 years ago, the big trend was Hadoop, and that was big data back in the day. Maybe oil was a great example. But now data gravity— the data's abundant, but what's scarce is getting the right data at the right place in the system. And you guys call it an AI operating model, which means it's kind of an operating system. More systematic, less search result. How is that playing out? Because now you have a scarce resource, you have a strategy there, and if you don't do it, you have fragmented datasets, there's a tax on that. Talk about this dynamic about why this piece is so important to the operating model.

>> Yeah, the Hadoop world that you were talking about, the big data world, all came around when suddenly storage and compute became abundant in the cloud days. Everyone's like, well, I know data must be valuable. Maybe if I collect all of it, somehow someday I will get useful data out of it. But no, I need the data out of it right now. And in order for an agent to be successful, I need it across all of my different systems. I need it to know about my frequent flyer status. I need it to know about my flights, my baggage, my hire car. I need all that information together right now to make good decisions. And so in order to do that, you're gonna need something different. the architecture that was created when cloud became a thing, Hadoop, then we had lakehouses because we had storage become abundant in S3. Now we need a new type of architecture, a type of architecture that can bring together all the state of our business right now and present that directly to LLMs. So we call that the streamhouse architecture.

John Furrier

>> You mentioned brain earlier. I love that analogy because we have, you can have heterogeneous compartments in your brain, a warehouse, a lakehouse, a streamhouse, you have pathways. I put my hand on the stove, not a lot of reasoning going on. Mm-hmm. And that is basically full muscle memory.

>> Yep.

John Furrier

>> So memory again, back into how should people think about this architecture? Because that's a heterogeneous data platform.

>> Yep.

John Furrier

>> This is kind of like, not a new concept, but now in full-blown AI. Yep. Speed of data. If you don't get the data in, it can't make that decision. Yeah. Actions will be not as good unless they have the best data. So I find that to be a very scarce resource.

>> Yeah, exactly right. It's about the speed of the data, but it's also about the accuracy of the data and it's actually about the right data. It doesn't matter if you get the wrong data at high speed to the LLM. It also doesn't matter if you get the right data but far too slow because it will have already made the decision. So you need all those things together. It's funny because people like to think of LLMs as, well, if I could throw more data at it, would that make the result better? But you wouldn't give your accounts receivable work to a nuclear physicist that worked on your energy production side of your business. So why would you take a very, very smart model, give it all of the possible information about your business and say, figure out accounts receivable? It just doesn't make sense. So you need the right context for the job. And people are bandying around this concept of context. What is context? Well, I like to think it's just the information you need to make a good decision, a fast, good, high-quality decision.

Dave Vellante

>> So let's talk a little bit about how you get there. You guys talking about Hadoop, no schema on write, just a bunch of, just a mess.

John Furrier

>> Hadoop failure brought in Databricks, made them successful.

Dave Vellante

>> But yes, exactly. And so we now had access to this data. It was relatively inexpensive to the point. Streamhouse is interesting. I think streaming, I think Confluent, I think the gold standard in maturity and reliability and breadth of capabilities. But context requires more than just real-time access. And I want to understand that. The North Star, when we talk to customers, is we want to build a digital twin of our enterprise and be able to understand the people, the places, the things, the assets, the processes in real time and make decisions. Yes. What happened? Why did it happen? What's going to happen next? And that is the context that's necessary. So how does that all come together? We use the term system of intelligence, which is a combination of the front end, what the client and the users are doing, and the back end, which is harmonizing all this data and understanding the processes and the tacit knowledge of the enterprise. That's a big nut that I just dumped on you. How do you guys approach that?

>> Well, I think the answer is it doesn't have to be centralized, right? You're right. It's not just about streaming in real-time data. It's about being able to take the data, connect it to all the systems, to be able to join it, enrich it, make it rich, to be able to govern it, to convert it from raw data into big data products or things you can reason about, and then to be able to bring that together in order to achieve a context. Our argument would be that you can actually move all of that very, very close to where the data is first produced, as close as you possibly can. So the earlier you govern it, the better. The quicker you produce the high-value data, the better. Then you need to just connect up to all the relevant places. So you can think of this whole thing as streaming plus plus plus, and the Streamhouse adds all those other capabilities I talked about to deliver on the ability to deliver the state of all your business right now.

Dave Vellante

>> It's okay, so the plus, obviously governance is there. So I get, you got the streaming, what are the plus plus plus? Governance?

>> Yes, you've got streaming, you've got governance, you've got connectivity. So the connectors to all the systems, you've got processing, So the ability to enrich the data, join the data, produce aggregates, et cetera. And then you have the ability to query the data. So serve the data to different places. So for example, you might serve the data directly in MCP to an agent. So it can literally say, what were your 5 most recent orders? What are your current flights? What is your rental car? Where is your rental car? Your agent can immediately know about all of those things. You don't have to tell it, it will just know that about you when it's trying to solve the problem.

Dave Vellante

>> And that enrichment, step that you just mentioned, that's where the data harmonization comes in. So I know when I ask for revenue, it's not NRR or ARR or annual revenue or calendar year revenue. That's where that occurs. And I don't need an army of folks running around in lab coats to really produce that.

>> Yeah, yeah. in general, you can push down this idea of, I guess, governed context or bounded context. So you can describe, hey, in this business process, what I mean is X. And then you can have these governed metrics and these governed datasets that effectively represent the foundation upon which the model will reason. One way to think about it is the LLM is a smart brain. It can see only the bits of the world you give it. So you better be very careful about what you give it for the good and the better and the worst, right? You want to give it the best possible data and also not the incorrect data or the wrong data or the wrong places.

Dave Vellante

>> And how much process knowledge gets into that learning loop? How does that all work?

>> You mean like the way the business processes work? Yes. Yeah, yeah. In general, you have two ways to do that. You can either explicitly describe that. So for example, in skills as part of the context, you can describe what you want it to do or you can give it a source of documents. So inside IBM's data platform, you could be taking all your unstructured documents, the ones related to that business process, and then you put that in as the context to that agent. So one way to think about the recipe for an agent is a set of instructions that you give it and a goal that you give it. The agent, the LLM itself, and then the contextual data in terms of unstructured data and structured data like these things I'm talking about, your bookings, your reservations, your everything. Once you have all of those things together, you have a world in which the LLM is realistically gonna be able to succeed at solving your problems.

John Furrier

>> You know, last week we were streaming 3 days at CoreWeave's first annual conference they've ever done. In fact, it was the first NeoCloud to actually do that. A lot of smart people, they're tight with NVIDIA. You guys work with them. We actually interviewed, IBM Research would tell some great stories. But when you start having this intelligence production system, whether it's CoreWeave, NVIDIA, however the intelligence is produced, and inject that intelligence, it changes things. One thing that we saw clearly was this idea that workloads are going to become intelligent. That's specialized intelligence, not general intelligence. The Frontier guys, they search the internet. Good for them. Generalize, they want to do tax accounting law, they probably have a lot of that law and all that stuff on the books. But when they come in and say, hey, how do you do it? This brings up the domain-specific features. It's being called specialized intelligence. There's a lot of different words for it. That's where the market's going, where agents— so when you see agents, you talk about workloads. What do intelligent workloads mean? How would you frame that? How should people think about it? How would you tell someone, you have a workload, you can make it intelligent?

>> Yeah, yeah. describe that. When this all first happened, when the LLM phase started happening, everyone was all very excited by RAG, for example, this idea, I will just index all of my organization's data and then I'll give it to an agent and they

John Furrier

>> can—It's not bad, it's good, but it's not the

>> answer.It's a cool idea, but it's not the answer. Like I was saying earlier, you wouldn't take a nuclear physicist and have them work on accounts receivable. So there's something here about the right size for the right job, the right intelligence for the job, the right knowledge for the right job, the right harness for the right job. And in fact, that's where the world's going. Have you noticed we've got— we started with agents, Now we have supervising agents, then we have sub-agents, now we have sub-sub-agents. That specialisation mimics what happens in the real world, right? Even inside your brain, there are various parts of your brain that are used for various different types of specialisation. So I would say that it's inevitable that we end up with workloads and workloads end up looking a bit like tasks. They're either business tasks or business processes. And as one big system, you'll be able to evolve, improve these individual pieces of the puzzle, make them better and better and better., and the final result, your final business capabilities will get better and

Dave Vellante

>> better as That's interesting. So how, walk us through that because one of my concerns is that you end up paving the cow path. You take a process that could be improved dramatically, but there's so much opportunity to improve. I'm sure you're seeing this in the field where organizations are just rethinking how they do process. It's not RPA, it's AI. A lot of it is RPA, actually. Yeah, right,

>> right.Yeah, RPA was the extreme end of this because it was literally implementing the clicking of a

Dave Vellante

>> mouse.Do this, do that,

>> right?That was the extreme end. Right. But there's a big difference between totally open-ended business solutions and the ability to flex within the goal. And so I would say that to some degree we overshot in one direction by trying to be super smart brains with all the enterprise data. We could undershoot it too by being too specific to, no, no, I want you to do the task in the following different ways. I think there's going to be room for both. Ultimately, what's our objective? We want to make the customer service better. We want to deliver better products to our customers. And we want to reduce our cost of operations. So I would argue you should start with the stuff that works, right? So the identified high-value things, get them automated in a flexible, cost-effective, highly impactful way, and then you can increasingly move higher up the value chain or where it's hard to do that right, you can move further down the value chain and implement the sub-agents that are better at that.

Dave Vellante

>> Okay, Tyler, so you're saying start with what works. So meaning a process that you know is tried and true, Isn't that what RPA did? What's different?

>> RPA never made it work. Besides that, what's different? The problem was the RPA system couldn't reason. It was literally just going through a script of click here, click here, click here, click here. It didn't even really encode the business process. So you had BPM tools that tried to encode a business process, but they were also too fixed. In other words, if they encountered an if statement and it was not X or Y, that was it. They stopped, right? There was an error condition. These agents can react, they can change, they can make a different decision, but they're still better in a bounded context than they

Dave Vellante

>> are.Okay, but start with that middle ground. Start with those just high probability workloads. Yep. Automate that. Yep. And then move on to the more complex.

Dave Vellante

>> Yeah.And take the learnings from

>> that.Yeah. And so you don't want something that is too narrow, but you also do not want something that is too wide. And once you have an agent factory that is successfully able to execute those business processes, you've nailed them. Then you can say to the supervising agent, do I even need these processes? Should I just do something totally different? The best bit about this whole system is it can learn at every layer.

John Furrier

>> It can learn at the top, it can learn at the bottom. The keyword you said was encode. You're encoding intelligence, encoding the process into the system.

Dave Vellante

>> Yep.

John Furrier

>> And then it learns, it's always on. This is why I like this always on because whoever can encode intelligence into their operation. Yep. that's why I love physical AI. The physical world we live in is, we work and we're online, digital, a lot of digital, but still we work. The word trust comes up a lot. And, I saw Rob Thomas's NVIDIA post about OpenShell. I'm like, okay, there's trust there, but it really wasn't about plugging into NVIDIA. IBM can be the trust partner because what we're talking about here is an operating model. So ask IBM and you can get answers— accounts receivable all the way up to nuclear reactions. Yeah. so I have to ask you, when I was Sitting there, I'm like, okay, I watch the presentation. Should I think about— I'm an enterprise. I don't want to be thinking about NVIDIA OpenShell, BlueField, Iverify Vault, OpenShift Fusion. I just want it to run. those are all involved. Okay, what's the strategy to do that? Because I think that is IBM's territory. You sit above NVIDIA. NVIDIA's got OpenShell today, but they're only one player.

John Furrier

>> Yeah.You're not plugging into— well, today you are because you're OpenShell support, but You can run all agents.

>> I feel like that's insanely IBM's home turf. If you look at IBM's pedigree as a business has always been about connecting technology all the way back from the early mainframe days, technology to actually impact, successfully helping people figure out, wait a second, this giant machine that takes up an entire room, what can it do that matters in the world? So this is IBM's moment in my mind. All these technologies you talked about are foundational. They are important because all the developers in the world need to know, okay, here's how I go from theory into reality. But just as important from the direction you're talking about as an executive, you're like, if someone's gonna be able to figure out how to have an AI factory that is constantly improving itself, that really is improving my business, who's that gonna be? I would argue IBM has exactly the pedigree that shows. Yeah.

John Furrier

>> And from a business standpoint, IBM International Business Machines was in their name.

>> Yeah.

John Furrier

>> How do you make that easy? 'Cause I think that's gonna be the key task for you and the team. Yeah, make it easy for me. I don't wanna say the easy button, we use it a lot, but I think that's even, it has to be even simpler than that. It should just be ingrained in. So it's a, you could choose an abstraction layer. How do you view that?

>> Yeah, yeah. So a lot of the people who are leading the charge right now in terms of change is honestly coming from engineering. So the builders are using these tools to build faster. So all of IBM's tools around Bob, Orchestrate, really what they do is they make this stuff like a cookie cutter in some sense. Like I have created this agent, Now how do I put it into production, monitor it, make certain it doesn't break out, make certain it doesn't do anything untoward, know the impact of the thing. So helping people do that along with the data foundation and HashiCorp products, that's great. But then on the other side, you also have IBM Consulting and IBM Consulting knows everything about getting into a business process, all the different verticals all over the world, knows all the most important things going on in those businesses so they can connect these technology foundations to the business.

John Furrier

>> I like IBM's Client Zero because it's a complex organization and every enterprise is complex, so they can bring that learning and learn new things. There's a thing that comes up a lot. I want to ask your thoughts on this because I hear tool calls come up a lot. I'm making too many tool calls. Well, remember when I was coding, I had library calls, like I would call a library and then I didn't have to code anything. I would just use it. I'm like, okay, so is tool calling just the nature of the system?

>> Yeah, tool calls aren't the problem. if you look at your average random program running on your computer, the number of different calls between different modules happening in any moment, there'll be hundreds of thousands of them every second. That's not the problem. I think what they really mean is wasted tokens. What they really mean is searching the world to try and find the answer to a problem that should have been immediately at the fingertips of the agent, which is what I mean by clear context. Give the agent what it needs to be successful, not everything else. When you watch the agent traces, you'll see them invoke 50 tools when it really, in the definition of the agent, should have been, you probably want this one. This is probably the source of the information.

John Furrier

>> It's interesting. A lot of the conversation, we did a word cloud on every single keynote over the past 2 years, you'd see, you'd hear kind of similar things. Actually, governance pops up a lot too. Low latency comes up a lot. And whether you're talking NVIDIA KVCache or data serving or a graph database, which is super beautiful when it comes to the accurate low latency recall. These are the new concepts. What are the things that are key as these layers? What are the consistencies that you see that are first principles? Low latency, obviously performance. Accuracy.

>> Yeah, I just think the ability to bring it together in a federated way that represents the current state is honestly the hardest problem because people always describe to you the problem of, oh, I have one thing that will solve everything, but how would the data get there? The question is always, if the data is coming from machines on oil rigs, it's coming from cruise ships, it's coming from wherever it's coming from all in the world, how on earth would you successfully be able to bring together all of that incredibly complicated context? I think so. I think that those technologies you're talking about are part of the solution, But you should always be asking yourself, how can I know the state of my business right now? Because that is the world in which you make decisions. That is the world in which humans would make decisions too. And so you need agents to have the exact same visibility into

Dave Vellante

>> your—And that really is the opportunity for IBM. You kind of just answered my next question, which is I was going to say, what's the prescription for avoiding islands of automation? Because you have this top-down edict, thou shalt do AI. So everybody's all right, I'll do AI. And they got personal productivity boom. At the same time, the organization is trying to figure out AI. And so there is a risk of having the islands of automation. And I think you just sort of touched on the answer and maybe IBM's opportunity. What would you tell the chief data officer or the CIO, how to avoid that?

>> Yeah, I think one problem I see people make, or one mistake I see people make over and over again, is they think that AI in production for autonomous agents will look something like AI being used by you or me, right? For personal productivity. They're completely different things. Completely different things. One thing I need to drive me safely to work every single day and hopefully not kill me. The other one, I'm perfectly happy for it to give me a bit of information. We'll work together. Those are different problems. And so under the latter circumstances, which is what we're actually trying to build, these autonomous agents that can run our businesses for us, we have to think very carefully about the guardrails around the system and the data it needs to be very, very

John Furrier

>> successful.Software for software.

>> Yeah, software for software or data for data or control for systems, basically. And I think those are really interesting hard problems.

John Furrier

>> Well, Shaun, I love the AI operating model. It's consistent with the computer science trends, math feeding math, data's coming in, it's like all the same kind of things emerging. Congratulations, thanks for sharing with the analysts and coming on theCUBE.

>> Really appreciate it, thank you very much.

John Furrier

>> John Furrier with Dave Vellante, special CUBE conversation here in New York City, IBM's One Madison, their flagship office in the heart of Manhattan. Thanks for watching.