theCUBE + NYSE Wired: Data + AI: Turning Data Into Knowledge for Autonomous Systems | Harsh Acharya & Amit Sawhney, Dell Technologies | theCUBE + NYSE Wired: AI Luminaries with Neo4j

Harsh Acharya & Amit Sawhney, Dell Technologies | theCUBE + NYSE Wired: AI Luminaries with Neo4j

This session examines graph-based knowledge layers and enterprise AI architectures that enable high-velocity generative artificial intelligence in enterprise environments. Harsh Acharya of Dell Technologies, Senior Vice President, Digital, Tech Operations and Infrastructure Platform Services, and Amit Sawhney of Dell Technologies, Senior Vice President, Services Enterprise AI for Data, join the Neo4j Luminary Series hosted by John Furrier on theCUBE Research at the NYSE Wired studio. The discussion explains how knowledge graphs, semantic layers and data productization form the context fabric for agentic generative artificial intelligence. It explores graph architectures, low-latency grounding, data mesh practices and integration of Neo4j into enterprise AI platforms to enable reliable scalable agent workflows. The conversation addresses grounding large language model outputs into systems of record to reduce hallucinations and improve trust and traceability. Acharya emphasizes prioritizing data products and ontology before building knowledge graphs. They state that temporal fidelity and low-latency access are essential factors to consider for commerce and seller workflows. Sawhney highlights context engineering, governance and return on investment tied to measurable outcomes such as sales uplift and operational expenditure reduction as primary factors guiding deployment and scale decisions. This conversation offers practical guidance for technology leaders and architects designing enterprise AI infrastructure, knowledge graph implementations and context fabrics to support dependable generative AI at scale.

Share this session

Clips
More from theCUBE + NYSE Wired: Data + AI: Turning Data Into Knowledge for Autonomous Systems

Amit Sawhney

VP

Dell Technologies

Harsh Acharya

Senior VP, Digital, Tech Operations & Infrastructure Platform Services

Dell Technologies

play_circle_outline Commercial B2B commerce use cases: product, pricing, contracts, policies, trust hierarchy

play_circle_outline Knowledge Graphs, Semantic Layers, and Context Fabric Powering Agentic Generative AI Decision-Making

play_circle_outline Temporal Fidelity in Real-Time LLM Grounding with Semantic Graphs and RAG for Low-Latency Agents

play_circle_outline Data products and data mesh as atomic units for graph readiness

play_circle_outline Embedding Trust, Security, and Governance in Data Products for Unified Graphs Across Structured, Semi-Structured, and Unstructured Data

play_circle_outline Measuring ROI, governance, business ownership, and tokenization cost accountability

Info

Harsh Acharya & Amit Sawhney, Dell Technologies | theCUBE + NYSE Wired: AI Luminaries with Neo4j

Amit Sawhney

VP Dell Technologies

Harsh Acharya

Senior VP, Digital, Tech Operations & Infrastructure Platform Services Dell Technologies

search

John Furrier

>> (INTRO)

>> John Furrier, host of theCUBE. We're here at theCUBE's NYSE studio, of course. We've got our Palo Alto studio connecting Silicon Valley to Wall Street, and this is our Neo4j Luminary Series in AI. As AI comes in, we're starting to see real value being created on top of the systems that are producing all the intelligence. Of course, graph databases are a big part of it. You're starting to see new architectures. We've got two great guests, luminaries here from Dell Technologies, Harsh Acharya, Senior Vice President, Digital Infrastructure Services. Amit Sawhney, Senior Vice President, Services Enterprise AI for Data at Dell Technologies. Gentlemen, great to see you. Thanks for coming on this series. This is a Luminary Series. Congratulations, you're a Luminary.

Harsh Acharya

>> Thank you for having us.

Amit Sawhney

>> Thank you for having us.

John Furrier

>> Well, first of all, everyone at Dell seems to be a Luminary these days. The business performance has been phenomenal. Congratulations to you and your teams over at Dell. We're seeing the AI factories, rack scales changing the game, but the real sleeper story that's emerging super fast is the role of graph database knowledge graphs, knowledge layers. You're starting to see a platform shift. So this has become quite the superpower. So I know you guys have been working on this because I've been briefed on some of it a couple years ago, and I know there's been a lot of, client zero, eat your own dog food, drink your own champagne at Dell. So let's get started. Why should people pay attention to knowledge layers and graph databases, and what is the key role it plays in this new platform that's emerging that's going to power this high-velocity GenAI?

Harsh Acharya

>> I'll start and then Amit, feel free to jump in. It's almost funny, we all have used graph for a long time now in different contexts. Part of my job is running commerce and we have used graph for search. We have used it for solving for complex pricing problems. And it was just another database technology that allowed you to drive relationships and make sense out of data with different dimensions. What's really happening with the change that's coming our way is that it's now becoming a singular point that is gonna allow us to define the organizational context that the agents are going to need. And when I break it down in my head from a commerce standpoint, there is a consumer commerce that Dell does. Everyone knows Dell.com, which is a much simpler use case. You need product graph, you need pricing, and you just put in something like Muse and you can put it in something like ChatGPT and customers can interact with it. I think where graph becomes a lot more critical and more complex to define and understand is when it comes to our B2B commerce experience. Which is really how do you actually define and dictate what exists, whether it's party, product, pricing, contract, agreement, policies, what's allowed, what's not allowed, as well as what can happen. Operational ontology, which is, what's the intent, what's the capability registry, what actions and preconditions and postconditions need to be true, what's the trust hierarchy, who can do what. And today, those things kind of exist in Outlook, Excel, disparate systems. And I think the major work that's in front of us is to figure out how to utilize Graph in a very different context, no pun intended, in order to make agents work in a very complex B2B environment. And that's why it's exciting again to be looking at things like graphs, both RDF and LPG, in a very different context. And that's kind of the lens that I'm wearing. And Amit can jump into

John Furrier

>> it.

>> Yeah, Amit, what's your angle? I'm going to come back to what he just said about the history of it all, but I want you to weigh in on why it's important.

Amit Sawhney

>> Yeah, look, Harsh touched on an important word there, context. Agentic AI needs context to operate at speed and operate at scale. Human beings have context. Imagine the context that we all have in our businesses, like institutional knowledge, process, policy, procedure, SOPs, systems, system knowledge, what have you. Now we expect the GenAI to operate at speed and at scale. So the fundamental difference between them and the GenAI and the human operations is context. So to me, it's everything. And while we are talking about graph, graph is just a component of the context fabric, as Harsh and I would internally talk about. And then you have to start going back to your atomic unit level even before you talk about graph, which is your data sources. Your data platform, your data products marketplace where your data products can be brought in, can be consumed, what ontology layer you want to establish so that we are all speaking the same language. Then you talk about graph, then you talk about the semantic layer so that you have the clear relationship established. Then you are building an enterprise context which agents can then leverage to drive the things that we expect them to do at speed, at scale. So to me, Understanding the nomenclature is important. Understanding the fact that it is one of the components in the grand scheme of context and agentic AI needs context fabric and context engineering to operate at scale and speed.

John Furrier

>> Yeah, context and reasoning makes decisions. That's the big buzzword now is decision models. Harsh, I want to come back to what you said because I think it's instructive. Talk about how much data Dell has, because if you look at the history of what Dell has done online for commerce, Okay, it's been pioneering from day one, so you have a ton of data. I know that personally, but there's also a lot of product configuration, so there's a lot of data. Okay, and I know you guys have done some good things, but now fast forward to now. What happens now when you have software that wasn't built for humans to click around an app? You have software built for other software, or agents to agents. Explain how you've done it in the past and how that's coming into this software-to-software, agent-to-agent, because you got to have low latency, you got to have functionality. You can't have hallucinations because that's bad data. So you have to hit all these checkboxes.

Harsh Acharya

>> Yeah. Yeah. So look, we have petabytes of data, but we also are not naive to think that all that data is usable in its own form. And I think that's what Amit touched on data products. I'll give you an example of what we did with our, you know, sales force when we gave them the Agentic tools. The first thing we had to do was to take terabytes of unstructured data that they dealt with. And when we kind of went deeper into looking at the data, there was just a lot of, you know, outdated information. There was a lot of stuff that was true at a certain point of time but wasn't true today. And so cleaning that up was the first task. The second thing was to ensure that, you know, we were truly looking at what I call the fidelity of temporal aspect of it, which is prices, for example. Prices change, contracts expire, your inventory fluctuates all the time. So we had to separate out the truth of the product, which can stay true as far as you are looking at the latest and greatest from what we call the temporal fidelity, where you are looking at things that can change at the drop of a dime. And then to your point, from a low latency standpoint, we basically said when we have customers or sellers interacting with an agent, they can't rely on, LLMs to provide the actual real-time data. So the way we architect our, whether it's vector and RAG-based model or the new agentic model, is that the graph provides an ability for taking the language understanding that LLM provides and then goes straight into the system of record to get the data on what the price of the thing is that was recommended based on the LLM input. And we have to manage that at millisecond level, right? And so really we had to take a look at both the quality of data We have to take a look at what actually comes from LLM. And then third, how do we ground it into the reasoning path and the temporal fidelity to make sure that we are never showing anything to the customers that can change using the LLM data. Right. And to your point on the actual latency, we have to build this out for both end customers and for sellers. And it's very critical for us, just like any other page load time, it needs to happen in seconds. It can't be minutes. And so again, like I said earlier, we used to look at things like Neo4j in a very different lens in the past. And now, well, you need both semantic and low latency databases to work together.

John Furrier

>> Yeah, that's a good point about the low latency, 'cause you can have a great low latency wrong answer, right?

Amit Sawhney

>> Yeah.

John Furrier

>> That's a problem. We hit the latency metric, but wait a minute, what happened? It's not a search query. It's a re— it's an answer that requires a lot of stuff. This is a different problem than just good search.

Amit Sawhney

>> Yep.

Harsh Acharya

>> Yep.

John Furrier

>> Okay. So let me just run this by you. So many organizations are layering AI capabilities on top of their existing data platforms like Snowflake, like Databricks. You guys have so much product information. This SKU has so many different features. I missed to keep that on. Catalogs. How does this new Neo4j or graphs and knowledge layers fit together? Because the reaction might be, "Oh my God, I don't want another database." No one wants another tool or thing they have to roll out. This actually has come out of these interviews that, no, this is actually a superpower. You kind of touched on it, Harsh, a little bit. Guys, explain this role of a graph and a knowledge— why does it play such an important— how do people get involved and how do they architect this platform? Because these are data platforms. These aren't just a database, it's infrastructure. So you got to look at the holistic picture. How do people get graphs going? What's your advice?

Amit Sawhney

>> I mean, I'll start with, again, like I said earlier, you have to start establishing your data products. The data products are the subatomic unit of a good graph. If you do not have a way of understanding the dimensions of what qualifies a product to be an AI-ready data product, then we will not get the full potential of graph. Graph at the end of the day is the relationship aspects of the entities and the different aspects of what the systems of record have. But what powers that is the data and the data products. So you have to have the data mesh principle with the data product as the first principle of the data mesh framework. That anyone who's coming into the data mesh is the product owner of their respective data. They are responsible for the sanctity, the quality, the observability, the auditability, searchability aspects of that data product. You've got that. You have a marketplace where again you can look at the data products, where different people can consume your data product. But at the end of the day, you are responsible for, like I said, all the different dimensions. Then you use that data product to build the graph which again is a relationship entity, but then it's powered with the right dimensions. We know exactly what we expect that graph to do. It's like a standard garbage in, garbage out. Yeah, if you don't have the right data product, you don't have a good graph. You will have a bunch of relationships, it's like we all have context about things, but we don't have any understanding of what anyone does. So that context is fruitful— fruitless, right?

John Furrier

>> So, so you're saying relationships are key context to company information or semantic data.

Amit Sawhney

>> Right, you have to establish a semantic layer before you even get to the knowledge graph aspect. Think about two human beings, right? Or multiple people in an organization. They can interact with each other all day long.

John Furrier

>> Yeah.

Amit Sawhney

>> But if they don't have any sort of common language about what they're interacting with, I call ontology. Yeah. They are, someone's speaking French, someone's speaking English. There's no common translation layer. They might be talking about the same thing, but there's no, If you don't have that established, what the semantic aspect is, what the ontology aspect is, there's no relationship that you can establish. So if you can't establish a relationship, you can't establish a context. You can't establish a context, you can't actually execute.

John Furrier

>> Okay. How about trust? How do you guys measure trustworthiness in the outputs?

Amit Sawhney

>> It's included in the data product dimensions.

John Furrier

>> So if you get the data products right, you've grounded the trust.

Amit Sawhney

>> The trust by design, the security aspect the zero trust architectural view of how you build your data flows.

John Furrier

>> I've heard people say, just connect your Slack in there too. You could do that. So this is kind of the operational piece. Explain that part because there's now new data sources that might look messy like Slack. That's just all this content.

Amit Sawhney

>> And again, if you step back, we've talked about our journey, internally, you've got 3 different sets of data: structured, unstructured, and semi-structured. You have to build the graph across all of them that traverses, that allows the agentic AI to traverse across a homogeneous bed of context. So while you can build structured graphs, you have to spend time thinking about how you build an unstructured graph with unstructured data. And the majority of the data is unstructured data, and AI follows data. So, there's a whole body of work that has to be established on unstructured data.

John Furrier

>> Alright, you guys are both SVPs, so I'm going to have to ask this question because a lot of luminaries that have been on this series who are doing a lot of great work have had, there's been a pattern. They have risen to the top of the superpower mountain in their company because they're working on a problem and then they just break through. They've been grinding on some SQL, joining rows and data wrangling and all of a sudden, poof, magic happens and they just get elevated. We're seeing that. I'm a little bit over the top on that, but that's kind of what's happening. So how are you guys measuring the success or ROI? Because people who are starting to put this together are catapulting up and working on harder problems. So the question is, how do you see that playing out on the success side? Evaluations, the ROI, how are you measuring your initiative and how do you sell it through to management? Look, we tie it to sales. That's an easy one. But give some examples of how this rise of relevance impacts people and then the outcomes.

Harsh Acharya

>> Yeah, look, we have defined certain strategic areas of the company where we do major investment in AI, but everything goes through our chief officer and our CIO. Looking at that both from a technology standpoint, security standpoint, and then our chief operating officer looking at it from an ROI standpoint before we let anyone take any AI workload to development and production. Amit actually helps us govern that, but the way we approach this is that we believe that you can do a lot of G&A-like functions through a lot of products that we have bought. And, you know, you just get those AI capabilities out of the box. And then in strategic areas where we really get benefits both at top line and bottom line, that's where we do all the investment. And look, if you don't have ROI, that just gets thrown out. You just don't work on it. And that, by the way, also helps us focus on the right set of things, both from a data and knowledge standpoint as well as our AI capacity standpoint, whether it's infrastructure or models or cost. But that's very much a top-down governance and mandate. We are a very operationally savvy company. I'm sure—

John Furrier

>> Yeah, of course. Yeah, go ahead, weigh in.

Amit Sawhney

>> One of the great things in our company, we measure everything. Yeah, we absolutely track everything, and it is part of their budget. You know, even when we talk about tokenization, it's coming out of people's budget. ROI is part of people's performance plan. So this thing is not just some sort of hand waving or some Excel math that doesn't add up. This is actual real stuff. And if you look at our performance, public information, uh, we've decoupled revenue from OpEx, right? Our revenue has grown and our OpEx has come down, and that is a testament to the great work of the AI and the data capabilities that our team members have driven. And again, of which Graph plays a critical role. So proof's in the pudding. The other thing that I also want to hit on is most companies don't understand the roles and responsibilities clearly between what BU or the domains are responsible for and what they expect their IT departments to do for them. We have been very clear. Business owns the outcomes. So if you are talking about your data sources, business owns the data sources. Yeah, IT will help you with building the platform. IT will help you build a marketplace, but you are responsible for bringing your data products, building your data products, and being responsible for the quality of those data products, bringing them to the marketplace. You are responsible for the ontology, i.e., business. Business is responsible for the knowledge graph at the end of the day, and business is responsible for the agent apps that they build that are leveraging those knowledge graphs. So that responsibility and roles have to be clarified from day one, or otherwise it becomes like, hey, I just passed it over to the IT department, and the IT department is expected to build me a knowledge graph and context engineering and context fabric. But no, we are very clear about what IT drives from platforming versus business or domain is responsible for from an outcome perspective.

John Furrier

>> I love that outcome perspective. But you brought up the challenge and the old way was, oh, build a knowledge layer. Oh, I got to build it from scratch. There's a lot of leverage going on, the things that, for example, you've done. So I appreciate that one metric of ROI, but talk about the tech breakthrough. So imagine your world because You guys have been developing this for years. A lot of people have run into the same thing. What have been some technical roadblocks or walls you've hit where graphs and knowledge graphs opened the floodgates of value? Can you share stories where you're grinding away and, then something happened?

Amit Sawhney

>> Between Harsh and myself, right? if you look at commerce and services, legacy way of doing it would be you build a commerce RDBMS, relational database management system, or services RDBMS, and then you try to connect the two together, and that would take forever, and then half of the data will not be right, and then the relationship will not be established, and you won't get the right answer. Now fast forward with the graph, you bring the two graphs together, right? a seller has information on what support tickets are open, a seller has information on what's working and what's not working in the customer environment that they're about to have a conversation about. And so the boundaries have actually been shattered. And we practice the philosophy very clearly, agent AI doesn't know the boundaries, organizational boundaries, those are human-driven. Agent AI doesn't know services or commerce or supply chain, that's gonna traverse horizontal. So for Harsh and I and several of our peers here in Dell Technologies, we have to start thinking about how we build the horizontal fabric across the company so that you have a true customer 360 view.

John Furrier

>> That's why I think this is—

Harsh Acharya

>> by the way, we put a lot of onus on starting with process first.

Amit Sawhney

>> Yeah, right.

Harsh Acharya

>> We do not let people use AI to automate things that shouldn't have existed in the first place. And so our conversation never starts with the technology breakthrough or what technology we can use to take your current process and automate it. Even within IT, in our own world to modernize ourselves, we are starting by documenting everything that we do. Then looking at how we would do it if we were starting the company today. And then you think about how to make it agentic, and then you talk about technology. And that's the biggest mistake I've seen people make in the industry is that they start with, I just want to use AI to do what I do today. And that's something we just discourage across the company.

John Furrier

>> Yeah.

Amit Sawhney

>> You have to reimagine your workflows. That's the way to start the AI journey.

John Furrier

>> Well, you have these little siloed projects come up, but the way you are thinking about it, you were just saying horizontal means it's a system. This is a data system. It's a data operating system. This is infrastructure. We're getting into a whole nother level. And so that's why I was asking about these breakthroughs because if you're in infrastructure, you have to have all the data at your fingertips. And the question will come up, and I think for Dell, this is a great one for you guys to talk about, is scale. Scale is hard. And so that complicates it. So if you're going to be horizontally scalable on the data layer, but yet be very domain and knowledge specific, you got to get that right. So could you guys talk about the scale challenge, how you think about it, how other teams might want to think about it? Because again, people are growing. You guys are well ahead of the pack for sure. I know that personally. But talk about the scale dynamic here.

Harsh Acharya

>> Well, look, it shouldn't surprise you coming from leaders of Dell Technologies, we genuinely believe that as more and more things move towards GenAI, the majority of inference will happen on-prem in our environment. And so the way we are architecting things, the way we are building things, we assume that, let's say I'm building a commerce workflow, or I'm building a workflow for our IT engineers. We assume that it's gonna be 90% automated. And when we think about how to set up our models, the context window, the hardware that we need, the storage that we need, how it's gonna work with the network, we plan for that scale. In our infrastructure that we put in place in our own environment. And so, we spend a lot of time doing that, and that is basically vertical scaling. And then from a horizontal scaling standpoint, if you think about it, like you said earlier, all the workflows that exist today were built assuming that purchasers or sellers or users are taking this horizontal route. They're clicking through every single, window in the pane and getting to the end. We are having to look at our API layer and abstraction layer completely differently. That's why we like GraphQL. It allows you to— it's almost like monoliths are awesome again. We went to this monoliths-based architecture like price is here, product is here. Well, in order for agentic AI to work really well, you need all of that to be in the same place. We kind of, vertically scale up our hardware infrastructure with the right models in place, which Amit actually helps us with. And then we do horizontal scaling using the likes of, Graph technology to scale it out.

John Furrier

>> I'm smiling because he said monoliths. I'm like, I feel like I'm having a cloud-native conversation because that's— monoliths had to touch all the resources. They had to understand what was going on. Cloud's horizontally scaled, but now with hybrid and on-prem, it's distributed computing. So if I'm going to implement enterprise AI and I'm an enterprise, I want to produce intelligence, I want to commercialize it, I have to look at everything. You got disparate systems right now, you got all kinds of infrastructure all over the place, you got legacy, all that comes together. This is a huge challenge. How can it become easier, I guess?

Harsh Acharya

>> Well, if it was easy, everyone would have done it by now.

John Furrier

>> Get a craft.

Harsh Acharya

>> Look, it doesn't make it easier. The only way to make it easier, like I said earlier, you have to start with reimagining your workflow and look at your processes from the ground up. There is no easy way. You can build, siloed agents. You can do Copilot stuff. You can do chatbots. But if you are truly trying to do, like Amit said, horizontal end-to-end agentic experiences, there is no—

John Furrier

>> Okay. So would you believe this to be a true statement if I said to you, turn your process into intelligence? Is that possible? And is that the right way to think about it?

Amit Sawhney

>> It is actually. It actually is. we'll give you a very relevant example. I'll speak for services domain that I'm responsible for as well. We started digitizing our processes with structured data in 2018, 2019 with ARIS, Celonis, and we now have unstructured data that we are bringing into it as well. We took all of our processes, about 3,000 unique workflows that we reimagined. We broke it down into tasks, 28,000 odd tasks, which is the subatomic unit level, because agentification has to happen at the task level. So one workflow could have one-to-many tasks. So that is the level of granularity by which what we are talking about, reimagining the process flow, that is what we're talking aboutis you have to understand your workflows, you have to understand your tasks that make up a workflow. Then you understand how many of those tasks can really be end-to-end agentified, if that's a word, without any human in the loop. What percentage would have some sort of AI assistance with human in the loop and what tasks can just completely go away if you were to actually do this or start from scratch as an AI-native organization. So we have that level of intelligence that we have now captured. And now we are building that to then bring in graph capabilities to build those relationships that allow us to capture the context that is across the different organizations of just taking services department at the seams, because at the seams is where the context is which we need to exploit. And then you have an agent capability that goes across horizontally.

John Furrier

>> Well, guys, this has been quite a great topic. We could go another hour, but I will just say one thing. The work that's being done at rack scale that's delivering the intelligence, these factories. They're producing intelligence in the entire system. That's enabling a whole nother kind of, another system, the data system, whatever you want to call it, it's my word, but that's the way I think about it. That's the flywheel, so you're injecting intelligence into the process, and your advice is just get your data in order and start doing it.

Amit Sawhney

>> Yep. That's

John Furrier

>> great.

>> Well, you guys are luminaries. Again, we've been tracking some of the success. I know you've been working on a lot of these projects, but you guys are also operating at scale. Congratulations. Thanks for coming on.

Harsh Acharya

>> Thank you so much.

John Furrier

>> Appreciate it. I'm John Furrier. This is the Luminary Series with Neo4j and the industry. It's really a discussion with leaders who are building the data operating systems for these companies, the AI factories enabling intelligence, and people are looking at the data structures Some say the brain of the enterprise. The moat is the data. That seems to be super valuable, and on-premise activity unlocked data that's happening is amazing. We're doing our part here on theCUBE to bring that data to you. Thanks for watching.

Harsh Acharya & Amit Sawhney, Dell Technologies | theCUBE + NYSE Wired: AI Luminaries with Neo4j

search

John Furrier

>> (INTRO)

>> John Furrier, host of theCUBE. We're here at theCUBE's NYSE studio, of course. We've got our Palo Alto studio connecting Silicon Valley to Wall Street, and this is our Neo4j Luminary Series in AI. As AI comes in, we're starting to see real value being created on top of the systems that are producing all the intelligence. Of course, graph databases are a big part of it. You're starting to see new architectures. We've got two great guests, luminaries here from Dell Technologies, Harsh Acharya, Senior Vice President, Digital Infrastructure Services. Amit Sawhney, Senior Vice President, Services Enterprise AI for Data at Dell Technologies. Gentlemen, great to see you. Thanks for coming on this series. This is a Luminary Series. Congratulations, you're a Luminary.

Harsh Acharya

>> Thank you for having us.

Amit Sawhney

>> Thank you for having us.

John Furrier

>> Well, first of all, everyone at Dell seems to be a Luminary these days. The business performance has been phenomenal. Congratulations to you and your teams over at Dell. We're seeing the AI factories, rack scales changing the game, but the real sleeper story that's emerging super fast is the role of graph database knowledge graphs, knowledge layers. You're starting to see a platform shift. So this has become quite the superpower. So I know you guys have been working on this because I've been briefed on some of it a couple years ago, and I know there's been a lot of, client zero, eat your own dog food, drink your own champagne at Dell. So let's get started. Why should people pay attention to knowledge layers and graph databases, and what is the key role it plays in this new platform that's emerging that's going to power this high-velocity GenAI?

Harsh Acharya

>> I'll start and then Amit, feel free to jump in. It's almost funny, we all have used graph for a long time now in different contexts. Part of my job is running commerce and we have used graph for search. We have used it for solving for complex pricing problems. And it was just another database technology that allowed you to drive relationships and make sense out of data with different dimensions. What's really happening with the change that's coming our way is that it's now becoming a singular point that is gonna allow us to define the organizational context that the agents are going to need. And when I break it down in my head from a commerce standpoint, there is a consumer commerce that Dell does. Everyone knows Dell.com, which is a much simpler use case. You need product graph, you need pricing, and you just put in something like Muse and you can put it in something like ChatGPT and customers can interact with it. I think where graph becomes a lot more critical and more complex to define and understand is when it comes to our B2B commerce experience. Which is really how do you actually define and dictate what exists, whether it's party, product, pricing, contract, agreement, policies, what's allowed, what's not allowed, as well as what can happen. Operational ontology, which is, what's the intent, what's the capability registry, what actions and preconditions and postconditions need to be true, what's the trust hierarchy, who can do what. And today, those things kind of exist in Outlook, Excel, disparate systems. And I think the major work that's in front of us is to figure out how to utilize Graph in a very different context, no pun intended, in order to make agents work in a very complex B2B environment. And that's why it's exciting again to be looking at things like graphs, both RDF and LPG, in a very different context. And that's kind of the lens that I'm wearing. And Amit can jump into

John Furrier

>> it.

>> Yeah, Amit, what's your angle? I'm going to come back to what he just said about the history of it all, but I want you to weigh in on why it's important.

Amit Sawhney

>> Yeah, look, Harsh touched on an important word there, context. Agentic AI needs context to operate at speed and operate at scale. Human beings have context. Imagine the context that we all have in our businesses, like institutional knowledge, process, policy, procedure, SOPs, systems, system knowledge, what have you. Now we expect the GenAI to operate at speed and at scale. So the fundamental difference between them and the GenAI and the human operations is context. So to me, it's everything. And while we are talking about graph, graph is just a component of the context fabric, as Harsh and I would internally talk about. And then you have to start going back to your atomic unit level even before you talk about graph, which is your data sources. Your data platform, your data products marketplace where your data products can be brought in, can be consumed, what ontology layer you want to establish so that we are all speaking the same language. Then you talk about graph, then you talk about the semantic layer so that you have the clear relationship established. Then you are building an enterprise context which agents can then leverage to drive the things that we expect them to do at speed, at scale. So to me, Understanding the nomenclature is important. Understanding the fact that it is one of the components in the grand scheme of context and agentic AI needs context fabric and context engineering to operate at scale and speed.

John Furrier

>> Yeah, context and reasoning makes decisions. That's the big buzzword now is decision models. Harsh, I want to come back to what you said because I think it's instructive. Talk about how much data Dell has, because if you look at the history of what Dell has done online for commerce, Okay, it's been pioneering from day one, so you have a ton of data. I know that personally, but there's also a lot of product configuration, so there's a lot of data. Okay, and I know you guys have done some good things, but now fast forward to now. What happens now when you have software that wasn't built for humans to click around an app? You have software built for other software, or agents to agents. Explain how you've done it in the past and how that's coming into this software-to-software, agent-to-agent, because you got to have low latency, you got to have functionality. You can't have hallucinations because that's bad data. So you have to hit all these checkboxes.

Harsh Acharya

>> Yeah. Yeah. So look, we have petabytes of data, but we also are not naive to think that all that data is usable in its own form. And I think that's what Amit touched on data products. I'll give you an example of what we did with our, you know, sales force when we gave them the Agentic tools. The first thing we had to do was to take terabytes of unstructured data that they dealt with. And when we kind of went deeper into looking at the data, there was just a lot of, you know, outdated information. There was a lot of stuff that was true at a certain point of time but wasn't true today. And so cleaning that up was the first task. The second thing was to ensure that, you know, we were truly looking at what I call the fidelity of temporal aspect of it, which is prices, for example. Prices change, contracts expire, your inventory fluctuates all the time. So we had to separate out the truth of the product, which can stay true as far as you are looking at the latest and greatest from what we call the temporal fidelity, where you are looking at things that can change at the drop of a dime. And then to your point, from a low latency standpoint, we basically said when we have customers or sellers interacting with an agent, they can't rely on, LLMs to provide the actual real-time data. So the way we architect our, whether it's vector and RAG-based model or the new agentic model, is that the graph provides an ability for taking the language understanding that LLM provides and then goes straight into the system of record to get the data on what the price of the thing is that was recommended based on the LLM input. And we have to manage that at millisecond level, right? And so really we had to take a look at both the quality of data We have to take a look at what actually comes from LLM. And then third, how do we ground it into the reasoning path and the temporal fidelity to make sure that we are never showing anything to the customers that can change using the LLM data. Right. And to your point on the actual latency, we have to build this out for both end customers and for sellers. And it's very critical for us, just like any other page load time, it needs to happen in seconds. It can't be minutes. And so again, like I said earlier, we used to look at things like Neo4j in a very different lens in the past. And now, well, you need both semantic and low latency databases to work together.

John Furrier

>> Yeah, that's a good point about the low latency, 'cause you can have a great low latency wrong answer, right?

Amit Sawhney

>> Yeah.

John Furrier

>> That's a problem. We hit the latency metric, but wait a minute, what happened? It's not a search query. It's a re— it's an answer that requires a lot of stuff. This is a different problem than just good search.

Amit Sawhney

>> Yep.

Harsh Acharya

>> Yep.

John Furrier

>> Okay. So let me just run this by you. So many organizations are layering AI capabilities on top of their existing data platforms like Snowflake, like Databricks. You guys have so much product information. This SKU has so many different features. I missed to keep that on. Catalogs. How does this new Neo4j or graphs and knowledge layers fit together? Because the reaction might be, "Oh my God, I don't want another database." No one wants another tool or thing they have to roll out. This actually has come out of these interviews that, no, this is actually a superpower. You kind of touched on it, Harsh, a little bit. Guys, explain this role of a graph and a knowledge— why does it play such an important— how do people get involved and how do they architect this platform? Because these are data platforms. These aren't just a database, it's infrastructure. So you got to look at the holistic picture. How do people get graphs going? What's your advice?

Amit Sawhney

>> I mean, I'll start with, again, like I said earlier, you have to start establishing your data products. The data products are the subatomic unit of a good graph. If you do not have a way of understanding the dimensions of what qualifies a product to be an AI-ready data product, then we will not get the full potential of graph. Graph at the end of the day is the relationship aspects of the entities and the different aspects of what the systems of record have. But what powers that is the data and the data products. So you have to have the data mesh principle with the data product as the first principle of the data mesh framework. That anyone who's coming into the data mesh is the product owner of their respective data. They are responsible for the sanctity, the quality, the observability, the auditability, searchability aspects of that data product. You've got that. You have a marketplace where again you can look at the data products, where different people can consume your data product. But at the end of the day, you are responsible for, like I said, all the different dimensions. Then you use that data product to build the graph which again is a relationship entity, but then it's powered with the right dimensions. We know exactly what we expect that graph to do. It's like a standard garbage in, garbage out. Yeah, if you don't have the right data product, you don't have a good graph. You will have a bunch of relationships, it's like we all have context about things, but we don't have any understanding of what anyone does. So that context is fruitful— fruitless, right?

John Furrier

>> So, so you're saying relationships are key context to company information or semantic data.

Amit Sawhney

>> Right, you have to establish a semantic layer before you even get to the knowledge graph aspect. Think about two human beings, right? Or multiple people in an organization. They can interact with each other all day long.

John Furrier

>> Yeah.

Amit Sawhney

>> But if they don't have any sort of common language about what they're interacting with, I call ontology. Yeah. They are, someone's speaking French, someone's speaking English. There's no common translation layer. They might be talking about the same thing, but there's no, If you don't have that established, what the semantic aspect is, what the ontology aspect is, there's no relationship that you can establish. So if you can't establish a relationship, you can't establish a context. You can't establish a context, you can't actually execute.

John Furrier

>> Okay. How about trust? How do you guys measure trustworthiness in the outputs?

Amit Sawhney

>> It's included in the data product dimensions.

John Furrier

>> So if you get the data products right, you've grounded the trust.

Amit Sawhney

>> The trust by design, the security aspect the zero trust architectural view of how you build your data flows.

John Furrier

>> I've heard people say, just connect your Slack in there too. You could do that. So this is kind of the operational piece. Explain that part because there's now new data sources that might look messy like Slack. That's just all this content.

Amit Sawhney

>> And again, if you step back, we've talked about our journey, internally, you've got 3 different sets of data: structured, unstructured, and semi-structured. You have to build the graph across all of them that traverses, that allows the agentic AI to traverse across a homogeneous bed of context. So while you can build structured graphs, you have to spend time thinking about how you build an unstructured graph with unstructured data. And the majority of the data is unstructured data, and AI follows data. So, there's a whole body of work that has to be established on unstructured data.

John Furrier

>> Alright, you guys are both SVPs, so I'm going to have to ask this question because a lot of luminaries that have been on this series who are doing a lot of great work have had, there's been a pattern. They have risen to the top of the superpower mountain in their company because they're working on a problem and then they just break through. They've been grinding on some SQL, joining rows and data wrangling and all of a sudden, poof, magic happens and they just get elevated. We're seeing that. I'm a little bit over the top on that, but that's kind of what's happening. So how are you guys measuring the success or ROI? Because people who are starting to put this together are catapulting up and working on harder problems. So the question is, how do you see that playing out on the success side? Evaluations, the ROI, how are you measuring your initiative and how do you sell it through to management? Look, we tie it to sales. That's an easy one. But give some examples of how this rise of relevance impacts people and then the outcomes.

Harsh Acharya

>> Yeah, look, we have defined certain strategic areas of the company where we do major investment in AI, but everything goes through our chief officer and our CIO. Looking at that both from a technology standpoint, security standpoint, and then our chief operating officer looking at it from an ROI standpoint before we let anyone take any AI workload to development and production. Amit actually helps us govern that, but the way we approach this is that we believe that you can do a lot of G&A-like functions through a lot of products that we have bought. And, you know, you just get those AI capabilities out of the box. And then in strategic areas where we really get benefits both at top line and bottom line, that's where we do all the investment. And look, if you don't have ROI, that just gets thrown out. You just don't work on it. And that, by the way, also helps us focus on the right set of things, both from a data and knowledge standpoint as well as our AI capacity standpoint, whether it's infrastructure or models or cost. But that's very much a top-down governance and mandate. We are a very operationally savvy company. I'm sure—

John Furrier

>> Yeah, of course. Yeah, go ahead, weigh in.

Amit Sawhney

>> One of the great things in our company, we measure everything. Yeah, we absolutely track everything, and it is part of their budget. You know, even when we talk about tokenization, it's coming out of people's budget. ROI is part of people's performance plan. So this thing is not just some sort of hand waving or some Excel math that doesn't add up. This is actual real stuff. And if you look at our performance, public information, uh, we've decoupled revenue from OpEx, right? Our revenue has grown and our OpEx has come down, and that is a testament to the great work of the AI and the data capabilities that our team members have driven. And again, of which Graph plays a critical role. So proof's in the pudding. The other thing that I also want to hit on is most companies don't understand the roles and responsibilities clearly between what BU or the domains are responsible for and what they expect their IT departments to do for them. We have been very clear. Business owns the outcomes. So if you are talking about your data sources, business owns the data sources. Yeah, IT will help you with building the platform. IT will help you build a marketplace, but you are responsible for bringing your data products, building your data products, and being responsible for the quality of those data products, bringing them to the marketplace. You are responsible for the ontology, i.e., business. Business is responsible for the knowledge graph at the end of the day, and business is responsible for the agent apps that they build that are leveraging those knowledge graphs. So that responsibility and roles have to be clarified from day one, or otherwise it becomes like, hey, I just passed it over to the IT department, and the IT department is expected to build me a knowledge graph and context engineering and context fabric. But no, we are very clear about what IT drives from platforming versus business or domain is responsible for from an outcome perspective.

John Furrier

>> I love that outcome perspective. But you brought up the challenge and the old way was, oh, build a knowledge layer. Oh, I got to build it from scratch. There's a lot of leverage going on, the things that, for example, you've done. So I appreciate that one metric of ROI, but talk about the tech breakthrough. So imagine your world because You guys have been developing this for years. A lot of people have run into the same thing. What have been some technical roadblocks or walls you've hit where graphs and knowledge graphs opened the floodgates of value? Can you share stories where you're grinding away and, then something happened?

Amit Sawhney

>> Between Harsh and myself, right? if you look at commerce and services, legacy way of doing it would be you build a commerce RDBMS, relational database management system, or services RDBMS, and then you try to connect the two together, and that would take forever, and then half of the data will not be right, and then the relationship will not be established, and you won't get the right answer. Now fast forward with the graph, you bring the two graphs together, right? a seller has information on what support tickets are open, a seller has information on what's working and what's not working in the customer environment that they're about to have a conversation about. And so the boundaries have actually been shattered. And we practice the philosophy very clearly, agent AI doesn't know the boundaries, organizational boundaries, those are human-driven. Agent AI doesn't know services or commerce or supply chain, that's gonna traverse horizontal. So for Harsh and I and several of our peers here in Dell Technologies, we have to start thinking about how we build the horizontal fabric across the company so that you have a true customer 360 view.

John Furrier

>> That's why I think this is—

Harsh Acharya

>> by the way, we put a lot of onus on starting with process first.

Amit Sawhney

>> Yeah, right.

Harsh Acharya

>> We do not let people use AI to automate things that shouldn't have existed in the first place. And so our conversation never starts with the technology breakthrough or what technology we can use to take your current process and automate it. Even within IT, in our own world to modernize ourselves, we are starting by documenting everything that we do. Then looking at how we would do it if we were starting the company today. And then you think about how to make it agentic, and then you talk about technology. And that's the biggest mistake I've seen people make in the industry is that they start with, I just want to use AI to do what I do today. And that's something we just discourage across the company.

John Furrier

>> Yeah.

Amit Sawhney

>> You have to reimagine your workflows. That's the way to start the AI journey.

John Furrier

>> Well, you have these little siloed projects come up, but the way you are thinking about it, you were just saying horizontal means it's a system. This is a data system. It's a data operating system. This is infrastructure. We're getting into a whole nother level. And so that's why I was asking about these breakthroughs because if you're in infrastructure, you have to have all the data at your fingertips. And the question will come up, and I think for Dell, this is a great one for you guys to talk about, is scale. Scale is hard. And so that complicates it. So if you're going to be horizontally scalable on the data layer, but yet be very domain and knowledge specific, you got to get that right. So could you guys talk about the scale challenge, how you think about it, how other teams might want to think about it? Because again, people are growing. You guys are well ahead of the pack for sure. I know that personally. But talk about the scale dynamic here.

Harsh Acharya

>> Well, look, it shouldn't surprise you coming from leaders of Dell Technologies, we genuinely believe that as more and more things move towards GenAI, the majority of inference will happen on-prem in our environment. And so the way we are architecting things, the way we are building things, we assume that, let's say I'm building a commerce workflow, or I'm building a workflow for our IT engineers. We assume that it's gonna be 90% automated. And when we think about how to set up our models, the context window, the hardware that we need, the storage that we need, how it's gonna work with the network, we plan for that scale. In our infrastructure that we put in place in our own environment. And so, we spend a lot of time doing that, and that is basically vertical scaling. And then from a horizontal scaling standpoint, if you think about it, like you said earlier, all the workflows that exist today were built assuming that purchasers or sellers or users are taking this horizontal route. They're clicking through every single, window in the pane and getting to the end. We are having to look at our API layer and abstraction layer completely differently. That's why we like GraphQL. It allows you to— it's almost like monoliths are awesome again. We went to this monoliths-based architecture like price is here, product is here. Well, in order for agentic AI to work really well, you need all of that to be in the same place. We kind of, vertically scale up our hardware infrastructure with the right models in place, which Amit actually helps us with. And then we do horizontal scaling using the likes of, Graph technology to scale it out.

John Furrier

>> I'm smiling because he said monoliths. I'm like, I feel like I'm having a cloud-native conversation because that's— monoliths had to touch all the resources. They had to understand what was going on. Cloud's horizontally scaled, but now with hybrid and on-prem, it's distributed computing. So if I'm going to implement enterprise AI and I'm an enterprise, I want to produce intelligence, I want to commercialize it, I have to look at everything. You got disparate systems right now, you got all kinds of infrastructure all over the place, you got legacy, all that comes together. This is a huge challenge. How can it become easier, I guess?

Harsh Acharya

>> Well, if it was easy, everyone would have done it by now.

John Furrier

>> Get a craft.

Harsh Acharya

>> Look, it doesn't make it easier. The only way to make it easier, like I said earlier, you have to start with reimagining your workflow and look at your processes from the ground up. There is no easy way. You can build, siloed agents. You can do Copilot stuff. You can do chatbots. But if you are truly trying to do, like Amit said, horizontal end-to-end agentic experiences, there is no—

John Furrier

>> Okay. So would you believe this to be a true statement if I said to you, turn your process into intelligence? Is that possible? And is that the right way to think about it?

Amit Sawhney

>> It is actually. It actually is. we'll give you a very relevant example. I'll speak for services domain that I'm responsible for as well. We started digitizing our processes with structured data in 2018, 2019 with ARIS, Celonis, and we now have unstructured data that we are bringing into it as well. We took all of our processes, about 3,000 unique workflows that we reimagined. We broke it down into tasks, 28,000 odd tasks, which is the subatomic unit level, because agentification has to happen at the task level. So one workflow could have one-to-many tasks. So that is the level of granularity by which what we are talking about, reimagining the process flow, that is what we're talking aboutis you have to understand your workflows, you have to understand your tasks that make up a workflow. Then you understand how many of those tasks can really be end-to-end agentified, if that's a word, without any human in the loop. What percentage would have some sort of AI assistance with human in the loop and what tasks can just completely go away if you were to actually do this or start from scratch as an AI-native organization. So we have that level of intelligence that we have now captured. And now we are building that to then bring in graph capabilities to build those relationships that allow us to capture the context that is across the different organizations of just taking services department at the seams, because at the seams is where the context is which we need to exploit. And then you have an agent capability that goes across horizontally.

John Furrier

>> Well, guys, this has been quite a great topic. We could go another hour, but I will just say one thing. The work that's being done at rack scale that's delivering the intelligence, these factories. They're producing intelligence in the entire system. That's enabling a whole nother kind of, another system, the data system, whatever you want to call it, it's my word, but that's the way I think about it. That's the flywheel, so you're injecting intelligence into the process, and your advice is just get your data in order and start doing it.

Amit Sawhney

>> Yep. That's

John Furrier

>> great.

>> Well, you guys are luminaries. Again, we've been tracking some of the success. I know you've been working on a lot of these projects, but you guys are also operating at scale. Congratulations. Thanks for coming on.

Harsh Acharya

>> Thank you so much.

John Furrier

>> Appreciate it. I'm John Furrier. This is the Luminary Series with Neo4j and the industry. It's really a discussion with leaders who are building the data operating systems for these companies, the AI factories enabling intelligence, and people are looking at the data structures Some say the brain of the enterprise. The moat is the data. That seems to be super valuable, and on-premise activity unlocked data that's happening is amazing. We're doing our part here on theCUBE to bring that data to you. Thanks for watching.