315 | Breaking Analysis | How AI Stacks are Rewriting the Rules of Business

0:00 >> From theCUBE Studios in Palo Alto and Boston, bringing you data-driven insights from theCUBE and ETR, this is Breaking Analysis with Dave Vellante. >> The shift from on prem to SaaS changed the technology business and operating models for IT, but it largely stopped there. Enterprises, they gained the ability or they gained agility and they had less friction coming from their IT departments, but how companies actually made money and the way they operated day-to- day, that really didn't change.

0:34 I mean, it stayed fundamentally intact. Now, SaaS companies themselves were the obvious exception, but big changes really only affected technology vendors, not the buyers. AI will be completely different in this regard, in our view. So this isn't just a technology change. The way organizations operate, the way they allocate capital and the way they generate revenue is being altered in ways that will permanently reshape how business works.

1:05 The core problem today is one that most executives can probably describe, but few have solved. This is the fact that organizations are running in silos and you think about it, each department has its own applications, its own data models, its business logic is trapped inside of those applications. And most of these work reasonably well for that department, but that's in isolation. What remains elusive is a single version of the truth across the enterprise. Let's face it, people still spend enormous amounts of time reconciling data, chasing down tribal knowledge and manually translating insight into action.

1:46 That's not an IT problem. That is an organizational tax. AI's promise is to eliminate much, if not most of that tax, and deliver dramatically higher productivity and the ability to scale without proportional labor growth. But unlike SaaS, the effect won't be contained to just the technology department or technology vendors. These changes will permeate the entire organization for the technology buyers across all industries, across all departments.

2:17 So our view is that the enterprises that get this right will become significantly more efficient, yet they'll also become platform companies that are generating network effects that create winner take most dynamics and much more sustainable economic advantage. In this Breaking Analysis, we preview the mental model that George Gilbert has developed to describe exactly how this transformation unfolds and we set up for a deeper dive on next week's Breaking Analysis. George Gilbert, welcome.

2:48 >> Good to see you, Dave. - Okay. Alex, please bring up the first slide. We talk a lot about bringing together the worlds of determinism, deterministic software and probabilistic software. Generative AI is probabilistic, bringing those two worlds together so that agents can act confidently. The fact is today's deterministic software, as we show in this slide that was developed by David Floyer on the left-hand side, it's a jungle of applications.

3:19 You see the ERP systems, the CRM, the Salesforce, the HCM, et cetera. And of course, importantly, the historical system of analytics, which is just that. It's a separated system. And we spend a lot of time, as we said front, just reconciling that with humans, the human glue here, this expert interpretation, tribal knowledge. There are a number of terms for that. And this all results in problems for the business.

3:49 You've got conflicting rules. You've got different guidance from different managers. You got a delayed truth. It takes a long time for corporate edicts to trickle down throughout the organization. A lot of the recovery tends to be manual. So George, picking up from there, this is the fundamental myth of determinism. And what we want to get to is a truly deterministic system that also has intelligence. Your thoughts? >> Yeah. This platform or this slide that David Floyer put together about how we have these islands of applications, it's really profound because you and I have talked a lot in the past about these islands of operational activity that the applications correspond to.

4:39 But what David has pointed out is that, even though these applications are deterministic, there's so much human glue in the processes around them that their output is effectively probabilistic. You can't guarantee that you're going to get the same outcome every time, despite the same inputs. That's a pretty profound insight that David had. And what we're trying to say is that's like a craft work model, where, before we had the assembly line, every part that came out was slightly different.

5:17 And then every finished product was slightly different because it was craft work. And that's what this is saying. You've got these islands, which are like the machines, and the people that are working on these machines in these silos and the result is you've got craft economics and craft work as an output. And we need to make knowledge work something much more repeatable and we need to change the economics so that the costs don't rise in line with revenue.

5:49 In other words, we need to add leverage to the model, repeatability. Repeatability to the quality and leverage to the economics, and then if we do it right, we can even add customization so that each offering can be different, but with the economics of repeatability. It sounds really abstract, but ... >> So let's make a finer point on that to help the audience. So craft economics really manifests itself in the form of human labor.

6:21 So your marginal economics at volume are more like services companies. You might get this economies of scale as you get larger. Companies who have gone through growth know this, especially services companies, it just gets more and more complex. There's more people that are out in the field and giving direction. What we've been saying now for quite some time, years actually, is the model will be not software as service, but service as software, where every company actually becomes a software company and software, like marginal economics, accrue to those companies that get this right.

7:03 Is that what you're saying, George? >> Yes. - That is your fundamental thesis. >> It changes everything. With this new technology model where we harmonize, as we'll explain, and integrate all these islands, then you get a new operational model where, instead of work being craft economics, either craft economics or industrial scale sameness, you can have the best of both. And then that changes your business model because then the business model for customers becomes like software companies.

7:39 They become like platforms and that changes every industry. >> So we're really scratching the surface here, but we're trying to bring together literally years of research that George has done, that I've been supporting and David Floyer has also done across theCUBE research team as well. So again, it's the technology model, the business model, and the operational model that George's model, his mental model of the future, has really dug into. And we want to summarize that and we're going to go deeper next week.

8:10 So let's focus now on the technology model and what that new AI software stack looks like, Alex, if you bring up the next slide. So again, we're talking about service as software and it requires people to really fundamentally change their thinking of what their technology looks like. We talked a lot about AI factories. We did a deep dive last week with David Floyer on the impact of AI factories. We've been talking about that for years now.

8:36 AI factories are fundamentally different. It's not just a server or a storage box or a network. The unit is now a rack and that AI factory produces intelligence in the form of tokens and those tokens will be monetized. That's where we get into the business model, but it requires a completely new software stack. George, I'll set it up and then I'd like you to go deeper and we can talk about each of these. So the bottom layer, you can see the traditional data platforms, where the modern data stack is now the new legacy.

9:09 But these are historical analytic systems. They largely, as George often says, tell you what happened and maybe even why it happened, but not what's going to happen next and what we should do next, and it certainly can't take action. So on the right-hand side at the bottom, you've got essentially the transaction systems, the systems of record, your CRM, ERP, those systems that are taking action. And that's what the agents are ultimately going to do. So those transactions have to be part of the system and has to be integrated.

9:42 And then the most important layer that we always talk about is a system of intelligence. People talk about the semantic layer. They talk about context. We'll come back to that. But not only why did something happen, but what is likely to happen and what should be done next? And then actually take action, that's the top layer, the system of agency. And then very importantly on the left-hand side, this is where humans are engaging. Some people might think of system of engagement as social networking.

10:12 That's not what we're talking about here. We're talking about reinforcement learning. We're talking about the agents learning from the reasoning traces of humans when there's an exception. So this is what we see. George has done tremendous amount of work and thinking and testing on this model. So George, I'd like you to pick it up from there and take us deeper into this concept. >> Okay. So when you and I first introduced this stack diagram a year ago, the goal was to synthesize vendor and customer activity into something simpler, where we could group things and make sense.

10:48 So the core was this system of intelligence, which is the green layer in the middle. The idea was that, in order to support agents and humans who are now driving end-to-end business outcomes, not functional siloed activity, but like order to cash or procure to pay or that sort of thing, we needed a layer to bridge the silos of operational apps and analytic data. And even analytic data, when we talk about data platforms, those are still silos because the data that's in there, when you aggregate it into cubes for business intelligence, those are silos.

11:27 The metrics and dimensions silo the data. And as you were saying before, that tells you what happened. The data platforms are usually the historical source of truth. The operational applications are the real-time source of truth and both of them are really just snapshots of people and resources. The system of intelligence above those, this new layer, it harmonizes and treats the business rules that, today, are entangled in the operational apps as assets.

12:00 And just the way we started to treat data as an asset many decades ago, we now treat rules as an asset. And once we harmonize that, those rules, we can ask questions or answer questions like, why did something happen? What's likely to happen? What should we do? And that becomes a digital twin, like a digital twin of the enterprise and the state of the enterprise. What happened? Again, what's likely to happen? What should we do? And that's what allows agents to be productive and effective because they use the system of intelligence to perceive about the state of the business, to reason, to decide, to act and to learn.

12:41 And that learning loop is critical. That's something we haven't had before because that sets up the compounding intelligence value and the winner take most economics. But we've gone beyond this, and maybe we should go to the next slide. >> Actually, before you do that ... >> Yeah. - Actually, Alex, bring that slide back up because I want to double click on some of this. So George, that green layer, the SOI, that is, we've said, going to be the most important piece of real estate in the new software stack and it doesn't exist today.

13:17 We often use several examples of organizations that have something substantially similar to this that harmonizes data and ingests process knowledge, harmonizes that. The typical example we use is Palantir Ontology. We talk about Celonis, we talk about Salesforce Data Cloud or Data 360, they call it now. ServiceNow has aspirations around this. Virtually every software company that is probably going to survive this SaaSpocalypse is going to be thinking about that layer, partnering with LLM vendors in a way where they can add value and determinism to that equation, but A, did I get that right?

14:05 And are those the right examples? And how do you se that actually manifesting itself? Will it be within more application silos? Will it be across the application silos? How will that occur? >> Yeah. So far, you were right to bring this back to particular vendors. So far, only Palantir has had a lot of success trying to build out that layer, but right now, currently the way Palantir is doing it's as if they were SAP 30, 40 years ago, but building out the application R/3 at the time for each customer.

14:46 In other words, it's not packaged and repeatable. So it's very expensive and they're Forward Deployed Engineers, which everyone's copying. The economics right now are very expensive. Each forward deployed engineer, I believe, is built out at the equivalent of roughly 950,000 a year and they're building the application for the customer. The question is can they take that work and make it repeatable and make the economics accessible, broadly? That's what Celonis is trying to do.

15:20 And then you look at, actually, what Salesforce has built with their Headless 360, which they announced at their developer conference a couple of weeks ago That is taking all the Salesforce applications and data cloud and saying any agent can access without the user interface. Now, the reason this is important, everyone's jockeying to be this system of intelligence, that layer, because if you are that layer, you have visibility into the work that the agents do.

15:55 You ideally drive outcomes and have visibility into customer support cases resolved or a customer upsold. And if you have visibility into that, you have much better pricing power. If you are just a feeder into that layer, let's say you're just data cloud for customer, you're just data cloud for customer data and you're the SAP business data cloud for back office data and let's say Palantir then has the layer above it, then your pricing is probably limited mostly to consumption, some variation of utility pricing, which is where the data platforms are today.

16:45 You need to be the one driving outcomes to have pricing power and margins and real, lasting differentiation. But let's, you were- >> One other point I want to put on this. Aaron Levie, a couple of weeks ago, put out a tweet saying, "Forward Deployed Engineers are going to be the hottest job in the future. " It reminds me of the data scientist tweet from the big data days. When you talk to these Forward Deployed Engineers, they will tell you it's a combination of art and science.

17:16 They will say the most important thing is really understanding the business need, the business requirement, not paving the cow path over existing processes, but really trying to reimagine what the most effective way to solve the problem is, number one. Number two, connecting to the right data sources and then obviously making sure that those data sources are governed properly. You see organizations or technology companies, some of them are trying to govern at the agent level.

17:49 I suspect that the right way to do this is to govern and secure and determine identity at the data level in combination with the agents, because you may not know, you may not have full visibility on all the agents. So these are the things that the Forward Deployed Engineers can go. The other thing they'll tell you is the most important thing to get it right beyond that is the system of engagement, i. e. the reinforcement learning, looping back and training, essentially, the system how to work with things like exceptions and how to adjudicate.

18:28 And so that, George, is, right now, a lot of heavy lifting. What we're envisioning is solutions. And this is why we say it's going to take the better part of a decade for this to play out because there aren't a lot of solutions. They will evolve, if I'm hearing you correctly, within these application silos, which is potentially problematic, is it not? >> Yeah. Right now, people are adding agents to the application silos, whether it's Workday or SAP or even Salesforce in their current configurations.

19:01 What we need is a layer that bridges them, because the whole goal is to be able to design processes that work end to end. So you're not optimizing within some local function. You want to be able to optimize activity across the enterprise. And you raised, actually, another point about the system of engagement that was very important and this is the learning loops, which is when an agent tries to execute a business process and it can't finish it on its own.

19:43 This is the equivalent of Tesla disengaging the autopilot and asking the driver to take over. That's a teachable moment. That's how the system gets smarter and that's what the system of engagement is doing. That's where the human enters and supervises and teaches and then that allows the system to have one more edge case that it knows how to handle on its own. That's the flywheel.

20:14 >> I want to make one more point because you brought up Headless 360. About a year and a half ago, maybe a little more, George came to me with this great idea for a Breaking Analysis, which was the face off between Benioff and Nadella. You may recall Satya Nadella or Benioff called Microsoft Copilot Clippy. Nadella's response that the future is going to be agents talking to a CRUD database, create, read, update, delete, implying, before the SaaSpocalypse, that SaaS was not going to be necessary.

20:46 The irony is twofold. One is eventually, the market caught up to it and we're seeing that in SaaS valuations, although they've done better recently. But the second irony is Satya, as we wrote, as George and I covered in a Breaking Analysis, his personal productivity apps were highly exposed because all the file formats were available for Anthropic to just go in and take over. And so injecting Copilot, embedding it into the individual applications is probably not the right model.

21:20 We've seen Salesforce take the move of making Salesforce Headless, meaning their pricing model is going to completely change. It's not necessarily seat-based through the application anymore. It is now, but it will increasingly be charging for consumption of the data, presumably. My understanding, and George, and you may or may not know this, is that Workday has chosen much more of an initial copilot model, which is probably not going to fly. Software companies and SaaS companies, specifically, are trying to figure this all out.

21:50 Any thoughts on that, George, to add? >> Yeah. No, Dave, you did tee it up right, which is like the Copilots in the SaaS apps today are the logical next step, but they're not the endpoint. And what you were talking about with the Copilot being a CRUD database, let's tee this up in order, which is it turned out when Nadella was trolling Benioff, because Benioff had trolled Nadella with calling Copilot Clippy, which was a disastrous little assistant that Microsoft had created 10, 15 years ago, something like that.

22:30 But the irony was ... So Nadella trolling Benioff said, "SaaS is going to be irrelevant. You're just going to have agents reading, writing/directly to the database," meaning you're not going to need your SaaS apps. But when you have hundreds or thousands or tens of thousands of people sharing an application, they have to go through an application. The agents have to go through an application to mediate all that activity. But the irony was that when you're using your personal productivity apps, that's when you can read and write directly to this CRUD file format.

23:08 That's why Office was more exposed first. That was like the worm turned and bit Nadella first. But the point about Workday and Salesforce is you can put agents into these silos and make them more productive for now, but that only gets you so far. You still need something that allows you to manage business processes end to end. And back to your point about the Forward Deployed Engineers and building the custom processes for every enterprise, you can't start at the bare metal.

23:49 You need higher level building blocks for the common processes. And that's what, ideally, Palantir would be building towards. That's what Celonis is trying to build. And then you have the Forward Deployed Engineers build the business-specific outcomes or processes, not the generic ones, but the ones that differentiate how a business works. We don't have that starting point right now, where people are starting from scratch with every enterprise and that's why the economics aren't really there.

24:27 It's not all that accessible. >> So one more time, Alex, if you would bring up that same slide, I want to share something that I've been learning in some of the events that theCUBE has been at. I'm going to start with Veeam. Anand Eswaran at VeeamON put forth the premise that there's a missing layer in the AI stack, and that layer is data and AI trust. Now that was obviously very Veeam- centric because that's what they do.

24:59 They do backup and recovery and they bought a company called Securiti. The reason I bring that up is because we often talk about knowledge graphs as a key layer and we use, oftentimes, RelationalAI as one of those other companies that is dancing around this green layer of the SOI, the system of intelligence. So what I shared with Veeam was you're not the missing layer. We're going to talk about this in a moment, supporting infrastructure around that missing layer for compliance and privacy and security, et cetera, but this is the missing layer, that system of intelligence, without which you're not going to be able to engage agents confidently.

25:44 So let's move on to the next slide here, which is the North Star. And the North Star is a digital representation of your enterprise, a digital twin. And there's a stack that we imagine emerging from that digital twin that brings together the deterministic piece, which is the bottom yellow, with the cognitive piece, what we know as LLMs and the hallucinations, but the intelligence there to eliminate those hallucinations is much more than hallucinations now.

26:20 It's running businesses. Bringing those two worlds together. And George, this is something the previous chart we put out several, probably well over a year ago in some form, others have certainly been talking about it. This is novel. People maybe talk about digital twins. What George has done is he's put together a very detailed description of each of these layers. He also, we showed this a couple of Breaking Analysis ago, around Google Next, the steps to get there.

26:51 Jeffrey Moore has been pushing George to tell us how we get from point A to point B. George came up with a nine step process and how that's going to evolve. And again, this is why we say it's going to take a better part of a decade for this to evolve. A lot of times the vendor community is saying, "No, no, it's here today. " It's not here today. You talk to organizations in terms of what they're doing, they're very much struggling with this.

27:10 So George, here's the picture of the full stack digital twin, the expertise refinery, your manufacturing intelligence in the AI factories and then refining it and monetizing it. Take us through your vision here as to what this looks like in the future. >> Okay. When we put that original vision of the system of intelligence, the system of engagement, system of agency together a year ago, we were really talking about harmonizing the deterministic applications, operational applications, SAP, Workday, Salesforce, and all the line of business apps, these are all deterministic.

27:52 And so we were talking about harmonizing the deterministic rules about how the business is supposed to run. And then all of a sudden this January, there was a whole lot of chatter about something called the context graph, which is how do you make decisions when the business rules break down, when there are no rules or when they conflict? This is all about the tacit knowledge, the tribal knowledge. And really for six months, those have been disconnected.

28:24 The deterministic digital twin and then this context graph about the why behind decisions, the tacit knowledge, where the rules break down, they haven't been integrated. And the point we wanted to make was that you can't really build that context graph, which we're calling the cognitive twin, unless it's integrated with the deterministic twin. Because if you look at all the talk about the context graph, it's all based on context that's captured in the deterministic twin.

28:59 It's like, what is the state of the business that allows me to reason through a decision? And without going through every layer, what we're really trying to say is that, right now, if you look at the frontier labs, they're spending billions of dollars on reasoning traces so that they can help the models learn, in a particular domain, how experts make decisions.

29:30 This is what's they're paying firms like Mercor, Surge and Scale, billions of dollars and they're expecting that to grow dramatically in line with their training costs. It's really amazing, but you don't see enterprises doing this, or only at a very small scale because all the discussion about the context graph, how to capture decision traces has been, they've been cheap, inexpensive compromises because the cost of teaching is too high.

30:02 And our contention is that if you set up that teaching in the context of the deterministic twin, business process by business process, you can cut the teaching cost by two thirds. And then if you get this gold standard reasoning trace, where a human expert has said, "Here's how I make a decision. " Let's say they're in a bank, they've underwritten a bond and the issuer has violated some of the covenants one quarter in their reporting.

30:43 They've violated the covenants determining what they're supposed to, the conditions behind their earnings and payments on the bond. You want something that an agent can reason through the evidence and say, "I know how to make this decision without a human. " It's not cookie cutter. I've seen this decision 100 times. You need a reason trace that allows the model to say, "Oh, I understand how this reasoning process goes.

31:23 And even if it's not exactly the same situation, I can make sense out of it. " And most of the context graphs have not done that. And what we're saying is if you put the two together, you could have much higher quality reasoning traces and that that allows you to capture not just the deterministic twin, but it allows you to capture the tacit knowledge that has really been the dark matter in the enterprise. And we're also saying you don't do this end to end.

31:53 You pick off one particular business process and then you do the rules, the deterministic rules, and you do the teaching for the reason traces that capture the high quality reasoning process. Anyway, by putting those together, if you integrate the deterministic twin and this context graph, this is what allows you to get what we expect to see with the Snowflake and Databricks, where you start seeing what they're capturing in their observability and in their evals and in their ...

32:43 Site reliability, the equivalent of agent reliability engineering. All that comes together in this what's going to radically remake the data platform. And anyway, so that's context for a preview of what we're going to see and it's a very different technology. It's not just a relational database scaled out. The amount of data you have to hold to capture, think of all the data that Tesla captures for all its cars.

33:19 Those are the equivalent of reasoning traces. And then your learning loop is every time the driver had to disengage the autopilot, that's the teachable moment. That's a new analytics. And then you have to learn from those teachable moments and then retrain the agents. That's a very different data platform. And the amount of data that you capture probably grows by a much greater factor than when we went from the Teradatas to the Hadoops and data platforms.

33:55 When we went from Teradata, we were just capturing transactions and we were looking at them for analytics. But when we went to big data, it was because we had to analyze clickStreams and we couldn't do that at the Teradata scale or Teradata cost. We needed something much bigger and much lower cost. And the same thing is going to happen as we capture the exhaust from agents and try and learn from them. The data volumes and the analytics, the data volumes grow by orders of magnitude, the analytics changes.

34:30 That's a bit of the preview that we're going to start to talk about next week. >> Okay. I want you to bring that back up if you would, Alex, because I want to dig into this a little more. For those on podcasts just listening, we've got five layers here. Layer one is the mapping layer and that's essentially, I believe, largely a data layer, where you're getting this single access to a single, you call it single canonical business object.

34:58 So it's across dozens of disconnected applications. So that begins to set up the single version of the truth. And then the next layer, layer two, is that rules layer and this is where the business process comes in. And this is still the deterministic level, which as we've shown here in yellow. And then up top, now you've got that tribal knowledge, the institutional memory is layer three and this is where people, why?

35:29 Why do we do it this way? Well, we've always done it this way. That might be part of the institutional memory, but maybe there's a better way. And so that brings us to layer four, which is the decision guidance. So you're taking that institutional tribal knowledge and then you're synthesizing that into specific recommendations and predictions on likely outcome. >> For that particular process. Yes, that's a recommendation for how should we handle this situation?

36:00 That's that layer four. >> So it's real-time advice. And then at the top is that learning piece, that feedback. Now this is the ideal state. This is the North Star. Nobody's really here yet, but we're starting to see examples. And I want to come back to some of the things that I was talking about earlier, things that we've been hearing at various events. So we challenge Google on this. This is the North Star. You announced Google, at Google Next, something called a Knowledge Catalog.

36:29 It wasn't even called a knowledge graph and it was largely the technical and operational metadata associated with the systems running on Google, but they fully had said, "This is where we're going. We are going to address the underlying application logic and process knowledge, and that's where we ultimately want to go in our full stack. " So eventually that'll evolve, presumably, into a knowledge graph and then ultimately this full stack digital twin. Veeam, again, saying that it is the missing layer, I tried to push them on, well, you've got this arrow here around scaffolding.

37:09 I was calling that Veeam piece the trust layer, compliance, governance, the data security, the ability to recover. I was calling that the scaffolding. I know there's industry terminology that maybe is different, George, but double click on the scaffolding. What is in the scaffolding? Would what I'm describing with Veeam, Dell recently made an acquisition of Dataloop. It was also as a knowledge graph, although they haven't yet applied it to their data protection business, but I got to believe Rubrik and Cohesity and others are thinking about this, and Veeam is actually marketing to it.

37:49 Commvault as well with their ResOps. But what's in that scaffolding? Is governance in that scaffolding? Help us understand that. >> Yeah. Okay, this is a great question. So teeing it up as the migration path, which we're not really showing today, and maybe we'll get into next week, but at the most basic level, you could think of your business intelligence metrics and dimensions, which is what customers want to talk about right now, which is like, help me just standardize how I look at my analytic data.

38:22 Those are called dimensional semantics. It's like common way of looking at sales and bookings or commercial remaining obligations. >> RPO. - Yeah, RPO, what we used to call bookings. You want to standardize how you define those and how you calculate those things. But then, like what we saw at Google, their Knowledge Catalog was starting to capture metadata about rules that cover slightly more of the operational semantics.

39:06 But what we're talking about here is that state where you're not separating the rules from the data, but when you combine them, when the rules are stateful, that sounds awfully abstract, but it's like when you combine how you're defining the rules about how the business is supposed to run with the data about the current state of the business, that's when you get to the digital twin. In other words, when you just have a definition, it's in a catalog and it's just a definition.

39:41 It's just a set of rules, but it's separate from the data that says this is the state of the business. And you can't then, when they're separate, you can't ask questions and answer questions like, "Why did this thing happen and what's likely to happen and what should we do? " That's not yet possible. >> Okay. And thank you for that. That's helpful. And I want to drink deeper on the governance piece because I'm still unclear on where that sits in this picture.

40:14 >> Okay. When you can marry the rules and the state, when you're modeling the business, when it's a live definition, you're essentially encoding business policy. And today governance is usually about resources, like who has access to this data, ideally at the column level or at the row level.

40:47 Who's allowed to see this? But when you have agents, it doesn't work that way because you don't know exactly what the agent is going to do ahead of time. That's the whole value of the agent. So you need rules, you need governance that are adapted, essentially, to intent. Like agents are allowed to do this sort of thing and that policy is encoded in this digital twin that says, " An agent that is performing this activity is allowed to handle these sorts of things or allowed to do these sorts of things, invoke these tools and take these actions.

41:33 " That governance becomes policy that's encoded, but it's governance not tied to a resource, it's governance tied to intent, which is what semantics and agents are about. >> And ultimately that intent is informed by the data. >> Yes. - Is that correct? Yeah. >> Yes. - And so that has been my premise, is that you want to govern at the data level, not at the agent level.

42:06 Obviously, you want to have an agent control framework, but the way to do that is you've got to do so at the data level or you're not going to have full scope. >> But you're not tying governance to the data. I mean, there is governance policy, like who's allowed to see what. You don't give up resource-based governance, but you have a new model on top of it, which is basically intent or action-oriented, which is it's not tied you can or can't do this action, but it's more like, within this action space, what are agents allowed to do?

42:54 I'm not crisp enough about it in articulating it, but it's a different model that has to be layered on top. >> So Alex, bring that slide up one more time. So I just want to come back to state. So again, I got this recency bias with VeeamON and last week at Dell Tech World, but what I said to these guys is that eventually this digital twin of your business, the digital representation of your business, people, places, things, activities, represents the state of your business at any point in time.

43:29 It's organic, it's growing, it's changing, it's guided by the dimensions and the metrics, which becomes a new organizational structure, essentially. But backup and recovery and resilience, business resilience will, at some point, come down to the ability to recover that state at a very granular level. So it's going to eventually require knowledge of that state. Today I can recover. I take a snapshot of my business.

44:00 I can do so at a very granular level. I can recover data or a file at some point and my RPO is maybe very small, five minutes or an hour or a day. State of the business, my premises, will have a similar level of requirement, where you will have to recover the state at a particular time and you will choose, based on your budget, based on the importance of the business, how much data you are willing to lose or how much of the state of the business you are willing to have to recreate.

44:32 >> Okay. - Please, George. >> You're getting at something really, really deep, which is, and I was alluding to it when we were talking about how agent development with evals and observability and agent reliability, they're all merging and it gets right to what you're getting at, which is the agent observability captures the traces of agent activity, of this. It captures what the agent was reasoning about, the state of the business that drove that reasoning, the tool calls, what the tools returned when they recall.

45:11 That is the new source of truth about the state of the business because in the old world, you had that in a write-ahead log in the DBMS. In the data platform, it was the data coming in. The real source of truth wasn't so much in the tables in the data, it was in the write-ahead log, which determined what was going to be in the tables. If you had to backup and recover, if something failed, it was the write-ahead log that was the real source of truth.

45:46 That new write- ahead log is this observability stack that, as I was saying, was capturing the state of the business, what made the agent reason about that, the calls and everything and the volume of that data is immense and it has to be cryptographically encoded and verifiable because if you want to prove what was the state of the business when a decision was made, you need that new observability platform.

46:17 And that is why what we see as data platforms today, the Snowflakes and the Databricks, are going to change dramatically because it's no longer just show me theCUBEs or the transaction aggregates or what did people click on. It's now what did the agents do? And that observability exhaust is how you learn from, how you prove what happened.

46:54 That is the new source of truth. >> Okay. Again, I want to put a finer point to that. What I heard from Veeam is they can tell you what the agent did, when it did it, who it did it to, et cetera, who had access to it, just the lineage of what occurred, the audit trail, if you will. What they can't tell you is the why and what the reasoning was. And so recovering the state of the enterprise requires you to recover the why so that agents can figure that out and then continue to take action.

47:32 And that is, again, it's futuristic, but it is going to be, if what we say about the operational model and the business model and the technology model is all true and the AI promise, this is going to ultimately be a requirement. And I suspect that companies like Veeam and Rubrik and Cohesity and Commvault and Dell and others are going to be tasked with doing this, or there's somebody else will emerge that will do it. Okay, let's come back to that whole business model, technology model, operational model.

48:09 Jeff Clarke, in his keynote this week at Dell Tech World, said a couple of things that caught my attention. He said that the cost per token dropped 80%. I think, George, you said it dropped probably 90% this past year. That while token spending is going up, but the cost per token is declining and we expect that to continue to decline as NVIDIA brings out more and more capable systems every year.

48:42 The other thing he said was that, but the consumption of tokens grew 320X. So without describing or saying the word Jevons' paradox, he was referring to it. Prices go down or costs go down, people consume more. This goes back to William Stanley Jevons who wrote a book about coal, and so look that up. But anyway, Alex, you bring up the next slide. So the other thing that Jeff Clark talked about is that he said that the tokens will be a line item on your P&L and you will be completely changing your operating model.

49:23 Now, Dell is a company that has leaned in heavily and invested heavily in AI. And what we're showing here is just, again, theoretical, but we're talking about a big boost in productivity. You're either going to get that from spending money on putting in AI factories or you're going to be tapping intelligence in the form of tokens through APIs or neoclouds, et cetera. But the whole point is you'll be scaling with less labor, whether it's double or triple or 1.

49:49 5X your revenue. The goal is to do that to scale, without labor. On the left-hand side, we're looking at the AI factory. We see, as we talked about last week, the x86 getting absorbed into that AI factory. The functions of and security and networking are going to the x86 functions will get absorbed. The AI factory will be producing that intelligence. And going back to George's stack that we showed in the last couple of slides, ideally the AI will be coordinating all the stuff that humans had been coordinating and reconciling and adjudicating and recovering, et cetera, using that system of intelligence, that semantic layer, which automates that glue.

50:39 So this is a very high level view, George, of where we think the industry is going. And again, this accrues advantage, confers advantage to the buyers of technology and it changes their business model and their operating model. I want to give you one simple example that I heard at Dell, and it came from Doug Schmitt. I was pushing him on, "Okay, Jeff Clark talked about the operating model. How has it affected your operating model?

51:10 " And he basically, I'll give it in my words, it's not exact, not verbatim, but basically Doug Schmitt is the CIO and he runs the services organization at Dell. And he said, "Well, here's something you could relate to. " And I'll, again, tell it my way. This is not verbatim, but essentially we can all relate to, we sat around in a meeting and you had the logistics people there, you had the parts people, you had the finance people, you had the sales, somebody from go to market or marketing.

51:33 And there was a question on the table and somebody presented data and then people started to question the data. "Where'd this data come from? I have different data. " And then somebody took the task or assigned the task, go off and get the right data. Come back next week at our weekly meeting. You come back next week. "No, that's the wrong data. " They present the data. Go, "That's not the data we're talking about. This is what we're talking about.

51:54 Let's do it through email now. " So two weeks go by before they get to the truth. He said, what Dell has done, and they've talked about this publicly, they built a data mesh, which is their term for connecting all their data together. They now have their "single version of the truth. " What they do today in the meetings is they prompt engineer their proprietary system and shape it so that they can get to the right answer in real time.

52:23 "Oh no, we don't want city level data. We want county level. " Or they just reprompt and reprompt at all that time the system is learning, but they get to the answer within, let's call it 15, 20 minutes, half hour, I don't know what it is, but they consolidate, they compress what used to take weeks or days down to real-time. I thought that was a good example, George, but I'd like you to pick it up from there and give us your perspective on how these three models, the operating model, the technology model and the business model, come together.

52:57 >> Okay. What you were talking about is a great example of the difference between a bunch of individuals who are highly productive and an organization that has a shared model of the state of the business that can help orient everyone's activity towards collective outcomes. In other words, how do we get everyone's activities aligned in pursuit of the business's goals?

53:29 Let me add to your anecdote, another one. There was an article that was posted on Sequoia's website that got a lot of attention a few weeks ago about, it was Jack Dorsey from Block had written a great history of the organization, the hierarchical organization. And what he was saying was, this was in support of his riff of downsizing Block by about 40% because he says, "The hierarchical org chart is no longer necessary when you have an intelligence that binds the organization, that's a shared truth.

54:10 " What he didn't get into in the article was that's what we're calling the system of intelligence, where it's a platform and a platform's core function is matching. And what becomes possible is for every activity and every business process, you have information that's needed, you have agents and you have, then, human expertise that's drawn in where the encoded expertise doesn't exist.

54:44 And then the system of intelligence matches all that with the outcome that's needed. The inputs are the information, the agents and human expertise, as needed. And when you need to achieve an outcome, this end-to- end platform, the system of intelligence matches all those inputs to achieve that. And what replaces the org chart is a hierarchy of business metrics. At the very bottom is activity level metrics and then there're business process level metrics.

55:16 And at the top, there's North Star metrics. And the point of this is there's another myth that's been going around for the last, I don't know how many months, that with agents, we're going to see the rise of the single person billion-dollar company. And this sounds a lot like that story that was going around at the rise of the web, which was the web would enable the rise of the solo merchant. And actually what we got was Amazon.

55:47 And the reason is there's a difference between personal productivity and a platform that organizes the collective efforts of a large number of people. That's the difference between a marketplace of contractors and the coordination, planning, control and resource allocation for collective outcomes that is what defines a firm. And I actually think that rather than single person billion- dollar companies with a lot of agents, that's personal productivity, we're actually likely to see winning firms that are orders of magnitude bigger, not smaller, because they can use the system of intelligence to coordinate agents and people through the system of engagement with collective activity that's just not possible with an org chart.

56:40 Whether it's a hierarchical org chart, a matrix org chart, divisional, we're going to see much, much larger companies. We're going to see platforms because that's what this software becomes. This ability to match the people and resources with these outcomes with this encoded expertise, it becomes software platforms. That's why when you teed up at the very beginning, your intro was this doesn't just affect the vendors, this affects the customers because now, rather than just seeing consumer online firms becoming platforms, we're going to see every business become a platform.

57:21 This may take a long time, but that's the end state. That's where you have high fixed costs in terms of setting up the system of intelligence to extract, refine, and deploy expertise as capital. You have low and declining marginal costs, because rather than costs rising with revenue, you have human expertise that gets pulled and matched and it's only used on edge cases. And this is where digital expertise gets encoded and that's what gives you low and declining marginal costs.

57:54 And then you get the network effects because with every transaction, with every business process that's encoded about the rules, what must happen, it adds value to all the others, with every expert teaching session to capture expertise, this is the why behind decisions that adds value to the network of existing expertise, and then with every transaction, your predictive models get better at seeing patterns and making more accurate prescriptive actions.

58:27 And all this compounds. So the economic model of every industry changes as it becomes more like a platform. >> And so we'll leave you with this notion that we talked about the technology model in the previous two slides that we're going to dig in deeper next week, but also we're going to dig into the business model and the operational model, aspects of it. As Jeff Clark said, that tokens are going to be a line item on your P&L, you're going to be tapping into those tokens, into the AI factory, or you're going to be building your own AI factories and spending that capital.

59:05 And then the operating model will have platform dynamics and those will accrue not just to technology vendors, not just to SaaS vendors, but every company out there. And George, thank you. I'm looking forward to next week. Why don't you give us a little teaser for it and where you see that going? >> I guess the teaser is that we're previewing the transformation of the data platform to see how far is it going to go into transforming into the system of intelligence?

59:39 Observability. Observability used to mean like Datadog, but now we need Datadog for agents and that's going to stretch the data volumes orders of magnitude, and that gets into your new system of truth that you were alluding to. And then that also becomes the substrate for the learning loop for agents, the reinforcement learning and also becomes how the CI/CD release process for agents, they have to go through the same rubrics and evals and judges, LLMs as judges that manage your observability.

60:22 Then there's a new form of data engineering and you're not just engineering operational data and turning it into analytic data, you're taking your knowledge assets and extracting structure from them. These are some of the things that we're going to talk about next week that's remaking the layers of the data platform and seeing to what extent it'll turn into the system of intelligence. And if it's not with the data platform vendors, then it's going to be other vendors.

60:51 >> All right George, we're looking forward to that. It's going to be our preview to Snowflake Summit, which, of course, is a preview to Databricks Data + AI Summit. Thank you, George, for spending some time with us on Breaking Analysis. >> Okay. Thanks, Dave. - All right. Thank you for watching. And Alex, thank you for hanging with us and we'll see you next week on Breaking Analysis.