316 | Breaking Analysis | Personal Agents Light the Fuse In the Age of Data Intelligence

0:00 >> From theCUBE's Studios in Palo Alto and Boston, bringing you data-driven insights from theCUBE and ETR, this is Breaking Analysis with Dave Vellante. >> The AI wave is starting to look a little bit like the PC era with some obvious differences, but we're talking about personal productivity, bottoms up and decentralized funding, and individuals taking control of their own work with agents, open tools and repeatable skills, much like power users once did with spreadsheets and PCs.

0:33 But unlike the PC era, these agents don't just create documents and dashboards, they act. So if every person and every department and every vendor builds its own island of intelligence, while this might be necessary in the journey to the end state of AI, it might be a necessary step, but we risk recreating the enterprise silo problem that has plagued software for decades. Only this time, it's going to happen faster with greater liabilities.

1:04 So the premise of this Breaking Analysis is that while the impetus for AI initially came from CEOs top down CEOs and boards of directors, the first phase of enterprise AI is largely taking shape and being driven by individuals and personal agents that are being deployed. But we think that the sustainable value is going to accrue to the platforms that organize enterprise knowledge into a true system of intelligence, what we call the SOI.

1:36 Now, ahead of Snowflake Summit and Databricks Data + AI Summit, we're going to focus in on these two companies in context of the industry. As we wrote about a year ago and a previous episode, almost exactly a year ago, both of those companies, Snowflake and Databricks, they've crossed the Rubicon. They're no longer just data platforms serving only analytics. They're moving into the layer where enterprise data, trust, context, actions, and even business logic become human readable, agent readable, and eventually executable by agents.

2:16 That same control point is being pursued by others, namely application vendors, hyperscalers, and even frontier model companies, because whoever helps organizations best model the business is going to shape how agents reason, how they decide and how they act. And this transformation, we think it's going to happen in two motions simultaneously, bottom up individuals that are going to build personal agents and skills that immediately improve productivity, and then top down.

2:48 Leadership has to guide that energy toward an AI native architecture so that those skills become governed assets that are connected to a shared ontology, not just another generation of disconnected tools. We don't want that. So the caution, by the way, is model makers, the big LLM vendors are going to try to bundle the model, the agent, the interface, and their own version of enterprise memory. We're seeing that early stages of that already. Frontier models, they're very, very important.

3:19 They're critical in fact. But owning the model is not the same as owning the enterprise operating context. So the system of intelligence is the world that the agent lives in. The model, think of that as the engine that reasons inside that world. And in this episode of Breaking Analysis, we're going to dig deeper into the emerging AI software stack and prescribe the steps as to how enterprises can evolve this model and what to watch out for and how they can be most effective.

3:49 Welcome George Gilbert, the architect of this enterprise model and deep thinker on this topic. George, thanks for coming on. >> Good to be with you, Dave, as always. >> All right, awesome. Let's bring up the first slide here and start by reviewing yet again the emerging AI software stack. Now the concepts behind this model are borrowed from Geoffrey Moore, who's been working with George to really envision the future state of the enterprise and how to get from where we are today to where we're going.

4:16 And George, we talked about this last week in Breaking Analysis, but set up the puzzle pieces of this AI stack that we're showing here. >> Okay. So let's start at the left. In the purple box, the system of engagement, which spans the system of agency, the system of intelligence and the data platforms, and frankly, the systems of record, this seems abstract, but what's coming into view is something very clear and very profound and it makes the rest of the architecture clear.

4:49 This system of engagement is the new front end. It's like Windows was, or it's like the browser, or even the smartphone. And the reason I'm saying it's that significant a transition is because each of those drove a transformation in the backend. Windows drove the need, ultimately, for servers, web browsers drove the need for highly scalable web backends, and then web and mobile really drove the need for the cloud.

5:28 And that's the premise that we're setting up today, which is to feed this new intelligent system of engagement, you're going to need a new backend. And that's the system of intelligence that we're going to talk about. And then the system of agency that sits on top, that understands the state of the business through the system of intelligence and then it operationalizes any decisions through that system of intelligence. The reason there's a system of intelligence is that this is where you harmonize all those islands of operational data and analytic data that's grown up over 60 years.

6:09 Humans can bridge those islands. That's what departments are and functional workgroups and matrix teams, they bridge those. Agents need something much more harmonized. They're not as tolerant of that ambiguity. So if you want to organize armies of people and agents toward collective outcomes, you need this new harmonized platform where you treat the business logic as an asset just the way, over the last 40 years, we've treated data as an asset.

6:47 >> Okay. I want to just point out that, Tony, if you'd actually bring that slide back up if you would, that left-hand side, that system of engagement, as we said last week, that's not the social media think system of engagement. It's the feedback loop. It's like George said, this kind of new browser that's going to feed into the system of intelligence so that the system of agency can learn and take action. Thank you, Tony. Okay. So it's useful to map existing software products into this structure that George just showed, this emerging AI software stack, to really help folks understand who's playing where.

7:31 And you'll see there are shifts, as we said. Companies like Snowflake and Databricks have crossed the Rubicon into new territory. So Tony, if you bring up the next chart, let's do that exactly. Here we're showing some of the leading players. George, we're setting up for Snowflake Summit and Databricks Data + Data AI here in June. So let's start on the bottom with the modern data stack shown here as data platform. But as we said in the open, both of these companies, they've crossed the Rubicon into new territory encroaching on BI, business intelligence, plugging into the system of record that we show in the bottom right, moving up into the governance layer with their catalogs.

8:08 We've talked a lot in Breaking Analysis about Horizon and Polaris and Unity. And to the left, that system of engagement, Snowflake Intelligence, Databricks Genie are encroaching there to allow talking to your data essentially. So you can see them expanding their aspirations. But George, walk us through this chart and the vendor mappings. >> Okay. So let's start again with the system of engagement all the way on the left because I think it becomes clear how everyone's coming together.

8:44 ChatGPT is actually going to become Codex just the way Claude has transformed into Cowork. Cowork is built on their coding agent, because when you want the agent to act on your behalf, it emits code to manipulate other tools, to manipulate other applications, to access data. So the original chat application, in order to turn it into an agent, it's actually going to be built on a coding agent harness.

9:16 So that's what that Cowork is for Anthropic and what Codex is becoming, ChatGPT is going to grow into Codex. Now, the reason I start with those is because they don't have anything structured. They don't have structured data to feed the intelligence behind them. They're good for either generating code or doing a web search. But if you want to go search your enterprise knowledge base or your enterprise data, you need that structured in a new way as a backend.

9:52 And that's what the data platform vendors, Snowflake and Databricks, are offering, they're starting to offer. Snowflake Intelligence is their front end. And I believe, and we'll find out next week, I believe it's built on Claude and the Claude harness. I don't think it's the raw Claude Anthropic model. I think it's the coding harness around it. And Genie might be a little different. But again, both of these are front ends that are first oriented around how do I go into the database backend and query data so that you can talk to your data in natural language just the way the other products allow you to talk to the model or talk to a web search.

10:40 Now, all the products on the right side in the green layer, in the system of intelligence, those are organizing your business processes and your personal productivity data into something that your personal agent can interact with. This was the great achievement of Glean where they built a personal graph of all the people and documents you interact with. And that was Microsoft's Graph, which they renamed Work IQ.

11:14 That's the foundation of Microsoft Copilot 365. That's their differentiation. And again, each of these ways of organizing information, those are currently islands. Eventually, we'll figure out some way to bridge them. But when you ask one of these vendors, "Well, why is your system of engagement, your chat interface different from anyone else," sort of like the Passover question, why is this night different from any other night, it's because of how they organize the data on the backend.

11:48 So you're offering up the data with some matzah. So the Microsoft now has something to overlay their data platform called Fabric IQ. That's an ontology, that's a digital twin, which is modeled on Palantir. Palantir's is much more mature. And as we've been talking about for a couple of years, relational AI placed in this similar layer in the system of intelligence where you're harmonizing not just data, but you're harmonizing all the business rules about how the business operates.

12:30 And Salesforce Data Cloud belongs there. So does SAP Business Data Cloud and Celonis. Now, let me just distinguish that from the governance metadata that tries to harmonize what's in the data analytic data platform. What's in Unity Catalog and Snowflake Horizon catalog, those are definitions that's different from having the business process logic modeled as live executable code.

13:01 That means you go to the catalog and it says, "Oh, this means this," but it's not running code. And that's the transition. When you get to running code, that's when it's in a system of intelligence. When it's below that, it's a catalog. So that sets up the taxonomy of who belongs where. And then the systems of agency, they're on top and that's where all the excitement is, but frankly, all the hard work and what differentiates this, each enterprise is in the system of intelligence.

13:36 >> Great. Thank you. And thank you, Tony. Appreciate that. So you can see there are a lot of moving parts in that new AI software stack and a lot of different companies are going at it different ways. We expect that there's going to be a lot of focus on that system of intelligence space. Right now, it's maybe some of the low hanging fruit is that system of agency. We're seeing companies go after the agent control framework, but really it comes back to the data, both governing that data and harmonizing that data.

14:09 So we're going to dig into that a little bit more. And we're going to use this concept of a north star. And Tony, if you bring up the next slide, we've said that the endpoint that we want to get to, of course you're never done in tech, but we envision the north star as a real- time digital representation of an enterprise, i. e. an enterprise digital twin is the term that we use. And it brings together the reliability and determinism that traditional software has provided for decades, those systems of record that we talked about, and even the analytics systems and the BI system.

14:44 Those are well known, they're trusted, and they're deterministic, and it merges those with the creativity and the generative characteristics of probabilistic systems. LLMs are probabilistic. They generate content. As we've pointed out in many previous episodes, a big portion of the activities that enterprises perform today involve human judgment. They're making exceptions, they're adjudicating, they're using that tribal knowledge that's known to a few to make decisions and act.

15:18 So initially we see AI taking on the non-judgmental tasks, the ones that are related to coordination. That's where AI is going to do very well in the near term. But over time, even the judgment oriented tasks are going to be incorporated to a fair degree into this idea of a digital twin. But that's going to take, George, many years to evolve. So pick it up from there and then let's dig deeper into this slide. Take us through it.

15:47 >> Okay. So just to be very clear, what we haven't done much of in the past is unpack the system of intelligence because there's a lot of richness in terms of layers in there. There's five layers. And as you pointed out, and as we originally conceived it, it was just the deterministic business rules that were, as I like to say, embalmed and entombed in all the operational apps. We need to treat those as a harmonized set of assets just the way we treat data.

16:19 That's in the bottom two layers, which I'll explain. The top layers, as you were alluding to, it's all the dark matter of knowledge in the enterprise, like what happens when the rules don't apply, or when they conflict, or when they break down and when they're not cut and dried. So let's start at the bottom two. So the very first layer is I've got hundreds of operational apps. In large enterprises, thousands. And I've got islands of analytic data.

16:50 I need something as a map to abstract all that so that when I refer to customer, it hits all the attributes of customer and all the different systems. That's layer one. It's just mapping. Think of it as a Rosetta Stone. The next layer up is, okay, so now that I have customer and order in one place, now these are the rules about how they interact, how my business must run. I must respect these business rules.

17:23 I cannot extend credit to a customer who is late on their payments and is a low value customer, something like that. Layer three starts getting up into the, how do we make judgment? This is where you start to collect evidence of what's happened before. It's a searchable record of expert reasoning traces that document actions that have been taken, the evidence supporting them, the contextual information surrounding them and the conditions considered during those decisions.

18:03 This is its institutional memory. That's layer three. Now, layer four, Dave, as you were describing, that actually gets synthesized. This is advice. This is when an agent or a human is in a situation about to make a decision. Layer four is where the system of intelligence, in conjunction with perhaps an agent, will synthesize guidance for a decision based on what's happened before in the institutional memory and the state of the business in the deterministic part of the digital twin.

18:43 It's like under these conditions before and the current state of the business, what recommended actions should I take? Then layer five, this is critical, and I don't think this is fully appreciated by at least those who are coming from the data platform up like our old crowd. This is the learning and feedback layer. This is the new big data. We went in, really in the 2010s, starting in the 2010s, we collected all the click streams from websites because we were trying to learn from the behavior of how people interacted with the websites.

19:26 This layer is all the behavioral exhaust of the agents. And as we're going to describe, this behavioral exhaust is the new source of truth about how agents reason and it's how you're going to teach the agents to do better in the future. And the amount of data collected here is 1000 or 100,000 times bigger than what, say, Datadog collected from tracking cloud native applications.

19:56 This is the critical layer. >> Got it. Thank you for that. And we're going to take you through, later on in this episode, how these layers mature. This doesn't happen overnight. This is why we've always said this is going to take a better part of a decade to play out. Some people have pushed back on that telling us, "No, no, it's here today. " It's not here today. This is a vision and the promise of AI.

20:20 And we'll see if our industry can live up to that promise this time around. It's broken many in the past. Now, as we said in our open, we saw CEOs like Michael Dell, certainly Jamie Dimon, Andy Jassy, Sundar Pichai, and many others, they sounded the alarm and initiated a top down transformation driven by AI. And while there are plenty of enterprise initiatives happening right now, what we're seeing is individuals are picking up on AI and they're building agents with things like NemoClaw or Hermes Agent, or others from LLM vendors and other open source tooling and they're making real progress with their personal agenda.

21:06 Meanwhile, let's show on this next slide all the vendors, they kind of have their own ideas as to how this stack evolves, and that creates an interesting dynamic and potential risk where we're just simply creating pockets of intelligence, which are going to impede that digital twin vision that we just talked about. And George, why don't you explain the islands of digital twins that may be a half step that we need to take, maybe how some of the vendors that we're talking to see this evolving, and why ultimately it might have to happen, but we want to get past that and talk about how organizations can evolve to that end state.

21:52 But please, pick it up from there. >> Okay. So this is really important, which is this digital twin or system of intelligence, which it's the backbone of everything because the agents, as we'll explain, really can't comprehend the business and then analyze it and take action without this intelligence layer. We're not saying that's going to emerge in one intergalactic unified step.

22:24 What we're seeing is Snowflake trying to harmonize the data it has and coalesce unstructured information related to the structured analytic data they have. We're seeing Databricks build a similar platform. And what I don't think customers, having spoken to customers, what they don't appreciate fully is that you take the Salesforce Data Cloud or the SAP Business Data Cloud and those are like overlays on part of your analytic data estate.

23:02 They're overlays because they take, in Salesforce's case, all your customer related data. They don't extract it from your analytic data estate. What they do is they put a layer above your analytic data estate and they enrich all your customer related data from their apps, from other apps, and from the data model that they prepare so that you still have the Databricks data, but they essentially synchronize or cache a copy of it into the Salesforce Data Cloud so that when you want to ask sophisticated analytic questions and operationalize agentic actions on your customer related data, you don't have to engineer that whole part of your analytic data estate.

23:55 It's done for you. Same thing with the SAP Business Data Cloud. And so we're going to see these islands, and eventually we'll bridge them, but this is how you start getting there pragmatically. >> Got it, and thank you. But I want to ask you, because right now these SaaS vendors are fighting for their life, fighting for relevance. They are going to try to bring together, partner with the LLM vendors, bring together the determinism that they are so good at with the non-deterministic and probabilistic emerging software.

24:31 They're maybe going to try to grab a piece of that or embed that into their environment. So in their vision of the world, they want to hang on to that install base, they obviously want to try to grow it, but they want to do so within their sort of safe space if you will. So how do you see that evolving? Because they're not going to buy into this enterprise nirvana. Now, maybe customers can force them to get there, but a combination of competitive dynamics or potentially horizontal platforms will emerge.

25:07 But how do you see that playing out, George? >> Okay. So allow me to bring up a potentially irrelevant analogy from when you're a kid when you play that game, when one hand goes from different people, one hand goes above the other. They're trying to make sure that their data layer is visible to their agents and that they're trying to be the ones where, when you want to operationalize, do analytics and operationalize decisions on customer related data or on operational data like SAP, you're using their platform.

25:49 Because if their platform feeds into, let's say a Snowflake or a Databricks, they lose differentiation and pricing power. They become a data feed. If they are the operational platform for agents, then they have visibility into the work product or potentially the outcome of the agent, and their pricing model is completely different. That's when they're charging essentially not for compute consumption, but they're charging closer to an outcome or a unit of work.

26:24 That's the argument about why they want their agents to have exclusive access to their data, or privileged access. So that's all that's going on there. But the great value is in who can model how the business works, and then the scope and fidelity of that model is where the agents will gravitate. >> Okay. And we're seeing, again, companies like Snowflake, and presumably Databricks as well, is benefiting.

26:59 They're not a public company, but Snowflake this past week had a blowout quarter. The stock was up, I don't know, 75 points in two days. So significant diversion from the SaaSpocalypse where Snowflake was sort of thrown in, that they're showing that they can actually monetize AI. What we want to do now is really dig into the details and, Tony, if you bring up the next slide, about how the foundational layers are going to mature.

27:30 And as we show where we see some of the key players here in the context of these layers that George has laid out. George, in this graphic and the one we showed earlier, we talked about the digital twin as the north star. How do you see this evolving? How would these layers mature into an enterprise architecture that can maybe avoid the trap or at least get beyond the trap of silos that you described earlier?

28:01 Walk us through the levels on this chart and how you've laid in some of the vendors here. >> Okay. So what we're doing here is we're unpacking the bottom three layers of the system of intelligence. First, we unpack the system of intelligence itself into five layers. Now, we're seeing how the bottom layers are going to evolve over time, how they're going to mature. And this is crucial because everything depends on the extent to which the scope and fidelity of what's called your data model reflects how your business operates.

28:37 That dictates the sophistication of the analytic questions you can ask, and then the analytic sophistication in turn dictates the confidence the underlying agent has in either recommending or taking action. And so I'm going to go through, in this slide, just the data model first. Then we're going to talk about the analytics and then the actions in two follow-up slides. So in this first one at the bottom layer, you're looking at business intelligence like reporting cubes, metrics and dimensions.

29:12 And Dave, actually what you mentioned with Snowflake's results being so high, it's because when you do agentic query, when you're basically using natural language to talk to your data, you're making it much more accessible for more people to query the data. And instead of having the simple dashboards with pre- compiled queries, this is like the equivalent of going from a quick chat to a deep research report where the agent asks follow-up questions, and so they're consuming hugely more compute per query and there's more people querying because it's easier.

29:57 That's why the results were going up like that. Now, that has implications for cost where someone's budget is going to get busted, but we'll come back to that another time. At level two, this is just where you harmonize it enough so there's one customer everywhere. At level three, I think we're going to hear more about this from Databricks and Snowflake this week. This is when events like real-time updates coming in through the equivalent of an event stream processing service, updates the tables in real time.

30:38 So your e-commerce interactions, as you're on a site, they update the recommended products in real time so that state and recommendations are real-time. The fourth layer, here is where you have behavioral interactions. This is you categorize customer behavior into high value shoppers or fraudsters. Fifth level, this customer is likely to buy.

31:10 This is a probabilistic prediction as data. At the sixth level, the sale is a web of related things. The model represents what is and how it's connected. Maria is a node, the order is a node. The tent that she bought for camping is a SKU. The promotion, the fulfillment center, the carrier, all that are nodes. And these are rich objects that you can query.

31:40 Okay, seventh layer. It starts to get really interesting. The sale that Maria, the purchaser in the last example, is a set of possible moves. The model now represents what can be done at any one point in time. Apply the promo, reserve inventory, authorize payment, split shipments, that sort of thing. It's not buried in API calls. The model knows that authorized payment requires a valid payment method and a held cart.

32:13 Level eight, the sale as it's happening right now all in one place. The model represents the current live state, not a snapshot. So as the sale to Maria unfolds, the model reflects it in seconds. The stock of inventory for the tent at this instant, her cart's exact contents, the payment authorization clearing, the fulfillment center's current load and the carrier's pickup window so that reserved means reserved now, not a picture of what things were last night.

32:51 At the ninth level, this is really in the future, but what we're trying to get to, this is changing how the sale runs. Changing how the sale runs becomes a data update, not a code deploy. As the system gains new power, it can reason about and improve the process itself, not just execute it. >> Excellent. - That's the level of maturity. >> Yeah. So this is next level work, George, that you've done.

33:22 And I want to now circle back to the data foundation, which is, as they say, the fuel for AI. Sorry for the bromide, but it is. And so, Tony, bring up the next slide. This focuses on the analytic data. It speaks to the increasing maturity and sophistication of the way in which the AI is able to exploit and leverage that data for competitive advantage, not just as a static historical version of the truth, i. e. what happened, but going well beyond into more advanced and accurate forecasting and planning.

33:53 So the spectrum, George, goes from level one to level seven to nine and moves from siloed metrics on one end of the spectrum, siloed metrics and dimensions that live inside separate departments moving to up that scale where you just discuss them. Again, a learning system, it's continuously improving based on feedback from humans and maybe other external factors. George, this is a major undertaking for organizations, but you're laying out the maturity model for the data foundation. Walk us through this.

34:28 >> Okay. So just as you were describing, if you distill the basic levels one and two, like Maria, the customer I was talking about, she's aggregated and hidden in a customer dimension, which is like, do gold tier customers who bought from the spring promotion spend more per order than bronze customers? But Maria's lost in there. So what you can't really do is explain why. You can see that Maria bought, maybe in a separate query, that she converts, but there's no sequence so there's no cause and so there's no recommendation.

35:11 That's levels one and two. These are static snapshots. Levels three and four start to get into why did it happen? So this is, with level three events, preserved. Remember this is where we talked about in the last slide, events now are peers to the records in the table. So events preserved in temporal order and they get behavioral labels layered on. So analytics can decompose an outcome into a chain of events that produced the outcome.

35:46 And you can ask, why did Maria's tent purchase convert so fast? And at level four, that pattern gets a name. Maria is a deal-driven replenisher. She got a deal and that's why she converted so fast. So now moving up to levels five and six, this is prescriptive. What should we do? Now, every entity, people, places, and things, they carry continuously updated predictions.

36:18 Maria has a 68% probability of repurchasing in this category within 30 days. And level six, the relationships are traversable as part of a graph. So the question sharpens to the individual with a quantified answer. Maria has a 68% probability of repurchasing in this category within 30 days and a 19% churn risk driven by two late deliveries. So as you can see, as the model of what you're representing gets richer, your analytics get richer.

36:55 You can ask and answer more sophisticated questions. Now, let me just skim through levels seven through nine. Now the model can represent actions, the live state of the business of the actions themselves and the process. So analysis can tell you about the business. So level seven, the business's possibility space becomes legible. The available actions are modeled with preconditions and effects.

37:29 This is like any operational application has to say, before this can happen, these things must be true and after this happens, these are the effects. So the analysis can answer what are we able to do now and what would each option produce? So for the spring sale across every order currently mid- checkout, which ones are eligible for split shipment and what's the delivery time effect? Level eight, the business's true present becomes knowable.

38:00 This is, again, where ... This is essentially when you can unplug your operational applications at some distant point in the future because the state is live. This is not a snapshot. And the analysis answers what is actually true right now across everything simultaneously. And so mid-sale, you can say, at this instant, how much tent inventory remains, how many carts hold one, which fulfillment centers are at capacity, and where are carrier delays emerging.

38:31 And then lastly, level nine, the business's own operating logic becomes inspectable because now the process, the workflow stages, the decision rules, the coordination patterns, it's now all data. So analytics can answer how does this business actually run and is that the best way? So for the sale, what is the defined path and order takes from cart to settlement? What rules govern the transition each transition? And where do orders actually stall versus where we assumed they would?

39:04 So the arc across this is level seven makes the business options knowable, level eight makes the present knowable, and level nine makes its operating logic knowable. So the through line is across that whole stack, layers on and two they count the sale, three and four explain it, five and six prescribe it and quantify around it, and seven to nine acts on and improves it, each span unlocked by the representational sophistication of the data model underneath it.

39:44 >> Thank you, George. And as we've talked about in previous Breaking Analysis', now we talked about the example of backing up data. In this state, you're backing up the state of the business, this level, you're backing up the state of the business. You're capturing what happened at each of those levels and not only what happened, but why it happened, what was the logic that went into the decision that caused the agent to take that action? So again, this is sophisticated stuff and it's not going to happen overnight, but in that previous section that George just went through, we focused on the value of the analytic data foundation and the maturity roadmap associated with that.

40:28 Here, Tony, if you bring up the next slide, we're going to dig into the range of actions that are supported by that data foundation and how that's going to evolve across the levels in Georgia's model. And the spectrum starts at level one with humans reconciling data snapshots from static reports. They're trying to extract meaning from all that and trying to figure out what to do next, all the way up to that level seven through nine where systems are operating autonomously based on top- down goals and constraints and objectives.

40:57 So George, walk us through this and help us understand how this matures and also the role of humans across this range of actions and how that evolves. >> Okay. So what's critical here is to reinforce, again, that just bringing ChatGPT or Claude into your enterprise is only the first step that the structured data ... And as we'll describe later, it's more than just analytic data.

41:28 That is what powers the intelligence of these agents. So here, we walked through first the sophistication of the data model, which is a fancy way of saying how well can you represent how your business operates. Then in the last slide, we went through the sophistication of the analytics. What questions can you ask about your business? Now, we go through what an agent can actually do. And I'm going to go through fo categories.

41:58 And the through line is that the action sophistication is gated by what the data model represents. And so each band can only support the kind of activity that its data and analytics makes safe. So first layer, there's essentially no agent action. This is what the vendors started to show last year where they're reading the data, not taking action. This is the talking to your data.

42:29 This is when agents can gather the information, Databricks Genie, Snowflake Intelligence, they use the metrics and dimensions to enable users to query the data in natural language. The output is either a dashboard or just conversational. And it's worth mentioning just here, when we say the output is a dashboard, this is actually really critical. Something happened in the last week, which didn't get as much attention as it deserved, which is that Microsoft shut off Databricks' access to Power BI, because Databricks was using its metrics and dimensions definition in the Unity Catalog to power the dashboards in Power BI.

43:20 And the reason that's important is in a world with systems of engagement where the system of engagement generates the user interface, the dashboard can be generated based on the metric and dimension. In other words, Power BI becomes irrelevant. That means if you own the metric and dimension definition, the BI tool is not that important because the dashboard gets generated on demand.

43:52 And a couple years ago, I talked to the guy who was running Power Platform at the time and he was like, "We're working with customers who are generating tens of thousands of dashboards, or are planning to, where they had handcrafted them before. " That's why Microsoft shut it off. They want to make sure that stuff happens with Fabric, not with Databricks. All right, moving on. Levels three and four, this is where you get segment level recommendations.

44:22 This is what you see in the Salesforce Data Cloud where you have tentative cohort level recommendations for a marketing program. Customers showing Maria's hesitation broken by promotion pattern convert three times better when sent a discount within 48 hours. It's a genuine recommendation and it's a heuristic, but a human marketer still pulls the trigger. So an agent might execute it, but a human has to initiate it.

44:56 Levels five and six, this is quantified individual recommendations. Humans still advise. Issue Maria, free shipping, predicted to cut her 19% churn risk to 7%. Levels seven through nine, agents discover, compose, and execute within governance bounds. So this is when action sophistication crosses from recommending to doing. And so like level seven, the actions are modeled. Retain Maria at acceptable margin.

45:27 Compose, resolve her open delivery complaint, then issue the retention offer, and autonomous execution begins here within governance bounds because the agent can finally reason about the action space rather than being hard coded to an API. That's really important. Level eight, this is live state. The agent executes against current real- time view of the business. Mid-flow, it observes that the tent has just gone out of stock and it either substitute ones or recommends an alternative rather than firing off an offer it can't fulfill.

46:07 And so now human oversight shifts from approving each step to handling exceptions. And then finally in the last layer, process as data, the agent acts on and improves the process itself. It observes that in resolving complaints. Before offering promotions, it retains deal- driven buyers like Maria better. Just the key is that we go from humans acting on static reports, the system recommending actions based on segments, to recommending quantified individual actions, finally to agents discovering, planning, and executing on actions and autonomously improving within governance bounds.

47:00 >> Okay, excellent. We're going to close now by helping our audience understand the reasoning layers and some of the most important parts of this digital twin model so they can work backwards from that end state and support the higher levels of maturity. And then in this slide, if you bring it up, Tony, we do that and we map in some of the vendors that we see as positioned to support the transition. Some of them new names that might surprise you a little bit.

47:26 So here, we're zooming in on layers three, four, and five. So George, explain the importance of these layers and how they set up the end state that you've envisioned. >> Okay. So what's really critical, and we mentioned this earlier, there's this observability layer that's merging to pull in a lot of other activity. When I said that ClickStreams drove the need for the whole big data movement because it was 1000 times more data volume than what we used to put in Teradata or Oracle for analytic purposes.

48:11 And it's not that Teradata or Oracle couldn't handle them, but we need essentially a system with a new pricing model because we were doing 1000 times the amount of data. When you take the exhaust from an agent, how it reasons, what context got pulled in, what tool calls it made, what results came back from the tool calls, how the tools talked to each other, the subagents talk to each other, that's 1,000 to 100,000 times more data than you were accumulating for cloud native applications like in Datadog.

48:47 Now, this isn't just for after the fact diagnosis. This has several separate disciplines that used to be several data planes. They're all collapsing into a single substrate. Observability captures the agent traces. Agents train on them to improve. Evals evaluate how they reasoned on that training and they score the agents, and that those evaluations are also used in continuous integration and deployment as gates for deploying the improved agents.

49:32 And those evals actually are also used in the equivalent of agent site reliability engineering so that in real time, another reliability agent can intervene if something's going wrong and correct the course of an agent in production. All these things now are becoming the new source of truth for agent reasoning. Just the way, in a DBMS, the log was the source of truth for what happened.

50:03 It wasn't really the tables, it was the log. The observability substrate, it's going to be cryptographically verified like what happened and what was the state of the business when an agent made a decision. And the reason I'm stressing this is this is why Snowflake bought Observe and why Databricks is putting a huge amount of effort into extending MLflow into a whole observability platform.

50:34 This is why Datadog went parabolic last quarter because the volume of data they're capturing from agents is increasing. I don't think Wall Street realizes how they're going from a diagnostic platform to what is essentially the learning substrate for agents. This is why everyone is competing on this. This is the new memory. The agent platforms themselves are going to try and offer observability, but if you're locked into their platforms, you're never switching from their models.

51:10 But the point is this is how your organization learns. So that's that top layer. And then just quickly, there's been a ton of chatter about the context graph. The context graph is itself two components. This is what happens when the deterministic rules about what's supposed to happen in a business process don't exist or conflict. It's where the tacit knowledge applies.

51:40 And there are really two sources. There's all the unstructured information, documents, communications, but it's not RAG. RAG is dying because it didn't capture all the structure in documents that represented so much of the meaning of the documents. And so Pinecone itself is offering now a new database that's not about creating vector embeddings that captures the structure of documents and you're going to see much more of this. And so it's a whole new form of data engineering.

52:10 Data engineering was taking your operational data and turning it into analytic data through pipelines. Now, we have a new type of data engineering, extracting, representing, and serving unstructured knowledge, turning it into structured knowledge. And then finally we have the expert teaching. This is where domain experts write out how to grade the reasoning process behind a good decision in a process. These are the rubrics and evaluations that agents then will follow in that observability layer. >> And that grading then gets fed back into the system and it just improves over time.

52:44 George, excellent. Really appreciate you taking the time to dig in to this work that you've been doing. We got to wrap. I just want to give a quick summary here. We talked about the AI wave starting out a little bit like the PC era with individuals taking control of their own productivity. But if every person, department, every vendor starts to build their own islands of intelligence, enterprises are going to recreate the same silo problem that has plagued software for decades, only it's going to happen faster.

53:13 The next enterprise stack is a new front end and a new backend. And that front end is that system of engagement that we talked about, personal agents, interfaces where people are getting work done. And then the backend is really the system of intelligent, that layer that we just unpacked in the last slide, and that organizes enterprise knowledge, it incorporates rules and context and actions and that business logic. Snowflake and Databricks, we said many times, have crossed the Rubicon.

53:39 They're not just staying in their little analytics sandbox. They are ambitious companies with ambitious leaders. And so they're moving beyond their traditional analytics swim lane and they've moved squarely into the system of engagement and they're moving toward the system of intelligence alongside the likes of those companies that we talked about, Salesforce, SAP, Microsoft, Palantir, Celonis, Oracle, Google, Anthropic, OpenAI and others, maybe AWS or certainly Microsoft. And then the richness of that data model is going to determine what AI can do in that level of maturity.

54:14 Remember, that transformation is going to happen in two simultaneous phases. Bottom up from the personal productivity side, individuals going to build their personal agents and skills, and at top down where leadership really has to make sure that those skills become governed assets tied to a shared ontology within the organization. And our warning is that customers shouldn't just blindly default to model makers, although there is some debate inside theCUBE and theCUBE Research community about this and we'll pick that up where some folks feel like the LLM vendors are actually going to be the place to partner with to own this sort of system of intelligence.

54:52 But the caution is that the frontier models, they're really important, but owning that model is not the same as owning the enterprise operating context and you don't want to get locked into that. That's a caution that we put forth. The long-term goal, of course, is massive improvements in productivity, not just personal productivity. And we think that the firms that thrive are going to move toward an operating model that's organized around intelligence in a shared business truth and a governed set of data and agents rather than traditional hierarchies that are built around humans.

55:27 We wish you luck. We'll be there evolving this model and reporting. This is Dave Vellante, for George Gilbert for this Breaking Analysis. Thanks for watching and we'll see you next time.