{
  "episodeId": "ep3",
  "episodeTitle": "Enterprise Context Graphs: The Next Trillion-Dollar AI Opportunity?",
  "speakers": [
    "Vikas",
    "Sachin"
  ],
  "speakerMap": {
    "Speaker 1": "Vikas",
    "Speaker 2": "Sachin"
  },
  "duration": 1671,
  "durationFormatted": "27:51",
  "wordCount": 4818,
  "segmentCount": 32,
  "segments": [
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 52,
      "timestampFormatted": "0:52",
      "text": "Hi, welcome to a new episode of the AI Valley podcast. My name is Vikas and I have with me."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 59,
      "timestampFormatted": "0:59",
      "text": "Hi everyone, I'm such an"
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 60,
      "timestampFormatted": "1:00",
      "text": "So today we have a very interesting episode on context graph, or rather enterprise context graph. And we are here to discuss if they are really a trillion dollar opportunity in 2026. And if they are, who's the one who will capture this opportunity? And if you're in the data world, you probably saw a bunch of tweets go viral on system of record and context graph and so on, right? lot of argument around the long-lived system of record by Jammin Ball who argued that despite the discourse of AI, agent won't kill the system of record like things like your warehouses, CRM, CRP and so on. Instead, truth will live in system of record with a semantic layer on top that tells agent on how to use that truth. In their viral post, Jaya Gupta and Ashu Gurk responded in their viral tweet with a new thesis about AI's trillion dollar opportunity about context graph. In their post, they argued that adding AI to a system of record with some semantic layer is not enough. Instead, the next trillion dollar opportunity is in what it misses and it introduced a new concept about decision traces or why behind the past decisions. So all of this will form what's called as a context graph, which they call the next trillion dollar opportunity. What do you think about it Sachin?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 148,
      "timestampFormatted": "2:28",
      "text": "Yeah, think, look, it's definitely something to ponder on, Because I've worked with some of these SaaS companies in the past, whether it's Salesforce, I've worked with Concur as well, doing expense management. Salesforce obviously is famous for CRM. Now, what happened is when these work-oriented software solutions came out, They were very, they obviously created a trillion dollar opportunity, right? Like the whole business changed because of that and work streams are very specific for that particular action or that particular workflow. Now the question really is how do you create action oriented systems? Because the world is moving in an agentic direction. Agents will be cross domain. You may need a system that let's say originates from CRM, talks to service, talks to marketing, and then, you know, does a lot of mix and mash between both of these systems. So I think definitely it's a trillion dollar opportunity. I kind of agree with what was said in that. Because if the erstwhile SaaS created a trillion dollar opportunity, then I think these context graph systems will create it. And I think this is a good play on the words that a lot of us talk about enterprise knowledge graph. But I think how do you create context sort of knowledge that's never been captured?  And these systems are not capable of capturing that today. Like if I'm interacting with the CRM today, does it capture marketing information? Yes, if you have the same system. But if it's a different system, then how do you associate that system of record that you've captured with, let's say, a cross system that's coming out from marketing? How do you collaborate this? Because that knowledge may have not been captured. There's also a lot of tribal knowledge, right? Like which is sitting in our heads. I'm a marketer, for example, I use a lot of jargon, which is very specific to my company, but I'm not capturing that anywhere. And so how do I take that knowledge and pass it on to an agent who's trying to do work on my behalf? And I think this is going to create a very  interesting system. It's going to create different, it's going to create very interesting ontologies also, right? Because how do I understand, how do I skim across these ontologies? How do I... make sense of what was asked of the agent, what action the agent was supposed to take, and did it take the right action based on what it understood. So I think, it's a long story short. think it's a very interesting domain. Many companies are going to now start playing into it. I also believe very strongly it's going to activate a lot of search, because when someone is searching for anything, in a product, on a retail site, or on a food site, wherever that is, it will also trigger some ontologies because users will bring their ontology, they will bring their tribal knowledge into the application. And when these agents who are trying to search items for them, they will have to understand what that means. So I think a lot of that mapping, that graph-based systems of what that context means so that it searches the accurate item. I think that's going to be super super interesting to see how this industry shapes up for it."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 318,
      "timestampFormatted": "5:18",
      "text": "Yeah, interesting point Sachin. I think while I understand the opportunity of it, the tricky question here is who gets to own it and how do we go about implementing this sort of a thing because it's one of those easier said than done class of problem, right? Because why we understand that it's not the system of record agreed. what my little bit, the bit that I don't understand fully is in the narrative that Jaya and Ashu built in their blog was that their understanding was that the agent that work in the execution path can see the full context when the decision is made. So they felt that the agent which are sitting in doing these activity or rather I think what they're implying is there would be a bunch of different vertical agent startup who will own the context graph for their domain. For example, the sales agent company will capture the context around renewals, support agent company will capture the context around escalation, and so on and so forth. So they believe that it won't be one central enterprise context graph, but versus different verticals. And that actually prompted a lot of debate and discussion on both X and other places on how to actually build a context graph.  In the tech circles there are discussions around something called as a two layer context architecture and things around operational context and execution intelligence and what this means for agent reliability, governance, observability, so on so forth. I believe, I understand why this article went viral. It's an elegant thesis, right? I think there's no two ways about it. And because they are right about the importance and opportunity of the context, I've like you eluded as well. I'm not able to understand or there have been multiple opinions on who gets to capture this trillion dollar because the reality of or rather messy reality in the enterprise is the heterogeneity of the data. Because anyone who's solved anything about enterprise because the data, the same copies of data lives in multiple places and they don't evolve at the same time, right? So, again, is, I'm not sure. I mean, for example, like, you know, when data warehouse, when we wrote application and databases, we saw data engineering or we saw data analytics or as a different class of product that sits outside the application and database. Will this be solved outside the application and the databases or will it be within? within the boundaries of the current application and databases. So yeah, those are things that I believe that we will see."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 487,
      "timestampFormatted": "8:07",
      "text": "See, I have a slight opinion on this. And this is just, I mean, again, it's just my opinion. Look, there are many contenders for the ownership of this context graph. For example, the CIO org, they traditionally have owned the infrastructure. They own the system of record. They own the connectors. Now, they may want to have natural ownership to these context graphs because in their mind, it's foundational infrastructure. Now, the problem that I see with that organization owning it is they are not typically in the execution path of any business decision. an IT CIO or can deploy Salesforce, but they probably will not take a call at whether this customer should get a 40 % discount or not. Yeah, yeah, yeah. That business decision sits with the business. And that's an operations team that decides what what workflow should a 40 % discount customer flow into? What should a different what should a marketing customer flow into? What's a large customer? What's a small customer? All these workflows are already pre decided and they sit with business ops teams because they understand how work actually needs to get done. But they don't lack the sorry, they lack the technical infrastructure. or the cross-functional authority to get this done. So we also have another contender, which is a business ops team, because they have to decide what happens for that particular customer in that particular workflow. There's also the data analytics team, which could claim ownership for this, because obviously they traditionally own the warehouse, they own the BI tools, they are considered to be the single source of truth. So their argument is also valid because they will say that this is data, it's about lineage, it's about governance. But again, the problem is they are in the read path. They are not in the write path. So again, by the time the data reaches them by some kind of an ETL process, the decision context is already gone. And then there's obviously your center of excellence, is your AI teams and all of these things. Now they may obviously want to own the AI strategy. They may want to own the prompt engineering. They may want to own how these agents are governed. But again, these teams are very new. They don't have a lot of budget authority. Their paths are slightly unclear right now because AI as a mainstream segment is fairly new. But let's go back a little bit in history. When a similar such problem happened with data platforms, that's when the CDO org came up all of a sudden. It originated with a new vertical."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 626,
      "timestampFormatted": "10:26",
      "text": "Right. I don't think before 10 years... Something like that."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 631,
      "timestampFormatted": "10:31",
      "text": "There was something. So high possibility that this vacuum may create something new. So  it's happened in the past. It could happen again. So I think it's going to be like a federated ownership model. It may craft out a new category. I think  that's where. So we'll have to wait and see how this plays out, how companies get intelligent about it. But I think to your earlier point, right, as to is this a domain that people will go after. think because it does not have boundaries today, this will create a lot of Because these startups can actually go cross-boundary. For the lack of a better word, startups don't respect these boundaries as much. In fact, if anything, they operate better when there are no boundaries. And I think this will create a lot of startups because this is exactly the playground that they like to play in. They will crash through the organization. They will create a lot of aha moments for these for some of these enterprises to realize how to create an organization. I think eventually enterprises will get smart about it and they will rally their organization around it. so I think that's, to me, I see this as major two opportunities."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 691,
      "timestampFormatted": "11:31",
      "text": "But tell me,  I'm just curious about the context platform market that is emerging here, I mean, I'm just thinking, One, you said this is not just a system of record. This is more on the activity, right? The activity goes, which is beyond the indexable content, right? Which is low level activity signal, time-stamped action within tools. the decisions could be taken on, like you said, slide chat meeting across different places, within, through a different system. I'm wondering if, obviously it will contain the decisions, the data, right? But what else will an enterprise context graph will contain? It will obviously have some business logic, it will have cross platform relationship or  of different function like you said, right? The finance org or the customer org, it will obviously understand your org structure. So do you think existing players like the data warehouse platforms have an unfair advantage or is it something for a new greenfield players to solve?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 758,
      "timestampFormatted": "12:38",
      "text": "I honestly think it's for... I think it's for new Greenfield players. And again, I'm just going by some analogy here. So, when the world started moving towards event-driven architecture, it created a Kafka. I see this context graph as a Kafka layer. The closest similarity that I can get to is thinking of it Kafka. It sits in between. So, it... captures decision traces as they happen, and it would feed that into a system. So I think it will be in the execution path, and it will supply that context to these agents. OK. Now, that said, it will create some new companies that will be able to break through a lot of this monotony. I'm not saying that the existing data warehouse companies may not be able to do it. If they are able to be as agile to understand how to break through these barriers, they could. But I see that this could, this is actually, the playing field could be very level here. And it could create a good opportunity for  any party to enter."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 816,
      "timestampFormatted": "13:36",
      "text": "I agree because we already have, even though we've been talking about agentic systems and agent, everybody's building their own agentic platforms and tools. But the funny thing about these agents is whether these are first party agents or third party agents, none of these talk to each other, right? Yeah. And  I'm assuming from, if you look from a customer point of view, they would see a merit in building a common layer of context because you know, they don't want to context engineer every single one of them. So, but the challenge is while execution paths are definitely local, right. Which each agent has their local context execution path. The context is definitely global, right. If we go back to Jaya and Ashu's core argument for agents owning the context graph. A system of agent startups have a structural advantage, right? To your point, right? Because they sit in the execution path. And this is true for capturing decision traces within a single workflow. But for most decisions, context comes from everywhere, right? And I'll give you an example, right? For example,  somebody wants to, you know, renew their observability tool, right? They would want to... understand what was their incident history, how many people are using it, what price did we buy it for, who approved it, who was the sales person that we were dealing with, and so on and so forth. There are different systems. And the kicker is every enterprise has a different combination of these systems. Somebody could run Salesforce, Zendesk, Snowflake, somebody runs HubSpot, Intercom, Data Break, somebody runs... you know, ServiceNow, BigQuery, CRM, or  the point is each company has different system. Some are not all of these systems are standardized. These are all, you know, bespoke for every enterprise. And this is the heterogeneity problem that a lot of people are already pointed out as a rebuttal to what Jai and Ashu's thesis, right? And because it doesn't address this heterogeneity, right? And because the vertical agent startup sees the execution path and captures context within their workflow, because enterprise have dozens of these agent across dozens of vendor each building their own context silo and the vertical agent can't see this full context web. which basically mean that to truly... capture this context graph, a vertical agent needs to integrate to say hundreds of systems within one org to just to cover the common cases, right? And this is basically an n cross n problem. So they have to have connectors to each of these systems and build out. so again, system heterogeneity just started the problem."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 976,
      "timestampFormatted": "16:16",
      "text": "Yeah"
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 983,
      "timestampFormatted": "16:23",
      "text": "because most people we think of context as a single concept because context itself is heterogeneous, right? And there is a great explanation by Thomas Tangs who gave an explanation on two emerging type of context databases. One is operational context and another is analytical context, right? Obviously operational context as the name suggests is operating procedure, your institutional knowledge and stuff like that. these are your customer's IP. Analytical context is the semantic layer that is the implied data, right? The semantic things like your revenue, customer acquisition cost, your EBITDA and so on and so forth, right? Something that is derived. So because these are deeply intertwined and distinct context, right? I think... I think that it's a hard computer science problem. Now, you think given the nature of debate will say a digital native org or say tech first org will jump on to solving this problem already or do you think they'll wait for industry to mature and solve this?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1047,
      "timestampFormatted": "17:27",
      "text": "So, look, I think this is going to create an interesting problem, right? And I see this a little bit, it's a little paradox, okay? So, now what happens is, today,  most digital natives have invested in agent architecture. They've created something that is an agentic system. Now, the challenge is that once you've deployed these agentic systems, you will obviously go back and figure out what's the ROI on that. And just by first principles, I think we have clearly demonstrated or we have spoken about the fact that without enough context, these agents will only execute until a certain point and drop off after that because they just don't have enough knowledge about what to do from here. So I think what will happen is the... these companies by investing in an agentic system, they've actually created some sort of a reverse problem for themselves, where they will eventually realize that these agents are not completing the task that they were supposed to do, because there is no decision, right? Like they don't understand, like an agent does not have enough information that if a customer asks for a 40 % discount, what happened after that? Exactly."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1116,
      "timestampFormatted": "18:36",
      "text": "Should it give or should it give?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1118,
      "timestampFormatted": "18:38",
      "text": "So the answer really lies in the question that if I'm giving the agent an activity or a task to perform, and it's an action-oriented agent, then how do I measure that action? Because the agent does not have information about that action. It did not know what decision was taken in the past."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1134,
      "timestampFormatted": "18:54",
      "text": "So it's an interesting point and I see where you're going with this, but come to think of it, don't you think this, this context as we're speaking about, right, is also the most, probably the most valuable piece of data within the org, barring the customer information. If you leave the PII data outside, if you look at the core mode of an org, right. Now the question is, will Will companies or will the enterprise trust a third party system to allow them to peek into all of this data and take it away for processing? So, yeah, because come to think of it, you you are sort of offloading your financial decision, your, you who knows they would be able to."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1172,
      "timestampFormatted": "19:32",
      "text": "Yeah. See where you're"
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1186,
      "timestampFormatted": "19:46",
      "text": "you know, if I if you if there's no guardrails around it, somebody would be able to, you know, train an AI to to predict how you behave in a certain situation, right?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1195,
      "timestampFormatted": "19:55",
      "text": "Yeah, absolutely. mean, see, look, the security risks of these are high. Okay, because if I knew that by following A, B, C as a flow, and I know that the result was D, I could reverse engineer it."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1209,
      "timestampFormatted": "20:09",
      "text": "Yeah, yeah. I think and if you also I realize that Thomas in his blog made this point, similar point because enterprise have gone through this this whole lesson when when they when the cloud data warehouses came around, right? So they handed over their both data and compute and they watched as the most strategic asset in their business, right? how they operate became somebody else's leverage. So, which is why 10 years after the new formats like Iceberg exist, which are open table formats, right? And they are kind of winning. If you see a lot of people are now moving to open table formats, because enterprises realize that, know, seeding control of their data to a single vendor creates a... an immense lock-in and could hurt them surgically, right? Because, that data itself could become leverage against them because in, because cloud models are around, you know, egress cost and data volumes and big data by definition was the volume didn't lend itself well to be taken out or moved around very quickly, right? So I believe context RAF will have similar set of challenges."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1282,
      "timestampFormatted": "21:22",
      "text": "Yeah, absolutely. think that's a very, very interesting point you've raised. Because look, it is... I guess this is a pattern that every new solution will go through. It will start with a little bit of denial and it will start with people pushing it back. But I think it will follow through the flow, people will understand how to use it. See, look, I don't see... How are we going to come out of this? I don't think there is a way out. I think this is a system that's a must have, especially given the fact that we're going to have a lot of agent systems in the future. So I don't see how we're going to come out of this. But I think as more and more companies start to evolve in this category, customers will make decisions about what to do. Should we give this to a startup? Should we give it to any, even a traditional company? Should we give it to them or should we own it? Because you're right, Until now, people just had the data but they didn't know what was done with that data, right? Because that decision logic was sitting with the human. Yes, so it's going to be interesting."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1339,
      "timestampFormatted": "22:19",
      "text": "The human aspect is, and I was thinking about it, see, there are two set of challenges here, One, an enterprise being afraid that their most valuable data is sitting with somebody else, and somebody's learning or sort of building insights on their data. Even though that may not be a public insight, but still somebody else has that insight. Now, the second aspect is what you mentioned is a people aspect, right? Now, imagine you and I are in a company with our set of tribal knowledge sitting with both of us, right? Now, as an org, the org is telling us that, look, you are to take the tribal knowledge out, the exception logic that you make for certain sort of engagement or deals. know, we always do X because of Y reasoning, which today lives in your head. The company is saying that take this out of your head and put it in a data store. Yeah, right. So I believe there will be a little bit or rather a lot of resistance from people as well who would be sort of because there is a anyways there is a little bit of you know fear about the AI led job replacements, right? So people are going to not jump as the first thing, right? So people will be very careful on if they give away this IP that's sitting in their head, the tribal knowledge, then they would probably be not required to make those decisions. Probably they are giving the autonomy to an agent to make the decision on their behalf. They'll be reduced to approvers of those decisions. one is the need While it's a trillion dollar opportunity, yes, but enterprises are not going to hand over that data to a dozen of  vertical agent startup and giving them their what's called as an operational DNA."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1448,
      "timestampFormatted": "24:08",
      "text": "Yeah,"
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1448,
      "timestampFormatted": "24:08",
      "text": "So because come to think of it, given how AI is going towards commodity, so LLMs are commodity and I don't think agents are going to be different, but context is going to be a strategic asset because that's not going to be a So enterprises will want to own their context with maybe an open and federated context platform that any agent, right?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1464,
      "timestampFormatted": "24:24",
      "text": "Absolutely."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1474,
      "timestampFormatted": "24:34",
      "text": "be it in-house or a third party agent or, you know, they can own and they can choose what data they expose to these agents. So it feels like it's going to be more and more of anybody who builds, maybe it could be one of our big techs or vertical, anybody who builds a truly open source or open data format context graph is going to have... going to see a lot more adoption than these closed enterprise contexts are."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1500,
      "timestampFormatted": "25:00",
      "text": "Absolutely, think, look, I completely agree. This is interesting, right? And the moat would sit with the person who manages to build this enough to have as much tribal knowledge of the company as possible so that these agents can take decisions, right? Because eventually you want decisions at scale. So yeah, I think...  This will obviously create a new vacuum for security of these agents. It may create a vacuum for, let's say, how do you provide infrastructure just to build the agents, to build the context graph, everything so that the ownership sits with the company itself and doesn't go out. So I think, look, there is this multiple roads and pathways that this can take. It is definitely going to be interesting on how companies are thinking about it. I mean, obviously in the capacity of our In our capacity, we've seen a lot of customers, a lot of companies talk about it. They have expressed interest. They've also shown the fact that this is important because now as and when they start employing AI systems to become the front end of interaction with their customers, they're seeing that agents need context, which was initially sitting in a human's mind. So I think it's going to be super, super... fun to figure out what roads these companies take. And again, like every technology, are always going to be players who will open source will thrive. People will create open formats to write some of this. There will be companies that will take a lot of these formats and build companies out of it to capture the trillion dollar opportunity. So I think it's a space where a lot of people will play. We saw this last time in the data world. We'll see it again in the content graph world. I think it's only a matter of time."
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1602,
      "timestampFormatted": "26:42",
      "text": "Yes, and I think we're not making any conclusions today, but I think it gives a framework on what enterprises will assess or think about these things. think probably in future episodes, we should also think about if somebody has to build something like this, what should be the building blocks? And we will have other sort of you know, contours to this problems as we go further, right? And then I think there are other class of problems, We are talking about enterprise context, like what about startups? Do smaller companies, do they need context as well? Should they start? Maybe it's hard to capture the historical information, but should they start capturing things?"
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1637,
      "timestampFormatted": "27:17",
      "text": "exactly"
    },
    {
      "speaker": "Vikas",
      "speakerRaw": "Speaker 1",
      "timestamp": 1646,
      "timestampFormatted": "27:26",
      "text": "as they make decisions should they capture in a special class of database. So I think there's a lot to be discussed as we unpack this. if you have any, you know, if you're listening to this and if you have any comments, feel free to write to us or comment. We're happy to hear what you have to think about this, about the enterprise context graph, right? So yeah, I think it's... It's interesting that suddenly everybody's piling on this debate."
    },
    {
      "speaker": "Sachin",
      "speakerRaw": "Speaker 2",
      "timestamp": 1671,
      "timestampFormatted": "27:51",
      "text": "Yeah. No, think we have left our listeners with some food for thought. If you weren't thinking about this, our recommendation would be to start thinking about it. If this field is in close alignment to what you do in your job, I think it's important to figure out a way to think a little hard, think a little deep about it, because this will come in some form into your company. So give it some thought. We are also open to some of the ideas that you think we know, maybe we have skipped a few things. Maybe we just didn't bring it up. So, but you know, write to us and see if there is, if you have some opinions about this. Have a great evening. All right. Bye."
    }
  ],
  "chapters": [
    {
      "title": "Hi, welcome to a new episode of the AI Valley podc...",
      "timestamp": 52,
      "timestampFormatted": "0:52"
    },
    {
      "title": "Yeah, interesting point Sachin",
      "timestamp": 318,
      "timestampFormatted": "5:18"
    },
    {
      "title": "See, I have a slight opinion on this",
      "timestamp": 487,
      "timestampFormatted": "8:07"
    },
    {
      "title": "Right",
      "timestamp": 626,
      "timestampFormatted": "10:26"
    },
    {
      "title": "I honestly think it's for",
      "timestamp": 758,
      "timestampFormatted": "12:38"
    },
    {
      "title": "Yeah because most people we think of context as a ...",
      "timestamp": 976,
      "timestampFormatted": "16:16"
    },
    {
      "title": "Should it give or should it give? So the answer re...",
      "timestamp": 1116,
      "timestampFormatted": "18:36"
    },
    {
      "title": "Yeah, absolutely",
      "timestamp": 1282,
      "timestampFormatted": "21:22"
    },
    {
      "title": "Yeah, So because come to think of it, given how AI...",
      "timestamp": 1448,
      "timestampFormatted": "24:08"
    },
    {
      "title": "Yes, and I think we're not making any conclusions ...",
      "timestamp": 1602,
      "timestampFormatted": "26:42"
    }
  ],
  "quotes": [
    {
      "text": "Like the whole business changed because of that and work streams are very specific for that particular action or that particular workflow.",
      "speaker": "Sachin",
      "timestamp": 148,
      "timestampFormatted": "2:28"
    },
    {
      "text": "Because if the erstwhile SaaS created a trillion dollar opportunity, then I think these context graph systems will create it.",
      "speaker": "Sachin",
      "timestamp": 148,
      "timestampFormatted": "2:28"
    },
    {
      "text": "And I think this is a good play on the words that a lot of us talk about enterprise knowledge graph.",
      "speaker": "Sachin",
      "timestamp": 148,
      "timestampFormatted": "2:28"
    },
    {
      "text": "So I think a lot of that mapping, that graph-based systems of what that context means so that it searches the accurate item.",
      "speaker": "Sachin",
      "timestamp": 148,
      "timestampFormatted": "2:28"
    },
    {
      "text": "I think while I understand the opportunity of it, the tricky question here is who gets to own it and how do we go about implementing this sort of a thing because it's one of those easier said than done class of problem, right?",
      "speaker": "Vikas",
      "timestamp": 318,
      "timestampFormatted": "5:18"
    }
  ],
  "fullText": "[0:52] Vikas: Hi, welcome to a new episode of the AI Valley podcast. My name is Vikas and I have with me.\n\n[0:59] Sachin: Hi everyone, I'm such an\n\n[1:00] Vikas: So today we have a very interesting episode on context graph, or rather enterprise context graph. And we are here to discuss if they are really a trillion dollar opportunity in 2026. And if they are, who's the one who will capture this opportunity? And if you're in the data world, you probably saw a bunch of tweets go viral on system of record and context graph and so on, right? lot of argument around the long-lived system of record by Jammin Ball who argued that despite the discourse of AI, agent won't kill the system of record like things like your warehouses, CRM, CRP and so on. Instead, truth will live in system of record with a semantic layer on top that tells agent on how to use that truth. In their viral post, Jaya Gupta and Ashu Gurk responded in their viral tweet with a new thesis about AI's trillion dollar opportunity about context graph. In their post, they argued that adding AI to a system of record with some semantic layer is not enough. Instead, the next trillion dollar opportunity is in what it misses and it introduced a new concept about decision traces or why behind the past decisions. So all of this will form what's called as a context graph, which they call the next trillion dollar opportunity. What do you think about it Sachin?\n\n[2:28] Sachin: Yeah, think, look, it's definitely something to ponder on, Because I've worked with some of these SaaS companies in the past, whether it's Salesforce, I've worked with Concur as well, doing expense management. Salesforce obviously is famous for CRM. Now, what happened is when these work-oriented software solutions came out, They were very, they obviously created a trillion dollar opportunity, right? Like the whole business changed because of that and work streams are very specific for that particular action or that particular workflow. Now the question really is how do you create action oriented systems? Because the world is moving in an agentic direction. Agents will be cross domain. You may need a system that let's say originates from CRM, talks to service, talks to marketing, and then, you know, does a lot of mix and mash between both of these systems. So I think definitely it's a trillion dollar opportunity. I kind of agree with what was said in that. Because if the erstwhile SaaS created a trillion dollar opportunity, then I think these context graph systems will create it. And I think this is a good play on the words that a lot of us talk about enterprise knowledge graph. But I think how do you create context sort of knowledge that's never been captured?  And these systems are not capable of capturing that today. Like if I'm interacting with the CRM today, does it capture marketing information? Yes, if you have the same system. But if it's a different system, then how do you associate that system of record that you've captured with, let's say, a cross system that's coming out from marketing? How do you collaborate this? Because that knowledge may have not been captured. There's also a lot of tribal knowledge, right? Like which is sitting in our heads. I'm a marketer, for example, I use a lot of jargon, which is very specific to my company, but I'm not capturing that anywhere. And so how do I take that knowledge and pass it on to an agent who's trying to do work on my behalf? And I think this is going to create a very  interesting system. It's going to create different, it's going to create very interesting ontologies also, right? Because how do I understand, how do I skim across these ontologies? How do I... make sense of what was asked of the agent, what action the agent was supposed to take, and did it take the right action based on what it understood. So I think, it's a long story short. think it's a very interesting domain. Many companies are going to now start playing into it. I also believe very strongly it's going to activate a lot of search, because when someone is searching for anything, in a product, on a retail site, or on a food site, wherever that is, it will also trigger some ontologies because users will bring their ontology, they will bring their tribal knowledge into the application. And when these agents who are trying to search items for them, they will have to understand what that means. So I think a lot of that mapping, that graph-based systems of what that context means so that it searches the accurate item. I think that's going to be super super interesting to see how this industry shapes up for it.\n\n[5:18] Vikas: Yeah, interesting point Sachin. I think while I understand the opportunity of it, the tricky question here is who gets to own it and how do we go about implementing this sort of a thing because it's one of those easier said than done class of problem, right? Because why we understand that it's not the system of record agreed. what my little bit, the bit that I don't understand fully is in the narrative that Jaya and Ashu built in their blog was that their understanding was that the agent that work in the execution path can see the full context when the decision is made. So they felt that the agent which are sitting in doing these activity or rather I think what they're implying is there would be a bunch of different vertical agent startup who will own the context graph for their domain. For example, the sales agent company will capture the context around renewals, support agent company will capture the context around escalation, and so on and so forth. So they believe that it won't be one central enterprise context graph, but versus different verticals. And that actually prompted a lot of debate and discussion on both X and other places on how to actually build a context graph.  In the tech circles there are discussions around something called as a two layer context architecture and things around operational context and execution intelligence and what this means for agent reliability, governance, observability, so on so forth. I believe, I understand why this article went viral. It's an elegant thesis, right? I think there's no two ways about it. And because they are right about the importance and opportunity of the context, I've like you eluded as well. I'm not able to understand or there have been multiple opinions on who gets to capture this trillion dollar because the reality of or rather messy reality in the enterprise is the heterogeneity of the data. Because anyone who's solved anything about enterprise because the data, the same copies of data lives in multiple places and they don't evolve at the same time, right? So, again, is, I'm not sure. I mean, for example, like, you know, when data warehouse, when we wrote application and databases, we saw data engineering or we saw data analytics or as a different class of product that sits outside the application and database. Will this be solved outside the application and the databases or will it be within? within the boundaries of the current application and databases. So yeah, those are things that I believe that we will see.\n\n[8:07] Sachin: See, I have a slight opinion on this. And this is just, I mean, again, it's just my opinion. Look, there are many contenders for the ownership of this context graph. For example, the CIO org, they traditionally have owned the infrastructure. They own the system of record. They own the connectors. Now, they may want to have natural ownership to these context graphs because in their mind, it's foundational infrastructure. Now, the problem that I see with that organization owning it is they are not typically in the execution path of any business decision. an IT CIO or can deploy Salesforce, but they probably will not take a call at whether this customer should get a 40 % discount or not. Yeah, yeah, yeah. That business decision sits with the business. And that's an operations team that decides what what workflow should a 40 % discount customer flow into? What should a different what should a marketing customer flow into? What's a large customer? What's a small customer? All these workflows are already pre decided and they sit with business ops teams because they understand how work actually needs to get done. But they don't lack the sorry, they lack the technical infrastructure. or the cross-functional authority to get this done. So we also have another contender, which is a business ops team, because they have to decide what happens for that particular customer in that particular workflow. There's also the data analytics team, which could claim ownership for this, because obviously they traditionally own the warehouse, they own the BI tools, they are considered to be the single source of truth. So their argument is also valid because they will say that this is data, it's about lineage, it's about governance. But again, the problem is they are in the read path. They are not in the write path. So again, by the time the data reaches them by some kind of an ETL process, the decision context is already gone. And then there's obviously your center of excellence, is your AI teams and all of these things. Now they may obviously want to own the AI strategy. They may want to own the prompt engineering. They may want to own how these agents are governed. But again, these teams are very new. They don't have a lot of budget authority. Their paths are slightly unclear right now because AI as a mainstream segment is fairly new. But let's go back a little bit in history. When a similar such problem happened with data platforms, that's when the CDO org came up all of a sudden. It originated with a new vertical.\n\n[10:26] Vikas: Right. I don't think before 10 years... Something like that.\n\n[10:31] Sachin: There was something. So high possibility that this vacuum may create something new. So  it's happened in the past. It could happen again. So I think it's going to be like a federated ownership model. It may craft out a new category. I think  that's where. So we'll have to wait and see how this plays out, how companies get intelligent about it. But I think to your earlier point, right, as to is this a domain that people will go after. think because it does not have boundaries today, this will create a lot of Because these startups can actually go cross-boundary. For the lack of a better word, startups don't respect these boundaries as much. In fact, if anything, they operate better when there are no boundaries. And I think this will create a lot of startups because this is exactly the playground that they like to play in. They will crash through the organization. They will create a lot of aha moments for these for some of these enterprises to realize how to create an organization. I think eventually enterprises will get smart about it and they will rally their organization around it. so I think that's, to me, I see this as major two opportunities.\n\n[11:31] Vikas: But tell me,  I'm just curious about the context platform market that is emerging here, I mean, I'm just thinking, One, you said this is not just a system of record. This is more on the activity, right? The activity goes, which is beyond the indexable content, right? Which is low level activity signal, time-stamped action within tools. the decisions could be taken on, like you said, slide chat meeting across different places, within, through a different system. I'm wondering if, obviously it will contain the decisions, the data, right? But what else will an enterprise context graph will contain? It will obviously have some business logic, it will have cross platform relationship or  of different function like you said, right? The finance org or the customer org, it will obviously understand your org structure. So do you think existing players like the data warehouse platforms have an unfair advantage or is it something for a new greenfield players to solve?\n\n[12:38] Sachin: I honestly think it's for... I think it's for new Greenfield players. And again, I'm just going by some analogy here. So, when the world started moving towards event-driven architecture, it created a Kafka. I see this context graph as a Kafka layer. The closest similarity that I can get to is thinking of it Kafka. It sits in between. So, it... captures decision traces as they happen, and it would feed that into a system. So I think it will be in the execution path, and it will supply that context to these agents. OK. Now, that said, it will create some new companies that will be able to break through a lot of this monotony. I'm not saying that the existing data warehouse companies may not be able to do it. If they are able to be as agile to understand how to break through these barriers, they could. But I see that this could, this is actually, the playing field could be very level here. And it could create a good opportunity for  any party to enter.\n\n[13:36] Vikas: I agree because we already have, even though we've been talking about agentic systems and agent, everybody's building their own agentic platforms and tools. But the funny thing about these agents is whether these are first party agents or third party agents, none of these talk to each other, right? Yeah. And  I'm assuming from, if you look from a customer point of view, they would see a merit in building a common layer of context because you know, they don't want to context engineer every single one of them. So, but the challenge is while execution paths are definitely local, right. Which each agent has their local context execution path. The context is definitely global, right. If we go back to Jaya and Ashu's core argument for agents owning the context graph. A system of agent startups have a structural advantage, right? To your point, right? Because they sit in the execution path. And this is true for capturing decision traces within a single workflow. But for most decisions, context comes from everywhere, right? And I'll give you an example, right? For example,  somebody wants to, you know, renew their observability tool, right? They would want to... understand what was their incident history, how many people are using it, what price did we buy it for, who approved it, who was the sales person that we were dealing with, and so on and so forth. There are different systems. And the kicker is every enterprise has a different combination of these systems. Somebody could run Salesforce, Zendesk, Snowflake, somebody runs HubSpot, Intercom, Data Break, somebody runs... you know, ServiceNow, BigQuery, CRM, or  the point is each company has different system. Some are not all of these systems are standardized. These are all, you know, bespoke for every enterprise. And this is the heterogeneity problem that a lot of people are already pointed out as a rebuttal to what Jai and Ashu's thesis, right? And because it doesn't address this heterogeneity, right? And because the vertical agent startup sees the execution path and captures context within their workflow, because enterprise have dozens of these agent across dozens of vendor each building their own context silo and the vertical agent can't see this full context web. which basically mean that to truly... capture this context graph, a vertical agent needs to integrate to say hundreds of systems within one org to just to cover the common cases, right? And this is basically an n cross n problem. So they have to have connectors to each of these systems and build out. so again, system heterogeneity just started the problem.\n\n[16:16] Sachin: Yeah\n\n[16:23] Vikas: because most people we think of context as a single concept because context itself is heterogeneous, right? And there is a great explanation by Thomas Tangs who gave an explanation on two emerging type of context databases. One is operational context and another is analytical context, right? Obviously operational context as the name suggests is operating procedure, your institutional knowledge and stuff like that. these are your customer's IP. Analytical context is the semantic layer that is the implied data, right? The semantic things like your revenue, customer acquisition cost, your EBITDA and so on and so forth, right? Something that is derived. So because these are deeply intertwined and distinct context, right? I think... I think that it's a hard computer science problem. Now, you think given the nature of debate will say a digital native org or say tech first org will jump on to solving this problem already or do you think they'll wait for industry to mature and solve this?\n\n[17:27] Sachin: So, look, I think this is going to create an interesting problem, right? And I see this a little bit, it's a little paradox, okay? So, now what happens is, today,  most digital natives have invested in agent architecture. They've created something that is an agentic system. Now, the challenge is that once you've deployed these agentic systems, you will obviously go back and figure out what's the ROI on that. And just by first principles, I think we have clearly demonstrated or we have spoken about the fact that without enough context, these agents will only execute until a certain point and drop off after that because they just don't have enough knowledge about what to do from here. So I think what will happen is the... these companies by investing in an agentic system, they've actually created some sort of a reverse problem for themselves, where they will eventually realize that these agents are not completing the task that they were supposed to do, because there is no decision, right? Like they don't understand, like an agent does not have enough information that if a customer asks for a 40 % discount, what happened after that? Exactly.\n\n[18:36] Vikas: Should it give or should it give?\n\n[18:38] Sachin: So the answer really lies in the question that if I'm giving the agent an activity or a task to perform, and it's an action-oriented agent, then how do I measure that action? Because the agent does not have information about that action. It did not know what decision was taken in the past.\n\n[18:54] Vikas: So it's an interesting point and I see where you're going with this, but come to think of it, don't you think this, this context as we're speaking about, right, is also the most, probably the most valuable piece of data within the org, barring the customer information. If you leave the PII data outside, if you look at the core mode of an org, right. Now the question is, will Will companies or will the enterprise trust a third party system to allow them to peek into all of this data and take it away for processing? So, yeah, because come to think of it, you you are sort of offloading your financial decision, your, you who knows they would be able to.\n\n[19:32] Sachin: Yeah. See where you're\n\n[19:46] Vikas: you know, if I if you if there's no guardrails around it, somebody would be able to, you know, train an AI to to predict how you behave in a certain situation, right?\n\n[19:55] Sachin: Yeah, absolutely. mean, see, look, the security risks of these are high. Okay, because if I knew that by following A, B, C as a flow, and I know that the result was D, I could reverse engineer it.\n\n[20:09] Vikas: Yeah, yeah. I think and if you also I realize that Thomas in his blog made this point, similar point because enterprise have gone through this this whole lesson when when they when the cloud data warehouses came around, right? So they handed over their both data and compute and they watched as the most strategic asset in their business, right? how they operate became somebody else's leverage. So, which is why 10 years after the new formats like Iceberg exist, which are open table formats, right? And they are kind of winning. If you see a lot of people are now moving to open table formats, because enterprises realize that, know, seeding control of their data to a single vendor creates a... an immense lock-in and could hurt them surgically, right? Because, that data itself could become leverage against them because in, because cloud models are around, you know, egress cost and data volumes and big data by definition was the volume didn't lend itself well to be taken out or moved around very quickly, right? So I believe context RAF will have similar set of challenges.\n\n[21:22] Sachin: Yeah, absolutely. think that's a very, very interesting point you've raised. Because look, it is... I guess this is a pattern that every new solution will go through. It will start with a little bit of denial and it will start with people pushing it back. But I think it will follow through the flow, people will understand how to use it. See, look, I don't see... How are we going to come out of this? I don't think there is a way out. I think this is a system that's a must have, especially given the fact that we're going to have a lot of agent systems in the future. So I don't see how we're going to come out of this. But I think as more and more companies start to evolve in this category, customers will make decisions about what to do. Should we give this to a startup? Should we give it to any, even a traditional company? Should we give it to them or should we own it? Because you're right, Until now, people just had the data but they didn't know what was done with that data, right? Because that decision logic was sitting with the human. Yes, so it's going to be interesting.\n\n[22:19] Vikas: The human aspect is, and I was thinking about it, see, there are two set of challenges here, One, an enterprise being afraid that their most valuable data is sitting with somebody else, and somebody's learning or sort of building insights on their data. Even though that may not be a public insight, but still somebody else has that insight. Now, the second aspect is what you mentioned is a people aspect, right? Now, imagine you and I are in a company with our set of tribal knowledge sitting with both of us, right? Now, as an org, the org is telling us that, look, you are to take the tribal knowledge out, the exception logic that you make for certain sort of engagement or deals. know, we always do X because of Y reasoning, which today lives in your head. The company is saying that take this out of your head and put it in a data store. Yeah, right. So I believe there will be a little bit or rather a lot of resistance from people as well who would be sort of because there is a anyways there is a little bit of you know fear about the AI led job replacements, right? So people are going to not jump as the first thing, right? So people will be very careful on if they give away this IP that's sitting in their head, the tribal knowledge, then they would probably be not required to make those decisions. Probably they are giving the autonomy to an agent to make the decision on their behalf. They'll be reduced to approvers of those decisions. one is the need While it's a trillion dollar opportunity, yes, but enterprises are not going to hand over that data to a dozen of  vertical agent startup and giving them their what's called as an operational DNA.\n\n[24:08] Sachin: Yeah,\n\n[24:08] Vikas: So because come to think of it, given how AI is going towards commodity, so LLMs are commodity and I don't think agents are going to be different, but context is going to be a strategic asset because that's not going to be a So enterprises will want to own their context with maybe an open and federated context platform that any agent, right?\n\n[24:24] Sachin: Absolutely.\n\n[24:34] Vikas: be it in-house or a third party agent or, you know, they can own and they can choose what data they expose to these agents. So it feels like it's going to be more and more of anybody who builds, maybe it could be one of our big techs or vertical, anybody who builds a truly open source or open data format context graph is going to have... going to see a lot more adoption than these closed enterprise contexts are.\n\n[25:00] Sachin: Absolutely, think, look, I completely agree. This is interesting, right? And the moat would sit with the person who manages to build this enough to have as much tribal knowledge of the company as possible so that these agents can take decisions, right? Because eventually you want decisions at scale. So yeah, I think...  This will obviously create a new vacuum for security of these agents. It may create a vacuum for, let's say, how do you provide infrastructure just to build the agents, to build the context graph, everything so that the ownership sits with the company itself and doesn't go out. So I think, look, there is this multiple roads and pathways that this can take. It is definitely going to be interesting on how companies are thinking about it. I mean, obviously in the capacity of our In our capacity, we've seen a lot of customers, a lot of companies talk about it. They have expressed interest. They've also shown the fact that this is important because now as and when they start employing AI systems to become the front end of interaction with their customers, they're seeing that agents need context, which was initially sitting in a human's mind. So I think it's going to be super, super... fun to figure out what roads these companies take. And again, like every technology, are always going to be players who will open source will thrive. People will create open formats to write some of this. There will be companies that will take a lot of these formats and build companies out of it to capture the trillion dollar opportunity. So I think it's a space where a lot of people will play. We saw this last time in the data world. We'll see it again in the content graph world. I think it's only a matter of time.\n\n[26:42] Vikas: Yes, and I think we're not making any conclusions today, but I think it gives a framework on what enterprises will assess or think about these things. think probably in future episodes, we should also think about if somebody has to build something like this, what should be the building blocks? And we will have other sort of you know, contours to this problems as we go further, right? And then I think there are other class of problems, We are talking about enterprise context, like what about startups? Do smaller companies, do they need context as well? Should they start? Maybe it's hard to capture the historical information, but should they start capturing things?\n\n[27:17] Sachin: exactly\n\n[27:26] Vikas: as they make decisions should they capture in a special class of database. So I think there's a lot to be discussed as we unpack this. if you have any, you know, if you're listening to this and if you have any comments, feel free to write to us or comment. We're happy to hear what you have to think about this, about the enterprise context graph, right? So yeah, I think it's... It's interesting that suddenly everybody's piling on this debate.\n\n[27:51] Sachin: Yeah. No, think we have left our listeners with some food for thought. If you weren't thinking about this, our recommendation would be to start thinking about it. If this field is in close alignment to what you do in your job, I think it's important to figure out a way to think a little hard, think a little deep about it, because this will come in some form into your company. So give it some thought. We are also open to some of the ideas that you think we know, maybe we have skipped a few things. Maybe we just didn't bring it up. So, but you know, write to us and see if there is, if you have some opinions about this. Have a great evening. All right. Bye.",
  "generatedAt": "2026-01-20T00:06:05Z"
}