Why context graphs are the missing layer for AI

Speaker A: Thank you, Jamin. Thank you Animesh, for being on the pod. Jamin, you're the first investor ever on my pod.

Speaker B: Uh, wow. I'm very honored. I'm very honored.

Speaker A: But you created such a stir with your systems of record post and Jay and I felt so compelled to respond. So maybe we should let you kick it off by just talking a little bit about your thesis.

Speaker B: Yeah, of course, of course. And the good news is you did bring a CEO, so there is someone smart on the call who will actually share insightful, insightful stuff.

Speaker C: The pressure is on the, the origins

Speaker B: of the post was basically a lot of conversations I have been having. I'm sure both of you have as well, with different folks who are productionizing AI workflows and applications and a little bit of this, uh, kind of like a letdown of like, oh, they're just not, they're not as good or they're not as impactful yet as, as I would have hoped. And, and then there was just kind of some triaging into well, what are the issues? And I think kind of what I found in having a lot of these conversations was agents are great, but when you have a lot of them running around in kind of like a very disorganized fashion, things start to unravel quickly. And it unravels because these agents are inherently kind of like read and write. They're reading context from somewhere, but also updating state somewhere else. And when that's happening in parallel across systems all the time systems are getting updated, what you end up losing is ground truth. And I saw a lot of posts about SaaS is dead, software is dead. And it all kind of felt like one way or another of saying systems of record are dead and the salesforces of the world and the servicenows of the world are going to be totally disrupted by some AI native equivalent. And, and I think, you know, I couldn't help but think my own, my own point of view was a little bit more of, that's a little too black and white. Like what we need, what agents still need, is a source of truth.

Speaker C: I mean we think about this in very similar ways. The companies that are able to actually capture what we're, I think colloquially right now calling a context, uh, graph, are going to be best positioned to actually own the work that happens on top of it.

Speaker A: Right?

Speaker C: And at Player Zero, we often think about like production engineering and what is the work that actually needs to happen in order to run and operate production. And you know, I think from a builder standpoint, the way this has evolved, right? Like less so from a market segmentation category creation. But just like, what is, what is the bottleneck and how has that shifted, um, over the life of, you know, agent companies that have evolved over the last two years of constraints moving around and models getting better, Um, a lot of what we have to think about when we're building agents is, you know, what are the priors that the agents have to have in order to be effective in any organization. And I think one of the most universal shifts that have happened, um, I think especially over the last, call it, six to eight months, as models have really kind of gotten really, really good, is that all of a sudden we're now seeing them actually be truly agentic, right? To actually have agency, to actually have authority and owning the work as opposed to being auxiliary or additive to somebody else's work. And I think what's been really fundamental in that shift is now all of a sudden they're the ones that are actually owning the decision making and the actual outcome as opposed to someone else owning the outcome and they being auxiliary to it. And with it has come this huge new set of possibilities to actually build around because the richness of how these decisions are actually navigating organizational boundaries. You know, why does this incident, why should this incident be resolved in this particular way? Or, you know, why is this drug being prescribed to this particular person? That reasoning is now all of a sudden being captured and encapsulated in a way that we really never actually had visibility into. And so I think what it implies is that. And, um, jam and just like kind of going back to your initial, you know, the bar for systems of record are increasing, right? They're just going higher and higher. I think that's actually what a lot of agent builders have just been reflecting on. We're sitting on top of this extremely rich set of trajectories that represent really interesting sort of decision making and organizations that otherwise could never even be observed. Um, and now all of a sudden there's an opportunity to build something greater, right? And it doesn't look exactly like the systems of record of the past. It's something different. It's represented differently, it's queried differently. And it also creates a different sense of leverage, um, for what the future of work actually might be.

Speaker B: And I think you said something super interesting. And when I boil down what's fundamentally changing, put aside the technology and if I had to just boil it down to something super simple, you, ah, have software and systems of record that are built and designed for People versus agentic systems that are built and designed for agents. And I look back to the pre. I always love looking at, like, historical, historical analogs. And I look back to the early days of the Internet and specifically the travel industry. And before the Internet, people, you booked travel through a travel agent or you called directly. You know, you had to call each individual airline. There wasn't an aggregation point. And for travel agents, they used something called a GDS system, a global, uh, distribution system, which, it was Amadeus, it was Sabre. And what did those platforms do? They went out to every single airline, all the hotels, and they aggregated the inventory, the flight data, the departure time, the arrival time, the open seats, the price per seat, the inventory. So as a travel agent, you had a desktop app, some terminal app on your computer that had a direct data link to an Amadeus server that you could go to one single place to look at all available inventory. And so the Amadeus and The Sabers, the GDSs of the world, those were the systems of record. And they also owned the ux, the front door. Well, then what happened when the Internet, you know, the Internet's invented. And then all of a sudden, you saw the rise of OTAs. Online travel agencies, your booking.com, your Pricelines, your Kayaks. Now, all of a sudden, as an end consumer, you could go direct. You could go direct to the OTAs. The OTAs on the back end might integrate with the GDS systems because, you know, in the early days, if you're a priceline or booking or kayak, you don't necessarily want to go build direct links to every single airline yourself. You might just go to the single place that aggregated because all that work has been done, the gds. And so what happened is now if we fast forward, we have the benefit of hindsight. What happened to that industry? Did the GDS industry go away? Was it a zero? Did software die? No, Amadeus is still a $30 billion company. However, Booking.com Expedia, Priceline, Kayak. I mean, there's been a lot of mergers there, but like, the market cap of the OTAS is now hundreds of billions. And the market cap of the gds, the old historical systems of record, is, you know, I don't know, 20, 30 billion. And so it didn't go to zero, but there was an order of magnitude larger opportunity that was created with a different. And it unlocked a different end user. Now consumers can go direct. I think the same thing's happening. Like a lot of systems of record will die.

Speaker A: Important thing. Important. I mean, not not to get lost in the discussion is the next generation, the interface in this particular case became 10x more valuable, 10x bigger.

Speaker B: Yes.

Speaker A: Next more valuable 100 and the, the GDS, uh, was unable to actually capture the interface off. So they didn't, you know, would you have rather been an investor if you, if you went back to 99, would you have rather been an investor in booking.com or amadips?

Speaker B: Yeah, exactly, exactly, exactly. And there's just, there's so many parallels there because even today you have Salesforce ServiceNow Workday, you have all these systems and the reality is, is like the people, so many workflows span multiple systems of record.

Speaker A: That's the core insight actually for decision traces, which is.

Speaker B: That's right.

Speaker A: The challenge with systems of record is because they grew up organically, systems of record are owned by individual functions, individual chains of command. But decisions cut across functions and chains of command. Animation should give the example in your case. But in most cases they flow through multiple functional organizations in a company.

Speaker B: Mhm.

Speaker A: Work off different systems of record.

Speaker C: So I mean at Player zero, we think about production engineering as this, uh, as a superset category. Um, and what we think about is, you know, there's all of these, you know, partitioned functions that operate production systems. Um, this could be, you know, sre, this could be support, this could be qa. Right. Or this could even be a faction of your developers that are all independently building a certain view of how your production system actually works. And they operate between all of these different systems of record, which are your ticketing system if you're a support engineer.

Speaker A: Right.

Speaker C: This could be JIRA or Zendesk. Uh, it could be Datadog if you're an sre.

Speaker A: Right.

Speaker C: Or it could be your code base or something like that if you're a developer. And there's so many decisions that are actually being strung across all of these different systems of record. But the kind of organizational dynamics of how is a ticket actually ultimately escalated, remediated and then resolved back to the customer, back to the user is, uh, is actually never truly captured. Um, and so this was actually kind of the core insight behind how and why we built what we built. But I think the really interesting thing that Ashu you mentioned there is in a world where these decision traces actually truly become centrally represented, you truly have a context graph, so to speak, then the entire shape of the organization actually changes as well.

Speaker A: Right.

Speaker C: These functions that I think were built based on labor constraints, for example. Right. We have SREs because SRE requires a specific skill set that is different than the skill set of a qa and that's different than the skill set of a developer. Um, these kind of functional groups of people that we built because we have certain labor constraints and certain systems of record that those pools of labor actually understand can actually start flattening. And I think we've seen kind of parallels to this in kind of past iterations of software. Right? I mean, I think Clay with the go to Market engineer or databricks with, uh, data engineer and stuff like that, this is not exactly the same where they're centralizing some sort of system of record. But if you kind of flatten the tooling and the access to priors to be able to do great work, then I think the functions themselves, cells actually start collapsing in a really interesting way to redefine what the shape of the organization looks like. Um, and I think that's actually something like jam. And I'm pretty curious on how you think about this, where our systems of record actually evolve. Like, how does the organization evolve?

Speaker B: Yeah, yeah. I think it's fascinating. And I go back, I mean the one thing I loved about the context graph post is like this concept of the decision trace and what I'm looking for. Like my wife was in software, uh, uh, um, I'm going to make an analogy and I'll connect it full circle at the end. But like my wife was in software sales forever and if there's one thing I know about her workflow, it's what she hated more than anything was data entry. I talked to this person, let me put the notes in the CRM. How many people are at the org, you know, what are they looking to buy? What kind of discount are they looking for? It was all of that and it sat in her head and she knew it because she knew her clients and she knew her prospects and all of that. But manually entering the data was just super hard. And you kind of had this like 80, 20, like 80%, like wasn't good enough. Like you either had all of it or you didn't. And when I look at the, what I think is going to be most interesting in this whole concept of, uh, a decision trace is, well, how do you capture that? Right? And what I love that you made a very clear analogy in your initial post about, you know, this concept of, hey, we only offer a 10% renewal or 10% discount at renewal. Um, that's standard. We can go higher than that only if there's been some sort of like incident or outage or you know, call it Like a bad experience. How do you go get that? Okay, well, I'm going to go to the ticketing system or the datadog system. I'm going to look for outages, I'm going to look for this, I'm going to look for that. And what's interesting is I might do that in my normal workflow. Um, what systems are going to naturally capture that so that I don't have to enter it? Well, I gave that discount because of this, because I'm just not confident in an organizational's, an organization's like entropy to be able to do that, to like enter that data. And you know, it's kind of like the, the old uipath, um, analogy where it's like, hey, you can just watch someone complete a workflow and then automate it, or you can kind of like describe it and build it with some gui. Um, I think that's what. Or either organizationally we're going to need to train people to like enter why you did things, or you're going to need systems to capture it. And I'm pretty confident we'll get to the latter. It's just a question of how. Like, how do we get there? Do we need new systems to do that? Is it new agents that monitor like that? That's kind of what's most on my mind, like organizationally, of how do we get from A to B?

Speaker A: So Jamin, I think you're spot on. I mean there will be exceptions. I don't want to generalize, but for the large part, getting human beings to explicitly train agents is as difficult, arguably more difficult than getting your wife to sort of give clean data to the CRM. No great sales rep ever wants to do that. That's just not in their DNA. And I think the same is true if you ask human beings to explicitly train agents. So I think most of this data capture will be implicit and it'll be a staircase that you have to deliver enough value in v1 of the product, where the value the user of the product gets far outweighs the training they're providing by being the human in the loop. And then that gives you an opportunity to go to step two. The early movers who move in 2026 and begin to capture these decision traces, translate them into a context graph will create this flywheel that then becomes the moat. I'd love to get your take on how do you think this plays out for systems of record? Are they all equally at risk, Jamin, or do you think certain categories are more at risk than others?

Speaker B: I think you will see the profit pool, like the relative profit pool, the decline for the classic systems of record at the benefit of, you know, modern players. Modern, modern, modern startups. Um, I think most are at risk. I mean even you know that Amadeus wasn't the only one. I mean Sabre was a huge GDS. It got acquired for 5 billion by, I don't, I think maybe private equity went public. Now it's got a market cap of like less than a billion dollars and so I guess it's just trading water, um, what industries, what types of system. I don't know if I have a point of view yet on like which are more at risk. I just think it's, it's going to be a, hey, most are going to be at risk and will kind of start a slow and steady decline and a few will remain.

Speaker C: In that vein, do you, do you think that these systems of record are actually eventually going to converge into a warehouse itself? And then if that, then you have different UX layers that actually operate on top of this. Right. I think there's something interesting to be said about the combination of the UX layer plus the database was like how we could actually overfit a particular.

Speaker B: I mean those used to be together. You had the data plane. I mean look, yeah, we're kind of coming full circle to like what is the comment Satya made like a while while ago about like all these SaaS apps are just like a dumb crud database, right? Like that's kind of what we're arguing like hey, well do these systems of record, are they just reduced to a dumb crud database that can kind of uh, do basic operations? But there's something else on top of the CRUD database that is orchestrating, handling the workflow, executing the workflow and that's uh, I think that's the big challenge for the classic systems of record. Do you slowly get reduced to a dumb database where someone else captures all the value or are you able to innovate yourself and capture that new front door, capture that new profit pool where the challenges are. You're going to have to build solutions that are cross functional and if you're selling to salespeople, you might not understand the workflow of the finance team, you might not speak the language of the accounting team and those are different buyers with different Personas and it's hard to sell to both. And it's like, it's, I think there's a lot of challenges, um, let alone maybe some of the innovators dilemma that all these Players are facing.

Speaker A: Hey, that's why startups exist. Lots of fun opportunities.

Speaker B: We're in the business of venture capital. We got to believe that uh, startups can pose a risk.

Speaker A: When I think about it, the two functions that come to mind or two business, I think about this as business. What business processes will get reinvented by agents and therefore what are the underlying systems of record? I think the go to market process, the business process of going from a lead to customers that are paying and renewing I think is hugely is a huge opportunity for automation and agents. 100% because that process is all done in high cost economies and it's a very communication intensive process. And if there's one thing we've learned about LLMs is they're incredibly good at both generating and processing communication 100%. The other process where that's equally a challenge is the entire development process from you know, uh, spec to managing code in production. There's so many different systems in that process. And you know people have built historically applications to help you do product definition, applications to help you manage code, you know, code repos like GitHub, uh, ticket databases like Zendesk, observability tools, blah blah blah. You can go on and on. And I think uh, animation, uh, you should definitely comment on that. I think those two business processes are hugely at risk. On the other end of the spectrum, I think processes that are really more tightly within the organization potentially, uh, finance, procurement. And by the way, there's nothing that's not cross functional. It's the big aha. They're also cross functional but they have less cross functional implications. I think the underlying systems of record probably have a better chance. But I want to open it up to both of you. What do you guys think? How would you segment the market if your firm picks some stocks too? Jaron.

Speaker B: Yeah, yeah, yeah, yeah. Animus. You should definitely take the code example. Um, I'll go back to my wife like 15 years ago this was her work flow. She hyper competitive person. She would get up in the middle of the night to look at the inbound leads because like how her company operated was you know, the day before there would just be like a set of inbound leads who you know, hit the contact us button, who like booked a demo on the website, who like went to the company blog and like downloaded the Gartner Magic quadrant that was gated by like enter email. Right. There's all these different forms of like inbound leads. You know I think they had some sort of like scoring system or some like basic data augmentation that you know, would add in. How big is this company? How big of a contract could they be? You know, maybe some predictive stuff around, like how likely they were to convert. And it was basically like a race, like how can I put as many in my name that I can go after? And you're like, that's crazy.

Speaker A: That happens in venture firms too. I hear.

Speaker B: Exactly. It does, it does. But it was kind of one of these like, okay, well let's look at an agentic future. I want to go, I hit like book a demo. Why shouldn't, instead of book a demo, why shouldn't I be able to just hit a button that says get a demo and pop up an AI avatar that can describe the product, that can answer questions about the product, that can give me a demo and guess what I'm ready to learn in that moment, Maybe my boss just yelled at me because I don't have some solution for this. And I'm feeling the pressure now and I'm ready to buy now or I'm ready to learn more now instead of like, hey, I hit that book demo and maybe someone reaches out to me in two hours, but maybe I've moved on, maybe I'm not thinking about it anymore or maybe it's not till the next day, right? Like so often in the go to market world it's, it's, there's so much urgency that's required because there's options. If I want to go buy a procure a piece of software, it's not like there's just one vendor that's going

Speaker A: to do it for me.

Speaker B: There's maybe five. Right. I think the extreme example here is like a plumber, right? Like when I'm calling a plumber, I probably need that plumber asap, right? Like my dishwasher just broke, my toilet's overflowing, I'm getting on the phone and I'm calling the plumber and guess what? If they don't answer, I'm hanging up and I'm calling the next one right away. And so that's lost revenue for the first plumber that I called that didn't answer. And that's an extreme example, but it's kind of similar in a lot of software sales, which is when I'm ready to get information, I have to, if I have to wait for it, odds are I'm going to go find a different vendor. So you have such an opportunity in the go to market world to just shrink that time from intent to like information Delivered to like software purchased because you don't have to wait for people.

Speaker A: Historically, the way this has evolved, and that's where the systems of record have evolved is they've chosen niches and then sort of, you know, kept adding capability and functionality. Uh, and I think that's a very plausible path and time will tell. I think the other path though is I think it's entirely possible that people will build cross functional workflows. Mhm. But do it for a narrow segment of customers. That segment could be an industry vertical, that could be a size segment, that could be a geography segment. They actually start stitching together. Not everything. We're stitching together multiple functions, multiples. Because that's how you capture the decision traces.

Speaker B: Yeah, that's right.

Speaker A: You do it for some customer segment where, you know, you don't have to have solved world hunger in terms of features and capabilities.

Speaker B: Right, exactly.

Speaker A: Anyways, you're doing this in real life. Jamie and I just, you know, we

Speaker B: just pontificate about it.

Speaker C: I have to sell it. I have to sell it at the

Speaker A: end of the day.

Speaker C: Yeah. Thematically, I think the best places or the greatest opportunities uh, to build this is in places where there's enough urgency and incentive, uh, to ultimately give agents authority to actually make these decisions.

Speaker B: Right.

Speaker C: And I think we talked about this earlier, right, where you know, these agents are now all of a sudden able to operate across all these different systems of record. I have to give the agents agency or I have to give them authority to actually go do this and make decisions ultimately to, to be able to capture this decision trace. And so this has been really interesting actually for us where, you know, JAM and you mentioned they ultimately start as like point solutions. And I think choosing that wedge is actually incredibly important. Um, for example, for us, right, Our wedge tends to be support engineering. Um, and the reason why support engineering is really interesting wedge there for us is because it is laterally available to all of the different functions that are adjacent to it. So it's available to sre, it's also available to qa, it's also available to developers. And the types of systems of record that support engineers need to interoperate between include the ticketing system and the observability system and the code. Right. And so they have to interoperate between all of these different things to be extremely effective. And by wedging there, even though, you know, it sometimes looks a little bit different, sometimes smaller, sometimes bigger than sre, right. It ends up becoming this really interesting advantage to actually start building. Right. This entire Kind of context graph. And um, then the other thing that we actually end up seeing a lot is, you know, ultimately we have to sell this, right? And like how do you sell something that's actually going to be cross functional that's going to change the way you work in a certain way. And I think the real like quote unquote alpha here at the end of the day is that you can actually do each one of these individual functions better because of the fact that you're doing all of them as opposed to doing any one of them in isolation. Um, and so an example of that, and I think you gave one similarly, uh, in go to market example of that in kind of the coding space is like if you understand the intent of how the software is actually built, you can understand how to debug it better.

Speaker B: Right?

Speaker C: I think that's sounds pretty obvious in hindsight.

Speaker A: Right?

Speaker C: But that's not, that's not an obvious thing. When you're thinking about, you know, an SRE's workflow in isolation and kind of going the other way around. When you think about the SDLC and you're trying to, you know, prevent bad things from going into production, understanding how things have broken in the past becomes an obvious, you know, prior to pull on, in order to say, hey, you're changing this area of the code. Here's what might break in the future, right? And so when you actually kind of create a cohesive pitch and you say, hey, you know what, I can do each of these individual kind of point functions better. And the advantage is the fact that I'm actually doing this in some sort of centralized kind of universal way, you end up being greater than uh, kind of the individual parts, so to speak. Like one plus one is equal to three.

Speaker B: You said something interesting too, which is I've, you know, working with a lot of these agent companies, whether it's a sales agent, a marketing agent, a voice agent. Like what I found is what people really want to buy, where you're able to extract a lot of value is when you sell observability. Like that's what people want. They want to know why did it do this when it failed, where did it fail and why did it fail? Like they want to be able to bug it. Maybe it's a voice agent and they want a transcript. Maybe it's a marketing agent and they want to know when did you send and why did you send this, uh, this proposal? It's so much of. And I think it comes down to trust. And you know, I wrote a post A few or last week and it was about like a third authority as the bottleneck. And on a mesh, it's kind of what you just mentioned, which is right now it still feels like we're in phase one. You know, AI is going to draft the email, but you got to review it and you got to click. A person's got to click send and AI is going to flag an alert, but a human has to review it and says, I'm going to go remediate it. Like I think we're in the trust building phase. Like in the same way in the early days of the cloud, like it took time for people to trust it. Is it going to be secure? Is it going to scale? Is it going to do this? And yeah, we just need to build trust with these AI agents and systems with a human in the loop that, you know, the human in the loop can help provide a lot of the context and the decision traces and you know, that will ultimately be the bridge that gets us to these kind of fully agentic and autonomous systems.

Speaker A: So I think you're right though, it's changing very quickly. Uh, what I pleasantly surprised is how quickly business users are willing to hand off trust to an agent to complete tasks. Now in many other cases people are saying, hey, I want to have a human in the loop. Yeah, uh, but 2026 I think is the year that that will change. Yeah.

Speaker B: And the irony here is that that uptake uh, is happening so quickly because the ROI is so painfully obvious so early on. And again, the irony here is all the AI pundits will say like, well, where's the roi? It's like, well the ROI is, is you see it in the uptick, like cursor. And companies like, they're only able to scale that quickly because the ROI is so obvious that the sales cycles are compressed to, you know, minutes, not days, weeks or months. And I think you are seeing that in so many of these agent companies where you see it in 30 minutes as a buyer, a procure of the software, like the ROI is just so obvious so quickly. You say, well, I'm going to spend X time doing this like, or I'm going to compress the time down to basically zero, to use the sales analogy, which, which that's going to help me reach and close customers faster. So not only is it removing monotonous time, but it's also driving business value. And like, like to me, like, you only have the speed of growth of all these companies if the ROI is very painfully obvious. So everyone who wants to kind of baker about where's the roi, where's this and that. It's kind of like, well, like the ROI is, is right in front of you.

Speaker C: I think it's so clear, um, after the first time somebody sees like an incident resolved or the first time somebody sees, you know, a problem that they didn't think about yet. Right. Kind of being detected days before they would have thought about it. Um, the first time any of these things happen, how quickly the rest of the sales process kind of accelerates. Um, I think this is actually a universal, uh, feeling among a lot of kind of AI agent companies going kind of back to the context graphs piece here though I think the question of why are they durable? Also needs to be answered at the same time. And so why is this agent actually going to stick around? And why are you the one to actually continue doing my work six months from now or a year from now when there's all of this other noise in the market? Right. That's where these context graphs actually end up becoming a really interesting factor to create stickiness at the end of the day. And this actually kind of takes me to an interesting question. I don't know if we necessarily want to talk about it right now or later, but, uh, earlier we talked about decomposing systems of record, kind of collapsing them back into warehouses and things like that, and then saying, okay, well the UX layer and then the, the persistence layer. Right. Could actually just be simplified and they can be decoupled. Whereas in the past we've actually coupled them like a Slack or a Salesforce, and now we're actually talking about something similar, right, with these context graphs and these kind of moats of understanding decision traces and decision making and organizational dynamics. And I think there's this question that came up a lot actually in the last week or two, um, after the post that Jaya made, Ashu made and then that I made, um, around is there a context graph infrastructure company or are these actually just the moral equivalent of intellectual property for each verticalized solution? So is there a context graph that is different for a production engineering company versus a go to market company, or is there actually going to be a central universal context graph for every organization? Organization? The same way that there's a single, let's say data warehouse. Right. Um, and I think this is a question that I don't think the Twitter sphere has actually come to an answer on. The companies that can capture the decision traces are the ones that have the UX to run the agents that can actually evaluate the outcome Take the action, do all of these different things and their ability to cooperate with the humans, uh, on the other side is why that UX is important. And so therefore the people who are best positioned to actually own the decision traces are going to be the ones who have at least staked out some part of the world. Um, to say, you know, production engineering is my world or go to market is my world or you know, um, I don't know, like health care reimbursements are my world, um, even though it is cross functional. But to actually say the decision end to end actually is owned here, looking kind of at the spanning kind of parts of a particular workflow, um, and that UX layer, while there are net fewer of um, those I think great context graph companies on the other end of this than There are distinct SaaS companies, um, they still are going to be verticalized and I don't think you're actually going to have a central context graph company, um, at the end of the day just because in order to build that context graph you need to be embedded in a particular workflow.

Speaker B: I think uh, an analogy here might be kind of like the process mining world from the RPA space where you know, everyone kind of knows of the, the UI pass of the world but like a company like Salonis is less well known even though they're kind of like almost like equally big. Um, and their whole pitch was like hey, we're going to help you identify which processes can be automated in the first place. We're uh, going to kind of do that process mining and it's almost like does there need to be an equivalent in the agents world that can help surface what workflows can be automated and how would you automate it? So specifically talking about context graphs, I think the challenge is creating uh, an environment that can capture the telemetry data and the usage. That's just hard. You might not have companies that are instrumented in a way to do it or you look at a classic systems of record, ah, many modern companies like, at least at the early stages their pitch is like oh, we built a different date, we built a flexible data model underneath our product that can cap, you know, it's not rigid, it's not a CRM that's going to have this field and this field and this field and if you want to add a new field, you're going to have to go get some salesforce admin that's going to come in and give you some custom version of it with some new field or some new data Type. We have a flexible data model. And that flexible data model allows us to capture this interaction, that interaction, this thing, that thing. Because first you kind of have to like capture. You have to have the infrastructure to capture the data. You have to have a system that is instrumented to automatically capture it. You need some human layer on top that can verify, yes, this was the reason we gave a larger than normal renewal discount. And so I still think the biggest challenge is going to be capturing that logic in an automated fashion. And it's not just doing that. It's making sure you have the infrastructure that's flexible enough to capture the range of decision traces.

Speaker A: I don't think there is one architectural solution to a context graph. The way you might build a context graph for software engineering or for production engineering may be very different from the way you build a context graph for go to market.

Speaker B: Yeah, and the other thing is, look, let me very quickly give. Not like the bare case on context graphs, but just like where they could go wrong. It's almost like semantic layers in databases. People have been trying to build semantic layers in your data warehouse forever. And all a semantic layer is to say, hey, you have all this data sitting in Snowflake or whatever database. It's kind of what we discussed at the top end of this podcast. How do you define ar? Let me create this, the definition of ar, and I'm going to create the table in Snowflake that has the current ARR and that is going to be delivered via semantic layer. That is the single source of truth, uh, for each metric. You know, and the challenge really isn't technology. It's like people, it's ashu. And I might just fundamentally disagree on, uh, um, how you calculate this or how you do this or when you give this and it's. How do you resolve that? It's not some mathematical equation, it's not some technical process. It's just you and I disagree. And so like the challenge will be how do you resolve those disagreements and context graphs in an organized way? And do context graphs remain this thing like a semantic layer and data that sound really good in theory and should exist, but for whatever reasons over decades,

Speaker A: we've just never, they never become real. No, I think that risk is real. Uh, the other risk I would add, just, you know, uh, uh, to build on that is, I do think in that an example, even when two people do agree, they agree today.

Speaker B: Yeah, that's right.

Speaker A: And truth changes. Yeah. Any agreement you have is out of date, minutes later. How do you manage that evolution or that, you know, uh, truth in many of these situations is not static, it's dynamic. Uh, with time being an important component. Often there are other dimensions other than time. You know, the definition of ARR in one part of a company can actually be very different from another part of the company.

Speaker B: That's right.

Speaker A: Even when there is agreement, that agreement depends on which part of the business or which lens you're looking at. Uh, so I think that will definitely be things that have to be figured out. Uh, but I think the key that sort of Anime said, what makes me confident that this will have an enduring value is these graphs are not being explicitly defined the way semantic layers work. You know, I remember these big workshops. You would get 50 people in a room and you'd all sit down and spend days, weeks agreeing on definitions of 100 things. And even if you had 50 people, you had only 50 people. They didn't represent the organization, they represented a slice. I think here the graphs will be implicitly created by creating value for customers. I want to keep coming back to that because I think it's that staircase of value. And folks that figure out how to build that right staircase of value for customers are the ones that will capture the data that becomes in a flywheel. Before we wrap up in a couple of minutes, I want to change topics a little bit. Uh, one of the things that Jamin, you mentioned in your original post that triggered all of this was the role of uh, uh, the data infrastructure. Companies like Snowflake and Databricks and for that matter, uh, Google, Microsoft, all the CSPs actually have a horse in that race too. Uh, what's your take on that category? What do you think happens to Databricks and Snowflake?

Speaker B: Yeah, I think the bull case is that they become almost the new, like transactional backdoor and the modern Salesforce, the modern service. Now, you know, in the same way, like Oracle bought netsuite. Why'd they do that? Well, because they wanted to own the database and the application. Like, they become kind of like the data app. Uh, future. Now the challenge will be, will be will people build those modern agents on top of the data bricks and snowflakes? Can they be the aggregation point? I think, as your post pointed out, when they're generally in the read path, right, not necessarily like the right path, like they are the aggregation point. Um, now you could build agents on top that can go modify underlying systems. Or maybe they become the new. They become like the OTAs, you know, they become that layer on top of the systems of Record where most of the profit pool ends up accumulating. Um, I don't know. Time will tell. I think there's definitely a possibility, like, will we see kind of like a modern salesforce, a modern servicenow, a modern datadog, a modern splunk built on top of these types of platforms? I think the applications that lend themselves probably look more like data apps, like a sim where you're just sending all your logs and you're searching on top of an observability platform where you're sending all your metrics and your traces and you're alerting. On top of. Um, I think you probably see more AI native applications like that built on top of these data platforms. Um, because I don't think anyone wants to be reduced to a dumb data platform that your finance team uses to build reports.

Speaker C: Um, yeah, I think the, it's hard for a snowflake or a databricks to actually become in the critical right path. Um, I think there's actually a different direction that I would actually think about this in, which is what are the actual primitives required to represent these context graphs at the end of the day? And I think a lot of it actually comes down to there's probabilistic data structures, think embeddings and things like that. Um, I think there's also an interesting blend with the conversations around continual learning, uh, and figuring out actually how to externalize some of the learning and continual learning right into these context graphs so that way it's quickly retrievable and then also internalize some of it. So that way actually the kind of decision structures or kind of like the operational dynamics of a company of who to go to and what are the kind of functions and stratifications, um, and the history of a company, so to speak. Right. That could actually get at some point even encoded directly into the model weights themselves. Um, and so it's an awkward thing to say but like, you know, there's to some degree a database that lives in the model weights. Um, and there's a, some, there's. There's some database that actually lives in the, um, you know, in the embedding store. And I think the, the fact that these data warehouses are not quite oriented around the right path, I think actually makes it difficult for them to actually be, um, the storage layers for these, uh, for these things. But I do think that uh, the front door For CRMs, the front door for like a lot of the canonical systems of record from the last decade are going to collapse into these data Warehouses, um, and I don't see any reason why not.

Speaker A: Fundamentally, this battle will boil down to the changes in the technology that allow you to collapse analytical systems and operational systems. There were underlying technology constraints that created the infrastructure we have today and that we replicate the same data in so many different systems. I think time will tell, but my instinct is the data infrastructure companies are actually better positioned to win than the systems of record companies just because of the relative cost economics and the way they've architected their infrastructure.

Speaker C: But the ergonomics of actually capturing like, so like the surfaces that for example, Salesforce is exposed actually capture judgment. Right. And, or even outcomes, I guess, uh, like more importantly right now, a lot of the premium that you pay for Salesforce is because of the services that they've built to do that. And if I understood Ashu, your point correctly, that you're saying that there's going to be an emergent kind of category of, let's say commodity companies that will basically just create those same abstractions, those same services to capture decisions from different front doors. Is that right?

Speaker A: Imagine a CRM, uh, app that's built on top of databricks.

Speaker C: Okay.

Speaker A: All of the different data infrastructure that databricks has, including HIPAA compliance, SOC 2, blah, blah, blah, they use sort of databricks infrastructure, combine structured, unstructured, uh, operational analytical use cases and build an application layer on top of.

Speaker B: Look at what both these businesses have done. You have databricks moving into the transactional world with Lake Base and beyond. You have snowflake buying, I think it was called crunchy data. You kind of have a lot of these operational analytical databases also moving into the transactional world.

Speaker A: Absolutely. And if you look at the CSPs, they already have both. They keep them in the silos and uh, they will potentially start to combine them in interesting and creative ways. So it'll be, you know, in some ways if the database war is back all over again. I remember the 80s and the 90s when Sybase and Oracle were duking it out. Some of that is going to happen all over again now in the next decade. If you both had to pick One prediction for 2026, what's at the top of your list?

Speaker B: I don't know. I mean, maybe one hot take is I think you see a very large AI IPO, whether it's like one of the big labs and OpenAI and our anthropic, or one of the big application companies like a cursor, where you just kind of have this like very quick founding to IPO Just one? No, no. At least one very large one. At least, at least. I hope we see a lot.

Speaker A: Hey, if there's only one, a lot of growth stage investors are going to be very unhappy.

Speaker B: Exactly, exactly, exactly.

Speaker C: Um, my bet is that the biggest slept on opportunity, um, in AI right now is this concept of world models. And I think we're going to figure out how to define it first, uh, for things that are outside of physics basically. M and I think it's going to be the biggest driver of um, enterprise AI advancement over the course of this upcoming year. Um, over continual learning and perhaps the

Speaker A: risk of stating the obvious, I really think this is the year of adoption of agents in the enterprise and I think it's going to be this massive tsunami that's going to benefit both incumbents and startups and attackers alike.

Speaker C: I thought we were doing hot takes.

Speaker A: These are boring things. Yeah, I'm going to leave all the disagreement to the two of you.

Speaker C: Yeah, well, I have, I have one more. I think actually a lot of AI point solutions are going to die. Um, and I think that's the case because the, I mean the point solution by definition is not one that is actually externalizing its learning right into what we're calling context graphs here and you know, understanding something larger than the individual function. I think with the changing kind of org structures that are natural with AI, um, a lot of point solutions I think are going to start withering away, um, if not actually completely die. And I think the, the durable part of enterprise AI actually comes from using these point solutions to actually build the flywheel. Um, it's a lot of what we talked about today, but I think the outcome of it actually means, means that if you're not thinking about this, it means that uh, there's somebody else who's going to go do it better than you.

Speaker A: Anything you would add jamin on that point? No.

Speaker B: I mean, maybe another hot take. I don't know if it's that hot because it feels like it's already true. I feel like this year the political discourse around AI and job loss is just going to hit a fever pitch and you kind of see already how that that's happening, but I just think there's that is going to be massively accelerated, um, on both sides of the aisle and it will become a huge sticking point for midterms elections. Um, I just, I feel like that topic in particular is really hitting a nerve.

Speaker A: I think it will, it will uh, reach a feverish pitch and it is why I think, you know, we'll be better off on this side of the pond.

Speaker B: Yeah, yeah.

Speaker A: There'll be a lot of debate, but ultimately I'm very confident that in the US we'll continue to move forward while a lot of other people around the world will get stuck in analysis. Paralysis.

Speaker B: Yeah, it's like I can't remember. I think I just saw this, um, or maybe this is old and it just happened to resurface. But there was some, uh, law in Europe that said, okay, the universal adapter we're going to stick to. You don't have a USB or USB C or this or that. And someone had a nice comment that was basically like, well, could you imagine if you struck this law like three years ago, you never would have had like the lightning cable or this or that. Which was just kind of a funny, funny analogy. Yeah. Look, I think the challenge will be like, you don't want to limit the potential, certainly. Um, but at the same time you want to be empathetic to like, what it means for like, real people. Um, and, you know, there always is creative destruction along the way, but that doesn't mean we shouldn't think, uh, about how to deal with that, um, creative destruction in the most empathetic way possible.

Speaker C: Well, have you, did you ready's, um, Aaron Levy's like Joven.

Speaker B: You know, it's on my reading list. It's.

Speaker C: I haven't yet, but yeah, I mean, it said something very simple. Just like, I mean, as, as we adopt agents, there's actually going to be more work than there's 100%.

Speaker A: Right.

Speaker C: I think that's like, we see this already everywhere. Right. Uh, like we have cursor now and we're not writing less code, writing more.

Speaker A: Right.

Speaker C: We don't need fewer developers, we need more.

Speaker B: I will always bet on the resilience of the American people and our ability to move forward with technology in progress and, you know, ability to reinvent what it means to work, how to work. And yes, we will have creative destruction along the way, but, you know, I think that will massively, massively, massively benefit us in the medium and long term.

Speaker A: Well, with that, hey, thank you so much, Jamin and Nmesh, for getting on the pod so quickly and ah, what a fun conversation on context drafts. And I think it'll be a fun year. And once again, thank you for joining us on the pod.

Speaker B: Thanks for having me. This is fun.

Speaker C: Super excited. It's going to be a good year.