The Trillion-Dollar Context Graph Debate: Jaya Gupta & Arvind Jain

Added: 2026-05-05

[00:00:00]I think what's going to happen is that enterprises where they actually have a layer where there is institutional memory that compounds and decision traces that, you know, improve through tracking them and ensuring them. I think those will will actually have like a durable moat and an advantage that >> [music] >> is harder to replicate.

[00:00:22]Thank you, Jaya and Ervin, for joining me today to talk about context graphs.

[00:00:26]Thanks for having us. So, context graphs were made popular by your post, Jaya, um where you called it AI's next trillion-dollar opportunity over the holidays. Um and here at Glean, when we saw your post go live, um we were immediately excited because it described a lot of what we've been building here at our company. Um and you gave it a name.

[00:00:47]Um so, we have been really focused on using graph-based structures to better understand the relationships within enterprises and also to understand processes, too.

[00:00:58]Um and your post took off we think for a couple reasons, um specifically in the in the technology space right now.

[00:01:06]For one, almost everything in the enterprise is now digitized. So, systems of record um live online and SaaS applications, recording the decisions that were made.

[00:01:18]Work is happening through all different productivity applications today from chat, email, project management tools, even meetings are now getting recorded.

[00:01:29]And so, you just have this really rich digital footprint within a company.

[00:01:34]But at the same time, a lot of that continues to be siloed in the same way that it has been in the past.

[00:01:43]Um and that is partially because there wasn't really a good reason to need to join all of this data together um and to understand these relationships and these processes.

[00:01:54]Um because a lot of the work that happens is distributed across these applications and it's a lot of work that is highly judgmental. Um for instance, figuring out when to escalate a support ticket very much goes into a lot of different judgment calls around the customer the issue with engineering um how the account team responds. And that's what ends up deciding whether to escalate something versus maybe what is just codified in internal documentation.

[00:02:27]But at the same time that we're at this place where everything gets digitized, we're also at a point in time where there's actually something that could use that digitization to better automate your work.

[00:02:39]And that's what we see with agents.

[00:02:41]Um and so, if we can give agents traces of what is happening as context cross work, they can actually take on more work in the enterprise and start to be able to understand and act like an enterprise does.

[00:02:55]Um so, Jaya, I know that I've just described context graphs the way that we've been interpreting it here at Glean, um but you were the one that came out with a seminal post. So, I'm curious to hear from your perspective, what made this urgent for you?

[00:03:09]Yeah, it's a good question. I think three things really changed at once and I think one was, you know, last year we kind of moved from chatbots to agents and I think in the chatbot world, you know, context helps model generate a better answer uh for a human to review, but in the agent world, you know, the model is actually taking the action. And so, I think a chatbot might give you a bad answer, you can ignore it, but an agent acting on a bad answer, you know, before anyone ships it, like that's it needs context. And so, I think that's one.

[00:03:37]Two, I think the models got a lot better and so, paradoxically, like that kind of made the problem actually more obvious.

[00:03:44]Like, I think a year ago when something failed, people were like, okay, the model just sucks, right? And then now it's like, I think the models are, you know, actually good enough that even when they miss something obvious, you kind of realize that the gap isn't really raw intelligence, but um it's it's something about institutional memory per se. Like, what, you know, a decent employee kind of knows what happened last time, you know, what were the constraints. I think you mentioned the support ticket example, it's a great one.

[00:04:09]Um and so, I think that is like point two. And then the third, you know, what we were seeing from all of our portfolio companies where founders in many, many different categories, whether it's like finance or clinical reasoning or um production engineering, support tickets, security, they were all sort of describing this missing layer, but, you know, in different language. Um and when we named it, I think, you know, as you mentioned, like people inside companies like, you know, OpenAI, Anthropic, Palantir, CEOs of public SaaS companies all told us they were kind of like circling the same concept for months without being able to to put a name for it. Um so, I think those are kind of the three three main uh things that changed.

[00:04:50]And Ervin, from your perspective, you know, you've built multiple generations of enterprise technology. What makes solving this context problem different?

[00:05:00]Well, I think the more than solving, first, the expectation of what people want to do with that context that is actually the uh the most challenging part because now everybody's seen the the reasoning, the intelligence of AI uh and they expect AI to to do all the work that we do today um in the in the business.

[00:05:24]And uh and it's a it's a it's a big expectation.

[00:05:29]And there's a lot of gap still, you know, in terms of connecting, you know, that intelligence uh of the model with the context and the ways of working of your company.

[00:05:41]And so, so so we are so I I when we think about like building this um context graph inside inside the uh like, you know, like, you know, at Glean, one one of the key challenges is that you know, we don't really have a uh way to sort of scope a problem. Like, you know, that, hey, this is this is the ultimate sort of way this context is going to get used and therefore, like, you know, let's prioritize and let's build, you know, start building the context in a in a in a particular order. You know, that that that's actually one of the most challenging things about like, how do you like fundamentally absorb like everything about your company, how people work, the ways of working because there's so much information, so much data. As you said, like, you know, like every conversation now is also getting digitized. So, so taking this really like uh broad set of information and you know, in different formats, different sort of styles, um how do you combine all of that with the power of the language models uh and then create a usable artifact from it. A usable artifact that actually does allow people to go and create those agents. Um so, it's it's a it's a challenging problem, it's also fun, though, like, at the same time. So, like, everything that we're doing here, it feels like um you know, things that we haven't built before. Like, brand new, exciting things.

[00:07:03]Yeah, I think another challenge you're getting at, Ervin, is when it comes to all this unstructured data, figuring out what to do with it has always been a problem in the enterprise and like, how to make it usable. It's really figuring out where those key use cases that the enterprise needs in order for AI to really feel like it can take on more of this work.

[00:07:22]Um I mean, it seems to be broad agreement both between you and and also what we're seeing in the space that context is really key for agentic performance.

[00:07:32]Um but I'd love to hear a little bit more about how this problem manifests specifically for enterprise customers and and even your portfolio companies.

[00:07:42]So, Jaya, from your perspective, where are your portfolio companies seeing agents stall?

[00:07:52]Yeah, it's a good question. I think some of it is um you know, missing data and then I think some of it is also a missing understanding of how decisions are actually made. Um I think, you know, after the article, I think like, you know, majority of F500 companies, either their CIOs or CFOs reached out and after talking to them, I was like, okay, your problem is definitely like data. Like, I think, you know, systems are definitely fragmented. I think they're complaining about like, migrations being incomplete and that I think the infrastructure was not really ready for, you know, agents to operate cleanly across it. Um but I think what's interesting, especially after this article, was that you know, you talked to some companies and even where the data problem is partially uh partially solved, the agents, they, you know, they would say are still kind of like stalled. And so, I think this was where, you know, many, many CIOs resonated with the idea of like, there's we actually don't have an understanding of how decisions actually get made inside the organization.

[00:08:51]And um I think, you know, the demand was like, hey, uh the demand was coming from these people from like, hey, we we really, really want to be able to track like, why decisions were made and um I think it's interesting because it's like, you know, we have a few portfolio companies that are in and around this area at the application layer. So, we have a company like called Player Zero where they're doing this in production engineering, effectively building engineering world model and saying, hey, when something breaks, um you know, the system should have an understanding of why it changed, what change, you know, what change and what might break next.

[00:09:28]Um and so, I think it'll be interesting to see what exactly emerges, but we're seeing, you know, application companies build the context graph uh as, you know, maybe a feature and then as well as like, platform companies starting to emerge and incumbents uh you know, from all of our saying, hey, we want we want to build this as a platform. So, I think it'll be interesting to see how this plays out.

[00:09:51]And I think you get at a really important point, which is there's data and then there's orchestration. And we haven't really seen a lot of that data move into how do we use it not for traditional forms of orchestration, but for agentic orchestration. And with these LLMs that are non-deterministic systems, and there has to really be this marrying between the two in order for this context to be helpful and actionable.

[00:10:19]Um Arvind, from your perspective, what are you seeing in terms of the decisions of work in workflows that are that are out of reach um for AI inside large organizations?

[00:10:34]Well, first um if you can really think about um like you know pieces of work that are happening like for example you know take health care and uh take the routine sort of high volume tasks um which you would think that you know you could automate fairly easily. Like take claims processing.

[00:10:56]Um the problem is that the like any given claim um that is submitted there uh there is there are so many unique elements of it.

[00:11:09]Like you know it doesn't fit any playbook. Like you can create a playbook, you know a playbook will say that hey look you know if it is on this procedure in this context and like you can you can like you can create these really really complex playbooks um but still like you know you always are thrown off like you know every new claim has something new. Like you know which was which could not be covered in the in sort of like you know those rules that you have. And that's why there you know there are those you know um the those folks who are actually then making that those human judgment. Um so these these kind of things you know will sort of continue to remain uh a challenge for systems. Like so and and when you think about AI AI is actually probably you know taking us an order of magnitude higher in terms of how much work it can automate. Like before like claims processing like maybe um you would take 80% of the claims you know you could automate you know the the approval or denial of them based on the on based on the playbook that you have.

[00:12:01]Now AI gives you like a little bit more subjective analysis and you can get from 80 to maybe 90%.

[00:12:06]Um the what I think we we will still continue to see ultimately um humans you know like who have to you know make those final calls on many of these things.

[00:12:18]Um and and I see this you know across the board like you know you you get MSAs uh that you're you know that you know that you have sort of like you know you're you're signing new customers, you're approving a new contract. Um finally like you know what I've noticed is that uh and I'll tell you this example like the we have we have the policy, we have the procedure, and yet the the person from legal will come and ask me that hey business you have to make a decision.

[00:12:41]Here are the pros of pros and cons of like you know you know like you know if I approve this is this is what's going to happen. If I don't approve this is what's going to happen. And and the like there is no like you know past example which matches exactly like you know this new contract, this new customer. And so then it like you know then it becomes a problem not so much of did I have the right data right the right context. It's a it's a problem of like ultimately like you know somebody has to make a call. Somebody has to make a business decision which is you know brand new. And it happens more often than than you would realize. And so so so when you talk about like you know what what remains outside of the scope of AI like as such like nothing like you know you can take most of these processes, you can start to have AI do more and more work for you, but I think there's there's probably still the final like in many cases a final sort of set of decisions you know that you know we will still need humans for.

[00:13:35]Yeah, agreed. And it it's interesting too I think with context and and kind of what you're you're getting at here is you can learn a lot from these different processes, but then at the end of the day you have to weight all of these decisions from different stakeholders to make that type of judgment call.

[00:13:56]And hopefully you can get examples of what this looked like before.

[00:14:00]Uh but you can't always be able to tell in this one new unique situation what can you do. And I think that is one of the the beauties and kind of what makes knowledge work actually very fun too is that you are responsible for making a lot of these decisions and making these tough calls and figuring out how do you up-level what you are given before in a way that is is unique um to your organization. Um and I definitely agree.

[00:14:27]I think humans are going to be a really big part of of agents. There's going to be a lot of great interactions between them moving forward.

[00:14:36]Um so Jai, you you touched on this a little bit in your a little bit earlier, but I wanted to talk a little bit about the role of orchestration and the role of data.

[00:14:45]Um so there's another debate that's kind of coming into play here around context graphs about where in the the AI enterprise stack does this fall. Um does it sit in the orchestration layer where agents are learning from the the traces of their past runs? Does it sit in the data layer um where you look kind of beyond those individual agent runs to the systems and the human interactions, understand what actually happened?

[00:15:13]Or is it you know some combination of both?

[00:15:16]Um so you've you've put out a a couple of pieces around it really needs to fall more in this orchestration side. So I'd love for you to expand around why you think orchestration ends up becoming the owner of context graphs.

[00:15:30]Uh I will I would actually say it's it's both. I think my opinion has now evolved after you know 25 more conversations the past like few few weeks. So I think um you know I I think it's both. I think uh you can kind of you know one of the things that's interesting is like where is state accumulating? And so I think you could kind of see that in like workflow automation companies. Like there's all these like Zapier or Workato, Tray.io or something like that.

[00:15:58]Um which are all super useful um but like I think what they mainly accumulated was just execution logic and not really building like a memory of like you know why did that workflow exist, you know whether it led to a good business outcome, what exception um came up last time or kind of like how some team um adapted the process over time. And so uh I think something similar is at the enterprise layer too. Like you know ServiceNow and Moveworks you know own the ticketing flow, but you know neither one really owns like that full decision trace.

[00:16:30]Um but I I wouldn't you know I wouldn't go all the way to say like orchestration doesn't matter. I think it's it's really really important, but I think kind of as you said like you know the power will kind of like sit where uh the execution as well as like the learning are kind of joined together.

[00:16:47]Um and I think the distinction is also collapsing and this is kind of why you see so many companies that are now like saying that they're competing or doing the same thing. Um and also why I think when this article got posted it was like a it was a huge mix of like data catalog companies, infrastructure companies as well as application companies that were kind of saying that we're all doing this and um and I think everyone is kind of converging towards like a very similar architecture.

[00:17:16]Um but it will be interesting to kind of like see how it plays out and where the power was. I think it'll be both though.

[00:17:21]And and Arvind, if you maybe zoom out a few years, where do you think the control point for AI in the enterprise ends up?

[00:17:30]What do you mean that I think like so if you if you collectively take you know like how work happens in the enterprise um there are many many layers um uh like of like technologies you know that sort of are going to contribute to uh ultimately like you know that business process getting completed. Um there's going to be the system of records. Um that's where um most of your uh information in the company is stored.

[00:17:58]Um there's going to be uh you know these um sort of knowledge graph you know kind of layers where uh it will basically store a high level of abstraction. And for example it will sort of understand like you know how a particular business process you know gets executed. Um >> [clears throat] >> and then of course there is the the layers above like you know which are making it taking advantage of the data in the systems of as well as you know there's sort of more derived artifacts like uh in the knowledge graph. And they make use of it to sort of actually you know make work happen in the enterprise. So when you think about power or like you know where there's most value like it's a hard question to answer. Like I think all of those are equally important. Um like you know building a system of record by the way like often times you know we start to talk about that as uh that's trivial like you know that's not value. It is actually quite a lot of value. Like this is like you know if you lose that like you know your business stops operating. So so so I so I so I so I I don't really compare that way. But I think if one thing that I would say is that the layers that are building that are sort of capturing you know the the human activity that are you know capturing sort of like these traces and learning like you know how work happens um those layers by themselves are also like just think of them like not more than uh a system of sort of like you know work or whatever you want to call it. Uh the real value gets you know gets generated when other um applications like specific agents in the enterprise make use of that information effectively. So so I actually like you know for us Glean like you know since we are one of those knowledge graph layers um I think of of us in the same way like you know we're one of the systems of record for this particular type of information which is how work gets done. But as as on on a on a on our own like you know we don't really add value to the enterprise. It only it only sort of becomes valuable when you know, all the individual agents that different departments in the teams, you know, in an enterprise and they are actually starting to build those agents and they're able to effectively tap into like you know, the information the traces you know, from from our system is when actual real business value gets generated. So it's sort of always like you know, it's it's not one place like you know, you the the value is sort of like you know, it's is it comes you know, from all these layers together.

[00:20:21]And Jai, you actually recently wrote about this too, which I think you're also getting at Arvin Rear. It's the layers, but it's also creating feedback loops across these layers, too. So if an agent takes an action, how does it learn from that action and how does it learn from the human interactions around that action and how does it learn from the applications that it that it touched? Um was that successful? Did it help with automation, etc. Do you want to expand at all on some of your more kind of recent writings on this area?

[00:20:53]Um yeah, happy to. Um I think the so the analogy that we made and published yesterday was really like I think if you think about all the consumer platforms and and I think Arvin, you were ex-Google, so you know this better than I do, but um you know, I think you think Netflix, Meta, Amazon, TikTok, all these other ones. Um I think that they instrumented behavior with a ton of granularity, you know, what what you clicked, what you ignored, what you hovered over, all these sorts of things and and I think this is actually interesting because Glean is doing this for the enterprise in in terms of search. Uh but um I think you know, B2B is is fundamentally different because you have like these negotiations across like different groups, like sales and finance, right? And um each of them kind of carries like a different incentive uh and and have has different constraints, right? Like your sales team, they want velocity or finance team wants to talk about margin.

[00:21:48]Um and so and I think today enterprises have really lacked the instrumentation of that reasoning that connected the action to the outcome, but but I think um you know, context graphs, decision traces, these are some of the things that are going to enable that for for going to enable B2B companies to sort of have um you know, I call it the ability to you know, build their own Google.

[00:22:13]One of the reasons I I think if if we go back to your original post, it got a lot of I got a lot of attention is because you also made a claim around context graphs.

[00:22:23]Um and that claim was that they were a trillion-dollar opportunity.

[00:22:26]Um and I think that implies something about them that they're not just a feature in part of a stack, but they're actually a platform.

[00:22:35]So I'd love for you us to pressure test that idea a little bit.

[00:22:40]What assumptions sat behind your trillion-dollar opportunity for me, Jai?

[00:22:46]Yeah, it's a good question. Um I think that one assumption is that this is going to be both the platform opportunity as well as an application opportunity. I think Arvin kind of said something uh a little bit similar, but I think you know, the at the application layer, the assumption is that companies can build um products that kind of sit deeply enough inside these real workflows to capture that sort of decision-relevant memory uh as a byproduct of doing these full work. Um I think at the platform layer, the assumption's a little bit different.

[00:23:15]It's that you know, there's enough shared infrastructure across workflows, memories, permissioning layers, right pads, feedback loops, all those other you know, enterprise things that you can kind of build a horizontal layer where many applications can come on top of them um and maybe be built on top of that uh horizontal platform.

[00:23:34]Um I think that's like the really big prize that you know, everyone has their eyes on. Um I think that there's like another third and fourth assumption, too. That the third might be like you know, will enterprises treat these decision traces as strategic assets that they own?

[00:23:51]Um like for example, the system captures um like how a finance team made some sort of judgment. I think the the legal example is actually really good. Like when your legal team comes to you and and you know, ask you to make a call, um what exactly is that asset? Who's going to own it? Is it is it enterprise data? Is that vendor exhaust? Is that model improvement data?

[00:24:13]And so I don't think that there's like a clear consensus here. I'm sure Palantir would say something different, Glean would say something different.

[00:24:20]Um and I think the fourth is that you know, I don't think foundation models will absorb too much you know, too much of this stack uh over time.

[00:24:30]Um and I think that you know, the hard parts here are you know, workflow embeddings, permissions, uh all the things that I mentioned earlier.

[00:24:39]However, like I think with with how fast I think these foundation models ship, it's I think it's naive right now to say that uh that they won't capture some value.

[00:24:49]Um so I think and I think those are the main things that um you know, we kind of assumed. Well, I'd love to hear your take, too. Yeah, I'd love to hear Arvin's take on this, too. Who who owns these these traces, Arvin? Maybe to start and where do you think the industry is going and where is it maybe getting a little bit ahead of itself, too? Well, I mean there's no question in terms of who ultimately owns um the traces and actually let's just so that you know, we're on the same page.

[00:25:16]We're talking about um the views, observations.

[00:25:22]You know, all of how work is happening.

[00:25:24]We are uh monitoring decisions that people are making and as a result now we know like you know, why you know, why we actually make certain decisions and how we do certain pieces of work. This is the this is what we mean by traces, right? And and this knowledge, this this um this sort of know-how, it belongs to the enterprise, you know, where this data is being collected.

[00:25:48]And so like we don't we don't like like personally, I don't believe in this model of you know, an enterprise software that um that a company comes in and you let that you know, you let that vendor sort of observe the way of workings of your business, they you let them observe you know, the data that's actually inside the company and they build these observations which are fundamentally about how your people work uh and what your business is and then they say that hey like you know, like you know, for you to have this information, you know, they're going to charge you for that and you know, and it's owned by them. That model doesn't make sense to me. Uh you will see um the I I think there will be like you know, more and more uh clarity on that front that enterprise data belongs to enterprises. Um all the derived learnings from it also belongs to the enterprise and like you know, if I'm a customer and I work with a vendor like you know, I know that they're learning on you know, learning about our business, but I need to actually own all of that data. I need to have the freedom to to actually replace you know, that vendor with some other um but my data, the intelligence, you know, the derived intelligence of my data belongs to me.

[00:26:57]Um so that's so that's that's that's that's the right model like enterprises have to actually uh demand and like even though like you know, we're a vendor like you know, that's like you know, and we should you know, we build a lot of that uh sort of learnings inside our system, but but we don't we never believe that you know, we own it. Like you know, it's really our customers.

[00:27:16]Um we also build memory uh when you think about like like Glean like you know, the once once once once Glean is deployed, it's actually learning about every individual and their preferences, how they like to work and and it's something you know, it's a we actually store that memory like in very in a portable format and you know, and it's you know, it's always sort of belongs to the customer and they're free to take that memory and and and and expose it to all the other AI applications that they have inside you know, their company. So so that's like you know, the these are like because we're not no longer talking about like you know, uh like a basic software system. Now we're really talking about like like the core of what your business is, AI is actually running it and that has to be owned by you.

[00:28:02]And I think you're also getting at something Arvin, too, which is in order for this to be useful, these traces have to be useful, they have to have some sort of application and that's also one of the value that you would get from going with with a vendor is that you get that continuous application and also that continuous learning.

[00:28:20]Um but as an enterprise, this is also becoming part of your your data stack um and it's a really important part of your data stack, too, um that you should you should always get to to be able to have control over.

[00:28:34]Um I'd love to close with maybe quick predictions um from both of you.

[00:28:41]Um if your view of context graphs it is is spot on, what does this look like 3 years from now?

[00:28:50]Uh I think if if this used right, the the question will be not like you know, who has the best model, but I think it'll be you know, who who has accumulated the deepest sort of decision-relevant memory uh inside the enterprise and um the second question being is that memory actually like making the system get better over time.

[00:29:15]Um and I think the companies that start capturing like decision traces really early will look very very different from the companies that don't. Um and I think even like beyond that, the bigger shift will probably be more organizational, too. Like I think right now um there's a large amount of like coordination overhead that exists because I think context doesn't really have like any place to like really live.

[00:29:39]Um and so uh I think I think what's going to happen is that you know, enterprises where you know, they actually you know, have a layer where there is institutional memory that compounds and decision traces that you know, improve through you know, tracking them and and storing them. I think those will will actually have like a durable moat and an advantage that is is harder to replicate.

[00:30:09]Yeah, and I and I would I would add to that the I think AI today like is still like you know, in very early journeys like you know, I I keep sort of like you know, as I talk to you know, enterprises um there's a lot of excitement, you know, with AI, um lot of use cases, people are aggressively deploying it like spending a lot of money for sure on AI.

[00:30:28]Um but it is still uh early days like in terms of like realization of business value for for most of those um enterprises.

[00:30:38]Uh and this is this is actually the gap like the gap today is that AI in most of these enterprises is starting um like a first grader.

[00:30:47]Um it has to learn everything, you know, on its own through feedback, you know, from the user within you know, within that AI tool as opposed to sort of sort of starting from a much better place where you're able to tap into that human intelligence um that sort of you know, decision making um uh like sort of like you know, intelligence you that you have inside the enterprise.

[00:31:11]So, I think this will be the key um reason like you know, why you will actually start to see AI make real impact um and in many uh business processes actually take the role of the humans um so that and like and and you'll start to see the fundamental shift, you know, where the shape of the organization as a result will start to also change.

[00:31:34]Now, you get you know, AI that looks maybe less generic too and and much more at the way that your enterprise functions and it becomes not just your data but a lot about your people, your culture, your processes too, which is part of each company's competitive moat.

[00:31:52]Um so, very interesting times ahead.

[00:31:55]Thank you, Jaya and thank you, Arvin for joining me. Really appreciate it.