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 improve through tracking them and storing them. I think those will actually have a durable moat and an advantage that is harder to replicate.
Thank you Gia and Irvin for joining me today to talk about context graphs.
Thanks for having us. So, context graphs were made popular by your post, GIA, 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, we were immediately excited because it described a lot of what we've been building here at our company. And you gave it a name. So we have been really focused on using graph-based structures to better understand the relationships within enterprises and also to understand processes too. Um and your post took off we think for a couple of different reasons specifically in the technology space right now. For one almost everything in the enterprise is now digitized. So systems of record live online and SAS applications recording the decisions that were made. Work is happening through all different productivity applications today from chat, email, project management tools, even meetings are now getting recorded. And so you just have this really rich digital footprint within a company. But at the same time, a lot of that continues to be siloed in the same way that it has been in the past. Um, and that is partially because there wasn't really a good reason to need to join all of this data together, and to understand these relationships and these processes. 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, 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. 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. And that's what we see with agents. Um, and so if we can give agents traces of what is happening as context crosswork, they can actually take on more work in the enterprise and start to be able to understand and act like an enterprise does. Um, so Jia, I know that I've just described context graphs the way that we've been interpreting it here at Glean. Um, but you are the one that came out with the seminal post. So I'm curious to hear from your perspectives. What made this urgent for you?
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, context helps the model generate a better answer for a human to review. But in the agent world, the model is actually taking the action. And so I think chatbot might give you a bad answer, you can ignore it, but an agent acting on a bad answer, before anyone ships it, like that's it needs context. And so I think that's one. Two, I think the models got a lot better. And so paradoxically, that kind of made the problem actually more obvious. 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 actually good enough that even when they miss something obvious you kind of realize that the gap isn't really raw intelligence but it's something about institutional memory per se like what you know a decent employee kind of knows what happened last time, what were the constraints. I think you mentioned the support ticket example, it's a great one. And so I think that is point two. And then the third, what we're seeing from all of our portfolio companies where founders was in many many different categories whether it was like finance or clinical reasoning or production engineering, support tickets, security, they were all sort of describing this missing layer but in different language. And when we named it, I think as you mentioned, people inside companies like OpenAI, Anthropic, Palantir, CEOs of public SaaS companies all told us they were kind of circling the same concept for months without being able to put a name for it. So I think those are kind of the three main things that changed.
And Arrvin, from your perspective, you know, you've built multiple generations of enterprise technology. What makes solving this context problem different?
I think the more than solving first the expectation of what people want to do with that context that is actually the most challenging part because now everybody's seen the reasoning, the intelligence of AI, and they expect AI to do all the work that we do today in the business. And it's a big expectation. And there's a lot of gaps still in terms of connecting that intelligence of the model with the context and the ways of working of your company. So when we think about building this context graph inside Glean, one of the key challenges is that we don't really have a way to scope a problem like, 'hey, this is the ultimate way this context is going to get used and therefore let's prioritize and let's build the context in a particular order.' That's actually one of the most challenging things about how do you fundamentally absorb everything about your company, how people work, their ways of working, because there's so much information, so much data, as you said, every conversation now is also getting digitized. So taking this really broad set of information in different formats, different styles, how do you combine all of that with the power of the language models and then create a usable artifact from it, a usable artifact that actually does allow people to go and create those agents. So it's a challenging problem. It's also fun though at the same time. So everything that we're doing here feels like things that we haven't built before, brand new exciting things.
Yeah. I think one of the challenges you're getting at Urban 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. And it's really figuring out what are those key use cases that the enterprise needs in order for AI to really feel like it can take on more of this work. I mean, it seems to be broad agreement both between you and also what we're seeing in the space that context is really key for agent performance. But I'd love to hear a little bit more about how this problem manifests specifically for enterprise customers and in even your portfolio companies. So, GIA, from your perspective, where are your portfolio companies seeing agents stall? Yeah, it's a good question. I think some of it is missing data and then I think some of it is also a missing understanding of how decisions are actually made. I think after the article, I think majority of Fortune 500 companies either their CIOs or CFOs reached out. And after talking to them, I was like, 'Okay, your problem is definitely data. I think systems are definitely fragmented. I think they're complaining about migrations being incomplete and that the infrastructure was not really ready for agents to operate cleanly across it.' But I think what's interesting especially after this article was that you talk to some companies and even where the data problem is partially solved, the agents, they would say, are still kind of stalled. And so I think this was where many CIOs resonated with the idea that we actually don't have an understanding of how decisions actually get made inside the organization. And I think the demand was coming from these people like, 'hey, we really want to be able to track why decisions were made.' And I think it's interesting because we have a few portfolio companies that are in and around this area at the application layer. So we have a company called Player Zero where they're doing this in production engineering, effectively building an engineering world model and saying, 'hey, when something breaks, the system should have an understanding of why it changed, what changed, and what might break next.' And so I think it'll be interesting to see what exactly emerges, but we're seeing application companies build the context graph as maybe a feature, and as well as platform companies starting to emerge, and incumbents from all over saying, 'hey, we want to build this as a platform.' So I think it'll be interesting to see how this plays out. And I think you get a really important point which is there's data 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, there has to really be this mirroring between the two in order for this context to be helpful and actionable.
Well, first, when you think about pieces of work that are happening, for example, take healthcare and take the routine high volume tasks which you would think that you could automate fairly easily, like take claims processing. The problem is that any given claim that is submitted, there are so many unique elements of it, which doesn't fit any playbook. You can create a playbook, but still, you are always thrown off because every new claim has something new which could not be covered in those rules that you have. And that's why there are those folks who are actually then making those human judgments. So these kinds of things will continue to remain a challenge for systems like that. And when you think about AI, AI is actually probably taking us an order of magnitude higher in terms of how much work it can automate. Before, like claims processing, maybe you could take 80% of the claims and automate the approval or denial of them based on the playbook that you have. Now AI gives you a little bit more subjective analysis and you can get from 80 to maybe 90%. But I think we will still continue to see ultimately humans who have to make those final calls on many of these things. And I see this across the board. You get MSAs, you sign new customers, you're approving a new contract. Finally, what I noticed is that we have the policy, we have the procedure, and yet the person from legal will come and ask me that, 'Hey business, you have to make a decision. Here are the pros and cons. If I approve this, this is what's going to happen. If I don't approve this, this is what's going to happen.' And there is no past example which matches exactly this new contract, this new customer. And so then it becomes a problem not so much of 'did I have the right data, the right context?' It's a problem of ultimately somebody has to make a call, somebody has to make a business decision which is brand new, and it happens more often than you would realize. So when you talk about what remains outside of the scope of AI, as such, nothing. 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 probably still in the final set of decisions that we still need humans for.
Yeah, agreed. And it's interesting too. I think with context and what 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 weigh all of these decisions from different stakeholders to make that type of judgment call. And hopefully you can get examples of what this looked like before, 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 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 uplevel what you were given before in a way that is unique to your organization. And I definitely agree. I think humans are going to be a really big part of agents. There's going to be a lot of great interactions between them moving forward.
Um so J, you touched on this a little bit earlier, but I wanted to talk a little bit about the role of orchestration and the role of data. So there's another debate that's kind of coming into play here around context graphs about where in the AI enterprise stack does this fall? Does it sit in the orchestration layer where agents are learning from the traces of their past runs? Does it sit in the data layer where you look kind of beyond those individual agent runs to the systems and the human interactions to understand what actually happened? Or is it some combination of both? So you've put out 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.
Uh I will actually say it's both. I think my opinion has now evolved after 25 more conversations in the past few weeks. So I think it's both. I think one of the things that's interesting is where is state accumulating? And so I think you could kind of see that in workflow automation companies like Zapier, Workato, Tray.io, or something like that, which are all super useful. But what they mainly accumulated was just execution logic and not really building a memory of why did that workflow exist, whether it led to a good business outcome, what exception came up last time, or kind of how some team adapted the process over time. And so I think something similar is at the enterprise layer too, like ServiceNow and Moveworks own the ticketing flow, but neither one really owns that full decision trace. But I wouldn't go all the way to say orchestration doesn't matter. I think it's really important, but I think as you said, the power will sit where the execution as well as the learning are joined together. And I think the distinction is also collapsing. And this is kind of why you see so many companies that are now saying that they're competing or doing the same thing. And also why I think when this article got posted, it was a huge mix of data catalog companies, infrastructure companies as well as application companies that were kind of saying that we're all doing this. And I think everyone is converging towards a very similar architecture. But it will be interesting to see how it plays out and where the power was. I think it'll be both though.
