AI is not capable of replacing you right now. Let's be clear, it's not ready to replace even a single human today. I actually don't believe that companies will be successful in reducing their workforce significantly. My high school kids, they can tell you very quickly what text is written by AI and which one is not. It's the pace of innovation that is actually creating that feeling of anxiety. You never feel this feeling of hey, you have arrived or okay, a company has been built, we have a stable business is completely elusive.
Right now for us, there is a war going on right now over who owns the AI assistant for knowledge workers. The one tool that sits across everything you do at work and actually knows your company. OpenAI and Anthropic want that seat. So do Microsoft, Google, Notion, and 100 startups. One company got there first: Glean. Arvind Jain founded it in 2019 after a decade building search at Google and co-founding Rubrik. Today, Glean is over $300 million in ARR and valued at $7.2 billion. In 2024, OpenAI reportedly told its own investors not to fund five competitors. Glean was on that list. Our conversation today with Arvind is interesting for two reasons. One, he holds that pole position against the most powerful companies in tech. And two, he's running one of the first and most AI-native companies in the world. This is going to be a power-packed conversation for everyone in AI.
Arvind, welcome to the show.
You've put yourself in this pole position which every company in the world wants to win: the AI assistant for knowledge workers. The big labs want it, Google wants it, Notion wants it, tens of startups are chasing this. But as all of them now make foray into it and especially the big labs, let's start with them. As they get to know every knowledge worker and what's happening in every organization, what is Glean's product edge in this world?
First of all, we started Glean in early 2019. That makes us the world's first enterprise generative AI company, and we've been focused on the knowledge worker since then, trying to understand how people work, what are their information needs, knowledge needs. It's very interesting that in the last seven and a half years, as we've sort of brought our product to thousands of enterprises, first of all, the sheer number of challenges that you have to actually solve which have mostly nothing to do with your actual product that you're trying to build. Like we're building an AI product, of course, right? You come into Glean, you ask questions, and we'll bring the right answers and the right information back to you. That's how we got started. Today we are a coworker. You come to Glean, you can actually go beyond asking questions. You can give us some work to do, and we'll do that work on behalf of you behind the scenes. Of course, we're connected with all of your enterprise systems. We understand how your business works. So we do that.
But the challenge is, like, to actually get this product up and running inside an enterprise, a lot of that has nothing to do with AI. It has got to do with how do you actually handle data safely and securely? How do you understand what is the governance architecture of an enterprise? What information can be seen by what users? What are sort of policy and compliance requirements? If you're in a regulated industry, what are the things that you can do with AI and you can't? So I think one thing that we always educate our enterprise customers is that this is a complex undertaking. Bringing a really good coworker to your employees while meeting all of your enterprise requirements requires effort and energy and that experience that we have learned over the years. So the main edge for us is that experience, the fact that we've been in this business for a long time. But then maybe I will also share a little bit more about where we spend our R&D resources.
So the company now is about 1,500 employees. That makes us one of the largest pure-play enterprise companies in the world, but we're actually not spending a lot of that time trying to train and build models. Glean is a multi-model product. We actually partner with the best model companies, whether that's OpenAI or Anthropic or Google. We actually also incorporate open source models built on Nvidia technology or another open source model out there. It's really great to be in that position where we can actually leverage that innovation that the labs are bringing, and then we get to actually build things that they're not focused on yet, so that together, with their technology plus our technology, we can bring the best experience to our users. So that's part of our core strategy.
So it's not just connectors and permissions. You're saying it's much more than just building connectors with CRM, Workday, every other system, and permissioning.
Yeah, it goes well beyond that. That's the basics. That's the nuts and bolts, like you connect with all the enterprise systems. Even that is actually, by the way, quite hard. And you should understand that today the industry has taken the direction of not wanting to build those integrations. We build it, but if you look at the strategy that the labs have taken, they are actually choosing this MCP-based architecture where they don't build integrations anymore. They're expecting that each individual enterprise application will put up an MCP server which allows an AI model to interact with those systems a little bit. But that's actually quite shallow, and in that approach you don't get to actually learn how your business works.
Ultimately, our viewpoint is this: if you're going to have AI do a complex piece of work that a human today does in your enterprise, then that AI has to actually learn from that human how they've been doing this work. Some of it is written down in SOPs and documentation, but some of it is in their mind and they exercise that judgment every day. So to be able to actually bring that automation in your business, you have to actually go much deeper—much deeper than connecting with the systems and understanding what data and what knowledge is where. You have to actually understand how work actually happens. And that requires you to actually even understand what work is, like what are the tasks, what are the key jobs to be done in this business. You have to observe who are the people who actually do that work, who are experts on it. You have to observe their activity to see these are the 10 steps that this person took to complete this particular task, and these are all the observations that we build. This is the context graph. And so Glean has been investing in that for many, many years. And that right now we still see as an area which is completely open, like we're not seeing other people actually building that yet.
So let me steelman the other side of the argument just for debate. One is with MCP, now the onus is on every enterprise company to build their own integrators into the model, right? So that makes life slightly easier. The big labs also, including Google, by the way, they all have endless resources. In Google's case, they also have Google Docs and the Google Work ecosystem, and this is one of the most important races, so they are going to put all the resources with forward deployed teams and everything. So everything you're talking about, from connections to permissions to getting to know how the knowledge work actually works with forward deployed teams, wouldn't they want to catch up to this in a pretty aggressive way? And let's say they at least solve through MCP and the connectors and integrations thing and memory, which then adds up to how this recruiter does their job and stuff. Is the edge still enterprise features? Is it the regulatory aspect you talked about? Is it the multi-modal nature of it? What is the cutting edge? Because my guess is there's something that even you built that is getting commoditized for you in front of your eyes, but then you are continuing to create more edge.
No, that's a very good question. In fact, you're absolutely right that technology today changes so fast that you have to keep reinventing, of course. But coming back to this question, we do believe that we have a structural advantage over the labs model companies. Talking to over a thousand enterprises over the last one year, some of the key trends that we have actually heard, like the top concerns that enterprises today have, are number one, there are all these different models out there, so many different AI products, and they're having a difficult time trying to figure out who they should work with, who they should bet on. Like one day you have GPT being the best model, the other day Gemini takes it over, then Claude takes it over, and then GPT comes back. And it's very hard to make that decision on who to work with.
Second is cost. And this has become a real issue now. For the last two or three years, we were in this pure "go and invest and we'll figure out how to get value from it," but that investment became too big. Companies have allocated token spend management and stuff, like they create these annual budgets and they've blown past that in one month or two months. So cost has become a real concern because business value is still elusive.
And then third, there's this sprawl problem, which is that the CEOs and executives pushed AI so hard inside the enterprise. In fact, there were things like token maxing—everybody is being encouraged to do more and more work with AI. And that has also resulted in now you are ending up with a lot of mess on your hands. You have a lot of AI slop, like a lot of AI-written documentation that nobody has time to read. People are building skills and agents, and there's a lot of duplication, and the sprawl has become a big issue. So these are the top three problems that businesses have.
So now when you think about models when they can't make a decision, and when you think about cost, open source now has to become a big part of their solution stack. So who do they work with? Working with a lab, they're going to actually direct them to their models. They could probably be too expensive, and you're not getting all the innovation. In fact, this is very interesting that a lot of this divergent set of technology that is being made available even by the labs—for example, Claude today, to the best of my knowledge, doesn't have great image generation models. For that, we have to use OpenAI models. So Glean can actually use all of these models. We can go to our customers, we can say that we bring all the innovation from all the labs because, A, it's available to us, we're their customers. We can also bring the open source model so that when a task doesn't require the top-end frontier model, we can actually use an open source model that is equally good at that task but costs an order of magnitude less. So we can bring that. So that actually is a story as being an independent AI company which doesn't have vested interest towards one particular model. It actually puts us in a very advantageous position because this is the core question that enterprises are trying to actually—they want to have that independence from a specific model.
What about then companies like Notion and some others who are platforms and maybe Granola is coming up in a different audience that are also collecting across the board organization? They're also independent. Notion has made a big foray. Actually, Notion is one of the few companies like you guys which is a pre-AI company that has completely revamped their platform in a good way, it seems to me from a distance. What do you think of them, or is that a different tier, primarily enterprise, they might be more SMB/mid-market?
That is true, but also one thing that's very interesting is that when you think about AI inside your enterprise, let's say you're building these agents for your go-to-market team: an agent that actually helps you prospect, an agent that helps you run effective meetings, an agent that nurtures opportunities better for you. All of these agents, you have your core intellectual property in that you've actually designed them. You've actually brought the intelligence, and this intelligence compounds over time as those agents get better, and you want to make sure that these agents keep working even if you change the underlying CRM. You want that layer, the AI agentic layer, to be independent of the system of records underneath. You don't want to be locked down as a better CRM comes about or a better document creation system comes about. You want to have that flexibility without losing all the intelligence that you've compounded.
So that's another technical, architectural sort of decision that businesses are making, that they want their AI systems to be independent of the system of records. So that sort of again is good for us. Glean, as not being a system of record, if you build agents on our platform, they will work whether you use CRM 1 or CRM 2.
What does the enterprise stack of 2027, 2028 look like in your view? What do Google and Microsoft have? What do Salesforce, Workday have? Do these systems of record continue and you are on top of them? Because there's also a thesis in the world that agents might just completely take away interface-based SaaS applications.
The truth is both of those things make sense to some degree. I personally believe that most of these systems of records will remain in the next few years. The Google productivity suite, Microsoft productivity suite, Salesforce, and many of your HR systems like Workday, ITSM systems like ServiceNow—each of them will be just fine. They will all be there. And in fact, they will all get very agentic in nature. So if they don't, then of course they will not stay there. In fact, those are the products that are going to go away which are not advancing with AI.
I think businesses will actually build more systems internally, and so the number of applications may end up reducing. Because if there was a very point use case, like one specific thing that a SaaS app did for you where that SaaS app is really just a UI wrapper on top of your system of record data in Salesforce, let's say, that kind of app, people are going to start to build those kind of systems on their own because it's very easy. In fact, I'm doing it myself. We had this asset management app and it was very, very hard to actually use. Say it's an Excel spreadsheet that you're trying to view in a few different ways, and that system didn't allow you to do that. It didn't allow me to sort, it didn't allow me to filter the way it was not convenient, and I couldn't actually ask a quick question off of that. So I just actually went in Glean and asked it to code up that application and just connect to the backend database, and it actually took me like five minutes. And even the first version of that app was actually pretty good. Behind the scenes, I don't know what model it used, probably must have used Claude to actually compute that app. But that's an example of where users are feeling more confident of building some systems themselves. So we will see that. We will see a lot of internally built applications.
How do you see the Glean product today? Is it an app-layer product in your mind? Is it a completely agentic offering, or is it actually the infra that enables all kinds of AI to be used, whether your own agents or somebody else's agents on top of it? What is your current view of Glean's platform?
So we are fundamentally the context and intelligence layer for your enterprise. It's a platform. You bring it, you connect Glean with all of your enterprise systems, and now we understand all the data, all the knowledge, where it sits. We connect with structured data systems like Databricks and Snowflake, we connect with unstructured systems like your documents, and once you've connected Glean with all of your enterprise systems, we now become that gateway. Think of us as an AI gateway, and it has a few different capabilities.
One of the capabilities is that we can actually retrieve any information from any system given any task, any question that is given to us. Another one is that we can actually help you take an action inside those enterprise systems with the right security. Yet another one is that you can actually ask us, "Hey, how is this task being done today in the enterprise?" and we can actually bring the right sort of—we mine how work gets done.
You can do that? Like you can ask Glean saying, "Hey, how does this person do again using this particular task?"
Yeah, we will actually go back and share that because behind the scenes we're building task-based memory, we're building personal individual memory. Like for any given person, we deeply understand who they are, what role do they play in the enterprise, how do they like to read and write, what are their preferences, what are the things that they're expert on, and what are the things they need done this week. So we know all of that. And this memory infrastructure is also another piece of infrastructure that we build as part of this platform.
And so now this platform is available for all of your AI applications inside your company to use. You may be building some agents. Let's say you're building those agents in Vertex in Google or Bedrock in AWS, and so you'll connect it with that Glean MCP gateway, which brings all of that enterprise context and intelligence into those agents. Now, these agent builder platforms don't have to go and figure out how to connect with all enterprise systems. As an agent builder, all of that connectivity is pre-established for you, so you just focus on what task you're trying to actually go and build.
Similarly, we can actually serve as a backend inside your Claude Cowork or ChatGPT, and we just make those tools better, make them more company-context-aware, smarter, you get better answers, and we actually also make them faster and cheaper. Because interestingly, what happens with AI today, Claude is actually super smart. You can go and give it an arbitrary task, and ultimately it'll figure out how to solve that, but it does it in a brute force manner. It will spend lots and lots of cycles first assembling just even the raw materials that it needs, the basic context to actually finish this task, and then it starts to apply the model intelligence on it. So that first part of that, we can actually short circuit. Claude can just instantly get all the context through Glean. Our customers have reported, we have actually done studies, that it actually reduces token consumption by 30 to 50%. So that's a big use case. In fact, a lot of companies are actually right now investing in Glean for that reason, because cost is a real issue for them.
So that actually makes us the platform, right? It's a platform that makes all your other AI agents and your AI products more capable, better, faster, cheaper. But then we do come with our own first-party products. We have our own agent builder. We have our own Glean coworker, which we think is more powerful, and in fact you can think of it as a superset of ChatGPT, Claude, Gemini, all combined into one. So that's our strategy: build that platform, build it in an enterprise-grade manner with the right security, safety, governance capabilities, but then also keep investing in our end-user applications.
So you will continue to encourage other agents to be built on top of Glean?
100%. In fact, our team is divided into two. There's the core platform team which delivers this context and intelligence platform to the enterprise, and then there is the Glean Coworker team which is building the first-party product experience on top of this platform. But they get no extra rights to use this platform than what Claude and ChatGPT get. In fact, that's the working model. The platform team doesn't even test Glean coworkers; they actually only work with Claude and ChatGPT as the primary test platforms.
How much of the use of Glean today is enterprise or people asking a question versus pure agents, and how is that ratio maybe changing?
Getting answers to questions, seeking knowledge, it's the top use case for Glean, but it's also the top use case for AI in the enterprise in general. So even in ChatGPT or Claude, that's the number one use case. Coworking and doing tasks is emerging as the new category which is growing at a much faster pace, and so we see now at least a third of our users going into very deep and complex tasks where they're no longer actually just asking questions, they're actually fundamentally doing a lot of their work with it. All engineers, of course, nobody's writing code from first principles anymore. AI is doing the first version of code writing. In general, coworking has fundamentally increased. For us, we're not the platform that people use to write code as much, although they can—again, it's the same model, so you can actually do everything that you can do with Claude Code, but that's not the use case that people actually use Glean for more. But think about a support engineer as an example. As a support person, all your cases, they're already using Glean as the first person to actually go and understand that ticket from the customer and resolve that case, and in fact most of the times it just does it all by itself and an agent doesn't even have to look at the ticket. So their coworking has now fundamentally increased.
Fast forward two, three years: what does Glean look like in that world? And maybe an extension of that question is, what does a knowledge worker's life look like in two, three years?
We want to remain the leading enterprise AI company in the world. Our mission always has been to help people do extraordinary work and expand their human potential. The way we think about our journey, next two to three years we want to keep extending this context and intelligence platform that we're building. We believe that we have an unbounded problem to solve. So we allow people to search and find things, we allow them to actually do work, but we have to make all of these products 10 times better, 100 times better than what they are today. So we'll continue this journey of investing in this platform, and we do expect us to become one of the de facto standards in the enterprise where a lot of other AI products use us as this underlying platform to build core AI capabilities within their own product experiences.
In terms of knowledge workers and how our lives are going to change, right now people are already doing that, by the way. In most of our customers, we have some of the most tech-forward companies as our customers today, and you would expect that everybody in those companies are basically becoming AI-first and using AI in a maximal possible way, but it's not happening. We see this power law where 5% of people are actually doing a lot with AI—they've completely changed their workday, in fact they've become largely a reviewer of the work that AI does for them—but the remaining 95% are still using AI, but they're using it for that use case that you first mentioned, which is knowledge, question answering, like "I have a question, give me an answer," or "I need some information, give me that information."
So first we have to actually fundamentally change that about AI. The reason why it is only 5% of the people doing so much and 95% not is because people don't have time, and not everybody has the same level of curiosity. AI today is very reactive. It only helps you when you actually ask for help. And so that's one thing that we're changing with Glean: we want to be that proactive work companion for every knowledge worker. Before you even think about, "Hey, should I do this task with AI?" AI has already figured out that you had to do these 10 things and has come to you and said, "Hey, by the way, I can do these five things for you," or "I've already done it for you, and here's my work." And so that is sort of going to fundamentally—
First of all, making AI a truly broadly adopted technology happens when it becomes proactive. So that's our vision, like to be that best proactive coworker, and that's what we think is going to happen to knowledge workers. Each of us will have this personal companion that knows everything about us in our work life, and it proactively helps us, and the burden is no longer on me to learn how to use AI. So that's the world that we're going to be moving towards in the future. And you can imagine that a lot of things that feel like drudgery to us, tasks which we're not excited about, our companion, our personal team of agents is going to do that work for us.
Yeah. Do you feel like almost a revolution is coming where companies quickly identify that 80% of the work is done by 5% of people because they've supercharged themselves?
I actually don't believe that companies will be successful in reducing their workforce significantly while still having the same level of market success or share of the market. Because if you think about it, take an industry, and if you are one of the companies in that and there are four competitors, all five of you have the same tools. All of you have all the great AI capabilities. If you can actually do the work that needs to get done with half as many employees, so could they. In reality, what's going to happen is that your customer expectations are going to double. You have to sort of build better products, stronger products, and you're still going to need that same staffing to do that. In fact, ultimately business success is indeed proportional to the human capital that you have inside the company. So that's my belief.
I would take a different stance where I do believe that this power law will change. Take a calculator: does everyone know how to use it today equally well? And I think that's what's going to happen with AI. So it won't be that 5% who have figured it out and 95% haven't, because AI itself will be that great teacher.
The thought that is expressed—I think Elon Musk has expressed this, Scott Wu recently expressed this in some interview—is that we are going as a human race from survival mode, where we were all doing work, to creativity mode, where it's an age of abundance because agents do all the work. It sounds like you're saying, correct me if I'm wrong, that that's not very imminent. It might be some futuristic vision, but in the near future, everybody will be just more productive, more work will be expected, and competition will take care that everybody will still need their people.
I think so. It is true that we are moving into a world of abundance—there will be enough food for everybody to feed and all of that. But as a business, the competition is going to increase. I really don't see this world where somebody can build a very flourishing, great business where their people only work two days a week and the competition actually works five days a week. I don't actually know how the workday comes down from five to four to three and a half that everybody's talking about, because companies are going to immediately see that as a competitive edge. I would want our team to be fully motivated and energetic and actually keep working the schedule that they are today.
But does the human work evolve from doing to reviewing mostly?
Human work completely changes across different functions. A lot of knowledge work is going to be done by AI for you. You don't need to, for example, do information synthesis; you can actually get AI to do all of that work. So I think our work will shift more to asking the right questions, actually thinking about what are the things we should be working on. And then AI acts as a very powerful companion to us. Think of it as, instead of being an individual contributor, I have 10 really, really good people available to me, and now we can actually build a product together with that team.
Let's shift towards how Glean is run as an AI-native company. So let's start with you actually. You were a co-founder of Rubrik and then you were at Google, so you've been an exec for a long enough time. How has your day changed as a CEO with AI? What are decisions that you are now actually fully giving to an AI system versus still keeping? How are you using AI in your daily life today?
Well, our primary tool to use is Glean because it's meant for this work, but I also use ChatGPT quite a bit partly, and I always compare these tools and see how we make Glean better as a result. The things that I do with AI: number one, I'm a very curious person, I have questions all the time.
I mean, that's why you started Glean as an answering machine in the first place.
