Hey, you amazing listener. Before we dive into this week's episode, I have a quick note to share. Since you're here because you're interested in hearing from leaders at the forefront of AI innovation, I thought you'd want to know about Canva's upcoming AI Vision event in Sydney this August. We're bringing together some amazing speakers and I'd love for you to be there. Scan the QR code on screen or search for AI Vision Sydney to find out more.
That's a real problem right now for businesses — how expensive AI has gotten. They had allocated a budget for AI for the year which they actually already consumed not even one month in, and that whole budget is already exhausted because these tools are very, very hungry.
Welcome to Prompted, a podcast about AI, people, and the creative spark. I'm Cameron Adams, co-founder and chief product officer of Canva, and I'm on a mission to talk to AI leaders about how they're practically using AI to be more creative and amplify their teams. Today, we are joined by Arvind Jain. Arvind is the founder and CEO of Glean, a work AI platform. Prior to Glean, Arvind co-founded and led R&D at Rubrik, one of the fastest growing companies in cloud data management, and he spent over a decade at Google as a distinguished engineer where he led teams in Google Search, Maps, and YouTube. Arvind, a huge welcome to the show.
Thank you so much for having me. Really excited to be here.
It is a pleasure to have you. Now, before we dive into the deep Q&A, we love to kick off the show with the prompt. With the prompt, we get you to sketch something from your past. Could be a childhood memory, could be something that shaped your career, and you need to tell us the story of that moment. Are you ready to pick up a pen and tell us that story?
Yeah, this is 1985. I think I was about 10 years old and my brother, my elder brother, had gotten a PC at home.
It was, I think, a 286 if I remember right. And it had the larger floppy drives, the 3.5-inch floppy drives. Actually, the whole concept was not really appealing to me, but a few days later he actually got a few games and one of them was Pac-Man. He showed that we could play these games — before that we didn't really have video games. I grew up in India, so this is actually even more unusual — most Indians wouldn't be able to tell that story. I think my brother was a pioneer in terms of understanding that there's something called a computer and it could be useful at home. He brought it there and we saw these games. He showed me how to play these video games on the computer. Pac-Man was my favorite, and that's what actually got me hooked into computers. I felt like, well, if I could actually also make games like that, because he also showed me that those are things you don't just buy — you could actually go and build games on your own.
Well, I would love to see what you have sketched. So, hold it up and let us see it.
That's kind of embarrassing, but I was just trying to show a picture of that old computer with a small tiny monitor and a huge box below it, and then a keyboard, and I'm trying to show myself sitting on a chair. But I think it exposed me and my drawing skills to you. This is...
That's quite a work of art. I'm pretty impressed with that. So, I think that gives something for Matt, our designer, to really work with. He's going to take it away. He's going to take your sketch, use a bit of AI to bring it to life, and we will see what the results are at the end of the show.
For now, let's dive deep on AI with you. So, you spent years building search at Google, but at some point, your focus shifted from helping people navigate the internet to actually helping companies understand themselves. I'd love to know — was there a particular moment that inspired that transition for you and led to the insight that looking inside companies could be just as valuable as looking outside of them?
So actually I left Google after about a decade of working there to start this company called Rubrik, which is a cloud data security and protection company. We grew fast as a company and in five years we were actually over 2,000 people. That's fast growth for any startup. That last year that I was there, we grew about 3x and it felt like everything was falling apart. We had a lot more people in the company but we were actually getting less work done. People were struggling. We were not able to release our software on time. Our business people were not selling the product as much as they were able to a year before. So we were trying to understand what happened — we saw a huge market opportunity which is what prompted us to actually 3x our team, and the returns didn't come. When we asked people what was going wrong, the biggest complaint from them was that they couldn't find anything in the company. Whatever information they needed to do their job, they didn't have it. They also couldn't find the right people who could help them figure out how to do these jobs. So this enablement problem was a huge issue for us.
And it was not surprising to me. We were built in this new SaaS era. We had about 300 different systems internally. Our company's knowledge, data, information was fragmented across all of those and even I couldn't find anything inside. And it's a problem that was not unique to us at Rubrik. Even at Google that was always the running joke — we were in the business of helping everybody in the world quickly find information and get answers to their questions, but we never really helped ourselves inside Google. It was incredibly hard to find a design document that you're looking for at Google.
I always remember those yellow boxes that Google tried to sell at one stage where it was like plug and play into your enterprise and it would help you find anything. But I'm not sure how that turned out.
It never worked because it was so hard to actually even put stuff inside of that box. So it's a long-standing problem — enterprise search was never solved. But I felt in the moment, this is 2018, I felt that now the world had changed. SaaS systems became a lot more interoperable and you could actually build a really amazing search product. The SaaS transformation combined with transformers and language models felt like two big developments that would make it possible to build a great search product. And it's super important because a third of all employee time is just wasted time to find information at work.
Were language models around in 2018 that you could utilize?
Yeah. So that was actually the time when Google had actually put BERT in open domain. It was a model that was trained — I mean it's a small language model now but at the time it was phenomenal. It was trained on all of the world's data and corpus and you could take that model and you could pre-train that at each one of our enterprise customers. You could build this language model that could semantically and conceptually understand information inside a company. And we felt that was actually a big unlock to making search more natural for people, where you can just come in and you don't have to remember exact words to find information. You could sort of freely express what you're looking for and we could use these embeddings and vector search technologies — these terms didn't exist at the time but the models were there. So we actually built the industry's first vector search solution combined with traditional search techniques to build this really effective product where people can go and ask questions and we do the hard work of finding the right information from across all these hundreds of different enterprise systems that you have.
Was it hard to leave a company that you had started and was so deeply ingrained in?
Yeah, it was actually. Initially, when I was thinking about this idea, it was not with the mindset that we're going to leave Rubrik and start Glean as a separate company. In fact, I actually thought there was a lot of technical synergy between what we're already building at Rubrik, because Rubrik was in the data protection business. So we did have to actually do the work of connecting with enterprise systems, bringing all of that information in one place so that we can just save it for safeguarding that data. So I felt like half the work was already done. I just had to now build a search index on top of it. So initially I thought I could build this product inside our company, but then it didn't make sense as you sort of go deeper. So it was a difficult thing for me, but part of me, the way I thought about it was — I never sort of felt like I was leaving, because I was actually working on this specific project precisely to help the company itself. So I kind of felt like I was part of the team even when I had actually now started working separately on this project.
When you were looking for people to help you start Glean, did you look to your partners in Rubrik, or are there other skills that you needed to bring together to make it happen?
Well, this is a very different product from the infrastructure product that we had built at Rubrik. So it required different type of people, but luckily of course I knew many of them having spent over a decade of my engineering career building search and AI systems. I had a lot of colleagues from Google who had all felt the same pain. I mean, I'm sure Cameron, you probably also have struggled in your career trying to find information at work. So the problem actually appeals to every engineer. I initially left Google to start Rubrik with these amazing folks who I wanted to work with, but I couldn't convince them to work on a data protection product. But as soon as I told them that we're going to build a great search product for people in their work lives, they were really excited. So we built that original team — a lot of the folks actually came from Google, like my ex-colleagues.
You said that working on things like Google Maps taught you how messy real-world data really is. I imagine that ended up being really good preparation for what you do now.
Absolutely. Enterprise data environments are really, really complex. Think about a business that's been in existence for the last 100 years or 50 years. There's so much information, so much knowledge that's been written over this long period of time. Much of that information now is obsolete, but as such it's not marked like that. And so when you think about connecting with all the information inside the company and still helping people, you better not share with them obsolete information. So you have to build all of that disambiguation inside the company. The internet is the same way — there is of course no controls in terms of who gets to produce what information on the internet. So the way Google was built always was with that same mindset that you have to deal with varying levels of quality of information and you have to extract the gems from the noise. Same concepts apply in enterprise, just like you have to use different techniques to solve the same problem.
When we're talking enterprise data, what is it? Is it cloud databases? Is it spreadsheets stored on someone's laptop? I imagine it's a wide array of things.
It's both of them, but also a lot of data that's in your email systems, communication systems. For engineers, documents that you write in Confluence, tickets that you store in Jira. These days also meeting recordings — a lot of context of how your business works is people talking to each other, which luckily now we're digitizing more and more of the enterprise. So really when you think about Glean, the way to think about it is that we are actually connecting to all the data that's inside your enterprise. Structured data, unstructured data, conversations, actual formal documents, tickets, customer interactions, all of it. Because ultimately, what we are building for you is one place where you can go and ask any question and we're going to actually surface the right answer back. And that requires us to be comprehensive.
You mentioned working with a company that might be a hundred years old and has a massive trove of information. And yes, some of that information is going to be obsolete because it was about the 1960s, not about the environment that we're operating in now. How do you figure out when data is obsolete and what is relevant to what someone's searching for right now?
There's a lot of knowledge which is not high quality — somebody maybe wrote a document for their personal use, it doesn't have all the great details. You have to sort of understand the quality of information, and of course being recent is one of the attributes, but how high quality it is. Is it being written by somebody who's a subject matter authority on it? And is it actually also the right information for what I'm actually asking you right now. For example, if I type in a search query which is 'onboarding guide' and I ask that question in Glean, well now the right onboarding guide is the one that's relevant for me based on my role, what I do — don't show me the guide which is meant for the marketing team if I'm an engineer. So when you start to think about people's questions and how to answer them correctly, it's not just about the knowledge itself. It's also about the person and understanding them and their role and what information actually matters to them. What's the right one that actually is going to help them in the current situation that they are in.
So what we do behind the scenes is, after we connect with all of your enterprise systems, we're building a knowledge graph of your enterprise. You're trying to understand your people, your other entities like your customers, partners, trying to understand the products that you build, different projects that are ongoing, the teams that work on those projects. And then we also look at all the knowledge, the data and the information and we draw connections between them. We try to understand that these particular documents are very meaningful for the marketing team, and how we figure that out is by looking at interactions. We look at who consumes what information, who spends more time reading certain pieces of information, about departments they are part of. So that's sort of how we start to build a true understanding of your enterprise, your data, and that powers ultimately this smart, context-aware, role-aware search system. But by the way, Glean has evolved far beyond being that search product that we started with. This foundation that we built has allowed us to actually do a lot more. And today you can think of us as both a Google or a ChatGPT-like search and conversational assistant, a personal co-worker that you can bring to every employee in your enterprise. You can also use us as this enterprise-wide RAG platform or an MCP gateway to all enterprise context that you can use and agents that you're trying to build inside your enterprise. So as AI has gotten better, so has the capabilities of our product.
What's the difference here between using something connected into ChatGPT or Claude? Because I can connect my Gmail in there. I could have my G Drive or my calendar. It can go out and talk to those things and bring me back answers. I presume that's in real time though, like it's constantly going out and fetching new data and bringing it in every time I prompt. It feels like Glean is a slightly different approach where you're constantly crawling, creating those connections and then using those every time someone asks. Is that an accurate kind of representation of it?
That's the right explanation for how Claude or ChatGPT works versus how Glean works. For some tasks, it's actually easy for you to actually go and fetch information live from an enterprise system and then have the LLM work on that data. But often times what we see is that these AI tools are very brute force in nature — you give them a task and they start to sort of go on that hunt to find the right information, the right context that they need to solve that task. They will make live connections into Google Drive or calendar, get some information, and remember that search has never been a solved problem. Each of these individual systems don't have great search APIs — you ask them questions and they're not always precise in terms of delivering the right information back. So the burden now is on these models — with Claude and ChatGPT, they have to sort of do this retrieval pass once, connect with all these systems live, and then they come back and see what results came back, and usually they're not satisfied. Then they will make another round and you sort of go into this loop of doing lots and lots of retrieval passes ultimately to get to a place where you're happy with the context that you've gathered, and then you do the work on it and help the user. But that ends up taking too much, a very long time. It also actually consumes and burns a lot of tokens, and that's a real problem right now for businesses — how expensive AI has gotten.
Versus if you contrast that with Glean, we've done the hard work behind the scenes already. On any given topic we have already aggregated and organized the right information on that topic so that when AI wants to now work on a given task, we can deliver the right context in one shot to AI so that it can work faster and be more accurate.
How many conversations are you having now with enterprises about token usage, about cost? Like is it a leading reason for people to look at your service?
Look, AI moves so fast. So nobody was talking about token costs in a big way two months back, but in the last one month that's become the center of conversation. Every conversation that I'm having, the AI spend at businesses have actually ballooned up so much. Companies are telling us that they had allocated a budget for AI for the year which they actually already consumed — not even one month in and that whole budget is already exhausted because these tools are very, very hungry in terms of consuming tokens because of the brute force nature of it. So Glean is the product that competes with Gemini, Claude, or ChatGPT in some way because we have the same kind of interface, but we can also be just the layer underneath these tools. With our MCP gateway we can deliver that enterprise context that these tools need as a backend, as a tool that's available to them. And so we do these studies where we have Claude work directly with enterprise systems through individual MCP connections versus when we have Claude connect with Glean as the only system where it gets enterprise context from, and we see that actually reduces Claude's token usage by more than 50% and makes it more than twice as fast.
You said the conversation has changed even in the last two months, and the last year has felt like an absolute explosion in the world of AI. New models, new tools, everyone's talking about agents. From where you're sitting, does this feel like a very different moment from any moment in your career, or is it just kind of business as usual for a tech company?
I would say nothing about today's environment feels normal for engineers like me who have been in the industry for over three decades now. The pace of the change of course is incredible. The appetite for AI — I've never seen an environment where budgets are so big and you can really go and convince customers to actually buy a lot of technology today. The appetite for AI is unparalleled. There's never been any technology where enterprises feel as compelled as they do today that they have to bring these technologies in. And then the landscape — we are living in this very, very competitive market and it has changed how we operate as a business. We're no longer a category creation or an evangelical play. Now we have to sort of spend all of our energy trying to tell our customers why our platform is more superior to some of the other products out there.
You mentioned kind of a change in product to be more behind the scenes rather than the first thing that people actually type into. Does that make it a bit harder to get people to understand what you're offering when it's not maybe as visual or as much of a product that they're using?
The conversation that we're having is because now there are so many different options — you were already talking about Claude, Gemini, ChatGPT also having some connections. They're not as deep, they're not as broad as Glean, but they do have some. And we don't want to be in this conversation where somebody has to think about making a choice between us and one of those other three players. Rather, we are always thinking about how do we add customer value no matter at what level of the stack, in what position of the stack we offer that value at. So right now the conversation for us is definitely leading more towards that we become this enterprise-wide context platform, one platform that they can connect to Gemini and Claude and ChatGPT, because a lot of these enterprises have all three of those solutions. And so now we just make each one of them so much better, more precise, and cheaper by being that secure context layer underneath them.
I love to let people know that we actually use Glean here at Canva. So it powers our internal enterprise knowledge base which scans Slack and all our documents so that people inside Canva can easily find stuff. So it is an amazing tool to use. You mentioned budget being unlocked now at unseen levels before. What do you think's actually driving that? Where's the budget come from? Why are they so happy to spend money in this area?
It's very fascinating how businesses are actually making room for all this AI spend. And of course it's also a huge concern now because while they were making room, they still didn't anticipate how much more expensive it was going to be for them. The first thing that drives all of this is the fact that people know that the world is changing and we're going to be in a very, very different AI-powered world. And if you as a business, first of all, if you don't have a workforce that is AI-first and is not able to leverage all the core strong capabilities of AI technology, then you're going to be left behind. And so businesses are very eager to make sure that they're training their workforce. They're actually getting everybody on this wave and then make that best AI-powered workforce be a competitive edge in the future for them.
What's the best way that you've seen of someone training their workforce?
You would see two types of users. There will be occasional users who will use it to ask questions, and then there will be power users who are really using these AI tools to actually share their workload with and actually get a lot of work done with these AI tools. And unfortunately 90% of the users are in the first category, 10% are in the second one. And so you do have to do more than just giving these tools to people. One of the things that we focus on at Glean is recognizing this problem and this gap that is emerging between the AI enthusiasts who are learning more and more how to leverage AI in their day-to-day work versus those who are not able to. So we think that AI can change, and instead of being reactive where you have to go to AI and you have to be the one who asks what AI is going to do for you, and you need to sort of feel and believe that AI can do a complex task for you and most times you won't do it — so change that and actually have AI come to you and understand you and your work life and what are the 10 things you need to get done this week, and how about if AI did three or four of them for you proactively and just showed the work back to you. So that sort of removes that cognitive load from an individual. It removes that burden and that fear of using AI, and it'll basically start to drive that education at a much faster pace because now nobody had to think what AI can do for them. AI actually came to them and showed the art of the possible.
Oh, that's a really interesting balance. Within Glean, you can construct your own agents and your own workflows, right?
Yeah, I'm interested — as people start making their agents and creating these mini workflows that are powered by AI, do you end up with AI debt? So you've got thousands of people at your company all creating their own mini tools. Who's keeping track of whether those are working effectively as new data comes in or new workflows establish? Are they...
appropriate for what's happening? Who's like looking at all that and making sure that it stays up to date? So, with Glean, we allow our customers to actually choose different strategies for how agents get built and shared across the enterprise. You could either choose a model where you have only a few people in the company who are official agent builders and they are the ones who actually go and build these departmental agents which have a full software development life cycle behind them. You test them, you evaluate them, you sort of monitor their quality on a continuous basis and these are staffed projects, right? Where somebody manages these tools actively and these get shared with the company. Or there's another model where everybody can go and create agents but there's a concept of a library and in the library there's moderation. So where there are few people who can approve an agent to go into the library provided it has the right quality. So you as an individual can build as many agents that you want for yourself but once if you want to share it with somebody else it has to meet certain criteria. So there are all these different strategies, we're seeing different companies and enterprises taking different approaches. But one thing is today without a doubt an issue. Like at many of our customers, users have built over 30,000, 40,000 agents and most of them are actually more like somebody who's trying to experiment and see what AI can do for them. They're not really that useful. There's also a lot of duplication and ultimately we came back to this conclusion that just like knowledge you will have this proliferation of skills, you'll have proliferation of agents and you will have to actually solve that problem of how to expose the right agent or the right skill to a user. You have to solve that again through the search problem. You'll have to figure out what skills are being actively used, have positive feedback, same for agents. And that's the world that we're going to be in. I think that's the right world in our opinion where innovation is unconstrained but then there are mechanisms to let people find the right things for themselves.
How are you seeing skills developed? They've almost seemed to be quite organic, they just popped up overnight and people were talking about new skills that they could deploy through their AI assistant. Are you seeing a quite structured approach to skills or people are just like sharing random files with one another because they worked?
It's kind of the latter right now. You know, people go and build these skills and then they share it with people who they know. There's actually very little work that has been done in the industry right now in terms of how do you sort of do this work at a departmental or a company level in a more organized or top-down manner. Like how do you actually publish official skills for different tasks. And so that's been the focus for us, how do we change that mechanism where you're ending up with this proliferation of duplicate, overlapping, conflicting skills to where you can actually create one, you have the mechanisms where you have a little bit of moderation and review built into this process and for any given skill there's only one that is kind of blessed or official. But second, in Glean, we run a process called skill mining where instead of actually having people manually go and build these skills, we're actually looking at how work is actually getting done in the enterprise, who are the experts on different topics and how do they do a specific piece of work and you can actually in fact have AI itself give you the first version of these skills learning from the best people inside your company. And then you can actually make that same expert the owner of that skill and publish it company-wide.
Is that skill mining within Glean itself or you're skill mining for all enterprises?
Oh no, we skill mine for all. That's a product feature that we bring to our customers now and that skill mining happens not through interactions in Glean of course, like that's very limited. That skill mining happens by looking at people's work in all these other enterprise systems.
Kind of loops back to the two models that you talked about earlier. One's a centralized model where AI agents, maybe AI skills are kind of managed by a dedicated team versus the kind of democratic proliferation of everyone being able to do whatever they want with AI. Are you seeing one or either of those models be more successful?
I actually prefer the model which is kind of a hybrid of the both where you let everybody go and innovate. I mean right now this is also the moment where people need to get educated and trained. So it's actually a very good idea for everybody to have that ability to build skills and agents and in fact let them also share it freely with their colleagues and don't create those guardrails. But at the same time you have a top-down effort maybe department by department where somebody is actually 100% dedicated to building and developing some agents or skills within their department and that's an independent effort. And then naturally what would happen is you will see if you have the ability to sort of combine the two and actually provide a good search on top of that then people can find the right skill that they want to use. It may be one of the departmental level ones or it could be something that somebody had a great idea and they built it themselves sort of as a bottoms-up. What we're seeing is the most success is when you let both of those things happen.
