The Neo Show: Founder Fridays #5: Arvind Jain CEO, Glean with Host Howie Xu

Yeah. Well, first of all, thank you so much for inviting me, Howie, and I'm excited to be here to share more about Glean with all of you. And before I start, like a quick background on me. So, I'm the founder and CEO of Glean. Before that, I've been in the tech industry for about 30 years myself. We started a company called Rubrik before Glean in 2014. I used to work at Google before that, starting with when Google was a relatively small, you know, privately held startup. So I've worked on search for a very, very long time and mostly on Google search, but now on Glean. And yeah, so that's my background. I'm really excited to be here to talk a little bit of, you know, like how AI is going to change the world. So we all know Google, right? Google does the search for the last 20 plus years. And one thing that's very interesting is the enterprise search never worked. Actually, Google started a company trying to do enterprise search. I don't know how many of you knew that, and then they gave up for one reason or another, and then they did a consumer search. Some of you were born after that decision. But the reality is enterprise search never worked. Why is that? And then why did you start a company? Because you know there were companies after companies trying out doing enterprise search. Why did you start this company?

Yeah. Well, I mean, enterprise search is a very old problem that has existed, actually it predates consumer search by many, many decades as well. Because in businesses you have a lot of information, you have a lot of data and people work and they need that information to do their work. So making that information easily searchable has always been a dream for every business that is out there. So, Google in the early days, we did actually have, we thought that the main way we're going to make money is with enterprise search. And we built this box, the yellow box. I don't know, most of you probably haven't seen it. It used to be an appliance that you would put in a customer's, you know, data center and somehow get it to connect with your internal enterprise systems and then make things searchable. So, like you have a Google-like thing inside your company. But Google failed at that initiative and not just once but like five times, you know, we tried to go and build something in the enterprise. I think the next four failures were more because I think Google had become primarily a consumer company and there was no real desire from the top, you know, leadership to actually solve this problem for the enterprise.

And when we were starting the company in 2019, this was after I had started a successful company called Rubrik. We had grown a lot, like, you know, our business was worth multiple billion dollars and so I was considered an entrepreneur who can potentially, you know, build another company. So I talked to a lot of investors about this new idea that I had for search and tell you why I chose to work on it. But the first thing I actually heard from many of them was that, you know, like, hey look, we would love to, you know, back you and invest in your company. You are a proven entrepreneur, but can you work on something other than search? Because nobody wanted to actually, you know, invest in this domain because there's so many failures. It was a graveyard of companies. But when you think about, like, why was it so hard to actually build a successful company in search? Search is a fundamentally hard problem. In some ways, you know, it's like you're trying to deliver magic because what are you telling people? You're telling them that, hey, come and ask a question and we'll surface the right information, the right answer back to you. And that's sort of like magic. That's like the oracle. That's the eight ball and people have very high expectations and it requires you to build a lot of technology behind the scenes. Now in the early days it was always hard.

Yeah. So the number one, like just connecting to enterprise information, is so hard, so difficult. And in the pre-SaaS world, like when we didn't have these cloud-based, easy-to-use software services, you actually would buy software on CDs, you'd have like hundreds of different versions of your enterprise software and your data was actually in private servers inside the data center and was extremely hard to just actually even go and connect with those systems. Cloud-based SaaS systems actually made it easy and I thought they made it really, really easy and that was actually the motivation for why I wanted to actually now solve the problem because I knew well if I have the data I can go and then actually make it searchable. I have like good experience with it and SaaS would allow me to actually have that data. So that was a big unlock. But then the second thing is what makes search special, what makes Google special, is that it gets to learn a lot from people. There are billions of searches that happen every day. Billion, you know, billion plus people using the product and people are actually constantly telling Google, like giving it so many signals that allows it to actually create that magical experience. For example, you ask for something and don't see good results and then you sort of tweak your search query. You change the words in it and now you're happy with what results you get back. But when you did that, when you issued those two searches, Google actually learned that those two phrases that you entered are conceptually similar, right? Even though the dictionary may not tell you so, but in colloquial terms, you know, they are similar because a user actually came and did the refinement on the query like that. So you learn a lot like that. And in the enterprise, unfortunately, you don't have that. You don't have enough documents to build PageRank, you don't have enough user activity to learn from people and that sort of then, like, when you accumulate a lot of information you could not disambiguate what is good information versus what's not relevant, you know, for any given topic. So that was another big issue in the enterprise and we solved it in a different way. I can go into the details of that.

Yeah. So like thinking about enterprise information, so you have it spread across hundreds of systems and also information has been produced over a long period of time. So there's also information that has now become obsolete, out of date. But when you think about like how humans are using that information inside the company, well, you are actually continuously doing things inside these systems. Employees are actually reading certain information, people are sharing information with each other, people are sharing documents. So if you start to actually observe activity and not just in your search system that you're building yourself, but if you start to observe activity inside your Google Drive or OneDrive, inside your SharePoint, observe what the actual conversations that are happening in Teams, or if you're using Zoom or meeting recording, like, see what discussions are happening there, you start to actually see there are a lot of valuable signals out there. For example, when somebody goes in Teams and asks a question and somebody else responds with an answer which is a link to a document, well, you can build that association that some smart human in your company thought that for that particular question that document had the right answer. And so if you start to observe and observe outside of the search product, there's a lot that you can tap into to sort of understand within an enterprise, number one, what information is up to date, what information has become stale, what information has high engagement versus what's actually sort of somewhere in the corner nobody's looking at that information. You can also see a given document, for example, is there a particular affinity to a team, a specific team. For example, there may be an onboarding guide but only your hardware team or your marketing team is looking at them, not other people. So it sort of helps you build that association that what knowledge matters to whom. So these are all the things that we had to solve. I think we built the product for the Glean product the same way as we built Google. It was just that the environment was different, the signals were different. But the core concept was the same. Try to learn as much from the human wisdom that is sort of buried inside your systems.

Well, look, I mean, yes, it is true that you have fewer number of documents and fewer interactions in the enterprise, but there are also some good things. One of them being that the information on the internet can be of very, very varying levels of quality. There is also a lot of people trying to fool you and a lot of people trying to actually get clicks and visits on your web page. So, you are in a different world. There are a lot of other challenges you have to solve and a lot of adversaries, but inside an enterprise you don't have adversaries. You also have highly trained workforce that is creating high-quality documents. And you also get to actually see over time, like one of the things that we do is on any given topic inside your company, we also build an understanding of who are the subject matter experts on that topic. For example, who's the latest expert on the LLM technology or who's an expert on verification of certain, you know, in certain domain. You get to actually see that, you get to actually build that knowledge graph behind the scenes and then you utilize that. So like, you know, why, like when you come and ask a question, we may not have a document that has a lot of visits yet, but we know that it's actually produced by a known expert on that topic and we still can actually sort of bubble it up and use that for the search.

Yeah. Well, so one, I actually, like, you know, we were talking about why we started Glean and what were the key moments or like, you know, like sort of insights that we had and I said that, you know, SaaS was one. But I think I didn't mention another one which is, you know, transformers and language models. They were, while like, you know, the rest of the world didn't know about it at the time, like today we all talk about LLMs and transformers and GPT. But in 2018 when we were starting, the only people who really cared about these large neural networks from a language perspective were search teams inside Google. And so me and like most of our early employees, we all came from Google and we actually saw that transformers were making a big impact on search. And so our first version of the product itself in 2019 was actually based on transformers. So like in that sense, you know, we're probably one of the first enterprise GenAI companies of the world. The term didn't exist at the time but we've been using transformers in Glean from those early days. All these terms got invented later, like GenAI or vector search, but that conceptual matching, like the ability of machines to understand what you're asking at a conceptual level and also look at the knowledge inside your company and understand that at a conceptual level, actually took search to the next level. So that was actually a key insight.

Well, I mean, I think we would have still been a really good product because I would say that transformers' impact at the time was maybe 10%. Even today, like in search, the impact of transformers is 20%. But like, you know, every one of those things matter at the end. If you have a product that actually answers questions three times out of five versus four times out of five correctly, that makes a big difference in usage and how people, you know, come back and start to trust your system.

Yeah. So when we started the company we used transformers at the time and we used it for conceptual or semantic matching. At that time the LLM technology was not powerful enough to generate their own words or sentences, their own answers, and the transformer technology was also not powerful enough to reason in a big way, but they could do fuzzy conceptual matching. So our first product, it actually looked, the look and feel of it was mostly like Google. Most of the times you would ask questions and we'll show you the best matching documents or conversations. Sometimes we'll actually extract answers because LLMs could not write on their own but they could actually extract some answers to specific questions. So that's how we started but then over the years the transformer technology has improved so much, the models have become so large and so much more capable and with those reasoning and generation capabilities it actually allowed us to actually do something different and evolve our product. So why, you know, we were actually before the ChatGPT launch and right after that as well, we were already on that path to actually utilize that latest capabilities from language models to reason and generate and change our core product experience fundamentally and move, you know, basically take it one step further. So before you could actually come and ask questions in Glean, you could ask them in natural language and we would actually surface the right documents from inside your company. But now, like just like ChatGPT, we could actually go one step further. We could take those documents that search returns and then make the model take the question and those results, the data from those results and actually synthesize its own response. So we could go to this more question answering mode where you are asking a question and now AI is actually writing the answer for you by reading the results that search returned. And so that was a RAG architecture that we launched in 2022 at the same time as ChatGPT. And it, like, you know, I think for a lot of people that was a big leap. Like, you know, people were not interested in sifting through, going through, you know, 10 documents, reading like long documents to get answers to a quick question they have. They want instant gratification, like, you know, I have a question just give me the answer, right? And so that was a big leap. LLM allowed us to actually advance our product. And so now if you think about Glean, think of it as ChatGPT, a more powerful version of ChatGPT inside your enterprise. You come to Glean, you ask questions and we will not only use all of the world's knowledge but also all of your internal company's private knowledge in a safe and secure way, knowing who you are, so only use the knowledge that you have permissions to use to then answer your questions or help you with your task. So that's how, that was the evolution and like these language models fundamentally played a role in sort of, you know, helping us like taking our product from a more Google-like to more ChatGPT-like experience.

Well, so first of all, every enterprise that we talk to is very keen on bringing AI technologies inside their company and they want to also ensure that their workforce is learning how to leverage AI, becoming AI-first, changing their habits and be ready for that future. So there's a very, very large appetite, like in terms of the desire from enterprises to do more things with AI. And this actually, and it's not limited to the leading adopters of technology like tech companies themselves or healthcare or financial services. It's actually every sector in the industry is like that. And I'll give you an interesting story. I went to meet with the CIO of a railroad company, a railroad company which is like 200 years old. And when I entered the office, the CIO actually asked me, why are you here? Don't you know that I don't buy technology? You know that I'm actually still using mainframes from 30 years back or like 50 years back and I've missed every single technology trend that you can name, whether it's internet or mobile or SaaS or cloud or, you know, you name it and I've missed it. So why are you here? And well, that was a difficult question and before I could think about an answer for it, he says, but don't worry, I'm not missing the AI trend and I'm going to actually leapfrog and I missed all of those, you know, technology revolutions but I'm going to catch up by actually embracing AI right away. So that's the level of appetite.

Now, how much are businesses succeeding? I think it's still a struggle. Like, you know, if you look into what businesses have, for example, doubled their revenue or reduced their workforce by half, you know, with AI, the answer is probably no. There's nobody out there that has done it. This is still early technology and what we're seeing as trends are, I would say, two different primary trends. Number one, like everybody is thinking about education first and how do I actually bring some AI tool that I can bring to everybody in the workforce so that they can just start, like people can start to bring AI in their day-to-day work lives, get comfortable with it and are not scared about it, like not scared about it and sort of be hostile to it because they feel like it's going to take your job away. So that's the first thing. And on that front, I would say that, like, you know, the most popular tools are like, you know, either ChatGPT or Copilot or Glean, because these are tools that like every employee can use on a day-to-day basis. Now, of course, we are the best, you know, that you won't be surprised to hear that from me, but among all of them, but that's sort of the first tool that you're bringing is a general-purpose AI tool that has all of the world's knowledge and ideally like also all of your company's knowledge that you can give to employees and is not specific, it's just ready for you to take whatever questions you have, bring the right answers, is ready to do some of your task for you. And that has seen success. So companies are deploying them at scale and employees are reporting good productivity savings from that.

But the other set of tools, like, you know, there's no shortage of AI products in the industry and enterprises are bombarded with thousands of these other pitches, you know, on a daily basis from AI companies. And largely, you know, people are struggling to see success on those vertical tools, I would say. For example, there is a tool that promises that, hey, you don't need business development representatives anymore, like you can just have AI call your customers and generate leads, you know, for your sales team. So there are products for that, but I think, you know, like when you start to think about going really deep and thinking about doing something that a human could do, I think AI is still not ready for that. And so whenever somebody makes huge big promises about completely eliminating, you know, a human from a workflow, from a business process, those tools are generally failing at the moment. Like there's still, you know, AI has issues with hallucination, it has issues with data connectivity. So things are still very early, I would say, in most enterprises. But some, you know, there are pockets of successes, I would say. Engineering, of course, like, you know, you're seeing how engineers are actually actively trying to use tools like, you know, these codegen tools, whether it's Copilot or others, to really sort of speed up your code development, to speed up code reviews. Customer support is a big domain where there's clear success with AI. We've delivered a lot of success with that with Glean. Large customers like T-Mobile, they have, you know, 50,000 plus customer service agents and they're resolving tickets, you know, 40% faster than before because AI is doing most of it for them now. So humans are still in the loop, you know, in each one of those use cases, but they're actually moving at twice the speed. And so there's significant sort of like wins across functions, across, you know, these use cases. But like, you know, when you think about where we will be, where we thought we will be today, we're behind. And I think where we will be in 5 years from now is going to be far more dramatic than, you know, the progress that we've seen today.

Yeah. Well, I think first I would say the concept of eliminating staff is not a positive construct for any business to think. You know, you have to always think about AI is not so much about efficiency gains. I mean, it does that. It does a good job at it. It can really automate a lot of things that you do today. But the bigger value of AI is, you know, augmenting all of us with new capabilities, allowing us to do, you know, work kind of work that we could not do before. I'll give you an example as a product manager. Like, you know, if you're building a product for enterprises, the way you think about your roadmap and new features that you want to build is that probably you'll go and talk to 10 or 15 of your most important customers and you'll conduct an interview. So you'll sort of see some features that they want from you and then based on that you build a roadmap. But that's very limited, that has a very narrow view. What we do now today at Glean is different. You know, we have, you know, tens of thousands of these customer and prospect calls where our potential customers or our existing customers are constantly telling us the problems that they're trying to solve, the needs that they have. But there's so much of that data that's across, like, you know, 100,000 customer calls and a product manager simply could not go through each one of them and listen to them. But now with AI, you know, you can actually build an agent that goes and goes through every one of those calls, understands the key feature requests that were actually talked about, organizes them by themes and gives you a prioritized roadmap based on the voice of, you know, those thousand customers as opposed to those 10 customers. And that's an example of a thing that as a product manager you couldn't do before but now you can.

So, in terms of like, you know, how long is it going to take for us to double our productivity? My guess is that, you know, we would get there. Like, you know, at a macro scale, I think I can make a very comfortable bet that the GDP of United States is not going to double in 5 years. It's not going to happen. But companies that leverage AI or departments that leverage AI, they will be able to easily go 2x faster in the work that they do without any doubts. So that's, and that is sort of like, you know, my take. And if you don't do it as a business then you are going to actually lose to competition.

Oh, like easily, like, I mean, like every one of us, like in this room, for example, like, you know, we should be looking to be able to do, like, you know, like at least double our productivity like within the next, you know, year or two.

Yeah, I think we are, overall metric-wise, I think we have, on the new lines of code that are being written, about 30% of them are being written by, like, whether it's autocomplete in the code or like, you know, a Copilot kind of...

A year from now, year ago it was probably zero. Like, I don't think we were using anything, any significant tools at the time. Like we had GitHub Copilot and it was doing some autocomplete for us.

Interesting, so you're not standardizing on one tool.

So you started a company doing enterprise search, but recently I noticed that you are kind of positioning yourself more than just enterprise search but also building enterprise workflow on top of the Glean system. Can you just give people some, like, what's the vision and then why did you do it?