May 1, 2023
Portfolio
Unusual

How Glean found product-market fit

Sandhya Hegde
No items found.
How Glean found product-market fitHow Glean found product-market fit
All posts
Editor's note: 

SFG 20: Arvind Jain on enterprise search and AI

In this episode of the Startup Field Guide podcast, Sandhya Hegde chats with Arvind Jain, CEO of Glean. Valued at $1B and serving over a hundred enterprise customers today, Glean leverages an enterprise-grade knowledge graph and LLMs (large language models) to help companies find and use their institutional knowledge across hundreds of internal tools and SaaS applications.  

Be sure to check out more Startup Field Guide Podcast episodes on Spotify, Apple, and YouTube. Hosted by Unusual Ventures General Partner Sandhya Hegde (former EVP at Amplitude), the SFG podcast uncovers how the top unicorn founders of today really found product-market fit.

If you are interested in learning more about the topics we discuss in this episode, please check out the Unusual Ventures resources on how founders can shape the future of generative AI and what AI builders should know about data protection and privacy.

TL;DR

  • Glean was created to help solve the problem of fragmented knowledge across enterprise information systems. Glean’s enterprise search tool has become indispensable for employees to find the answers they need at work.
  • Glean's PMF moment: To build their enterprise search product, Glean did a lot of custom work to connect with different cloud-based SAAS applications. They also had to understand who had permission to view particular types of content within a company. They built an enterprise graph to map out the relationship between a company’s knowledge systems and its people so that search results would provide relevant information to the searcher. 
  • LLM technology helps Glean understand user intent and synthesize and summarize relevant knowledge to answer questions effectively, potentially changing the search experience fundamentally.
  • Glean solves for hallucination by constraining its model with the right body of knowledge. Glean also provides citations for the responses that it provides, so searchers know where the knowledge is coming from.
  • Present-day AI technologies have unbounded capabilities and are capable of making products much smarter than they might have been even a year ago. The shift to AI has forced product builders to rethink how they build their products.

Episode transcript

Sandhya Hegde:

Today we'll be diving into the story of Glean, an emerging startup in the enterprise search space with over 100 enterprise customers and last valued at over $1 billion. The CEO of Glean is serial founder, Arvind Jain, a guru in enterprise software who was a distinguished engineer at Google before leaving it in 2014 to start Rubrik, a cybersecurity company last valued at $3.3 billion. In 2019, he started Glean and is joining us today to talk about everything from the internet to security to search.

Arvind Jain:

Thank you so much, Sandhya, for having me here.

Sandhya Hegde:

I am so excited to be chatting with you. You've had just an incredible 25-year career in technology, starting at Microsoft during the dot-com boom, so you saw the birth of the internet, and now you're building Glean at a time that AI is making search a new space of red-hot interest again.

So I'm really curious, how would you describe the platform shifts you saw in technology as you've been a part of these companies, the internet, mobile, cloud, and deep learning? And I'm curious whether this modern AI platform shift feels different to you in any way.

Arvind Jain:

Very much so. I guess we can start from the beginning of my career in the mid to late-'90s, and that was a time when the internet was just taking off. When we started, Amazon didn't exist. They just were getting started as a bookstore online. And I remember being at Microsoft at that time where internet was not a big thing for us. Of course, you know the cool businesses, they were Windows and Office.

And from there, I joined a small startup called Akamai, where the objective of that company was to help people serve their content on their webpages, their media assets, on the internet, because it was hard at that time. The internet was coming up, so it was very slow. And I remember being part of some of those early, for example, the video on the internet use cases, one of them being in 2000. We were working on bringing March Madness online. So first, it used to be on TV, a big event, and nobody imagined, at that time, that hey, you could actually be on the internet and watch it live, even if you're at work where you don't have access to TV.

So those are the early days. I remember with that, all you could do on the internet was have a really small tiny window in which you can actually stream the video because the web was slow, it just didn't have enough capacity.

Sandhya Hegde:

Right. You're given the 32-pixel box and you had to do it with what you can.

Arvind Jain:

Exactly. You had to squint through it, go back from watching it on a reasonable... We didn't have big-screen TVs, but you still had the 27-inch TVs. From there, you sort of go into a small window on your computer, but the good thing was that you could actually see it, you could see it at more places and then on-demand. So that was phenomenal, but that was, of course, we saw the internet come up and fundamentally change how people build applications, how people build businesses.

And then the next big revolution was mobile, which were driven by iPhone and Android. That was also very interesting. When these big transformations happen, it brings a lot of interesting technology shifts, especially for the products that you're building. So when mobile came, suddenly the way we would design our web applications at Google fundamentally changed. We built our applications to work on 13-inch screens or 15-inch screens on laptops or even bigger monitors. And now suddenly, there was a use case for you to actually present your application experience on these really small four-inch screens on the mobile. And also, you had to make it work on networks that were much slower than the DSL or cable internet speeds that you enjoy at home. So that creates a fundamental shift in terms of how you build and architect and design your services.

And now here we are in the 2020s, and the big revolution now is AI. And to me, it feels very, very different from those two big phenomena that happened in the '90s and the 2000s, like the internet and mobile, in the sense that this seems to me an even bigger transformation. In fact, I think the capabilities of the AI technology today are so unbounded in the sense that we just have to completely rethink how we should build our products or what our products should even be.

If you think about going back in time when mobile came, and say that I'm running an E-commerce business, so I know that yes, I need to make my business now work on that mobile device, I need to change the design of my website so that it can render nicely on a small screen. I need people to be able to order on those small devices, so we need to make sure that my website is fast and works well on these low-power devices. But it was very clear what I had to do as an owner of business to adapt to that new world.

But now with AI, I think it fundamentally changes. Your product changes completely, its ability, what it can do. It can go up 10 levels of shift in one go. So that's sort of the big difference that we are feeling with AI. It is, in our line of business, which is we are here to answer people's questions, so that sits right into the heart of what these large language model technologies, the technologies that they're actually bringing into play. So it actually helps us make our product much more smarter, more smarter than what we thought we could be one year back.

Sandhya Hegde:

That is fascinating. I think you said it so well, which is, previous technology shifts are very much about, "Okay, how will my product work with this new change?" Even with the move to cloud, it was like, "Oh, how will our backend adapt to this?" Whereas with AI, it's, "What will our product do even?"

It's hitting the strategy at every level at the same time, which is very disorienting.

It's fascinating that you started Glean in 2019, which is just before a lot of the transformer model application went from research-grade to consumer-grade. What was the inspiration behind starting Glean at first?

Arvind Jain:

Before we started Glean, I was doing another startup called Rubrik, and we were lucky to build a strong business in that company and as a result, we had really fast employee growth and in four years, we were more than 2,000 people in our company.

And one thing that happened with that growth was that we didn't get similar gains in our productivity. In fact, on a per-person basis, people were not able to get as many things done as they used to before. Everybody was struggling. So we felt that we grew fast but we couldn't actually figure out how to scale, get the same amount of output for the investment that we're putting into the business. And so as part of trying to figure out what was wrong, we conducted a pulse survey and asked people what can we do better as a company, and what are the key problems that you are facing today, which is hindering your ability to do great work?

And on this question, the biggest complaint that we got back from people was people said that, "Hey, I cannot find anything in this company. I don't know where to go and look for information. I don't know who to go and ask for help." And so this was the single biggest employee environment issue. Now with me, my background, I'm a search engineer and I was not surprised that nobody could find anything in the company. I myself couldn't find anything. And part of the reason was that we were built in this modern SaaS era where our company technology infrastructure was basically 300 different cloud-based SaaS applications. So our company knowledge was just completely fragmented across all of these different systems, and no wonder. Who has the time to go one by one into these different places to find the thing that you need?

So as a search engineer, the first thought that came to my mind was that, oh wait, this is a problem that we can solve, yes. We can put a Google-like search engine in front of all of our business content and that would make it easy for people to find things. And as part of trying to actually find a product like that, we realized that there's no product in the market that is solving this problem.

And in fact, not just us, but like every company in the world, every employee and every company in the world was struggling with this massive fragmentation of knowledge across many different systems and it's becoming very hard for people to find things at work. So it felt like a very important problem to me at that time. So I, along with some of my ex-colleagues from Google, we decided to solve for this problem and start Glean.

Sandhya Hegde:

One of the things that I have observed in my experience with larger companies that I would say pretty much all face the exact same problem that you described at Rubrik, which is if you scale fast, which is by the way, the best case scenario for a startup is that you're scaling fast because you have a lot of customer demand, your product works, you have investors at your doors. So this is the good bit, that you get to scale quickly, but it comes with this incredible feeling, especially for the folks who are there at the early stages of the company that everything is slowing down.

And I'm curious, I think I've often felt that one of the reasons why people don't prepare for this eventuality, given that it's happened to so many people before you, you should almost assume it's going to happen and you start hiring but you don't prepare for it, is that because it's very hard to put a number on it? Can you even value exactly what is the loss of productivity happening? Because you don't have good knowledge base, you don't have infrastructure. I feel like this is a very undervalued investment. Because people can't quantify it, they just undervalue it, and they can't imagine what “great” looks like so they undervalue the solutions potentially.

Has that been your experience at all?

Arvind Jain:

It has been. In fact, when we were starting Glean, one of the common feedback I would get from people who I would talk to was that, "Yes, hey, this is an important problem. I know I face it, I know people in our company face it, but I'm not sure how to quantify it. Exactly the problem that you're saying. I don't know how much value to associate to it, and what are the kinds of business value I'm going actually derive from it if I were to put this tool in production?"

And that's sort of understandable. Whenever there's something new, you're used to doing things one way and you have to take a leap of faith to change your ways and buy something different. So absolutely, we face this problem, but our belief was very simple. And then I asked people this, "Think about your personal life for a minute, do you think you can function without Google?" And everybody will immediately tell me that, "No, that's an absolutely necessary product that we need to function in our day-to-day life." And no surprise, that one product allowed Google to make itself the largest technology company in the world.

So it's a really important, really fundamental problem. And so our belief was and is that every person should have a great search and question-answering experience at work. So we stayed with that belief, but I think now, people are seeing the value of it more, especially once even when we went into a more distributed work model because of the pandemic. So suddenly, leaders see that hey, their employees are alone, at home doing the work. They don't have help available to them. So when they have questions, the alternative of waiting for somebody scheduling a meeting and then getting their answers, it's just going to completely slow you down. So it's important to self-enable people, make sure that they have tools where they can quickly ask questions and get their answers.

And then now finally, I would say with ChatGPT, I think it has really struck executives' heart because now you can actually see it's not just that, "Hey, can a search engine help me find a document?" Now you can see that the technology can actually help answer complicated questions and actually do a lot of work for you. For example, you can ask this question now: "Give me the top 10 customers who are likely to churn." And believe it or not, but with the AI and the LLM technology, it's going to figure it out. It's going to actually go hunt through information in your enterprise corpus, figure out what makes it likely for a customer to churn, what metrics are important to watch, and it can go actually in your metrics dashboard, look at whether those metrics are looking bad for some businesses, and actually give that answer back to you.

The reasoning capabilities of these AI technologies have actually gone to such level that now people can clearly see that this is going to have a transformational impact on our employee productivity. So it is no longer a question of is it valuable, how valuable it is, now it has become indispensable.

Sandhya Hegde:

One of the things we talk about a lot on this podcast is what you said about taking a leap of faith. When you're doing something new, your early customers have to be people who had always been looking for a solution for the problem, are willing to take a leap of faith on a half-built, just-ready product and give you a lot of feedback.

What was your hypothesis on who that customer segment would be, given that often people would just start with small startups for early adoption, but you are working on a problem that is much more relevant to bigger teams than tiny startups that can literally just work out of one room? So I'm curious, what was your hypothesis for who will be the desperate early customer for this? And how did you go about defining what would even be a minimum viable product?

Arvind Jain:

That's a great question. So one thing that we are lucky on is the fact that Glean is a product that we use every day ourselves. Everybody on our team uses this product, and every one of us has fundamentally faced this problem ourselves in our entire work career.

Even at Google, by the way, when we were at Google, it was incredibly hard. It was easy to find things on the internet, but it was incredibly hard to find things within Google. So it's a problem that we understand people like us face. So we started naturally with that as our first target customer segment, technology companies, because we're very familiar with them. We know that people, whether it's engineers in these companies or support agents or marketing people or sales engineers, or account executives, we know exactly what their information needs are. What do they struggle with on a day-to-day basis? And so we still felt comfortable that that's the segment that we wanted to start with first.

And in general, we feel like this problem actually starts to impact you fairly quickly. Even a 50-person company or a 100-person company ultimately accumulates a large number of apps and a lot of content. And at that point, a product like Glean already helps them because we saw the experience with ourselves because we adopted the product as we built it. When we were 50 people, we already knew that we could not live without it. That's how it indispensable it becomes because we built our habits to just... It was so easy for us to find things and that's what we need now.

We had a good sense of who the end user is and who the target buyer is, just based on our own experiences. And so we started with fast-moving, fast-growing companies. Typically, companies with between 500 and 2,000 employees, which are still, as you said, are able to move fast, are willing to experiment. Because I think for those startups, three or four years back, they were in that same place where they were looking for other progressive companies who would buy their product, so we see more willingness from them to buy products from ours. So that was the initial target segment for us.

Sandhya Hegde:

Lots of follow-up questions. So first, you said something really fascinating, which is even within Google, it was hard to find things about Google that you needed internally as a team. So maybe that's a good segue to ask, what is fundamentally different when it comes to consumer search and enterprise search? Obviously, the knowledge lives in different places, but from a search technology perspective, what was fundamentally different about building enterprise search with Glean?

Arvind Jain:

Just from a technology perspective, the internet is sort of homogeneous in some ways. Everything is a webpage, they're all linked with each other. And so when you start to build a search engine, the first technology is to actually bring all of that content into your search index. So you build these web crawlers and they're basically hitting every website in the world, fetching the webpages from them, and then saving the content of those into your search index.

And enterprise is very different in the sense that there are many different business applications and there is no standard way to bring content from all of those apps into your search system. In the past, you couldn't even bring content from them because they were closed systems. But now with the advent of the SaaS model, you have APIs through which you can actually connect with these applications and bring content so that it can make it searchable.

But the issue is that every app has unique and different APIs. So that's a big part of building a great enterprise search product is that you need to actually build these connections into all the different popular cloud-based SaaS applications, and it grew that custom work to bring that in.

The second big difference in technology is that enterprise content is governed. It has a notion of entitlements. If you think about your company information, not all of it is available to everyone in the company to make use of, unlike on the internet where every piece of information is available to all of us. So you actually build these measures of when you actually connect with these applications and bring content, you have to understand who has permissions to view that content versus not. And now, when somebody comes and does a search, you can only give them results and answers that are in the knowledge that they have access to. So that's a pretty significant technology challenge that we've solved.

And then finally, I would say Google is really smart. It almost reads your mind. You give it two words and it can make sense of exactly what you wanted and it's going to bring those answers back to you. And the reason it's able to do that is because there’s massive data and usage volume on the internet. Google learns from human intelligence, it learns from our behaviors all the time. I ask for something and I actually click on the third result because that, I think, is the best one, and a lot of other people did the same, it gives Google this, that I can move that up to number one.

In the enterprise, it's a little bit different. You don't have that much data, you don't have that many people, that much usage. So you use a different set of techniques to figure out what knowledge is still fresh and relevant, what knowledge is actually more important today versus not. So what Glean does there is we actually build what we call an enterprise graph where we actually take all of the company knowledge and then we also take all the company's people and then we start to build relationships between them. So we build this notion of, "Hey, this document or this piece of knowledge is more relevant to people in this particular engineering team. So when they come and ask a question, you should actually prioritize those resources.”

That's an example of how when you think about relevance, you have to think very, very differently as well. It's a lot of different, like you said, fundamentally different challenge, I would say, but the goal is the same, which is to help answer people's questions.

Sandhya Hegde:

Right. And how do LLMs change the game? Is it more that just, okay, now you have a much better interface for people to be able to navigate this knowledge graph or does it change something about how you would want to construct the knowledge graph itself to begin with? How big of an impact was it on your strategic thinking for how you are building Glean, and when was it? When was it that you were like, "Oh, okay, this changes everything. Got to go rethink our stack?"

Arvind Jain:

Yeah. So the language models, in fact, large language model technology has been in development for quite some time. When we started in 2019, we already had in our mind to use large language models to do semantic matching. And in fact, we use open-domain, large language models, which are published by Google, this model family is called BERT. And so we used these BERT models, which were trained on all the knowledge on the internet, and so they have these really awesome capabilities. They can actually understand, at a semantic level, that these two things are equal. And for example, if you're looking for a user guide or a product manual is one and the same thing. This model understands those kinds of things.

So we use these models when we started out in 2019 to power semantic matching in the Glean search engine. You don't have to actually exactly know the right words that were used in the documents to find those documents. You can actually freely express your questions in natural language and get the answers.

So that technology, we've been using for a while, but I think what happened last year, so with the GPT-3.5 or GPT-4, now with Google PaLM, the capabilities of these models, they're at a very, very different level compared to where the language models were in 2019. I don't think I could have imagined that this is how smart or, in fact, I guess what the right word is. It's actually, it's scary, I would say, how much they understand language.

Sandhya Hegde:

Yeah, I think I had the same reaction where I was like, "Yeah, no, I understand this. It is just token prediction and we'll do classification. It'll do better classification, it'll do better semantic matching, it's predicting tokens, it'll do some generation. Great." And then suddenly, you see it doing chain-of-thought reasoning and it's horrifying because I'm not sure anyone feels like we perfectly understand how that happened.

Arvind Jain:

We still are trying to fathom all the things that we can do with it. So some of it is how this technology plays into our core product. So number one, it actually helps us understand what a user is trying to do. When they come and ask a question, our ability to understand their intent goes up significantly if you use this model technology.

The second thing is about presentation. So the user asked a question, you sort of have the right set of documents or conversations, like a body of knowledge, that actually answers their questions. And in the past, we would actually show them that body of knowledge in form of search results. Sometimes we would extract specific answers if it was easy for us to extract them and show them at the top. But that depended on only for a limited set of queries.

But now, what you can do is you can actually take all of that body of knowledge that is relevant for that question and you can actually take the element and ask it to consume that and actually precisely answer the question for the user. So synthesize it, summarize it, don't make them go through 10 different documents, don't make them go through 30 pages on a single document. Instead, answer that question in a very effective way. So that's where the power of these models come in. So yes, it's actually, in some ways, it changes the search experience fundamentally.

Sandhya Hegde:

How do you think the ecosystem will evolve in terms of this idea of knowledge management being a very kind of narrow market and use cases in terms of the idea of, "Oh, we'll spend software on knowledge management?"If you talk to people about the category of knowledge management, it was, "Oh yeah, Atlassian has a business," and it's a few specialized things like sales enablement, which is kind of knowledge management, but it was considered a small space. And I feel like this will essentially create a much larger category of software that is around, whether you call it research or knowledge management or search, whatever the names are, but it's basically automating this act of I'll go consume 10 documents, which are hopefully the right 10 documents, and synthesize what I'm learning, either for myself, or to present to any other people. That was just literally the work that pretty much every white collar knowledge worker does, either a small or a big part of their workflow. Is that, right? It's definitely a considerable part of my workflow, for example.  So it feels like this will just make that category of software 100x bigger. So I'm curious how you're thinking about what are all the ramifications and does this change Glean's vision in any way? Does this mean there is an even bigger opportunity for you to chase?

Arvind Jain:

So I think our mission and the way we have actually looked at our opportunity, we've always thought of ourselves as that we could be the de facto search solution for all businesses in the world, which means that every person, every worker in every business is using Glean 10 times a day to get more things done. So the vision has always been there, but now we feel like it's actually closer in terms of how we achieve it.

And in some ways, the way to think about what is this new product, when we started out, we always called ourselves the work assistant. It's not just that, hey, when you have a question, we will answer that question for you. Part of helping you be more effective at work is also to surface relevant information to you so that you can be knowledgeable in your job. Part of it is actually helping you understand that what things need your attention today right now. So there's always been that discovery element, which is also part of Glean.

And now if you think about it, so you think about the way to describe Glean is that, imagine a human, an expert in your company that has been there in the company since the beginning from day one, and they've actually read every single document that has ever been produced in the company. They've been part of every single conversation between any two people and they have the magical capability of not forgetting anything. And now they're actually available to you 24/7. Whenever you have a question, they can use all of that knowledge that they know about, and they can also use their human expertise and actually quickly answer your questions. So imagine what that does for you. What does that do for your work?And I think that's the vision that's not becoming reality with these advances in AI technologies.

It's a big market and for us. If you think about it, it's not just knowledge management, it's not just knowledge access. It actually goes beyond that because now the technology is also like a command center for you. It can actually do things for you. You can ask it to write something, you can ask it to actually take actions on your behalf. And that's where we are still trying to figure out: “How do you focus and how do you use this technology in a focused fashion?” Because I think... And that's what I was referring to you before about this technology is so powerful that I think everybody feels like there's so much that you have access to, you can bring great experiences, but then you have to figure out how do you focus?

Sandhya Hegde:

How do you focus? How do you control? You almost want to put more guardrails in place. And maybe speaking of focus and this broader vision, how are you tackling hallucination given that that's an uncomfortable, but okay, feature if you're just generating content, but probably a deal breaker if you want to really rely on accurate answers for questions about facts like, oh, who are the 10 customers at risk of churn? I want to be absolutely sure that it made 10 customers that actually work with my company and have risks involved. How are you dealing with the hallucination problem?

Arvind Jain:

If you think about these models, they feel like they know everything because the way they're built is that they're going to actually predict the next piece of text. The natural instinct of these models is that you can ask any question, it's going to actually produce an answer. And it could be random, it could be complete, like a false statement that they can make. And so there is obviously work happening in terms of how do you control that and how do you actually teach it to say that, "Okay, no, I have not seen this. I don't have the ability to answer this question?" But I don't think it's there yet, so you have to work hard on it.

So we use these models in a different way. We use them in a very constrained fashion. And what we do really is we combine the power of a good search technology with the reasoning capabilities of these models. So the end-to-end flow for us is a user comes and asks a question. Now, this is the use case where you cannot give them a random answer, you have to give them a precise correct answer because they're going to rely on it.

So the first thing that we do is we take that question that the user has asked and then we figure out, through our search system, what is the body of knowledge that you can actually use to answer this question? Try to find the most relevant documents for this particular question, and also limit them to documents that I personally have access to, so that don't answer questions using knowledge that I'm not really authorized to use. So we do that part and it sort of generates the body of knowledge for a given question.

And then when we invoke the model, we actually tell the model that, "Use only these documents to answer this question," and that is how you ground it. That is how you ensure that the certain piece of knowledge that you're actually okay with the model using is being used to produce answers. And then we actually also make sure that we are creating citations so that we actually always answer questions in Glean using these models. We'll actually say that, "Hey, we built up this answer from these two resources." So, basically, our solution for hallucination is to constrain the model with the right knowledge and tell it not to use the inherent knowledge that it has from the internet.

Sandhya Hegde:

Right, right. Yeah. Don't trust the pre-training data, if that makes sense.

Arvind Jain:

Yeah. Use only the language capabilities of it, not the knowledge, not the knowledge that is inside the model.

Sandhya Hegde:

Yeah. Don't use that Reddit feed for answering our questions.

Arvind Jain:

Exactly.

Sandhya Hegde:

It makes sense. So we're switching gears a little bit to go-to-market strategy, and I see a lot of AI startups really trying to leverage more product-led growth given that they have to teach customers what the capabilities of the product even are and being able to let them play with it themselves before pitching a deployment seems to be very effective. But then there's kind of issues around security, privacy, all of that. So what are you trying to do with Glean in terms of at-scale go-to-market strategy?

Arvind Jain:

Yeah. So Glean uses a traditional enterprise sales model where we are actually not having individual users sign up for Glean and start using it. We typically go and work with the CIO or the technology team at our customer and work with them in a proper framework and bring the technology to their entire organization. So that's the model that we follow.

And it's really important for a product like Glean because we connect with all of the company information. And so with that, security is a very, very important consideration. We need to make sure that we are compliant with whatever policies are in place at that organization in terms of where data can decide, and who has access to it. So it's very important to be careful in that sense, and the product-led growth model where individuals are signing up and sharing their sensitive company data with Glean, we believe that's not the right model. So, therefore, we've actually gone with that traditional work with the security team at our customers, work with the technology team, and roll out the product in compliance with their requirements.

Sandhya Hegde:

And maybe a broader question about the ecosystem, the growth funding environment right now is incredibly tough and if past data is a guide for us, it's not over. We haven't probably really hit the bottom of this bear market yet and venture growth funding or maybe even the public markets. So it's going to be at least another year, at least that's kind of what we are planning around.

Having been a founder several times, having seen a lot of the ups and downs of the technology market cycles, what's your advice to founders who only have 12 months or less of runway right now and are thinking through, "Okay, I don't have a lot of runway. Also, there's been this big technology shift that has potentially created even more opportunities for what my team can try in terms of either pivoting or  expanding of our product vision?” What would be your advice to founders in this situation right now?

Arvind Jain:

Well, so if you are an early-stage company, which is still looking for product-market fit, then I feel like you're not as impacted. I think we are, especially with the AI revolution that we're in the middle of, the amazing opportunities for companies to build new products that will have capabilities that people have just never seen before. So I think there's plenty of opportunities and I think folks should actually just focus on building a great product. I guess I'm an optimist in that sense that you build a great product and you will basically get the funding to further grow the company in the future.

But if you are in the somewhat, like you mentioned, growth stage before and now you have 12 months of runway left, and for me, I don't know the details of this because we have not been in the market. Luckily, we had the funds to actually keep investing in our technology and build the company for the long run, but that's a tough situation to be in. And I mean I think sometimes founders are very sensitive to taking a new round of money with lower valuation than before, and I think that should probably be removed out of the equation. I think building a company is a long-term thing and I think these short-term fluctuations don't matter. What's important is to keep your company capitalized to the right extent and it's okay to sometimes go and do a round, in my mind.

But I think the usual advice that of course everybody's saying today is that run the ship in a tight mode and conserve as much cash as you can, try to grow the business with few people because you don't know when you're going to be able to raise money again.

Sandhya Hegde:

Awesome. Well, thank you so much for joining us, Arvind. This was absolutely wonderful. I know our founders will also enjoy listening to it a lot and I definitely did and took away so many ideas around how my own workflow can change in the future and it looks like we'll all be using products like Glean.

All posts

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.

All posts
May 1, 2023
Portfolio
Unusual

How Glean found product-market fit

Sandhya Hegde
No items found.
How Glean found product-market fitHow Glean found product-market fit
Editor's note: 

SFG 20: Arvind Jain on enterprise search and AI

In this episode of the Startup Field Guide podcast, Sandhya Hegde chats with Arvind Jain, CEO of Glean. Valued at $1B and serving over a hundred enterprise customers today, Glean leverages an enterprise-grade knowledge graph and LLMs (large language models) to help companies find and use their institutional knowledge across hundreds of internal tools and SaaS applications.  

Be sure to check out more Startup Field Guide Podcast episodes on Spotify, Apple, and YouTube. Hosted by Unusual Ventures General Partner Sandhya Hegde (former EVP at Amplitude), the SFG podcast uncovers how the top unicorn founders of today really found product-market fit.

If you are interested in learning more about the topics we discuss in this episode, please check out the Unusual Ventures resources on how founders can shape the future of generative AI and what AI builders should know about data protection and privacy.

TL;DR

  • Glean was created to help solve the problem of fragmented knowledge across enterprise information systems. Glean’s enterprise search tool has become indispensable for employees to find the answers they need at work.
  • Glean's PMF moment: To build their enterprise search product, Glean did a lot of custom work to connect with different cloud-based SAAS applications. They also had to understand who had permission to view particular types of content within a company. They built an enterprise graph to map out the relationship between a company’s knowledge systems and its people so that search results would provide relevant information to the searcher. 
  • LLM technology helps Glean understand user intent and synthesize and summarize relevant knowledge to answer questions effectively, potentially changing the search experience fundamentally.
  • Glean solves for hallucination by constraining its model with the right body of knowledge. Glean also provides citations for the responses that it provides, so searchers know where the knowledge is coming from.
  • Present-day AI technologies have unbounded capabilities and are capable of making products much smarter than they might have been even a year ago. The shift to AI has forced product builders to rethink how they build their products.

Episode transcript

Sandhya Hegde:

Today we'll be diving into the story of Glean, an emerging startup in the enterprise search space with over 100 enterprise customers and last valued at over $1 billion. The CEO of Glean is serial founder, Arvind Jain, a guru in enterprise software who was a distinguished engineer at Google before leaving it in 2014 to start Rubrik, a cybersecurity company last valued at $3.3 billion. In 2019, he started Glean and is joining us today to talk about everything from the internet to security to search.

Arvind Jain:

Thank you so much, Sandhya, for having me here.

Sandhya Hegde:

I am so excited to be chatting with you. You've had just an incredible 25-year career in technology, starting at Microsoft during the dot-com boom, so you saw the birth of the internet, and now you're building Glean at a time that AI is making search a new space of red-hot interest again.

So I'm really curious, how would you describe the platform shifts you saw in technology as you've been a part of these companies, the internet, mobile, cloud, and deep learning? And I'm curious whether this modern AI platform shift feels different to you in any way.

Arvind Jain:

Very much so. I guess we can start from the beginning of my career in the mid to late-'90s, and that was a time when the internet was just taking off. When we started, Amazon didn't exist. They just were getting started as a bookstore online. And I remember being at Microsoft at that time where internet was not a big thing for us. Of course, you know the cool businesses, they were Windows and Office.

And from there, I joined a small startup called Akamai, where the objective of that company was to help people serve their content on their webpages, their media assets, on the internet, because it was hard at that time. The internet was coming up, so it was very slow. And I remember being part of some of those early, for example, the video on the internet use cases, one of them being in 2000. We were working on bringing March Madness online. So first, it used to be on TV, a big event, and nobody imagined, at that time, that hey, you could actually be on the internet and watch it live, even if you're at work where you don't have access to TV.

So those are the early days. I remember with that, all you could do on the internet was have a really small tiny window in which you can actually stream the video because the web was slow, it just didn't have enough capacity.

Sandhya Hegde:

Right. You're given the 32-pixel box and you had to do it with what you can.

Arvind Jain:

Exactly. You had to squint through it, go back from watching it on a reasonable... We didn't have big-screen TVs, but you still had the 27-inch TVs. From there, you sort of go into a small window on your computer, but the good thing was that you could actually see it, you could see it at more places and then on-demand. So that was phenomenal, but that was, of course, we saw the internet come up and fundamentally change how people build applications, how people build businesses.

And then the next big revolution was mobile, which were driven by iPhone and Android. That was also very interesting. When these big transformations happen, it brings a lot of interesting technology shifts, especially for the products that you're building. So when mobile came, suddenly the way we would design our web applications at Google fundamentally changed. We built our applications to work on 13-inch screens or 15-inch screens on laptops or even bigger monitors. And now suddenly, there was a use case for you to actually present your application experience on these really small four-inch screens on the mobile. And also, you had to make it work on networks that were much slower than the DSL or cable internet speeds that you enjoy at home. So that creates a fundamental shift in terms of how you build and architect and design your services.

And now here we are in the 2020s, and the big revolution now is AI. And to me, it feels very, very different from those two big phenomena that happened in the '90s and the 2000s, like the internet and mobile, in the sense that this seems to me an even bigger transformation. In fact, I think the capabilities of the AI technology today are so unbounded in the sense that we just have to completely rethink how we should build our products or what our products should even be.

If you think about going back in time when mobile came, and say that I'm running an E-commerce business, so I know that yes, I need to make my business now work on that mobile device, I need to change the design of my website so that it can render nicely on a small screen. I need people to be able to order on those small devices, so we need to make sure that my website is fast and works well on these low-power devices. But it was very clear what I had to do as an owner of business to adapt to that new world.

But now with AI, I think it fundamentally changes. Your product changes completely, its ability, what it can do. It can go up 10 levels of shift in one go. So that's sort of the big difference that we are feeling with AI. It is, in our line of business, which is we are here to answer people's questions, so that sits right into the heart of what these large language model technologies, the technologies that they're actually bringing into play. So it actually helps us make our product much more smarter, more smarter than what we thought we could be one year back.

Sandhya Hegde:

That is fascinating. I think you said it so well, which is, previous technology shifts are very much about, "Okay, how will my product work with this new change?" Even with the move to cloud, it was like, "Oh, how will our backend adapt to this?" Whereas with AI, it's, "What will our product do even?"

It's hitting the strategy at every level at the same time, which is very disorienting.

It's fascinating that you started Glean in 2019, which is just before a lot of the transformer model application went from research-grade to consumer-grade. What was the inspiration behind starting Glean at first?

Arvind Jain:

Before we started Glean, I was doing another startup called Rubrik, and we were lucky to build a strong business in that company and as a result, we had really fast employee growth and in four years, we were more than 2,000 people in our company.

And one thing that happened with that growth was that we didn't get similar gains in our productivity. In fact, on a per-person basis, people were not able to get as many things done as they used to before. Everybody was struggling. So we felt that we grew fast but we couldn't actually figure out how to scale, get the same amount of output for the investment that we're putting into the business. And so as part of trying to figure out what was wrong, we conducted a pulse survey and asked people what can we do better as a company, and what are the key problems that you are facing today, which is hindering your ability to do great work?

And on this question, the biggest complaint that we got back from people was people said that, "Hey, I cannot find anything in this company. I don't know where to go and look for information. I don't know who to go and ask for help." And so this was the single biggest employee environment issue. Now with me, my background, I'm a search engineer and I was not surprised that nobody could find anything in the company. I myself couldn't find anything. And part of the reason was that we were built in this modern SaaS era where our company technology infrastructure was basically 300 different cloud-based SaaS applications. So our company knowledge was just completely fragmented across all of these different systems, and no wonder. Who has the time to go one by one into these different places to find the thing that you need?

So as a search engineer, the first thought that came to my mind was that, oh wait, this is a problem that we can solve, yes. We can put a Google-like search engine in front of all of our business content and that would make it easy for people to find things. And as part of trying to actually find a product like that, we realized that there's no product in the market that is solving this problem.

And in fact, not just us, but like every company in the world, every employee and every company in the world was struggling with this massive fragmentation of knowledge across many different systems and it's becoming very hard for people to find things at work. So it felt like a very important problem to me at that time. So I, along with some of my ex-colleagues from Google, we decided to solve for this problem and start Glean.

Sandhya Hegde:

One of the things that I have observed in my experience with larger companies that I would say pretty much all face the exact same problem that you described at Rubrik, which is if you scale fast, which is by the way, the best case scenario for a startup is that you're scaling fast because you have a lot of customer demand, your product works, you have investors at your doors. So this is the good bit, that you get to scale quickly, but it comes with this incredible feeling, especially for the folks who are there at the early stages of the company that everything is slowing down.

And I'm curious, I think I've often felt that one of the reasons why people don't prepare for this eventuality, given that it's happened to so many people before you, you should almost assume it's going to happen and you start hiring but you don't prepare for it, is that because it's very hard to put a number on it? Can you even value exactly what is the loss of productivity happening? Because you don't have good knowledge base, you don't have infrastructure. I feel like this is a very undervalued investment. Because people can't quantify it, they just undervalue it, and they can't imagine what “great” looks like so they undervalue the solutions potentially.

Has that been your experience at all?

Arvind Jain:

It has been. In fact, when we were starting Glean, one of the common feedback I would get from people who I would talk to was that, "Yes, hey, this is an important problem. I know I face it, I know people in our company face it, but I'm not sure how to quantify it. Exactly the problem that you're saying. I don't know how much value to associate to it, and what are the kinds of business value I'm going actually derive from it if I were to put this tool in production?"

And that's sort of understandable. Whenever there's something new, you're used to doing things one way and you have to take a leap of faith to change your ways and buy something different. So absolutely, we face this problem, but our belief was very simple. And then I asked people this, "Think about your personal life for a minute, do you think you can function without Google?" And everybody will immediately tell me that, "No, that's an absolutely necessary product that we need to function in our day-to-day life." And no surprise, that one product allowed Google to make itself the largest technology company in the world.

So it's a really important, really fundamental problem. And so our belief was and is that every person should have a great search and question-answering experience at work. So we stayed with that belief, but I think now, people are seeing the value of it more, especially once even when we went into a more distributed work model because of the pandemic. So suddenly, leaders see that hey, their employees are alone, at home doing the work. They don't have help available to them. So when they have questions, the alternative of waiting for somebody scheduling a meeting and then getting their answers, it's just going to completely slow you down. So it's important to self-enable people, make sure that they have tools where they can quickly ask questions and get their answers.

And then now finally, I would say with ChatGPT, I think it has really struck executives' heart because now you can actually see it's not just that, "Hey, can a search engine help me find a document?" Now you can see that the technology can actually help answer complicated questions and actually do a lot of work for you. For example, you can ask this question now: "Give me the top 10 customers who are likely to churn." And believe it or not, but with the AI and the LLM technology, it's going to figure it out. It's going to actually go hunt through information in your enterprise corpus, figure out what makes it likely for a customer to churn, what metrics are important to watch, and it can go actually in your metrics dashboard, look at whether those metrics are looking bad for some businesses, and actually give that answer back to you.

The reasoning capabilities of these AI technologies have actually gone to such level that now people can clearly see that this is going to have a transformational impact on our employee productivity. So it is no longer a question of is it valuable, how valuable it is, now it has become indispensable.

Sandhya Hegde:

One of the things we talk about a lot on this podcast is what you said about taking a leap of faith. When you're doing something new, your early customers have to be people who had always been looking for a solution for the problem, are willing to take a leap of faith on a half-built, just-ready product and give you a lot of feedback.

What was your hypothesis on who that customer segment would be, given that often people would just start with small startups for early adoption, but you are working on a problem that is much more relevant to bigger teams than tiny startups that can literally just work out of one room? So I'm curious, what was your hypothesis for who will be the desperate early customer for this? And how did you go about defining what would even be a minimum viable product?

Arvind Jain:

That's a great question. So one thing that we are lucky on is the fact that Glean is a product that we use every day ourselves. Everybody on our team uses this product, and every one of us has fundamentally faced this problem ourselves in our entire work career.

Even at Google, by the way, when we were at Google, it was incredibly hard. It was easy to find things on the internet, but it was incredibly hard to find things within Google. So it's a problem that we understand people like us face. So we started naturally with that as our first target customer segment, technology companies, because we're very familiar with them. We know that people, whether it's engineers in these companies or support agents or marketing people or sales engineers, or account executives, we know exactly what their information needs are. What do they struggle with on a day-to-day basis? And so we still felt comfortable that that's the segment that we wanted to start with first.

And in general, we feel like this problem actually starts to impact you fairly quickly. Even a 50-person company or a 100-person company ultimately accumulates a large number of apps and a lot of content. And at that point, a product like Glean already helps them because we saw the experience with ourselves because we adopted the product as we built it. When we were 50 people, we already knew that we could not live without it. That's how it indispensable it becomes because we built our habits to just... It was so easy for us to find things and that's what we need now.

We had a good sense of who the end user is and who the target buyer is, just based on our own experiences. And so we started with fast-moving, fast-growing companies. Typically, companies with between 500 and 2,000 employees, which are still, as you said, are able to move fast, are willing to experiment. Because I think for those startups, three or four years back, they were in that same place where they were looking for other progressive companies who would buy their product, so we see more willingness from them to buy products from ours. So that was the initial target segment for us.

Sandhya Hegde:

Lots of follow-up questions. So first, you said something really fascinating, which is even within Google, it was hard to find things about Google that you needed internally as a team. So maybe that's a good segue to ask, what is fundamentally different when it comes to consumer search and enterprise search? Obviously, the knowledge lives in different places, but from a search technology perspective, what was fundamentally different about building enterprise search with Glean?

Arvind Jain:

Just from a technology perspective, the internet is sort of homogeneous in some ways. Everything is a webpage, they're all linked with each other. And so when you start to build a search engine, the first technology is to actually bring all of that content into your search index. So you build these web crawlers and they're basically hitting every website in the world, fetching the webpages from them, and then saving the content of those into your search index.

And enterprise is very different in the sense that there are many different business applications and there is no standard way to bring content from all of those apps into your search system. In the past, you couldn't even bring content from them because they were closed systems. But now with the advent of the SaaS model, you have APIs through which you can actually connect with these applications and bring content so that it can make it searchable.

But the issue is that every app has unique and different APIs. So that's a big part of building a great enterprise search product is that you need to actually build these connections into all the different popular cloud-based SaaS applications, and it grew that custom work to bring that in.

The second big difference in technology is that enterprise content is governed. It has a notion of entitlements. If you think about your company information, not all of it is available to everyone in the company to make use of, unlike on the internet where every piece of information is available to all of us. So you actually build these measures of when you actually connect with these applications and bring content, you have to understand who has permissions to view that content versus not. And now, when somebody comes and does a search, you can only give them results and answers that are in the knowledge that they have access to. So that's a pretty significant technology challenge that we've solved.

And then finally, I would say Google is really smart. It almost reads your mind. You give it two words and it can make sense of exactly what you wanted and it's going to bring those answers back to you. And the reason it's able to do that is because there’s massive data and usage volume on the internet. Google learns from human intelligence, it learns from our behaviors all the time. I ask for something and I actually click on the third result because that, I think, is the best one, and a lot of other people did the same, it gives Google this, that I can move that up to number one.

In the enterprise, it's a little bit different. You don't have that much data, you don't have that many people, that much usage. So you use a different set of techniques to figure out what knowledge is still fresh and relevant, what knowledge is actually more important today versus not. So what Glean does there is we actually build what we call an enterprise graph where we actually take all of the company knowledge and then we also take all the company's people and then we start to build relationships between them. So we build this notion of, "Hey, this document or this piece of knowledge is more relevant to people in this particular engineering team. So when they come and ask a question, you should actually prioritize those resources.”

That's an example of how when you think about relevance, you have to think very, very differently as well. It's a lot of different, like you said, fundamentally different challenge, I would say, but the goal is the same, which is to help answer people's questions.

Sandhya Hegde:

Right. And how do LLMs change the game? Is it more that just, okay, now you have a much better interface for people to be able to navigate this knowledge graph or does it change something about how you would want to construct the knowledge graph itself to begin with? How big of an impact was it on your strategic thinking for how you are building Glean, and when was it? When was it that you were like, "Oh, okay, this changes everything. Got to go rethink our stack?"

Arvind Jain:

Yeah. So the language models, in fact, large language model technology has been in development for quite some time. When we started in 2019, we already had in our mind to use large language models to do semantic matching. And in fact, we use open-domain, large language models, which are published by Google, this model family is called BERT. And so we used these BERT models, which were trained on all the knowledge on the internet, and so they have these really awesome capabilities. They can actually understand, at a semantic level, that these two things are equal. And for example, if you're looking for a user guide or a product manual is one and the same thing. This model understands those kinds of things.

So we use these models when we started out in 2019 to power semantic matching in the Glean search engine. You don't have to actually exactly know the right words that were used in the documents to find those documents. You can actually freely express your questions in natural language and get the answers.

So that technology, we've been using for a while, but I think what happened last year, so with the GPT-3.5 or GPT-4, now with Google PaLM, the capabilities of these models, they're at a very, very different level compared to where the language models were in 2019. I don't think I could have imagined that this is how smart or, in fact, I guess what the right word is. It's actually, it's scary, I would say, how much they understand language.

Sandhya Hegde:

Yeah, I think I had the same reaction where I was like, "Yeah, no, I understand this. It is just token prediction and we'll do classification. It'll do better classification, it'll do better semantic matching, it's predicting tokens, it'll do some generation. Great." And then suddenly, you see it doing chain-of-thought reasoning and it's horrifying because I'm not sure anyone feels like we perfectly understand how that happened.

Arvind Jain:

We still are trying to fathom all the things that we can do with it. So some of it is how this technology plays into our core product. So number one, it actually helps us understand what a user is trying to do. When they come and ask a question, our ability to understand their intent goes up significantly if you use this model technology.

The second thing is about presentation. So the user asked a question, you sort of have the right set of documents or conversations, like a body of knowledge, that actually answers their questions. And in the past, we would actually show them that body of knowledge in form of search results. Sometimes we would extract specific answers if it was easy for us to extract them and show them at the top. But that depended on only for a limited set of queries.

But now, what you can do is you can actually take all of that body of knowledge that is relevant for that question and you can actually take the element and ask it to consume that and actually precisely answer the question for the user. So synthesize it, summarize it, don't make them go through 10 different documents, don't make them go through 30 pages on a single document. Instead, answer that question in a very effective way. So that's where the power of these models come in. So yes, it's actually, in some ways, it changes the search experience fundamentally.

Sandhya Hegde:

How do you think the ecosystem will evolve in terms of this idea of knowledge management being a very kind of narrow market and use cases in terms of the idea of, "Oh, we'll spend software on knowledge management?"If you talk to people about the category of knowledge management, it was, "Oh yeah, Atlassian has a business," and it's a few specialized things like sales enablement, which is kind of knowledge management, but it was considered a small space. And I feel like this will essentially create a much larger category of software that is around, whether you call it research or knowledge management or search, whatever the names are, but it's basically automating this act of I'll go consume 10 documents, which are hopefully the right 10 documents, and synthesize what I'm learning, either for myself, or to present to any other people. That was just literally the work that pretty much every white collar knowledge worker does, either a small or a big part of their workflow. Is that, right? It's definitely a considerable part of my workflow, for example.  So it feels like this will just make that category of software 100x bigger. So I'm curious how you're thinking about what are all the ramifications and does this change Glean's vision in any way? Does this mean there is an even bigger opportunity for you to chase?

Arvind Jain:

So I think our mission and the way we have actually looked at our opportunity, we've always thought of ourselves as that we could be the de facto search solution for all businesses in the world, which means that every person, every worker in every business is using Glean 10 times a day to get more things done. So the vision has always been there, but now we feel like it's actually closer in terms of how we achieve it.

And in some ways, the way to think about what is this new product, when we started out, we always called ourselves the work assistant. It's not just that, hey, when you have a question, we will answer that question for you. Part of helping you be more effective at work is also to surface relevant information to you so that you can be knowledgeable in your job. Part of it is actually helping you understand that what things need your attention today right now. So there's always been that discovery element, which is also part of Glean.

And now if you think about it, so you think about the way to describe Glean is that, imagine a human, an expert in your company that has been there in the company since the beginning from day one, and they've actually read every single document that has ever been produced in the company. They've been part of every single conversation between any two people and they have the magical capability of not forgetting anything. And now they're actually available to you 24/7. Whenever you have a question, they can use all of that knowledge that they know about, and they can also use their human expertise and actually quickly answer your questions. So imagine what that does for you. What does that do for your work?And I think that's the vision that's not becoming reality with these advances in AI technologies.

It's a big market and for us. If you think about it, it's not just knowledge management, it's not just knowledge access. It actually goes beyond that because now the technology is also like a command center for you. It can actually do things for you. You can ask it to write something, you can ask it to actually take actions on your behalf. And that's where we are still trying to figure out: “How do you focus and how do you use this technology in a focused fashion?” Because I think... And that's what I was referring to you before about this technology is so powerful that I think everybody feels like there's so much that you have access to, you can bring great experiences, but then you have to figure out how do you focus?

Sandhya Hegde:

How do you focus? How do you control? You almost want to put more guardrails in place. And maybe speaking of focus and this broader vision, how are you tackling hallucination given that that's an uncomfortable, but okay, feature if you're just generating content, but probably a deal breaker if you want to really rely on accurate answers for questions about facts like, oh, who are the 10 customers at risk of churn? I want to be absolutely sure that it made 10 customers that actually work with my company and have risks involved. How are you dealing with the hallucination problem?

Arvind Jain:

If you think about these models, they feel like they know everything because the way they're built is that they're going to actually predict the next piece of text. The natural instinct of these models is that you can ask any question, it's going to actually produce an answer. And it could be random, it could be complete, like a false statement that they can make. And so there is obviously work happening in terms of how do you control that and how do you actually teach it to say that, "Okay, no, I have not seen this. I don't have the ability to answer this question?" But I don't think it's there yet, so you have to work hard on it.

So we use these models in a different way. We use them in a very constrained fashion. And what we do really is we combine the power of a good search technology with the reasoning capabilities of these models. So the end-to-end flow for us is a user comes and asks a question. Now, this is the use case where you cannot give them a random answer, you have to give them a precise correct answer because they're going to rely on it.

So the first thing that we do is we take that question that the user has asked and then we figure out, through our search system, what is the body of knowledge that you can actually use to answer this question? Try to find the most relevant documents for this particular question, and also limit them to documents that I personally have access to, so that don't answer questions using knowledge that I'm not really authorized to use. So we do that part and it sort of generates the body of knowledge for a given question.

And then when we invoke the model, we actually tell the model that, "Use only these documents to answer this question," and that is how you ground it. That is how you ensure that the certain piece of knowledge that you're actually okay with the model using is being used to produce answers. And then we actually also make sure that we are creating citations so that we actually always answer questions in Glean using these models. We'll actually say that, "Hey, we built up this answer from these two resources." So, basically, our solution for hallucination is to constrain the model with the right knowledge and tell it not to use the inherent knowledge that it has from the internet.

Sandhya Hegde:

Right, right. Yeah. Don't trust the pre-training data, if that makes sense.

Arvind Jain:

Yeah. Use only the language capabilities of it, not the knowledge, not the knowledge that is inside the model.

Sandhya Hegde:

Yeah. Don't use that Reddit feed for answering our questions.

Arvind Jain:

Exactly.

Sandhya Hegde:

It makes sense. So we're switching gears a little bit to go-to-market strategy, and I see a lot of AI startups really trying to leverage more product-led growth given that they have to teach customers what the capabilities of the product even are and being able to let them play with it themselves before pitching a deployment seems to be very effective. But then there's kind of issues around security, privacy, all of that. So what are you trying to do with Glean in terms of at-scale go-to-market strategy?

Arvind Jain:

Yeah. So Glean uses a traditional enterprise sales model where we are actually not having individual users sign up for Glean and start using it. We typically go and work with the CIO or the technology team at our customer and work with them in a proper framework and bring the technology to their entire organization. So that's the model that we follow.

And it's really important for a product like Glean because we connect with all of the company information. And so with that, security is a very, very important consideration. We need to make sure that we are compliant with whatever policies are in place at that organization in terms of where data can decide, and who has access to it. So it's very important to be careful in that sense, and the product-led growth model where individuals are signing up and sharing their sensitive company data with Glean, we believe that's not the right model. So, therefore, we've actually gone with that traditional work with the security team at our customers, work with the technology team, and roll out the product in compliance with their requirements.

Sandhya Hegde:

And maybe a broader question about the ecosystem, the growth funding environment right now is incredibly tough and if past data is a guide for us, it's not over. We haven't probably really hit the bottom of this bear market yet and venture growth funding or maybe even the public markets. So it's going to be at least another year, at least that's kind of what we are planning around.

Having been a founder several times, having seen a lot of the ups and downs of the technology market cycles, what's your advice to founders who only have 12 months or less of runway right now and are thinking through, "Okay, I don't have a lot of runway. Also, there's been this big technology shift that has potentially created even more opportunities for what my team can try in terms of either pivoting or  expanding of our product vision?” What would be your advice to founders in this situation right now?

Arvind Jain:

Well, so if you are an early-stage company, which is still looking for product-market fit, then I feel like you're not as impacted. I think we are, especially with the AI revolution that we're in the middle of, the amazing opportunities for companies to build new products that will have capabilities that people have just never seen before. So I think there's plenty of opportunities and I think folks should actually just focus on building a great product. I guess I'm an optimist in that sense that you build a great product and you will basically get the funding to further grow the company in the future.

But if you are in the somewhat, like you mentioned, growth stage before and now you have 12 months of runway left, and for me, I don't know the details of this because we have not been in the market. Luckily, we had the funds to actually keep investing in our technology and build the company for the long run, but that's a tough situation to be in. And I mean I think sometimes founders are very sensitive to taking a new round of money with lower valuation than before, and I think that should probably be removed out of the equation. I think building a company is a long-term thing and I think these short-term fluctuations don't matter. What's important is to keep your company capitalized to the right extent and it's okay to sometimes go and do a round, in my mind.

But I think the usual advice that of course everybody's saying today is that run the ship in a tight mode and conserve as much cash as you can, try to grow the business with few people because you don't know when you're going to be able to raise money again.

Sandhya Hegde:

Awesome. Well, thank you so much for joining us, Arvind. This was absolutely wonderful. I know our founders will also enjoy listening to it a lot and I definitely did and took away so many ideas around how my own workflow can change in the future and it looks like we'll all be using products like Glean.

All posts

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.