Thanks everybody. We'll go ahead and get started. Arvind, I just wanted to do a little bit of background so folks have a good understanding. You spent much of your career working on the problem of how people find and use information. Before founding Glean in 2019, you co-founded Rubrik, a data security company, and spent more than a decade at Google, including work on search, YouTube, Maps, and Google Fiber. The idea for Glean came from a different kind of workplace problem you encountered at Rubrik. Employees said they couldn't find information inside the company, and you've said you couldn't either. You founded Glean to help companies make their own knowledge more accessible through AI. And through Glean, you've now had the view into thousands of companies, enterprises, how they're trying to go from AI pivots to scale deployment. So today we're going to talk a little bit about that: what separates companies from getting real results to those that are stuck in proof of concept, and what leaders can do differently now. So, I'll just jump in. As I mentioned, you've watched thousands of companies try to move from AI experiments toward deployment. What has changed the most, would you say, in the last six months? And can you point to a company or team that illustrates what scaling looks like in practice?
Yeah, well, first of all, it's really nice to be here. Thank you, everyone. In terms of changes, AI, of course, has been this technology where change is faster than we can imagine. In the last six months, the capability of language models has fundamentally changed. I mean, they are capable of doing so many things that we couldn't do six months back, or in fact even three months back, with the latest models. So how is that translating into what enterprises are doing with AI? We are seeing a few different shifts. One, people are no longer talking about pilots. They're not thinking about that, hey, like I need to experiment, I need to actually maybe try a few different things, you know, and see what sticks.
The narrative has shifted, and businesses now have very clear official plans for all the different agents that they're going to build, the automation that they're going to bring, and these are primary OKRs, primary objectives for the business. We're also seeing significant growth in agent development. AI has moved from its primary use case, which was helping people find information, helping people get answers to their questions. This is how AI really started at a broad basis, like in the companies, that's the first thing that every business does is bring a tool like ChatGPT or Claude or Glean, for that matter, that can actually help every employee quickly get productive with AI as their personal assistant, as their information finder. But it has evolved from that into real business impact and an agent roadmap. So what we're seeing now are businesses are building agents across all of their business functions, which are adding significant productivity improvements to bottom line or top line. As an example, maybe I'll pick Zillow. You know, that's a company that started their journey with Glean about a year back. The initial use case with AI was they wanted to bring a ChatGPT-like experience to everybody inside the company, but one that was fully connected to your company's context. So that, of course, you know, Glean does. But then in the last one year, the story has shifted from Glean being an employee tool for every employee at Zillow to a platform that is actually helping them agentify their business. So today they have 3,400 agents that they've built across all of their business functions, everything from RFPs to performance reviews that people have to fill out to internal employee assistance. All of these, you know, business processes are now being done by agents, you know, at the company.
That's amazing. I guess I wonder as well how you keep track of that many agents with that many things going on. Is there any thought you have there, just in terms of keeping track of the volume?
Yeah, it's an interesting question. Like AI, while it is very useful, like in terms of it allows you to do a lot of things, create a lot of things, we're seeing across the businesses it's also becoming a problem. There's a lot of AI slop that is being created because, like, creating is easy; in fact, consuming is hard now. A lot of skills that are being developed, a lot of agents, there's a lot of duplication around that. So different businesses are using different strategies to cope with that. Sometimes, you know, you regulate; you don't allow everybody to build agents. Sometimes you have a model of moderation where everybody can build, but you can only publish once a moderator approves these agents so that other people can find it. Or they use, you know, the free-for-all approach where let everybody go and build everything, but then have a good search on top of your skills, agents, and of course, documents that allow people to actually find the best skills, the best agents for the use cases, you know, that they have personally. So it's sort of like different strategies, you know, that businesses are using. And I think ultimately what we will see is that, as has been with knowledge, where businesses can't really control like how much documentation gets produced inside a company, I think we'll run into the same issue with agents and skills, and search will be the final answer in terms of how do people actually find the good ones.
I'm curious about companies getting real outcomes from those that are stuck in pilots. We touched on this a little bit, but I'm curious if there's a hurdle that you've seen or you can think of that you've helped a company overcome some way, where there was a blockage and you were able to get past that.
With AI, the key sort of challenges that prevent widescale deployment of it inside your enterprise has, number one, it's been security. Businesses don't feel comfortable. AI agents, as an example, do all kinds of things inside the company. Give them enough permissions to take actions, mutate your company's data and knowledge. So security is always a tricky subject to deal with. You have to bring the AI platforms that allow you to sort of properly sort of guardrail, you know, all of your AI workloads. So that's been one big blocker.
Second one is context, which is as people bring AI solutions in, they build agents, and they realize that these agents are not really that smart. You can point an agent to your internal company's data and knowledge, and then let them execute workflows that humans were executing before, and you realize that, you know, at best these agents are performing at the level of a new employee inside your business. Because if you think about it, a new employee also has access to all the information that they need to do their work, but they don't have that context and that experience and that judgment yet to actually do work, you know, that is complex in nature. And so we often hear this, like, you know, that my agents feel very naive. They feel like, you know, day-one employees. How do I actually get them to the level of human experts? And context is the key there. Like, you know, you have to actually take all the data, all the knowledge that's inside your company, you have to bring it in one place, you have to start building a deep understanding of your business and how work actually happens.
In fact, even if you were to take all the data and all the documents, that still probably only represents 10% or 20% of your overall human wisdom, the collective intelligence of the enterprise, because a lot of it is still in people's heads.
But the good thing is that it's actually getting expressed. Like people are expressing intelligence every day as part of their day-to-day work. Like somebody may ask me a question on Slack or Teams, and I answer that, and as part of that, I actually share my experience. Maybe I'm approving an expense report which violates policies, but I put in a note there expressing why it's okay to approve it, and I'm using my judgment there. So people do express a lot of their intelligence in these informal sort of interactions, you know, whether it's meetings or conversations or working in different enterprise systems. So that's like, if you're able to tap into that, that is when you start to get close to human level of intelligence, and that has been very hard for companies. Like, how do you bring all of that knowledge, all of that informal, tacit sort of wisdom, you know, that's in your enterprise, and bring it to AI in a way that AI can consume it? So these are some of the key challenges, you know, why AI sort of remains in this pilot land or not driving real high value.
And what do you do about those challenges? I mean, if you can track a Slack message, an agent can read what you're saying, gets a little bit more information, but that's still only 10 to 20% of what you know in an enterprise. How do you get that other part that is maybe just in someone's brain, but that they never actually say out loud, or that they never Slack to somebody else, to then have the agents understand that context?
Yeah, well, I mean, that's what I was saying. So the 10, 20% number that I was mentioning, I was referring to like actual structured data and long-form documents.
But then these days, like as an enterprise, you can actually ensure, like you want to make sure that you have the right mechanisms, the right tools in your company to sort of capture more and more of that tacit knowledge that exists.
So, for example, recording like every meeting that you know internally you have between team members or with your customers. The more you record, the more you can now tap into that through AI. So that's sort of one thing, like, you know, how do you digitize more and more of your interactions, and you don't have to think about like processing or doing more complex things. It's a matter of just recording. Once you record, AI is actually pretty good at taking large amounts of data and knowledge and extracting those, you know, one or two insights, you know, that are in that data, and then use it, you know, bring it to agents so that agents can operate at the level of human experts.
Yeah, it sounds like an ongoing challenge, I would guess. Speaking of that, given some of the security discussion that we've had in recent weeks, you know, thinking of as companies give AI tools access to more internal knowledge, how should leaders balance usefulness with speed and security? And thinking about things like employee trust, all that. I mean, I'm curious about the risks there and what maybe people underestimate.
Yeah. So as we work with these large enterprises, we see the full spectrum of very conservative approach to AI, where there are businesses who have not given, you know, agents any power to mutate or change your systems, and also like given agents very limited access to your company's data and knowledge. So that's sort of one end of it, and the other end of it is, you know, fully drinking the AI Kool-Aid and letting agents go wild, like running long-running independent agents, making them your teammates. And as companies try to stay on these two ends of the spectrum, you know, there are problems on both sides.
Of course, if you're too conservative, you're not going to see business value from AI. If you are too aggressive, then bad stuff is going to happen. And the approach, which is kind of obvious, is you have to invest in the right platforms that can both bring the right context to your agents, but at the same time have the right guardrails in place. Guardrails which are dynamic guardrails, you know, where you bring the human in the loop when AI detects that, you know, there's an action that an agent is going to take which could be sensitive, which could be damaging. So it's sort of like, you know, you have to use that judgment as you build agents. We always guide our customers that you start slow in the sense that you can build agents that will replace a human completely, let's say that's your goal, but still, like, don't replace a human yet, and instead let that agent run and do everything that it needs to do, but let the final review happen through human in the loop. And that's actually very important because AI by itself is not intelligence. You know, the definition of AI is it has to learn from human behavior and kind of mimic it. And so if you build these agents where the first few months you have humans in the loop that provide active feedback to your agents, tell the agents when they are actually doing something wrong, something destructive, something bad, tell them when they're off, when they're not using the right information. If you have that collaboration between AI and humans for the first few months, they will bring the agents to a place where you figured out the right guardrails and you've also actually trained the agent to actually do the right things. So that's sort of the right approach, where every agent that you build, like, start with that human playing a bigger role in the beginning and then slowly sort of wean off.
When you mentioned human in the loops, it makes me think of a lot of the discussion that we have now around bottlenecks happening with that, and, you know, this idea of getting past the pilot stage to scale it broadly. Is it going to be possible to maintain human in the loop when we're trying to scale at such a big, you know, to such a big degree? It just feels like there's so much on the agent side but not enough on the human power side, if you will.
Well, I mean, right now, 100%, because if you think about what we're doing with AI, we're still mostly talking about replacing some of the work that we were doing as humans. And so from a workload perspective, if I have an agent, and let's say that I'm actually resolving a customer ticket as a customer support agent, or I'm a paralegal who's reviewing, you know, a commercial contract. Before, I had to do the whole thing myself, and now I have an agent that does most of the heavy lifting. It does 90% of the work for me, and all I have to do is, you know, do the final review of the work it did, and that's actually still giving me a 10x uplift. So the workforce is there, like you're actually making them more efficient. So human in the loop is absolutely not an issue in our mind. But I think like as you change the use cases and you start to do things which were just not simply possible, you know, for humans to do before, that is when the scale factor will become harder. Because yes, with AI you can do thousand times more work than you could do with your human workforce. But that is still, I think we're pretty early in the journey, like with enterprises, that I don't see that as being a bottleneck today.
Okay. And I guess I wonder about another potential cause of a bottleneck would be actual executive buy-in versus, say, technical hurdles or even workflow hurdles. Can you just give an example of a leader who is taking a concrete action to actually push something through, and if that's the case where it's maybe a human bottleneck rather than the work itself?
So, like as we were talking about, what are the roadblocks for AI in your enterprise? Security, you know, concerns, privacy concerns, often become a big factor in bringing AI at a big scale inside your company. And so we've seen this, like, you know, especially with very large enterprises, you know, that are very security sensitive, ultimately they will not allow, the security teams will not allow AI to have access to the right data or access to right actions capability in your enterprise systems. And sometimes, you know, somebody has to break that logjam. Like, you know, like a vendor like us, we're trying to get more things done, the security team, you know, is pushing against it, the business team wants to do more, and that's where we see sometimes the CEOs or the business leaders ultimately making that call, that decision, that we are going to do it, and now we have to actually go figure out, like, how to solve for security. So sometimes we see that happen, sometimes we see that doesn't happen. But that's an example where leadership actually plays a key role in driving AI adoption inside the company, because these challenges are going to keep coming. Like, you know, like today's security, tomorrow it may be privacy. We see that a lot, for example, in Europe. So what we've seen, like, you know, as effective strategy for enterprises is if you today, if you make AI a function of its own, have, for example, a chief AI officer who doesn't do other things and they're fully focused on only one objective, which is bringing AI across all of your different business functions inside your company. When you actually formalize a role like that, then we've seen that, you know, you start to make fundamentally more progress compared to when you tell, like, everybody in the company that, hey, look, everybody should embrace AI, you know, we will actually create funds for it, go and adopt and bring the right tools. If there's no centralization, then we see that oftentimes it lacks because people are busy doing, you know, the work that they're doing, and AI just becomes an additional responsibility for them. So that's the point-in-time strategy that works well. Create, you know, an AI function within your business, but that's also temporary because, you know, within a year or so, AI is going to become so ingrained in each of our business functions that we won't think of it as a separate thing. We won't think of AI roadmap as a separate roadmap to build. It'll get fully sort of absorbed into the core planning, you know, across all different business functions.
