Heat up. Here. Heat. Heat. Heat. Heat. Heat. Heat. Heat. Heat. Heat. Heat.
All right, we are getting ready to get started with Glean Go 2026. Please come grab a seat in the keynote area. Thank you.
Heat. Heat. Heat. Heat. Back. Hey. Heat. Heat. Heat. Heat. Heat. Heat. Heat. Heat.
Good morning, San Francisco. My name's Judith Bramwell and I lead our Glean business in EMEA. I'm delighted to be here today all the way from sunny London to kick off our annual Glean Go conference. Thank you so much for coming all the way here. And to all those joining us around the world on the live stream, welcome to Glean Go. It's actually also my wedding anniversary today. So, what better way than to celebrate it with you? Given this special day, I was wondering, is there a way I can make a link between a wedding and AI? So, here goes. Like AI, all good wedding speeches start with good context. Who you are, how you met, what's important to you. Like AI, the cost of a wedding will probably be more than you originally planned. Maybe the wedding was expensive, but what about the ROI? Surely, it's got to be worth it in the long run. Unlike AI, you typically only get one model choice at your wedding. Well, we certainly have a great congregation here today. I'm feeling a lot of love and a lot of energy in this room. It's going to be an amazing two days.
When we came together last year, enterprise AI was already moving quickly. 12 months on, quickly feels like a massive understatement. Frontier and open models now coexist, with frontier systems pushing the boundaries of capability while open alternatives broaden access, customization, and innovation. At the same time, the price performance curve is resetting as the cost of increasingly powerful AI continues to fall. The number of ways employees can access and use AI has multiplied. Agents are moving into day-to-day work, and experiences from what AI can do inside every company have changed enormously. But most important progress is happening inside your organizations. What if every employee could tap into the full context of their company right when they need it? That's what we see our customers doing each and every day. Putting that context at every employee's fingertips across every function. In the sales function, before a seller walks into a meeting to discuss a renewal with a customer, they know the promises made, the issues still open, and what could put that deal at risk. For developer teams, a developer opens a pull request. Glean shows the systems it touches, a similar outage from last year, and who needs to review it before the code ships. If you look at support, when they receive a major escalation, instead of searching multiple systems, the team sees the account history, the product issue, fixes already tried, and the engineer most likely to solve it. And for new employees, when they join, Glean instantly connects them to the people, tools, best practices, policies, and passwords they need to get up to speed. So work no longer moves through your organization as a series of handoffs, creating bottlenecks, losing context, and waiting for someone to connect the dots. Instead, knowledge moves ahead of the work. Every person across every function has the context of the whole company behind them so they can take the next best action to move the business forward.
And that progress is already represented in this room. Over the past year, you've pushed the boundaries of AI and put it to work against the real business problems. You have connected knowledge across your enterprise systems. You've turned years of institutional knowledge into usable intelligence. You've built agents for real workflows. And you have worked through the hard questions around security, governance, and adoption. You've not just simply embraced AI, you are helping define what it becomes. That is why Glean Go is not only about what we have built, it is a celebration of what we're building with it. Throughout today, you'll hear directly from some of the companies leading that change. From our customers at General Motors, DaVita, Ericsson, and eBay to our partners at OpenAI, NVIDIA, Snowflake, Cisco, Databricks, and more. They'll share what they're building, the results they're seeing, and what they've learned along the way. Over the next 90 minutes, we are going to show you where this goes next. You will see how AI is moving from simply answering questions to anticipating what needs attention and helping move work forward. How companies can get more value from every model and every token, and how an open secure platform gives organizations the freedom to keep using the best technology as it evolves. So to begin, please join me in welcoming Glean's founder and CEO, Arvind Jain.
Hello everyone. Thank you so much, all the 5,000 plus people who are joining us today for Glean Go from this room as well as online. And a special thank you to all of our customers who will be sharing over the next two days how they're using Glean to drive business value within their own organizations. General Motors, they're using Glean to help engineers design better cars faster. McCarthy is transforming construction management so teams can arrive at the job site ready with project plans and safety procedures on hand. DaVita is rethinking patient support, helping healthcare employees surface trusted information that improve patient outcomes. What you could do with AI a year ago doesn't resemble at all things that we can do today. But there's still a lot of work to do. Every enterprise that we talk to has three things on top of their mind today. They're struggling to bring the right context to their agents, which results in poor quality and makes it hard to realize business value. And of course, AI costs are running very, very high, absolutely more than what most enterprises can afford today. And while AI is solving some problems for us today, it's also introducing some new ones. Proliferation of AI-written documents, also known as slop, skills, and agents, which are all getting very, very hard to manage and secure. We'll spend the next two days talking about how Glean will help you tackle these problems and make AI work for your enterprise.
Let's talk about context first, since bringing enterprise know-how to agents isn't easy. Today, employees spend a lot of time coaching AI agents and are feeling overburdened with that. This is bot-sitting. We are spending too much time working for AI as opposed to AI working for us. This is not how it's supposed to be. And bot-sitting is just one negative consequence of shallow context. There are many more. It's also why token costs keep climbing. It's also why every new AI app that you bring in your enterprise feels like one more disconnected front door. To understand shallow versus deep context, we wanted you to hear from experts in the field. So we asked them what context means to them. So let's take a look at that.
Context is what transforms intelligence into relevance.
The most critical part of getting the right and precise output from the model.
The understanding of the tribal knowledge of all of the information across all of your systems that fundamentally makes up your business. Anything fed in alongside your prompts to help shape the response you receive. Context is what gives information meaning.
Many different things to each one of them. But for us at Glean, we recognize context as the collective human intelligence and the experience of your enterprise, packed in a form that AI agents can use to operate at the level of your human experts. As a pioneer in enterprise AI when we started in 2019, Glean has been thinking about the need for rich context in AI years before the industry caught on to this concept. We built the context foundation that now feeds enterprise AI across millions of users. Why is context so hard to build and where is the data to build it from? Structured data and data lakes capture only 5 to 10% of all enterprise knowledge, and documents like playbooks, SOPs go only as far. Much of the real know-how of how work happens lives in people's heads. But fortunately it does surface in emails, messages, meeting notes, exception approvals. All of that data is available somewhere. People are expressing it every day as they do their regular work. The challenge is in capturing those moments of human judgment and turning it into something useful.
So, Glean's context engine combines advanced retrieval with a powerful enterprise graph. Retrieval surfaces the most current high-quality information across your systems. And the enterprise graph maps the people, customers, teams, projects, and relationships between them that essentially define how work actually gets done in your enterprise. Glean builds these ontologies largely automatically, but we do give you the ability to visualize it, to work with it, and also to change it when you need to change it. That deep understanding of your business, how it operates, is what sets Glean apart from the industry's typical approach of delivering shallow context via MCP. Glean is the only context and intelligence company that truly understands your enterprise. I would like to talk a bit about other things that powerful context can enable AI to do. It can make AI proactive as it understands you, your role, your work, things that you need to get done today, and Glean can do work for you before you ask for it. It also allows AI to become a true teammate as opposed to an agent that does a very specific task when you're commanded to do that. Glean's long-running independent agents use rich context to learn how to become as good as an escalation engineer or a customer support agent or a technical account manager and deliver business value at a completely different level. Think of agents now, these long-running independent agents, as real human beings who are operating at the level of that expertise and that experience that everybody in the company has learned over the years.
So, let's switch topics now. I want to talk a little bit about AI sprawl that we have all been experiencing. Most businesses I speak with have ended up with multiple AI front doors and agent platforms that are all connected partially and inconsistently with your company's data. There's also a proliferation of content, skills, and agents as AI has made it so easy to produce them. Glean's AI gateway makes your AI tools think alike and work consistently. And how do you bring the right content, skills, and sub-agents to your agents? That's where our core search expertise comes into play. Glean identifies the highest quality content, skills, memories based on human and successful agent usage and solves that problem. Search is more important in AI than it has ever been.
This year, Glean also evolved into becoming an enterprise AI platform. In fact, 78% of our customers now access Glean through APIs and MCP. That means we are making Claude Code, Microsoft Copilot, ChatGPT, all of these tools better with Glean. AI front doors, even our competitors, they're all getting better because of Glean. While we're talking about Glean's AI gateway, I want to highlight two more capabilities that it has. First is our memory platform. Glean learns about your employees, how they work with AI, their working preferences by looking at their activity across all of your enterprise systems. It also learns how agents compound their intelligence over time with that human feedback. This memory is a key asset and it is available to you as a platform that you can leverage in all of your AI tools and agents even as your AI tooling actually changes. This is AI sovereignty. You being in control of your own AI. While we're talking about that, Glean also brings all closed and open-source models via this AI gateway so that you can choose the models that meet your data retention and other sovereignty requirements.
Well, there's one more important issue that we need to address today, probably the biggest one, which is token costs. Solving for soaring costs starts first with model choice, of course. Our goal is to bring the best of all AI innovation happening in the industry to our customers but at the cost that they can afford. To do this, Glean aggregates intelligence across all open and closed-source models. Most systems still point the most expensive models at the simplest of tasks. That's not really a limitation of AI. It's an architecture choice. Glean doesn't want you to be spending tokens where you don't need to spend them. But choosing the right model for a given task is not an easy problem. It requires you to fundamentally understand how models perform across the huge range of enterprise work. Glean has built a sophisticated evaluation and training framework on real-world enterprise data with all the customers that we've been working with, which has helped us build a really effective model routing technology which we use every day in our products today.
We released a benchmark today showing that Glean saves 81% on token costs compared to Claude Code, with Glean responses preferred 78% of the time. Context produces higher quality results and it delivers token efficiency. That's because quality context stops AI from having to reconstruct it at runtime, which burns tokens and hands you back answers you can't use. Now to a very important topic. 2026 is the year of open source in AI. And I'm not just talking about models, which as we discussed are allowing us to optimize AI costs. They're really getting as good as closed models. I'm also talking about harnesses. Open-source harnesses often beat closed ones in benchmarks. And they're allowing us to create amazing AI products very quickly. And one of these products is Glean Tow, a new desktop app that does everything from code generation to general co-working. It's named after one of its creators. We took an open-source coding harness, combined it with Glean's context platform and AI gateway, and built this incredible product which now the majority of our company internally is already using as a primary code generation and co-working product. Alan is going to talk about Glean Tow more later. To conclude, Glean is an enterprise AI platform that truly understands how work gets done. Now, let me invite Amrajan, our head of product, to show you the products that we're building on top of this context foundation. Thank you.
Welcome everyone. I'm John and I'm thrilled to be here with you all today to cover what we have been building at Glean. Let's get started, shall we? AI has made our lives easier. It has made such that we all can accomplish a lot more work from our laptops. As product builders, we are able to ship more features faster. But our human attention hasn't really changed. We all have a million tasks on our plate, a rapidly growing number of AI features, AI tools available to us. And yet, in our moment of need, when we actually need AI to help, we are left not knowing what feature to turn to and what job it solves. So if human attention is now the bottleneck, then we have to meet the user the moment they need it and help them get the most value out of AI. This is AI that works unprompted.
Today I'm excited to show you what the future of work looks like with AI that is proactive. We are announcing proactive AI that first prioritizes and executes work with task management, email triage, meeting coach. Second, informs business decisions with interactive dashboards. Third, coordinates and executes work before you even ask with independent agents. Fourth, improves over time with memory and skills that learn and self-heal continuously. And we are moving AI from single player to multiplayer. The team chat product, a shared space for teams to build and collaborate with AI. Let me invite Selene so that we can show you some of what we have been building.
In this demo, we will show how a fictional HR SaaS company might use Glean to surface renewal risks, manage customer escalations, and bring teams, AI, and agents together to resolve critical business issues. Interactive dashboards are our newest artifact. It provides a dynamic, sharable view combining business data and insights, letting users drill into what matters. Let's begin in Glean. I'm going to open up an interactive dashboard. This one is focused on renewal health, providing a live view into all of our renewals. Before Glean, I would be limited to only seeing CRM data. Now we combine structured data from our CRM, usage telemetry from our warehouse, combine it with unstructured customer data, all in one view. What we see here is that Frasier Automotive's renewal is red, putting revenue at risk. Glean proactively identified this risk and shows that they have been impacted by an API failure right during the performance reviews. Now, we need to make sure our teams are aware and act quickly to save this renewal.
Team chat is that cool new feature that lets teammates and Glean proactively work together. So, collaboration builds on shared context. Instead of copying it between tools, in a few clicks, I add the AE, I add the CSM into a shared thread. I tag in the AE to understand what happened. The AE asks Glean to find out more and asks for help to draft a communication plan. Pay attention here. This might appear simple but represents a huge unlock for your business. To date, AI has been predominantly single player. One person going back and forth with AI alone. Team chat is what multiplayer AI looks like. Teammates working together with AI to drive better customer outcomes. The CSM asks Glean to update the next steps in our interactive dashboard to keep everyone aligned on what's happening without any manual work. So we saw how Glean is providing a living view, living understanding into which renewals are at risk and we have a strategy to save this renewal. Now let's see how our engineering team is collaborating with an independent agent to fix the API failure directly in Slack.
Independent agents are long-running agents with their own distinct identity, memory, and context. They can be trusted to act proactively and they can coordinate work across teams. But they also have governance in place so that they don't go off the rails. Let's see this in action. In Slack, Glean's independent on-call agent is working to resolve that very same critical API failure we looked at alongside the engineering team. The agent is proactively identifying the escalation, analyzing the issue, looking through the code, opening a pull request, and tagging the right engineers to fix the issue collaboratively. Glean's superior enterprise context is what makes this possible. This agent learns how to proactively fix API issues because of its memory. So, it can use that deep knowledge to resolve future issues even faster. This is AI not doing everything but making sure everything gets done. Now that we have a fix from engineering, let's see how we can now communicate this back to our customer.
Proactive task management. It uses enterprise context to surface what needs your attention. It prioritizes it and prepares the next step before you even ask. Let's go back to Glean assistant. It looks like Glean has prepared something for me here. Once I click on this task, Glean has already built the first draft of the slide deck for our upcoming customer meeting. It is even bringing that context from engineering about the root cause and the resolution of the API issue, the resolution we just shipped. Again, this feels really easy because it is. This is how context unlocks better productivity for the enterprise, faster, higher quality execution, and definitely not bot-sitting.
What Selene and I just walked you through is how Glean delivers proactive AI. Interactive dashboards that help leaders get a full view of the business and coordinate work across all levels of the enterprise. Independent agents coordinate work across teams to resolve issues faster. Proactive task management gets work done on your behalf. And team chat shows what is possible with multiplayer AI. These new capabilities are unlocked all with Glean's powerful enterprise context. Now I want to pivot a bit and show you how that very same context can identify which processes to transform. Our forward-deployed engineers work closely with our customers on their AI transformation journeys and strategies, and our customers tell us the same thing. Enterprises need a reliable, real-time view of how work actually gets done in their environment. Transformation starts with that visibility, knowing what actually needs transforming. And to date, that visibility has been rather fragmented, still leaving immense opportunities completely untouched by AI.
With the enterprise graph, Glean already has that visibility and we now can identify which workflows are the strongest candidates for that transformation. Today we are announcing Glean Transform. This is a living representation of how work gets done in a company. Let's take a deeper look. We'll show how we bring AI transformation to our teams at Glean. Here, Glean has already transformed 78 workflows and highlighted 272 opportunities across all departments. But let's actually dig into the sales department. In Glean Transform, you can see there are 38 opportunities ripe for transformation just within our sales team. One click, Glean highlights how our sales team gets work done along with a prioritized list of transformation opportunities. Let's dig into one of them: our discovery call and use case mapping process. Here we can see specific recommendations for each step of the workflow mined from the enterprise graph. In the account research step, you can see that our sales development reps are constantly context switching between different tools like Gmail, Google Drive, and it is slowing down progress. Glean Transform understands this work at a granular level. So we can see how teams gather account information and the shared context and why this is worth automating.
Transform now has built a blueprint for a call research agent. In one click I can see that agent that we have built. Now this agent includes all of the correct instructions, tools, knowledge, other agent config details so that we can set up that automation with one click. And because Glean Transform understands the way that work really gets done in my environment, I have confidence that the agent will fit into my process and will make a real impact. But it is one thing to feel like an agent is having an impact. It is another thing to prove you have made a difference for your business. If we go back to the main Transform dashboard now, you all can see how Glean Transform measures AI impact across our engineering workflows. Our customer escalation workflow has become 46% faster. For example, one click deeper and you can now see how that happened, why that exactly happened. With the new diagnostic agent that we deployed, you can see that we are resolving escalations faster, assigning tickets more accurately, saving 40 minutes per ticket. So, Transform enables us to see all of that impact in one place. Transformation is not a one-time initiative. It's a constant evolution, opening up new possibilities to reimagine how work gets done in your enterprise. Now let me turn it over to Amar and Phil to talk about how GM is transforming their business with Glean. Thank you.
Hello everyone. I'm Amar Malitira. I'm the Chief Operating Officer at Glean, and today with me is Phil Lutka, Executive Director for Core IT at General Motors. Thank you, Phil, and thanks for being here. GM is driving an incredible transformation with AI. So let's start with a big picture strategic question here. What strategic problem was General Motors trying to solve for when you decided to invest in Glean? And why did Glean become a platform of choice for that challenge?
It's a good question to start with. First, I'm just so thankful to be here and just in this incredible energy. But for GM, the problem we were really trying to solve was an access issue. We needed to have a platform that could help us find all of this vast knowledge that exists within our enterprise corpus. GM's a big company, thousands of employees, tens of thousands of employees, tons of business units, just so much knowledge, 100 years of history, and finding the relevant important information, it's hard. And Glean has sort of bulldozed that challenge to the ground and made it very easy to access. The second piece was we needed a product that we could grow with and learn from, and one that we felt we would get a good partnership from. And I think we've seen that for the last eight months now with Glean, as we're building together, we're learning together, and we're delivering just incredible value to the business. So those are the core areas that were really part of that initial decision. And I think the last sort of key piece was we wanted a great user experience. This needed to feel clean, modern, easy. And that certainly is what we found with Glean.
That is fantastic. So talking about user experience, one of the things I noticed was the incredible pace at which you guys drove adoption. They went from zero to 60,000 users in just 90 days. That was just amazing. There's a lot of lessons to be learned on how you guys drove it. So how did GM really plan for, validate, and roll it out so that every single employee could use it, as opposed to what we generally see in most companies, which is the technical teams, the early adopters using it? Give us some thoughts on it. How did you do it?
I mean, first, I'm just really proud of this accomplishment. This is, of all of the IT projects that I've done in my career, Glean has been the highlight, the most impactful, iconic, and successful, which is a pretty big deal. It took us a lot of work. This was not like a 'we'll just turn it on and make it easy for everyone.' Obviously, we started with a product that just was easy to use, but we came at it with our customer and our end user in mind. So we really were curious. We wanted to understand the problems they were trying to solve. What's the friction that exists in everyone's daily life? How do we help people find the person that knows how our battery chemistry works or a 10-speed transmission works? How do we help them find a relevant, current document? And all of those sort of scenarios. And we also wanted to bring it to folks in the sense of we had a lot of employees, thousands, that had really not used AI before. And we had another huge population that had and was ready to get cooking, building agents and all that fun stuff. And so we really met people where they were in that journey and just spent a tremendous amount of energy and calories on enablement, training, joining all-hands meetings, staff meetings, and just again leading with curiosity: what problems can we help you solve? How can we help you get comfortable through that? On top of just the reality that Glean is super easy to use, we had it integrated into all of our systems. We made it so that that first experience on day one through day 90 was great. And I'll say too, when we talk about 60,000 users adopted, we're pretty serious about that. We sort of look at the typical metrics, but 5D30 is where we anchor. And I looked this morning, we're still at over 80% 5D30 for our whole organization. So it's real adoption, sustained continued use throughout the day, the month, and the week.
That's fantastic. So one of the questions every single enterprise around the world is grappling with is how do you measure the impact of AI. And so how are you guys doing it at General Motors? And more importantly, if you could give us one Glean use case that had material impact on your business.
There are a lot of use cases so it'll be tough to pick, but we do the baseline measurements. Obviously, adoption we talked about. We look at the impact that it's having, ease of use, is it finding the right relevant information, sort of all the great tools that Glean provides us, we leverage. And then on top of that, we really focus on impact, and not just did we save a bunch of time, but did this really enable the next great idea, a faster decision, or just a better outcome altogether for the business. Nquille in room 11 at 11 o'clock. You should join. It's going to be a great session. He's going to talk about our DR agent, which is the design release engineer agent that we built in partnership with Glean's forward-deployed engineering team. It effectively takes this massive amount of issue, warranty, telemetry, engineering data, all of this data across just a ton of disparate systems, and it brings it into one clean user experience, and it helps our 6,000-ish design release engineers bring all of that context together, solve problems quicker, get to customer outcomes faster, and really just deliver good value and improve their overall enjoyment with the work they're doing. They're spending less time swiveling trying to find information and actually shipping good quality outcomes.
Fantastic. Thank you very much for that. And looking ahead, given what you guys have already done so far, how is GM thinking about continuing to innovate and stay ahead of the curve?
Gosh, I mean, everything that we've been seeing here is a big part of that. I think transformation is the big thing. We don't want to just bolt AI on to processes that maybe aren't great or maybe we don't need anymore. And so I think that's really the mindset we're driving across all of GM: look at this problem that you have or this way we work and assume it can be completely different. And so really that transformation is a big part of it. Obviously, continuing to focus on the outcomes. Arvind talked about token use. We want to do it efficiently, but we also want to make it easy to experiment and safe. So those are really the key areas that we're looking at and just really learning together with you. And I think again, that's the partnership that we appreciate so much and why we select vendors the way we do: we want to be able to build together and deliver those cool experiences.
Phil, we are so grateful for your partnership. Glean is going to be with you in your transformation journey. We're going to learn a lot. So thank you for being here with us and for sharing your perspective with the audience here. Appreciate it. Well done.
All right. Good morning. Hi, everyone. This will be a fun panel. We have about 12 minutes, so we're going to do some speed questions. I'm Kate Rooney. I'm a tech correspondent at CNBC covering AI, which has been not busy at all. It's been a very chill couple months. Nothing going on. But thank you so much for having us, Arvind. And I want these guys to introduce themselves. I was just saying Alexander's long resume at OpenAI running Codex. But let's go around the horn. Everybody probably knows Arvind, but Alexander, tell us a little bit about your role at OpenAI, what your day job looks like.
Hey folks, great to be here. Thanks for having me. For the past year or so, I was leading a lot of our agent work with Codex, and now I lead enterprise products for OpenAI.
Terrific. A tiny startup you may have heard of, OpenAI. Arvind, I think everybody knows Arvind, but thank you for having us. We'll kick off with you. One of the big topics we want to talk about is agents. We were talking a little bit about this backstage, and this move from reactive to proactive. And I think a lot of people, myself included, want to live in a world where AI is a little more proactive and can do a little bit more for us. It seems like there's a gap still. I wonder what you hear from your partners, from customers, and what needs to change in order to bridge the gap, make AI a little bit more proactive and a little bit more useful, I think, for the average company, the average executive, and just the average Joe like me who wants it to do more.
Yeah, I mean, this is absolutely, we believed AI can make true impact when we don't have to go and direct it all the time and give it tasks that it can do for us. Because what we've seen, what we've noticed, even at the most advanced technology companies in the world, people are barely using AI to its full potential because they just simply don't know what AI is capable of doing for them. So we always thought about this: why is it like that? Why can't AI actually come to you and figure out who you are and what you need to get done and actually just start to propose and actually do those things for you? That would sort of completely change the dynamics of adoption, and you will see a lot more AI sort of helping everybody and the adoption becoming more uniform inside your enterprise. But that's the trick though. What's missing is to make AI proactive. Think about personal co-working. If AI knew everything about me as an individual, what my role is, what I do in the company, what are my OKRs for the next three months, what are the weekly tasks I need to get done, and it also kind of knows how I like to do that work, how I write, how I read, with that full context, it can then actually come and figure out that these are the priorities for me, these are the things I need to get done, and actually propose and get those things done for me. So that's what it needs. It needs that deep context, deep understanding of your people and your business, to actually become proactive.
Alexander, I wonder what it looks like at your day job. You have conversations, I'm assuming, with Fortune 500 companies. OpenAI works with some of the biggest companies in the world. What are they saying about adoption, about AI agents, and this conversation about maybe making them a little bit more proactive?
Yeah, so I actually kind of want to start with the risky move, which is polling some people here. But you know, I totally agree with what you were saying about how AI just requires so much thoughtfulness. I like to joke that it's sort of a little bit bottlenecked by our own creativity and even typing speed. So raise your hand if you think here you are running enough loops in parallel in the background. Okay. Raise your hand if you feel slightly bad about that and you're like, 'Oh, I should be running more loops.' Yeah. Okay. Okay. Thank you for reacting, team. Yeah. So with the right context, agents can be much more powerful. And then the thing I would add to that is the form factor is really meaningful, right? Like right now we've built a lot of tooling, and tooling is great. We can achieve mastery over tooling and we can make it awesome for ourselves. But that puts a lot of onus on us. And so a lot of what I hear from enterprise leaders when I'm talking to them about AI is they want to know, hey, how can I reduce the amount of change management required for AI to be helpful for me? How can I make it so that I don't have to train every single person in my company how to prompt the right way, how to manually inject the right context, and how to even know that AI could help for a task? So we think a lot about, okay, how do you get the right context by default? We think about what is the right form factor. You know, you use a coding agent differently than you use a product that's shaped like automations, which you use even more differently than you talk to a teammate, right? Where it's really natural. If we were colleagues, I could delegate to you or you might proactively suggest an idea to me. So, this is kind of the journey we're trying to take people along. Starting from, okay, we all know how to use ChatGPT, we know how to use Codex, we know how to use Glean just like search manually. How do we then build products such that people at a company can sort of by default get value from deeper AI workflows?
One of the things you mentioned, Arvind, was sort of the context. I think about security is such a big issue and that tradeoff between, at least as an average consumer, yes, you might want to give AI more information so that it can make better decisions. But you're also worried about personal security, cybersecurity for companies. How do you think about that entire topic and keeping companies' data safe?
Yeah. Well, so first, I think security has so many different areas where we need to actually do a lot of work in. When you think about agents, agents can run either on behalf of me or they can have their own identity. If they're running on behalf of me, these agents should not be given access to enterprise data and information that I did not have rights to see or use, and on the same token, they should not have the ability to make any changes into the system that I could not make. So that's the first thing that you have to think about as you build your agent infrastructure, is making sure that you're properly restricting your agents to things that they should be able to do. But also, when an agent is running on my behalf, I actually have a lot of powers as an employee. I can go and change things in many different systems, but an agent has probably been built with a specific set of tasks that it's supposed to do. So you have to further go and restrict these agents. You have to think about the range and the scope of tasks that it's supposed to do and accordingly restrict their permissions to those kind of actions. So these are some of the basic things that you have to actually do to make sure that agents suddenly don't start to actually create problems inside your enterprise. But then the new thing, which I'm actually really excited about, is this concept of independent agents. So these are agents that are not running on my behalf. Treat them more like employees or colleagues of yours. Say, for example, take your customer support team: maybe you have 10 people in the team today and now this agent is the 11th member of that team. So when it is like that, when the agents have their own identity, you have to again think about what is the right permissions model, the security model for these agents, what information can they see, what they can't, and what's again the guardrails in which they should operate so that as a team member they know they're nice to people, they're actually adding positive value and not actually escalating and fighting with the humans inside. So these are interesting topics that now we have to actually solve for as agents become more and more independent.
It seems like a lot needs to happen behind the scenes to take out some of the friction. If you need to have certain permissions, I would think that's probably one of the challenges too. I wonder how you see it at OpenAI. And also if you want to respond to any of Arvind's commentary on it.
No, I agree with that. I think we're about to see the rise of, maybe two big shifts are going to happen, right? Independent agents and agent execution shifting to the cloud off our devices, which will be way more secure, and a lot of our enterprise customers are very excited about. I do have to pause for a beat though and say if we're talking about security and cyber, a lot of our customer conversations today start with cyber. One of the things they're most interested in is just how to prepare for the new cyber landscape. And we obviously have great cyber models, but coming back to running agents in enterprise, we kind of see there's a few layers and tackling them sort of incrementally seems to be one of the best ways to go about it. So one of the layers is just control: how do you decide what agents can do, what they cannot do. And interestingly enough, this is somewhat easy by default because over time you want to connect the agents to all your context so they can actually be incredibly useful, but when you start, they're not connected to anything, right? And so we've actually seen that with a lot of our customers, just starting with the most important pieces of context or tools to connect them to and incrementally building your way out. It might not be exactly what your CEO or the average employee wants, but your security team will be very happy if you add one and then you add the other one. You can actually move really fast if you do it incrementally. So okay, so you have to lock down what configuration, what permissions the agent has. The next side of it is actually observability, right? You can be much more confident in giving your agents access to things if you have really good understanding of what they're doing. And any manager or IT admin can go in and say, okay, here are the different actions this agent took. I like that. That's great. Let me actually optimize that or let me lock this thing down so it can't happen in the future. So basically pairing these two ideas, or maybe three. One is the control of what the agent could do. The second is the observability of that. And then lastly, starting to shift away from agents always impersonating us as people and having more agents that have a specific role with exactly the right permission set for that workflow, and that's an independent agent.
So is that like an agent has an avatar and is its own employee? What does that actually look like?
This is a fun one. I'd be curious what you think. We talk a lot about embody, not embody. I think the answer is both, right? Like some things you just want an automation. Like every time someone writes code, we want a code review to happen, you know, that doesn't need to be an embodied thing. How do you think about it, Arvind, too?
Well, I think with independent agents, I like the concept of them sort of having a human-like identity, like an avatar and a name. Because when you do those things, of course they're gimmicky in some way, but when you do it that way, that's when people will start to truly treat them like a real teammate of theirs and it'll unlock different types of collaboration with agents.
I wonder, we're running out of time, but big picture implications for companies or investors in the audience, what are the economic implications? What are the implications for productivity, profitability for these companies if there is actually a world where companies can harness this well and sort...
to close that gap that we talked about in the beginning. Arvind, maybe you want to take this, and Alex.
Well, AI of course promises significant productivity, and fundamentally, in the long run, we're going to see work change. We're not going to be doing most of the things that we do today anymore. We're going to have these amazing agents and assistants and co-workers that actually help us do all of this work. So the economic impact is very hard to actually even measure — it's so huge, so big. We just have to make sure that we are getting there in a way that we actually can. Today, unfortunately, what's happening is AI is often becoming more expensive than the business value that it is delivering in the short order. So that's the challenge that we have to solve. Long term, there's no confusion — this technology is going to be great, it's going to create a lot of economic value. We need to make sure companies don't run out of money before that happens.
We didn't even talk about token maxing. I can't believe it didn't come up.
Well, last word, Alexander. How do you think about this, and how should companies and some of Glean's partners, investors maybe even who might be in the audience, think about the implications and how to sort of prepare their businesses?
I think individuals and teams are able now to be vastly more productive, and that is only going to increase. But there is now going to be — and by the way, this is not a new thing. I think since the history of our species, we've created tools, and tools always mean that someone who is good at using tools is vastly more productive than someone who is not. So we've seen this problem before, but we are now going to enter a situation where the most productive individuals could be running unicorn companies, and folks who are less productive are where they are today. And so I think the most important implication and takeaway is just to have really high agency. Make sure you're hiring really high agency people and think about how to structure your team to unlock the creativity of those high agency people and then scale that to the rest of the company.
Fascinating. We're past the shot clock. This flew by. But thank you all for listening and tuning in. Arvind, thanks for having us. Alexander, thanks for being here. I'm going to welcome Tony up next. Let's hear it for Tony. Thank you. Thanks, guys.
Good morning. I'll start with something structural about Glean. Our business model gets better when your tokens do more work, not when you burn more of them. A lot of vendors are built the other way around. Arvind showed you a few minutes ago what happens when proactive takes on more of the work. But that only pencils out if you're spending AI on tasks that actually need it, matched to the right level of intelligence. Tokens are money. With unlimited tokens, we could all do millions of things, but everything has a cost. So our goal is simple: drive down the cost of AI doing work, and then dramatically more things become worthwhile doing with AI. Token yield and what I'll talk about today comes down to three levers: model choice, auto routing, and AI usage controls. Now, I'm going to invite one of the smartest engineers I know, Eddie, to the stage to talk to me about our evals and our latest benchmark results.
Hey, Tony. So yeah, let's start with the first one, model choice. You really can't optimize anything with token spend without having a real model choice. That's why Glean gives you over 40 open and frontier models and why we work so hard to make sure literally every other week as they drop that we're getting that day zero support.
Yeah, definitely. Speaking of models, this is the year that we think open models really caught up. Mostly because we stopped measuring them on maximum intelligence and instead measuring them on cost performance, and on that curve, open models have closed the gap.
Yeah, you said curve. I want to take a look at this. This is called the Pareto frontier — not potato frontier, I thought that the first time I heard it. Pareto frontier. So what is this? This is a model that shows us price performance. A model is Pareto optimal if nothing else beats it on performance at the same cost or matches its performance for less. So we see cheaper, less capable models in the lower left quadrant, and we see the more expensive, more capable models in the right quadrant. And when we talk about the frontier, that's that curve that traces the best option at every cost point in between them.
Yeah. So this chart you guys are looking at, we built on enterprise data capturing the variety of tasks and different work across departments that are in the enterprise today. So it's important to note, these are our results on our benchmark. They involve our query set composition. They do convey these models across the frontier, but they're conditioned on the choice of harness as well as the evals in the environment. So if you guys look closely, in this setting, GPT-5.6 Luna does dominate a big part of the frontier given its incredible bang for buck. In the middle, we see an appearance from open models like Kimi K3, GLM-52, and actually as of this morning — we haven't recreated this plot yet — but that Ox Alpha model from the weekend is known to be GLM-53 Flash. Now, I have a strong suspicion it's going to be at the frontier here as well. And then on the top right there, we have Opus 5 giving us the strongest intelligence for those who are willing to pay for it. And if you did go back about a month, the graph would look pretty different, right? Luna would be at a different cost point, so it might not be shaped exactly like this because OpenAI hadn't done that awesome big price cut. Kimi K3 didn't exist yet, so it hadn't shipped, and those two wouldn't be these sort of intelligent, affordable options.
Yeah, that's really the thing. I mean, we just made this slide, right? And it's already obsolete. The frontier doesn't hold still. And if you're locked into one provider, if you're looking at one ecosystem, you're not taking advantage of every cost savings or every intelligence opportunity that comes up. You simply cannot bet on one provider here.
Yeah. So in our data, the same pattern can kind of play out task by task. So over here we have information-seeking tasks, and you can see for those, a lot of them get served by pretty inexpensive models. And you can see that in the sort of slope of the chart being pretty flat. In contrast, for coding, you see the opposite — the slope is a little bit steeper, and you can see that performance jump as you move up to those frontier open models as well as Opus 5. And this is sort of in contrast to that overall curve. So this kind of analysis makes the case for auto routing, not just by tracing that overall Pareto curve, but also task by task.
Yeah. And we don't just run all these eval benchmarks so we can put a pretty slide up on screen. These evals are what we use to guide our auto routing decisions, which brings us to that second big lever that I talked about. So today we're announcing auto routing in Glean, and it played a key role in Glean saving 81% on token costs versus Claude Co-Work, a benchmark that we also released this morning.
Thank you. Yeah. Let me put that 81% in perspective here. So that means we're talking about 58 cents per task with Glean versus $2.98 with Claude Co-Work on a similar task.
Yeah. So Glean's advantage comes from routing across those 40-plus available models, picking the right model at the right reasoning level for the given task. So for example, Opus 5 we know excels at visual tasks and visual artifact creation. So for those kinds of tasks, we route to it there. But it's not just choosing the model, right? It's also deciding how hard the model should think. So for simpler tasks, we might use 5.6 Luna but on low reasoning rather than high reasoning. And to make that all happen at runtime, we've post-trained our own specialist model, Glean Waldo, that's post-trained on strong open-source base models such as the Neotron family, and that model's capable of routing to the right reasoning level on demand.
Yeah. And you know, as an engineer, as a product builder, the part I really love is we didn't win that benchmark on just the price per token alone. We actually use 70% fewer tokens. Can you tell us why?
Yeah, totally. All these things kind of compound for token efficiency, right? So it comes back to our roots in search. With search, we have a very purposeful choice to precompute a central index, build this enterprise graph that has all these signals, and that way the model can much more efficiently find that relevant context at runtime instead of sort of building it up incrementally, wasting a lot of loops collecting it. Claude Co-Work and other harnesses, other setups might take a different approach when they rely on federated MCPs. Those services will end up searching live across a lot of sources, and all those different sources have different search qualities. They're not reconciled. All that data then needs to be deduped, filtered, joined, all while your tokens are burning, right? So that's a pretty token-intensive operation, and it shows up in this token difference.
