BoxWorks 2026 Keynote: The keys to a successful AI transformation

Please welcome CEO and co-founder of Box, Aaron Levie.

Okay. All right. How's everybody doing this morning? Okay, very, very, uh, we're very grateful to have everybody here today. We're, um, uh, we, we've been working super hard on some brand new capabilities that we're excited to, to share and, and show off this morning. Um, obviously, uh, we're at the center of one of the most transformational shifts that I think we've all seen in the enterprise and in IT. So we wanted to share a little bit about what we're, what we're building to help enterprises, uh, to really be able to take advantage of all of their data and unstructured information in this new era of AI. So, um, so, you know, we started Box, uh, with really a fundamental mission, which is to power how the world works together, and that mission has been able to impact so many organizations all around the world. Companies like Samsung Semiconductor, transforming, um, how obviously the future of chips and memory and so many other core fundamental capabilities are built out. Uh, organizations like Morgan Stanley and leading financial institutions on how they work with their clients and how they transform, uh, private wealth management and, and, uh, and, and the investment, investing ecosystem.

Uh, really proud to be able to work with incredible healthcare and research organizations, transforming medical experiences and research breakthroughs that affect all of our lives. Um, and, uh, a little bit of a shout out. I know we have some folks from Japan. Anybody from Japan here? >> Okay, good. Thank you for, for the, the trip. We are very excited to work with, um, with some incredible, uh, leading, uh, innovators like Toyota and many more, um, that are transforming obviously our daily lives, um, in the automotive and, um, and technology spaces. And what all these companies have in common is, uh, their businesses fundamentally run on unstructured data and the information, uh, that, that we all work with every single day. And the way that we store and secure and protect and work with that data, uh, is incredibly important to these organizations. And we know that we're at the center of a very big shift around what we're going to be doing with this information at scale, uh, as a result of, of AI. And AI is transforming everything about how we work. Um, you know, in all of our daily lives at this point, we are seeing the impact of what AI is going to do in our personal lives and in our workplaces and in our workflows. Um, and we, we sort of feel this in census, I think, in a very visceral way. But what's interesting is this is only going to be accelerating from here. So it's something we have to all collectively prepare for around what's going to be happening next and what's going to continue to, to happen as, uh, as the world transforms with, with this new technology. And we're seeing, and, you know, we have many folks from, from Silicon Valley here, and we're going to see, um, uh, Jensen and, uh, in just a little bit this morning around what, what the state of AI progress looks like. But AI progress is, is only accelerating. It's, it's, uh, uh, you know, incredible the speed at which this technology is advancing. Um, what, what's amazing is if you kind of think about, you know, where we were just a couple of years ago, it already felt exponential in those early days. Um, that jump, you know, when we first saw ChatGPT and then we saw, you know, some of the latest models of GPT-4 and their early versions of, of Claude that were rolling out, that already felt exponential. When you zoom out and you look at the progress since then, just again in the past maybe one year or so, uh, it's, the, the climb is even steeper. Uh, so the models are being trained on far more data. They're being trained with far more compute. They're being trained on, on greater and greater levels of capability in almost every single domain. They're getting better with tool use. They have better reasoning capabilities. They can now use computers, whether we're sort of seeing them literally use the computer or that's happening in a sandbox. They, they're, they're able to use all of the tools and code, uh, that, uh, that we've been using for, for, for decades. And so these AI systems are getting better and better at doing work across a wide variety of domains. And what's incredible is, is given how steep this curve already is, we have new advancements on the horizon. There's this thing called recursive self-improvement. And the rough idea is, well, as AI gets better and better at being able to do engineering and being able to do research and being able to do reasoning and do cross-domain analysis, then actually the, you know, AI researchers will use AI to accelerate the development and training of new AI models, which will then sort of recursively self-improve because that model will get better and smarter, which helps the next model that is being researched become even smarter, and that sort of feeds itself at a much faster rate than what we were dealing with before. So this is only accelerating. So if we thought the curve was steep already and the impact was already relatively, uh, significant, this is just going to accelerate. And so the thing that we think about at Box, and I think, you know, many of us in this room are, are thinking about in our workflows, is, you know, what if we had, uh, an assistant or a digital worker that, that was an expert in legal or an expert in engineering or an expert in manufacturing or an expert in finance, and they could work a thousand times faster than us? And, and what would that mean to our business processes? What would that mean to the daily workflows that we have inside of our organizations? Um, and we've, we've already seen the sort of rough sketch of that. We saw it with the initial sort of era of maybe the, the chat experiences that we've had. So AI is already transforming, you know, a lot of our daily productivity. Um, we're getting maybe 10, 20, 30, 40, 50% productivity gains by being able to ask an AI system a question and, uh, it can reason on that question using the, the expertise of the underlying model, or we could ground it with existing data and get information back. This is already transformational, and it would already be meaningfully impactful to how we work every day if this was just where we stopped. But it's not where we're going to stop because what we're doing now and what we're all kind of seeing is this idea of what if I took that same, uh, AI model, that same agent experience, and instead of just requiring a user to sort of prompt it and get an answer back, what if I deployed that agent and, and that model inside of a workflow or a business process? What if I fanned out the agent across multiple parts of that process? What if it sort of replicated and it could do a lot more work in parallel for me? Um, and that's really the kind of new era that we're about to enter with AI. We're moving from just chatting with an AI system sort of one by one and sort of working in a kind of synchronous fashion with AI to a world where we're going to deploy AI inside of our workflows and our business processes, and AI sort of running behind the scenes, or again, it can fan out and parallelize across a workflow that we want to do. And so now instead of just sort of single user productivity that, that is again very meaningful, we can now sort of deploy AI across much broader and much harder business problems. And so not just, you know, asking one question about a contract, but asking about the business risk of a deal or, or inside of our organization. And so instead of the AI just reading one document or finding one, one piece of information and reviewing it for you, it goes out and it sort of fans out and looks at all the relevant data and finds all the relevant documents and starts to write code for solving a particular task or connects to our CRM or ERP system and reviews data and then triages that and combines it. Uh, this is sort of, you know, this idea of we now have swarms of agents or fleets of agents that can go off and execute work for us, and those will be deployed by the end user. The end user might ask a question and a fleet of agents will go out to try and solve it, or maybe it'll be deploy, deployed inside of a workflow and inside of a background process. But the significant impact of this is we now have a decoupling of our sort of prompting of the agent to get an answer back and our ability to deploy agents at scale on any kind of business problem we're trying to solve. So this means that an agent could go out, or many agents could go out, and review all the literature in a particular research area in life sciences, and it could look at lab data and it can look at patient information to accelerate drug discovery. Or, uh, if we're working on an M&A deal or a private equity deal, and you want agents to be able to look at all of the, the due diligence data and all the market information and all the investment prospectuses, all of a sudden we can now scan a way greater set of information for what is the exact right investment, uh, profile we want to be able to make and what the exact deal we want to execute, because agents can, uh, parallelize that and, and be in way more places than we can at once. Um, or again, this idea of agents running in the background and, and automating our workflows. Agents can dramatically accelerate many steps inside of processes because they can review documents much more quickly or review data, uh, inside of an organization in any step in a process. So we could streamline, uh, insurance claims or client onboarding or loan processing, agents running in the background in these workflows that are executing more and more steps, um, inside of our processes and puts really us as people inside of the more impactful and more meaningful parts of those workflows. So this is what it means when we start to scale out agents across these workflows.

And, uh, the really, you know, significant, you know, point of this is that obviously we can now do this across almost every single domain. So AI models are being trained again on legal and manufacturing and life sciences and financial services. So this begins to impact almost every part of our organization, and, and certainly any part of our organization that is more information-centric, um, that, uh, that we're doing work with. And so there's a, uh, you know, a pretty fundamental element here, which is for the first time in history we have this sort of separation between our time and energy into a process and the value that gets created out, because there's a decoupling of the actual productive work that's being done in that process. The productive work can now be done with computers, and we can send again one prompt or one workflow, and that, that could be equivalent to hours or tens of hours or hundreds of hours of work that we would have had to do. So it's really the first time in history where knowledge work has had that same kind of impact. We've seen this in manufacturing and other areas, but this has never really happened for knowledge work before. We've never had probabilistic systems that can actually help us make decisions, help us accelerate work, um, in all of the kind of enterprise workflows that we do every day. Which raises a kind of an interesting question. If you have this sort of new level of super intelligence that's available and accessible to anyone because of these AI models are relatively broadly available between frontier intelligence and open source models or open weights models, then we actually have a really interesting question that we have to sort of contend with, which is how do we ensure that our companies are uniquely competitive, uniquely effective at leveraging AI? Uh, how do we think about competitiveness in the future when everyone has the same substrate of AI, they have the same substrate of intelligence for their business process? You know, every law firm or every manufacturing institution or every bank has access to the same level of models or the same level of intelligence if they so choose. Um, then how do we start to make sure that we're designing our businesses to be competitive and unique and differentiated in this new world? Um, and so this is obviously going to be then the new role for everybody in this room and beyond, which is the, the unique approach is going to be how do we apply this in our organization with our data, with our workflows, with our people? How do we design our workflows and processes, uh, to be able to best use AI in a unique way in our organization? And a massive part of that is that models themselves don't, don't deliver that result because the models are trained on obviously largely broadly public information, but sometimes, you know, privately acquired, uh, data in, uh, in a bunch of domains. But again, that super intelligence knows everything about the world, but nothing about your specific business, or at least nothing that's not available online. And so models alone are, are quite limited in the impact that they can deliver. And agents thus need a lot of context. They need a lot of context about your organization. They need context about all of the critical data and information that is unique to your business. And the 21st century, I think, will be defined by the companies that sort of can best ensure that their agents have access to this context and their workflows in a unique way. So let's double click on the data for just a second. So this data that, that we all work with and we've, we've been kind of accumulating for years and years, we know how important it is. Um, it's the same data that we all work with as, as end users. It's the same data that's powering our processes already. But agents really, really need this information to be effective. They need the same data that we work with if they're going to help us accelerate decisions and help, help us accelerate execution. And as a byproduct of all this, agents are going to create even more of this data, by the way. So this only grows and kind of it has its own kind of recursive loop, uh, in that, in that sense. But when you parse out the data, and this is sort of where we get really, really excited, about 10% of that data is, is sort of classic structured data. This is the data that's living inside of your ERP system, your CRM system, your databases. Insanely important data. Like obviously we cannot run our businesses without the data inside of our ERP, inside of our CRM, inside of our customer databases. Critical data for powering, uh, our operations. But by weight, by volume of information, 90% of our data is unstructured. This is the data that is, you know, classically stored inside of all of the files and, and objects, uh, inside of our enterprise. And this is the data that obviously at Box we've been very, very focused on since the founding of the company. And it actually is the critical information that agents need to be able to help us accelerate our workflows, help us make decisions, and help us be more effective. So let's look at the context that agents need for just a second that's coming out of the unstructured data. So first of all, they need our product designs. They need our product roadmap information. They need to see where there might be delays. They need to see where somebody changed what we're actually going to go build. That's all inside of our specifications and our road maps. Uh, they need access to our projections, our plans, the, the potential plans that we, we might have. This is largely not living in, inside of our ERP system today. It's, it's usually in the prospectus that we generate. It's in the memo that we create. It's in the PowerPoint that is presenting a possible projection. Agency that exact same information if they're going to help you with a planning process. Agency access to your marketing strategy, your campaign assets, if they're going to be productively used in answering questions for your customers or, uh, or your employees around what the latest, uh, product development is. Uh, they need access to, uh, you know, employee data that might be employees you're onboarding or existing employees. They might need act, they might need access to how do we enable our sales teams or our go-to-market engine. And then they need to know what, what have we signed up for, what customer obligations do we have, what are the critical risks or liabilities that we've created that's inside of our contracts. So all of this unstructured data contains the critical information that agents need if we're going to successfully use them and put them into our workflows as an organization. And this is why we're insanely excited about at Box around the role of unstructured data and AI. Now there's a slight catch, of course, which is, you know, this makes it look really, really, you know, straightforward and easy, but I think we all kind of know, especially if you're in larger and larger enterprises, that to get agents this data, you know, our information is fragmented, it's in a lot of different systems. This, you know, generally even underrepresents the scale of the fragmentation that we see in the enterprise. And all of these systems have been deployed and implemented for very good reason. Um, each of these tools sort of serves a, a specific purpose and is very useful in its, in its particular right. Um, but the challenge is, of course, even as employees or as users of these tools, we don't always know where the right information is. We don't know which, which system has the latest, you know, updated draft of a document or a critical contract that we've signed because that data is relatively fragmented. Um, but now agents kind of enter and arrive on the scene and they're trying to find this data and they're sort of working across all these systems and they can't quite figure out this environment. So it's far harder for agents because they don't have all the same implicit knowledge that we have about our environment and where somebody put that last file or that last document or that last decision. So agents have a hard time sort of finding what is the most relevant and up-to-date information across these systems. But that's just really the start of the problem. So this is, this is a fundamentally existential challenge that we face as enterprises, which is if our data is fragmented, we know that it's incredibly important. We know that we have to find the most important information for, for our agents. And this becomes now a fundamental challenge that we have to contend with as organizations. The first big problem with this is that agents could just work with the wrong data. The more fragmentation we have, the more sprawl that we have means it's, you know, we're liable to have an agent go grab the wrong version of a contract, the wrong, you know, draft of that financial projection. Uh, pretty meaningful if, if the agent grabs the wrong financial projection document in your planning process. Pretty meaningful if it grabs the wrong patient information inside of one of our life sciences reviews. So that fragmentation now becomes a, a pretty meaningful challenge when agents are sort of roaming around these systems. It gets a little bit, uh, harder and crazier when you think about the cyber challenge. So all of a sudden we have all these agentic systems and we, we're probably approaching somewhere on the order of 10 to 50 to 100 times more agents than people inside of companies. These agents will very easily grab the wrong information as the right information when they have access controls, uh, that are overexposed. And so maybe you have an overprovisioned or overpromisioned environment because one fragmented data system wasn't kept up to date, or maybe even worse, if you have any kind of malicious actor, uh, where agents are trying to find any, any kind of crevice in the organization where they can find data. Now all of a sudden agents are actually leaking information. So that's obviously clearly going to be a huge problem for organizations. And then finally, a lot of our systems just weren't built for the AI era that we're now in. Our data is in on-premises systems, or maybe we have a sort of vertical stack of technology that doesn't play nicely with another set of tools, um, or agents that we want to be able to use. So you have an interoperability challenge, you have an integration challenge. Um, you know, a lot of our on-rem systems don't work really well with MCP or have CLIs. So now you have a challenge where the agents can't actually literally technically get the data that they need. So this is a massive challenge that we're going to be, you know, facing as, as, uh, as we move into this new AI, AI era. And, and it's interesting because a lot of times when we talk to customers and have conversations, we end up starting with an agent conversation and an agent, uh, transformation conversation, but very quickly it becomes a data conversation within, you know, one to two minutes. Now, of course, we're a little bit biased because we're having the conversation, but you can very quickly see how the role of unstructured data and context really, really matters and how we govern that, how we manage that, how we organize that really, really matters. And so we fundamentally believe that enterprises are going to need a, a seamless and secure way to connect enterprise content to AI. And again, well-governed, compliant, secure way in the organization. It's got to work across all of your, your, your systems. It has to work across all of your users, all of your agents, and all of your applications. So this is the platform that we're building at Box. This is our intelligent content management platform. And when we started this journey, uh, uh, you know, really helping serve enterprises, it was started on this backbone of how do we help manage, um, any of the scale of data in the organization. That's the sort of foundation of Box. We built a layer of data security, compliance, protection, governance to ensure that enterprises could manage their most important data inside of our platform. That's things like threat detection, data classification, data governance, retention management, much more. We then built a layer of content services to make it so you could throw any amount of data or, or unstructured data and files at us. The ability to organize that data with metadata, the ability to automate workflows, all of that information, you know, built into, um, all those capabilities built into the platform and our content services layer. And then finally, we've been building an AI platform and an agentic layer where you can do everything for, uh, from prepare your unstructured data for the use of AI agents, but also a leading AI harness, uh, specifically focus on unstructured data use cases. So the ability to have long-running agents that can, um, take any kind of document-related task and execute work on that. That could be data extraction. It could be AI workflows, uh, and, and, and be able to leverage any leading AI model within that harness. That's our AI platform layer. And then again, very importantly, we want to take that platform and make sure that it connects to all of your users. That's fundamental. That, that was sort of the origin of the company. Any of your agents. So it doesn't matter what agentic platform, uh, or, or which, uh, of many you've invested in or, or are deploying. You want to make sure that all of those have that same seamless access to information. And then ultimately any of the applications you're using, any of the, the, the sort of tools where content and information needs to be able to show up. So this is the platform that we've been building out, and this morning we have a set of announcements that take this platform even further around enabling you to get the most out of your unstructured data in Box. So I'm not going to sort of, you know, kill any of the thunder. Um, we'll leave that to the team, but, uh, but I'm incredibly excited for, uh, the, uh, the major product announcements that you're about to see that really transform the way that we work with our information in this new AI era. So to kick things off, I'm going to bring up Ben Kus, our CTO.

Hi, I'm Ben Kus, chief technology officer at Box. Now, as Aaron illustrated, one of the key things for your agents is to make sure that they understand your business. And to do that, they need to access your business context, your data. And for most companies, they store their data in a number of different data platforms: structured data, enterprise applications, and your unstructured content, your files. And it's these files that present some unique challenges for agents. And they're constantly changing. People are creating them, they're editing them, they're sharing them, they're moving them. And these files often have some of the most sensitive data in your whole organization. So you have to really worry about security and compliance. And in many cases, these are complex formats, and agents need to make sure they understand them. And of course, agents don't store the data themselves. So they need a secure way to access this information. And these are the exact challenges that Box is built to solve. So it starts with the layer of virtually unlimited storage, and then on top of that we build the security and the governance: your permissions, your auditing, your threat protection, much more. And then on top of that we layer the rich content services to organize, collaborate, automate the, uh, workflows on this content. And then on top of that we expose these capabilities via our agentic platform. This allows third-party agents or the integrated Box agent to access this data via MCP, via CLI, via ADA, via API, and more. Now, everyone though who's ever experienced enterprise content and files has always had a similar question: How do I find the right information in all those files? ... in relevant and accurate information, because if you give agents wrong context, they're going to give you the wrong results. In the last few years, um, some enterprises, when confronted with this challenge, they tried to solve this problem themselves by building vector databases, their own retrieval layers, syncing data, syncing permissions, using maybe some of the traditional retrieval augmented generation techniques. But this has often presented major challenges. Not only is it a lot of work to get this, uh, working properly, but sometimes the vector databases struggle when you get millions of files, terabytes and petabytes of data. You have to sync permissions and data constantly, especially as everything is changing rapidly. And the infra costs associated with this are significant, often cost prohibitive just to be able to store these kind of indexes. And even if you figure out how to solve these problems, you run into a challenge where the technology that powers retrieval is evolving so rapidly that oftentimes when you finished the work of making this system production, you have to just go back and redo it again with the latest techniques. This is why I'm excited to announce Box Context Retrieval. It's integrated directly in the Box platform, and it's designed to solve the retrieval challenge at enterprise scale, giving agents secure access to the right business context. So the way this works is that, um, for your data in Box, um, for different file types, we prepare the enterprise content. So, uh, you have your document-based files, you have your multimodal files, your images, your audio, your video, and we take those and we use multiple indexes. Not only do we, the, the traditional text-based index, but we also add vector embedding-based indexes that makes the content more searchable. And then we, um, anytime any file changes in your organization, we'll continuously update it so that at all times when you go to retrieve information, you're getting the latest, uh, information. And then when it comes time to actually do the queries, the, we do a hybrid retrieval mechanism. So it combines the precision of the traditional lexical-based, uh, searching to the new meaning and intent understanding of semantic retrieval. We combine these using a rank fusion technique, and, um, this gives the agent the best of both worlds, where the agent can actually be able to retrieve, uh, the best information. So whether you're looking for an exact customer number in a contract or you're doing research on a business concept, the agent can retrieve all these. And very importantly, this is not just a search for a list of files. This is pulling out the most relevant information from the content so to give to the agents. So they don't have to go read everything and look through hundreds of pages of files just to find the one page that, that matters. We, we provide that directly to the agents. And these capabilities are integrated directly in Box Agent in the agentic search capabilities, but they're also available via MCP. So third-party agents, custom agents, OpenAI, Anthropic, other vendors, the agents that you build yourselves, these are all can access the and leverage the power of the Box Context Retrieval. And the results are significant. When we run the recall benchmark test, you see a 20% or more accuracy about getting the right data in the query. And because we're providing only the most relevant information from these files, you see a dramatic reduction in the input tokens that agents need to process, oftentimes 90% or more. And for the same reason, the responses are much faster because agent has less to look through. So you get up to four times or more, uh, faster answers. So you have better context, lower cost, and much faster responses. And because this is built in a Box, it's 100% permissions aware. It'll directly sync with your source of truth of, of file permissions. You don't need to sync anything. You don't need to worry about vector databases. You don't need to worry about indexing pipelines. All of it's directly integrated in that Box. And I would expect that context retrieval techniques continue to evolve, and we'll evolve with them. So we'll always provide to you the latest context retrieval capabilities. You don't have to worry about it. We'll take care of it in the background. Of course, these kind of retrieval models, uh, and the retrieval quality depends on the underlying model that powers them. This is why we are proud to partner with NVIDIA to use their Neotron retrieval models to power Box Context Retrieval. We're also excited to announce that we're, we incorporated the open-weight Neotron models into the Box Agent, so customers have a choice of very accurate, cost-efficient, and fast models from NVIDIA. Bringing this together, Box is a single platform to securely manage enterprise content for your users and for your agents. With that, I'd like to bring Normal Normal on the stage to tell us more about how these capabilities power advanced business process. Thank you.

Thank you, Ben. I am super excited to be here and show you all how Box is innovating with AI agents and workflows to power how your teams work together. With models getting better every day, the workflows you run right now will look completely different tomorrow. At their core, all major business processes are centered around enterprise content. That's how teams onboard new clients, research new opportunities, collaborate on contracts, and ultimately make business decision. Let's now look at a typical workflow: customer onboarding. If you look at the workflow, customer document comes in, someone reviews them, routes it to the right department for approval, and a business decision is made. But if you observe closely, there are bottlenecks throughout the process. First, customer information is hard to find as it's fragmented across multiple documents. Teams spend a lot of time finding the information you're looking for manually. Not to mention handoffs adds delays and inefficiencies, making it very difficult to give you a visibility on where things stand. When every process depends on searching for content, extracting information by hand, and manual handoffs, your business grinds to a halt. The fact is that your organization is being held back by outdated processes. And that's where Box comes in. We've spent years helping organizations move content-heavy work forward, and we are bringing the same expertise to AI-powered workflows. With AI at the center, Box does not just add intelligence to your content; it actually puts the intelligence to work. And content services in Box makes it easy to find and reason across your content, give you the information you're looking for, makes your process visible, and orchestrates the work for you. Let's begin with our engine behind our workflow capabilities, the Box Agent. The Box Agent is purpose-built for unstructured content to understand and act on your enterprise content with secure access to your data. It supports long-running process reasons across hundreds and thousands of documents, and more importantly, is designed specifically to protect you against agentic security threats. With the Box AI Studio, your teams can customize the agents for your business use cases as well. So what you get is a very powerful, secure agent for researching and analyzing complex content. And today, we're excited to announce new capabilities for the Box Agent. Very soon, teams can search relevant web information safely and securely on Box. You can teach the agent how they work through reusable skills and instructions. You can multitask across your complex content for up to 20 hours, and more importantly, convert all your analysis into deliverables using spreadsheets and PowerPoints and other content directly in Box. Now the thing is, the vast majority of enterprise information still lives inside unstructured content, and if you want to automate the business process around that content, you would need to unlock the critical data inside it first. And that's why we launched Box Extract. As you know, Box Extract combines the latest AI models, advanced OCR techniques, and to, to accurately and automatically extract structured content at scale. And today we're announcing new capabilities to Box Extract to make it more intelligent, efficient, and more trustworthy. We're introducing eval cells to help you measure extraction accuracy. Agentic cost and accuracy optimization to improve your efficiency. Human-in-the-loop reviews to make your workflows more dependable. And adaptive model routing to balance cost and accuracy. And because Box Extract is natively built on Box, it inherits Box's security permissions and governance controls. With that, I want to invite Paige on stage to show us all the new capabilities for Box Extract.

Thanks, Normal. For this demo, I'm playing the role of an operations lead at an insurance company. Our underwriters rely on our customers' claims history or loss runs to decide what coverage and price to offer. But reviewing all those documents manually can take hours. That's where Box Extract comes in. To get started, I'll create a new agent and select representative sample documents. From here, Box Extract takes on all the heavy lifting, proposing the custom metadata fields most relevant across this data set and creating specific prompts for each extraction. In our business, accuracy is critical. One wrong dollar amount could leave us miscalculating risk and pricing our policies all wrong. So, before deploying this agent, I need proof that it performs reliably at scale. To do this, I create an evaluation set to benchmark our agent against verified ground truth data. This gives me an instant report on accuracy across different layouts and file types. I can also compare how various AI models perform, then choose the one that gives me the right level of intelligence for this workload. Now, I'm seeing consistent errors across my total losses field. Instead of spending hours manually tuning my instructions, the optimization agent analyzes how extractions can be improved and tests those fixes directly against our benchmark set. I'll accept these changes, instantly boosting our accuracy to 100%. But what if the document itself has conflicting data points? With review rules, I can define business logic in plain language to flag potential issues, like with this loss report saying that this claim for water damage is fully resolved but still listing a $45,000 outstanding reserves amount. Extract catches this immediately so we can get clarity before issuing a quote. And once deployed, this agent doesn't sit static. It continuously learns from everyday underwriter feedback to keep building accuracy. And that's the power of Box Extract, turning hours of document review into decision-ready data in just minutes. And it all happens seamlessly and securely in Box. Now we can quote faster and more accurately because while our applicants might enjoy accidental discounts, our chief underwriting officer certainly won't. Back to you, Normal.

Thank you, Paige. Uh, it is incredible to see Box Extract in action. Now that you've have your content is structured, the next step is automating the critical businesses that runs on it. And that's why we built Box Automate. Box Automate was designed from the ground up to orchestrate content processes across agents, teams, and your enterprise systems. With Automate, teams can natively add AI agents to their workflows to drive both deterministic as well as agentic outcomes. You can connect workflows to your business applications and scale automation of all your business process across your enterprise. With Box Automate, today teams can customize their agents for their use case, be it instructions, models, outputs, knowledges, everything. You can add human-in-the-loop reviews by assigning tasks to people based on set criteria. You can apply advanced guardrails to set clear boundaries on what your agents can perform and connect with external applications using enterprise integrations. With Box Automate, teams can bring AI into their workflows, keep people where judgment is needed, and work with systems they want. Now that you have a content-based process in place, your teams need a clear view into their workflows. And that's the purpose of Box Apps. With Box apps, teams have everything they need to bring intelligent content experiences to life. With Box apps, teams can create custom apps in minutes using an unified no-code builder. You can power views using metadata to surface key insights and tasks. More importantly, put agents to work directly within your apps to review information and look for answers without actually having to open a single file. All of this allows you to manage your critical business process in one central place. And our customers are using Box apps to create tailored dashboards for the work they do every day, from contract life cycle management to insurance claims to customer onboarding, vendor onboarding, and digital asset management. And these are just some of the examples of what you can build with Box apps. But instead of me describing how Box is delivering value, let's hear it directly from one of our customers. Please join me in welcoming Kevin Shin, head of information security at Samsung, to the stage.

Hello, I'm Kevin Shen, head of security at Samsung Semiconductor, Inc. Thank you for having me today. Today I'm excited to share how our GRC governance, risk, and compliance team completely transform our compliance and third-party risk assessment process using Box. The vendor risk assessment process is known as one of the most difficult and manually intensive work that the security team has to do. No one likes it, whether it's a security personnel or vendors themselves, really, really hate it. However, this is a mandate that many of the government compliance requirement as well as the security framework mandates it. So we have to do it and do it really well, as this is auditable. The process used to be we email security questionnaire form to the vendors via email, manually track down the responses and evidence that's sent back to us, and calculate the risk score by hand manually, and chase down for additional mandatory evidences and follow-ups again and again. As some of you probably know, it's a very painful process. And in the meantime, the vendor themselves is asking, is our information being secured or is this going all over the place? So very, a big, big problem. So as our vendor risk footprint grows, the process become very slow, inconsistent, and impossible to scale. We needed a modern and automated solution that is not fragmented all over the internet or all over the fragmented solutions exposing our vendors' very sensitive information. So this is where Box comes in. We transform our entire GRC life cycle into an automated AI-driven workflow. So the first step, we distribute security questionnaire directly through Box Form. So no longer email anymore. The vendors now have one centralized place to send their, submit their responses, upload their evidences without relying on emails or spreadsheets and worrying about if the information is being secured. Step two, Box Automate automatically triggers Box AI agents. This means AI agents analyzes the submissions and automatically categorizes the vendors into risk tiers from low, medium, high, and critical. So higher risk vendors are automatically routed to our team for human review and collect additional evidences that are required through Box File Requests. Again, no email, just centralized area where they feel comfortable utilizing and adding their information. So low-risk vendors bypasses this manual process altogether, move straight to the final decision stage, again streamlining the process. This step is very powerful, as you just seen the new capabilities coming in related to AI. The Box AI agent generates strict metadata on every submission. So our team can review AI insight. Think about that, AI insight that our team is utilizing to understand what's happening to the vendor risk situation instead of us having to manually analyze. Also, from the head of security standpoint, I could now compare the vendors and compare the vendors and analyze not vendor per one vendor at a time, but I could compare in an entire spectrum of the vendors and then accurately measure the risk that's associated. We want to continue to grow this setup leveraging the latest Box AI and AI agent capabilities. Boxure Pro, very important to me, ensures full security protection for the entire workflow happens. This means it enforces granular access control and data governance policies across the entire workflow to ensure sensitive supplier information, vendor informations are protected every step of the way. So this process have been truly transformative to us. Our team receives information faster. Assessments are consistent and accurate, and our overall compliance structure and posture is significantly stronger than before. Best of all, my team built and configured the entire workflow using Box. This means without needing to comp, needing to depend on the complex custom code or need to manage servers. We all went through the PHP scale where we have to patch. None of those matters anymore. Everything lives in one trusted AI-powered platform. So if you're managing third-party risk or compliance framework, consolidating your workflow into Box will transform your operations just like it did for us. Thank you.

This demo is inspired by Samsung's vendor evaluation process and adapted to show how today's announcements can further enhance their work. With that,

I'll be playing the role of the head of governance, risk, and compliance at Samsung. My primary goal is to ensure that every third-party vendor we work with meets our strict security and compliance standards. With hundreds of partners supporting our operations worldwide, our extended third-party network introduces risk every day. To solve this, we manage the entire process using Box Automate. I just received a request to evaluate a new vendor. And instead of collecting sensitive content over email, we send the vendor a standardized questionnaire using Box Forms, where they can provide required details and upload supporting documents in one place.

Upon submission, Box Docgen turns the form field inputs into a structured document that is placed in a new folder alongside the included attachments. And behind the scenes, Box Automate is driving this workflow end to end, coordinating people and AI agents the second a new file lands. So instead of having analysts spend hours combing through pages of technical audits, Box Automate kicks off the evaluation in real time, invoking the Box agent to immediately start our review. Now, to figure out just how risky a new vendor might be, we set up an evaluation rubric that the Box agent can use to score them. The agent analyzes the submission and evaluates it across key areas such as data sensitivity, operational criticality, and past security incidents. Each area gets a score which dictates where extra reviews are needed. And running in parallel, another Box agent looks for vulnerabilities or inconsistencies in the submitted documentation. If anything is identified, Box Automate is on it, tasking the assigned reviewer to request additional information and even drafting our email. My team isn't tracking down the findings or wondering what to do next. Box uncovered the issue and initiated the follow-up automatically.

So, now what? Instead of kicking off a frantic chain of who's reviewing this emails or digging through folders, we jump straight into our GRC Triage Box app. At a single glance, my team can see exactly what vendors are currently under review, who owns each assessment, and all the key details. Looking at my new vendor, we immediately see they're assigned to our senior risk analyst. She doesn't need to dig through nested folders or spreadsheets to understand the assigned risk level. The AI-extracted metadata, risk factor breakdown, and audit summaries are all right in front of her. And that's how we replaced endless email threads and manual audits with intelligent automated orchestration in Box, driving fast, compliant onboarding without a single spreadsheet in sight. Back to you, Nurmal.

Thank you, Paige. Thank you, Kevin. It's so incredible to see how Samsung is using Box workflow capabilities to streamline their vendor assessment process. In addition to Samsung, we've seen massive success from customers like SSAI, Bolt, Insperity, Argonne National Lab using Box workflow capabilities. Using Box, customers have accelerated their growth, scaled massively, and reduced the time taken to get work done from weeks to days to just minutes. And here's the exciting part. We are entering an extraordinary new era where frontier models have reached a new level of intelligence, enabling AI agents to train and continuously improve one another. So we asked ourselves a question. What if we use the frontier intelligence to help design and optimize all of our business processes from end to end? Imagine a world where superintelligence looks at all of your content, data, and workflows and runs millions of permutations to design the fastest, most accurate, and optimal process for your business. That's exactly what we're working on. I am super thrilled to introduce Box Architect, an autonomous agentic system that partners with your enterprise to reimagine workflows for you. Box Architect designs AI-first workflows for your enterprise with superintelligence. It autonomously builds the right process for you with a combination of agents, Box, and content. Post-deployment, it continuously monitors and optimizes the workflow for you. This is much more than a faster way to build and design with AI. This is your path to becoming an AI-native company. I want you all to share my excitement. So, please welcome Scott on stage to show us a sneak peek of Box Architect.

Thanks, Nurmal. This is a sneak peek of Box Architect. For this demo, I've been asked to modernize every workflow at our credit lending company. That's hundreds of processes, and each one contains countless combinations of possible steps, agentic actions, and handoffs. This is neither an easy nor quick task. Luckily, I have Box Architect. I'll start with our customer onboarding process and give Box Architect all the key files. Architect is now analyzing our existing credit applications, reviewing our reference files, and checking for metadata to understand how our business operates. It will then provide a summary and make a recommendation on what to build. Great. Now let's have Architect reimagine our customer onboarding process. Box Architect is now applying frontier intelligence to understanding how we decide credit, mapping the key stakeholders, and then running through all of the permutations of workflow steps and components to determine the fastest, most accurate, and optimal process that's tailored to how we run our business. And this is the process that Box Architect has proposed to build in Box. Now I have the opportunity to review and approve it based on our policies. And the best part, Box Architect doesn't just design the process, it can build it as well. Architect is now autonomously generating all of the Box components, including a new metadata template, an Automate workflow, and a Box app. And it even creates custom Box agents for you. And just like that, Architect has built the entire workflow for me. Let's review the Box app first. With this app, my team can easily see the exact status of each application in one place and click into specific views of the content. Box Architect has also created an Automate workflow for me that can be tracked in this app. So, let's go check that out. You can see that Architect has built an entire multi-stage Automate workflow that's optimized with the ideal set of outcomes for how we work. And it doesn't stop at building. Box Architect will continuously review the process and, with my review and approval, update individual steps within Automate and/or any other component that I built today. With frontier intelligence on my side, modernizing this workflow was much easier and faster than I expected. Back to you, Nurmal.

Thanks, Scott. Amazing demo. Seeing Box Architect in action really brings home the power of this moment. Now, let's bring it all together. The Box agent understands your content. Box Extract structures it. Automate puts it to work while Box apps surfaces it. And because it's all built on Box, it connects natively to Forms, Docgen, and Sign while extending seamlessly to your third-party applications. And soon, Box Architect will help you reimagine how every single process runs with superintelligence, transforming your content operations into autonomous workflows that move your business forward. Everything you see and all your workflows your teams run is possible because of the platform underneath it. So to talk to us more about Box as a platform, let's please invite Fernando to on stage.

Hello everyone. So, we've seen what AI can do when it's connected to all the context coming from your enterprise content. Next, we're going to see how to apply all that context anywhere where you're doing your work, building your agents, building your custom apps, or automating your workflows. Let's get started. So every workflow you just saw was running on top of the Box UI where admins and users are clicking, dragging, typing whatever they need. But Box is also a headless platform and this provides a secure standard connection to allow you to extend all Box capabilities including content management, security and governance, and AI. So you don't even have to use the Box AI because developers and partners can programmatically access and act on your enterprise content. They can deploy AI integrations or connect Box to third-party agents like ChatGPT, Claude, or Gemini, and all of that while maintaining your existing permissions, compliance certifications, and audit trails. Let's take a look at some AI projects.

Builders and developers assemble all these pieces by themselves. They put together memory, databases, vector stores, orchestration frameworks, and models, but they still need to figure out governance and access to all that content. So for every one of these projects, you're going to need access to content. So are you going to bring all those files into a specific separate data store, which then means that you will need to convert all those files, chunk them, create indexes so you can run low-latency and high-speed search at scale? I mean, this may work for a prototype, but it's just not ready for production. So while models can reason and orchestration frameworks can coordinate agents, Box provides the secure and governed content foundation underneath that allows all your apps and agents to be connected with the enterprise content that you need to make it available everywhere you're building your apps, connecting your agents, or automating your workflows. So instead of having to rebuild content infrastructure from the ground up, any developer can get started with our developer portal, create a free developer account, and start using our APIs and SDKs, our CLI, the Box MCP server, and now Box Mount. Anyone can get started and start playing with our APIs in our developer sandbox, which then means that you can try all the Box features without having to mess with your production data. Let's take a look at each one of these in a little bit more detail.

If your application requires content management, we have APIs and SDKs that let you do that. Let's say you're trying to build some UI, let's say a supplier review application. We have UI elements. UI elements give you immediately accessible and easy-to-integrate web components for common file operations. So let's say your procurement team is trying to bring some data into the application. We have an uploader for that. Then you want to show that data back to the user just to make sure that everything looks good. We have a previewer. So this means that you can really focus on your business actions, your business rules, and not having to reinvent the wheel by building file handling from scratch. If your application or workflow requires Box's agentic capabilities, we have the Agents API. The Agents API allows you to have full conversational access to our Box agent and any custom agent you create. With the Box Agents API, you can also run a deep research project across all your enterprise content and bring that result into the broader business process. You can even ask the agent to structure the data so you can connect it to workflows and have a more programmatic access to it. We also made this API fully A2A compatible. And this means that every Box agent will have an agent card that you can just drop in anywhere where A2A is compatible, including Gemini Enterprise. And this gives you vast possibilities, for example, in bringing multiplayer modes for Box agents running with your colleagues, say inside of Slack.

Now, let's talk about the MCP server. MCP is quickly becoming the standard for AI interoperability if it's not there already. And we have an amazing MCP server where we've curated the most relevant operations from our platform into an experience specifically targeted to agents. With the MCP server, you can connect your account and have peace of mind that all the operations that will be occurring will be respecting your permissions and your access controls. Developers can also use the MCP server to connect to their favorite coding agents or working runtime environments. But these are some of the partners that make all of these possible. We have a huge MCP ecosystem where you can go and today find the Box connector already available in all of these applications or developer runtime environments. Click a button, get started. And by doing that, you're immediately making your agent way smarter because now it'll be connected to all your enterprise content. It's incredibly easy to use.

Now, let's talk about the CLI. If you mostly live inside a terminal or you're a developer or vibe coding, we have a CLI that has access to every single API we have. With the CLI, you can programmatically access our API stack directly from the terminal by invoking commands. But we know that agents are really good at using the same tools that people use. So when you combine the CLI with an agent skill, you can have your agent have full access into the APIs without you having to think about how to instruct it. It's very powerful. But today, we have something very special for you because we made APIs available, the MCP server, the CLI, and we thought we were missing something specifically for agents that are running in sandboxes or runtime environments that are collaborating, where you're creating a sprawl, spawn of agents all fanned out tackling a certain process. That's why I'm so happy today to introduce Box Mount.

So very exciting. So you might be wondering, so what is Box Mount exactly? So think of it simply as Box Drive for agents because we brought the same sync engine you've been using for Box Drive into the terminal. So you can mount a Box folder and have all the changes that happen to that folder immediately and automatically synced back to Box in full bidirectional manner. So this means that you can create a deal room where you are spawning a bunch of agents to collaborate on a given folder, each in its own sandbox, each mounting the same folder, and that can also be a collaboration between agents and humans. Humans will be using it via Box Drive, agents that folder via Box Mount. That means that if any agent or a human makes a change, that change immediately gets updated in every other environment without you having to focus about that custom file transfer pipeline. So it's very easy to use, follows your permissions, and the additional value that it has is that during this collaboration conflicts may arise. But Box Mount takes care of all of that for you because it will flag if there was a conflict and you'll know if you need to retry the operation, merge it, or whether it requires further human review. We're making Box Mount available in the following runtime environments and development tools. And in particular, we'd like to give a big shout out to NVIDIA for making Box Mount available in OpenShell and DigitalOcean for their being launch partners for Box Mount inside of managed agents. Thanks.

So, Box works with the models, tools, and apps your organization already chose or will choose in the future. But we know that these tools are constantly changing. If you look at the last year or so, many of the tools we're using now didn't even exist. With Box as the file system for AI, you can have that single secure governed platform that can allow you to find the right access to information across millions of files in seconds while keeping all your security controls in place. So Box gives developers a foundation where they can connect their models and frameworks to the enterprise content, and admins are the ones who control and govern how that content is used across all the apps, agents, workflows that the organization is choosing. In this partnership between admins and developers, you make company knowledge available to all the people that need it across all the apps, all the tools, all the workflows with all the security controls in place. With that, I'd like to introduce Kevin Bronc from BTS and Dave Brown from Anthropic. They'll be telling us... Yes. Yeah. All right. Morning.

Good morning, everyone. My name is Dave Brown. I'm a member of the go-to-market staff at Anthropic supporting our productivity, collaboration, and content partners and customers.

Hello, I'm Kevin Bronc. I'm from BTS. We are a professional services firm and for 40 years we have spent time with our clients on the people side of strategy. So this looks like building capabilities, helping shift mindsets and align strategy and culture to get to the execution that they need. And we work with some of the world's best organizations around 35 countries. And it probably wouldn't surprise anybody in this audience or watching at home that the number one thing we're hearing from all of our clients is how do we start really driving P&L value from our AI? And so at BTS, we've been talking about this as an idea of spinning up innovative flywheels that can turn into AI diamonds where we see impact to the bottom line and when that happens then we can institutionalize them, then we can renew them and we can continue to innovate. So as we help our clients with that, we've been looking internally as well. And our latest diamond is set to answer the question, how do we scale BTS's collective brilliance. So if we fast or go back all the way to March, a few months ago, finding the right knowledge at BTS globally could be very painful. It could take a long time. It could take a lot of relationships, a lot of back and forths. In fact, it felt a little bit like this.

Tired yet? So, I don't know if that's how it feels in your organizations when you're looking for knowledge. But we wanted to solve that pain, right? And to do so, we created something that we call the BTS Brain. And what the BTS Brain is designed to do is take all of our portfolio knowledge. We have a complex interconnected portfolio and really visualize it and bring it together. So we have Box as our data layer. That's where we're pulling all of the trusted content. And then we've used Claude design and Claude skills on top of it in order to bring the entire portfolio together to ensure that consultants have the best information at their fingertips at any moment. I'm going to give you a taste of what it looks like.

So we call this the brain and in here is our interface that consultants can jump into at any moment and jump in and look across our entire portfolio across all of our centers of excellence. Look and see all of our client needs and the solutions and how they all come together and connect and pair together. And at any moment we can go straight into Box where we can then find the source of truth. And these are all curated active files that we have owners across the business that are looking at to ensure that any consultant can access what they need. On top of that, we have Claude that has skills and ability to then pull into these trusted files to be able to shape the content in real time. And through this, not only are we able to look at different things, but we can add them back into the brain so that more consultants around the globe can then access them. So, it's a really, really powerful system, and it allows us to not only visualize our portfolio, but actually bring it to life to drive speed, quality, and scale.

So, why does this combination work? Well, as Kevin mentioned, Box and Claude are better together. Box has the trusted content, unstructured and structured content in every type of file. And Claude is the frontier reasoning engine. It provides you with model intelligence, unstructured data and reasoning that's combined and able to be put into the workflows that your team and your apps are already using. So for us practically at BTS, what this looks like is consultants can take all of our client context that lives in a secure Claude project or in a secure Box folder and they can use all that context and put it into the trusted content of the brain to then tailor and customize to meet our customers exactly where they are. What this actually means for us is we're able to customize at 10 times the speed we were able to do before and we're doing it with a significantly higher quality because we're bringing the best thinking from around the globe to each and every client engagement at a moment's notice.

And we're bringing the best of Box and Claude deeper together. In fact, Box has curated an incredible MCP server with plugins and skills. In fact, we're actually doing more looking at appmentioning files in Box. And our governance layers are complementary. Box Shield, Box capabilities, the Box MCP server, and Claude respect all of your existing permissions for your users. So, you're able to know and your admins are able to trust and ensure that your content is secure for agents and humans alike. So when I go back and think about AI diamonds, it's only been a few months, but already what we're seeing at BTS is we're seeing not only do we have clear portfolio understanding across the firm, but we're seeing increased velocity and onboarding of sellers. We're seeing higher quality proposals and higher quality deliverables. And in some cases, RFPs that used to take weeks to put together are getting done in mere days or hours. We're going to see how things continue to innovate. Box MCP and Box APIs are coming together to not only allow us to access our content, but we're also starting to explore a concept of knowledge health and really being able to understand how healthy is the knowledge across our firm so we can continue to drive innovation from the field. But at the end of the day, AI is not just a technology opportunity. It really is a people opportunity. So, we've leaned into new ways of working and new roles. We've established an ecosystem of working that we call owners who are owning the content within the Box and within the brain. We have navigators who are searching and leveraging the brain. And then we have givers who are innovating with our clients and feeding that knowledge back into the brain. These roles exist at all levels of our organization from the front line to the CEO. And some days we play one role, some days we play all three. This is a really, really exciting moment and it couldn't be done without the foundational unstructured data we can pull from Box and the partnership with Claude to be able to access and bring that data to life. So this is what it looks like to drive speed, quality, and scale and do so with pride. I'm now going to turn it over to Paige who's going to show what this might look like in your organizations. Thank you.

We just heard how BTS partnered with Anthropic to build and utilize their knowledge brain. My demo was inspired by this work and shows how a consultant can use Box and Claude to build a tailored client blueprint. Right now, I'm in the thick of a high-stakes engagement with a global manufacturer who's rolling out a new production tracking system to hundreds of frontline team members. But plant supervisors are hesitant about this change. So, the client's leadership team wants a concrete 30-day adoption blueprint, and they want it by end of day. Our firm has solved this exact challenge before, but the proven playbooks, recipes, and learnings that make up our collective brilliance aren't doing us any good if they're trapped in slide decks or one-off message threads. That's why we built our knowledge brain on Box Hubs. Unlike a generic folder where files can get dumped and forgotten, Box Hubs gives us a curated source of truth. The hub references our content without duplicating it, meaning permissions, governance, and version histories remain completely intact. So while Box maintains all previous versions of this playbook, connected AI agents only reference the most current. Now, I don't have a lot of time, so let's start building our blueprint. The team supporting this client has been consolidating discovery notes, metrics, and more in our client's Box folder. This information is accessible to Claude through the Box MCP server. So I can simply ask this agent to consult the client folder and my transformation hub to build our blueprint. Box automatically segments and indexes all assets that live in our hub into semantic chunks. So when Claude searches our hub, it performs high-precision semantic retrieval using the specific concerns raised in my client folder to pull the information straight from our verified playbooks. Now, for the sake of time, I ran this prompt earlier today, and within minutes, Claude used our approved template to create a complete blueprint backed with Box citations, so the team can double-check every methodology. And our blueprint was saved in the right folder, so the required permissions and governance are automatically in place. Using Box and Claude together, our knowledge is live, secure, and easily activated to solve our clients' toughest challenges. Not only did we build the exact blueprint we needed, but it saved me so much time. I might actually make it out of the office by five. Back to you, Fernando.

Are you ready? Our mission is to power how the world works together. AI models are trained on public data, which means they don't know anything about your business. The real value comes from giving agents context by connecting them to your unique business content. Box is the intelligent content management platform. Think of it as the secure content layer underneath your whole AI stack to be able to work with any agents, any model, any workflow.

Box AI helps us surface the right information at the right moment from decades and decades of content. We're spending more time being creative, more time on the project, and really more time being human. When agents are grounded in secure enterprise context, they can orchestrate entire workflows and drive end-to-end outcomes. Box Extract will check the documents and fill in the metadata fields. Belmark went from doing 4,000 data points a year to 240,000 because of Box. Box is the file system for AI. Any agent plugs in through open protocols like MCP right where your content lives, whether that's from Claude or ChatGPT or Cursor. Box has a security layer architected specifically for the age of AI agents. For the CISO, you get full visibility and control over every agent action. They built the infrastructure to be future-proof. I need to know that the data that I'm leveraging and using is something that I would consider trustworthy. Box is the solution tool I use every day, all day. Box connects your content to the people, apps, and the AI agents that need it. Box is game-changing.

Please welcome to the stage founder and CEO of NVIDIA, Jensen Huang, joining Aaron Levie.

Okay, I got something. Yeah. Yeah. Okay. Oh, you're going to be on the side. Okay, we're just going to screw with the AV people. Okay. Um, all right. Uh, I take the box to work. This is BoxWorks. Jensen, how you doing, man?

Um, you've been at the center of some interesting stuff recently, this whole AI boom.

What, was there a moment? Actually, we're going to talk about obviously AI progress and everything going on. I appreciate you coming out here. I am kind of curious because there is a pretty cool story here. Obviously, the founding of NVIDIA and the centrality that you guys have had in this movement. What was the moment that really, really felt like the AI thing is happening?

And then all of a sudden, you know, big breakthroughs in tool use, so now we can actually use physical tools.

Right. And all... and we've actually even seen that there's actually some pretty good transfer happening, like the latest version of Astra is just as good at robotics as like the fine-tuned models. So this stuff is transferring across, like you can, like the same thing that's powering computer use is now able to translate into robotics.

Yeah.

What if you maybe with your hand drew the exponential curve? I don't know what you're seeing on the RSI front, but like do you think things go vertical from here? Do we tap out at some point and then it's just a pure diffusion era? Where do you think we are in the pure just AI model progress from what you're seeing on the ground?

So that kind of brings us to a little bit of the diffusion era of AI. So what are you seeing? You have had some pretty breakthrough announcements obviously in the chip architectures that will continue to keep the efficiency going and you know it's incredible the amount of tokens that these agents are going to consume. So we need as many chips as we possibly can get. When you think about now the layer above that and how does AI diffuse into the enterprise? What are the building blocks that you guys think about where obviously you know we've announced a major partnership on context retrieval so agents can get access to the underlying knowledge that they need. You've done a lot of work in sort of this idea of long-running agents that sort of have their own identities and can spin that up. You've done a lot of security work. Where are we in the diffusion phase? What are the building blocks that we need to get right in the next three years or so to actually diffuse AI across the enterprise?

And what that allows us to do is get token cost to come down.

Therefore, the intelligence value will go up.

The quality of the answers improve.

And that's what Box is about. The thing that I'm super excited about is all of a sudden you got what, one exabyte, how many exabytes of unstructured...

Okay. And so let's say you have a few exabytes of unstructured data. The world generates, I mean this is how incredible your opportunity is. All of that data, if you could put a semantic engine on top of it that understands the meaning of that data, and that's what the context engine is.

Right. And what makes NVIDIA special of course is all the chips that we make, but what we're really obsessed about is the libraries that create these amazing systems that people build. And one of the areas that we focus on is the semantic database, the vector database, the embedding of it, the indexing of it, the search of it, the ranking of it. We have world-class libraries in those spaces and partnering with you, we turn it into a context engine, this intelligence engine that sits on top of Box. That context engine is connected to people, of course. And you know, there's a few hundred million users of Box around the world, but now it's going to be a few hundred billion of agents that are going to be connected to that context engine. And so the amount of data that's going to be created, the amount of data that will be stored, the amount of data that will be retrieved and used is going to go exponential. And so I'm super excited about that.

When you look at, let's say we...

Oh yeah, sure.

It felt like there was an answer in there. So I kind of pulled it out of you in my head.

I appreciate you, you were walking me there.

Yeah.

No, by the way, it's good.

One of the smartest CEOs in the world, you guys, is the reason why. Yeah. Okay.

Yep.