Box的Aaron Levie：模型到工作流的桥，值一万亿美元

id="podcast-transcript" data-framer-component-type="RichTextContainer" style="transform:none">INTROAaron Levie: I want to send an emergency alert to everybody who's a sophomore or junior in college and just be like, follow these 20 accounts on Twitter—and also join Twitter—because this will just help your career. You're either a year ahead or a year behind simply based on your feed. And my feed is so wired in. I'll still talk to 20-year-olds that are like, "Yeah, I see some articles," and I'm like, what do you mean? How do you see articles? I don't even know what that means. Do you just get lucky that somebody emailed you an article? You just have to be wired in. I enjoy it. It's a lot of fun. I play with everything. MAIN CONVERSATIONSonya Huang: Today I'm excited to welcome Aaron Levie, founder and CEO of Box. Box is building a collaboration and content management platform for the enterprise that has now taken on new life with AI. I'm excited to chat with you today about Box and AI, about your general thoughts on AI—because you are such a thought leader—and about how founders can reinvent themselves for this AI wave. Thanks for joining us. Aaron Levie: Thanks for having me. I'm a big fan of the podcast. I religiously watch every episode, so great work. I am wondering, though—I know this is a different podcast, but when do you talk about root canals on this one? Sonya Huang: Oh my gosh. Are we going to call in Doug Leone? Aaron Levie: Have we figured out how to weave that in yet or not? I mean, I'm happy to talk about a different kind of pain and suffering that I've been through. Sonya Huang: We should just do this pod with a live root canal. Aaron Levie: Let's do it. See how it goes. Actually, that would be like Hot Ones, but you get a root canal and you have to talk about your strategy while you're in a dentist chair. And Doug is just on the other end of it. Sonya Huang: Amazing. Doug is so pleased with himself right now, by the way. Aaron Levie: Have you seen every tweet? Sonya Huang: Oh, yeah. He's very pleased with himself. Okay, good. Let's start with—I'm curious your take on this: application companies are the hottest neolabs. Agree or disagree? Aaron Levie: Two years ago, I think it would not have made that much sense—like, what does that mean? But very clearly, I think this is what's playing out in the market, and it's all working out mostly because of open source. What's pretty amazing is right now, the whole concept of being an LLM wrapper, a model wrapper, is actually working out. What I think people underappreciated was that in the real world, in the enterprise, what you need is some bridge from the model's capability to the actual workflow that the enterprise has. And on that bridge, probably $1 trillion has been bet on basically one of two outcomes: either that bridge is very limited, or that bridge is actually very vast and needs to be able to take on a lot of depth within organizations. That bet basically looks like: are you only long the model itself and superintelligence, or are you long this application tier—or, maybe previously, it would have been just the pure neolab. But I think it's very clearly playing out that there's a lot of gap between the model and the workflow. And as you bridge that gap over time, you get to a point where you realize, oh, I should also do the model. Then you have enough data, you have enough domain expertise, where that becomes its own flywheel. So I'm very bullish on this idea. And what's cool is it's opening up multiple layers of opportunity for startups, because you could either be the actual applied company itself—i.e., the neolab—or you could be the infrastructure provider to the neolab. You have multiple layers of going and attacking that space. But I think it's a huge update for the market and everybody's views of it. Sonya Huang: Box Labs, let's go. Aaron Levie: We already have it. It's a little bit secret. We pay very close attention. Right now, the main focus is: let's make the agent really, really good on any model. But over time, obviously, you would peel off certain use cases, either on a per-customer basis or across the whole data set. Sonya Huang: Yeah. It seems like there are two forces happening. One is people don't want the fox guarding the henhouse. They don't want the seller of the token to be the one that is also metering and gating what the best token is for each use case. And then the second is there's actually a lot of work to do on that bridge. There's real research involved in it, and it's pretty bespoke to the exact workflow and the exact end customer that you have. Aaron Levie: Yeah. I think what I tend to see happen in the Valley—and for very good reason; it's actually why these companies have been so successful—is that everybody is so research-pilled. Which is, again, totally awesome. Big fan. It leads to all these breakthroughs. But there's this sense that, okay, the model and the intelligence in the model is the only form factor that matters. And then you go to the real world and you see how intelligence actually gets rolled out in people's workflows. The model could be the most intelligent superintelligence in the world, but that workflow still requires you to connect up to other data systems. It still requires these moments where there's a human-in-the-loop interaction. There are delays in the workflow, so the agent has to sit idle. There's change management of the actual business process. There are legacy systems that haven't been modernized. So you go through these five or ten things that are much more operational, much more blocking and tackling, than just the pure superintelligence of the model. And the last thing a classic research organization wants to do is go attack every single one of those things. This is not a one-off in history. We've always had this relationship between infrastructure and application. Obviously AWS or GCP or Azure have created trillions of dollars of market cap in infrastructure. But guess what? There are also trillions of dollars of value in software that only exists because of that infrastructure. If you were to go back ten years ago and look at what was happening in the data space, and you looked at what early versions of GCP or AWS were building, I guarantee you would not have predicted Snowflake or Databricks existing. You would have been like, the infrastructure already does that. Why would you pay another $10 billion of revenue to all these other products that are just making it so you can work with your data? The same thing is going to be true for intelligence. The models will be insanely valuable, but the application of bringing those models into real workflows—in banking and life sciences and healthcare and government—that's just going to be a lot of software. Now, the challenge for the next, let's say, two to five years is: how much do the model providers need to move up that stack and also try to compete at that layer? Or do they leave open—intentionally or accidentally—that entire space for the application ecosystem? To some extent, this is a big strategic question for them, because on one hand, you want to be closer to the customer. On the other hand, you also want to have an ecosystem so people trust you. So that's going to be a really interesting tension over the coming years. Sonya Huang: How do you think it'll play out? I feel like that's the question we spend every day wrestling with. Aaron Levie: Yeah. I don't want to speak for Sequoia, but you can sense some of the existential dread of a VC right now, which is: should I just put another billion dollars into Anthropic? Or should I attempt to see how the applied layer is going to play out? So nobody's that envious of your position having to figure that out. But obviously, it's even harder for the entrepreneur. I'm pretty long, obviously, the application layer. I'm also equally very biased. I have a very concentrated bet with very limited diversification on it working out that you still want to buy technology that understands your workflow and can get to the core enterprise data. But I just don't see a different event happening than all of history. We only have 50 or 60 years of computer software history, but basically, when you go to that law firm, and you go to that pharma company, and you go to that bank, they need something that bridges the core technology to their workflow and their business process. AI has not meaningfully changed the need or the shape of what that looks like. The best manifestation of this is the chatbot versus agentic workflow demarcation. The chatbot can be totally universal and totally horizontal. But all of a sudden you look at that and you say, well, my workflow needs something to ping me at the right time in the process, or it needs access to a certain kind of data that the chatbot can't natively get access to. So somebody has to go into that organization and get it set up, and somebody has to provide domain expertise to this model so it really understands our particular business process. So then, unless you really just underwrite the big labs at—honestly, I'm not exaggerating—like 100,000 employees, if you don't underwrite that, then the diffusion economy is going to be massive. Every single one of those companies, whether it's a 50-person firm or certainly a multi-hundred-thousand-person firm, is going to need an army of people to go in and help them with that transformation, that change management. Whether it's the FDE phenomenon or just, again, understanding that domain expertise, that's going to be a very big deal. And then you alluded to this point about the fox guarding the henhouse. That might actually be singularly the biggest reason this has to happen. Even under complete benevolence—nobody's actually doing anything sneaky—it just stands to reason that if I'm going to give a task to an agentic system, I want that task to be cost-optimized, with accuracy held constant. So who can do that? It's the company that doesn't care which of ten different models is performing that task. Definitionally, you would want that to be the company that does not have a preference. Sonya Huang: And then the force you have going against that is that the model companies can subsidize their models or offer them at different rates. Aaron Levie: I don't know how long that lasts, though, because when these companies become public, I think they will be held to basically the same laws of capitalism as everybody else. The subsidization is working very well up to a certain threshold of spend, and we might have exceeded that spend when you're at tens of billions of dollars of capital. In fact, if anything, it might even be worse, because eventually you have to pay for your training runs as well. So the subsidization of tokens, I think, just has to be a temporary phenomenon. Sonya Huang: You mean from a gross margin perspective or from an antitrust perspective? Aaron Levie: Entirely gross margin. Sonya Huang: But they have such high gross margins on the inference right now. Aaron Levie: But then what are they subsidizing? Then they're actually charging decent rates. Sonya Huang: Yeah, but they can afford to price the API higher than their own first-party products. Aaron Levie: Totally fair. There's an interesting dimension, which is: well, if the API is the high-margin thing that is paying for the subsidization, but you've moved all your customers over to the applied product and there's no API revenue—there's an equilibrium you have to strike with this. And then all the while, if you have some non-economic actors, or people with a totally different game theory in this—Meta being one, maybe SpaceX being one, China certainly being a giant one, even Nvidia being one—that changes the calculus as well. Those four cohorts don't necessarily need to make money on inference at the same margin structure that an Anthropic or an OpenAI needs. They might be fine bringing inference down to a 10% margin because they just want to pay for massive compute clusters. As long as that happens, and as long as there aren't insane, closely held proprietary secrets, then no matter what, you're going to have cost per token go down on a like-for-like basis. All of which means more value accrues to the application layer. Which is a very long-winded way of saying I think there's value for everybody in the stack. The only thing I probably wouldn't bet on is: okay, one or two labs get 95% of the value creation. I think there's going to be a much more dynamic environment. And honestly, if I were one of the two or three biggest labs, I think I'd prefer this outcome too, because—back to your antitrust point—at some point you'll just be nationalized if you're the only thing that exists as intelligence. So you want a little bit of healthy competition in this ecosystem anyway. Sonya Huang: Yeah, totally. Let's transition to talking about Box. I want to come back to tokens and Jevons paradox and China and all this stuff, but let's talk about Box for a second. Aaron Levie: I love that topic. Sonya Huang: I'm guessing a lot of people who listen to this podcast use Box, but give people a brief explanation of the history of Box and how you're reinventing yourself with AI. Aaron Levie: We started the company as a way to securely store and share data in the cloud. It was a very simple idea, but we cracked a nut or struck a nerve, and—for Doug out there—we were able to scale up quickly. We pivoted rapidly into the enterprise. The idea was that enterprises would be moving from on-premises systems to the cloud, and they would need a better way to store, share, collaborate on, and manage all of this unstructured data—their corporate documents, their financial documents, their marketing assets, their research materials—in the cloud, securely. So that was the company. We had been flirting with AI products and experiences really since about 2015. If you remember the first rise of—at least in modern times—the AI winter that happened in the 2015 to 2018 period, where we thought it was going to happen and it didn't. That was a period where we were like, okay, these very early AI models are showing us signs that if you looked at an image and you could classify the image, that's pretty useful if you're in an enterprise. Now maybe you take all of your image data and label it, or maybe you'd OCR something and be able to pull out the text in there. That's enormously helpful. The problem was it was insanely expensive, and you had to have a model for every single use case. It required a hyper-trained model for each workflow you wanted to do. So we shelved it. A few years later, we started paying attention to the GPTs. We had some hackathons where people were like, oh, we could do type-ahead in one of our note-taking products. Those were early versions. We did some early work in text detection and classification, which helped with security use cases. But then, obviously, the ChatGPT moment hits, and that was the big head-exploding moment—if for no other reason than they figured out a form factor that opened up everybody's mind to, oh, these could be interactive systems where you just ask a question and get an answer back, of increasing complexity and length. So we looked at that very quickly. We jumped all in. We did the whole company pivot. Everything was exactly like, academically, what you should do. We had a team carved out. We put the best people on the team. We met every day, looked at the updates, and slowly but surely built out what today is our AI stack, and then, basically, the Box Agent. For us, the use case is very straightforward. We sit on hundreds of billions of files. Every single one of those files contains critical information for an enterprise. That could be their contracts, their research files, their marketing assets, their loan documents—all of this critical information. The problem is they rarely know what's actually inside of it. Unless you literally look at the document and search and find it, you just don't know what's inside of it. So now agents can be farmed out to answer questions about that data. They can pre-process it and extract metadata from those documents and turn it into structured data. You can use agents to automate steps in workflows. So we've built a platform that lets you deploy agents against all of that unstructured data, and that's been the core focus. Sonya Huang: Are you using agents to create new content? Aaron Levie: We are. There are a couple of modalities where that shows up. One is we have an online collaborative product that an agent can generate any amount of content in. And then we've done most of the—I think probably more exciting—work with OpenAI and Anthropic on how you do advanced document creation, PowerPoint creation. We've decided that their tech is, at this point, always going to be frontier. So we have an agent that interacts with those systems to produce a high-quality PowerPoint, etc. Sonya Huang: Awesome. Your tagline: "Your business lives in content. Unleash it with AI." What are the hero, home-run use cases for how people are unleashing it today? And then, if you had to fast-forward a few years, what do you think people will be doing with AI in your product? Aaron Levie: Probably the easiest hero for a more traditional enterprise to think about is as simple as: you have a million contracts, why don't you find out what's inside them? Or you have a million research documents—be able to pull out all of the critical structured data, put that into a database, and then be able to query, analyze, and automate workflows around that. That's the thing that knocks it out of the park every single time, because it's been a long-standing problem that people have never been able to apply human labor to. It's just too expensive to read every contract, every research document. Maybe you could do it if you had a loan document process, but most other data just never gets read at that scale. And then I think the stuff that we're probably as much, if not more, excited by is really the equivalent of what we see with coding agents or other complex agents, which is long-running agents that are executing your entire workflow or process. This would be in the form of: you go to a bank and you're onboarding at a bank, and they've automated every step that is possible to automate, and then it jumps out to a person at the steps in the process for extra review or extra verification. But now, instead of that one- or two-week back-and-forth, it just happens in an hour. That's the dream state of most of these enterprise workflows. What if we could onboard a client faster? What if we could discover critical data inside of our research much more quickly? What if we could alert to a security event much more quickly? To do that, you need these background agents or workflows that are pre-established for those processes. Sonya Huang: Totally. It's the year of the long-running agent. Aaron Levie: It is. Sonya Huang: I'm curious. You made the analogy to Claude Code. It seems to me that in the coding domain, using AI is not only accepted, it's embraced. Aaron Levie: Yes. Sonya Huang: In the content domain, which I think is where a lot of the content in Box sits, using AI to produce content—there's almost this allergic reaction to it. All the Pangram stuff on Twitter. There's this concept of workslop. I'm curious what you think about workslop, and will this still be a thing in a few years? Aaron Levie: I'm going to separate the Box corporate hat and just riff as a consumer of—gosh, I wish there was a better term—workslop inside of an enterprise context. Sonya Huang: I get board decks that are entirely written by AI these days, and it kills me. Aaron Levie: So here's the difference, I think, on the acceptability. There's probably more symbolism to this topic than just the slop element. Actually, the diffusion of AI in general almost ties to this. Take code. Other than the top engineers that we hang out with—who have deep taste in the code, and the judgment is incredible, and for them it's as much an art as a science—take that group aside. For most of the world, code is a utility. It's just trying to accomplish something. You're just trying to automate something. You're just trying to put an interface up there so somebody presses a button and moves to the next step. For most of the world, the value creation of code has been to automate things and have it as a utility. We'll probably still use the term slop for a while, because there's taste in front-end design and there's taste in systems, and you don't want vulnerabilities in your code. So that's going to exist for a while. But at the end of the day, if you can tell an agent, please go generate my entire back-end system or my front-end system, it's not only acceptable, it's preferable. That was the thing that was blocking us from moving forward, so we need to go do that. At least the way society functions and the way our brains work at the moment—maybe this changes—when you get a presentation from somebody, there's still this association: I'm trying to decide if I can trust that person to execute on that thing, or deliver that result, or understand that topic. So when you see workslop, you're like, I'm losing my ability to know for a fact how much of the thought process was them versus how much was the AI. How much should I even care about that? Because I myself am doing the same thing. So we have this very weird collective issue, which is: I'm doing workslop for some of my brainstorms and decisions, but when I get it from somebody else, I'm like, should I trust you? I don't know. It just might be a thing that, as a society, we have to keep cranking through over the next three to five years and end up at the other side. I hate to use these totally busted analogies, but obviously you don't care when you see somebody's financial model. You're like, yeah, that was clearly generated by a macro. You did not personally compute all of that. But you're showing it to me and we're talking about it. So why can't the same exist for a strategy deck or whatnot? I think right now we're going through this evolution of: what is the person's role? What is the content a proxy for? Is it supposed to be a proxy for how much that person knows? Is it a proxy for what we think they can execute on? We're just in this very messy period where we have to figure that out. Sonya Huang: Totally. Did you read the Stan Druckenmiller Wall Street Journal piece? Aaron Levie: I read the discussion about it—the reaction to it—but I didn't read it. Was it very sloppy? Sonya Huang: I don't think it was sloppy. I loved it. To me, it was a nice counterexample. I have this allergic reaction— Aaron Levie: How many "it's not X, it's Y"s were in there? Sonya Huang: I don't think there were any. Aaron Levie: Okay, okay. Sonya Huang: But it does show up as 100% AI in Pangram. Aaron Levie: Okay. And dashes? Sonya Huang: I think they're okay. Aaron Levie: You can't do that. Sonya Huang: But it was a nice counterexample to me, because normally I read something that's clearly written by AI and I just have this allergic reaction. Whereas with the Stan piece, I didn't. And I'm not sure how much of that was just, you know, it's Stan, therefore I trust in Stan, versus— Aaron Levie: No, but it is psychologically weird, because I'll read these X articles and I'm now doing 2x the amount of work to read them. I'm reading it, one, for the substance, and I'm also reading it, two, for the calculation of: did the person write it, or am I just literally reading a Claude prompt? And then my mental processing is: does that upweight or lower my judgment of the person or the post? I think we're in for some weird times because of this. I would hate to be a college professor. I would totally quit, because you're just like, I don't know anymore what you did. Sonya Huang: It does seem like the calculator is the closest analogy, though. Aaron Levie: Yeah, except that was more finite in terms, and you still had to piece together so many more things. Some of these analogies are breaking down—"it's just a task"—because at some point this thing is doing at least ten tasks at once. But yeah. Sonya Huang: Let's talk about harnesses. How does the Box— Aaron Levie: Transition to harnesses. Okay, speaking of calculators, let's talk about harnesses. This is back to the neolab, applied-layer thing. There's a bunch of things that we know about our file system, our permission structures, our search engine. And by all means, we'd love all the labs to train on our understanding of this, because it would only make external agents as good as possible. We always talk to labs like, hey, we'll give you as much data as you want about how the system works. But that aside, we have a lot of depth of understanding of: what do people do in Box? How do they search Box? When they look through ten files, how do they decide which is the one to pick? What is their internal calculus or heuristic for figuring out the most relevant document to look at? We know all of that, and we've built an agentic harness that attempts to understand that set of domain understanding about our system. It obviously has access to our search system, our file system. It has a bunch of mechanisms for pulling out just the text of a document, pulling out chunks from the document, doing embeddings on the document on the fly. So there's a set of tools it can use, and effectively it's a harness for asking questions of a large data set. In my Box account, I have—I don't even know the latest number, but on the order of tens of millions of files, just because it's everything that has ever accumulated over 20 years. I can now ask any question of all that data using the Box Agent, and it goes around, it does multiple searches in one, it ranks them, it very quickly pulls out the most relevant information, and then in some cases reads the full document. It does all the steps. Then we compare that against, well, what if we just gave Claude our API, or gave OpenAI our API? And we see meaningfully better results on accuracy and latency, because we know exactly how to tune it for our workflows. So that's effectively the harness we built out. Sonya Huang: And then what evals matter the most to you? Aaron Levie: I have a couple of funny personal ones—I just keep track of my own use cases. But we have, I don't know, hundreds of different tests that we do on every single model. We actually have two evals at the moment. One is we put out a thing called the complex work eval, which is a set of domain-specific work in life sciences, financial services, public sector, tech, etc. It's exactly what you'd think of as a document-centric eval: given these five documents and this set of problems, what would your answers be? We test every single model against those with our agent. And then we have a holdout eval—actually, the first one is holdout also—but the second one is just our Box instance and how Box employees use their data. We eval every model again on that. So we're able to roughly keep track of all the incremental progress. We see when things move by half a point in terms of model improvement. Then we roll out default models based on different cost and accuracy thresholds, and we let customers choose any model they want from, effectively, our model garden. Sonya Huang: What's your current view of the race and where all the horses are in terms of model performance on your use case? Aaron Levie: They more or less closely correlate with code, with one exception, which is that in some of our use cases, Gemini is disproportionately better than what you would see from coding. It might be just better tool use—given the Gemini ecosystem and what they need to build for, it solves a strong set of general knowledge work use cases as well. But I think mostly correlating to code. So Fable 5.1 was clearly state of the art and the best model that we've seen. There are obviously rumors about other models, so we'll see how the race continues on this front. But by and large, when you look at GDPval, Mercor has their APEX eval—these things will all generally follow the coding models. And so I think we're just neck and neck on Grok, Muse, the Fable class, and GPT-5.6-slash-whatever they're building next. It's a total race right now. Sonya Huang: And do your customers typically express a preference on which model they want to use, or do they just use your default? Aaron Levie: By volume, they use our default, because it's just easy and it works extremely well and it's tuned. There are a few ways our agent manifests. The way you'd most commonly experience it as an end user is you'd just be searching and asking questions of your data. But by volume, the volume of tokens tends to go through more of our workflow agents or data extraction. That's where you actually have customers doing evals, and they're basically saying, okay, I want to make sure that at this cost profile, I can get 98% accuracy on data extraction. That's a place where we'll have an FDE that goes in and helps you understand your data environment and test against five different models. And then you're basically at the mercy of the eval. Sonya Huang: What are you seeing in terms of the adoption of open-weight models in your customer base? Aaron Levie: Probably higher than people think, lower than what enterprises actually want, and much, much, much, much lower than what it will be in five years. Some mix of that would be the message. Sonya Huang: And it's primarily cost that's driving that decision? Aaron Levie: I have to probably attribute 30-plus percent to just the sexiness of— Sonya Huang: "I want to try GLM." Aaron Levie: Yeah. I think there's that. I've heard CIOs of Fortune 500 companies say, "We're playing with open source here." And I look at that and I'm like, well, I know for a fact that Gemini or Muse would have been just fine at that particular cost profile you're trying to hit. Or probably even GPT-5.6 Luna or Terra, or whichever one had the crazy discounting they just did. It probably would have been totally fine. But you want to be able to be like, okay, I'm a little hedged. It's cool. We're at that phase still. Over time, I think it stands to reason that you'll see meaningfully different costs, because you'll be able to peel off workloads that only make sense when you're grinding down to the cost of inference, in which case open weights will have the economic advantage. Right now there's this challenge where sometimes it's more token-inefficient. Sometimes, randomly—I've heard stories—it'll just speak Chinese mid-chain. So you're like, okay, well, that'll be weird for a bank. We need to work on some of those things. But long term, I think it has to be the case that you're going to peel off those workloads. One of the more interesting posts on this, that I totally subscribe to, is Jesse at Decagon's. You probably read that post—this paradox where you're going to see closed continue to go exponential, but what's going to happen is each use case that matures, you can peel off to open source. Once you have stability in that use case, it starts to make sense to veer it toward an open-weights model, assuming one of two things is true: one, that it's actually literally cheaper, or two, having some post-training gets you X percent more performance. So I think you will be in a reality where—and this is going to be very confusing, probably for the press more than people in the Valley—you'll be like, wait a second, the revenue of Anthropic, OpenAI, etc. is off the charts, but somehow open weights is also growing exponentially. And you're like, how is this— Sonya Huang: The pie is growing so fast. Aaron Levie: Yeah. And it's like, the pie is growing so fast, but what's happening is actually there's an interesting duality. It's not even just "rising tide lifts all boats." It's like, no: we either use Fable or 5.6 for orchestration and then farm out all these long-tail tasks to a cheaper model. Or the opposite is true—you have some orchestration agent that by default does the cheaper stuff, but occasionally sees something that is just way too hard and pops it out to one of these heavier models. So you might have a blended 50% spend on each, but ten times the amount of tokens on the open-weights model. Everybody's winning, but there's an interplay between why they're winning. Sonya Huang: I'm curious how you think about memory and customization or personalization, and where that's going to go. It seems like today the dominant architecture is still RAG-based systems. You can get fancy on the RAG, but it's still context lookup, where the weights themselves aren't fundamentally changing. If I listen to my friends at the labs, there's continual learning—this idea that the model's weights should adapt as it gets to know you. We had Engram on the podcast. I don't know if you know Dan. Aaron Levie: I just got introduced to him. I listened to the podcast. I would have loved to have been the fourth person in the room. Sonya Huang: Amazing. I think they're working with customers to help bake some of the context into the weights themselves. What direction do you think this is going to go? Aaron Levie: You're catching me right before I actually do my call with Dan, so I wish I could have talked to him first, and then I'd have a way more eloquent answer. I'm extremely fascinated by the approach. I have no reason for not wanting it to work and exist. We live in a world at Box where we see this high degree of complexity on permissions and access controls and data that tends to be the rub on a lot of these types of approaches. And I'm going to put them aside, because I'm sure they've already thought this through. So I'm going to talk more generically, philosophically. I think sometimes you will talk to a researcher who imagines the world working the way they work, which is: I'm a researcher, I have access to everything. So if I had a model that was trained just on my world, this would be amazing. And then you're like, let me introduce you to a lawyer. The lawyer has this tiny little access point of just the five projects they're working on, because somebody one door over is working on the competitive project for another company in the space, and they can't have any overlap in what they see or what they know. There can't be a single document that passes between those two walls. They have to be hard barriers. So, sure, you could still train a model just for that one user. But what happens if every single day they get added to or removed from something that adds important context to what they need to understand? Again, I think there are probably going to be breakthroughs in continual learning that resolve all of this. But this is why previously there was no way you could pull this off five years ago—it would be insanely expensive, impossible to wrap your head around how those access controls are supposed to work. But obviously, as the cost curve goes down, as open weights get cheaper, smaller, faster, better, I think this becomes super interesting. One thing on the podcast that I found very fascinating—and I just need a T-chart, honestly—is: what is the decision point of what goes in context and what goes in the weights? You have to be a little bit thoughtful about where the massive performance gain is that you get by baking it into the weights. There's probably some incredible calculation, like: when the rate of change of the data is not that high, but the upside of the weights dramatically changes the accuracy of the model. You'd have to land on some sort of rubric like that. Sonya Huang: If you could wave a magic wand—I think Karpathy said this in some prior interviews—if you could almost remove all the memorized information from the models and just have it encapsulate the specific reasoning capabilities, like the ethos of how we do things, for example, at Sequoia, and then you have all the actual content in a lookup system. If you could wave a magic wand, that almost feels like what the system would look like. Aaron Levie: That one's super interesting. I think the question will be: how much are enterprises different at that level, versus it's actually their literal IP that makes them different? How many different styles of execution are there in the world, versus, no, it's the depth of knowledge about that particular legal case, and how do I apply it to this other project I'm working on? That's where so much of the value sits. But again, if you can just wait till my Zoom call with Dan, then I'll really know the answer. I'm a fan, because no matter what—I've jumped right into the individual, but at a firm level there are probably ways to take this approach. I'm a big fan of what Trajectory or Applied Compute or Prime Intellect are doing, because there's no question that if you're Eli Lilly, you want a model for how you do drug discovery. That probably does need to go farther or deeper or be more specific than what you're getting off the shelf. And there aren't a lot of church-and-state problems for drug discovery workflows. They probably want as much of that information available to as many people as possible. So I think it's going to be domain-specific. You're going to have different outcomes based on which vertical or type of use case, and where the firewalls need to be in that process. Sonya Huang: Okay, so you hinted at the beginning that there's a Box Labs. What type of work is Box Labs doing? Aaron Levie: It's the equivalent of Box Labs—I don't know if we've used a capital L yet—but basically, it's our applied AI team. Sonya Huang: And what research areas are most interesting to your team right now? Aaron Levie: On the continuum of engineers, there's some cluster that is more on the research bent, and of that cluster, the things we spend time on are, again, at the applied layer. But it's a lot around: how do you take agents and get another ten points of accuracy improvement given X problem? How do you build a map of the problem set, with a given set of data, to best execute on that task? We spend a lot of time on that style of work. We have a team, for instance, working on how you do, at the agent level, at the harness level, some form of autoresearch on hill-climbing accuracy—answering questions or sets of problems on a given set of client data. If you're a bank and you have a bunch of loan documents coming in, and these are 100-page documents, whether you're getting 70% accuracy with an off-the-shelf model or 97% is, obviously, a world of difference in whether you can actually automate that process. So somehow you have to hill-climb from the base model to the 97%. There's a lot of work going into the system to pull that off. Sonya Huang: Maybe zooming out—and this can be a Box question or a non-Box-specific question—the role of systems of record in a world with agents. I'm sure you saw some of the Twitter discourse: every software company is trying to sell me their own agent right now. I don't want another agent from them. I want their system of record to work well with my agent. How do you think about that? Aaron Levie: Hashtag Claudeforce. Sonya Huang: Such a catchy name, by the way. Very catchy. Aaron Levie: It's one of those things where for the first three minutes you're like, man, that seems funny. And then four minutes later, especially when you see the stock, you're like, brilliant move. This is great. We're doing this. And then—I think somebody said this best—when they heard Matthew McConaughey say it out loud, that sealed the deal. Sonya Huang: That was the moment. Aaron Levie: That was the moment. I was like, man, he can sell software. His voice is so good for selling systems of record and agents. If you're in our contemporary group—you built a SaaS platform, and you have some set of data and workflow that your customers operate in—there are effectively two things you just have to do, and I think anybody attempting to do one over the other is going to lose. You have to build an agent that is insanely great at your product. That agent has to be provably 10 or 20 points better than an off-the-shelf agent at using your system—not because you've hobbled the other side, but because you are so eval-maxed and so tuned to your particular workflow that you can improve your system. You have to have that. And because you understand your domain, unless you're totally asleep at the wheel, you probably have use cases that no one has thought to build products around, because you talk to customers every day and you see what they run into and you're like, oh, we could just have our agent do that for you. I probably talk to a couple hundred customers a year in a variety of capacities. I've had at least two dozen times where the customer has a use case that is a breakthrough moment for me, like, shit, that would actually be totally insane. Sonya Huang: What's an example? Aaron Levie: Unfortunately, since you put me on the spot, I don't know if my example will pay off the level of excitement I just had. Sonya Huang: I'll put you on the spot. Aaron Levie: The thing I was thinking of—I think it's going to be a womp womp for the podcast—but there was a customer who had this idea: they wanted an agent in the background figuring out when documents met their governance policies. Does something need to go into some kind of archive, or into some kind of legal hold, or whatnot? And see—exactly. That's exactly the voice I was worried about. You couldn't even pull it off. But in our world, this is awesome, because think about it: every company has a head of governance. Sonya Huang: I meant that sincerely. Aaron Levie: Okay, no, I believe you. Listen, you do enterprise, so I think it was at least half serious. Imagine you're an enterprise. You have a head of compliance and a head of governance. They can only oversee the whole enterprise. They've never been able to be everywhere at once. Now imagine if they could sit next to the employee and be like, oh, you're about to do something that breaks our governance policy. The idea was: what if there was just an ongoing agent that automatically said, nah, that's going to break your governance policy—instead of the user having to try to predict or understand this stuff. Anyway, those are the kinds of things where, if you have an agent within your product, you're going to be able to identify them sooner and better than the rest of the market, and/or do things that maybe would be impossible to do off-platform. On the other hand, obviously you have to go headless. You literally have to make sure that your APIs are exposed to Claude and ChatGPT and all the different platforms, and you have to make sure you have either a direct way into deterministic APIs, so those agents can use your APIs and make calls via MCP or whatever, or at least make your agent headless and exposed in those systems. The only reason this is even remotely a hard debate is you have to make sure, as a system of record, that you can find a way where commercially it makes sense on the other side and is valuable and interesting. The reason I think a lot of people who weren't in these companies got that wrong was underestimating the amount of new use cases that are just total upside—complete white-space opportunities for these systems of record. In the Salesforce example, I use Salesforce more today, probably by an order of magnitude, than I ever have, because I MCP into it via Claude or ChatGPT. I'm always asking questions about the data inside of our CRM system. Sonya Huang: Do you think that means the systems of record become more tollbooth businesses, then, to make sure they're capturing the opportunity? Aaron Levie: I don't love that term, because no one's had a good experience at a tollbooth. Sonya Huang: I love tollbooths. Aaron Levie: Yeah, exactly. You love it more than governance agents. I would say that because they have a depth of purpose—organizing the workflow, managing the data, securing the data, providing guardrails—then, yeah, you have to have some kind of volume-oriented business model on that other side. And I just think if you're solving real problems for customers, it'll make money. This is so cheesy, but I've told LinkedIn product managers I'd probably pay 10x more for LinkedIn if I could just MCP into it—if I just had a way of always understanding, okay, the CIO is doing this thing and I need to reach out, or whatever. Take my money. So these systems actually have a tremendous amount of value based on the data they have, and customers will absolutely find some way to reward you for that value creation if you're doing the job. Sonya Huang: Super interesting. Maybe related—let's talk about product UI. The big generic chatbot, chat box, agent: is that going to be the dominant UI for how people use AI in the future, especially when it comes to the application layer? Aaron Levie: This is why I think the applied layer has so much room to run. The universal chat system that you ask a question to and get an answer back, or it does some work in the background—that's going to be a mainstay. That UI will always exist. It'll be incredibly powerful. The horizontal products will have it, the vertical products will have it, everybody will have it. It's just like your product has a search box—obviously it does. So that's always going to be here for these one-off asks of an agent: go find this thing, answer this question, produce something for me on demand. But most of the enterprise is made up of processes and workflows that are happening behind the scenes. Sometimes they're happening with computers running these things. Sometimes they're happening with other people doing these things. Sometimes they should be happening with people, but you could never afford to have them happen with people, so they just didn't happen. That's a slightly different metaphor than a chatbot where you ask a question and it comes back with an answer. It's more like, okay, I want agents in the background to do things for me. Read every contract. Look at every log. Triage every security incident. And then, instead of me chatting, maybe I'll chat as a means of catching up. I want a dashboard. I want a workflow. I want a queue. I want a task list. So then the challenge becomes: does the horizontal product take on every one of those components and manifest every one of those experiences in one? In which case, I think you'll start to be like, man, that thing is pretty heavy. And then we'll start to be like, oh, this is no longer the simple, easy, delightful thing anymore. So then the vertical players, who actually understand the process, can manifest all the right buttons and tabs and names of things for that particular workflow. As you have agents doing more work in the background, doing more async work—I've farmed out a bunch of agents to review things as they happen, or whatnot—that leans more toward the applied companies that understand those workflows and those processes. I think you're going to have these in every field. We already know how they're going to look in legal, with Harvey, Legora, etc. We're seeing them start to emerge in areas like security. We've seen them emerge in the long-running coding agent, with Cognition and Factory. So I think that will be one of the bigger applied AI use cases. Ultimately, five years from now, I would bet 90% of all tokens in the enterprise are things a user never kicked off and they just see a result. They see a task show up and they have to review it. It's just happening. Sonya Huang: Makes sense. Okay, let's talk about AI diffusion. Coding agents—it was like, boom, January 1st, 2026 happened, and the fastest diffusion of anything into the economy we've ever seen has happened. The diffusion of the rest of the AI magic into the rest of our jobs seems like it's been a lot slower. What are your thoughts on that, and where are the areas where you think we're going to see faster diffusion, and how is that going to happen? Aaron Levie: You always have to compare and contrast coding versus everything else to really understand the dissimilarity. In coding, essentially—and this is back to the utility point on slop—the utility of code is almost 100% represented by the amount of text that you can generate. Obviously, an insane amount of value went into the text, in knowledge and expertise and meetings and everything. But ultimately, the text is the thing that produces the program that is actually the thing you're trying to do. If you could have the world's greatest programmer who never had to sleep, never had to eat, could intuit what to build, and could just sit at a computer all day long, your value creation would be 100% correlated with how many hours they could sit at that computer and type lines of code. Ideally, good code is the thing most correlated to whether you produced software that people wanted. So it's all text. The models are hyper-trained on it. Everybody in AI labs treats coding as a competitive benchmark to constantly try to exceed. They get to do their own evals on it every single day, because they are the ones coding the models themselves. And it's the most technical audience of all time, where when they deploy an agentic system and run into a bug, or a problem, or some MCP server comes back with "connection invalid," they fix it. They know how to triage the problem. They don't call IT. They're just like, oh yeah, I didn't open up that port. Sorry, I'll fix it. So that's five things. Oh, and maybe the sixth: it's a very, very high-paying vertical, so it's automatically valuable. If you could get a 10% or 20% productivity gain, let alone a 5x productivity gain— So take those five or six things that coding has as beneficial properties for automation, and compare that to every other form of knowledge work. You'd probably have a histogram—I don't know if anybody's published this; maybe you can—of the similarity to coding, and what the domains are that start to look less and less like coding as you scale out. And lo and behold, legal is interesting, because there's a lot of value creation from somebody sitting at a computer reviewing legal documents, writing legal documents, processing large amounts of information. Okay, so that's blowing up. Then you go down the list. Now take something like a sales rep—so much farther down the list in terms of likeness. The sales rep's value creation is basically convincing an external customer to buy software or technology or a Caterpillar truck from them. That is the value creation to the economy of the sales rep. Let's say we brought the world's best automation to them. First of all, again, they'd have to figure out how to technically wire it up. They'd have to make sure they give it all their data, all these kinds of things. But no matter what, they're still rate-limited and constrained by: did the customer respond to them? Do they want to meet? Can they meet next Tuesday, or can they meet today? Does the customer have budget? All these other things. So that's maybe the entire continuum of knowledge work—and obviously this is not even touching working with atoms. On one end, you have somebody rate-limited by so many external factors. On the other end, you have somebody who could sit at a computer all day long and just type text. Your ability to automate things is that continuum. For the real world, we have to bring intelligence to these workflows in ways that somewhat feel like the shape of their work, and then find a way to deliver the change management, deliver the implementation, and get data into a format and into an environment that actually works with these systems. Asterisk: one of the other big things is, if you go to most engineers in 2026—maybe minus two months ago, given the latest phenomenon—the code's in GitHub. You just connect it to GitHub. Remember, there was this period where when you launched a coding agent, there was no sign-up or register. It was just, give us your GitHub. That doesn't exist in knowledge work. There's no "give us your GitHub" for knowledge work. Sonya Huang: Give us your Box. Aaron Levie: Well, Box customers have a much easier time with all of this, unfortunately. We're only $1.3 billion in revenue run rate, so that means there are a lot of people not using Box. What are they using? Their data is in on-premises systems, legacy file shares, legacy infrastructure, enterprise environments that don't talk to agents particularly well. So just think about that distinction between implementing coding agents versus everything else. Even something you'll fall asleep about: access controls in the enterprise are totally different. I actually totally forgot that point about coding. In coding, you get access to basically most of the stuff ever relevant for your job. In knowledge work, you're like, hey, Sally, can you open up that file share for me? Can you open up that project, because I didn't get access to it? How do you make sure the agent has access to that set of things? All of that work has to get done. So the thing I think we have to prepare for is two things. One, Silicon Valley has to prepare for diffusion taking a lot longer than they think. Or than we think—but I really think they, because I know how long it'll take. And the second thing—the good news—is this is all correlated to applied-layer value creation. The companies that will have the patience, the full domain expertise, the sheer work ethic—because it's not like everybody's just coming through the floodgates; you have to pound pavement and get out there—that will be the applied layer. So I think this actually represents $1 trillion of applied-layer AI value: how do you get the technology to the lawyer, or to the sales rep, or to the life sciences researcher, or to the person who runs the customer support team? That's all opportunity that exists right now. Sonya Huang: Awesome. I'm going to close by asking for some advice for other founders. Let's start with founder advice, then company-building advice. On the founder side, it seems like you are in every AI cap table—every cool new company, like Engram. How did you get yourself in the middle of the AI conversation? Aaron Levie: There are probably two parts. One, I was just very well primed for it. Working with unstructured data for 20 years, you can instantly see the benefit of agents on that. Honestly, it took longer than I would have wanted for us to get to have this conversation, because we tried to have this conversation eight, nine, ten years ago, and now it's finally happening. So first of all, just super well primed. Our product's shape, and what people do with our product, already lends itself extremely well to agents. Obviously, we had to bet the company on that. And lo and behold, we had the positive feedback loop of customers actually saying, yeah, that would be very powerful if I could read every document and answer any question. So that's the first, and certainly the biggest, by a factor of ten. And then the other is just that I'm extremely fascinated by the technology, and it's fun. Sonya Huang: Where do you learn about it? Aaron Levie: Your podcast. The Dwarkesh Podcast. Twitter. Unfortunately for brain cells, it's probably 95% Twitter. I have a routine where at the end of each night, I just go through the feed. I probably look like some sad meme, just scrolling and scrolling and scrolling and attempting to triangulate all the information. Sonya Huang: That's amazing, because it is the global town square for AI. Aaron Levie: It is. And I want to send an emergency alert to everybody who's a sophomore or junior in college and just be like, follow these 20 accounts on Twitter—and also join Twitter—because this will just help your career. You're either a year ahead or a year behind simply based on your feed. And my feed is so wired in. Sonya Huang: This is the number one advice I give to people when they're asking how to get current on AI. Follow these 100 accounts. Aaron Levie: But unfortunately, first step: join Twitter. You'll still talk to 20-year-olds that are like, "Yeah, I see some articles," and I'm like, what do you mean? How do you see articles? I don't even know what that means. Do you just get lucky that somebody emailed you an article? Just join Twitter. What are you talking about? So you just have to be wired in. I enjoy it. It's a lot of fun. I play with everything. Sonya Huang: What's your favorite new AI product? Please say Instinct. Aaron Levie: Okay, full disclaimer: I have not done the Instinct invite code yet, simply because I have a backlog of three other personal assistant products. Sonya Huang: You gotta try Instinct. Aaron Levie: I know. Everybody's—I'm very excited. Sonya Huang: And I'm not an investor, so— Aaron Levie: Oh, you're not? Okay, so this is totally genuine. Okay. I 100% will have—I don't know when this is going to run, but I'm sure I will have played with it by the time it runs. I'm in pre-release on a couple right now that I'm spending some time with. I think the personal assistant stuff is super exciting. At least in some of my use cases, it's still showing some of the limits of browser use, as an example. We still have some work to do there. Probably the entire internet needs a CLI for their product. So there's still some blocking and tackling at the infrastructure level for these things to be totally awesome. And then I look like your average AI-pilled knowledge worker: every day I'm asking one of five different AI systems 20 to 30 questions—doing research, looking for talent, looking for what a competitor is doing, what's happening in this market, how do you expand there. So kind of that. Sonya Huang: On the company-building side—you're a 20-year-old company at this point—what's your advice for other people trying to reinvent their companies to make sure they have max adoption of AI, not just at the individual level but at the company level? How do you make your business legible for AI? Aaron Levie: Some of this we have as a byproduct of how we've always thought about information systems in the company. To no exaggeration, if you have a question you'd like to ask about the business that has ever been documented in a form of unstructured data—a meeting note, a project plan, a roadmap, a presentation, a financial document, a financial planning session—it's 100% in Box. We benefit from a data architecture that is already insanely tuned for this, because this is how we've run the company. We didn't let anybody use anything else. So the data is very easy for us to work with at scale. And then, of course, we have Salesforce and all the other core canonical systems, and we've been able to have, I think, pretty good data hygiene. So agents running on top of that makes it a little bit easier to be AI-first in how we operate. Maybe a couple of best practices or things we've seen. First of all, trying to figure out where the highest-leverage-impact workflows are going to be, and targeting those. Our CIO's very AI-pilled. We have a bit of a center of excellence on AI. We've hired some internal AI PhDs to help with these processes. Sonya Huang: Do you have leaderboards? Aaron Levie: We don't token-max. We do actually have—literally, we have a list of people by number of tokens—but it's usually to inspect: okay, do we think that's useful, or is there a learning there that we should take back to some other function? Of our top three AI users at Box, probably one of them is wasting half the tokens, and two of them are like, oh shit, whatever they're doing, we need to do an internal training session for everybody else. We had this thing two weeks ago where I was like, can you just get everybody in a room on this particular team and show them how this one person is using AI? And six hours later, they were in a room and the guy was doing a full demo of what he was doing. Shout-out to Mick. That's the kind of stuff we're trying to do—how do we show everybody what it looks like to work in this way? But for as fast as we're moving—and we're shipping, at some parts of the stack, two or three times more actual customer-facing product; I don't care about how much code, but did we actually deliver more functionality that customers are asking for? Some parts of the stack, we're doing that—but then you'll talk to a friend at Anthropic, and you're like, oh my God, we still have a ways to go. The particular meta constantly changes. Two years ago, you'd be like, you're just using a plugin in your IDE? We're not ready to fully be AI-first. And then, okay, everybody roll out Cursor, and everybody rolls out Cursor, and that finally happens. And then you're walking around and you're like, you're not only working from Slack, just at-mentioning bots doing your coding? What are you doing? We're constantly changing what the workflow paradigms are on this. Sonya Huang: Do you guys have a Slack coworker agent, for lack of a better term? Aaron Levie: We have a few that have that shape. I'm pretty excited about Claude Tag as a form factor. You still have to get the team construct right and the data right. But we have a variety of ways that people work with agents in Slack. I don't know if it's as Slack-pilled as Benioff would like us to be, or as Anthropic or OpenAI are, but we're heading in that direction. Sonya Huang: Yeah. I think history books will be written about the art of business in this time. I would love to read the new Art of War with everything that is happening here, because I think it's pretty extraordinary stuff we're seeing. What do you think it takes to win in AI versus pre-AI? What does it feel like to be a founder right now versus when you started Box? Aaron Levie: I am both jealous of and also not jealous of the young founders you meet. You're just like, oh, to be young again. The whole world is your oyster. You can go in any direction, and the leverage you have—you'll meet with them and you're like, oh my God. I saw a product a week and a half ago. They did a demo of their product, and I was like, this would have been a 40-person project five years ago—or, especially when we were starting out, easily a 40-person project. And it was two people. You're just like, how do you have so many tabs that work? And they all seem to have stuff behind the tabs that all seem very functional. This is not fake. So I'm very jealous of that. It's incredible, because you can start your company from scratch with that as the design principle. Now, we will get there, because we're just going to muscle through it. There are a couple of things we just can't do. We're very uncomfortable with the idea of removing the code review, and some of these things get talked about, because our customers can't possibly entrust us with their data security and compliance if we don't take that seriously. So we're always going to have a little bit of a discount on the productivity because of where we are in the stack and what we do as a business. But so jealous of being able to be fresh in that. On the other hand, at the same time, for every great idea, it's instantly five competitors. We didn't have that problem. We had a good couple of years where we could just grind on our product and our experience. It wasn't like every three days you were like, oh, Sequoia funded this thing, and Benchmark funded this thing. We weren't going nuts with that. Now, we had our own version of that—at the time, I probably was going nuts—but in retrospect, it was not worthy of going nuts. Now it's like, oh man, this is a real race in every one of these markets. So I think I'm probably pretty consensus on this: in a world where AI builds things so much faster, the shift probably moves to whoever can actually get it to the customer being in the best position. It's fun talking to founders who are pretty pilled on that. Scott or Matan—they get the mandate. They're just like, this thing is going to be an enterprise diffusion play, so you have to get it to the enterprise. Anybody who mistakes the mandate right now is just going to lose. It's game over. Sorry. There are quite literally trillions up for grabs at the applied layer, and the companies that know how to build the teams and get to the enterprise will be the ones that win. It's obviously guaranteed. Sonya Huang: Well said, Aaron. This was a very fun conversation. Thank you so much for joining. Aaron Levie: Thanks for having me. <div class="framer-mcvrsd" id="