We also don't really have these kind of core sources of truth for data in most enterprises. Like, you know, where is the contract for that one customer stored? It might be in five different places. That's five different chances for the agent to get the wrong one. That is heavy, heavy human work. There's no way to automate it. You have to have people go in, understand the workflow, understand the data, build out evals. I don't—I think we need to not be waiting around for the one agent to rule them all.
So Aaron, nice to meet you in almost person finally, and it's been a while that I followed you from the early days of being a founder. And I'm looking really forward to our conversation. And I actually have a question that I'm thinking of for a long time. So organizing company knowledge in order to make it workable for agents seems to me as the key question. If I compare it to the industrial revolution when you had a steam engine changing to electrical engine and people were playing around with it but not really realizing how much you can reorganize and how much you have to reorganize—and I know you have a pretty strong opinion on that—and I would like to start right with that.
Yeah, I mean, I think this actually cuts to one of the core questions of our kind of agentic era that we're entering, which is how do you get the agents to participate in the workflows that you're doing, and how do you get them access to the knowledge that they need to be effective in any business process you throw them into. And if you just kind of keep doing things the way that you've always been doing and you just kind of drop in agents into your existing processes, oftentimes you're not going to get the upside that you would expect because there's just a bunch of, you know, kind of challenges inherent in agents. Like they're very forgetful. They have to be reprompted with exactly the right context to be effective. They need access to a whole bunch of resources in your organization to be useful. They need to have very, very kind of well-documented sort of understanding of your process. So these are a bunch of things that people don't have to have. You know, people—you can kind of throw them into a new problem and very quickly they can kind of intuit the behaviors of other people in the organization, what's the style of work that we do, they can, you know, relatively easily find what they're looking for by asking a colleague or kind of spending a bit of time navigating the corporate systems. Agents don't have a lot of those benefits. They're just going to kind of go try and find what they're looking for instantaneously, come back maybe with the wrong thing and run with that. And so then the big question is how do we get agents the most important kind of data and knowledge that they need to be able to be effective? And so I think knowledge management and information management in the 21st century will become one of the kind of core raw ingredients of this industrial revolution that we're entering.
Who do you think gets it right at the moment? I'm thinking of the first ChatGPT moment and Microsoft came out with Copilot right afterwards and saying, 'Oh, we got your SharePoint, we know what we're doing,' and then you were trying like, 'Okay guys, this is not how it works.' Who would you say, three years into the ChatGPT game, who got that?
Well, there are sort of different kind of parts of the ecosystem. I think there's sort of companies that are trying to deliver on that vision as closely as possible. So, for instance, what we're working on at Box is make it really easy and secure to get the right knowledge in your company to agents, again, in a very kind of secure and well-governed way. So we've been working on a bunch of technology to really kind of help with this problem. Then on the other side there's the sort of, let's say, the practitioners or the users of the tools. And I would actually say like generally you can go and look at small startups to sort of see what are the new practices that those startups are running with or using to be able to go and work in this kind of more agent or AI-forward way. And what they're doing is these small startups are kind of running their businesses entirely differently from how we would have run a company five or ten years ago. So first of all, they're sort of running the business with agents as first-class citizens in their company, which means that they're documenting everything. They are making sure that all of their decisions, all of their corporate knowledge is built out in such a way that agents can on day one have access to that information. They're constructing their workflows in a way that agents can natively participate in them, as opposed to them feeling like they're sort of an afterthought or add-on to the workflow after it's already been kind of designed for people. They're sort of incorporating the right ways that humans now need to be participating in processes where they have to review the work that the agent is doing, where the agent has to be kind of effectively making sure that you're looking through all of the output that the agent produced and then incorporating it into the rest of the workflow. So they're designing their workflows in a sort of an AI-first way. So most of the lessons are coming from small startups that are kind of operating in this way, and then the big question is how do big companies end up operating like these startups are, given how much change management is required, how much scale these companies have. That's a much harder problem than I think a lot are anticipating, but you know, we're—I think everybody's kind of collectively on this journey together to go and do that.
I can pretty much remember the day when you were a start scale-up of like 70 million revenue. Now you're 1.2, 1.3 billion in revenue. So you're a huge company. And as I know you from the outside, I would assume you play around on the weekend, you test stuff, you sit there at night and you go into that—these things, but the scale is much different. And how do you—how would you start in transferring and what did you do in playing around making Box an AI-native company? Or did you shy away and say like, 'We will never become an AI-native company, there are other companies out there'? What's that state at the moment?
Yeah, no, no. We're—I'm betting everything on us being AI-native. And so that means that every process in our company has to be positioned for being AI-first as an organization. So everything about how we run our company, how we operate, has to be delivered in an AI-first way. You know, that takes on different characteristics. So in engineering it's relatively straightforward because we are all kind of collectively learning the best practices where you're going to have maybe Cursor or Claude Code or Codex, and every engineer is going to be deploying agents and then some of your agents will be local and some of them will be in the cloud. And you need a kind of modern kind of AI-first product development lifecycle, which means you need a workflow engine for these agents to participate in that have access to your product resource documents and your design files and your codebase. So we've had to do a lot of work to kind of design a modern agentic kind of coding system for the company. Then we have to make sure that all of our kind of the more knowledge worker tasks in product management or go-to-market or sales or solution engineering—how do we make sure that all of our Box employees have access to agents that help them work as effectively as possible? So agents that can access all of the kind of corporate data in our enterprise. Agents that help you automate account research for customers you're going after. Agents that help us accelerate our product campaigns or our marketing campaigns. So that's a lot of the rest of the work. And I think mostly what we're doing is we want to encourage everybody to use AI as much as possible, as long as it's kind of practical in the use case. We want to make sure that we keep people in the loop with everything. So one thing that we've been very clear on is that you still are accountable for the work that the agent is producing. You can't kind of relieve that accountability and say, 'Oh, no, the agent did it.' We're also very—I think biased toward the human-plus-agent collaboration modality as opposed to just move to everything being agents and kind of get rid of the people. We want to make sure that we find a way to augment the work that people are doing as opposed to kind of really think about replacing that work with agents. So we really think about it as a force multiplier of the people side. So you know, we haven't looked at like how do we cut jobs with agents. We kind of do the opposite. We sort of say how do we get more output per person and how do we just make sure that we're getting more done as a company as a result of adopting agents.
How do you—like, I mean this is much more, as I also see it, a compute revolution instead of like doing communication and a little bit of efficiency. What you're describing is how do I leverage the intelligence of a person through an agent on top, and this is much more difficult in my experience for people to adapt to because it's so different regarding the job. So let's take an example. On the way here, of course, I prepared with Codex. This is my usually go-to spot because I like to code, but I also like to do knowledge work. And I find it a very, very reliable agent. And there are certain things I can delegate to it and know like that'll happen. And then there are certain things I have like feedback when I really want to review it and say, 'Okay, you double-checked it.' Most of the time the problem is finding the right source in the vast tool set that we also have that many other companies have. And I know that you were always saying you're the security and government slayer. How does it practically work? Because what tools and on what level the rules can clean up the mess that many companies have out there? All the Excel files, all these dead artifacts—they're so important in order to get a good outcome. And this is to me one of the key question marks that I usually have.
Yeah. I mean this is—you know, this is sort of one of the big problems for an enterprise is a large amount of their data, as an example, is fragmented across, you know, 10, 20, 30 different systems inside their enterprise. Agents don't easily have access to that data. The protocols don't work well. The systems are on-prem so they don't really kind of connect to MCP server. I mean they can't really kind of show up as an MCP server very easily. They're not built for, you know, agents to go and kind of navigate very effectively. So our data is fragmented. We also don't really have these kind of core sources of truth for data in most enterprises. Like, you know, where is the contract for that one customer stored? It might be in five different places. And so if it's in five different places, that's five different chances for the agent to get the wrong one, the wrong version, go off of the wrong answer to a question. So we need our information to be—we need better sources of truth. We need more authoritative sources of documents. We need to move this data to the cloud effectively. So that's a lot of the work that has to be done if we're going to be able to use agents across our workflows in knowledge work. That's actually why we're seeing so much demand right now is more and more companies are realizing they have to go and modernize their systems and be able to move that to the cloud effectively.
Do you have also—I mean the famously forward-deployed engineers for like doing this for Box customers, so they are with clients and doing the cleanup of the data and stuff like that. How did that job then change compared to—
Yeah. So we do have our equivalent of FDEs and what they're basically doing is they're going into an organization. They're saying, 'Okay, let's understand where the data is in this workflow.' You know, we go to a bank and they have a bunch of customer documents for getting loans processed or onboarded in a wealth management relationship. There's a bunch of personal documents, a bunch of loan documents, a bunch of bank statements that they need to go in and get into the cloud. Well, that data might be fragmented everywhere. So we have to first help the customer understand where it is. We have to find a way to kind of map it from the prior sort of architecture and move it to a modern architecture. So we have to kind of get it into the cloud. We have to get it in a data-ready, in an agent-ready data format. Then we need to actually figure out what models are going to work well on this data, how should we configure agents to work well. We have to do a lot of eval for that customer. We have to help them sort of figure out is this workflow actually going to work for you inside your organization. So all of that work has to be done. That is heavy, heavy human work. There's no way to automate it. You have to have people go in, understand the workflow, understand the data, build out evals. That just takes a tremendous amount of work. And that's why you're going to have humans go and get utilized for just all of these processes in the future.
But how do you see it when—I have many questions regarding that. Let's start with one that is—a new model comes out, in that moment you need someone and you need the skill set in order to evaluate again the new model against the data, against the system prompts you might run in your system. So who would do that then? Customer side, or do you see that as a service in the future that will be continuously there?
I think vendors will have that kind of talent and they'll help support that. And then they will—you know, certainly vendors will lean in with their version of FDEs. You'll probably have systems integrators that have a role for that as well. So they'll have a bunch of talent that is sort of FTE talent as well. And then individual customers will also probably have some of this homegrown within their IT organization. So the IT organization of the future will often be really kind of these FDEs in the organization and their job will be to constantly kind of maintain and curate these agents for each of the lines of business to make sure that all of the workflows and the agents keep running effectively across the enterprise. So I think you're going to see a lot of interesting roles emerge inside the enterprise to help deploy and manage and observe and improve the quality of agents inside of organizations.
And what about the a little bit juicy topic of building, let's say, relationships to the providers that have models that are maybe locked off for the public but are available to a certain amount of people. Is this the new kind of PR relationships that has been built or what do we see here? What do you think? I know you have a strong opinion on that one.
Well, this is a very—it's a very interesting question because I think there is a little bit of a debate—is, you know, will there be these kind of trusted access layers to get access to Claude or to get access to GPT 5.6. I think there's an argument that that's one scenario that plays out. Alternatively, there's a lot of value in this idea of kind of being a model router where, you know, sometimes maybe it's frontier model access that you want and other times it might just be an open source model that you want to make sure that is kind of tuned to your particular workload. And who's going to be in the best position to be able to give you access to an open weights model? It's going to be the abstraction layer between the workflow and the model. It's not going to be one of the model providers. It's going to be that kind of routing layer between the actual business process. And so that's—you know, that's the layer that we provide. That's the layer that Cursor provides or Factory or Cognition or Salesforce or ServiceNow. So that's going to be a lot of value that that kind of applied AI layer will deliver.
Yeah. No, I get that and I can see that, and it's still new. It was definitely new to me that a model has been rolled back and hold back. I was like, 'Okay, this is a bit awkward.' And I told our clients in the end, the stronger your layers in between are—like the context, the prompts and so on—the more independent you are. But still, it has a little bit of a taste that this happened. And it also dawned to me, okay, there is a new age of models because there is something to—I always see it as a triangle. You have the model as the engine, you have the context, and you have the prompt, and it all needs to be in balance in order to provide something. But there are certain moments when the model becomes so strong that you sit in front and you might have had this too in December when everyone thought like, 'Oh wow, Opus 4.5, right, this works finally now,' and this is different to technology we had in the cloud age. And if there are now companies having a big advantage, do you still see that as startups running behind you can still catch up, or would you say this is a moat that is so strong it's hard to catch up? Because these companies are partners on the same side, right? Like they're partners, providers, and mix it up.
Well, sorry. The moat being so strong on which side?
The moat being so strong on the provider side, let's say OpenAI with 5.6 or Anthropic having such a strong not only responsibility but also moat saying having a stronger model and providing that to certain players but not everyone.
Well, I actually think that the model providers probably would like to make sure their models are broadly accessible. You know, if you were to talk to Sam Altman right now, my guess is he wants to make sure that GPT 5.6 is massively available because anytime that they constrain access to their model, that's just less demand that they can kind of go and have driven. So I would say the models want to be freely available, but the routing layer is what can actually understand the workflow the best to be able to figure out what model works best with the specific business process. And so the reason why the routing layer adds a lot of value is because there's a lot of scenarios where maybe you want Claude 5 for really kind of being a planning agent or an orchestration agent in a workflow, but then you want GLM 5.2 for doing heavy token processing of a lot of the unstructured data in the workflow. So in this case Anthropic is still going to generate quite a bit of revenue because Claude will be involved, but by token volume more of the tokens will flow through an open weights model and the routing layer will be the reason why you can choose different models for those different tasks. So I think this is going to—this idea that as an enterprise you probably want to have some kind of abstraction between your knowledge, your workflow, your data from the models, so that way you can always swap in the different models, and the routing layer basically will provide you that abstraction.
If you—but still with this router, let's think that through. And I think you were the one when Uber had the headline out, 'We cut back on Claude Code because it became too expensive,' and I think you mentioned, 'But don't forget, they're levering up on Codex now.' So you cannot give up these models, but you need to budget it. You need to make a decision of what value do you get out and how do you teach people on the one hand to make this decision to have an idea for the value, and how do you budget it next to the IT budget, for example, saying, 'Okay, we have to see it in the controlling at least and plan it ahead a little bit'? This is totally new.
Yeah. No, the ROI of agents is this really fun topic because, you know, how do you measure the ROI when it's a brand new technology, it's changing very quickly, it's being deployed in like sometimes a diffuse way throughout an enterprise where you can't measure every single person's kind of use of AI at all times. A lot of it is sort of obfuscated and not something that you can kind of just go and deeply understand. So in general, what you have to do—what that leads to is you probably need to have different org owners have budgets for AI like they have budgets for anything else in the business, right? And at the end of the day, you have to give your head of engineering a budget or your head of marketing a budget or your head of sales a budget, and they have to use AI. They have to think about AI as one of the budget line items and they kind of have to be responsible for how effectively do they want to deploy AI in their organization. There's almost no methodology in the history of management science other than having accountability be driven by the team responsible for the delivery of the thing that you're trying to deliver. So if you want AI to accelerate your sales team, you want your head of sales to be the one that owns how much should they be spending on AI. You don't want a different organization owning that decision because then you kind of remove the accountability from that org owner. So I think the reality is what's going to happen is the IT team is going to own managing the technology. They're going to own kind of bringing in the latest version of it, the best-in-class version of it. That's their job. And then the management and effective use of the technology will go out into the organization. So the head of engineering has to figure out how are they deciding on the ROI and the productivity that they're getting from AI. That's not the CIO's responsibility because the CIO also can't be responsible for how effective is your sales team on the people's side. And AI right now is just an augmentation of the people's kind of ability to get work done. So it's going to be this kind of joint responsibility. The tech has to be owned somewhere, but then the accountability of the ROI really has to be coming down to the business owners who ultimately own the outcomes of those job functions.
Do you already see Fortune 500 changing in that direction, or is this something more close to you and anyone in the Valley at the moment?
I think it's—I think it's early for the Fortune 500.
Has generally seen this as an IT technology initially.
Yeah. Yeah. But now is starting to figure out that no, this is really a technology that augments people, so that requires that kind of joint tie at the hip dynamic. But yeah, I mean, we you can't hold a department responsible for something that it can't see the entire workflow through. And so this is going to, I mean interestingly, it's going to put the IT organization in a more powerful position because the IT organization is now responsible for the outputs of and the sort of leverage that the entire company has. Because if the IT department brings in technology that's not good, then obviously the whole AI strategy is going to fail. Conversely, if they bring in technology that's amazing, then the company could run way faster and way more effectively. But still, the budget decisions, the ROI decisions have to come from each individual organization.
Which makes perfect sense. And let's stick with that example and say the IT has a decision saying okay, we also have to come with something that is secure. And then you come from the departments and they, let's say, started with cloud code or cloud co-work or something, but from the IT side it has been the tool of choice. And now even further you come with claw tech and it even enters your Slack account, and for normal people at the moment it's not easy to overlook what is happening. You definitely know what is happening and how strategically important this can be. How do you balance these two sides that are racing up the... Yeah.
