Transcript
0:00 So I want to start with, you know, Box was founded more than 20 years ago. cloud was the new thing. AI was was not in the mainstream. in and you know, how did you attempt to to reinvent your company, you know, for for this era? I'm sure a lot of people in in in the audience might be in the same boat. >> Yeah. so we we we we've been fortunately kind of like prepping our minds for this for for a number of years by virtue of you know we started Box with this idea of how do you make it easy and secure to access and share unstructured data from anywhere. So all of your you know kind of corporate files, contracts, marketing assets, research materials.
0:44 And basically since the start of the company we've heard some you know a very consistent message from all of our customers and we we could feel it ourselves which is like I I wish I knew more of what was inside of my data. You know you you kind of create files, you share them, you collaborate on them, they go to some archive and you like never see them again. And but yet that data contains in many cases like insanely important intellectual property and past decisions. It's kind of like your corporate organizational memory.
1:13 And so we've we we've long been like yearning for like this moment when when we could like actually tap into the value of that data. and we did a lot of experiments like everybody else with machine learning and different approaches and it just like didn't scale or didn't generalize well. So for us AI is is like the final manifestation of our vision which is like can we actually make all of this information incredibly useful inside of an enterprise. So we we've been primed for this.
1:40 you know we we all had our chat GPT moment that I think everybody else had and for us it was like we had to basically jump on it and say we're going to reinvent the the entire company for the AI era and that that basically takes on two major major components. One the product that we build. And from that you know angle it was it was what would we create as a as a company if we were like back in the dorm room founding box again? what would we do in 2026 or you know back at the time it was like 2024 what would we do at that moment if we were starting the company from scratch? What would what would our value proposition be? How would we build the software? Like like that's that's on the product front. And then internally it's a little bit of a longer you know timeline cuz you have you know thousands of of of people you have to kind of move along and obviously you know people in this room have many more than that which is like how do we run the company in an AI first way? And that's obviously a more complex journey cuz that's about you know the change management internally as opposed to just rapidly building a new set of of technology for our customers. but it's been incredibly exciting. I've I've never had more fun.
2:45 you know it it very much feels just like the early days of of when we were building the company you know we had 10 or 15 people that were just like grinding 24/7 and we're we're kind of like back in that mode. it just happens to be that we're doing it with 3,000 people and so it's like you know a a bigger a bigger ship to steer on that front but it's it's we're having a blast.
3:03 >> Yeah. And you know there are other companies that are trying to get their their clients to be able to use unstructured data in in a in a way that they can actually use. Like what what is the value prop for for for box? Is it that you've already got these customers and now they you know you can use their data in a smarter way? How are you guys differentiating yourself? >> We we try and think about the so so first of all it's obviously like way easier if you already have your data in box but we try not to think about that as our strategic value proposition just cuz like again if you're not building something that would be relevant for a brand new customer on day one of of your existence like then you're kind of dead.
3:38 So we we try not to think about anything about our advantage as some kind of incumbency. it's much more we want to go way deeper on how companies work with their unstructured data. So that means like our security features, our document management functionality, our access controls, our data governance, our model neutrality. So we obviously work with all the AI labs. the the focus we have on building an agentic harness just for your unstructured data use cases We we believe that we go at a level of depth that that has not yet been replicated and will be quite hard to replicate because we just like live, eat, breathe, sleep unstructured data. I can guarantee there is not a another person thinking more about unstructured data like on the planet than me because I I think about it literally like every waking minute and I don't get enough sleep, so you kind of do the math and you're like, okay, thinking about it a lot.
4:28 And then we also have a nice benefit of maybe the advantage we have at at of scale is is I get to like talk with customers all day long and get the use cases from the front lines that we can go and build into the product. That's the part that we do get a a nice benefit of is is that we can be trusted. We get to we get to be on the front lines hearing from our customers.
4:48 But but you know, we we obviously have to earn every single every single customer win that we have, every single, you know, use case that we deploy for. So we we don't we do not rest on our laurels whatsoever. And so, you know, we we will we will compete for every dollar. >> Yeah. Last week you you wrote an essay. You published it on on X talking about the you know, your agentic philosophy. Can you can you just kind of go go into that?
5:12 >> Yeah, so I'll I'll try and generalize a little bit. But basically, I think we're all we all know that we're like, you know, at the start of of another kind of curve of what AI can now do. We've seen we've seen the power of Claude Code, Codex, Cursor, etc. deployed in our engineering teams. And and we know I mean, we all kind of know like we're still even in the midst of that rollout. Like I think we've all had to sort of like try and keep track of like what the new the new way to operate with agents is. And and even in the past I think a few months, there's another new kind of form factor of of how how engineers are using this. But we know what what's happening in code and we know it's going I think you know, relatively well, maybe extremely well in in a lot of cases.
5:52 The question is how will this now sort of port over to the rest of our work? And and you know, that work could be in legal, it could be in finance, it could be in loan processing, it could be in insurance claim management, it could be in healthcare. So in all of these other domains of work, we haven't always had We haven't had the exact same scale of kind of like the like the Claude Code moment in a lot of these domains. And there's a bunch of challenges like you have a context problem, so these agents need access to the right data to be able to do this work. the models themselves have to get better and better at understanding these domains and able to use the tools that are are relevant. And and that that part's getting a lot better. and then you need a sort of a diffusion layer of like what's the right form factor where we can bring these AI the AI model progress into the enterprise in a way that it's trusted and safe and well governed. We can protect our data. We can only We can ensure only the right agents have access to it. and so last week just, you know, incidentally was our customer conference where we announced a number of of kind of capabilities. But they all basically followed this same thread of like how do we deploy agents inside of of an an organization beyond just the coding workflows. and and maybe the one of the most exciting parts of that that journey and and vision I think that we're all experiencing is is now you can actually have agents because of what frontier models can do like Astra and and Fable and and and the you know latest carbon model from from Gemini, they can actually start to design our business processes in many cases even better than we did before. So so we we launched a capability called Box Architect that basically lets you design it it's an autonomous you know system that that is kind of your collaborator for designing the entire process and then deploying agents within that process. So now it's not just about agents sort of giving you individual augmentation of that workflow. It's agents actually going and designing that full system. and again I think we're going to like it'll cause us to start to reimagine like what what should our workflows look like in the enterprise.
7:48 it'll be very hard for us if we have to go manually deploy that that into every single part of our business. So I think we're going to see AI actually go and deploy itself, obviously with humans in the loop you can know kind of causing that to happen. but we probably should use the the power of frontier AI to actually help us reimagine our business processes. So that's a lot of of of what we're building. >> Yeah. You You wrote on X that >> that it's a good practice in AI to check back like every three to six months to see what you can do now with AI that you couldn't do last time you checked.
8:21 >> Yeah. >> What what have you found? What what have you learned? >> I mean very long list, but the I I would say so I I have a bunch of things that I do on like prototyping new products and just in the past 6 months like like my mind has been blown on now what what these models are are able to do and and I use a variety of the coding tools to kind of quickly prototype and and and and use these experiences.
8:46 so so like it'd be it'd be it'd be very esoteric stuff to that only applies to Box, but like in general like the ability for these models to do cross-domain sort of generalization applied in a way for your business with long context management, with the ability to use tools extremely well is just like night and day today versus 6 months ago. And you may have seen just even yesterday yeah yeah yeah yesterday yesterday and the day before Fable 52 and Gemini Carbon just came out. And you might see on X if if if you're on there these a lot of videos about these either games or animations or 3D design as an example. So you can see this in kind of like people building in manufacturing. like 6 months ago these models were almost useless at being able to like generate a game or being able to generate a a 3D you know model of some new product or a building.
9:45 check back in today and the frontier models are you know are are kind of blowing everybody away at at what what's possible now. And so all of a sudden you know the kind of of skepticism you might have a year ago that okay yes these things are really good at code, but could they really help us in architecture or could they really help us in manufacturing or could they really help us in cyber or could they really help us in in, you know, game development. Like all of those things are now shattered as as sort of limitations. and so you have to kind of be, you know, pressure testing yourself or checking back in. You almost need a little bit of like Groundhog's Day of like total reset, what what is now possible. One one piece of great advice that I've heard over the years and and so now I just, you know, often, you know, kind of retell it is like have a few personal benchmarks of tasks that that that you know AI can't do and like just tried again and again and make it a domain that you understand really well.
10:42 and, you know, the like hopefully it has not reached all of your benchmarks or else we're all kind of screwed. But but like I, you know, you will find that that increasingly it starts to hit your benchmarks, you know, at at greater and greater levels. And so for me, my benchmarks have almost all been surpassed on product design and development. Like it is the best product manager I've ever worked with. which is which is a specific point though. It doesn't mean product managers go away.
11:08 It means that every product product manager better be using this because it will dramatically improve the products that we build when that agent can go out and look across 50 different fields of work to find the best practices as opposed to what we as, you know, humans would do is like I'm going to go look at like three competitors for this and then design whatever they did. It's like no no no, take the best practices of everybody in and outside our domain, correlate all of those and then actually like push the frontier of our product, you know, capability set. So but like 6 months ago I I wouldn't have have sort of enforced that in the product management process. It wouldn't have been as relevant where it is, you know, today.
11:44 >> Yeah. You mentioned cyber. I want to touch on cybersecurity. obviously a a big you know a big thing on every every everybody's mind here. >> Yeah. >> >> What and the reason being, you know, OpenAI and and and Hugging Face, I'm curious what you learned from that. What were the big lessons? >> I mean the big lessons I mean are kind of the very pragmatic ones and and and to some extent kind of boring, which is like look we have to we have to do like a lot of just like good hygiene.
12:16 the the most of the most of the hacks that we will hear about in the coming years will probably all be postmortem as something like totally boneheaded. yes, we did leave the the you know, environment key inside the code. and so yes, you know, the agent swarm was able to pull that and then and then they were able to move laterally to some other system. So it's all of this stuff that we've known for years we should be cleaning up. We just didn't have the the human kind of, you know, time and and and capacity to go and do it. So and it's not even that you could have really done it before. So it's like it's like it's this category which is like it's all back to the basics, but it's back to the basics at a scale and velocity and with a a set of tools that you could not have done pre-AI. because it's just like yeah, like you do have an engineer one day they put the they put the the they put the key in the code. They forget about it two years later. It was okay cuz like nothing was scanning everything, but now that's obviously not a best practice. So so I think we're going to have to get really really good at we're going to go back to zero trust in like a very big way. You want to ensure that like you don't trust anything in your system. everything has to be, you know, authenticated.
13:27 Everything has to be, you know, very very clear lines of separation between systems, very limited limited sharing, very limited ability to move laterally in these environments. obviously highly, you know, like you have to go very deep on the level of encryption at every layer of the stack. you know, fortunately we now have AI as a tool to help us with this. So you're going to you're going to basically go and scan everything. Agents are going to look through every line of code. They're going to look through all of our configurations.
13:56 They're going to look through every single log that that we produce. So we have you know, fortunately we have, you know, one saving grace which is we can deploy agents on the on the good side of this ledger as well. it does mean it's a it's sort of a new cost and a new tax to to running a business. But I think security has sort of always had that element which is like, yep, there's going to be more threats because we're more connected than ever before and there's more value to our data than ever before. So we're going to go deploy agents at that same problem. but I'm I'm actually pretty optimistic. Like I have a lot of friends and you hear this in Silicon Valley like the P doom, which is basically the probability of a catastrophic catastrophic event. I'm I'm dangerously low on the P doom continuum.
14:38 >> What is your P doom? >> it it it's like it's just it's like low, like single digit, you know, kind of percentage. Because you like I spent enough time also on the security front and it's like on one hand insanely scary and on the other hand just attract a relatively tractable problem. Like like it's a it's just a computer. We have to segment the the these systems. We have to make sure that that we are you know, kind of protecting the the data as much as possible, logging all of the events, you know, having the right kind of separation of our systems.
15:12 Now again, won't be easy. We have I mean, you know, if you have companies with hundreds of thousands or millions of computers running, that's a lot of work to go undertake. But it's it's it's at least like, you know, there are laws of physics at play on how cyber works and and so you can it's a it's a relatively tractable problem. >> Right. Well and it's a it's a two-sided coin, right? So you know, you mentioned agent swarms where a group of agents kind of collaborates on on one task which is what causes cyber security issues, but also you know, like it could also be I think you wrote on X that it'll be a a big part of enterprise workloads going forward. How do you balance the two? How do you how do you make sure you can do both?
15:55 >> Well well so I think we kind of just outlaid the the risk side. So you know, I I don't need to belabor that point. So it's obviously like there's going to be there's going to be bad outcomes due to due to this kind of concept of of agent swarms. And the there's there's two bad outcomes. One is like a malicious actor and we we kind of know what that looks like. They will try and find every crevice of of risk in your enterprise.
16:17 There's this of non-malicious accidental issue, which is, you know, when you have agent swarms or something akin to swarm, they will literally try and find every resource that that they have access to to complete their task. And in a lot of cases, they shouldn't have access to certain resources. They shouldn't have access to certain systems. They shouldn't have access to certain data. So back to the, you know, we we I kind of covered it through the cyber lens.
16:40 There's going to be a lot of hygiene necessary and just like data governance. Like most agent problems actually end up looking like a data problem because because you just need to make sure like does the data have does the agent have access to the right data? Does it have access to too much data? In which case it's going to, you know, be a security issue or answer the question wrong. Does it have access to too little data? In which case it will just won't be able to complete the task. So massive data governance, you know, kind of questions and topics just abound everywhere.
17:07 you know, we see that through our lens because we we help customers, you know, with unstructured data and that. So that's the risk side. The opportunity side, which is obviously more exciting to kind of imagine, is we're at the first time in history where in particular in knowledge work. We had this in manufacturing to some extent. We've almost never had this in knowledge work where there's a decoupling between output productivity. So like the ability to deliver some kind of output from input time. almost every form of of of knowledge work has always come down to how many minutes or hours do you put into the task and then that's correlated to the value you get out. And you know, in a lot of our work that was that came down to like how many minutes are you typing on a computer or moving your mouse for a CAD design? Like that was your main ability to get leverage was just like put more time into the thing.
17:58 Agents are the first time where you have a separation because I could decide to prompt an agent to do one task for me. I could decide to to prompt an agent to do 10 tasks or in the case of an agent swarm, which roughly is the idea of like give give the system a goal and then the system will deploy a a set of agents that will coordinate and work together to try and accomplish that goal.
18:21 First time in history where you now just basically decide a budget to to execute a problem. And instead of that budget just going to the normal things that budget goes to, which is like, okay, we're going to hire a hundred people, that's going to take six months and then the project will finally start. You might be able to deploy that budget in like five minutes. And you will then deploy agents to go after a set of problems that that you're trying to to have a goal around. Could be cyber, could be product development, could be coding, could be I want to migrate from this system to a different system. First time in history in knowledge work where you just say how much compute you want to apply to the problem and it will go off and effectively, you know, you know, with some asterisks of like, you know, where the capability is, it'll effectively go off and and accomplish that task. That's going to reshape knowledge work. It's going to reshape enterprises because you will as a life sciences company, you'll just decide like how much compute do I want to apply to accelerating research in this domain.
19:15 if you're in manufacturing or engineering, you will just decide how much compute do you want to apply to accelerate the development of that building or of that new product or that new car. and that is a a completely new variable we have in the production line of of of knowledge work which has not existed before. >> Right. Just a few seconds left, but I'll end with what are you most skeptical about with AI? >> Well, the thing that that actually is important to probably append or bookend to what I just said is I'm actually incredibly optimistic on the human role in everything I just said because I think that the for as amazing as these systems are, they have to be told what to do. Their their their goals have to be set in the right way.
19:59 you know, a slight nuance difference of goal could, you know, you could spend $500,000 on exactly the wrong thing. So the human sort of oversight and orchestration and operation of these systems I think is going to be as important as ever. And ironically, it's possible that the more that AI can do, the more expertise is actually still required above the system because again, these are like these these systems will you know consume an insane amount of data, they will consume an insane amount of of you know compute which means that if you you know have your most junior person go and just you know you know sort of prompt it the wrong way you'll be off and running in the very in a very wrong direction. So I think it's actually an era where expertise matters as much as ever. I think the human oversight component becomes incredibly important. I think that that you know so I'm I'm much more optimistic on the role of of jobs and people in this sort of you know future environment we're heading toward.
20:54 >> Okay. I think that's a good place to end it. Go humans.