Aaron Levie on why AI is 'total upside' for SaaS and Box's agent roadmap

guys doing? Why are you Why don't you stand up? You can stand up and we can I'm just worried the clips are going to make me look really small.

No, no, we'll zoom in. We'll zoom.

Okay, thanks. Okay.

Uh how's your 2026 going? Uh actually um you know minus the stock market great.

Yeah.

So [laughter] it's um uh you know the the pace of change in AI is incredible.

Yeah.

Uh we are having an inc just you know an insane amount of fun on uh we're building a set of future agents that are going to be able to do much more complex work with your unstructured data

and just the rate at which uh the the best practices of how to do that are changing and the research out there. Uh so we're having a great time.

Yeah. What's uh what what's your process like? like are you actually tinkering with different models yourself? Do you have teams that are dedicated to transformation in AI? Is everyone working on is distributed? How are you thinking about that?

Uh well, the main thing I focus on is just uh the agents we're building, which is a relatively small team. So, you can kind of, you know, you can see them across three rows of of engineers and we just spend, you know, basically 247 pushing the limits on what you can do with an AI agent that has access to your enterprise content. And um and so the things that I've seen uh that that we get excited by are just, you know, we we would never have we wouldn't have been able to do a brainstorm that this would be possible two years ago. Like you it just wouldn't you wouldn't know

you wouldn't know architecturally how you could pull off what we're what we're now able to literally do. It wouldn't have been in the in the sphere of of possibility. So that that's what makes us so excited is um is you know when you look at things like cloud co-work uh codeex obviously cloud code and you see this idea of longunning agents that you know can basically use any amount of tools work with any amount of data they don't really run into the same context limits that that you know we would have run into maybe a year ago

now imagine that for any form of knowledge work with all of your enterprise data that's that's what we get excited by. Do you think the labs are arms dealers today? Ben Thompson posted. He said, uh, what did he say here? He said, uh, in other words, for the last decade of the SAS story, the last decade of the SAS story has been about growing the pie. The next decade is going to be about fighting for it, and the model makers will be arms dealers.

Yeah. Um, well, I I think they they I mean, almost empirically are going to be arms dealers. Um, I think the pie the only thing and I haven't read the the whole piece yet, but I think the only thing I might take exception to is I I think the the markets are are still in positive sum territory

because what's going to happen is you're going to use software for now the labor side of of that workflow. And I think people kind of tend to miss this which is which is if I um if I have software that helps me manage uh contracts and you know we have a lot of customers that put their contracts in box.

Now all of a sudden agents running through those contracts lets us at Vox tap into another form of spend that we couldn't have tapped into before. So that that's just full TAM expansion when you kind of look across all of the different categories of work that AI agents will now be able to go and augment. Um, and you know, to be fair, we're still very early in that trend. So I I understand why maybe Wall Street hasn't kind of fully priced in that dynamic, but we're seeing it within our customer base. So uh that's what gives us obviously the the confidence of uh of this direction. Yeah, there was an interesting take about uh the the fact that yes, you can vibe code a point solution for a specific problem, but uh if you have a database, if you have a relationship with a company, uh there's a reason uh there was one take that I think it was John Gruber was saying like, oh, people will just vibe code all the software that they hate and it's like, no, the reason that they hate the software is because they can't get off [laughter] of it no matter what they want. They can't move to a startup. They they're stuck there. uh you know companies have hostages not customers sometimes and I'm wondering about the value of a database the value of actually being deeply integrated into an enterprise and how sticky that is.

Yeah. Um, obviously the setup of the hostage thing I'm [laughter] I might take some exception with, but uh

Alex, he wasn't talking about you.

No, I I you know, it's very easy to move your files around. So, um, so we have to earn our keep every single day, but um, uh, but I do think that there's truth to obviously the more data you have inside of a system of record, the more, you know, effectively locked in you are. And it has historically been hard to change those systems. But I don't think that would be the my defense of of software. My my defense of software would be that you've specifically defined your business workflows uh in a in a deterministic manner in these systems. And obviously if your workflow is changing, you know, pretty rapidly, then then it would make sense maybe to change a vendor. Uh but for, you know, if you're Ford and you're doing your supply chain on an ERP system, you want that to work the exact same way every single time. You know, the the you know, billions of transactions going through that ERP system,

you cannot take for granted. And so the idea that you're going to go vibe code that is is to me sort of you know not not not not possible or at least very you know not likely. But then the other point is that your your company has a fixed amount of IT resources

and you have to decide what you want to go spend your time on as an organization and do you want to go spend time on you know rebuilding something that the market can supply you and they've seen the best practices thousands of times or do you want to go and build that out with your end of one experience? Obviously, you know, maybe trusting the agent has uh uh has seen of enough examples or do you want to spend your limited scarce resources on building software and building experiences that will make you more money and that will actually be used by your customers. I think you know on the margin the average enterprise is going to spend their time and energy on the ladder. So, interestingly, I I end up in this weird spot which is I'm 100% bullish on vibe coding. I'm 100% bullish that we're going to have 100 times more software, but it that still doesn't yet cross the threshold where I would want to go and build our own CRM system. It's just not worth it relative to all the other things that we can.

How is your software buying process changed or priority set changed?

Yeah, I mean we're um I would say that the the you know we are relatively locked into a core set of vendors. Uh we are going to deploy agents uh on those vendors. we might we might buy slightly toward the agents that those vendors offer assuming that they offer competitive agents and then we'll have a set of agents you know uh we build you know ours for a large portion of knowledge work use cases but we'll use agents from from different kind of you know SAS providers as well and I think the um that's why I I think that a AI is basically total upside for SAS because uh because agents are going to need a system of record to work within there needs to be a traffic cop of what that agent can access and what how did you define the workflow and there's got to be a user that can access an interface that the agent is is providing updates into. So so you still need software for all of that and assuming that you have an incumbent vendor that that remains you know very competitive and very engaged and they're able to build for where the world is going then I would basically bias on existing software that owns those workflows or data. Uh, at the same time, I think there's going to be a large number of new categories that emerge simply because there's no incumbent or because the incumbent is asleep at the wheel and that's going to produce all of these new startup opportunities. So, I'm just generally bullish on software broadly, assuming that that it's from vendors that understand the um the mandate on what they have to build with AI.

What segment of software are you like if you're bullish broadly, is there a subcategory that you're particularly bearish on? because I think obviously everyone's just selling software today. If you sell a digital if you sell a digital product at all uh you're getting sold. Uh but uh obviously I think a lot of people in the industry feel that it's very oversold at this point, right? There's great great companies trading it. I

I think you could probably design this like perfect quotient that captured how like what's the network effect within the software. So how many users are are sort of touching the tool? How much data is being stored in that system and how much gets added?

How many connectors across other applications are there? And then how valuable or mission critical is the workflow that that software is involved in. So if I and then maybe with one X factor of like is that company, you know, sort of priced uh to perfection in terms of what their seat price is and what they're charging their customers. But I look at, you know, those four or five variables and you could kind of look across SAS and I see a lot of names that are being sold that it just is not on the list of things that I think get disrupted by AI. And if if anything, probably they're things that that you use more of as you have more AI. Um, and

when I think of something that like if you if you if you could open up a fresh install of that piece of software and do the same work that that you would have done with

with with something that you've had installed for a decade

and it's the same, it's like that can be replaced. But if you're like, "No, I don't want to open up a brand new CRM because it won't have the history of a decade and and and all my workflow customized and stuff like that."

Yeah, I think that's a perfect heristic. And so so like you know that obviously puts a lot of pressure on if your personal productivity with no network effect with limited sort of data that's aggregated that's a that's a danger zone

and then and then it all kind of is like a you know function of that versus the opposite end which is like your Oracle and it's an ERP system and like your whole business runs on it

and you know what we spend our time on is thinking about okay you know when a customer has a million or 10 million or 100 million documents in box how do we make that data you know a 10 or 100 times more valuable than it was before y

and so we come in from the perspective ive of you've spent you know years in many cases building up your security permissions, your access controls, your workflows and the amount of data in that system.

So our job now is to make sure that agents can run within that environment and add more and more value

um and uh and and as long as we can you know you know deliver that we we believe our position is very defensible.

Yeah. The perfect example would be like uh style transfer for images. Like people were using Lensza, then they all went to Studio Gibli moment. Then as soon as Nano Banana came out, they were like, I got to turn myself into a dinosaur. And it doesn't require any like pre-work or like knowledge of who you are. It's like you upload a photo, you get a photo back and you're good. And then if then if there's a new app, you can just do the same thing because your camera roll is actually where the photos live. And that's

that might be a hard company to go public. Uh yeah, if it was just that, of course, they have a million other things [laughter] going on, but like yeah, if you're just if you're just that it's like, okay, open the fresh app, get the output, and then move on with your day. That's that's risk.

Uh companies of of boxes size, where do you think uh are are generally getting leverage that is under discussed right now? Like is there is is your like in-house legal saying, "Oh, we actually can, you know, adapt headcount planning because we're just way more efficient." Yeah, I I think um my my general take is just what are the what are the fi three five ten things that would have been in your work queue that you didn't get around to that now agents let you go and deploy. So if you're in legal it's you know what are the contracts that you were waiting on that were bottlenecking a deal.

What are the size customer that you wouldn't have supported you know reviewing a contract for because the revenue threshold wasn't high enough to to make the ROI worth it? what is the marketing campaign that you couldn't deliver because you didn't translate it in, you know, X language because it was too expensive. Um, so you kind of go through in enterprise and you kind of look through at all of those use cases and and that's what we're spending our time on. So we use agents to go and do more and more of the work that I just mentioned. Obviously a lot of agents on the coding side. Uh so my expectation is our roadmap should be you know let's say two to three times larger maybe a year from now than it was a year ago. Mhm.

And the reason you can do that is because every engineer should be able to produce two or three times more code. Um and while that's not like you know the most important metric, uh ultimately that does ult, you know, correlate to how much software we're producing. So I think what's going to happen and um uh you know uh uh I I think you're seeing this trend already which is which is we all collectively are going to have much more ambitious product road maps. We're going to all build out way more software. I don't think that will mean that that we charge, you know, an amount that is correlated with how much more software we build. I do think it means more competition in these spaces, but I think many of these markets are still largely untapped. Um, and so that that's still why you have a positive sum dynamic as a result of that.

Makes sense. Uh, talk to me about the huge big models, expensive frontier stuff versus like smaller models or maybe even just uh an earlier model that you implemented in some sort of like minor workflow and then it just like stuck around like transcription. I'm sure you've been doing transcription of documents for a long time. that's gotten better, but do you really need to throw Opus 45 or, you know, 5.2 Pro at it? Like, you can probably leave some things in place like how are you deciding that? Is this showing up in in cost at all?

I think it is. I think that the um it probably might maybe wouldn't be as extreme as as you've set it up. Um but I think the general maybe the continuum I'd argue would be like you've got the Gemini Flash. Yeah.

You know, family on one end and then you've got the Opus family on the other end. That's kind of your continuum. Y

and that continuous sort of moving up over time in terms of capability. So maybe something was a Gemini 2.5 flash uh you know use case a year ago

we'd probably move that to Gemini 3 flash just because that extra two or three points of even if it's transcription or data extraction still valuable and you can now deliver that at the same cost that you could have previously. So it's less of an area that we want to lower cost. It's more where if we can sustain current cost level but add you know incremental value to the customer we'll probably do that all day long.

Yeah. Um I I think we're going to be in this in this kind of you know maybe you know general uh uh sort of point in the curve for a couple of years and then I think you'll see real real bifurcation which is the the stuff that is just fully solved will just get cheaper over time and then the frontier work that is basically where you know you are making a $100 an hour knowledge worker two to three times more productive we will just continue to use the best-in-class model for that work as a as a you know as software companies and generally as a society and I think that will kind of will sustain for you know the next decade. I I don't see that slowing down at all simply because these models just keep getting better and better. Um and uh but but it will be this really interesting biodal effect which is like if you're in if you're a pharma researcher you're going to want you know you're going to want whatever opus 5 is and whatever GPT6 is and and so on and if you're doing some backend you know transaction processing you'll be fine with whatever the Gemini flash of that of that period is and that's how we'll kind of split the costs.

Yeah. What industry outside of tech do you think is the most AGI pill? [laughter] Because I'm assuming you you have a customer that uh that you get on let's say with a large customer catching up with the CEO

and maybe they're in some you know not on not on the coast maybe they're in energy something like that or or uh

financial services and and who who's like as fired up as you are.

Yeah. Uh I I would say um it's a fantastic question that I will totally evade because because I I think it's actually much more sort of firm. Yeah, it's company by company as opposed to like it's a like a specific sector. So [clears throat] uh there's some I mean like you know my bias in general would be informationheavy businesses. So, you know, businesses where a CEO sits around and says, "I just know we're sitting on 10 million documents that have the answer to every single customer question we could ever, you know, ever imagine. And if I could only have an agent go and mine that information, then my employees would be 20 or 30 or 50% more productive." So, so the companies that have that kind of information are are the ones that obviously are are going to increasingly be more AGIP. That's professional services firms. Uh that's consulting firms. That's financial services firms. Uh that's um uh that's law firms in a lot of cases where they're they just know that like we we seem to spend a disproportionate amount of our time redoing the work that has already been done in the past. Why can't we take advantage of all of the information we've produced so that next incremental project is 90% faster

and you know 2x the you know kind of quality of output because it wasn't some new person having to kind of get caught up and then we can go and deploy our time on better customer relationships or creating more value in new ways. So I'll have things like CEOs will call me and say I know I'm sitting on 10,000 contracts. I want to figure out how can I go and sell this new capability to uh to someone based on the the the terms I have in the contracts and um and that's simply an exercise of do you have an agent that could go through every contract and just tell you the insights of you're trying to find a sponsorship for this brand uh and this particular client in this contract you know has the rights to go do that. How do you connect those dots amongst a large data set? Uh those are the companies that I think are are extremely excited. Mhm. Are you seeing any uh any effects of the buildout or bottlenecks show up in your business? Like we've seen the prices of memory spike. Uh Western Digital stock is way up. Like I imagine like you're probably a few layers removed from these prices moving. But is it affecting you or are you worried about it? Are you losing sleep?

Well, first of all, as a storage guy, I I love that they're finally getting their day. [laughter] So if you've been if you've been

giving up for Washington and Seagate, can we give another Can we get another for Seate?

I mean, think about how forgotten these companies and they're just like, oh, there's little circles on a disc and [laughter] and they are now the coolest companies.

Um and uh and so storage is hot. Um and it's actually very funny because like you know we we we are literally in the storage business and for years everyone thought well that's a commodity. In the AI era the data is the most important asset you have. Sure.

So, so the ability to get the right data to an agent is simply the most strategic thing you can do. So, wherever you are in that stack, if you're the infrastructure, if you're the software layer, if you're building

a shame that data is the new oil was kind of ruined.

Yes. The wrong [laughter] it was it was not the oil back then. Can we just bring it back or um

we need to rehab that phrase. we uh um uh yeah so um so I I think now we we actually can justifiably you know sort of say that and actually in most companies when you when you say you know I want to go deploy AI or I want to have an AI strategy usually what underpins that is a data strategy you don't have an AI strategy if you don't have a data strategy so um so that's the um I think that'll be the um uh the story of of you know the next decade uh what was your question

my question was like is it actually show is it keeping you up at night where you're like okay because you sell a service It's not just saying, "Well, I bought this." You're not a Western Digital or Seagate reseller. So, it's not like I just put% on top.

We have pretty locked infrastructure um from our long-term kind of public cloud contract. So, stupid after being locked in.

Okay. Okay. [laughter]

Yeah. Okay. Are there is there a bingo card that I should Did I Are there other things? When do we tell Grind one? If I found

Grind would be good. If you talk about private equity, we'll give

private equity usual things. Everyone we love.

So um okay. So uh so we we have locked in contracts for our public cloud.

Um I would say more the inverse uh in a world of of complete abundance of let's say data center capacity. We would just you know we would be able to deliver even more to our customers because the rate of a the price of AI would just go down. So so you know I want a world of just I I want solar data. I want you know space data centers. I just want everything because that will just mean

there we go. Everyone's pumped up about space data center. Well, read the timeline. You know, you know people are excited about this.

Okay.

You always knew that Twitter would end up in space. You called this, right? That's why you got on the platform so early. That's why you got a million. I need a million followers on the SpaceX social network. Space influencers. I'm buying early.

We are. I'm glad that my tweets are training, you know, the future of of space. So, uh, um, uh, can you just imagine you're like an alien tweet? This is your first access to information.

You run ads [laughter] that then pay for the next space rocket that goes to Mars.

Actually, we're not tweeting enough. We We need to be It's a It's like a oni come, they're just going to they're going to desperately want to learn about enterprise SAS. They're [laughter] like, "We don't care about your religions. We don't care about anything else."

They're like, "Why did SAS uh, you know, what was SAS pricing in January of 2026? Why was it a problem? Um, yeah. So, I just want AI to be super cheap and the cheaper it gets, the the more we're going to use it.

Well, that's a great place to end it. Thank you. Let's give it up for Great. Thank you. This is fantastic. Enjoy the rest of your day.

Do I just leave now?

Yeah. You leave because I'm going to tell everyone about Century. Century shows developers what's broken. It helps them fix it fast. That's why 150,000 organizations use it to keep their apps working. And up next, we have Chuck Robbins, the CEO of Cisco, live in