AI Will Change Jobs Faster Than You Think — Aaron Levie (Box) & Michael Grinich | init() 2026

We believe so, yeah.

Yeah, so, for us, we had... We already had a pretty, sort of, sizable focus on AI going back, like, six or seven years prior. So, if you remember, there was this moment in, like, 2016, 2017, 2018, where vision models got to the point where they could much better classify information. You could do much better OCR. You could detect what was inside of images. And so, we had a whole team focused on, basically, how do we bring vision models and, you know, kind of very early language models to all of the data in Box.

But the massive flaw was you had to have a model that was trained on each specific problem that you were going after. And, obviously, it was just way too expensive. So, a customer would come and say, hey, I'd like you to look through all of our images and classify them all. And, you know, you both needed a proprietary model for their kind of data. And it was, you know, millions of dollars to be able to go process all that data. So, we kind of shelved it. Didn't really take off. But we were, we had our kind of mindset in a spot where we were ready for, you know, when the time would come for AI to actually work and be cost effective and be able to work in a more generalized way.

So, ChatGPT emerged. Honestly, you know, to some extent, I kind of feel bad because, like, we could have probably identified it six or 12 months prior just by looking at some of the early use cases of GPT-3. But you needed, like, the ChatGPT form factor to kind of pull it all together. And it was just very obvious. Like, this is finally the form factor where an end user would be able to go and ask questions of their data. And then the only question was, like, you know, how do you get the right information? How do you get the right context into the model to be able to answer the question? And so, that was mostly an infrastructure project of, you know, at the time, RAG. Now, it's been able to evolve, obviously, to much more tool use, you know, type capabilities. But that was, we just jumped on it right away within, you know, about a week and a half. And I wouldn't say pivoted the company, but certainly started to build out the kind of core of the next gen of our platform.

Without psychosis, though.

I think I know because, like, when I'm talking to non-AI people, I don't sound... Well, I guess I wouldn't even know that, but, like, I think I don't sound totally insane. So... But you're right. I don't 100% know, but I'm, like, extremely confident I don't have AI psychosis.

That's right, that's right. That's exactly what you would say.

I'm pretty glued to X. I'm sure your target demographic is very glued to X. And I have... I'm very proud of my feed. I have it so dialed in. The For You tab is just so dialed in a way that I think I could go head-to-head with anybody here on how strong my For You tab is on information. And there's just a bunch of tricks. Maybe you already know this, but, like, if you, like, share the right tweets... And I don't even share them with anybody. I just pretend to share them. And then the algorithm is like, oh, this is really what he wants. So you share. You like the right things. You do the right, you know, kind of massaging of the feed. And so my For You tab is super wired in. That's how I get probably 80% of the news and information.

We have, obviously, a now pretty sizable AI team internally that's constantly working on, you know, breakthroughs within the domain that we're focused on, on unstructured data and AI. Fortunately, by being in Silicon Valley, you're kind of always running into research in some way, either directly or indirectly, in some kind of dinner conversation. So you're getting some kind of downstream visibility into, you know, how is RSI going, and what are people worried about on AI safety and alignment and these types of things. And so I guess I try and just calibrate it. Like, you know, taking all that, triangulate the different information sources, and then... and then build a view of what I think is happening.

And then my calibration comes because, as probably you do as well, most of my time is spent by volume of hours with customers in the real world. And that's your instant way of now gut-checking all of the stuff that we're seeing online versus now the real insurance company, the real bank, the real life sciences company. What are they actually doing? How much progress do we actually have? What are the capability gaps? What are the diffusion gaps? And what is it going to look like to actually roll this technology out at scale? So that's why I mostly confidently say I don't have a psychosis. If I was only in the online version, I'd probably go crazy because I would think we're in some takeoff. And then I go to the real world, I'm like, yeah, man, these people are just still in meetings and talking to humans. And so we still have a lot of work to bridge the stuff that we all see online and see in Silicon Valley with the rest of the world. So that's mostly what I do. And then I'm just constantly playing with nearly every new tool that comes out. Just to give it a little bit of... Just to stay current with kind of where things are.

I try and play with, let's say, 75% of the new things that we all see. Instinct comes out, you play with it. Muse comes out, you play with it. Grokbot comes out, you play with it. But I haven't obsessively automated my life or something. I have friends that do that. It's cool to hear their stories. But I don't get it into that level. I treat it a little bit probably more academic. Like, let's play with the tool. Let's see, where's browser use at this point? Where's computer use? How are things going on the capability progress? And then I try and kind of reflect on the things I'm hearing from customers. Does this finally unlock this new use case that wouldn't have been possible before? And I think it's kind of the probably informal routine I have at this point.

Yeah, and do you mean personal agents as of the past 60 days?

Yeah, I think it's going to be really interesting on this as a form factor. What is the idea of this kind of persistent agent that has its own goals, kind of identity, access controls, sort of universe it lives in? And I think there's actually some argument that it'll actually work better in our personal life than in the enterprise in a weird way, because in our personalized lives, there's potentially less risk of sort of like, you know, breakage of the barrier of information of what people should have access to. So I can have a Muse agent or an Instinct agent or Grok bot that I give it tasks that go run and do them. And I'm basically in control of what data I give it. So there's a kind of limited blast radius.

But for that same agent to be useful in an enterprise, it probably needs to talk to other colleagues. It needs to be able to gather information from other systems, which then has some degree of exposure of more kind of, you know, cyber risk and other kinds of challenges. So weirdly, we might be in one of the first use cases that could work better in the consumer world than in a professional context, because of the sort of parameters on data access and controls. But I'm very excited by the space. I think I would be less confident in it two years ago, even six months ago, maybe, because I didn't think browser use was at the stage or computer use was at the stage where it could handle the long tail of tasks. But now you can kind of squint and you could imagine one to two years from now, like probably any arbitrarily, you know, maybe not any arbitrarily complex task, but like any like five to ten step real world task. Like it's got to make the phone call. It has to talk to a person. It has to gather information. It has to do another two steps. You could probably pull that off with where things are going.

So then the interesting question is like, what's the scaffolding we need to build out in the economy for that? Well, you know, how do you do the identity, you know, kind of and security handshake between my agent and a commercial vendor that that agent is trying to work with? We saw some protocols, you know, just the past couple of days, I think to try and support that. I was just with a bank this morning and, you know, the question is at what point should an agent be allowed to do a wire transfer? And it's a really interesting question that three years ago was not really on the table as an interesting problem for a bank. But now it's just incredibly relevant. Like should a bank to... Let Muse do a wire transfer. You know, is the capability ready? Do we know how to fully trust that that hasn't been something that, you know, kind of stole a token or session, you know, somewhere? Like, how do we make sure the security handshake is good? How do we, how do we do the, do we have the right form factor for the notification that the user should get to do the approvals? Do we have an industry? Like, there's a whole bunch of industry standardization that's going to be needed. Like, like, I don't think we've got great uniform universal, you know, cards for interaction, interacting with, with, you know, how do you verify that something's happening? Like, we have lots of stuff to build out as an industry. But I'm quite excited because it's like, obviously, incredibly, you know, exciting technology, if you can just tell an agent, go do some real world thing for me. And, and then it just, it just gets executed. Like, that's, that's all completely net positive for society.

Maybe, um, I don't know, I wouldn't, I'm probably not that, I'm not probably not that consumer built. So I think that, I think AI, by and large, is an industrial technology. I think that if you, if you think about like, where can intelligence be best monetized? Where is intelligence sort of most of a scarce resource? Where if, if applied, could it have the greatest kind of upside and outcomes? It's like, just like, inevitably, inevitably, in an enterprise context. So, so I think we've got the right kind of ratios of like, you know, token consumption in the economy, I still think will be 80 to 90% enterprise workloads, because that's who, you know, could literally spend a trillion dollars a year on, on compute is going to be enterprises that want to do drug discovery faster, that want to be able to have much better experiences, you know, for their banking clients, that want to be able to go and automate more manufacturing. I just think by volume, that's going to be so much more impactful.

I don't know that we need a consumer breakthrough for, for the enterprise stuff to take off. I think the enterprise stuff taking off is actually much more of the blocking and tackling of a decade of diffusion work that just has to happen inside the world of the enterprise, which is like, you've got to change the workflows, you got to get data into an environment that agents can work with, you have to make sure the access controls and data security are all put together in a way that they'll let agents get access to the right information. You have to figure out where's the human supposed to be in the process? You have to figure out like, what do we do when the accuracy rate is 95%? But we don't know what the 5% is like, so, so like, how should we do the loan, the loan workflow or the drug trial workflow when, when, when, you know, it's extremely, you know, great on the evals, but the real world is, is just, it's just existential getting it right or wrong. And we don't yet know, you know, which again, you know, in, in the non, you know, verifiable environment, because you're now in a real new net new use case for the customer, you know, is it accurately performing its job. So like, that's all the work that I think we have to do for the rollout.

I think the consumer thing is kind of just, I would just say, a different category altogether that has a lot of powerful use cases. The one modality that might be shared is this idea of, okay, like, like, maybe, maybe I do want an agent that is sort of always on and always running, can perform these complex tasks for me. But I would argue for a lot of people, this is going to maybe sound like I'm missing something, but like, I think for a decent amount of people, when they look at a Muse, or, or, like, I mean, you kind of introduced this idea, like, like, when you look at a Muse or an Instinct, I think there's plenty of users that have been on Claude Code or on OpenClaw, obviously, that are like, yeah, like, we've had this in the enterprise. And so now that, now it's the consumers' turn. But I don't know that, that, that then flips back into then the enterprise in a completely different way that, that, that would look like the iPhone starting in the consumer and then, and then being brought in.

Yeah, I saw that. I saw that tweet also.

Well, let me tell you, let me tell you, like, so our air conditioning broke this weekend. And the one or two people that we called to fix it, you know, kind of, like, one person came and they had this whole thing, and, like, the part was, like, gonna be two weeks out or something. And there's, like, only two authorized resellers that could, you know, find this part. So we went on Muse, set the same problem. Lo and behold, it turns out the part is, it can be, you can switch it with a commodity part that anybody else can go and do. And we found another vendor, you know, through AI, and they came same day and fixed it. And so the... So, you know, it's tricky because, like, could you squint and probably have just pulled that off with ChatGPT before? Yeah, probably. Like, but, like, it's also one of those things where it's, like, I don't know, maybe Muse will need to call somebody. Maybe, like, there will need to be some vendor on the other line where you're asking for inventory. And that's, like, really cool. So I think probably just the industry just needs to market more things than, you know, booking flights.

Yeah, I mean, well, it's funny because we've been doing demos of this shape for a decade. Chatbots have been around for a while.