Inside Box’s AI playbook with founder & CEO Aaron Levie

Aaron Levie, CEO & co-founder of Box, joins Azeem Azhar to explore how an “AI-first” mindset is reshaping every layer of Box – from product road-maps to pricing – and what that teaches the rest of us about building faster, smarter organisations.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  We're going to look at the last year and the next, let's say, two to three years as probably the most defining periods of workplace technology change that we'll ever live through. What would happen if many parts of your business, you could start to do things in an hour or two hours that would have otherwise taken a week?

• 0:16Azeem Azhar?

  Which part of your business do you expect to disappear or become unrecognizable in the next 12 months?

• 0:23Speaker 00

  I think actually very few roles will disappear outright. I think it's what we will be doing in those roles that will fundamentally shift. We should have three, five, 10x more leverage as individuals inside of an organization. Everything that we do will be accelerated.

• 0:36Azeem Azhar?

  How should we think about a world of resourcing when agents start to appear and become more reliable?

• 0:44Speaker 00

  In that world, the IC is now a manager and there's no upper limit of how many agents they could be deploying. Then the question is, what is that manager of human managers of agents? What do they do? Should they also be managing agents? Open question. I mean, this is in the category of things that will change in a five-year period that might redefine 100 years of management science.

• 1:05Azeem Azhar?

  I mean, it's super complicated because you have model releases every two weeks right now. The velocity is extremely fast.

• 1:12Speaker 00

  It's this incredibly exciting, stressful, crazy time. But for those that enjoy learning and changing and evolving how they work, this is a very exciting time to be diving into that.

• 1:29Azeem Azhar?

  Today, I'm really delighted to have Aaron Levie, the CEO and founder of Box as my guest. Box is a company, a product that I first came across, I reckon, 20 years ago, before it had the dot com domain. It was one of the first cloud storage companies in the world. Of course, now it does much more for enterprises. And Aaron has been leading a transformation of Box into an AI-first company, both for the company itself and for its customers. And Aaron, you've been really public about the benefits that AI will bring to all of us. If Box is fast becoming an AI-first company, which part of your business do you expect to disappear or become unrecognizable in the next 12 months?

• 2:15Speaker 00

  I think what's gonna hap- the, the part that will be unrecognizable, I want a first caveat that like, I'm probably gonna be wrong on ninety percent of anything that is sort of twelve months out. You know, anything in AI that you can't like see and feel right now is just so hard to predict. And it's actually, I think relatively apropos of, of the name of your core content right now, it is this idea that, that everything that we do will be accelerated. And so I, I somewhat try and readjust when we get asked like, okay, what roles won't exist, for instance. I, I think actually very few roles will sort of disappear, uh, kind of outright. Now, there, there are some that, that we can get through as, as in a couple of categories. But for the most part, I think it's what we will be doing in those roles that will fundamentally shift.

• 2:56Speaker 00

  And then ultimately, the, the amount of output that we get from many of these roles will, will be the biggest difference. And so I think there's like two ways to think about AI. I'm sure there's more, but like, one is, is when you look at work, if you kind of imagine like a mass of work. And then you kind of think about AI taking part of that mass. That's sort of like, I think our default instinct because of, you know, kind of what literature says and sci-fi. Sci-fi doesn't sort of anticipate like dynamic economies. It usually says, okay, we're working in the factory, robots took the jobs and now nobody's working. That, that's kind of the-

• 3:28Azeem Azhar?

  It's the lump of labor fallacy, as the economist called it, right?

• 3:31Speaker 00

  That, that, that's exactly right. So like every sci-fi book ever written has a lump of labor fallacy, you know, built into it. I think there's a different way to think about AI, which is imagine work is actually on a timeline. If you kind of zoomed out over a two hundred year period, let's say back a hundred years, forward a hundred years. And you just sort of look at progress in society. All we're really doing in, in work is sort of moving through some progress in society, right? We're trying to like cure cancer faster. We're trying to, you know, give people better transportation so they can get places. We're trying to entertain people in new ways. My view is AI is an accelerant to the timeline as opposed to a reduction of the kind of mass of work.

• 4:10Speaker 00

  And so if you, if you start to think about it more as an acceleration on the timeline, then it, it gives you a little bit of, of some instruction of then what to go focus on with AI, which is do, use AI to move faster as an organization. If as a byproduct- That you save money, fantastic. Or if you can now do things that that are now cheaper because you're moving faster, also fantastic. But use AI to actually move faster as an organization. So if you roll out a year from now, but then I think even probably, you know, it's almost easier to think about five years from now than one year. But at some kind of point in time in in the not not crazy amount of future. I would just argue that we should have, you know, three, five, 10x more leverage as individuals inside of an organization.

• 4:53Speaker 00

  Which means that when we go and brainstorm a new marketing campaign to go do. Instead of us having to have three brainstorms over a two week period and go in and, you know, talk to lots of other companies for best practices and benchmarks. That might be a 30 minute exercise, you know, just involving your marketing people and and AI kind of, you know, coordinating on that. And then boom, you just, you know, shrink that process by a couple of weeks into 10, 20, 30 minutes. And so imagine that across engineering, product development, marketing, you know, going and and coming up with the sales plan for a customer. What would happen if if many parts of your business, you could start to do things in in an hour or two hours that would have otherwise, you know, taken a week? And and what is that? Yeah.

• 5:38Azeem Azhar?

  Sorry, pardon me, but there's this idea then embedded in that is that we all have big backlogs in our Kanban boards. And, you know, can you move through that backlog much, much more quickly, which of course talks about one of the behaviors that's important. Can you cultivate useful things to do further down the backlog? I think we'll get into that idea as well. So there is this notion of.

• 6:00Speaker 00

  Well, but and and and to that on the lumber labor policy, like the Kanban board might go off on a completely new direction as a result of this. I like as an example, I have personally, and this is like to the chagrin of my my colleagues, I have lit up more projects internally. Because of my access to AI. Because what I do is I go experiment on some idea, then I then get convinced it's much easier to go do now. Because AI kind of let me start to do the experiment on it. And then now all of a sudden I've lit up a project that now people are working on. And that was. That wasn't in their Kanban board a week ago.

• 6:38Azeem Azhar?

  I think it's something you may have said on X. I think you said it makes things much easier to start projects. And I remember reading that X post thinking, yes, and who gets to finish them, Aaron?

• 6:48Speaker 00

  Right, no, a hundred percent. I literally, so we have this project that we're doing right now, we're on the tail end, where basically I shaved about three days to a week off of the initial brainstorm process that would have taken. And I did that in a couple of minutes with AI. Then I rallied the team around it because the AI did the upfront part and it sort of showed what it might look like as a result. We still had another three weeks of work to go make it happen. I shaved off a week at the front end. But by shaving off the front end, I lowered the barrier to us sort of realizing what would be possible. But then I added three weeks of work to everybody else that wasn't on their plate. So AI is going to do that a lot more than we realize because what's going to happen is it's going to expand the aperture of now all the things we can go work on.

• 7:32Azeem Azhar?

  I completely agree. And you know, there is, there's this idea of Armdal's Law, which is one of the famous computer science engineering laws, which says that, you know, if you have a process that's got a number of steps and you accelerate one part of the process, if that's only 10% of the duration of the project, you've, you've only shaved 10% off the entire time. And so we have to turn that into reality. And I think there's this notion of what is the propagation curve of AI within, within an enterprise? So it feels like you've described individual productivity, your individual productivity, maybe it applies to my individual productivity as bosses who generate work for other people accelerating.

• 8:13Azeem Azhar?

  One of the things I found quite curious in talking to enterprises and CIOs around the world is that many of them report lots of individual improvements, right? Individual developers, writing more code, whatever the metric happens to be. But they're not seeing the sum of the parts being necessarily greater than the individual contribution. There seems to be some mechanism that breaks down from individuals being able to do really well using AI and a team, a function or an enterprise being able to return the same result. Does that ring true in what you've seen with, within Box or your customers?

• 8:51Speaker 00

  I think embedded in there is an interesting kind of point about we will see stats of okay, individual productivity goes up 10, 20, 30% with AI. But like clearly at this stage in 2025, organization-wide productivity is not 25 or 30% higher, or already the economy would already be, you know, completely impacted. And so there's some kind of dilution that happens through each kind of turn of the crank of incorporating the AI productivity into the larger system. And similarly, there's, um, I don't know if you've ever read the book, The Goal, but it's all about basically process improvement. And the book basically goes through exactly what you just said, which is it sort of goes step by step and finds each of the individual bottlenecks in a particular kind of system.

• 9:35Speaker 00

  And until you sort of figure out where is the bottleneck as the next sort of gating factor in productivity, you don't get any benefit at the other end of the amount of productivity you get. So I can light up a project incredibly quickly. But if every other corresponding part of that process is not also benefiting from AI, then all I've done is in many cases, actually added more work, which then won't inherently even show up in productivity metrics, because I've actually now added even more tasks to the organization. And we haven't yet gotten efficient enough to then sort of see the productivity gains across the organization for that. So I think this is probably very much just a natural evolution of, I wasn't around for it, but I'm sure that like there was a multi, you know,

• 10:18Speaker 00

  probably decade plus evolution of like, the first people with email in a corporation were like super productive, but then they were ultimately bottlenecked by everybody else that didn't have email. And that was a probably a decade-long process for everybody to finally get wired up, where you could efficiently communicate with everybody to the point where that showed up as a productivity gain.

• 10:37Azeem Azhar?

  And then reasonably quickly after that, we drowned in full inboxes because everyone had had access to email. So you went through this sort of funny, funny cycle. But you've raised something that I'm curious about. What is the balance between you as the CEO, being good with these AI tools and coming up with new projects, and needing your frontline workers to be really, really great at them? So here, sometimes we Think of when it rolled out, the internet was very much in many companies sort of frontline driven and slowed by the center. If you look at things like ERP, that was driven from the top because frankly, who wants to live their life in an SAP interface? And sort of forced down in the center. And you've got these two different models.

• 11:19Azeem Azhar?

  To what extent do you need to enable the absolute frontline workers? And what's that balance between you as a boss, leading them from the front?

• 11:26Speaker 00

  Yeah, so I benefit from just, I'm already pre-wired to just like love technology and and always, you know, want to explore new things. That's that's why we created Box. And so so I have a little bit of a natural advantage because like this is just, this is like what I do for fun. These are like my hobbies usually are just like technology, like that. It's not like I don't do, I don't do fly fishing. So I already love technology. I love to play with technology. Most of my social friends are are doing technology or AI. So I I get it in every sort of part of my life. And so then the question is, how do you make sure that you're enabling all the folks that that don't have that as as sort of the same built in default way of way of operating? We spend a lot of time figuring out how do we enable kind of the front lines or the broader population of colleagues?

• 12:10Speaker 00

  To become better and better at these tools. And it's it's interesting because this is a space that is moving so fast that, you know, most people are, you know, probably fatigued actually by the amount that's coming at them. To try and figure out. And in particular, if you've tried something with AI a year ago or six months ago and it didn't work. You might have be like, you might have like written off that, well, okay, AI can't solve that particular thing where I want to review a contract or go and summarize. You know, five hundred documents into a into a report, you might have been like, okay, AI doesn't do that. And actually, you know, check back in today and it does. So then the the need to kind of stay current on everything that's happening is just unbelievable.

• 12:51Speaker 00

  So we're we're doing a lot of things like some of our practices are every Friday, we have an internal all hands. And we'll have somebody demonstrate how they're using AI. We're trying to make it more of a social collaborative exercise where you're demoing in front of your colleagues. How you're using AI to get some, you know, particular type of work done. We're working on more internal enablement, which would Include kind of agitation. We're pushing everybody to read Substacks and subscribe to newsletters and listen to podcasts so they can kind of stay as current as possible. I think about this as if you're working in technology, you know, in particular, but then probably the broader economy, certainly. You know, we're going to look at the last year and the next, let's say, two to three years as probably the most defining periods of workplace technology change that we'll ever live through.

• 13:36Speaker 00

  And so it's this, you know, incredibly exciting, stressful, crazy time. But but for those that enjoy learning and changing and evolving how they work, this is a very exciting time to be diving into that.

• 13:49Azeem Azhar?

  So there's a lot in that, a couple of different, you know, questions. Well, the first I wanted to pick up is, give us a flavor of those Friday afternoon learnings. I mean, we do something similar with my team. In fact, starting in December two thousand and twenty two, we started to do a daily all hands, uh, which is only a team of five, right? There's a smaller group, fifteen minutes. What did you do the day before? And back in those days, of course, to show how random it was, it would be a bit of mid-journey. Well, I did this image in mid-journey. And now it's getting more and more sophisticated. You know, people are showing, uh, you know, workflows that might involve eight or ten agentic steps. They might involve decision making by the agents and

• 14:31Azeem Azhar?

  things that break rather randomly. Like, you don't know. It worked. It worked really well with the two point zero version of Gemini Flash. And in the two point five version, it's it's stopped working. And, you know, of course, but we're a we're a small organization. We can, you know, we know each other very, very well, six people. Box is much, much larger. So what do those actually look like? And how do you get contributions? And is it still, oh, look, I I've got a great prompt in Claude. Let me show you. Or is it, we've built something in crew.ai that involves five hundred steps?

• 15:05Speaker 00

  Yeah, the focus is we want to push the limits of our platform. So it's a narrowed scope, um, because it's just box AI functionality.

• 15:12Azeem Azhar?

  All right. You've got agents and other things like that built into box.

• 15:16Speaker 00

  Exactly. And so and so it's all going to be related to, um, automating a document process. Generating content, reviewing content, summarizing or adding intelligence on top of data. As the benefit of the surface area is somewhat narrowed. But what happens is we'll have a sales rep come and it's all on Zoom. So come present for, you know, 500 plus people that tune in live and then thousands that will watch it. They'll present their workflow for, hey, here's how I take all of my transcripts from my customer conversations. And I turn it into a summary and then I sort of tune it for that customer's industry and I send off an email.

• 15:55Speaker 00

  Or I take all of our sales material and I produce some best practice guide for my sales team. Or I'm in engineering and I want to take a chunk of code and have the AI look through it to improve it or do documentation. So these are the kind of use cases that are emerging organically. And then we just have those basically boxers go and show that to everybody in the organization. And then for us, again, it's like very selfish because both we want to work in this new way. So we can be helpful for our customers. But we also want to see then the upper limits of our own technology. So then we know what to go improve on the product. If we can shorten the product feedback cycle that normally happens, which is you release something, you listen to customers. It takes a couple of months.

• 16:40Speaker 00

  If we can internally just learn those things very quickly, we can recycle that back into the product roadmap. And then improve our own internal development velocity. So we're using it selfishly for our product teams to then also see, oh, wait a second. We ran into this issue where people want to share prompts, but we don't have a useful mechanism for sharing those prompts. Let's go build a prompt library, like those kinds of things.

• 17:01Azeem Azhar?

  Yeah, I could do with a prompt library. I mean, it sounds great. It sounds very, very inclusive. It sounds like there's a lot of activity, but activity doesn't necessarily beget results. So let's take that upper level to your board. They need clarity. They don't have a ton of time. So what is the single metric that cuts through all of this activity, all of this hype? To really reveal benefits for box from these activities, like that's showing the productivity.

• 17:29Speaker 00

  I think as an industry, we're very early in what is a normalized sort of metric to quantify AI driven productivity. I think there's a range of approaches that you can take. We've actually explicitly said that right now we're still in the journey to figure out what metric we're going to land on. There's an easy one, which is sort of like, how much money did I save directly? Then there's another one, which is how much money in productivity or hours or or system time did I have cost avoidance on? So that's like, if I had done this with, you know, human labor, it would have cost this much. And then a third easy one is just like, what are the hours, tokens, you know, some some metric of just sheer output coming from the AI that that now you're utilizing?

• 18:13Speaker 00

  Right now, what we're actually way more focused and I'm very comfortable being more in the in the, we actually want like a thousand flowers to bloom mode. I I don't know if I would feel comfortable if I were like at JP Morgan with that strategy, but being on the front lines of a software company that that is fundamentally, you know, becoming an AI first company, I don't mind a little bit of a period of let's just, let's let everybody go and test and push these systems. And I think, I think the best scenario for us right now is is this idea of revealed preference, right? Like, let's just like, see what people actually do. And almost by definition, the market, you know, we we set very high goals for anybody working in the company. And so it's quite easy just to then let the market decide how to use AI to accomplish those goals.

• 18:57Speaker 00

  And so if we're actually setting our goals properly, and if the managers or leaders know what AI is capable of, so if if we know that AI is capable of making a product development process go thirty percent faster, and we set that goal kind of built into that, then we don't need to tell people to use AI, or how to use AI. By definition, the only way to hit that goal would be to use AI. So for us at least, I'd I'd rather ratchet up the goals, I'd rather accelerate the timelines, hold people to a high bar of how quickly things should happen, and then revealed preference will just be what tools do they use? How do they incorporate that in their workflows? We don't necessarily need to hand everybody the exact template to go follow.

• 19:36Azeem Azhar?

  Do individual workers have a budget of tools that they they can access? I mean, what we've done within my team is that everyone has a quite generous a thousand dollar a month budget for AI. Well, once you've got O one Pro or whatever the ChatGPT Pro thing is, that's two hundred bucks out out of the door, right? And then you need Claude and and so on. But but, you know, and no one's really approaching that that level right now. But the idea was to to give people enough freedom to experiment. And then every three months, you can do a triage and say, what worked, what didn't work, what don't we now need? What has fallen off the agenda?

• 20:10Speaker 00

  Yeah.

• 20:10Azeem Azhar?

  And what is working? But what is the right way to allow them to explore? So you do have a thousand different flowers that are blooming. So you're exploring the possibility space for Box or any enterprise, but without losing sense of kind of financial control or security or data.

• 20:30Speaker 00

  Yeah, so I love that idea of a of a user budget or employee budget. And that's exactly probably what I'd recommend to anybody outside of our company. Right now, we have not kind of created any sort of stated limit. And mostly because I don't know what I would I would have our organization do. And for example, so within Box, Box AI basically connects almost any leading AI model to your data. And so that means that as an employee, you can go into Box and let's say, create a document and use Claude three seven to generate, you know, any of the content in that document. So you're effectively able to use any of the leading AI models, you know, within within our company. And we don't have a budget limit, only because I think it's very provocative. And I think it's the right thing to do for us.

• 21:15Speaker 00

  We've just said, go run wild. We want to learn what what everybody's doing. We might now, who knows? If you could kind of like parallelize a hundred coding agents all using Claude three seven, doing long running agent, you know, agentic coding tasks. Maybe we'd be like, OK, sorry, you've just spent $200,000, you know, building some feature like that. Then that'll be a problem. We are not there yet. And so maybe we'll have to check in at some point if we've had to kind of evolve our practices on that.

• 21:43Azeem Azhar?

  That's an important segue. Thank you. You're such a wonderful guest because we wanted to move towards what happens in a world of agents. And that has been the sort of the rhythm of 2025. I guess what happens with the world of AI agencies, these things can spawn multiple instances. They get to run much more autonomously because you move from the prompt, respond, prompt, respond. The human is always in the loop. So that creates an issue around reliability. And trust, but it also affects how you think about the design of the organisation.

• 22:16Speaker 00

  Yeah, yeah.

• 22:17Azeem Azhar?

  I I I noticed that Moderna, which is the the wonderful biotech firm, there was a story in the the Wall Street Journal a few days ago saying they've merged their tech and H R under a single person so that you can align workforce planning with with A I. Moderna had famously deployed more than three thousand different G P T to tasks in the previous couple of years. How should we think about a world of resourcing in general when agents start to appear and become more reliable?

• 22:49Speaker 00

  Yeah, I mean, this is definitely in the category of, again, things that will change in in a five year period that might redefine a hundred years of management science. This is one of those. You know, if you think about like the Alfred Sloan typical management science approach of, okay, you know, we we we developed all these hierarchies and and everybody's an expert and and everybody has sort of division of labour and and, you know, very clean systems. You know, this is sort of blowing that up a little bit because now you do you do have some fundamental questions of what happens when an I C or individual contributor effectively can become a manager of agents. And there's sort of no upper limit of how many agents they could be deploying. In that world, the I C is now a manager.

• 23:33Speaker 00

  Then the question is, what is that manager of human managers of agents? What do they do? Should they also be managing agents? A open question. The other question is, how much is is this the responsibility of a reinvented H R department? How much of it is is a responsibility of a reinvented I T department? Where you can either have H R sort of thinking about, I have a new type of labour force, à la agents, and I have to think about them in coordination and concert with people, or Is it the IT department saying, hey, we're like the technical experts. We're going to always be on the front lines of what's happening in the AI space. And we're going to bring that to the organization or the lines of business or to HR and say, hey,

• 24:16Speaker 00

  I have a new type of labor I can supply to the organization. Here's the latest set of offerings that we can bring to the business. I'm probably partial to the latter because the technology is moving so fast that I think you want, you're going to need to have deep technical expertise to understand where this is all going and how to incorporate it into the rest of the business workflows. Because these AI systems will not show up in a sort of any kind of like physical form, you know, for the foreseeable future, you know, like we can get into robots or whatever. But like, that means that they're only going to be deployed from software and IT is going to manage that software. So, so I think that this puts IT in this unbelievably interesting role, which is you're now kind of responsible for the digital workforce of the company,

• 25:00Speaker 00

  which means that, that you need to then also be incredibly astute at the business processes of the company. Mm hmm. Um, you need to be incredibly astute at, at basically, you know, if I could go and add a workforce to the sales team to go in and do lead generation or lead review, what part of the sales productivity metric would that move? That's not usually a classic IT conversation that they would normally go in and be that partner on. In five or 10% of companies, definitely. But for 90% of the world, what, what usually happens is the sales leader says, hey, IT, I need a new tool for helping my sales reps go and do lead gen. But in this moment, IT should probably be coming to the sales leader and saying, hey,

• 25:43Speaker 00

  I can deploy this kind of digital labor at this part of your set of workflows. So, so the premium on understanding the business, the business model, the business processes is massive right now in IT. And, and I think this is going to be a major moment for companies to kind of get around that.

• 25:58Azeem Azhar?

  Wow, I mean, it's super complicated because you have model releases every two weeks right now. So prior to ChatGPT, the big five model companies released a new model every six months. Now it's every, every two weeks.

• 26:12Speaker 00

  Yeah.

• 26:12Azeem Azhar?

  Within that, you have their specific product releases. So OpenAI announced a new coding agent just before we went on to do this recording. You also then have the growing plethora of applications that, or wrappers as they used to be called, right? That are built, built on top of these. So the velocity is extremely fast. And IT has, I suppose in recent years, also got a sense of being the chief technical authority of the company, right? What, what can we use? What has got the right trade offs between the investment we put in, the result we get, set against the risk? Because IT has a risk component to it. So there's an entirely new discipline.

• 26:57Azeem Azhar?

  So I, I think internally, it ends up having to be the CEO.

• 27:00Speaker 00

  I, that, that I totally agree with. But very quickly, I've met the CEOs of these big companies, like, they're not going to spend more than two hours a day personally trying to read all the AI news. So then the question is, who are they going to go and say, I need you to be my expert to help me figure out what we're going to go and deploy. And, and I guess my, my point being, because of how technically complicated and quickly this space is moving, I personally wouldn't want to rely on, on a non-technical function to kind of, you know, guide me through that. The kind of vocabulary that, that we've had to develop just to even make it through this conversation is like, it's already two years of us having to be wired into this to be like, okay, you know, Claude three seven Sonnet is better at coding than, than O one.

• 27:45Speaker 00

  And, and, and think about the productivity, you know, difference if you deploy one AI coding tool versus another. That is not something you want to leave to chance. So, so that's why eventually I still think that it becomes, you know, IT or engineering or, or whatever, whatever we want to call it. But it's a, it's a technical function that will have to basically be that counterpart to the CEO, to be very clear. Or maybe there'll be a lot of organizations say, okay, HR, you should go define what this looks like. But the CTO, the CIO, the CDO, the Chief AI Officer, that's the right hand to kind of getting through that journey.

• 28:20Azeem Azhar?

  And when you think about most large companies and most of your customers have already been in existence, they have systems of record, they have established processes, there seems to be a question between are we just able to throw AI systems, whether it's L L Ms or it's L L Ms based within agents on top of our existing decade-long relationships with our core sort of enterprise systems of record? Or do we need to start to rebuild these right from, you know, function at a time? How do you help CEOs think about that particular question?

• 28:55Speaker 00

  Yeah, I mean, this is super interesting and I I I would say that my thinking on this continues to evolve. And so I, you know, in five years from now, I I reserve the right to be totally on a different, you know, have a different view. Right now, I think that for existing categories of technology, a lot of the incumbents are very good. They're with it in terms of this latest tech breakthrough. This is not in an environment where I don't meet a lot of incumbent CEOs of software companies that have their head in the sand on this movement. If you compare that to, let's take a snapshot of 2007, let's just say. And if you were to go to ask the incumbent software companies, how much are you going to invest in cloud? So you go talk to the CEO of

• 29:39Speaker 00

  SAP, the CEO of, you know, at the time Larry Ellison, et cetera, and you say, how much of your business can be cloud first right now? There was this kind of culture of resistance. Larry was actually a big thought leader on on cloud. There was a little bit of a time where they took longer than they probably should have to move to the cloud, but but he was obviously within the 90s, he was he was a big believer in network computing. So so credit to him, but the company still took a while to shift. If you ask the equivalent incumbents today, and you say, hey, Benioff, or Bill McDermott, how much are you going to do AI? Like, these companies are all being completely pivoted to being AI-first platforms. So the the good news

• 30:17Speaker 00

  for an existing large enterprise who is using one of these vendors, your vendors are 100% motivated to becoming as AI-centric as possible and ensuring that

• 30:30Speaker 00

  Behind the scenes, they're they're making their data models, you know, improved. They're making their architectures improved to be able to support AI agents. And just in the past, just to give you an example, just in the past month and a half or two months, we've announced agent interoperability with Salesforce, with ServiceNow, with Google. We just announced something with Microsoft. And so the good news is everybody has gotten the bug. Every platform is going to be an agent first kind of software platform. So that's existing software. The neat thing, though, is AI is also offering the ability to have new categories emerge very regularly that are not just CRM or ERP or ITSM. These are categories of software where it's kind of quasi-professional services.

• 31:13Speaker 00

  It's it's a new sort of category of digital experiences where you say, now I want software to translate my marketing asset into 20 languages. Or I want software to, um, obviously, you know, be an AI agent for for developing code. These are going to come from new companies by and large. There might be incumbents that try and insert themselves into that space, but these are going to be lots of new startups. So those will be areas where you can kind of reinvent the process fundamentally. But I'm I'm bullish and optimistic on how the incumbent companies sort of work through this. I know that that might come as bad news for some disruptor startups where they would prefer that Salesforce wasn't with it. But but I think Salesforce is gonna still maintain their dominance on the on CRM plus agents.

• 31:57Speaker 00

  I think ServiceNow will maintain their dominance on ITSM plus agents. There's still lots of startup opportunity to be very clear, but I wouldn't anticipate those guys going away in the process.

• 32:06Azeem Azhar?

  Yeah, I I mean, I think it does feel like they have leaned in much, much more quickly. And I think with changes at the FTC, the M and A window reopens.

• 32:14Speaker 00

  Yes, that's a great point. I mean, Move Works being acquired by ServiceNow, like that may have not happened before. Then maybe there's one other byproduct. This era of enterprise software, for the most part, you still have leadership teams or even founders running companies where they them they themselves were the disruptor. And so everybody's like very clear that like you do not want to get disrupted. That's scaling up to the biggest companies. One of the fun stories. Of of software right now is Sergey Brin goes in every day and is working on AI.

• 32:46Azeem Azhar?

  Right.

• 32:46Speaker 00

  This is like, think about the era that we're in, where Sergey is just like coding and doing AI training right now. Because not only is this probably, I I'm in, I'm all, I shouldn't speak for Sergey, but probably the most exciting, you know, moment that that they've had, you know, as a technology company. But also, he probably understands what's at stake to make sure that Google makes it through this journey and Sundar, you know, as well. So you have these kind of like leadership, Sundar fully gets it, Sergey, you know, on board, driving things. It's been rumored that Bill Gates is, you know, back doing AI stuff at Microsoft here and there. So so what what an incredible moment that founders are coming back to their incumbent companies because of how important this period is.

• 33:27Azeem Azhar?

  I don't know, absolutely. And even non-founders, I mean, I heard, I spoke to someone who knows Satya Nadella very well. And he said that Satya is coding and he keeps hitting the rate limits with his coding assistant, which I think like, you're the CEO of the company. Somebody turn a turn a switch. On that question of agents, world of agents and these software platforms, we've come from a world of per desk pricing. And the these disruptors like Benioff, they disrupted a world where you paid an enormous license fee in maintenance. That's what on-prem software was like in ninety eight, ninety nine. And so how do we, how do we think about about agents and what this software, this business model turns into? We spend a hundred billion dollars a year roughly on enterprise software.

• 34:11Azeem Azhar?

  And, you know, I think I think we, people have started to say, this is going to be the business of selling outcomes and not just tools. So maybe we should even start with your experience at Box, right? If you're going to be selling outcomes rather than tools, how are you going to charge your customers?

• 34:27Speaker 00

  Yeah, so I think it's probably going to be a hybrid business model ultimately. Because I I don't think AI disrupts the still that that sort of desk worker dynamic. So so I think we're going to have, it's going to be seats plus. So basically, you know, think about it as a seat model for the end user that's interacting with the system. And then you've got this kind of like potentially exponential in some cases. But certainly augmented business model of basically, you know, agents doing additional work. So now all of a sudden, and this is like why I think it's the most exciting time to be building software. If you were a startup and you were pitching, I'm just gonna make up the example, but you were pitching legal document or contract management software 10 years ago.

• 35:12Speaker 00

  And you go to a company and they would say, okay, well, I have 15 lawyers, so I'll buy 15 seats of your software. And that software startup is like, okay, obviously it's not that big of a TAM because it's only the number of lawyers. So, you know, we'll probably have to price more than like, you know, the office subscription, but like, obviously there's an upper limit of what you can charge for software. So maybe it's like 200 bucks a seat, you know, something astronomically high relative to productivity software, but like 200 bucks a seat or 300 bucks a seat. That's your upper limit. You have 10 seats that you can sell to that company for a couple hundred bucks a seat. Now you're that exact same company in 2025. What you're gonna do is you're gonna go to that company, you're gonna say, we have software that not only makes your lawyers more productive.

• 35:52Speaker 00

  We actually start to automate some of the work that they otherwise do or never got around to or have to kind of outsource. And, you know, we're gonna charge at the rate of $5 a contract or $10 a negotiation or a hundred dollars a kind of a legal review. You know, some metric that an agent is gonna go off and do. And that metric has no upper bound for an organization. It's not capped by any inherent thing other than that company's own volume of customers or, or, or sort of business that they want to go through. So now that same software company, you could underwrite five or 10x the amount of revenue that they could generate versus what they could have generated before. To make this not even a hypothetical, think about Cursor.

• 36:35Speaker 00

  If I told you in 2010, there's an IDE company that's a fork of VS code. That's gonna get to 300 million in revenue in two years. You would be like, none of the words would, you'd be like checking which words. Yeah, which word did I miss here? Like, like, it doesn't make any sense, like, even it, like, $20 a user for an IDE, like, how many people could there even be? But it's because agents have an unbounded amount of utilization that they can go and drive. And so I don't think it disrupts the seat model, it expands on top of the seat model of software.

• 37:10Azeem Azhar?

  Yeah, I mean, I mean, if the if the TAM, the total addressable market is growing so rapidly, it almost doesn't matter what the seat model is because there's there's so much upside. There's a few things that just unpack here there. So one is that, of course, the underlying cost of delivering this input is declining really rapidly. So we saw the cost of a GPT four equivalent token decline about 200 times in an 18 month period to to to June, July last year, and they continue to decline. So that moment is really, really noticeable and that trend continues. The second is the utilization of those tokens, especially with inference time or test time scaling, which is what the reasoning models do, especially with agents, dramatically, dramatically increases.

• 37:54Azeem Azhar?

  As a little aside, I was sitting with one of my colleagues and we would we just thought we would map our company use of AI tokens. A lot of it comes through APIs and he he used 400,000 tokens in 15 minutes, just doing doing some little thing. And so you you get these two variables that are inputs, like the cost is coming down, but the the amount we use is going up. It's very clearly got positive elasticity of demand. OpenRouter, which is an aggregator of LLMs, has seen a 45x fold increase in per token usage over the last 12 months, which is which is pretty dramatic. And and on the other hand, you know, agents are getting more reliable.

• 38:35Azeem Azhar?

  So so meter, which measures this stuff, says that on certain coding tasks, an agent can do 50 minutes of work, unattended 50% of the time, but that is doubling every seven months. So you're not far away from the five hour task, but it's only at a 50% reliability. Yeah. So we're trying to make sense of all of that, right? In in terms of what does the shape of the the market actually look like? If I'm paying for outcomes, I want a certainty that it's been delivered. So if I'm the agent provider, but my agent is only 60% reliable, I'm going to have to run this agent several times to give you a guaranteed verifiable outcome, which raises my cost and my time.

• 39:18Azeem Azhar?

  So so when I start to package all of this, I'm thinking, hmm. Does this actually slow down that future that you and I have talked about? Or is the rate of technical progress so fast that this becomes a kind of a tractable problem to get to the future that you described?

• 39:34Speaker 00

  Yeah, I'm gonna go with the latter. I think you can outrun it with with just the efficiency gains of the models. Here's a fun example. We just upgraded one of our default kind of use case models from 4.0 to 4.1, and we saw about a 15-point improvement, basically overnight, because all we had to do was switch the model, maybe tune the prompt a little bit. 15-point improvement in basically favorability score of answers. You know, I'm gonna make up the numbers, but we went from like mid 70s PDFs to 90, you know, kind of more or less on PDFs. So asking a question, getting a summary, kind of adding expertise. And so basically, in less than a couple quarters, going from 4.0 as our default to 4.1 as

• 40:17Speaker 00

  our default, we got that level of boost. I haven't checked in on the latest on 4.1 costs, but like, was not, you know, didn't kind of astronomically change the the financial variables of this part of the product. So I think what's gonna happen is you're gonna see improvements to these models just continue at the exact same rate that we've seen. The cost curve come down at basically exactly the same rate. We've got this great dynamic, which is you're now seeing TPUs sort of bring down costs on the Google side. That's gonna obviously continue to drive competition for the Nvidia's of the world to have to continue to improve their pricing and their their kind of gains. So I think, I basically, asterisks that I could totally be wrong, I think costs are not gonna be the underlying problem

• 41:02Speaker 00

  with AI. The things today where, to your exact point, you've got to run something through a model 20 times, either in a year from now, you'll run it through a model five times, because the model will be 4X better, or it'll be 25% the cost, because some other efficiency gain has happened that has caused the amount of times you run through the model to be just much cheaper. Like, you can solve it through either variable. Either the token goes down, or the quality of the token goes up. One or the other, you're gonna end up being able to outrun this.

• 41:29Azeem Azhar?

  I suppose where we've seen the tremendous growth on the agents has been in coding. And coding as a domain, of course, is very, very verifiable. There's a, did it work or did it not work?

• 41:40Speaker 00

  Yeah.

• 41:40Azeem Azhar?

  Lots of things you've discussed in your business, unstructured data in contracts, in sales proposals, in market research. A summary of that is not necessarily, you know, right or wrong. There's a verifiability problem, but verifiability of a softer question is just a more expensive verifiability. But it does seem to me that that needs to be tackled because a lot of the value in the world is not the lines of code. And it's quite difficult. We've discovered this, that, you know, we run tests, I know the box does this as well, where, you know, you have lots of the L L Ms and you're trying to get it to do some work with unstructured text. And we just discover that seven times out of 15, oh, three is best.

• 42:23Azeem Azhar?

  Four times out of 15, Gemini 2.5 is best, except sometimes when it's Claude 3.7. And then occasionally your deep seek outperforms on the remaining tests. It's really hard to make a choice. Now, we're not putting them in mission critical roles right now. There's always a human in the loop. Is it the underlying capabilities of the models improving on this exponential curve that solves that problem? Or is it, you know, what the vendor offers to their customers in making sense of it? Or guarantees or assurances?

• 42:52Speaker 00

  I'm talking my own book for a second, but this is where I think having the software layer be different from the the model layer ends up being useful. Because we don't have any, I mean, we have got great relationships and friendships and partnerships, but we don't have like a technical bias toward one of the AI vendors. So we will go where the model is best able to solve a particular problem in that domain. And so you can easily imagine an agent in the future that says, okay, for legal contracts, I like Claude 3.7. For financial documents, I like Gemini 2.5. For, you know, large code bases, I like Grok 3.5 or whatever. And and it will route, you know, based on whatever data type it sees or the length of the data type or some other nuance. It's like not that hard. Like, we can technically do that today.

• 43:37Speaker 00

  We don't, it's not that elaborate because we haven't. We haven't found the need, but like you would want to rely on your vendor to do that. So Allah box. But the good news is that these are increasingly tractable problems. And if you had said this two years ago, I would be pretty nervous because, you know, think about the state of the world of like Chat GPT three point five. You know, you, you're, you're like rolling the dice, whether the answer is accurate or not. And the context window, it's bizarre to even think back this much. The context window at the launch of Chat GPT three point five, I think, I could be wrong. Somewhere between four and eight thousand tokens.

• 44:13Azeem Azhar?

  That's right.

• 44:14Speaker 00

  That's like, you know, a couple of pages of a document. And now we have million token plus models. That you just give the, the whole, the whole document to the model and it comes back in two and a half years. To see a two hundred and fifty X improvement in how much context you can give the model is, is, you know, truly insane. Um, there's not a technology that's ever improved at that rate. So I'm increasingly confident that any of these types of problems are, are going to be tractable. And then, and then there's the whole category of unstructured work that to your point is not, it doesn't get compiled like code. But almost anything in the realm that we're seeing today of two, five, three, seven, four, one, three, five. Which is funny because like, you know, anybody listening to this will know every single one of those numbers, what it meant.

• 44:59Speaker 00

  And to the outside world, it's like the most insane sounding thing. Um, uh, but, uh, uh, there are many problems with any of those class of models. Will do ten times better than any human for that task. Like if I go into, I, I, I love all my doctors. Uh, they're, they're amazing. But respectfully to them, there are many times where I've been to a doctor two times in a row over a two month period. And the notes that I get from the second meeting are like forty percent accurate relative to what I said in the first session. And so God bless to all of the doctors in the world, but like making our doctors have to take copious notes to transcribe an interaction. That then is going to miss like many key bullet points that, that actually I, I shared, you know, in that process.

• 45:43Speaker 00

  My God, I would love everybody to be required to have a transcription of every conversation with a doctor. Run that thing through Gemini two, five pro. And that is the note that's in the doctor, you know, E H R system. So I think for a large body of work that we do today, every day, these things will already exceed what the humans are doing. Now, maybe they'll interpret the wrong insight as the key insight, but their ability to capture, synthesize, diagnose, summarize, transcribe is absolutely above human capability, you know, at this point.

• 46:18Azeem Azhar?

  Yeah, well, I mean, very, very true. And we have that experience within all of our meetings. You know, we're always using one or two different transcription methods to help us actually speed up how those meetings go. And of course, in talking of speeding up, I can't believe we're already top of time, Aaron. And normally I summarize some of the things that we have described and discussed. I will have to do that in the show notes because it's been such a rich set of guidance for people in the enterprise and bosses and workers trying to figure out how to make sense of of A I. I I want to just leave with one question for you, which is, what is the one thing that you are now doing with A I? As we're in the middle of twenty twenty five, that has really started to change your life outside of work?

• 47:02Azeem Azhar?

  That is surprising that you didn't think you'd be doing a year ago.

• 47:06Speaker 00

  I think I've reached the point in my personal use of A I that is very much akin to, it's very early in this, but it's very much akin to probably what sci fi had intended. If you think about the Jetsons or any, you know, author would have kind of predicted. I have a little command center where I'll be talking to my, my six year old about, about some historical thing or something interesting. He'll ask a question and it's the kind of thing that normally you just punt on. And you'd be like, I, you know, I don't know, man, like, I don't, I don't know why that happened. I don't know how that works. And we just like a hundred percent of time pull up Gemini or Chat G P T and just ask the question and go back and forth with it.

• 47:49Speaker 00

  And it's like, it's this remarkable thing where just anything you want to know is now at your fingertips as this personal assistant that's just always with you. So that's how it's changed my personal life to the point where I think we'll just get to a point in a couple of years where we won't remember what it was like to not be able to have every piece of information at our fingertips. It would, it would just be like, wait a second. So you just had to like guess how much Tylenol to give a five-year-old? You know, when they're sick, like what were you supposed to do back before you had these assistants where you just ask any question?

• 48:21Azeem Azhar?

  I think that summarizes exactly where we are. It's starting to become so deeply normalized for us. We are still at that moment of wonder though, where actually we recognize that we have it and we used to not have any of this. Aaron, thank you so much for making the time this morning.

• 48:36Speaker 00

  Thanks, man. Good to see you.

• 48:38Azeem Azhar?

  Okay, cheers. Bye bye.