Aaron Levie 谈 AI 的企业级采用

AI is going to take over the enterprise. We know this is going to happen, and it needs to happen to us faster than it happens to our competitors, which is a totally different dynamic than we saw with cloud.

What is the journey over the next decade? It's about the speed at which humans can change their workflows. How fast can somebody use a computer to do something—to type an email, write code, or generate a marketing asset? When that's no longer a limiter, how do these jobs begin to change?

It's so strange to me how many disruptions are happening all at the same time.

Aaron, thank you very much for joining us. Everybody here already knows you. However, I still think you should introduce yourself, just for completeness.

Okay. I'm Aaron Levie, CEO and co-founder of Box. At Box, we help enterprises take all of their unstructured data, or enterprise content, and turn it into valuable information. AI is absolutely this incredible accelerant for that problem.

I just learned that we're investors, didn't you?

Well, many years ago.

Many years ago. So, no claims post-IPO, actually. Ben Horowitz had this early blog post—basically, I think the title was The Fat Startup.

Yeah.

In response to The Lean Startup, right?

And we, let's just say, very much took that to heart. We basically deployed every single lesson. The name of the game was: You get big fast, you scale aggressively. That was a very important period in our company's journey.

The nominal topic of this is AI in the enterprise, but I think it's good to be nuanced about this because it's less obvious than people think. You've been talking a lot about AI on X, but you're also thinking about it in terms of your business. So let me set up the first question as follows: AI has historically been this very B2B, enterprise thing—chatbots or whatever, personalization systems. What's unique about generative AI is that a lot of the use cases are actually consumer or prosumer, right? Think about creativity or developers. It actually hasn't made as much inroad into the enterprise yet. It's just starting now.

Does that match with your experience? And how are you thinking about this transition to the enterprise?

I think if you were to look at the idiosyncrasies of AI and then reverse-engineer why that was the journey up until, let's say, the pre-ChatGPT moment, AI was extremely hard to use. It required, in many cases, having custom models for basically every problem you tried to solve. There was almost no way that a consumer ecosystem could flourish based on that. It was just not generalizable enough. There were really few products, other than maybe Siri and Alexa, that you'd interact with that would even have some sense of AI. Enterprises were the early adopters of AI systems to bring automation and workflow automation to their companies.

Then, boom, ChatGPT happens, and all of a sudden it's the exact right form factor for mass adoption. There are no startup costs. It took 2 seconds to learn the product. It's just a chat interface, so it was perfectly ripe for taking off in the consumer space.

You also have these incredible conditions set up for mass adoption. You have billions of people on the internet, and it was set up as a free product. Again, it kind of solved this latent question mark that everybody had: When are we going to see AI touch work and touch our lives? Everything was kind of the perfect set of conditions for mass consumer adoption.

On the enterprise side, unfortunately, you have kind of the opposite. You have lots of workflows that have been ingrained for decades and decades. You have lots of legacy IT systems with data that's not set up well to be accessed by AI. You have a sort of shadow IT problem, which is that most corporations don't want end users just injecting text into prompts that might contain information that the AI models could learn from.

So it's a difficult environment for that same level of virality, with the exception of a few of these prosumer categories. I've talked to large-corporation CIOs who are seeing people just show up with Windsurf, Cursor, and Replit. You're getting this sort of shadow IT version that we saw.

Dev tools have always had that.

Yeah, 100% fair. Dev tools have had that. But I think you're still seeing that now in the ChatGPT kind of leakage into organizations.

I'm sure their prosumer usage inside a corporate firewall is off the charts, even separate from the people that pay for it.

Totally.

So now the question is: What is the journey over the next decade for the real change management of deploying AI systems that drive the more GDP-changing productivity gains?

That's something where I do think we have to be prepared for this to take many years. It's about the speed at which humans can change their workflows, as opposed to how quickly the technology can evolve and advance. In Silicon Valley, and certainly for anybody tuning into this, we imagine, “Why doesn't the breakthrough that we just saw get released? Why doesn't that permeate every corporation within 6 months?”

It's because people have meetings and budget processes. They have to go through a governance council, get compliance on board, and figure out who has the liability when the thing recommends this stock and then the financial-services provider shares that with a client. That takes years, and there will be case law that needs to happen. We still have lawsuits going on about who owns the IP of this stuff. So that part is going to take years.

What's interesting, and I think you'll especially appreciate this on the cloud side, is that I remember when we first were scaling up in the enterprise, let's say in 2007, 2008, and 2009—let's say that 3-to-5-year period post-AWS, post-cloud starting its journey. Basically, to a T, every conversation you'd have with a CIO or a group of CIOs was: “Yeah, that's nice. Maybe some little corner of our organization could use this. We are never going to go fully to the cloud.” They had their arms wrapped around their servers.

I remember.

Yeah. Basically, they did not want to give up the infrastructure. There were too many questions and too many compliance issues. There were existential job questions of, “Well, what happens when this gets delivered as a service?”

Here's something super interesting. Let's say we're now 2.5 years into the ChatGPT moment. That same group of CIO conversations—none of that. It is basically assumed. It is fully assumed that AI is going to take over the enterprise.

The CEO, the CIO, the CDO, every job, every org leader is basically like, “We know this is going to happen. This is not a situation where we're trying to push it off. It is purely a sequence of events: Who do I deploy? How do I deploy it? How do I drive the change management? Is the model ready?”

What's really interesting is that I think the level of buy-in you have now in the enterprise is 5 times greater than we had in the early days of cloud. You could even see it. To me, the classic litmus test was that, if you remember, 15 years ago, Jamie Dimon was probably most famous for saying, “We're never going to go to the cloud.” They basically said JPMorgan would never go to the cloud.

Today, the equivalent commentary—although I don't have a perfect Jamie Dimon quote—is that David Solomon at Goldman Sachs has given this anecdote that they can now write an SEC filing or an S-1 for an IPO in a few minutes, when that used to take a number of analysts a few days.

The fact that those are the anecdotes already coming out of the biggest banks means that we're not in a fear-of-AI world. We're in a “we know this is going to happen, and it needs to happen to us faster than it happens to our competitors” world, which is a totally different dynamic than we saw with cloud.

Do you think this has implications for companies today that are building products that are pre-API or pre-AI products? With the cloud wave, you basically had a bunch of cloud-native companies that ended up taking over. Snowflake is a great example of this: The ones that decided not to go all in and were hybrid—hybrid kind of became known as meaning it won't work, right? Anything called “hybrid” hasn't worked, and they have to do that.

Do you think that because the buyer in the enterprise is more ready, companies that are pre-AI have more of an opportunity? Or do you think you're going to see the same thing with a lot of AI-native companies?

Well, I'm going to give you the non-answer: I think both. One benefit that the cloud cohort—or the SaaS cohort, after we all understood and agreed on what SaaS would look like—has is that, whether we adhered to this perfectly or not, we basically all tried to build API-first platforms.

Yeah. And so, or at least, API kind of equals platform. So we have the UI, and we have the API. And if you think about it, AI and AI agents are the perfect consumers of an API, right? They basically become these superusers within your system, on your APIs.

If I had to say, “I want to deploy agents to go and automate my ServiceNow workflows,” I think I’m better off just deploying the ServiceNow agent to go do that than doing an entire reinvention of my ITSM system to solve that use case. You could just go down the list, like Workday: if I want an AI agent to do some kind of HR-related task, I think I’m better off doing that within Workday than building an entire new system.

So you have a bunch of different factors versus the pre-cloud days. Going from pre-cloud to post-cloud was an entire rewriting of your software. You had to go from single-tenant to multitenant. The scaling of the systems was totally different. Even the functionality and application logic was different, because it should be real-time and collaborative. It shouldn’t be as asynchronous and batch-oriented as on-prem systems were. And so, in a cloud world, it is a reinvention of the user experience and what you’re doing in the system. We should definitely get to that.

Well, I just want to make sure I tease this out, because this is actually a very interesting point. Your claim is that to go from pre-cloud to post-cloud, that ripped through the entire stack, all the way down to the infrastructure—for example, tenancy. You had to rewrite everything. And then what you’re saying about AI is more of a consumption-layer thing, which is like, you just treat the existing systems as they are, and then AI becomes the consumption layer.

Do you think this is a 1.5 step, and the 2.0 step kind of rips through the entire stack?

Okay, so let’s bookmark that one for 1 second. If you do a pure Clay Christensen sort of approach—sustaining innovation, disruptive innovation—disruptive innovation is this thing that looks so much harder, so different, so less profitable. Sustaining is like, actually, no, I’d like to build that because it’s incremental. It’s better for our business overall.

The on-prem guys had a disruptive innovation. Everything about the business model of SaaS looked different, harder, stranger. “I don’t have the talent. I’m running a service-delivery operation as opposed to, ‘I ship you a CD-ROM with my code.’” Everything—go-to-market, finances, pricing model, the business model, everything.

AI, again, with the bookmark being the really big disruption that you could contemplate right now with AI, everything kind of looks like a sustaining innovation if you’re an incumbent. Instead of a user pressing the buttons in the application, let’s have an agent run through the API and operate as if they were that user. And so all of a sudden, for a lot of SaaS providers, this looks like TAM expansion, because now, for the first time ever, I can actually deploy my software for use cases where the customer didn’t have users on the other end before to do those things.

So I think you have a lot of TAM expansion. Now, the good news, with 1 caveat—which maybe we’ve bookmarked and we’re going to get to—but let me just say the 1 caveat is, you now have a component that has a very different COGS model if you’re a software provider. And so now it’s almost like when we went from on-prem to cloud, we went from perpetual to recurring. And it feels like with AI, you kind of have to go from recurring to usage-based, just because—

Yeah. Okay, so the business model will shift for some of the use cases, because even if you look at Cursor, Replit, and Windsurf, there does seem to be this baseline seat price, and then your consumption usage is there as an add-on. SaaS providers are well structured to be able to have that kind of dynamic.

If it was 100% usage and the user seat goes away, I do agree. Then you have a little bit of a business-model crisis.

Oh, so you think—but right now it’s not clear that that’s going to go all the way over.

Well, until the human literally is not a seat on the system, I don’t think you remove the end-user license as a component.

Okay. But again, that could be the much bigger disruption.

Yeah. Now, just to fully lay out the market dynamics, I think SaaS incumbents have a couple of other idiosyncrasies right now versus the on-prem days. Another idiosyncrasy is, I would say, on the margin, you tend to have founders still leading the SaaS companies.

That’s a great point.

We didn’t really have that in the on-prem world. Siebel already had 3 CEOs later, and PeopleSoft already had multiple CEOs later, so it was a different leadership structure in these organizations. A lot of the time, you still have the founder around. They’re poking around, they’re really into AI, so there can be a more natural pivot of the company from the leadership standpoint. So, a bunch of different factors.

Now, to the benefit of startups—which is why I can hold both of these in my head—I’m very bullish on the SaaS incumbent being the natural place for that AI agent relative to that category. I just think we have this incredible expansion of categories for the first time that we haven’t seen in probably 15 years.

The SaaS 1.1 wave actually expanded the software universe. We had these new categories of software that we didn’t expect before. Nobody would have predicted the Confluences and the Snowflakes in the pre-on-prem days. We didn’t have all of these different cuts of how do you work with data, how do you do this workflow, how do you do that. Lines of business didn’t have 15 different applications they got to use. Post-SaaS, they did.

For startups in the AI world, the equivalent of that is, I think, there are a lot of categories now where there’s no actual software incumbent in that category, where AI agents all of a sudden let you go build software for that category: legal, healthcare, education, and so on.

That’s definitely true on the consumer side, right? If you look at the top use cases of OpenAI, it’s almost like the top of the pyramid of needs. It’s creativity and fulfillment, et cetera. I think number 5 is professional coding, but everything above that is one of these. So on the consumer side, that’s very clear. Is that clear on the enterprise side?

I absolutely think so. If we did a snapshot 10 years ago of the size of the contract-management market or the legal-document market, it’s sub-$2B. I’m making up the numbers; it could be plus or minus $1B.

Yep. Yep. Would you agree that in 5 years from now, AI-agent-related spend on legal services should be in the many, many billions, into the double-digit billions?

Absolutely. No question. So all of a sudden, there aren’t these natural incumbents that were like, “Oh, we captured all that market.” AI agents all of a sudden expand the size of the software-related spend in that space.

I can underwrite that for healthcare, legal, and consulting services. I think there are entire areas of financial services. We always think, “Oh, finance has been wired up for so many years.” No. Banking—you know, consumer banking has been wired up, trading has been wired up, investment banking never went digital, wealth management never went digital. These were not categories where you ever had major software platforms to help these entire categories of the economy. And the reason was because the work was unstructured, very ad hoc and dynamic, with lots of unstructured data, as opposed to stuff that goes into databases.

All of that is now ripe for AI, and that will then largely be ripe for many startups, because there won’t be a natural incumbent in those spaces.

I mean, it’s so strange to me how many disruptions are happening all at the same time with AI, right? If you think about everything you said, which is basically vertical SaaS or vertical use cases, a lot of that is actually human budget, right? That’s being disrupted. There’s a bunch of new use cases that we never really thought about before, which is creativity. I mean, who would have thought that 2D images would be some massive market? But it’s a massive market, right?

I’ve been a programmer for 30 years. In that time, software would disrupt other things—we’d disrupt all of these things—but we never got disrupted. We were like, “We’re safe. We’re screwing you guys.” But clearly now software is being disrupted, right, for the first time that I’ve ever seen in 30 years.

Do you think this level of disruption is something that existing companies will not be able to manage? More to the point, you are a business leader right now. You have to think about product, and you have to think about your organization. Does it require you to think about too much? How do you structure your company as well? How do you structure your product, or do you think this is actually all pretty manageable?

I think your R&D literally—I’m putting myself in your shoes, right? You’re a CEO, your R&D is changing. Every part of the stack is changing, everything.

Yeah. Yeah. I think the reason that I’m probably, frankly, more distracted by what we’re building is that I don’t have enough time to stress out about the actual organizational side, because I’m stressed out enough about just literally the actual delivery of the product.

I think if I had a little bit more time, I'd get more stressed out about all the other change. We are very much leaning into the idea of being AI-first. We have a twofer on this: one, by being as AI-first as possible, we'll see the use cases that our product should go solve for customers. So, check that box. And then, second, I'm just a believer in the efficiency and productivity gains.

I do think it changes basically everything about work, and there are lots of interesting examples of what it means. In the future, does the individual contributor basically become a manager of agents? That's a totally different job, right? My recent go-to is just thinking about it as a lot of the productivity of your organization was rate-limited by literally how fast somebody could use a computer to do something—to type an email, to write code, to generate a marketing asset.

When that's no longer a limiter, how do these jobs begin to change? Your job is now orchestration, integration of work, planning, task management, reviewing, and auditing. That will radically change work.

Interestingly, it probably behooves us not to over-rotate on transforming yet internally for any given company, simply because the technology is changing so fast that you probably wouldn't want to snap the line right now and run your whole business on this technology, because 2 years from now it's going to be so much better. I think progressively figuring out which workflows have high-impact upside and getting them rolled out in a decentralized way so people can experiment—I think you want to do a few of those kinds of things first.

I can't imagine a listener not knowing what Box does, but just for completeness, maybe can you talk very quickly about what Box does and how you're thinking about how that dovetails with AI?

We started the company with a really simple premise: make it easy to access and share your files from anywhere. We pivoted about 2 years into the journey to focus on the enterprise market, and the whole idea was that enterprises are awash with all this unstructured data: corporate documents, research files, marketing assets, M&A documents, contracts, invoices—all of this. As companies move to the cloud and move to mobile, they need a way to access that information, collaborate securely on it, and integrate that data across different systems. So we built a platform to help companies do that.

We have about 120,000 customers, including about 65% or so of the Fortune 500. What's incredible right now is that we've had this ongoing problem since the creation of the company: with structured data, the stuff that goes into your database, you can query it, synthesize it, calculate it, and analyze it. Your unstructured data, the stuff that we manage, you create it, share it, and look at it, and then it basically gets forgotten about. It goes into some folder, and you almost never see it again. Maybe you find it once every 5 years for some task you're doing, but that's about it.

Most companies are sitting on data that is largely unstructured and getting the least amount of value from it relative to their other structured data. AI is basically the unlock. AI lets you finally say, “Okay, we can ask this data questions. We can structure it. We can look at a contract and pull out the 10 most important fields. Once we have all that data, we can analyze that information, get insights from it, and then start to do things like workflow automation.”

That was never possible with your unstructured data. If I want to move a contract through an automatic process, I can't do it if I don't know what's in the contract. Previously, the computer was not able to know what's in the contract. For us, there's a huge unlock in what you can finally do with your information and your content.

We're building an AI platform to handle all of the plumbing and the user experience to make your content AI-ready, effectively.

I don't want to be too bullshitty and provocative, but I have to ask this: I've been in enterprise software for a very long time, and a lot of the business model is predicated on the fact that building software is hard and takes a long time. To what extent do you worry about that not being true going forward? Do you think we're entering a time when bespoke software is upon us?

I'm bearish on the extreme version of that. If you imagine the poles of this, the extreme version is that all software is prepackaged. It's the Ford Model T: it's going to work only in one way, and everybody uses the same thing. We get that. That's not going to happen.

The other extreme is that everything is homebrew. You wake up in the morning, utter something, get your software for the day, give your software for that thing, and then the next day you do it again. You change it. The downsides of that model, and why I think it doesn't work, are that if you ask 90% of the world's population, you probably find that 90% or more just don't care enough. They don't care about the tabs on their software or the modules on their dashboard. They want someone else to say, “This is what you should look at in the morning.” They don't even want to have to prompt the AI to tell them what to look at.

Given that that's basically guaranteed to be where 90% of the world is, no matter how you cut anything, 90% of our software should largely be, “Okay, you log into the HR system, and it just looks like an HR system.”

In fact, there's another interesting dynamic, which is that over many years, our software and the actual way that we operate companies have had this flywheel relationship between them. The way we run our HR department isn't so different from the way Workday wants us to run our HR department, and that's fine, because that's not the area where we're going to have a lot of upside innovating. The way that we do our ticket management for customer tickets is the way that Zendesk decided to do ticket management, and that's fine, because that's not the core IP of the company.

In a way, it solves an operational problem for you: you don't have to figure it out. People miss that about software. I don't want to have to think about the workflow of an HR payroll process. I just want the software to do that. That's what people are buying. Nobody wants to customize those things.

Again, given that we're going to be in this world of many different outcomes playing out, the reason I'm still bullish on Replit and vibe coding is for a different category. I'm the IT person, and I have this crazy queue of tasks. Then someone's like, “Can you build a website for this thing? Can you code up some inventory, random plugin for this product?” That now becomes 10 times easier.

It's the new prototyping, scripting, and long tail of stuff that people never get to. That long tail is so long, and people never get to any of those things in that long tail. I could underwrite a 10x growth in the amount of custom software that gets written, and the fact that these core systems don't go away, because there's actually going to be way more software in the world that gets created.

Let me pressure-test this. I can imagine why it would be hard to rebuild Box, because what you do is actually hard. This is core infrastructure. You store data, and that's really important, so I don't think you just vibe-code that away.

From my perspective, a lot of SaaS apps just look like CRUD to me. CRUD—I don't know what the acronym stands for—but it's basically reading and writing data from a backend. Do you think there's a world where the consumption layer evolves to just using AI and this class of companies goes away? Or do you actually think, if I heard what you just said, that the durability of these companies is that they basically teach you what the workflow is?

I'm still going to say the latter. Now, I don't know if you need to bleep it out, but if you want to share a couple of examples of who you put in the not-hard-CRUD layer, then we could parse that. But it's up to you.

The not-hard-CRUD layer. Yeah, I mean, I would say most vertical SaaS companies I see—the technology is trivial.

Yeah, but the understanding of the domain.

No, no, this is what you said before. This is what I want to present.

The thing is, that's actually a great insight. I've always underestimated vertical SaaS relative to the outcome. Twenty years into doing enterprise software, I'm just no longer going to underestimate vertical SaaS. It's not about the technology. It's the fact that somebody else has figured out the business model that works, and they have 10 people from the pharma industry sitting next to the engineer, saying, “This is how you should do the clinical trial workflow.” That becomes so much of the IP.

That translates fine to agents, but I would still bet on that vertical player doing it, as opposed to somebody prompting their way into ChatGPT to build an FDA-compliance agent.

I would still largely bet on Compliance Agent AI to do that over the pure horizontal system that has no particular domain expertise. I still think that there’s a relationship between some amount of GUI, the agent, and the APIs, because you don’t want to go to a blank, empty screen every day of your life and say, “What’s our revenue today?” You just want a dashboard at some point that shows you the revenue.

That’s right. Of course.

Then it’s almost like cached queries, in a way. Somebody has made the decision that this is a known way to solve this problem in the enterprise. That’s why I don’t think the theory of full abstraction away from the interface, where everything is an API call, happens. Ironically, what will probably happen is that, in a couple of years, we’ll see agents rebuild entire web pages and dashboards. Then we’re going to find ourselves asking, “Wait, why are we having an agent? Why do I have to spend tokens to create something that is a config on a dashboard?” We’ll just be back to where we started for some amount of software, which will mean that these things are going to live together.

Cool. Let’s move from software to decision processes. I won’t say the name of the company, but I just spoke with a very, very legitimate company—a household name. It’s a private company, though; it’s not a public company. At the board level, for every decision, they ask the AI for more information to inform the decision. The founder was telling me, “It’s literally better than half of my board members.” It’s been great as discussion fodder, to be provocative, and it also shows how fundamentally unoriginal the board members are. How much have you thought about bringing AI in to help with decision processes?

Yeah. And by the way, I think the board is low-hanging fruit because boards tend not to have a lot of context about the business, and so the stakes are probably lower, anyways.

But is this something that you’ve thought about?

Well, the board one is an interesting one, so maybe we can unravel that one. I already use it for our earnings calls. We’ll do a draft of the initial earnings script, and because Box deals with unstructured data, I’ll load up the earnings script and use a better model to say, “Give me 10 points that analysts are going to ask about this, and how would I improve the script?” It just spits out a bunch of things.

How good is it at predicting?

Oh, 100%, because it has access to every public earnings call in history. At the end of the day, analysts can only ask you about tailwinds, headwinds, and who’s buying what. It’s not because analysts are smart or not smart; those are just the things you would try to deduce from an earnings call when buying a stock.

You wouldn’t have thought of these questions beforehand, or is it just a margin on the margin?

No. What I’m using is the specific parts of the document that are missing the answers to those questions, so I can inject the answers into it. You’re typing out a thing and realize, “I forgot to give 2 case studies in this section,” or whatever. It’s a quick way to do some analysis on something.

It’s funny: Bezos famously had this memo-oriented, essay-oriented kind of meeting structure. We never did that, but I was always fascinated by the companies that could do it. We’re entering a world where you could probably just pull that off. Whether it’s a board meeting or a product meeting, you could do a quick deep-research essay on the topic. Obviously, every strategy meeting in history would be better off if you had that as a starting asset.

I think the argument against that would be that the reason Bezos said to do it was because it forced people to think clearly about what they were doing and write it down. The exercise meant that the people walking into the meeting had more context.

This would almost argue that they would have less context because something else did the thinking.

Two things. It was to make sure that the person doing the thing had the clarity to write it, for sure, but it was also to inform everybody else who didn’t do that work. It certainly would have helped everybody else in the room.

I’m not 100% sure. We should do a full longitudinal analysis of whether the people who wrote the essay actually had the better products. There are some Amazon products I don’t like, and so did they obviously write an essay for those, too? I don’t know the hit rate on the essay specifically as much as I like the idea of writing down a strategy and thinking it through. Why not have an agent do 90% of the heavy lifting?

A lot of my workflows are like this: If I have a topic where the direct change in my workflow might be the kind of thing that, 3 years ago, I would have lobbed over to the chief of staff and said, “Hey, can you go research the pricing strategy of this ecosystem or something?” that’s just a deep-research query now. I can wake up and have it.

But what does that now take? What are the trade-offs?

I just do it for the most random things, which means I’m expanding and exploring way more spaces mentally than I would have before. This is equally why I’m actually more optimistic on the jobs front, because what we do too many times with AI is look at today’s way of working and say, “AI will come in and take 30% of that.” No, no, no—we’ll just do totally different things with AI. I wouldn’t have researched that thing before, when it required a person to research it, because that would have been an inane task to send to somebody.

Yeah. So one thing: When we run the numbers—and by “run the numbers,” I mean look through how AI companies are doing—where does the value accrue? There’s basically one takeaway, and that is that these markets are very large and growing very fast, and value is accruing at every layer, from literally chips up to apps. The only real sin is zero-sum thinking: “The models aren’t going to be defensible,” or whatever your zero-sum thinking is. That just hasn’t proven out.

Now, this is still largely a consumer phenomenon. What I’ve been thinking about, and I don’t have an answer to this, is enterprise budgets: You can’t just create budget out of thin air. You have a limited resource. As budgets get reallocated, to what extent do you think this is zero-sum—the old budgets get robbed—versus budget-accretive? How do you think about that? Where we’ve come from, that has not been an issue. I think in the enterprise it probably will be.

Fully logical. A large number for a startup can also be a very small number for a large corporation. You have that dynamic playing out. You could probably take a meaningful engineering team, and for the price of 5 or 10 of those engineers, you could probably pay for Cursor licenses for the entire engineering team. This would argue that it’s actually coming out of headcount.

Here’s where the asterisk is: There’s an infinite set of ways that this actually plays out. This is why you can never take a point-in-time snapshot on these kinds of things. Next year’s planning process might work like this: In a perfectly parallel universe, the salary increase that year would have been 3.5% for employees, but this coming year it’s 3% because we’re going to take 0.5% and deploy AI for the company.

Or maybe next year we’re not going to add the 50 engineers we would have added; we’re going to add 25 and pay for AI. But guess what? The year after that, we’re going to see engineering productivity gains, so salaries increase because it’s still a competitive environment. We then add engineers the year later because we’re getting higher productivity gains.

I think that most companies of any reasonable scale—past 100 employees, let’s say—have enough dynamism in their financial model within a 1- to 2-year period. This is where it doesn’t look like what an economist would think it looks like.

Can I just spit this back? I think this is actually a very good point that’s buried in there. I just want to make sure I’m following along, which is: The software license cost to a startup relative to a large people organization is relatively small. It’s just a couple of headcount, which, if you look at normal performance management, normal attrition, normal variability, and even hiring timelines, is kind of in the noise. You already have an annual budgeting cycle to fix that up.

Basically, within the noise even of just headcount planning, all of this could work out without some massive disruption.

Totally. There could be an upper limit to this point, but let’s say the going rate in Silicon Valley for a new engineer coming out of college is somewhere between $125,000 and $200,000. I’m just making that up. Let’s say your most aggressive Cursor usage or something is $1,000 or $2,000 a year. So you’re at maybe 1% of salary.

And that’s just not—here’s the question. Again, let’s do this crazy apples-to-apples thing. If you went and recruited from Stanford right now and said, “Okay, you Stanford grad, have a choice: You can work at this company and get paid $125,000 with no AI, or you can get paid $123,000 with full access to AI. Which one are you going to do?” They would do the $123,000 all day long.

Yeah. But even then, your argument, which makes a lot of sense to me, is that it’s kind of on the margin when it comes to the total. But just as a way of exploring why these things are not the high-order bit of the cost increase on budgets—

No, I love that. That’s great. I did one kind of late-night modeling exercise once, and I’m afraid to say all the numbers here because I think they’re just going to be so wrong. But something on the order of $5 trillion or $6 trillion in knowledge-worker headcount spend in the US.

Yeah, everybody says for developers—I think they say 40 million. Let’s just say it’s 30 million. Let’s say the average is $100,000; you’re at $3 trillion. These are just massive numbers. So it’s many trillions.

Yeah. So you have many trillions of dollars. If you take a couple percent of that, or 5% of that, you’re already doubling the entire US enterprise software spend. You can just make it work within that. This is why I don’t think people will make cuts because they have to pay for AI. They might make cuts for other reasons, but even in those cases, I think you’ll often have it be for myopic reasons temporarily.

There’s enough flexibility to basically consume this and then actually recoup the productivity gains. I think that’s great. I try to parse everything you say through the lens of where you’re landing on AI coding, and you seem to have a very pragmatic view of where things actually are at. Where are you landing right now?

Well, it’s been an evolution. I would say, in the entire AI thing, the biggest surprise to me is how effective it is at code. My sense is—I’m just going to say a couple of facts, and then we can back out what this means in aggregate.

One fact is that I do think AI helps better developers more than not-better developers, and the reason is you just have to be able to know what to ask for and know how to deal with the outcome. Someone said it—I thought beautifully—on X. I forgot who it was, but I thought it encapsulated it. He said, “Ninety percent of what I know, the value of it has gone to zero, but 10% has tripled, more than 10x, or whatever it is—100x.” I think that’s exactly right.

I do think that for a lot of rote use cases, AI can do it and it doesn’t need to be double-checked. There’s a lot of things, to your point, like prototyping and scripting. If you look at usage of OpenAI, the primary use is actually professional developers, which means it’s part of a developer workflow.

And then probably the most controversial stance I have—and this is probably sunk-cost fallacy because I’ve been a programmer for a long time, and my PhD is in computer science, so maybe this is sunk-cost fallacy—is that I just don’t see a world where you get rid of formal programming languages. They arose out of natural languages for a reason: We started with English, and then we made programming languages so that we could formally describe stuff. It would be kind of a regression to go back.

I still think we’ll use languages. Maybe they’ll change, maybe they’ll be more like a scripting language, but I think the existing tool set will evolve and it’ll still be professional developers. I think we’ll still have developers and developer tools. That’s kind of where I am.

No, I’m fully on the exact same page. The fun thing to me is how AI coding is just at the tip of the iceberg of agentic automation. It’s the best thing to first get experience with agentic automation, but I think you’ll see this in basically every other space.

What’s so fun is that, in a 1-year shift, let’s say, the nature of the relationship with the AI has changed. If you think about the GitHub Copilot moment, it was, “Oh, this thing is incredible. It’s going to type ahead and predict what I’m typing,” and then you’re basically using it to work 20% or 30% faster, deciding which parts of it you take on or not.

Now the relationship is totally different within, again, a 1- or 2-year period, where you’re using Cursor, Windsurf, or whatever, and the agent is generating this chunk of output. Then you’re just reviewing it. What’s incredible is that none of your expertise is any less valuable in that review. In fact, it’s probably even more important than ever before, because in some cases it’ll just be wrong 3% of the time, and you review it, but then you’re literally doing 3x the amount of output.

The nature of how that changes both programming and everything else—why not have that for basically everything? This is sort of this new way that both software should work and actually will work. The big joke a year after ChatGPT was, “Okay, this thing generates a legal case and it’s wrong 10% of the time.” It’s like, actually, hold on—that’s not necessarily a problem. If you think about what this new paradigm of work looks like, sure, it should be wrong 1% of the time.

But the job is: You deploy a task, it generates a thing, and it comes back. You should expect that it’s 2% wrong, and then your job is to go and fix its errors. It’s such a weird inversion. It used to be that the AI was fixing your errors. That’s what we thought the AI was going to be. Now it’s a total flip: The human’s job is to fix the AI’s errors, and that’s the new way that we are going to work.

Right. So this begs a very obvious question, but I’m going to work up to the question. There’s a great paper at NeurIPS from an MIT team that basically says you can optimize a running system with agents.

The way they did it is they basically had a teacher agent and then more junior agents. The more junior agents would go try a bunch of things, and of course they had much more knowledge of the literature than any single human being. So they tried all different things, and then the senior agent would say, “Oh, this is good; this isn’t good.” Once it optimized the system, they would use it.

The person running it—the human being—was then helping the teacher agent decide what the parameters were, what was good, what was not good, and providing high-level direction. You’re already starting to see cases where human beings are running multiple agents, and even that is already starting to have some kind of bifurcation.

One way to think about it is that in any R&D organization, of course people start as ICs, but then they very quickly get interns and go into management. Maybe we’re just skipping that step. The obvious question is: What happens to entry-level engineers? Does this change how people get introduced to computer science, for example?

The cool thing is probably more people will now get introduced to computer science, because anybody can learn it. It’s been 25 years for me, but in the early days of programming basic applications or putting up websites, it was just extremely frustrating. You would spend days and days being like, “Why does that thing not work?” I had very few resources for figuring out why the thing didn’t work.

It would have been 100x easier if I could have had an agent write the thing. I would have learned 10x faster. Honestly, what you did—not 25 years ago, but 10 years ago—was go to Stack Overflow. It’s the slow version.

Think about how many people missed the window before Stack Overflow and got pushed out of the ecosystem because they were just like, “This is too frustrating.” You’re going to have a way bigger funnel at the top of people now learning programming and computer science.

I think a similar percentage of people will fall out, so it’s not like you’re going to get a 10x increase in programmers, because you still have to enjoy it and like solving problems. It’s going to change the nature of the incoming class of engineers that you hire. They literally will not be able to code without AI assisting them.

And it's not 100% obvious that's a bad thing because, assuming you have the internet and the site stays up, we should have access to the agents. I think it's mostly just that we have to adapt how we think about the role of an engineer and what these tools are giving us in terms of productivity gains.

I meet with a lot of larger, non-tech-oriented companies as customers, and generally the thing I'm recommending is: hire a bunch of these people, because they're going to flip your company on its head in terms of how much faster the organization can run. I do understand—I want to be sympathetic to the job market for anybody coming out of college, because I don't think it's easy right now, and it probably hasn't been easy in a number of years.

If you're graduating, the thing I would be selling to any corporation, some way or another, is that if you're AI-native right now coming out of college, the amount you can teach a company is unbelievable. Conversely, if you're a company, you should actually be prioritizing this talent. It's like, why does it take you guys 2 weeks to research a market to enter? I can do that with Deep Research and get an answer to you in 30 minutes. They will be able to show companies much faster ways of working.

Do you think there's any stumbling into problems this way, where you kind of adopt too quickly and get into a morass you can't get out of? Or do you think at this point it's pretty clear the stuff can be practically consumed?

What would the morass be that you'd get into?

You hire a bunch of vibe coders, and then they create something that nobody can maintain. It's really a total mess, which, by the way, I will say I have seen.

Yeah. You could easily overdo this whole thing. I think, as with anything, deploying these strategies in moderation while we're all collectively still getting the technology to work better and better is super important, as is understanding the consequences of these systems.

This is not a moment to just have your whole company vibe-code. I will say, one of my favorite things that I'm witnessing in the whole coding thing—and I don't know, the point of this talk is AI in the enterprise generally, but the coding thing is so salient—is that a lot of the OG programmers that I've known for a long time, who are off creating companies or are CEOs of public companies like yourself, are all back to programming.

You talk to them, and many of them are like, “I code most nights with Cursor,” just because it's really enjoyable. The reason I didn't code before is because I just couldn't keep up with the fucking frameworks. I'm like, “Dude, I don't know how to install the fucking thing. What is this Python environment stuff?” I just didn't want to learn all of that. You're literally learning bad design choices that somebody else just made up. They're not fundamental to the laws of the universe, and they don't make you any smarter. It's just a waste of brain space.

Well, the amount of frustration I have when I look through, let's say, our product roadmap—pre-AI, although this still obviously happens because we haven't fully transformed everything about how we work—is when you would see things like, “We have to upgrade the Python library in this particular product,” and it's 3 engineers for 2 quarters.

Exactly. At the end of that project, 0 customers will notice that we did something. We resolved some fringe vulnerability that is not even going to happen, but you have to do it because there's some compliance requirement where you have to make sure you're on the latest version. It's super important, but the thing is never going to happen, and all of a sudden you're wasting hundreds of thousands of dollars of engineering time.

The fact that that's now a Codex task is just unbelievable. The amount of things that you can now relieve your team to go and work on is incredible.

And the other big boon for the economy—and this is again where the economists just totally missed this stuff—is to think about every small business on the planet, of which there are millions, tens of millions, whatever. For the first time ever in history, they have access to resources that are somewhat approximate to the resources of a large company.

They can do any marketing campaign. Did you see the NBA Finals video from Kling AI? The Veo 3 video?

You can now put together an otherwise million-dollar marketing video for a couple hundred dollars of tokens. Apply that to every domain and every service area: I can run a campaign that translates into every language. I can have this long tail of bugs that I never got around to automatically get solved. I can have the analysis of a top-tier consulting firm done for my particular business.

For the people or companies that are resourceful, creative, and imaginative, the access to resources right now is truly unprecedented.

What do you think is the best metric for anybody interested in tracking this stuff, as far as how fast it's going? Is it GDP? Is it margin? Is it top-line? Is it headcount growth? Is it all of the above? How do you measure it?

Internally, we've explicitly taken the stance that we want to use AI to increase the capacity and capability of the company. Just do more. Whatever you track, make sure it happens—do more or do it faster in a given time period.

That somewhat relieves the pressure from people who think this is about cost-cutting. It's just, no, do more right now. Let's figure out what works. Some things won't work. We want experimentation, so just use AI to do more.

Okay, so that's us. The way you should measure that, then, in a couple of years from now is that either the growth rate of the company should be faster, or the amount of things that we're collectively doing should be greater.

Yeah.

The only reason that wouldn't show up in growth rate is that every other company also does more, and so that gets competed away, which is also a very viable outcome. This is just the new standard of running a business.

But there's no shift in the equilibrium, right?

There's no shift in the equilibrium. You just have to do it.

And the ultimate product of all of that is some other kind of metric of satisfaction. Our products get better. It could be the Consumer Price Index or something.

But did the iPhone show up in GDP? I don't know, but my life is better with the iPhone than without the iPhone. I'm pretty sure it did.

Okay, fine.

But it would ultimately show up in new cures to diseases and better health care. I don't know that the dollars would move around all that differently as much as life expectancy should go up, the cost of housing should go down—these are weird metrics that productivity gains will drive, but that economists wouldn't naturally associate with enterprise software and AI.

And this is, by the way, where I am: clearly there's a disruption because marginal costs are going down in a bunch of things, like writing code and language reasoning and whatever. Some companies will take advantage of that, but I don't think the fundamental equilibrium changes. I think, to your point, we just do more tech, products get better faster, and we solve problems that we haven't solved before. But it's not asymmetric.

The way I think about it is that if we go back to 1985 and just looked at how everybody works, I think we would be totally stunned by how slow everything is and how long it took to research something, analyze a market, create a campaign, or whatever.

It has now been baked into our human productivity that we just do all those things really fast. In 10 years, when we all have AI agents running around, we will look back to today and be like, “How did we function? You spent 2 weeks deciding on the message for the marketing campaign? How is that possible?”

What we do now is run 50 experiments with AI agents. They all come back with versions, we look at them all together, and then we make a decision in an hour and move on. That's obviously how work works, and that's what we will be saying 10 years from now.

Do you think we'll ever saturate the consumer? I caveat this by saying this comes up at every one of these inflection points, so I wanted to ask it again for the umpteenth time.

I'll say yes, just because at some point, maybe. But my list of purely consumer demands has not gone down. Health care is a totally unmet need that I have. I don't like to go to doctors or dentists or anybody because of just how hard it is to get scheduled.

I mean, buying a car, man. There are so many things that just need to be sold. The cost of housing—we clearly don't have enough houses. Now, where will AI drive that? Maybe robotics would be the play there, but I don't think we're anywhere close to consumer satisfaction or satisfying all the needs of consumers.

Well, I actually meant more that things change so fast that they saturate the ability to adopt new things. I do think that is certainly possible.

I think I track my parents as a decent kind of proxy, or even just friends—college friends who aren't particularly in tech—and they're still in their ChatGPT phase of adoption. They haven't moved on from that. They haven't made a Veo video yet; they're just using ChatGPT to ask questions about the life experiences they have. So, ironically, maybe one of the problems was that ChatGPT was so good that, if you imagine what people thought AI should be able to do for them, it already met 80% of what they would have projected. We know it can probably still do 10 to 20 times more, but their needs are going to be satisfied for some time on those core use cases.

Yeah. So I think this is true for the most basic consumer-query-type things. But this is the opportunity for startups: AI will show up in ways where maybe the person isn't even in the market for an AI thing; they just want a better version of that. This could simply be another market constraint: as soon as it saturates, you just make the product better.

Yeah. If I could just get better healthcare, I don't need to think about that as an AI problem or not an AI problem, but AI will be behind the scenes delivering that. Then I don't think you're saturated anytime soon.

The consumption capacity becomes another market constraint, but there are a ton of other ways that you can improve things. That's great. I love that you're so optimistic.

I am. I'm 98th-percentile optimistic.

All right. So I think we've had a fairly pragmatic conversation about the current impacts and the near-term impacts. If you take a longer view, can you dare to guess what things look like in 5 to 10 years?

I think Sam Altman and Jack Altman had a podcast recently, and it was very good. I'm going to paraphrase, probably in some wrong way, but they were going back and forth about how we just got what we would have predicted as AGI 5 years ago. We use it, and it's now built into the most anticlimactic UI.

I think that's my instinct for a lot of this: 5 years, 10 years, whatever your number is. And this is why I'm so optimistic about society and jobs and all this stuff: I don't think it's the crazy Terminator scenario where we automate everything away. I think the human capacity to want to solve new problems, create new products, serve customers in new ways, deliver better healthcare, and try to do scientific discovery is built in, and it will continue. AI is this kind of up-leveling of the tools that we use to do all those things.

I think the way we work will be totally different in 5 years or 10 years, but you're already seeing enough of what it will probably look like that I think it's an extrapolation of that. When you want the marketing campaign done, you have a set of agents that go and create the assets, choose the markets, and figure out the ad plan. Then you have a few people review it and debate and say, “Okay, let's go in this direction instead.” Then you deploy it, and you're on to the next thing.

Each company's units of output grow as a result of that growth. We're all still in competitive spaces, so some of it gets competed out, and others will keep growing faster than they would have before. So they'll hire more people, and you'll have new types of jobs. We'll have jobs for people just to manage agents, and you'll have operations teams. Adam D'Angelo had this cool role that just got announced.

Really cool.

Yeah, the role is to work with Adam at Quora and figure out which workflows can be automated with AI. I think you'll have a lot of those kinds of functions.

I think one of the exciting things about at least being in Silicon Valley—or anybody tuning in and being in this ecosystem—is that we're seeing the change happen faster here, and it's going to be 5 or 10 years of this rolling out to the rest of the economy. I think we'll spend the next 5 years making the technology actually deliver on the things that we're all collectively talking about: making it more and more robust, increasing the accuracy, bringing the costs down, and improving the workflows it can tie into. We'll be working on that for quite some time.

And you think ultimately this leads to the biggest piece of the dividend being better products for users and a better user experience?

Yeah. I think the software gets better, our healthcare gets better, and life sciences discoveries increase. I think it's all a net positive for society.

I love it.