a16z Podcast - AI Inside the Enterprise

Episode Date: April 24, 2026

Steven Sinofsky, board partner at a16z, Aaron Levie, CEO of Box, and Martin Casado, general partner at a16z, discuss the reality of AI inside enterprises. They cover the gap between Silicon Valley and the rest of the world, why most AI initiatives fail in large organizations, and how agents, infrastructure, and workflows are evolving beyond the hype.   Resources: Follow Aaron Levie on X: https://twitter.com/levie Follow Steve Sinofsky on X: https://twitter.com/stevesi Follow Martin Casado on X: https://twitter.com/martin_casado Follow Erik Torenberg on X: https://twitter.com/eriktorenberg Stay Updated:Find a16z on YouTube: YouTubeFind a16z on XFind a16z on LinkedInListen to the a16z Show on SpotifyListen to the a16z Show on Apple PodcastsFollow our host: https://twitter.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

So the board goes to the CEO, and what does the board say? We need more AI. And what does the CEO said? Oh, okay, I'll get like a consultant to do more AI. And then they have some centralized project that nobody knows how it works. They haven't aligned their operations and those things will fail. The funniest concept that the more code we write, the less we would need engineers, it would be the opposite because now your systems are even more complex than before,

which means that you're going to be running into even more challenges of when you need to do a system upgrade or when there's downtime and you have to figure out, like, well, how do I fix that problem, when there's a security incident. So, yeah, we're just getting started with the jobs on this run. They're going to hit a wall at integration. The thing that's not different about AI and that agents don't fix, that nothing fix, is that any enterprise of a thousand people or more or that's older than 10 years is just a massive stuff that's sitting there waiting to be integrated.

And you can't just say it's going to integrate. AI actually doesn't help to integrate anything. AI feels like it's moving fast, and for many companies, the real transformation is just getting started. There's a growing gap between what's possible in Silicon Valley and what's being deployed inside large organizations. Engineers are already shipping with agents and new workflows, while enterprises are beginning to adapt those capabilities to more complex systems and real-world use cases. That creates a moment of opportunity. The tools are getting more powerful and,

Companies are learning how to integrate them into existing workflows, data systems, and decision-making processes. At the same time, there's a deeper shift underway. AI isn't just another layer of software. It's starting to act more like a new kind of user, one that pushes companies to rethink how systems, permissions, and workflows are designed. In this episode, Steven Sinovsky, board partner at A16C, Aaron Levy, CEO of Box, and Martine Casado, General Partner, at A16Z, discuss what's working in Enterprise today and where the transformation is heading. Hey, we are here monitoring the situation live, and we're very excited to talk about a bunch of AI stuff.

And we have three of us are here today. There's me, Steven Sinovsky, and Martin Casado, who will wave and say, hi, I'm Martin. Oh, hi, Martin. And Aaron Levy, who is working on the elevation of his hair today. So we're excited about that. It just keeps getting more vertical. And I thought I could kind of tame it, but it didn't work. And is that just a token issue or a parameter, number of parameters issue with you?

It's too many parameters. Too many parameters. Okay. I have the same thing but in reverse. Okay, so. Listen, you have a distilled model. There, yeah. I run local.

So we had a lot of, there's been a busy week of things, but we want to bubble it up a little bit and just start talking about where, where things are heading. But I'll let, I just kick it to you, Aaron, and you start where you are the most excited this moment because you have visited a ton of customers this week and have learned a lot. You've shared a lot on X,

but I think you're the most in the trenches CEO who is really talking to customers every single day in the enterprise, which is what the three of us tend to look at the most. Yeah, I think my, it feels like my job these days is just bring reality to the valley and then bring the valley to reality as much as possible.

And it is a kind of a crazy divide that exists at the moment. Well, take a step back. I actually think it's super interesting. What is it? What's the gap caused by? What is that? Yeah, well, I think the gap is,

and Martina, I'm sure you see this, but I think the gap is caused by the styles of work that exist in Silicon Valley and in engineering roles. versus sort of the rest of the world. And we've talked about this a couple of times in different forms, but the technical aptitude of an engineer is just like insanely high. The level of wired in this to what's going on on the Internet is insanely high.

The ability to use your own tools and make your own choices is insanely high. And when things go wrong with the systems that you choose, you can just like quickly debug them and then make them sort of work for you. And then obviously you have all the benefits of just the models are really good at code and the work is verifiable. So you have like, you know, five or ten things that make agents work in an enterprise context for engineering, or at least even a startup context for engineering, that tend to be a gulf between the way you work that way in engineering and the rest of sort of knowledge work.

And so a lot of what I see is trying to figure out how do we kind of, you know, bottle up all of the greatness that is, you know, what we are seeing from coding agents, what we're seeing from agents that can use computers, to how do you bring that into the enterprise, where the workflows are, you know, quite different. The users are less technical. The data is much more fragmented. The systems are much more legacy. And so that tends to be the divide.

So it's not even that we're like talking past each other, like in one of those kind of classic like government versus industry. It's just literally like there is just a pure workflow and technology set divide. And that's why it's going to be, you know, a number of years for this sort of diffusion to roll from what we're seeing in Silicon Valley. What we're seeing is tech startups all around the world into the rest of knowledge work. Martin, just to build a – you have a ton of experience in big companies. One of the other issues, though, is scale.

And the difference in scale that Silicon Valley operates at the startup level versus everyone else. Yeah, I also think that – I mean, these secular trends, like the Internet was like this, actually start with individuals. And big companies tend to make decisions. centrally. And this is one of the fastest growing secular trends. So like, there's probably a lot of individuals in big companies that are doing it where like, yes, the big companies themselves don't know even how to think about it. And so when you hear stats like, oh, like MIT had this stat, like 95% of AI efforts in big companies fail, like that's clearly silly because I am sure

everybody is using chat GPT very effectively. What they really should be saying is, you know, whatever. Like I sit in these boards too. So the board, goes to the CEO, what does the board say? We need more AI. And what does the CEO said? Oh, okay, I'll get like a consultant to do more AI. And then they have some centralized project that nobody knows how it works. They haven't aligned their operations and those things will fail. And so I don't, you know, when we say scale, often we think about things like system scale or number of people. I think the secular trend is scaling wonderfully, which is being reflected in the numbers of these companies. But organizations don't know how to adjust these kind of, you know, ageal processes

that have been, you know, worked on for a decade around, you know, data and governance and operations and compliance, et cetera, that's kind of right now where I think, like, Aaron is right between the secular trend and the organizational decision body. And this is something that we actually tracked very closely because we're starting to see now, I would say, in the last few months, finally some real kind of inroads into the enterprise, but it's, it's tepid. Because, and the last thing I'll say at this, one of the reasons is there's a lot of sketch. because the board wants AI, CEO, AI, failures have created some amount of bruising,

which is, you know, requiring these companies to get past it in order to do kind of the second go at it. And so I feel this is exactly where we are. Yeah, I 100% agree with that, which is it's good to start with agreements because we know how quickly that fade. We'll disagree the rest of the show. Exactly. Exactly. It's the only time we're going to agree.

I think maybe one more point on the board for agreements, Maybe you guys would agree. There's also this very interesting dynamic. I would say this is a minor one relative to everything else. It's probably 5% of the problem. I think it would be more fun to talk about the real problem. But there's a fun kind of, as an aside, there's a fun dynamic where you go to an engineering team

classically for the past, you know, and, you know, Stephen, you can take us back in history on this one. And one of like the easiest ways to stall a project was just getting the architecture, you know, kind of the fights on, you know, language to use, what architecture path to go down. That could take months and months to kind of work through as your teams work through that. Because of the pace of change in AI, you actually have this incredible dynamic where the labs are obviously leapfrogging each other so frequently, but

with not the exact same paradigm of how you should deploy agents and how they will work, and is the agent harness in the computer, is it outside the computer, do you run it in your cloud, What is it hosted? What tools does it have access to? Like, we are, like, this is not a point where these are completely fungible technologies. And so that actually creates a bit of paralysis because now as an enterprise architecture team in the real world, you're like, man, like, what horse do I want to, you know, kind of get behind?

And which architecture path do I want to get behind? Because I've been burned by doing the wrong thing in AI maybe three or four years ago. And I went down at some path that now is deprecated or not the right strategy anymore. So to some extent, the speed of our change in tech actually reduces the ability for the tech to get diffused into the really, really important workflows because now you have a lot of paralysis in just making decisions.

So I actually think it's kind of fine because there's still so much upgrade work people need to do in their infrastructure and their systems and their data. But this is kind of an interesting dynamic where I'll go have conversations with CIOs and their AI teams and I'll say, hey, what are you using for your chat system

or your core agent orchestration? and they'll say, yeah, we're in the middle of a debate between these two or three paradigms. Right, right. And you hear that across almost every single customer because there is a little bit of a nervousness of like, who do you get in bed with and how much do you sort of, you know,

fully lock yourself into one particular path? And we also know that if you don't lock yourself into a path, it's always, then you're building for this sort of duality, which is, you know, also takes a lot of work architecturally. Oh, actually, I hate to jump in, Steve. I just like, there's a, like, so Aaron is totally correct, and there's a very, specific instance of this playing out in product companies right now.

And I'll tell you what it is. So software product companies, you know, circa six months ago, they viewed integrating AI was like, you're actually integrating it into the product, right? So everybody was like adding like whatever, this chat feature or like, you know, and so it's kind of like this fusion or this hybrid model. What we're seeing instead is instead of viewing AI as software, like just view it as a user. And so instead, like, take your product, make it a CLI tool, and then have the AI be an agent that actually uses it. So you're not fusing the two. You're just making it more useful for

AI. This is a very, very significant architectural and mental shift, right? So we started as pure product, and then we didn't quite know what the end thing looked like. So we created this, you know, this, you know, AI software hybrid that hasn't worked. And now we're kind of going to the agentic model, which basically means the agent is going to be whatever, it's going to be whatever, be cloud code or whatever, and then my product now just should be something that consumed by that, and like, that's the actual modality. But, you know, within a year now, you've had to re-architect your software twice. And so I think no matter, but many places that you look in the industry is having this dilemma of actually trying to figure out what the

final form looks like. And Stephen, you will remember, remember all the hybrid versions of cloud? Yeah. Yeah. Remember, like, you know, like remote desktop and all these things? Like, I think kind of like speed running that evolution to the final form. Right. And I think that that people in Silicon Valley don't quite appreciate when a big company says, well, we have to map out our bet that we're going to make because like that just seems stupid. And, you know, if your job history is, you know, five two-year stints at startups that went from seed to series A to aqua hire or something.

Yeah, he didn't learn anything. Well, you never, you don't, your frame of reference is not, you know, picking an account's payable system that's going to last 40 years. Yeah, I actually, I have like all these visual aids today. So here's like the ultimate, the ultimate engineer if you're in Silicon Valley is lower, lower, lower, lower. Yeah, yeah, is Gilfoyle. And Gilfoyle is like, I don't want to talk to anyone. Yes. And I will just write the code and you go do your thing.

And the thing is that you have people in enterprises that are saying, I'm going to use the model and do my thing, but they're going to hit a wall at integration. And the thing that's not different about AI and that agents don't fix, that nothing fix, is that any enterprise of a thousand people or more or that's older than 10 years is just a massive stuff that's sitting there waiting to be integrated. And you can't just say it's going to integrate. AI actually doesn't help to integrate anything.

And so even if you change everything, the people say, oh, no, if you make it an agent, then it can just go ahead and be a user. But if you're a user, like if you've ever called customer service for something, like literally you get bounced to another human if the system that you're talking to doesn't work. And they're like, well, that's. a manager or no, you're talking about payment, not reservations. And so, like, we're, we're, what I think is so exciting is that now we have proof of this technology that everybody likes it.

I mean, you see all the people who don't like AI are saying, look at what's happening in law firms because people are seeing hallucinations and it's ruining legal cases and all this. And, and the reason that's happening is because the 25 year old associate is the one using AI successfully already and had been using it for a year. Steve, it's actually a little worse than that, where it is right now many companies are incentivizing people to use AI by counting tokens. Oh, yeah, yeah, yeah.

And so I'm not going to say the name of the guy. I spoke to someone yesterday who worked for one of these large companies that famously does this, and he's like, me and my coworkers have agents do useless tasks just so that we can, no joke. No, no, totally, well, you get whatever you measure. Yeah, yeah, that's right. So, like, it's like the extreme form of what you're saying, Steve.

Oh, yeah, yeah, yeah. People that are, like, being fake productive and producing a lot of, you know, you can say perhaps problematic artifacts just because they're using these models. When the Internet happened, all of a sudden, every company needed websites. And so, like, a very famous moment in time was not too long ago when every internal team had, like, a team website. And they went out and they got, like, a vendor to write HTML and to create their site. And then there was a team.

But like, there's nothing dumber than having a team website at a large company because a team gets reorganized like six months later. And so companies were just filled with like, with thousands of these dead web is what the expression was. But I think, but I think, but go ahead, good, no, no, but we should, we should drill into your integration point because I do think this is something for, you know, sort of some reality to settle in in the valley on on the real world's sort of journey to fully being agentified and what that's going to take and what that's going to look like and your your point about

being passed to the different human you know based on the role that you need to interact with you know agents basically don't have any there's no real exception yet for the agent having the same problem because you basically you know as you pass through a different human it's it's a different set of access controls that that that human has and if an agent can sort of buy bypass any of those steps, then that's how you instantly get the security risks that, like, you need to kind of pass through those steps so that way you don't accidentally, you know, get to the wrong piece of information and there's verification. And so there's a lot that you need to kind of build out for agents to be able to go and work with all these systems.

And we've talked about this, but like most legacy environments don't have the most authoritative, you know, access controls. So you're always as a human going and saying, hey, Sally, can you share that thing with me that I don't have access to or, hey, Bob, what's the number inside your data system for this question? And so if agents just get the exact same permissions that you had, then they'll just run into these walls everywhere and they won't be able to complete the process. And unlike a human, they're not going to know to go talk to Sally or ask the question of Bob. So they're going to just be kind of stuck. So what's going to happen is you're going to have a lot of agents that don't have access to the right data. They're kind of working through systems that

that are, you know, not the real sources of truth for the information. They're getting the wrong number. They're getting the wrong document. So this is the real work that enterprises have to go through right now. The good news is that it's actually a great time, again, if you're a startup, because you can just, you get to know all the problems right out of the gate. So you can design your organizations, you know, to try and avoid this. But for big companies, there's real work that goes into, how do I upgrade my systems?

How do I modernize my technology environment? How do I make sure that, you know, agents will have access to the, right data, the right documents, the right context to be able to do their work. And that's sort of the work ahead. And there's, you know, there was this, you know, kind of headline of open AI working with, in codex, you know, working with Accenture, Deloitte, all the major system integrators. And there were some kind of like, you know, snarky comments online around it that I was fascinated by because it sort of showed how, how maybe, you know, great that divide is from the rest of the world versus those in tech because to me it was like the most obvious announcement of all time,

which is a large enterprise is going to have to go through the change management, the systems implementation, the integration of technology for these agents to be able to go and work. And so there was this sort of like, you know, people thought it was somewhat ironic that, oh, we need people to implement the agents that are going to go automate the people. And it's like, no, that's exactly how it works. Right, right. You actually do need to do lots and lots of work to be able to be in a position where agents can actually go and help you do, you know, any of the job.

of the automation. So, so that is, and that's going to be, there's going to be businesses that are doing that for decades. Like, it's going to be an incredible opportunity for this kind of next generation set of firms as well as existing ones that, that lean into that. Let me throw this out there. Well, first, I think the other thing that people shouldn't celebrate when those fail because they will fail because they're, as Martin was describing,

a lot of them are going to be these sort of top down mandates where they picked like the most acute problem in the company and think, oh, AI is going to go solve that. And the IT people are going to be like, oh, God, that's the worst. Don't do that. That's the worst system to try to do that. But the CEO or CFO or whatever is going to be obsessed with solving or the, most likely the customer service person will be obsessed. But I do think if I were advising a startup specifically in order to sort of enter the

enterprise space in that way, definitely would be thinking about not just like building a company that step one, I only work with all the headless SaaS software that's out there. because there just won't be any. Like, but the thing you can do is structure the value that you offer. And also this applies to what you go do in a company is, it's really a fork. And the fork is, is this an agent that is seeking information and presenting it to some human? Or is this an agent that's supposed to go act and do something? Like is it acquiring or is it doing?

Because if it's, it turns out that's what happened with the internet. the internet got very, very valuable when the first step was just providing access to things to people. Yes. And like all of a sudden, all the sites that were like that literally did integration. Like, hey, I need expense reports but viewed by department or I need to see our current inventory status across like the two companies we've acquired. All of a sudden, the web became the integration point. And so I do think that if you just show up first and just say, hey, we can. can actually use agents to learn stuff about what's going on in a company. And in particular,

because you're here, Aaron, like learning across the files becomes way more possible than it ever was before. In fact, AI might be the first time that inside a company search can actually provide immediate value because the web just wasn't structured to deliver those results. And then you start to think, once you could bring them all together, then you can add like an agent that has an approve button or a reject button or something like that. Let me just try and provide, finally the point where I get to disagree. Uh-oh. We're in trouble now.

No, no, no, no. Well, you can get invited back. You're invited back, so good. No, no, I think this is a very legit view, but it's not the only view. And in light of AI, I think it's not the only kind of compelling view. So here's the other. So let me just try and rephrase it. So the current view is we've got like AI is software.

It works in a different way. we have a current set of systems and we have to integrate this new type of software with an existing system so that it can get access to data, it can do things, but in a safe way, right? So here's the kind of end-to-end argument of why this isn't

about evolving software systems. The end-to-endom argument is these LLMs are nondeterministic, they are smart, they deal with the long tail of complexity, and it turns out those are all things humans do too, and we've spent 40 years building

interfaces, processes, and design to deal with messy humans. And, you know, we know who to access and we have access control. And so if you have the mindset that an agent is more like a human and you hire the agent, you give it its own email address, it can access documents like humans can, it can log in, it can request the things that it needs, then it will be drafting on all of the process that we've put in place for humans, not for software. And so I would just encourage us, If we have this discussion, like, listen, I grew up like you guys in software. I always think of every system like software. But these models don't integrate well with software, actually.

I think it turns out in what we're learning as an industry is if you view them more like humans and you draft on the mechanisms we put in place for humans, they're much easier to integrate. Well, that's – and I think we're saying – I love that point. I think we agree with that for sure. I think the issue is humans have a bunch of extra benefits that the agent doesn't have. the human has a lot of context that we get for free by virtue of we can keep track of the myriad relationships that we've built in our organization

and the person to tap on the shoulder when we need something done or we need to get information, that's not documented in a company yet in a way that the agent can just sort of draft on. And so I think we all would agree that you can't treat this like software. You treat these as people accessing systems and tools, but they are at a, they're both at a massive advantage

that they can work in parallel at infinite scale, and they're at a disadvantage in that they don't know who to tap on the shoulder. Hey, I, listen, Aaron, I am all for agent onboarding. Like, you know, the agent comes and it goes to orientation. Yeah, yeah, yeah, yeah, yeah. And the CEO gives it the culture discussion and that every, I'm not kidding. No, you're probably right.

Every department does their pitch. Yes. This is what we do. Yes. I mean, I think, I actually honestly think, given the technical nature of these agents. Yeah. and how much entropy they have and kind of how unruly they are, we're going to have to go through the processes that we've refined around humans.

Because humans have all of those things. And so I just, you know, it's more about providing schools for them than somehow building some, you know, fancy index database. No, no, totally agree. I mean, of course, what I love about that is you just keep going with the analogy because what that is is it's the same argument that humanoid robots will be. the best kind of robot, which is we have a whole world designed for humans. And like I saw at the Consumer Electronics Show, I saw this robot go into an elevator. And then there was a button pushing

robot on the elevator. So because the robot was a tiny little thing that, like a Roomba on the floor, it couldn't push the button. So the same company that invented that robot invented a device you buy for the elevator that pushes the button. But then I asked, why did you need a device to push the button. And it was very interesting. They said because the elevators don't have systems that they can hook into as a robot, like there's no Wi-Fi press the button in the elevator capability. Yeah. There's no API for that. There's no, there is no headless version of the elevator. Yeah. And that's actually a great metaphor for like the problem, but I think that we're actually solving in the enterprise with these agents, which is we just, you know, we have two types of systems,

those for humans and those for software. And these tend to be more like humans, so we should draft on those as much as possible rather than try to retrofit them. Well, and so the big news last week was, I think, you know, Salesforce, I don't know if they surprised people or not, but, I mean, based on the reaction,

it seemed like it was maybe a surprise, they went full headless, and they basically said, you know, like we want to be used everywhere across all of the different agents. And I see that as a little bit of a bell weather because I think as Salesforce goes, so does a lot of enterprise software.

And I think a lot of people are going to try and, you know, have to figure out what is the new business model in this headless world. You know, do you charge a little bit of a small just API tax? Is there a seat for the agent? So there's obviously some work to do with that. And Stephen, I saw one of your tweets on, you know, some of the, you know, complications there. But I think as a moment, it's a big deal because I think it's a recognition that, you know,

software will be running in the background. It always has for machine users and applications, and now it is for these sort of probabilistic machine users or non-deterministic machine users. And what's cool, and where I think this gets pretty exciting is, you know, as soon as I saw that announcement, like I had like five to ten personal use cases where I would need, you know, the headless version of Salesforce

because I'm always doing just a tremendous amount of customer-related intelligence work. I'm going into a meeting. I need some information. I need to do, I'm going into a city, who should I be meeting with? And so if you imagine, you know, being able to run compute in the form of agents across all of your data systems, like the use cases become, you know, pretty wild around what that opens up. So I think this gives, I think this gives a lot of software platforms, all new use cases that they can tap into that where, again, you were normally constrained by the number of people on these platforms. But now the headless user can be, you know, 100 or 1,000 X the scale of those human users. So this is, I think, an exciting moment because as you have more of these agents running around and the headless user, you know, software modes, you just have, you know, way more use cases for these tools.

I also think, I just think on this one, what's so, what's so super cool is, is that, of course, the first step is doing exactly what you described, which is just looking stuff up. And so the most interesting thing is using this notion that an agent is just an entity, it's incredibly obvious to me that it's another license. Now, it might have a different license model, but it has to have an identity. Like when you go look something up in the box CRM system, I don't know if it's Salesforce or not. When you use the box CRM system,

it has to be a person like with a certain amount of access rights. And you presumably as CEO, you might have access to a bunch of stuff, but also there's a lot of ways that they actually don't want you to have the rights at the right time. Like you might be able to look and see who is on the account, but you don't need the up-to-date quota of those salespeople and stuff. And that might be HR sensitive. And you should probably have some other level to,

go see that. But as you go down the org, the agent is never going to have more permissions than the person who's getting it to go do something. And in fact, it's just going to be like a peer to somebody else in an organization. Because otherwise, you have all of these issues where the peer, where a human can just say, oh, get me the super smart agent that knows everything that I'm not allowed to know. And to the, in the IT architecture sense, what's so fascinating about that is you have to build, you can't let the agent get the results and then try to figure out what works or not. But suppose of all the points that Martine made about,

about the LLM stochastic model, which is you're not going to be able to figure out. It's not like a record in a sequel table that you could just apply Ackles to. It's actually like, it could be words in a sentence or just the number that shows up. And so I actually think that whole discussion about HeadList for me made the SaaS

Apocalypse seem even dumber than it was already, and it was already dumb. So, like, at first it was dumb. And then I'm like, oh, my God, it's actually much dumber than I thought it was in the first place, because you're just going to have this explosion. Now, someone might come with a very clever pricing scheme and that agents, you know, somehow cost less because maybe for the first five years the read only or they're always tied to a person or something. But it is another seat.

There is no way around it. And like, if you're a SaaS company, you're crazy to try to say, oh, just use the credentials of another human. Like, that's just, that would be like bad security practice from the get-go. Exactly. So actually, in fact, so this is playing out in many domains. You can even make the argument that, like, a headless SaaS doesn't make sense. And here's the argument. The argument is, well, let me give you an example.

So if you use OpenCla, do you know why you use a Mac Mini with OpenClaugh? Well, for two... Is that what you're getting at? It's not. It's two things. It's number one for iMessage. Right, right. It's for the integration.

Yeah, yeah, yeah. Because there is no headless version, so you're just going to, like, use it. And then the second one is very interesting, which is if you've tried to use headless browsers with agents, the problem is all of the websites have anti-scraping measures. Right, right. So they don't work.

And so the reason to use a back mini, so it can actually use Safari proper. So to do anything headless kind of assumes that the entire internet is going to go headless when I think all of these models, like all of the data is humans working on the actual apps that are not headless. Like that's all of the data anyway.

So I think these models are going to be very good at just using apps like they are today and we're already seeing this happen. And rather than the headless versions, the non-headless versions are what's actually being used. So you could argue that it's just sales. force. Like, not a headless.

Like, yeah, it will go to a browser. Wait, wait, actually just to clarify, right, do you literally mean the agent goes to the browser? Yes. Oh, no, no, I'm taking the other side on that one big time. Yeah, yeah. But let me simplify the argument so we could actually have it. Yeah, yeah.

So, so today if you use an agent like nanoclaw or open claw, you could use a headless browser. Let's say I wanted to like look up the value of my house on Zillow. The headless browser simply doesn't work. because Zillow's so tired of people scraping it, so it will detect headless browsers. So the thing that works is, you pop up Safari,

and it uses a proper Safari directly, right? And so then all of a sudden it works. And so... No, but, but, but, but, I think, hmm, I mean, I would just say that... And any software that has a good API, the agent would absolutely prefer to use the API, and then you pop into the browser

the moment that you run into some execution problem with it. And, you know, set as a fantastic long-term computer science software guy. However, these models are trained on data and RL environments from existing software that didn't have those APIs. Yeah. And right now, if you actually look at the adoption and the use of these agents, they look far more like what a human would do than what, like, a program would do. So maybe you're right, but A, that's not what we're seeing. and you can honestly make the end-to-end argument

when it comes to data and all of the controls in the internet, to Steven's points, all the existing controls, to just be like, these are going to actually have the same actions as humans. Well, the APIs most that, I mean, I think the APIs of any software provider

will follow the same access controls of whatever the user is that is. Right, but they have to rebuild it. I mean, it's like you've got this existing app and all the models trained on all of the people using the app. Well, on that point, and it's a totally fair point, but I would guess over time you're going to have really, you're going to have very, you know, kind of accurate, rigorous data sets

for, you know, for models to be trained against the MCPs of every SaaS platform, the APIs of every SaaS platform. Already it's in the, you know, they're already training against all of our documentation on our products and our APIs. But I just think, to me, it's more just an inefficiency of navigating through pixels versus just, you know, you can just do a quick. Yeah. An adage in systems, Aaron, is that layers never go away. They just get layered.

Well, so, but on that, I'll support your point, 50%. Oh, well, there you go. Yeah, yeah, yeah, yeah. I just think if you need to do a search for a document, our search API is going to be a faster way to do it than, you know, you know, clicking through a. 100% yeah. Right, but, but the, but the, to half support the point, like the new codex, the computer use that on the desktop is, is, you know, just insane. And I, I mean, Stephen obviously knows, you know, everything about how it would work.

I, when I saw my ability to move a mouse and then this other sort of mouse moving and clicking things, I was like, I don't understand computers anymore. Right, right. So there is a pretty, and to your point, Martine, my first instinct was to use it for something where I know there's no available API. So I did actually use it right away for a thing that I don't have access to the API. And an agent over time is going to probably have to figure out, is there an easy MCP or CLI for this action?

And if not, then I'm going to pop into some kind of cloud browser or cloud computer or maybe local thing that I can sort of parallel track and then go and execute that. So that does seem like a reasonable architecture. But I still think that I'm going to pound the Salesforce API massively in headless mode just because that'll be an efficient way to go look up records. Yeah, I mean, I think that you're sort of,

I think you're both saying the same thing, but there's just a time to, but no, but there's a time to mention. And I think, like, there was a moment at the internet that I really was thinking about when you guys were, when I was seeing the timescale difference, which was suddenly the 8 million, quadrillion pages of how to use word in Excel that we had written over the years

that we posted on the internet, we had used to ship them with the product and people would have them on their hard drive, not connected to anything, and they would say, like, how do I make a nice chart or whatever? And it never worked. They could never find the thing that they wanted.

But what happened with the internet was the net result of everybody finding it caused us to make better documentation, but it also caused Google search to be better at finding the information that it needed, which then completely changed the way that we thought about doing documentation. And I think that with headless, especially for the kind that's just finding things,

it's going to really change the way that information is exposed. And so the way that Salesforce sees today of exposing a headless API is I'm almost certain if I were to go look at it. It's going to look like the developer API in front of it behind a CLI. And it's going to look a lot like that. But in fact, that's not at all how humans using Salesforce interact as a human, trying to solve like I'm standing in the elevator, waiting to go see a customer,

what is the stuff I need to know? Like that mapping is completely different. And so that API is going to really change as a result over time. Yes, I think the API changes for sure. I agree that. But I do think that unlike the humanoid, you know, kind of comparison,

where sort of the physical world has interesting physics issues that you eventually run into, the digital world doesn't. And so at some point, your agent can run in parallel, you know, 500 times. And like I'm going into, I want to do a market map of customers across the Fortune 500. That agent can fan out and do that work in a way that I can't as a person in a browser. Right. So to some extent, agents get to let you, you know, sort of bend the laws of normal, you know, human-based workflows. And so then, like, that's why, and I think that that means the APIs maybe eventually evolve,

but not obviously in the direction of the end user product, but maybe more toward an agentic sort of set of workflows of what is that agent looking to do. Well, but Martine, I think, would jump in and just say, wait, you didn't describe anything new. You described an architectural, no, but you described an architectural problem with today's software, which is its API and Performance Gate was based on how much I can type. which is sort of the point I was making, which was our help system was designed on how much we could ship on one CD and had no data about what it is that people were trying to do and no context.

But it didn't change like the problem, which is I needed to make a chart. Yes, exactly. Well, so like one real example of this, we launched a box agent that gives, you know, that has, you know, a bunch of more capabilities built into it. And one of the capabilities is that it searches across your whole box environment. it doesn't have the same limitations of a human-based search where you type in one query, you get back a set of results,

you look through them. You know, it fans out, does multiple queries. It can look through hundreds of results instantly and do its own re-ranking of that. And so that's just like, you know, again, you wouldn't want to be rate limited by the same process that a human went through, which is where the humanoid robot is, you know, you're kind of willing to be like,

okay, the humanoid is still going to walk into the elevator, it's still going to press the button when actually, you know, In an agent world, you're like, no, no, I just want you to go and instantly press the floor that I'm going to. Yeah, yeah, but we should just be very clear that, by the way, I very much agree. But we need to make a distinction between, like, would you ever build an indexing that's only for AI and not for a human? And I think that's less obvious. Yeah. So clearly there's like performance gains based on automation.

We've got to involve our architectures for those. But if you find a great way to index documents and you don't expose it to a human, I think. Yeah, you got to. Yeah. Exactly, 100%. Well, this is, I mean, it's actually, and I think this kind of probably reinforces some of Stephen's kind of internet analogy on documentation.

There is this really interesting thing where, you know, it started out where we, as we've been building our next set of agents, we first gave it the current set of tools. We saw it how it used those. And then eventually we realized, oh, there's actually an even better way that the agent could do it. So we improve the underlying scaffolding. And then, oh, by the way, that will actually help the end user also. So it does let you sort of contribute back into the mothership of technology improvement

that does sort of lift all the boats of your users. Let me ask this. It occurs to me as you're saying it. Like I sort of got all tense when the idea became, no, like, oh, we have 10,000 people hitting our SaaS system today. And we've got it all working and it's all great. But now we're going to have 10,000 new people, which are the agents, for each of those 10,000 employees.

And they're actually hitting it 500 times as much. Okay, so that SaaS product will collapse. So, like, that's the first order, because it wasn't architected for that volume. Like, we saw this with all the BI tools. Like, when all the BI tools came out, all of a sudden, they were looking at the SAP data and trying to snapshot it and absorb the whole thing every night

for a new kind of set of slices and dice that your view across all 500. And like all the people making ERP were like, well, we don't do that. And so they had to go build all of this themselves because they had the knowledge of the data. Their API just couldn't, was not designed for that kind of lookup. So my sort of thing to throw out there and fight about is,

what is it, what does the change management look like in a company? Because you can't let loose an agent that hits the system at 500. X the humans. And it's not a token thing. It's an actual like, wow, we don't have the network bandwidth and the throughput to handle 500X for any one of our customers. So what happens? So I've got a provocative adjacency, which you guys can tell me if I'm doing too much on a tangent here. But here's my provocative adjacency, which is I don't know if having more agents is that big of an architectural shift. I just feel like we understand like whatever. If it's read-only data,

you cash it, you know, like all the state issues are around mutable, mutable globally shared state. We understand the limits of those. We know how to architect around this. We had to tackle all of those things when we went to the internet. And so if you built your system not to handle it, like you suck at building the system and you deserve to go down and just go build a system that doesn't suck. And like I just feel like this is kind of standard computer science. However, I do think agents do introduce something that organizations technically have to deal with. And let me just give the analogy in code.

I think, Stephen, this is what we call maugging on it. I don't know. You've been a question mocked. I have no idea what he just did, but I'm just looking forward to how he magically made the problem go away. But go ahead. No, no, no, no, no. No, the problem is there. I just think, like, we know how to go from 10 users to 1,000 users.

We just got rid of the stupid people. So now everybody is smart. No, no. Okay, so let me give you an example for coding. So this is where I actually think there's a shableness. shift in how work gets done. So when you code with AI, your code kind of gets worse over time pretty materially. And so it's almost like you're introducing as many problems as you are solutions. And I don't think we've actually figured out how to manage that. Does this make sense?

The whole world right now. Yeah. I mean, like this is this kind of reasonable question is, you know, if you're using AI, yes, you're productive, but are you creating more problems than you've actually solved for a solutions. And I do think that there is this actually, you know, open question when it comes to using agents on existing systems for creating things, which is like, like, do we know how to wrap the growing set of entropy around that? And I would say anecdotally, watching companies struggle with AI coding, which of course, I'm, you know, listen, I'm very close to many AI coding companies. I'm clearly very bullish on it. I don't think we know how to do that yet. And so the, you know,

the agents on a system, I think we can tackle those with known techniques, using agents for long-running things organizationally, where like, you know, the universe is kind of as clean as it was three days after than you started. I'm not actually quite sure we know how to do that at all. Well, I love that point because that gets back to where we started, which is the difference between scale and not scale. Yes, yes. And why it's perfectly rational for big company people to be, like,

like no freaking way is this coming into our company because a big company is about to, the wheels are going to come off a big company or a division in a big company or a product in a big company at any minute. Like if you're, Marcy, we were both giant company executives. Like literally we woke up every morning thinking, oh, the wheels are coming off today. This is the end of it. I'm getting fired by the five o'clock. And whatever started, I left yesterday thinking we were three months late and we're now nine

months late. And that's a typical day. And so, but the reason that that doesn't happen is because you put constraints all over the place, which is exactly why Guilfoyle can't work at a big company. Because he thinks he knows, and it's also why all the one-shodding, vibe coding kind of people have no problem saying it's fine, because they've never had to live in an environment where the constraint was to prevent the whole thing from imploding. I feel this is so critical, Steve, that you're like, so again, this is going to sound like a little tangential, but it feeds into this, which is I feel like core technologies kind of catered to like some human need, like the internet catered to like connectivity and social

networking kind of catered to vanity. And I feel like AI caters to our need to be productive. So I feel like we feel like we're being very productive and we do all of these things, but we may actually be creating like mounds of extra work to do. And well, Aaron, you're deploying AI right now. Like boxes all in.

So tell us, tell us, share a story like of the wheels coming off or not coming off. Yeah. Well, well, I think, I think,

uh, we're probably in the more pragmatic part of the continuum. So, so, which is why we don't claim that it's a 10x productivity game to our engineering team. It's like, no, because we have a lot of guard rails in place that create these constraints

automatically in our system. There's, we, we still rely heavily on, on code reviews. We still rely heavily on security reviews. So are you, are you guys coding with like a rock and a chisel and stuff?

I know, it just sounds like, we have chalkboards and like, but, no, but like I, I, uh,

we had this new feature that we, that we launch and I was like, go, go, go, go, go, and AI built probably 80 to 90% of the feature. And the, the thing that slowed down the release of it was we have to do a full security

review because we can't let there be any, you know, accidental code injection into the thing that we created. So there's a lot of stuff where you kind of go super fast, but then you get still rate limited or constrained by some other part of the process. And I think that's sort of, you know, relatively natural until we figure out then that other part of the process, security reviewing one, or the actual code review being one, or just even your pipeline for, you know,

getting things into production being another one. So we're doing quite a little, you know, quite a bit of retooling of the whole product development lifecycle. but I don't think that it's a 5 to 10x gain. I do think it's a 2 to 3x gain, maybe across the board. You are still rate limited by how quickly can you review this stuff and check on the work.

I do think that Martin's sort of pointing at, though, a thing that is the big open topic across enterprises, and to some extent, engineers will face it first and we'll find the right equilibrium. The harder part still remains in the rest of knowledge work. This is why if you're in accounting, you know, we don't quite yet know when you can take your hands off the wheel

you know, doing a full accounting audit you know, because of AI. What you can do is have the AI go and like comb through unlimited amounts of data to find anomalies that maybe are, that would alert your accounting team to, oh, we actually have to go dig into this. That's awesome because that's only net new level of visibility

versus the part of the accounting process where you're doing a fine-tooth comb on making sure every single number is accurate. that's probably still humans right now. So I think the key is where do you find the productivity gains? And I do think that if you are a CEO or a board of directors or a management team, you're kind of trying to figure out. And you're also getting confused because Silicon Valley is telling you all the things.

And so you have to sort of figure out where is the productivity most potent, where I actually can get the gain, I can get the success with less of the downside. And I think as an industry, we're all sort of figuring it out. by the way, this is actually why I remain unbelievably optimistic on jobs because I don't think you, like, I just think we've gotten it wrong on thinking, you know, all the places where you're going to remove humans from this, because you still need a human in that, you know, somewhere in the loop. Maybe the abstraction is a little bit higher and you don't need the human in a loop at every single stage that you needed a year ago. But, but you do need a human sort of kicking off the process, reviewing the process, and incorporating whatever the work was. And so that creates just still a tremendous amount of opportunity and jobs across these organizations. Oh, let me, I have to jump in because I have, I have a whole bunch of, like, visual aids I brought today to make it exciting. We got, you got a bunch of comments on, on the MTS Live thing about people agreeing with you. So I don't want to let that slide because, you know, we complain about

not agreeing with you. But, but like here, to your point, to your point, this was a book in the 80s called The End of Work. Yeah. And I, this, so, actually, sorry, it was in the 90s. It, it came out like six months before the internet hit. And the whole thesis was that technology revolution was a complete bust and we got no gains in productivity, but now there's going to be no more jobs because the economy is stagnant. And this was a guy. He called himself a futurist. And like, so the whole notion that it, that's like, one of the neat things about this

whole AI moment is like the number of things that when you hear them the first time you think they're stupid. And then you go back and think about it, you're like, oh, my God, it's way stupider. And this idea that, like, AI just gets rid of jobs. It's as ancient as, like, you talk about the accountant. Like, one of the things people thought was that computers would get rid of accountants. Yes. And that was like IBM's pitch in, like, 1965. But what it actually did was like, oh, my God, we could do so much more with accounting now that they're not like literally just adding numbers all day. And I think when you look at like, just the notion of, of like creating information, synthesizing and all it.

Like, AI is an accelerant for that for a person who knows what they're doing, and companies are suddenly going to want more of those people creating more of that information, not to mention the fact that if AI is creating valuable information and there's more of it, then more people will need to consume it to do something. And the essence of a company is acting on information. And so this idea that information is just going to get produced easily and be in surplus and not used makes no sense at all. Because as you know, like in the unstructured information world,

the problem is that you can make it, but the consumption of it effectively is the gaining factor. And that's the gaining factor now. We, I think, we had a conversation with one of our board members who's chair of our audit committee, and so he's CPA, and he was telling us, you know,

kind of early in his career, I can't even retell it because it felt so manual. But so I don't, I don't even know how the world worked before all of the modern technology, but he was explaining the CPA's process. And I was like, yeah, it seemed like the most manual thing of all time. But, but, and Steve, I think this is right out of your book is like, it was actually quite simple in sort of the amount of things you could do.

Yeah. Because of how undigitized and relatively manual the whole thing was. And computers actually only made it more complicated, more comprehensive, and thus created even more jobs. because of that complexity that we introduced. And you can just sort of see how easy this is to show up in so many areas of work. It's like we can just now,

we can afford to make things more complex. And so if you make things more complex, then actually you eventually still run into now new constraints of who can understand that complexity. And so like, you know, it's like to me, it's like the funniest concept that the more code we write, the less we would need engineers would be the opposite

because now your systems are even more complex than before. which means that you're going to be running into even more challenges of when you need to do a system upgrade or when there's downtime and you have to figure out, like, well, how do I fix that problem or when there's a security incident? And so, yeah, I mean, this is, this is like, we're just getting started with the jobs on this front.

Right. Listen, we're a few years actually into this, and you can actually look a bit at the data too, right? Like, what are the companies that are hiring the fastest? Like the AI data companies? And they're hiring like crazy. Yeah. But not only that, like I remember there's this early prognostication,

which is AI writing code will get rid of infrastructure. Like it's going to commoditize infrastructure, which is this kind of very strange prediction given the fact that there's more software than ever before been written. And sitting on the board of a bunch of infrastructure companies, some that have been flat for a while, they're all doing fantastic because there's so much software and there's so much more software out there now.

And so, listen, if you look at the data on the ground from the companies, it's more software. The A.N. Native companies are hiring the most, it's very clear to me that we're in an expansion phase. And maybe just my only final point on this one at least is, is I think people, we have a little bit of a myopic view in Silicon Valley on thinking that, you know, engineering jobs are, you go to work at Google or name your, you know, tech company and startup.

And that's an engineering job. And look, we're so wired into that because that's obviously the ecosystem that we're all part of. And then you sort of forget, well, like, John Deere is trying to, make automated tractors and Caterpillar is trying to have AI systems and Eli Lilly is trying to design even more pharmaceutical, you know, kind of, you know, therapeutics. And just you can go through 5,000 other companies. They're going to now have the next set of engineers that are going to use Claude Code and Codex and cursor to be able to automate even more of their businesses and be able to

design and develop even more software for their workflows and their systems. And so it just might, you know, be that you don't go and work on a social network and improve the social network algorithm, you go work at John Deere and you improve the, you know, intelligent farming algorithm. And we just have to, you know, I mean, this is sort of like completely, you know, like Mark Andreessen predicted this, you know, 15 years ago of like software's going to eat the world. And what that means, though, is that everybody's going to have lots of software. And this gives everybody the ability to finally have lots of software, but you still need

then an expert or semi-expert to be actually going and prompting the agent on what to do, reviewing its work and managing the system that it builds. So all of the predictions on don't go into coding and don't go into software engineering, I think will be proven quite wrong. I think, I mean, look, that was super good, Aaron. And I think that the base case of all of this is just that there's too many people out there right now that don't like technology and have a static view of the world. So when they look to whatever it is that they think AI is going to do, and people hear

automation, they just assume it's going to take things away. Like here's a... We have a lot of people who like technology, though, that are also creating that... Right, right. So here's like, this is article fighting the paper chase. Lower, lower, lower. Well, I'm looking on a feed. Even lower.

I'm looking at a feed. What are they doing? Oh, you're looking at a feed? Okay. Oh, you're looking at my Mac camera. Yeah, yeah. Oh, that's why. Oh, you're fancy.

Right. So this is like Time magazine. Every kid in high school read it. 1981, but the whole view of what computers would do would be they would automate the paper in a company. And so the idea, like the whole first generation of computing was literally taking paper forms and turning them into something on a screen, then printing them out and then making it all easier.

And you fast forward, and it's all of these things that you just said, Aaron, like, you know, there was an era when lawyers didn't type. And so what happened was they just, they had people who were legal assistance, they called and paralegals, and they did all the typing. And then, like, some students at Harvard, they brought a computer into the classroom. And so this is, I'm lowering it so you guys can see. Yeah, yeah, yeah, yeah.

So they brought, that's an original laptop in there, in the early 1980s. And they brought, they brought this computer into the classroom, and then they got thrown out for using it. But they were literally, they used to do law school, and you'd write the essays in longhand in a book, and then the professor would have to read them. And now, of course, you just type them,

and you have access to the database of all the science, citations, but that's exactly, like, nobody deals with a lawyer who isn't in track changes with your contract. Right. And last I checked, there are way more lawyers today than there were 30 years ago. And they all are, every human lawyer you talk to is a computerized lawyer. Their citations come from the internet. Their information and the brief comes, and they type the brief.

And I think we, you know, kind of going back to the myopic approach, I think we maybe over. Like, I mean, as a big lover of technology, I wish this was true, but I think we just over assume that, like, everybody's job is just, they're just inside of Microsoft Word, and they're just typing a Word document. It's like, like, I mean, most of the time with lawyers, I'm like, you know, strategizing something or they're like working through a complex analysis of a situation. And it's not, like, I could go to an AI for advice, but that would probably only increase the chance that I go and then call a lawyer to say, hey, what do you think about this,

this situation that, you know, that we're dealing with? And so a lot of these jobs just have a lot of contexts that aren't sitting just, you know, literally on the, on the computer, doing all the work. They do have to kind of touch grass as a part of the job. And so then AI, yeah, AI will help automate the creation and production of the content and the review of the information, but then it still has to be incorporated into the real world of real value production. I feel like we're live and we're supposed to end it four. So what I'm going to do is just say we're live and it's four. And I guess that means we just stop and some lights fade or something.

None of us have done this before. We don't know what's supposed to happen. But someone is waving at me and smiling saying, yes, I think you're right. The smile means stop talking. Okay. All right. Well, it's great to see everybody. Bye, everyone.

Thanks for listening to this episode of the A16Z podcast. If you like this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X and A16Z and subscribe to our substack at A16Z.com. Thanks again for listening, and I'll see you in the next episode. This information is for educational purposes only and is not a recommendation to buy, hold, or sell any investment or financial product.

This podcast has been produced by a third party and may include pay promotional advertisements, other company references, and individuals unaffiliated with A16Z. Such advertisements, companies, and individuals are not endorsed by AH Capital Management LLC, A16Z, or any of its affiliates. Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy.