Good. I don't like that I can never approve my intro.
No, you can never approve your intro.
Yeah, but thank you for that.
It doesn't work that way.
Freak of nature. Okay.
You're a freak of nature. You, how's the family?
Family's good. Family's good. They love AI, so 6½-year-old thinks everything is something you can talk to.
What's the use case Max is using for AI?
Kind of everything. So, we actually do these. He generates Sora videos, which is like a fun, fun little project, and I'm sure many folks know this app called Suno. It does probably the best quality music generation. And so you just do like a Mad Lib of a Suno song of a song, and you know, it'll like it's endless entertainment for a six-and-a-half-year-old. So totally inappropriate stuff that they produce. So I'm happy to play one if you'd like.
It'll be great.
Okay.
So, all right, so, let's talk about. Let's start with a really interesting topic that you have a lot of opinions about, NotebookLM.
Yeah.
Actually, can we, can we show Aaron's tweet that he had recently?
Oh, no. Okay, okay.
So he said Oh, my gosh, they're making fun of us.
But you'd have to read the whole thing, so-
Yeah
D o you guys have time to do that?
Yes, you should-
Okay
Y ou should talk about the tweet.
This is the tweet you chose of all of my 10,000 tweets?
Well, I had a request-
Okay.
I had a request for you.
Yeah.
I think, can you just, with-
I literally, I literally had a tweet-
Hang on, hang on, hang on.
Like, I had a tweet three hours later that was at least five times better than this tweet, and this is what you show?
With the advent of AI, what's happened is you've become very serious-
Yes
A nd your tweets have gotten very long.
I know.
Can I just-
Very academic.
We have a collective request-
Yeah
W hich is, can you at least have three funny tweets for every long tweet about why AI is going to change the workflows?
You know, the algorithm right now has, like, already solved the humor side of Twitter. Like, you don't need me to-
I don't need you?
D o that.
Okay.
But this is... You're not helping the cause .
I'm not helping the cause?
Yeah. Yeah.
All right, let's talk about, let's talk about ROI.
Yeah.
What's happening with-
Does everybody know what NotebookLM is?
I think we've talked about it.
Have you guys already talked about it today? Okay.
I mean, they all have Twitter accounts as well.
Okay, great. Okay.
All right, actually, do you have any points of view on NotebookLM and what's happening over there?
No-
It's significant
It's just fascinating. I mean, it's exactly what you'd expect. You know, you give agents access to your computer and each other, and they are building their little society online right now. And, you know, we're gonna have an agent-to-agent economy probably. They're gonna start businesses. It's gonna be insane, so.
So, let's go back to—
Yeah
E nterprise software and how enterprises are actually thinking about this because on one end, there's this crazy stuff that's happening where there's such an exponential curve of innovation that's occurring, and on the other end, when you start talking to enterprise CIOs, ROIs are still pretty slow. The adoption is still not quite to the degree that we need it to be, and I think you've got a bunch of CIOs and CSOs over here.
Yeah.
What does it look like when enterprise cracks the code on adoption?
Yeah. I think AI is a tale of two cities at the moment, right? You know, maybe even three, I guess. If you go talk to an engineer, AI has become basically an unstoppable force in the future of engineering. You can't—you know, probably this year, it would be impossible for the average engineer to build software without AI. Like, it's just like they will not be capable of doing so. And you now can regularly see tweets from people at Claude or OpenAI, et cetera, where 100% of the software is written by AI.
Yeah.
And-
We had our first product, which is gonna be 100% run by AI in, like, 2.5 weeks.
Wow, okay! So this is just happening everywhere, and that will become the new trend, and I think we all kind of, you know, understand what that looks like, which is you are telling an agent what to go build. It goes builds it, you review it, you ship that. You know, there's a lot more intricacy, but that's kind of the future of engineering. The challenge, I think, is almost every other form of work-
Mm
A ll of a sudden is harder than that. And there's, like, a wide continuum. And the reason why it's a lot harder is in coding, you've got these incredible properties. One, the code is verifiable because it either runs or doesn't run. You don't really have that in, let's say, legal or contracts. Like, you can't instantly verify if the contract, you know, sort of is, you know, verified. You have to go, like, read through the thing. You often don't have the same level of security, permissions, access control issues in software development. You know, generally, most engineers get access to all of the code relevant to what they're working on. That's not exactly how knowledge work happens.
Knowledge work is sort of this infinite matrix of what do I have access to, and what does G2 have access to, and what are we working on together, and what are we not allowed to see from each other? Code doesn't really have that. And then code is effectively an entirely text-based medium, where, you know, since the beginning of coding, we just work inside of a little text editor, and we type out code, and that's exactly obviously what agents are purpose-built to be able to go both generate and read from. Now, take every other form of knowledge work.
Mm-hmm.
Also, I should add one more idiosyncrasy. In engineering, you do things like you document your work for other people to be able to see it. You know, these are not the practices of most areas of knowledge work. So take every other area of knowledge work, marketing, legal, sales, finance, et cetera, and think about how much more hostile every other one of those forms of work are compared to coding. So we have this incredible take-up of coding with AI agents, and then everything else basically is a much messier way of producing value in the economy. It often requires a lot of person-to-person context that gets generated. We're on a video call, we're talking to each other, we're in a meeting, and we write down something, but it doesn't get digitized.
That's how, like, the real world of work actually happens, which is totally different from how we code. So I think what's happening in this sort of tale of two cities is you have AI coding, which is obviously exploding, and then you have everything else, which is either you have the option of attempting to kind of deploy broad agents across your organization, which often doesn't work as well, because the way that we work is not sort of purpose-built for agents. Or you have the success scenarios where you take specific business processes, where you're willing to effectively re-engineer that business process to support agents being effective.
And I think to me, the biggest mental jump that I've had to go through is, we have to probably imagine a world where instead of thinking agents will adapt to how we work, we will have to adapt to how agents work. And it's not like a subtle point.
What's that gonna look like in sales?
It is going look like building systems and building workflows where we are effectively producing information and creating context in support of making that agent effective at what it does. And so, you know, we're, we're a bit biased, because all we do is think about unstructured enterprise data, so we think about context all day long. But this issue of context for agents is gonna become the defining issue in the next decade if you want to deploy agents in an enterprise, which is agents-
Which is to feed context in the most efficient way to agents as fast as possible-
The most-
A s economically as possible.
The most efficient way, the most accurate way, the most comprehensive way. The problem is, there's, like, 10 variables you have to get right, right? So, you know, if you try and just do the least amount of context possible, then the agent doesn't have enough to work with. If you give it the most amount of context, then we have something called context rot, where the agent starts to sort of do worse and worse the more information it gets. You have to make sure that the way your systems are constructed are in such a way where the agent isn't accessing information that it shouldn't have access to.
Mm.
So it obviously has to both take, you know, on your permission set, but you have to make sure that the user's permission set is actually up-to-date and authoritative. So there's all of these sort of ways that we're gonna have to upgrade the way that we work in our systems to be able to make agents, you know, be as effective as we know is possible, and I think that's gonna mean changing a lot of our workflows in an organization. And just like in coding, where we've adopted this paradigm, which is, you know, if you go talk to the engineer that's building 100% AI-driven code, the way that they build software is completely different than how you would've built it a year ago.
It is. Totally. Yeah.
You are working on behalf of the agent.
Yeah.
You are-
All you're doing is creating specs and, in the Markdown files.
You're creating specs, you're making sure the agent is following those specs. You're sort of self-correcting it when it's going-
That's right
A nd veering off. Tell your person in marketing or sales or legal that that's how they're going to work. They're not prepared for this. This is the journey that we will collectively have to go on to be able to get the full potential and power out of agents. So I think, I think, you know, there's good and bad news in all of this. The maybe the bad news is it takes time. The good news is that nobody's really that far behind, and it's a chance of real alpha for the teams and the companies and the people that are willing to actually go and reconstruct or change their workflows to make agents effective.
So you could have an incredible leg up if you're willing to go do the work to go and transform that process. But it means that we're gonna have headlines for years, which is sort of this cognitive dissonance of: Why is it we're seeing this incredible AI transformation, let's say, in coding, but not as much in other areas? It's because we haven't done the work to change those other areas to make them ready for agents. And that's the journey I think we're all collectively on.
What's the most impressive example you've seen outside of coding in a customer of yours, where they've done something really thoughtful on automating some of the other functions of the business?
Yeah. Yeah, so, again, from our lens, we're very biased because what we think about is all of your unstructured data. So, you know,
Yeah
If you think about, like, where is your company's context, a large portion of it is sitting inside of your enterprise content. It's your research materials, it's your memos, it's your marketing assets, it's your contracts, it's your financial documents. That's what we've been storing for now 20 years. The problem is, you've never really been able to tap into that value at scale. Agents let you finally go and do that.
So we have a lot of customers where they say, "Hey, I have a, I'm processing 10 million medical records per year, and inside of every single one of those medical records is critical information that will either, you know, let me, let's say, provide better care, or do a better insurance claims process, or better understand my patients." And previously, you basically had to throw human, you know, time and energy at solving that problem of reviewing all of that data. So you've constructed your workflows either to have humans go review all that data, which is slow and inefficient, or you've constructed your workflows basically to never see that data and never get any value out of that information.
So we have a number of customers that are saying, "Okay, I'm gonna go deploy agents," looking through all of that information and then surfacing that up in the appropriate way, usually through some other kind of interface or software, to be able to better make decisions or route information or automate that workflow. But it takes, again, the real change management to go and re-engineer that process. On the other end of it, you have new revenue opportunities, you have cost-saving opportunities, you have ways of sort of, you know, getting through bottlenecks in your workflow, but you do have to go do the work to actually change the workflow.
But I think, what you're starting with is saying, "Okay, where am I sitting on a large amount of data, that today has gone underutilized and that we're not exploiting? And how... What can I do with agents that let me actually finally take advantage of that information?" That's a pretty good starting point to start to unlock, you know, major areas of value in the enterprise.
What does good look like in one year and three years?
Just for, like, anything?
No, for-
Oh, okay
... for enterprise software-
Okay.
with this, this topic that we're talking about.
Politically, I can... What are you talking about? Okay.
No, I'm talking about for workflow automation-
Yeah, yeah
W ithin enterprises.
Um-
For adoption acceleration.
I think that there's probably this interesting bimodal thing, which is, you probably you know should establish the general productivity, you know, sort of set of systems, so people just get the one to two hours of productivity gains a day, which is, okay, all of a sudden, I have better general intelligence. I can get, you know, content generated for me, et cetera. So like that, you've got to check that box. And then on the other bimodal side is, like, the very deep workflow reinvention, which is, you've picked off three, five, 10 workflows, which is, if I could go and automate that contract process or that financial process, you know, I could add a point of revenue growth because I'll have way better insights into my customer base, and I'll know exactly what to upsell them with.
Or, if I can re-engineer my you know, wealth management onboarding process, I will be able to streamline that from a two-week process to a one-hour process, so I'll be that much more competitive. And so being targeted largely probably top down in terms of where you deploy the energy, on that second sort of end of the barbell, I think having you know, three to five wins over the next year or two would be pretty key.
Because I wanna talk about two questions in the time we have. The first one is around this notion of, like, the role of enterprise SaaS with these agents coming out. What happens with systems of record, system of engagement, all of that? And just talk about the narrative of SaaS is dead. Like-
Yeah
... walk through that for how you think about it. And the second one is, I wanna talk about pricing.
Yeah.
Commercial models, how are they evolving?
Yeah. Well, I think-
They might be related.
Yeah. You might imagine I'm gonna say SaaS is not dead. So that's my message. I'm gonna leave you with that. I'm sympathetic to the concept. You know, a couple things are gonna happen. Software will be cheaper to build-
Mm-hmm
... which means you're gonna get more SaaS, and if you get more of something and you have the same budget, then prices go down. So I'm sympathetic to the idea that you're gonna see more competition in software. And I'm sympathetic that agents will do some capabilities that we've currently had in software. Those two things have to happen. They have to be inevitable. The part that I'm not sympathetic to, that I think you'll see on Twitter, which is... Actually, there's two parts that I'm not sympathetic to.
One is that we're just going to vibe code our own CRM system or ERP system, and I largely don't think that's gonna happen because, you know, we all collectively have a fixed amount of resources in our organization, and our customer only pays us more if we can deliver more things that they wanna buy and that better sort of support their needs. And us, you know, vibe coding an ERP system is sort of not on that list of things our customers care about. And there's a couple categories where maybe you have to go and customize your system or build your own software for that, because there's nothing in the market that solves your issue. But for the most part, that doesn't seem to be where you'd wanna deploy your scarce resources.
Geoffrey Moore, who wrote Crossing the Chasm, kind of perfectly, you know, kind of, just, you know, sort of, you know, created this concept of core versus context.
Mm-hmm.
Which is, you deploy resources on the things that are core, and you outsource the things that are context, and outsource, or rent, or you have service providers, and I don't think that fundamentally changes due to AI.
The second, and probably the bigger reason, is that agents, if you imagine a world of 100 times or 1,000 times more agents in an enterprise than people, and those agents are all gonna be non-deterministic, they're all gonna be probabilistic, they'll all make the wrong decision 5% of the time, or 2% of the time, or whatever, then the value of the system that traffic cops what those agents can work on, when they add data to a system, when they take data from a system, when they produce something, the systems that are traffic copping all of that workflow or the data that those agents have access to, I would argue, go up in value. Because you now have 100x more either opportunities for value creation or risk, if it's not managed well.
And so if you're imagining, you know, an ERP system with 1,000 times more agents than people were ever logging into it, then that ERP system has to do a really good job of coordinating what agents can work on, what data they have access to, et cetera, and that's the deterministic side. So the non-deterministic are the agents, which are kind of what we used to do with software, and the deterministic side is the workflow that says it's gonna happen exactly the same way-
Exactly the same.
E very single time. We will never accidentally have G2, you know, look at data that he shouldn't have access to. That's the power of software, is we've codified our workflows and business processes into a system that will run the same way every single time. And that, I don't think, you know, reduces in value, even in a world of cheaper or more, you know, volume of software.
How do commercial models change, pricing models change-
Yeah
I n the future?
I think we should expect some pricing pressure on software, because, again, if we can build... You know, let's just say for us, we're gonna add way more features because of AI. We're gonna build just way more software. We will not, you know, charge the commensurate amount of what all of the rest of the market would have charged for every single incremental piece of software we produce. We're gonna add that as more and more value for our customers. So then, on a relative basis, software gets cheaper if everybody in the economy does that. That's sort of, you know, obviously, kind of microeconomics on that front. Conversely, though, I think we're all gonna be deploying agents.
We have a set of agents, you have a set of agents, every company will have agents that basically augment the sort of labor side of what used to be required to use our software. And for that, we'll charge probably a consumption model, and you'll pay for it like you've paid for service providers in the past, which is, "Okay, I want a certain number of contracts reviewed," and I used to maybe, you know, think about that as a you know outsource service provider. Now, that would be an agent doing that work. And you know, we'll all collectively charge some kind of consumption model. And then the big question is: What are the flow through of the economics, right? How much goes to the software stack? How much goes to the AI model providers?
How much goes to the infrastructure? Ultimately, Jensen, you know, makes it all anyway... so it's all basically gonna go to Jensen. So, but, like, I, I don't know if he's-
Lisa is here, man.
Okay, yeah, Lisa will get it all, like.
It was AMD.
Lisa. I mean, everybody's gonna get it. So don't worry, everybody in this room will get the money. So, I don't know who's... you can't see everybody, but I'm sure every chip provider is here. Just, just please let us have a little bit at the software layer, if you can, if you can leave some room for us. But that will be the, the flow through, I think, of economics.
And so will the pricing... Is there a theoretical maximum? Will the pricing be more subscription-based, user-based on a continued basis? Will it be more-
Yeah, I mean, I think-
How is that going to evolve?
I think this is sort of, I think customers actually get to decide this, because what's gonna happen is you're gonna say, on the early stage, when you're experimenting with AI, you're gonna say, "I don't wanna commit a lot upfront, and so I actually want consumption-based," because and I wanna be able to have some flexibility of testing this-
Once I get to scale, consumption doesn't make sense.
Yeah, exactly, and then you're at a certain scale, and you're like, "Actually, I'd rather lock in a rate, because I don't want my EPS to go up or down 3% every single quarter, and so I actually wanna lock something in." And I think that will basically be the trend, as we've seen with SaaS, and compute, and et cetera.
Is there anything that you wanna say to the CIO, CISO audience that you, you haven't covered, that you'd like to say to them?
Definitely buy more software, 'cause the SaaS market needs that right now. I think it's an incredibly exciting moment for all of us in technology. The ability... And maybe the final thing I'd leave everyone with is, really think about AI as a way of augmenting what you're doing. You know, there's gonna be plenty of ways to save money and cut costs, but I think the real potential is when you start to say, "Okay, I'm living in this world of abundance, and I can have agents go and do anything that we've never gotten around to." And so how do you go through that list, just way lower in that list than you would've been able to get to otherwise?
And ironically, we think about that list as, okay, those are then lower priority things. Totally the opposite. What will start to happen is people will say, "Because agents have lowered the cost of doing something by 10X, I can actually now do the more ambitious things." Because the more ambitious things previously were, like, too hard to, like, mentally get around, because you're like, "It's gonna take three years to figure out if that piece of software actually is gonna produce value for our customers." But you just go and build that now in two weeks, and you find out. And so, I would argue, use agents to be more ambitious, and to actually do way more as an organization, not just the smaller things, but actually the bigger things.
I think that will be the future of the economy, and so you might as well get ahead of that anyway. So thanks for having me.
Aaron Levie, CEO of Box.
Thanks, buddy.