010: Why AI Won't Kill Salesforce | Aaron Levie (Box)

Hey, good to be here.

Okay. No pressure, I guess.

I think that one part of me says I think the trajectory is pretty good around how various SaaS companies are figuring this out. And on the other hand, just having seen history in these sort of disruptive platform shifts, I recognize that there's a meaningful chance that in 10 years from now we'll look back and we'll laugh at kind of how naive we were on the architecture we chose or the billing or pricing model we chose. The question is basically that we all have to kind of figure out is it 2008 in the era of the mobile disruption or is it 2005? With the distinction being in 2005 we were all building these sort of very weird janky mobile Frankenstein interfaces that were like Java based or like weird WAP protocol based things, or is it 2008, there's an app store, we know that you should build iOS or Android native apps and we've all kind of figured this out, right? And I would say categorically two years ago it was 2005. The ChatGPT moment was the iPhone of AI but it actually didn't introduce the ultimate end-user architecture for most SaaS products because if you had just seen ChatGPT and you said oh the future of all software is you just query everything in an interface, that was sort of the wrong insight to take from that moment. The right insight really kind of came about a year later which is wait a second we use these models to basically develop agents and those agents basically do work in the background and whether you're talking to it in a chat interface or you're kicking it off in a normal workflow pattern that doesn't really matter as much. And it's really about this new unit of value that your software provides, which is agentic workflows or agentic work. And so I would argue that we're definitely probably more likely in the 2008 kind of moment of mobile where we do understand these patterns. We're all kind of coalescing around this idea of the agent. It's hard for me to perceive that there's something kind of more disruptive than agents on the horizon because this alone has uncapped productivity potential for what you can deliver now in your software. So I think the industry and existing SaaS companies are coalescing around the agent idea and then you have this big question of okay well how do you monetize it? How do you price it? How are these deployed? Which have unlimited amounts of interesting kind of contours around those questions.

Well, so first of all, I love the conversation that is out there of okay, is the seat model dead? Does the AI agent kind of abstract all of that? I think it's a super super interesting and enjoyable conversation to debate because I love debating anything related to enterprise software and business models. My take is probably going to be a bit more nuanced. And partly because you just get kind of more pragmatic the more kind of gray hair you have in this industry. And you live long enough to see Oracle as a $600 billion company when your naive days you would have said well you know everything's obviously moving to SaaS and this business is going to get totally disrupted and you're constantly reminded that Larry Ellison's the goat and Safra is amazing and these companies just find a way through. So, in the more pragmatic version of now the era that I'm in, I think realistically we're going to see a hybrid, which is I think that the seat model is still so well suited for the human part of the interaction of agents. And so I don't see a reason why that goes away. You might have pricing pressure on the cost of the seat because on a relative basis the kind of consumption part of the software is adding more value than the seat itself. And so could that introduce some downward pricing pressure? It's plausible. And then you have this sort of uncapped unlimited upside of the consumption model of the amount of work that the agent is doing or agents are doing which each company I think has to navigate to their own version of what that looks like. And I think you guys certainly see every version of this and I'm very fascinated by the different approaches because some basically say okay we're going to take a cost plus model and we're paying a certain amount for the AI and we'll add 20% on top of that. Others take a sort of like what was the human work equivalent and try and say okay we're going to do 20% of the cost of what the human would have cost and so it's a value-oriented model. And then there's sort of stuff in between which is okay, let's just see if we can squeeze kind of what's the upper limit of a cost plus approach where we get 80% or 90% gross margin on the cost because the value that we're providing is so massive, but maybe that actually still ends up being 95% cheaper than the human version of what that would have cost. And so you've ended up in this kind of interesting quasi middle ground. And then you have another axis of this which is is it the amount of hours that the work is happening? Is it the amount of chunks of business outcomes? Is it each task that the agent is responding to in a customer service request queue or is it the amount of lines of code written? So I think it's like super fun how many different areas you could go down. I think as an industry we're going to run to a little bit of an issue with customers at the end of this which is potentially some degree of fatigue which is I have 20 different AI vendors that all have five different ways of pricing. So I have a hundred different ways of trying to figure out what I'm going to be paying for all of this AI agent work. That's going to eventually be a really interesting problem that SaaS actually solves for the customer because you could just say, 'Okay, that thing's $20 per user. That thing's $25 per user. That thing's $40 per user. I kind of understand it.' AI agents unfortunately have the ability of having uncapped potential expense on the customer side. So as an industry you have to also have the customers playing another role in this that we all have to navigate is we're a set of software vendors we might have a set of theories that eventually has to get tested by the market and we will have to kind of coalesce around what that looks like.

Yeah. So, the amazing thing about our position is and I think that there's lots of SaaS companies that have a version of this, but we get very excited about ours is we have always been basically limited by the number of end-user humans on the other end of our customer for the amount of value we could create for them. And so we have software that helps companies manage their contracts and their invoices and their financial documents and their marketing assets. And so we have a very wide amount of use cases that we can help customers solve. That's the really fun amazing part. The challenge is we might go to a legal team and we might say, 'Hey, we've got a really great way to manage all of your contracts.' But they might say, 'Oh, well, we don't have enough people to go and label all the contracts to make sure they go in the right place, so we can't really fully utilize the full power of your platform.' And so then they use us in a maybe a more simpler, less powerful way. Well, this is the first time now we can actually bring labor with our software. And so we can go to that customer, we can say, 'Hey, your legal team is five or 10 people, but what we now pack into our platform is the equivalent of a hundred lawyers that are doing way more work for you to be able to go and label all of your contracts, extract all the data from them, and give you a really well organized way to manage your contracts.' Now you can go to the procurement team and say, 'Hey, all those invoices that you process, well, it used to be that you were manually processing all of that. Now we actually bring the labor that can automate those workflows with our software.' Or we can go to the marketing team and say hey all of those marketing assets that you have well we can now organize them and basically extract all the relevant data from them so you can search them and find them in a way that again you couldn't before. So for us, we can in many cases do the equivalent of doubling, tripling, quintupling the staff of our customers by bringing agents with the software that they're buying to solve way more use cases. So the kind of quick back of the envelope way that I think about this is which categories of software were historically underutilized because the customer didn't have as much talent on their side to fully take advantage of all the capabilities of that software and those will be the areas that have the highest degree of upside with AI agents. And so if you kind of have that mental model, you can almost look across all of software and figure out where's all the potential of these products, where have you had software that could be way more powerful in a world where the customer had unlimited capability to go use that software and that's where you're going to see so much upside. So this is why I think we're seeing breakthrough startups in legal and professional services and financial services and healthcare and insurance because these are spaces where the company never really was able to fully digitize their workflows before. They had lots of manual processes. So, that means you have lots of software that has basically gone underutilized relative to what it could go power.

Yeah. Let's maybe separate two things. There's the in a vacuum version of this and then there's the okay what actually happens in the economy as this plays out. You know in a vacuum if you kind of perfectly extrapolated one role or one category of roles and you saw what AI did you might say okay that could replace 10% of a category of work. But the important thing is to realize what actually happens in the economy and within a business when that occurs. So let's just take a real life example of what we're seeing day in and day out. Most of the time when we are going to a customer and we're showing them what AI agents can do on their software with their data, with their unstructured data in Box, their first instinct is to solve use cases that they were not deploying humans at before. Right? So, this is not the case that they had 10 people that were labeling all of their contracts. It was that actually they were just manually working through their contracts before and they were...

Right. And so if you think about it, in most areas of an organization, the cost to doing something new is usually correlated to hiring a new person to do that job. And so in the vast majority of workflows in a company, you're kind of sitting around being like sure if I had unlimited resources, I would go solve that workflow. But the starting cost of now solving that problem is hiring a person to go do that thing on day one. My fixed expense now is $50,000 or $100,000 or $200,000 just to start doing that unit of work. And so you have so much work that never gets done because the incremental cost to even starting is that fixed full-time human or outsource or whatnot but you still have a high startup cost and so actually most work in an enterprise is not happening because you have that high fixed cost. So, what if the first expense it ever took to label, extract data from a contract was $1 instead of $100,000? Like the very first contract that you send to your business process outsourcing firm is going to be a $100,000 deal you have to do with them to get that first contract labeled. So, that means that on day one, that's how much you're spending to get started. Well, with AI, you might spend a dollar or 10 cents as your first starting point. So, all of a sudden, a hundred times more potential areas of automation emerge in an organization that you were not deploying humans at before that now all of a sudden AI agents can go and kind of slip into. A great example is what we're using even with AI coding. A significant portion of the AI coding that we're doing within Box is usually just getting to the next task that that engineer would have had to do anyway, but we're just getting to it faster. So we don't see that as well, we just saved an engineer. We see that as we now accelerated our roadmap. And so now if we can accelerate our roadmap, what happens next? And this is where the economists get this all wrong. If we can accelerate our product roadmap, that actually encourages us to hire even more engineers because now we have higher productivity in this part of the organization to deliver even more value.

Yeah. So I think you're going to see plenty of that. There's no question. So let's assume that 20, 30, 40% of the economy initially thinks about AI like that, right? I think for every company that has that conversation, there's going to be another company that says, I'm not going to measure this on how many people I reduced. I'm going to measure this on how much output I increased. And all of the companies that have the paradigm or prism of I'm going to think about output increases, those companies have a totally different frame of reference or way of operating through this, which is when I go talk to the marketing team and we say, 'Let's go launch this campaign.' I'm going to ask, 'Why can't we launch it in 20 markets instead of two?' I'm going to ask, 'Why can't we personalize it to way more demographics instead of only one or two?' I'm going to ask why can't we do this in a week instead of three weeks. None of those questions related to I want fewer people working on this. Every single one of those was why can't we do more because AI is going to help accelerate the execution of this effort. This is a total N of one example to be very clear but we've had multiple things even in our marketing department where we will use AI to help accelerate a project that I can almost guarantee we will be adding more headcount to that area because AI made it now more plausible that we will do more of those types of marketing activities. So by actually lowering the barrier of doing a category of work, we've now decided to invest in it for the first time which now means more people are going to go and work on that particular type of, in this case it's content strategy, marketing production, and now we will have more people working in that area than we did pre-AI because AI has now made it so we can better execute that part of our strategy and I think you're going to see as many of those scenarios as you see the scenarios that are okay, I just want to specifically reduce expenses using AI. And that's why I think it balances out. And then the only other point to make is you can't last that long as a business if your only metric is reducing expense because you will have a competitor who decides that they don't want to be on the expense reduction side. They want to be on the output increase side and that company will just better serve your customers and so for a single cycle you might be able to reduce expense but over any meaningful period of time the company that delivers more value or has more output will be the one that ultimately wins that specific market.

Well, you know, right now, so we're 100% committed to going AI first as a company and essentially that means let's look at every opportunity, every workflow in the company and figure out do we have an opportunity to use AI to accelerate that. So yeah, that's the general mental model is let's use AI to do more faster and every project that we see let's ask ourselves could AI kind of help with this effort and that's the approach that we're taking. We've generally had this stance that right now we just want to do way more to experiment way more to use AI way more and not overly worry about either the budget of that AI or the direct kind of A/B test of the ROI because we just want maximum usage of this technology to get familiarity with the tools. We want employees to have the right vocabulary around this. We want them to test different models within our product and figure out what works better. So right now we're kind of in just go fast, try anything you want, with a bunch of constraints around compliance and governance and whatnot, but not being overly prescriptive in a tops down way about how to use this. I think where I'd like to be in a year from now, two years from now is the magical end state at least in this sort of transition period while we all collectively figure out what a fully AI first company looks like would be I'd like to be able to go to an engineer's roadmap and see that basically 0% of their time is spent on the classic stuff that we always got slowed down by which is like well I'm doing a bunch of testing work over here, I'm refactoring this thing, I have to upgrade this library from some old version to the new version and you add up all of that and they have like 15% of their time left for the actual feature that the customer is asking for and you're like wait how did we lose 50, 60, 70% of our engineering capacity to basically run the business things that are table stakes you have to do them to survive but your customer never experiences the end result of that value proposition. And so what I'd like to be at as imminently as possible is that's 0% of the human time. Maybe it's a couple percent because you got to oversee it all. And all of our time gets to be spent on what feature to build next, how it should look, how it should work, ideating through that. I like the idea of you can see with these latest kind of coding systems, the ability to have kind of parallel agents running in the background, which basically gives you the ability to almost kind of do multivariate testing of how you'd even build a feature. Imagine a world where the typical debate of how you'd architect some product that you're working on. You spend days, hours, weeks debating the architecture. What if you had an agent go try every version and then you empirically get to just see like which one works, which one scales better, which one works better within your environment. Like that becomes super interesting now new ways of thinking about coding. So I'd like to take that kind of approach and apply to every part of work. How do we streamline every legal contract review that we do? Well, we have Box AI that automates how you review the contract and get all the insights from it. How do we onboard every sales rep or employee to instantly become an expert at how you work with customers or work within your particular field or...

Well, that's an example where you take all of the kind of corporate knowledge in the company and you turn that into insight for your employees with AI. So these are a few of the ways that we think that will transform how we end up working in the future.

Well, I don't know that anybody's getting it like, probably the most wrong you could get it is by just avoiding it or ignoring it. I'd say that's an increasingly very rare small camp and so that almost doesn't matter because they'll figure that out soon enough. I think that the companies that are overly obsessed with cost cutting I think get it wrong because again you should be focused on output increase or capability expansion. But also I think that you're also creating an inherent tension with employees that I just think creates an environment where people then have to feel like, okay, anything I go and automate, I'm probably then risking my own job in the future if I go and automate that thing. You actually want people to be encouraged to automate their work, right? And not feel like at the end of that your goal is basically to try and replace me. That is not going to be a way to get adoption of an AI system within an organization which is why we are so maniacally focused on and we talk only about how do we drive output because that's actually the thing that we care about.

Yeah. But they do that with or without AI. So like, and I think we're in an environment where two things can happen. One, the press will conflate those things, but also the, you know, companies sometimes have an incentive to also conflate those things as well. And so I'm not saying that there was zero impact of AI to be clear but unfortunately we're in an environment where it will be very hard to untangle, you know, which outcome actually drove, you know, take the Meta cuts that happened two years ago let's say. I don't think AI was mentioned once in those cuts yet they were able to do a massive efficiency gain because they realized that they overhired during the pandemic they had lots of redundancy there was lots of people not being able to do, you know, daily execution work that was creating value at least in their estimation. And so then they had to, you know, do some major efficiency cuts and again that was pre-AI. So unfortunately we will be in an environment where when these cuts happen it will look like AI was a big part of that and I guess in like the most micro sense, you know, they might actually say okay well yeah now we are able to do that thing with AI but that cut was probably driven by some other efficiency program that was going on in the organization. So I wouldn't over rotate on these sort of one-off examples.