Bold Names: The CEO Who Says Cheaper AI Could Actually Mean More Jobs
29:44 Share

You want a straightforward path to your goals, but at Merrill, we know things may get in the way.

Or new opportunities can put you at a crossroads. With the bull at your back, you get a personalized plan and a clear path forward. Go to ml.com slash bullish to learn more. Merrill, a Bank of America company. What would you like the power to do? Investing involves risk. Merrill Lynch Pierce Fenner & Smith Incorporated. Registered broker dealer. Registered investment advisor. Member SIPC. A wholly owned subsidiary of Bank of America Corp.

Aaron Levy is the head of the kind of company that we all rely on every day, but few outside his industry really know. His customers include most of the Fortune 500, movie studios, automakers, consumer electronics giants, marketing firms, and also the Department of Defense.

They've all got endless reams of documents from contracts to top secret R&D plans. It's the kind of stuff that used to go into filing cabinets, but now it's all stored by Levy's company called Box. Box lets companies put that data into the cloud, organize it, and give access to those who need to see it and keep out those who don't. Now more than ever, the business of Box is also AI and all the ways it can make us more effective.

Levi is both a full-throated booster of AI and one of the most vocal critics of the hype surrounding it. But here's the twist: Box isn't making its own AI model. Levi says they'll work with anyone else's product. OpenAI, Anthropic, take your pick. That's something the industry calls being "model agnostic." And he says customers are buying his data storage for exactly that reason.

They don't charge for the amount of AI customers want to use. Something Levy says is unlocking the technology's true potential. I think Sam Altman has this line of AI is becoming too cheap to meter. And that's like a very profound concept. Like what happens if intelligence is just literally infinitely available and basically free? What will that mean to software in the future? It's got some very, that's a very kind of compelling prospect. In

In many ways, Box typifies the way that AI has taken once-sleepy sectors of tech, such as cloud storage, and strapped a rocket ship full of investor interest to them. This year, the stock has shot up nearly 40% and stepped with the fortunes of other hot AI stocks like Nvidia.

Much like NVIDIA's CEO, Levy is part of a growing cadre of tech leaders focused on what comes next for AI. How will it change how we live and work? In his world, AI is about speeding us up, not replacing us.

We are still seeing hallucination. We are seeing cases where mistakes will be made. So I think we're at a stage where humans do need to be in the loop for the most kind of mission critical, high severity type use cases. But I think that will be something where over time, you'll just see less and less of that. From The Wall Street Journal, I'm Tim Higgins. And I'm Christopher Mims. This is Bold Names, where you'll hear from the leaders of the bold named companies featured in the pages of The Wall Street Journal.

Today, we ask, where's the line between hype and reality for companies racing to incorporate AI into their ways of doing business? And we'll hear why Aaron Levy thinks every country is going to have its own sovereign AI, especially once the technology reaches human-level intelligence. Aaron, always a pleasure. I feel like we've been spending a lot of time together lately. It's been a lot of AI talk. Let's just start with what y'all are doing these days, because I feel like...

You guys are, I always call you an arms dealer. I'm teasing. I'm fine with that. But pick and shovel dealer. Yeah. No, actually let's go with arms dealer. That sounds a little more badass. Okay. Arms dealer in the AI wars. So for those who last updated their knowledge of what box is a while ago, what all do you do now? Yeah.

So at Box, we help enterprises manage their most important data. So this is their unstructured data that is their enterprise content. So their financial documents, their contracts, their marketing assets, their R&D files, everything that goes into, in many cases, how these organizations run, how movie studios create their films, etc.

how marketing agencies launch campaigns, how major automakers or consumer electronic companies deliver their products. All of that data, all that content has to be stored somewhere, it has to be secured, workflows around that content have to be automated.

You have to collaborate on that data. So we have a platform that handles all of that for over 100,000 customers, nearly 70% of the Fortune 500, and many of the most important brands on the planet. And so that's what we've been building out. And AI for us is this breakthrough moment because for the first time ever, you can actually tap into that information and understand what's inside of it in a way that wasn't possible before. Yeah. So what specifically are you helping customers do with AI? Yeah.

Yeah. So, you know, in our particular category, so if you think about the world, this is oversimplification. But if you think about the world of enterprise data as really having kind of two big parts. So there's structured data. This is the data that goes into a database. And with that data, you've always been able to query it. You've always been able to synthesize it. You've always been able to effectively ask it questions. You can put that data into dashboards.

The interesting thing, though, is that 90% of our data inside of an enterprise is unstructured data. And much of that is in the form of content.

So, with unstructured data, this is what's in your email or your Slack channel or all of your files, you've really never been able to ask that data questions. You've never really been able to sort of summarize that information or synthesize it in any interesting way. So, AI, and generative AI specifically, is really the first time where at scale you're able to use computers to understand all of that unstructured data.

And so now for the first time ever, you can just ask a question of all this information in your enterprise and get an answer back. I think in other words, it gives you great insight into how real companies, not just companies in Silicon Valley, are trying to use AI in their day to day. Is it working? Is that promise happening?

First of all, there's no question that there's a lot of hype right now in the space. Anybody in Silicon Valley, we can't help ourselves from being just wildly excited about whatever the new technology breakthrough is. This is just how we operate. This is why we come to Silicon Valley. This is why we join. The tech industry is, we're generally extremely excited and optimistic and in some cases, over-extrapolating the change as soon as we see a new form of innovation.

But there is actually a tremendous amount of very realized gains and real promise that is now playing out. You just have to figure out where AI is actually going to be extremely useful and where is it maybe too early

And we're still years away from the breakthroughs that will enable certain use cases. So in our space, we tend to be very pragmatic as a company. By virtue of selling to enterprises, you can only really sell an enterprise something that's very real and actually delivers its promise because that enterprise usually has a procurement cycle. They have an evaluation cycle. They tend to only implement or adopt things that actually are technically playing out in practice.

And so by virtue of that in our business model, the only things we get paid for are things that are real. So a really straightforward one that companies have just been really trying to solve for decades and decades is how do I take a lot of content? Let's say it's contracts or invoices or marketing assets or financial documents. How do I take all this content?

and be able to label that data and pull out the structured data from those documents. So if you take a contract, contracts have dozens of important data fields inside them, but they're just unstructured text by default inside of a contract.

But the data that you really want from a contract is what's the amount of the contract? What's the renewal date of the contract? What are the clauses of the contract? Who are the parties that are named in the contract? What's the liability in the contract? All of that, if that can be inside of a database, that contract becomes 10 times more valuable to you because now you can ask –

a whole bunch of contracts, a set of questions. You can have a set of business rules in your organization that you get alerted on when different contracts come up for expiration. But you can't do that if all I give you is just a Word document that has a bunch of text in it. You have to actually extract the information from that document and put it into a database that is a structured database.

And so AI is this breakthrough for us where we can finally go through that document and pull out the structured data. So that's a use case right now that customers are deploying. We're in the early stages because it just got released. But now, for the first time ever, you can begin to extract data from your contracts, and then you can begin to automate your workflows that really were just incredibly hard to automate before. You know, before you –

Because you're talking about putting a non-deterministic system in AI into a very deterministic workflow, which is in this case a kind of data entry. Do you keep humans in the loop? I know in the latest models they're getting better at minimizing hallucinations. I also know it seems like we're never going to eliminate them completely. How do you deal with that? Are there humans reviewing this work?

Yeah, so first of all, it's up to the customer on how much human in the loop that they want. We are building more and more product capabilities that allow humans to be in the loop. So we'll have functionality where we'll alert you when different fields don't get filled out or when there might be lower probability of the field being correct.

Right now, you just get to decide how much human in the loop that you want to have. But for a meaningful portion of use cases, the AI is performing dramatically better than what humans are able to deliver. Well, I'm curious when it comes to having the human in the loop. I mean, is there a trend you're seeing with customers out there, real businesses? Are they saying, yeah, we want to have Bob involved? Or is it like just let the AI do it and we'll –

figure it out at the end? Yeah. So, so, uh, in general, you know, um, for it's really the severity of the use case, um, is, is probably the simplest way to think about it. If you, uh, want to have an invoice where there's an amount in the invoice, there's a shipping information, the invoice, uh, there's, um, you know, dates in the invoice, um, you know, any of the latest breakthrough models, uh,

can basically solve that problem without any human in the loop. Um, uh, that is just like now a solved problem, um, for the vast majority of, of any kind of structured document, um, that has some, some degree of kind of consistency with it. Uh, if you had a 200 page contract that had a lot of, uh, esoteric clauses, uh,

And very confusing math of, you know, at this point, this clause is triggered and please see, you know, item, you know, 17A for when that gets triggered. And there's a lot of kind of sort of judgment that has to be used to assess, you know, when that might actually be triggered. That's something where you'd want Bob.

or Sally involved in still reviewing what the output is. Interestingly, we've actually found a lot of examples where even the humans that we give that contract to, they can't figure out actually when something would actually be triggered and AI in many cases is doing a better job because it can simply retain more in its memory at once than the human can.

And so there's actually plenty of scenarios where the AI is doing a better job than what the average human, or not even average, but 90th percentile human is able to do because of its sort of the full amount of reasoning that it has inside of its system. But, you know,

We are still seeing hallucination. We are seeing cases where mistakes will be made. So I think we're at a stage where humans do need to be in a loop for the most kind of mission-critical, high-severity type use cases. But I think that will be something where over time, you'll just see less and less of that. You get to take advantage of all these advances in the so-called frontier models from OpenAI, from Mistral, from Anthropic, right? Because you are model agnostic, which means you can just plug whatever...

you know, chat GPT clone you want into this process. From the perspective of the folks building those models, you know, the price per word that gets processed is dropping like a rock. All these different models seem to be reaching parity with each other in terms of their abilities. You have venture capitalist Mark Andreessen saying things like, it turns out anybody can build a large language model like chat GPT. What's going on here? I've never seen a technology go quicker from,

This is cutting edge and incredible to this is electricity. Yeah. Well, I mean, this is, this is partly a driver of my excitement. If this was a single vendor that had all of the proprietary, you know, capabilities and we were at their, their mercy on how they priced it and they wanted 98% gross margin, you know, this industry would not exist right now. It would be a very scarce technology company.

The business models in AI wouldn't work because you'd be paying so much to the AI vendor. You wouldn't, you know, there would be a leftover for the individual software providers to make any money. But that's just not what played out. And what played out was a heavily competitive industry of at least five or six major players that are incredibly well capitalized.

all competing for price, quality, and performance 24/7. That is an incredible outcome for anybody developing software because the moment you're above that layer of the stack, the AI model layer, all of the competition accrues to your advantage because that means your pricing is going to get lower, the quality of the AI will get better, the performance of the AI will get better, which means that you can solve harder and harder problems

And just know that it's only going to get easier to solve those problems. It's only going to get more – your software will only get more intelligent. So you have to, to some degree, kind of work backwards from where things are going a couple years out and anticipate a rate of change that we're just not used to in technology. So we've made some strategic kind of business model decisions anticipating that. So one example is we decided to make unlimited –

AI queries, a default offering in our kind of core enterprise plan, our Enterprise Plus plan, as opposed to sort of charging customers for that and having overage fees because what we realized was our underlying price of these tokens are dropping so fast

It doesn't even make sense to meter the customer's usage at some point. You know, I think Sam Altman has this line of AI is becoming too cheap to meter. And that's like a very profound concept. Like what happens if intelligence is just literally infinitely available and basically free? What will that mean to software in the future? It's got some very that's a very kind of compelling prospect.

Aaron Levy just described for us how AI is becoming a commodity, where the price for access is dropping to nearly zero. But could that ubiquity come at a geopolitical cost?

Stay with us. You want a straightforward path to your goals, but at Merrill, we know things may get in the way.

Or if new opportunities can put you at a crossroads, with the bull at your back, you get a personalized plan and a clear path forward. Go to ml.com slash bullish to learn more. Merrill, a Bank of America company. What would you like the power to do? Investing involves risk. Merrill Lynch Pierce Fenner & Smith Incorporated. Registered broker-dealer. Registered investment advisor. Member SIPC. A wholly owned subsidiary of Bank of America Corp.

Let's talk about something, kind of the bigger impacts here, the economic ones. NVIDIA's CEO has talked about how every country needs to have this idea of sovereign AI, which basically means they can produce and control their own AI, everything from data to the compute required. He was just in Denmark celebrating a big AI supercomputer coming online. I'm guessing you have feelings about this idea of sovereign AI.

how you'd unpack that. Is that what's really necessary? Why? Why not? I guess help me understand it. I lament that it's probably necessary. It sort of defies my...

you know, hopes and dreams of technology being kind of free flowing and available everywhere. Uh, but I just think the reality of the world effectively has made this be the, the, the likely reality. Um, China is not going to rely on anybody else's AI. Uh, many countries in the EU are not going to rely on, uh, the U S is AI. Uh,

The U.S. is certainly not going to rely on AI from any kind of, quote unquote, competitive or adversarial countries. So I think almost by virtue of the geopolitics of technology, trade, and kind of these economic sanctions, I think you're going to see sovereign AI as just a necessary reality. Great if you're anywhere in the chip stack and the data center stack, because it means that

You know, we're going to be building out these data centers everywhere and for either training or for inference. And so I think that's a tailwind that is going to last a long time. I guess I'm of two minds. I think technology and bits and information should be free flowing on the internet. But to be fair to the other side, AI in its sort of ultimate form being, you know, some deep intelligence, it would be very hard to kind of compartmentalize this.

in some kind of like shared infrastructure, I think that would be a very kind of unlikely outcome. So I do think sovereign AI is a reality that's going to stay. You posted the other day on X, this social media platform that I have noticed you're active on often. I like the clarification of what X is.

Just for some of the audience, maybe my mother out there listening. And I quote here, if your AI startup doesn't have its own nuclear power generation deal, are you even an AI startup? But I guess I'm curious, are you going to get into the nuclear game here? I think you're making a joke just given the fact that so many of these AI companies are looking for power because...

training AI just sucks up a lot of juice. But are we in an arms race for, is it a nuclear arms race to power AI? Yeah, exactly. Fortunately, where we have explicitly decided to be in the stack, we don't have to be in the nuclear power generation game. Maybe just a coal plant? Yeah, exactly. We have a solar farm that we're building out, but

I think with internet speeds just getting faster and faster, you don't have to put these data centers right next to the ultimate end user that is going to utilize the AI. So you can actually start to have these clusters of basically AI data centers move closer to the power wherever that power might be.

And so this idea of kind of reopening nuclear reactors, that has happened quickly because actually one of the most important factors is just the cost of the energy for either training these models or running them. And so, you know, you could imagine we could create an island somewhere in the world that we all agree is like the power generation island and put data centers there and they could train the models. And then you could ship the trained model anywhere and run it anywhere. I think one of the other fears of AI that is in popular culture is

Yeah.

maybe can be more productive and why that doesn't necessarily mean we're all going to be losing our jobs here in the next couple of years. Yeah. So this one I'm extremely convinced by. And, you know, short of a major black swan event, I'm a firm believer that AI as a productivity driver will be neutral to net positive on job creation. Because I think actually so much of what holds back the economy and just the dynamism of the economy is the

is simply people's access to different kinds of resources or talent or skills to do things for them. And that the lower the cost is of doing those things, the more demand there will be for those things. And so whether that's somebody having an idea that says, hey, I want a website that can give me a storefront for this new business I want to start.

If we can shrink the cost of doing that by 3x or 5x, I think you see a dramatic increase in the people that then utilize whatever that set of coding skills was. If I can shrink the cost of creating a marketing campaign, if I can shrink the cost of translating product literature into new languages to serve more markets,

Anytime you bring down the cost of something that has more demand than current supply at any kind of given price point, you're just going to see growth in those particular sectors. So there's a – let's say it costs $10,000 to build a website, making up a number. But you go and ask a contractor to build a website, it's $10,000.

Let's say AI makes that contractor, you know, 50% more efficient. So now technically it costs, you know, $5,000 to make that website. The question would be, do we get more than two times the amount of websites built per

because we lower that cost by 50%. And maybe it won't be websites in this case, but if you look at across the economy, the amount of things that we'll do more of, because AI finally made it more kind of price competitive to do that thing, my bet is that we'll get more things built. We'll get more marketing campaigns launched. We'll get more content translated. We'll get more healthcare inquiries happening. We'll get more tutors getting

to be utilized because we'll have made all of those underlying resources more efficient and thus more affordable to the end consumer. I'm a big believer that this will be a growth driver of most segments. There'll be some, I can, you know, we can all think of a few that maybe won't be growth economies, but I think as a whole, AI will be a driver of growth of jobs.

If AI is going to be an invisible driver of productivity, one we someday take for granted, companies that don't jump on the bandwagon risk falling behind their competitors. When we come back, Levi gives us the precedents that explain how companies will adopt AI or else.

That's next.

Think scaling AI is hard? Think again. With Watson X, you can deploy AI across any environment. Above the clouds, helping pilots navigate flights. And on lots of clouds, helping employees automate tasks. On-prem, so designers can access proprietary data. And on the edge, so remote bank tellers can assist customers. Watson X works anywhere, so you can scale AI everywhere.

Learn more at IBM.com slash Watson X. IBM, let's create. One of the paradoxes for decades after companies were investing, you know, at the time, billions and billions into IT was that throughout the 80s and the 90s, there was this productivity paradox where consumption of those goods didn't lead to

or even substantial increases in per worker productivity. And economists were baffled. They were like, how do we, what is the root of this mystery? If this is going to have such a big impact, when are we going to start hearing, you know, let's say CEOs on earnings calls saying, hey, you know, this quarter we saved this many millions or billions of dollars thanks to AI, or we saw this increase in productivity and that's the reason we're paying a higher dividend. What,

When does that actually show up? My hunch is more likely than not...

AI gets so embedded into how we work that you almost won't be able to decouple it from the individual kind of productivity outcomes in that way. You'll have plenty of cases where somebody will say, hey, I saved 2% on our total expense because customer support became that much more efficient. That'll happen for sure. But the version that I think what will actually happen is in the higher volume scenario is

your sales reps will just become 10% more productive or 5% more productive because they'll go after the leads that are that much more likely to close with that much better of a pitch that is that much more tuned for that customer's industry.

And what will happen is the companies that implement AI to better execute, they're not going to just bank those savings and then become 3% or 5% more profitable. They're going to reinvest those dollars back into the business. And over a couple of generations of that, I think it will be so baked into how we run our businesses.

That it won't even be something you can kind of compartmentalize or decouple. AI is sort of one of these productivity enhancers where it'll just get baked into what we're doing to a point where we can't imagine working without it. But it's sort of equally strange to call it out as an explicit driver of the 20% beat we did last quarter in our sales.

It would be more like you'd have to almost call out if you weren't using AI because it's just so assumed that obviously your engineers are writing code 10% faster because they can go look up the answer or have AI generate the code faster for them. And that's just built into our kind of new benchmark of productivity. As long as you're using AI.

As long as you're using it. As long as you don't.

and solving problems faster. We were trying to solve this UX problem and I went to Anthropic and had Anthropic put together the solution so we could see it really quickly. And for me, that saved...

You know, three days of back and forth of trying to make a decision on a particular capability, and Anthropic did it in four minutes. Not four minutes. You personally used Anthropic to do this. Yeah, but, like, it was easy. Anybody could have done it. And so maybe that four minutes was actually a lie, like probably 25 seconds. And so...

I don't know how to price in going from 25 seconds to a couple days of back and forth, but you multiply that by hundreds of people or thousands of people in an organization, and the company is just moving faster. It's shipping more. It's serving its customers better. But again, I think everybody in the economy will have to do that to stay competitive. And so that just becomes the new baseline that we're all participating in.

Aaron Levy, it's always a pleasure to talk with you. And yeah, I just want to thank you for joining us. Hey, thanks for having me. Obviously, very exciting stuff to come. And that's Bold Names for this week.

Michael LaValle and Jessica Fenton are our sound designers. Jessica also wrote our theme music. We got help this week from Julie Chang, Catherine Millsop, Scott Salloway, and Philana Patterson. For even more, check out our columns on WSJ.com. I'm Tim Higgins. And I'm Christopher Mims. Thanks for listening. Say this is your financial life. Over time, things can get more complex. With a personalized plan,

Merrill can help you navigate it all. Learn more at ml.com slash bullish. Merrill, a Bank of America company. What would you like the power to do? Investing involves risk. Merrill Lynch, Pierce, Fenner & Smith, Inc., registered broker-dealer, registered investment advisor, member SIPC, a wholly owned subsidiary of Bank of America Corp.