A host on Sequoia's Training Data put the puzzle to Box chief executive Aaron Levie bluntly: coding agents arrived and became, in the host's words, the fastest diffusion of anything into the economy we have ever seen. The rest of "the AI magic," spreading into everybody else's job, has been much slower.
Levie's answer was that you always have to compare coding against everything else to see how unalike they are. He then counted the reasons — "like five things. Oh, and maybe like the sixth."
Six things coding had going for it
First, code is text, and the text is the product. Levie's thought experiment is a programmer who never had to sleep or eat, who could intuit what to build and sit at a computer all day: the value created would be almost perfectly correlated with the hours spent typing. Lines of code, ideally good code, are the thing most correlated with whether you produced software people wanted.
Second, models are, as he puts it, hyper-trained on code. Third, the AI labs treat coding as a competitive benchmark they are constantly trying to beat. Fourth, they get to evaluate their own systems on it every day, because they are the ones writing the models' code.
Fifth — and this is the one that decides whether a rollout survives its first week — programmers are, in Levie's phrase, the most technical audience of all time. When an agent hits a bug, or an MCP server (the connector that lets an agent reach an outside tool) comes back with "connection invalid," they know how to triage it themselves. They don't call IT. They say: "Oh yeah, no, I didn't open up that port. Sorry, I'll go fix it."
Sixth, it is a very high-paying kind of work, so a 10% or 20% productivity gain is automatically worth having, let alone a fivefold one.
A continuum, from typing to waiting for a customer
Levie suggests imagining a histogram of how similar other jobs are to coding — something he says he has not seen published. Legal work sits fairly close: a lot of value is created by someone at a computer reviewing documents, drafting documents and processing large amounts of information. That, he says, is blowing up.
Further down the list is a sales representative. Their contribution to the economy is convincing an external customer to buy software, or technology, or a Caterpillar truck. Bring them the world's best automation and they still have to wire it up and hand it their data — and even then they are rate-limited by whether the customer replied, whether the customer wants to meet, whether Tuesday works instead of today, whether there is any budget. That, in Levie's framing, is the whole continuum of knowledge work: at one end, someone constrained by external factors they cannot automate away; at the other, someone who can sit and type all day. And this, he notes, is before you get to work involving atoms.
No GitHub for knowledge work
The second obstacle is where the data lives. Levie recalls the period when launching a coding agent involved no sign-up and no registration — just "give us your GitHub," because an engineer's code is simply in GitHub. "There's no, like, give us your GitHub for knowledge work," he says. There is "give us your Box" — and Box customers, he argues, have an easier time of it — but he immediately discounts his own case: at a $1.3 billion revenue run rate, a lot of people are not using Box. Their material sits in on-premises systems, legacy file shares and enterprise environments that, as he puts it, don't talk to agents particularly well.
Then there is the point he says he had forgotten about coding: access control. An engineer generally has access to most of what is relevant to their job. In knowledge work, getting at a file share means asking a colleague to open it up for you — "Hey, Sally, can you open up that sort of file share for me?" Someone has to work out how an agent gets that same set of permissions.
All of that sits on top of the operational list Levie ran through earlier in the conversation: connecting to other data systems, moments that require a human in the loop, delays in a process that leave an agent sitting idle, change management for the business process itself, legacy systems nobody has modernized. Five or ten items, he says, that are much more blocking and tackling than the raw intelligence of a model — and the last thing a classic research organization wants to do is attack every single one of them. In the Valley, he says, everyone is "research-pilled," for good reason, and it produces the breakthroughs; but the model can be the most intelligent thing in the world and the workflow still needs all of that.
Why he thinks the slowness is the opportunity
So Levie draws two conclusions. One is that Silicon Valley should prepare for diffusion taking a lot longer than it thinks. The other is that the delay is, for him, good news, because the work it creates is exactly what an application company can sell. Getting the technology to the lawyer, the sales representative, the life sciences researcher, the person running a customer support team requires patience, domain expertise and, in his words, willingness to "pound pavement" — it is not a floodgate opening on its own.
He reaches for a precedent rather than a projection. Cloud platforms such as AWS, Google Cloud and Azure created trillions of dollars of infrastructure value, he says, and also made possible trillions of dollars of software that exists only because of them. Looking at early Google Cloud a decade ago, he argues, nobody would have predicted Snowflake or Databricks; the obvious reaction would have been that the infrastructure already does that. He expects intelligence to work the same way, with the models insanely valuable and the job of bringing them into real workflows in banking, life sciences, healthcare and government amounting to a great deal of software.
Levie is explicit that this is a wager, not a measurement. He estimates that roughly a trillion dollars has been bet on one of two outcomes: that the bridge between a model's capability and an enterprise's workflow is narrow, or that it is vast. He is long the application layer and says so with the caveat attached: "I'm also equally very biased," with "a very concentrated bet with very limited diversification." A host described the mirror-image dilemma from the investor's side — whether to put another billion dollars into Anthropic or try to work out what plays out at the applied layer. The open strategic question, Levie says, is how far the model providers move up that stack over the next two to five years, caught between wanting to be closer to the customer and wanting an ecosystem that trusts them.
Where the advantage moves
His closing argument comes out of something he watched a week and a half earlier. A demo of a product he says would easily have been a 40-person project five years ago — built by two people. "How do you have so many tabs that work?" he remembers thinking, with functioning things behind all of them. He is jealous of founders who can start from that as a design principle. Box, he says, will muscle its way there, with a permanent handicap: it is not willing to drop code review, because customers cannot entrust it with their data security and compliance if it does, so it will always take "a little bit of a discount on the productivity." What he does not envy is the other consequence: every good idea now draws five competitors within days, where Box once had a couple of years to grind on its product.
Put those together, and Levie lands on a position he admits is probably consensus: if AI builds things much faster, the advantage shifts to whoever can actually get the thing to the customer. He describes founders, referred to only by their first names, who have accepted that mandate — that theirs is going to be an enterprise diffusion play, so the work is getting into the enterprise. Anyone who mistakes the mandate, in his account, loses: "It's just like, it's game over, sorry." There are trillions up for grabs at the applied layer, he says, and the companies that know how to build the teams and reach the enterprise are the ones that will win.