Two weeks before this conversation on Sequoia's Training Data podcast, Aaron Levie, the chief executive of the file and content company Box, noticed that one employee was using AI in a way the rest of his team was not. His request was simple: get everybody in a room and have him show them. About six hours later, Levie says, the team was assembled and the colleague was running a full demo of his own workflow. "Shout out to Mick," Levie added.
That small episode sits at the centre of how a twenty-year-old software company is trying to convert individual experimentation into something a whole organisation does. Levie's argument, laid out in the episode, is that the hard part is not getting people to try AI tools. It is finding the handful of people who have already worked out something useful, and then making that knowledge travel.
A leaderboard that is not a competition
Asked whether Box runs leaderboards for AI usage, Levie's answer was careful. "We don't token max," he said — meaning the company does not treat consumption as a goal in itself. Tokens are the units of text a language model reads and writes, and they are how usage of these systems is metered and billed. Box does keep a literal list of people ranked by the number of tokens they use. But the list is a starting point for inspection, not a scoreboard: the question is whether a given person's heavy usage is actually useful, or whether there is a lesson in it that should be carried back to another function.
By Levie's rough reckoning, among the top three AI users at Box, one is probably wasting about half the tokens they consume, and the other two are doing something that warrants an internal training session for everyone else. A list that only rewarded volume would promote the waste alongside the insight; looking at what the leaders actually do separates the two.
Around that sits some ordinary organisational scaffolding. Box's CIO is, in Levie's phrase, "very AI-pilled." The company runs what he calls a little bit of a centre of excellence on AI, and has hired internal AI FDEs — forward-deployed engineers, in the industry's shorthand — to help with these processes. The first task he names is targeting: working out where the highest-leverage workflows are going to be, rather than spreading effort evenly.
Why Box's own files made this easier
The interviewer asked how a company makes its business legible for AI. Levie says Box benefits from an accident of its own history. With no exaggeration, he claims, any question about the business that has ever been written down in unstructured form — a meeting note, a project plan, a roadmap, a presentation, a financial document, a planning session — is in Box. "Unstructured data" here means the ordinary documents people write, as opposed to the tidy rows of a database.
That did not happen by itself. "We didn't let anybody use anything else," Levie said. Alongside Box sits Salesforce and the other systems he calls canonical — the single agreed home for a particular kind of record. The result, he argues, is data hygiene good enough that agents running on top of it can work at scale, which makes operating in an AI-first way easier than it would be for a company whose records are scattered.
Features delivered, not lines of code
On output, Levie is specific about what he is counting. In some parts of Box's stack, he says, the company is shipping two to three times more actual customer-facing product. "I don't care about how much code," he said — the test is whether Box delivered more of the functionality customers are asking for. He is equally quick to add the other side: talk to a friend at Anthropic, he says, and the reaction is that Box still has a long way to go.
Part of the difficulty is that the target keeps moving. Two years ago, Levie recalls, the state of the art was a plugin in a developer's editor, and adopting it felt like a stretch. Then the instruction became: everybody roll out Cursor, the AI coding environment. By the time that had actually happened, the question had shifted again — why are you not working from Slack, mentioning bots to get your code done? Box has a few agents in that shape, and Levie says he is excited about the Slack-based form factor, while adding that Box is not as Slack-pilled as Anthropic or OpenAI. What counts as the workflow paradigm, he says, is constantly changing.
The productivity discount Box accepts
There is one thing Levie says Box will not do. Some of the fastest-moving teams talk about removing code review — the step where another engineer checks changes before they ship. Box is, in his words, very uncomfortable with that, because its customers could not possibly entrust it with their data security and compliance if it did not take those obligations seriously. The consequence is explicit: "we're always going to have a little bit of a discount on the productivity because of where we are in the stack, what we do as a business."
That is also why he is openly envious of founders starting now. A week and a half before the interview, he saw a demo of a product he says would easily have been a 40-person project five years ago, built by two people — with working tabs behind working tabs, and nothing faked. A new company can take that as a design principle from the first day. Box, he says, will get there by muscling through.
The personal version
Levie's own habits run on the same instinct. He describes himself as an average AI-pilled knowledge worker who asks one of five different AI systems twenty to thirty questions a day — researching markets, looking for talent, checking what a competitor is doing. He plays with everything: he is in pre-release on a couple of personal assistant products, with a backlog of others he has not got to yet, and says browser use still shows its limits. His fix, offered half-seriously, is that the entire internet probably needs a command-line interface for its products, so agents have something to call instead of clicking through pages.
His learning diet is blunter still: probably 95% Twitter, scrolling the feed at the end of each night and trying to triangulate all the information. He says he would like to send an emergency alert to every college sophomore and junior telling them to join and follow a set of accounts, because on his reading you are either a year ahead or a year behind based on your feed. It is the same bet he is making inside Box — that the useful knowledge exists somewhere already, and the work is finding whoever has it and watching what they do.