Box Holds 100 Billion Files. Its CEO Says Enterprises Only Move for a 10x

Box is an intelligent content cloud that, Labenz notes at the top, securely manages more than 100 billion documents for more than 100,000 customers — AstraZeneca, Teach for America, Morgan Stanley, Airbnb, Broadcom, the United States Air Force. Aaron Levie, its co-founder and CEO, has spent two decades in enterprise software and rode the cloud wave as an early SaaS pioneer. He tells Labenz that leadership at those organizations is more energized about AI than about any other technology in his career, cloud included. Before they get to the product, Labenz asks for a personal warmup: what does Levie actually use, and does he think AGI is close?

The definition of AGI is too amorphous for a forecast, Levie says. He is downstream of whatever Ilya, Sam, or Greg are saying, with no better crystal ball than anyone else. What he will stand on is the slope. The rate of exponential improvement in AI models is incredible. If that pace holds, the systems will increasingly perform any general task you give them. Reasoning models are already posting strong results in math, complex logic, and coding. Those foundations, if the benchmarks keep rising, are the building blocks for models that can keep learning on their own. "Within the next few years we could be getting some form of whatever we would have previously defined AGI as."

His phone is a quieter piece of evidence. For a decade and a half, the home screen changed at a crawl: Uber, then Spotify, maybe a social app, WhatsApp, something new every year or two. In the past six months he has added at least five apps — talking to Gemini Voice, using ChatGPT with video, querying Perplexity, playing with xAI and Grok, prototyping with Artifact and Claude. AI is leaking into both personal and professional life. That infusion, he says, is already more visible on a lock screen than any consumer-software wave he can remember.

Labenz points to a LinkedIn post in which Levie called this the most energized he has seen enterprises about a new technology. Cloud was the last wave people remember as enthusiastic. Levie was in the room for that one, and he would not describe the early cloud as super excited or aggressively pursued. The conversations were skepticism, resistance, friction. Enterprises had to move infrastructure out of data centers they managed, trust vendors they had never treated as enterprise suppliers — Amazon was atypical — and swap a physical idea of privacy and compliance for APIs, dashboards, and audit reports. Reluctance was the mood.

Snap the line at two or three years into cloud versus two or three years into AI, after a recent trip to New York where he met a couple dozen CIOs and customers, and the environments could not be more different. Enterprises are looking for almost as many AI use cases as they can find, in many cases more than is practical. Cloud was utilitarian: take this piece of infrastructure and virtualize it. AI sounds like: what if we could solve a business problem we have never been able to attack, or put human talent on more interesting work? Creativity and energy are in a different register.

He puts a large asterisk on the excitement immediately. At-scale deployments in the enterprise are still very early innings. For someone in enterprise software, though, the implication is an insane amount of opportunity: Microsoft, Oracle, and Google at the top of the stack; Box, Salesforce, and ServiceNow in the application layer; and brand-new startups building AI agents that solve a whole new set of problems. The window is relatively narrow. Inside it, he expects a tremendous amount of change and a tremendous number of new companies, with enterprises ready and excited to pursue them.

Labenz hears a reversal of the cloud story. Cloud took physical management away from IT and put uptime on the vendor. Levie is saying IT departments will go from supporting the work to doing more of the work. That will take a dramatic transformation. For decades the pattern was stable: marketing, sales, or finance asked for a CRM or an HR system; IT reviewed vendors, deployed the technology, and enabled the business; the business still owned execution. Sales still closed deals. They just used IT's CRM to do it.

AI flips that. The business will go to IT and ask it to deploy AI labor against the problems of the day. A head of sales might ask whether there are agents that can spin up a new campaign. A back-office team might ask for agents that review contracts, invoices, or client onboarding. IT is no longer only deploying software so the business can solve a problem. It is being asked to solve the problem. That means understanding the business more deeply, becoming more strategic, and tracking the ecosystem of companies that produce agents. Jensen Huang at Nvidia put it best, Levie says: the IT department becomes the HR department of AI.

If IT is HR for AI, Labenz jokes, Box is applying for a promotion. Heading into 2025 the company has been shipping generative features as Box AI, an abstraction layer that connects foundation models to enterprise content without dropping security, privacy, access controls, or permissions. Three capabilities sit on that layer. First, talk to your data: take a hundred sales presentations and ask for best practices selling a product to a new customer, or talk across life-sciences research for patterns — retrieval-augmented generation over many documents. Second, extract structured data from unstructured files — renewal dates and party names from a contract — into a database you can query, dashboard, and automate. Third, agentic workflows: a contract agent that pulls the riskiest clauses, routes them, and connects into Salesforce, ServiceNow, or Microsoft. Levie calls it the most the company has ever built, because new things that were never possible with the data show up on a daily basis.

Labenz's first digging-in question is the industry's RAG complaint. Companies rushed to ship a version of exactly what Levie described, and it often has not worked. The usual failure, once you look under the hood, is that embedding search never retrieved the right sources, so the answer never had a chance. Levie asks whether that trough of disillusionment is worse on broad, heterogeneous sets — email plus files plus calendar. Labenz says it pops up even when the situation is not that complicated.

Levie would love to call the fix genius. They got lucky. About a year before ChatGPT launched, Box was building Hubs, unrelated to AI: a way to organize content topic by topic with many-to-many virtual pointers. The same document can sit in twenty hubs without being moved or having its permissions changed; one update propagates everywhere. A sales hub, a Japan-region sales hub, and a product-line hub can all point at the same source of truth instead of forked copies that drift. Search and discovery were already on the punch list. The day ChatGPT shipped, they realized you should be able to talk to everything in a hub.

That turned out to be the architecture RAG needed. Imagine 100 million files and someone asking for last quarter's revenue figure. The corpus has `earnings_final`, `earnings_draft1`, and `earnings_draft1_sally_edits`. The embeddings all look relevant; only one version is right. Multiply by years of drafts and signal from email, and generic questions on internal data get hard. Public tools such as Perplexity inherit something like PageRank: a CNN article outranks a random blog. Corporate data has no authority score. Hubs make the user declare it. You put the official sales decks in the sales hub and you ask sales questions there, not HR questions. Two or three such hacks, doubled down on, have made Box's RAG "at least a hundred times better" than a broad deployment across all your data.

They still pinch themselves. The file-system work took a year to a year and a half; there was no way to build it faster given the complexity of the platform. If they had not already been that far in, Levie fears the speed of AI would have made the architecture look too daunting. They might have shipped a less optimal version. They got lucky, and they built the right thing for the right moment.

Labenz's working contrast is autonomy: how much discretion the model has, versus an intelligent workflow with the path pre-wired. Levie subscribes, then deliberately takes the broadest definition so users are not staring at 17 flavors of the same product. In Box, an agent is a mix of one or many models, tool use on the platform, underlying skills — system prompts plus some proprietary architecture — and access to your data. That collection is an agent. Talking to a single document counts, even when nothing sequential is happening. You can also build a custom sales agent with instructions about language and workflow. That first generation, he admits, is what people would have called assistants two years ago. They wanted one word.

Where it is going is closer to Labenz's definition: multi-step, probabilistic flows. Review a document, then kick off another process to an agent or a human; collate a lot of data and produce something. Software has always been good at deterministic workflows. The vast majority of work is not. It requires judgment and changes with new inputs. That majority, in the future, is non-deterministic, judgment-oriented, agentic work — which is an incredible opening for enterprise software, because that is what can finally be digitized.

He periodizes the industry in three eras. Forty years ago, systems of record: CRM and ERP, rows in databases, 100% deterministic. Then systems of engagement: Slack, Box, collaborative tools, less structured, more fluid, messy in a human-to-human way. Now systems of intelligence: AIs automating those workflows, with the structure of a system of record and the flexibility of a system of engagement. Labenz offers a working definition of intelligence that Levie likes: the ability to do useful work when there is no explicit algorithm for every step. The enterprise corollary, which Labenz also presses, is that if traditional code can do it, use traditional code — faster, cheaper, more reliable. AI is for the fuzzy zone. The annoying fact is that many old systems do not even have APIs, which is why browser access suddenly looks exciting: automate five systems that cannot talk cleanly to each other.

Labenz's thesis is that the models can already do more than enterprises are deploying. Tyler Cowen recently told an audience that the humans are the bottleneck. Levie agrees that AI can do vastly more than most enterprises think, or have the near-term appetite to roll out. There is more technology available than people realize. He has to remind himself: the task he was about to ping someone for, try Artifact, try AI first. He sits in the center of the industry and still has to trigger that reflex. The gap for everyone else is probably massive.

He will not let the model providers off the hook. You cannot run an enterprise workflow in a regulated industry at 98%. You would not find it acceptable if 98% of scheduled flights succeeded and 2% of the time you showed up at the airport with no ticket. Enterprises want 99.99999% reliability on anything that really matters. Creative work — a marketing campaign, a blog post, an email — can tolerate some hallucination because a human can edit. A billion financial transactions a week or a month, with AI wrong on 0.01% of them, is a non-starter. Benchmarks are climbing, but evals still have to be reset because everything is not at 100%. Cost is the second gate. Customers say they would love 10,000 agents, then look at the bill and take it stepwise. Privacy and security concerns remain. And there is a person in that seat today; until you can move them or teach them a new skill, you will not automate the workflow. This is a decade-long change. The cloud already wrote the roadmap.

Labenz pushes. If intelligence is what you use when there is no algorithm, and if with context and examples you can often get AI on par with a human at lower cost — medical diagnosis is a recent example — why wait for a 10x lower defect rate, the way the industry waited on self-driving? Levie says he nailed it, as advice to any vendor. You cannot walk in and say you will do exactly what they do today, 40% cheaper. An economist would take that deal all day. In real life that person has 17 other projects. Forty percent cheaper might be ninth on the list, and then they have to go to the AI council, run a six-month test, and weigh it against everything else. You need multiples on the status quo. Not incrementally better, cheaper, or faster — an order of magnitude on one of those dimensions. A tenth of the cost to review contracts, or 10x better at finding the cancer. That is also why the biggest AI opportunities are work that is not automated today. You are not replacing a process; you are layering on. Copilot and Cursor broke through because the job did not change and productivity showed up immediately. Replacing an existing process to save a little money is a much harder enterprise sale.

The narratives on enterprise software contradict each other: never a better time to start a company; actually incumbents will capture the value because the tech is easy to copy and they already have the customers; Klarna reportedly shutting down systems of record and maybe building its own AI HR stack; seats versus outcomes. Levie says if he had heard that pile at the start of Box he would have been too stressed to found a company. He leans on timeless competitive strategy. A thin layer on OpenAI is a bad idea. A thin layer on Salesforce is a bad idea. Salesforce will build the CRM AI thing; Workday the HR thing; ServiceNow the ITSM thing. A slight gap in what OpenAI does today is a risk that the model layer moves up. Opportunity sits where those players will not naturally go: a lot of non-AI interface and workflow labor, cross-platform AI workflows none of them will lean into, agents orthogonal to Salesforce or Workday strategy, enough traction for a moat. "There's going to be a tremendous amount of AI startup opportunity but it will not come from just doing an AI-first CRM system, because you should anticipate that Salesforce is an AI-first CRM system."

Klarna, as he reads the reports, is the exception, maybe a little overplayed, and nothing about it is impossible. Building your own Workday with AI is simply not a priority for most companies, the same ranking problem as before: they will not homegrow an HR system to save a couple hundred thousand dollars. He finds it provocative and invites more companies to try and share lessons. He is not convinced 90% of corporations would ever do it. On pricing he has no doctrine because Box is testing every model. AI's gift is that software can get closer to an outcome, so pricing can too: pay per lead an agent generates; pay for the compute that generates 10,000 leads; or keep a classic subscription that averages volume. Salesforce and a few others invented pay-per-seat twenty-plus years ago, and that was SaaS. The business-model question is open for the first time in two decades. Each company has to learn what its customer is actually paying for.

Labenz asks about a niche: will buyers keep Salesforce and still purchase a separate safety or compliance layer to red-team agents? Levie's yes is instant. Think in stacks. Salesforce, HR, ERP, Box — which things cut across all of them? Those need a vendor independent of any one player. Things that only make sense inside one system are an incumbent game; things that must work across every app in the enterprise are a startup game. Internally, Box has committed to being an AI-first company, in part because it sells the stuff. A new hire learning to sell goes to the sales hub and gets answers as if they were talking to a top expert, 24/7, without waiting on Slack. Coding tools show anecdotal gains of 5 to 10% for veterans and about 50% for new hires who ramp faster. The metric he cares about is shipping more software. On the famous productivity paradox, he will take Tyler Cowen's over on more than half a percent of GDP a year — if you start the clock in a few years. Diffusion through human-mediated environments always takes longer than it should. He leaves with the fact that still startles him: anyone, anywhere, can get 90th-percentile expertise on any topic instantly. Three years ago that was not conceivable.