TL;DR The gap between what AI models can do and what enterprises actually use is widening, because capability moves at the speed of compute while adoption moves at the speed of people and process. Aaron Levie's read is that the real blockers are context and data access, and that as agents come to outnumber people, software goes API-first, pricing shifts to consumption, and the applied layer keeps most of the value.

Check out the full podcast with Aaron Levie below.

Aaron has watched three platform shifts from the same seat. Cloud gave enterprises a place to put their data, mobile made that data usable in more places, and AI is what happens once you can finally throw compute at all of it and ask questions of everything at once. Twenty years and a public company later, Box sits inside a large share of the Fortune 500, which is exactly why his read on where AI stands is worth slowing down for. From inside the enterprise, the view looks different than it does from Silicon Valley.

What is the model overhang?

Every meaningful jump in model capability quietly opens another tranche of work that agents can now do, across legal, financial services, healthcare, and life sciences. Aaron's rough estimate is that something like a Fable 7 or Fable 8 will open up another 70% of the next tranche of enterprise use cases through capability alone. Yet the models already do far more than they've been deployed to do, and that distance between what's possible and what's in production is the model overhang.

His counterintuitive point is that the overhang keeps widening. Model progress and diffusion run on different clocks, and only one of them is speeding up.

Compute compounds. Org charts, procurement cycles, and security reviews don't, so the delta between what a model can do and how much of it an enterprise has actually wired into its operations grows even as the models improve.

Why did AI coding take off first?

Coding is the tell. AI took off inside Silicon Valley faster than anywhere else because writing software happens to have close to the ideal shape for an agent. The codebase is already text, engineers already have access to the code they need, the frontier models are trained heavily on code, and the output is verifiable through tests and CI. On top of that, the people doing the work are online and sharing what works in near real time, so best practices spread in days.

Step one pace outside that bubble and the same models stall at the door. A less technical user, legacy systems, data fragmented across a dozen environments, access controls that won't let an agent roam freely, and unresolved security and compliance constraints all sit between the model and the work. Aaron thinks bridging that gap is a five to ten year project, and the encouraging part is that the models keep getting roughly 4x to 5x better across those years, so by the time a real-economy enterprise rolls out its agent strategy, it's building on far more capable models than exist today.

Why is every AI problem really a context problem?

When Box talks to customers, Aaron estimates that about 90% of what sounds like an AI conversation is actually a data and context one. The reframing underneath it is that agents use data much more like people do than computers ever could. A decade or so ago the only thing you could really automate was structured data sitting in a database, which is a tiny slice of what an enterprise generates. The other 90% or more lives in PowerPoints, research documents, contracts, log files, and marketing assets, and traditional compute could do very little with it. Agents can finally work with all of it, which is why the file system is having a rebirth as the place agents read from and write to.

The unsolved part is getting the right context into the agent at the right moment.

An agent can move at a thousand times human speed, but it only gets that short window to be handed what it needs. Solving for it means giving agents reliable access to the systems that hold enterprise knowledge, through APIs where they exist and through another path where they don't.

How do agents reach the systems that were never built for them?

Aaron named that other path directly. When we asked what's still blocking agents from proliferating, one of his answers was the messy mix of access, data formatting, and API coverage across an enterprise's systems, and then he pointed at the browser.

Most enterprise software was never built with an agent in mind. Some of it exposes an API and most of it doesn't, but nearly all of it has a browser interface, because that's what a person needs to use it. If an agent can drive that interface reliably, the long tail of systems that never got an integration suddenly becomes reachable, with no vendor cooperation and no net-new API work required.

That's the work Browserbase does. We give agents a real browser they can operate at scale, with the proxies, stealth, and session observability that keep automation working once it leaves the happy path, and the concurrency to run many of them at once.

What is headless SaaS, and why does per-seat pricing break?

Matthew Prince of Cloudflare recently noted that agents had crossed over human users on the internet, and Aaron's read is that the ratio only compounds from here. The number of humans online roughly doubles at best, while the number of agents can grow by an order of magnitude a year, so you end up with something like a Moore's Law for how many agents there are per person. Extend that inside the enterprise and the same ratio shows up: if a company runs a hundred agents for every human, its software has to be built API-first, with the machine interface as the primary way in rather than a bolt-on. Aaron calls this headless SaaS, and Box arrived at it early almost by accident, since it has run a serious API since close to day one.

The harder consequence is what it does to pricing, and Aaron thinks Wall Street has this one backwards. A person logging into Workday is predictable, gets near-unlimited usage, and should stay a seat, since there's no reason to blow that model up. An agent behaves nothing like that predictable user.

Seats and consumption end up running side by side, seats for the humans and metered usage for the agents, and most SaaS business models still have to be upgraded to hold both.

Why the applied layer wins, even if you're bitter-lesson-pilled

There's a real question hanging over every applied AI company: won't the frontier labs eventually go deep into your vertical and swallow it? Aaron reaches for the hyperscaler era to answer it. AWS, Azure, and GCP had every incentive and all the raw infrastructure they needed to commoditize the layers above them, and yet a decade later Databricks is worth around $175 billion, Snowflake around $90 billion, Cloudflare around $100 billion, with Mongo, Twilio, and a long tail behind them. That's roughly a trillion dollars of market cap sitting on top of the hyperscalers, doing the last-mile work the infrastructure layer structurally doesn't do.

The frontier labs will run a similar play, picking a few verticals and going deep enough to win them, and those wins will get read as proof that the frontier eats everything. Aaron's bet is that outside those pockets a healthy ecosystem of vertical applications sits on top of horizontal models, handling the compliance, security, data access, and change management that a model alone doesn't touch. The applied companies bring the specialists and forward-deployed engineers, wire the right data systems into the agent, and make sure a model upgrade doesn't quietly break the plumbing.

This year added a second reason that only started to bite recently, which is cost. The spread between the very edge of the frontier and the frontier of three months ago can run about 10x, and routing to a tuned open-weights model can push it to 50x. That turns the ability to route each task to the right model tier into a real economic lever, and only the companies that understand a task deeply can pull it off. As Aaron put it, you can't do intent-based routing on a contract review in an M&A deal unless you understand the use case well enough to know when to reach for a frontier model like Fable or GPT 5.5 and when a tuned open-weights model like Nemotron will do the job.

Why distribution is the new moat

Aaron's advice for founders starting from zero today lands in the same place. The cost of building software has dropped by something like 10x, so high-quality, well-built, tasteful product is now table stakes rather than an edge, and job one is separating yourself from the flood of slop that cheap building produces. Most companies stop there, ship the product, and wait for users who never arrive.

In a world where almost anyone can build the product, the scarce resource becomes reaching customers and understanding their problems well enough to actually solve them. That's why Aaron sees the forward-deployed engineer as a durable trend rather than a fad, since real-world enterprises need help standing agents up in their own environments and often lack the internal depth to do it alone. The startups that invest early in FDE motions, sales, and consultative delivery are the ones he'd bet on to win the accounts that matter.

Where this leaves us

If Aaron is right, and the pattern from three prior platform shifts suggests he usually is, model capability keeps compounding faster than enterprises can absorb it, context wins where compute already has, and the applied layer thrives, especially the parts that can route intelligently between model tiers. Every business gradually rebuilds its software surface for agents to consume, and the founders who win the next decade treat distribution as seriously as they treat their code.

The browser sits underneath much of that. It's the interface nearly every enterprise system already exposes, the way into the long tail of software that never got an API, and where agents will do a lot of the last-mile work Aaron describes. If you're building agents that need to operate the web reliably and at scale, you can start building on Browserbase today and grab an API key in a couple of minutes.

Watch the full conversation with Aaron Levie on Navigators, where he goes deeper on Box's own agent rebuild, model routing, and where the real economy is quietly starting to catch up.