Author: Lincoln Wang | Founder of MindsLeap | Global Partner at Founders Space | Founder of Founders AI Club
Aaron Levie's father recently got hooked on using AI to write code. The elder Levie, hardly a programmer, clicks "yes" without hesitation every time Claude Code pops up asking whether to download a package. Levie later joked with a pained smile: "His machine is probably full of malware by now — every NPM attack that's shown up in the last three days, he's taken every single one."
The story came up during a recent LangChain conversation. Levie is CEO of Box and a veteran of enterprise cloud storage. He used this slightly absurd anecdote to引出 a judgment every entrepreneur should take seriously: the stunning capability AI coding agents have demonstrated does not mean enterprise knowledge work will see equivalent gains in the same timeframe.
Between them lies a chasm. And it runs far deeper than most people realize.
The four privileges of the code world
Levie posed a rarely discussed question publicly: why is AI so good at coding?
His answer: the code world enjoys four near-perfect conditions. First, models were overtrained on code because it is one of the most abundant content types on the internet. Second, the output of code work is verifiable — it either runs or it doesn't. Third, the people using these tools are themselves highly technical; when the agent makes mistakes, engineers know how to fix them. Fourth, codebase permission structures are typically open — engineers naturally have access to most of the data they need to do their jobs.
"But knowledge work is nothing like that."
In an enterprise, every team and every individual has a different combination of data access permissions. Agents inherit these permission constraints, and complexity grows exponentially. When a person needs a file, they can go ask their colleague Sally for access. But agents lack the social skills to navigate a complex permission maze.
This contrast matters because it explains why most entrepreneurs get excited watching coding agents perform, then get frustrated trying to bring AI into sales, marketing, finance, and other core business functions.
An overrated analogy
Levie reminded the audience: "You have to step outside the tech bubble for a moment and think about this."
There is a massive cognitive gap between the cutting-edge capabilities people in tech see every day and what the vast majority of knowledge workers can actually use. This gap is not temporary — it is rooted in the nature of the work itself.
Knowledge work output is not as easy to verify as code. Whether a contract is well-written, whether a marketing plan is effective, whether an insurance claim decision is accurate — these are hard to confirm with simple tests. Model capability in non-coding domains is still improving, but it is far from the maturity level seen in code.
The signal is clear: when setting AI strategy, enterprises should not extrapolate the entire organization's intelligence timeline from the pace of coding agent progress.
Who builds the bridge
So how do you cross this chasm? Levie sees two directions.
One: highly vertical, domain-specialized AI agent startups are rising. These companies have an edge because they already understand the context of specific domains, know which data sources to connect, and typically come equipped with "forward-deployment engineers" to help enterprises integrate. "If you're doing vertical agent work, this is a massive opportunity."
The other direction comes from inside the enterprise. The teams that understand their own business processes best are always internal teams. Levie noted that some companies have already started creating internal "forward-deployment engineer" roles for AI agents — people on the technical team dedicated to bringing agent technology into the organization, understanding business needs, and completing system integration.
Both directions point to the same conclusion: AI diffusion in enterprises is not a pure technology problem — it is an organizational design problem.
Box's answer: fix the plumbing first
As a company that has long provided content management for large enterprises, Box occupies a unique position. Levie said that even before thinking about AI, they had already done extensive "core plumbing" work — document conversion pipelines that transform enterprise content into formats agents can process, security permission controls that ensure agents can only use data within authorized boundaries.
Box's agents are designed as enterprise content processing experts, then connected to domain-specific agents like Harvey (legal) or general-purpose agents like Claude via MCP servers. It does not try to cover every domain. It excels at what it does best, then collaborates with other specialized agents.
"This is a plumbing problem." Levie said. Before agents can truly run, data, permissions, security, and governance infrastructure must be in place first.
Cost will be the next big story
In the second half of the conversation, the two addressed a rapidly emerging reality: token costs are exploding.
Uber's CTO has publicly said they ran out of token budget. ServiceNow issued similar warnings. Levie mentioned that the "unlimited token consumption" meme lasted only three days before being busted — because only companies at Facebook's scale can truly afford unlimited usage.
"For normal companies, we need budgets. We need planning cycles."
VC-backed startups in Silicon Valley can convert venture capital directly into token consumption. But Wall Street banks, public companies, and traditional enterprises cannot. They have EPS targets. They cannot suddenly add ten million dollars in AI spending in a single quarter. As cost pressure mounts, companies will be forced to optimize — adopting multi-model hybrid strategies, routing simple tasks to cheaper models and reserving the most powerful models for core tasks.
Levie's judgment: cost may be the defining story of the next three years. When automating an insurance claims process could mean fifty million dollars in additional revenue or savings, it is worth dedicating the optimal model and custom instructions to that workflow. But not every workflow has that kind of ROI.
This is a five-plus year deployment window
Levie closed with a timeframe: "We're going to go through years of technology diffusion, change management, and adoption."
Coding agent capabilities are still accelerating, but the intelligentization of knowledge work will be much slower due to all the issues above. This cognitive dissonance will persist for years. Which is why he disagrees with those who think the AI bubble is about to burst — not because the technology has problems, but because enterprise-grade diffusion moves far slower than people expect.
For Chinese entrepreneurs, this judgment carries several implications.
Don't rush to promise your board a company-wide intelligence timeline based on coding agent performance. Knowledge work AI-ification will take longer — that doesn't mean the direction is wrong, just that the path is longer.
Invest in the "boring" infrastructure work. Data governance, permission systems, document standardization — these are not the sexiest parts of an AI project, but they determine whether agents can actually run inside your organization.
Consider the vertical path. If you are an industry practitioner, your domain understanding is itself a moat. General-purpose model companies don't understand your business processes. You do.
The most ironic part of Levie's father story: when a non-technical user faces AI tools, the biggest risk is not inability to use them — it is being too willing to trust. Enterprises introducing AI agents also need to find the balance point between trust and governance.
This is not just a technology deployment challenge. It is a rebuild of organizational capability.
About MindsLeap
MindsLeap is an AI-native organization transformation accelerator.
In deep partnership with Silicon Valley innovation incubator Founders Space, MindsLeap connects cutting-edge global AI insights, the Silicon Valley tech entrepreneurship ecosystem, and real-world transformation scenarios for Chinese entrepreneurs.
MindsLeap is building a transformation ecosystem for entrepreneurs, startup founders, AI engineers, industry experts, and investors — helping enterprises move AI from awareness, strategy, and tools into organizational capability, business processes, product innovation, and growth systems.
This article was translated and adapted from the Chinese original with AI assistance.