Aaron Levie on Why Frontier Labs Will Be Way More Valuable Than They Are Today

Why does it matter? Because the AI job apocalypse narrative has it exactly backwards.

Aaron Levie thinks the entire tech industry is having the wrong conversation about AI and labor — and the cost of getting it wrong is a generation steered away from careers that are about to become more valuable, not less. The 85% of the economy that tech has never served is about to need engineers, lawyers, and doctors at a scale Silicon Valley can't currently imagine.

• AI doesn't destroy jobs — it exposes bottlenecks that were always there, invisible because everything was slow and manual • A new role called the "agent operator" will generate 500,000 to 1 million jobs this decade • Token budgets migrating from IT spend to OPEX will effectively double global enterprise technology spend • Frontier lab valuations may still be underpricing the terminal market — Levie would "load up" on rounds even at current prices

More lawyers, more doctors, more engineers — AI reveals the bottlenecks we never knew existed

There are going to be more lawyers in five years than there are today. That's Levie's actual claim — not a hedge, not a hope — and his logic is airtight. AI made it easy to generate legal content. What it didn't do is make it easier to get anything approved by a court, file a patent, or navigate the real machinery of law. The ultimate constraint was always the number of lawyers who could review the work. Now that everyone thinks they're a lawyer, that constraint just got much more visible.

The same dynamic plays out in healthcare. One of Levie's customers automated patient referrals entirely — no more week-long phone loops. Then they discovered the bottleneck had simply moved: the appointment is still 18 months out, because the limiting factor was never the referral process. It was the number of doctors.

His core argument is that tech represents somewhere between 8% and 15% of GDP. The other 85% — tractor companies, pharma firms, banks — doesn't have enough engineers to automate what's about to happen to their industries. AI coding tools don't eliminate engineering jobs; they let John Deere and Eli Lilly finally access the kind of engineering capability Silicon Valley has always taken for granted. A CS grad who once headed straight to Google will now go to Caterpillar — not because options dried up, but because that's where the scale of the problem actually lives.

The "agent operator" is the decade's highest-leverage career move — and almost no one is training for it

There is a role that doesn't fully exist yet, and Levie puts the job creation figure at 500,000 to 1 million positions: the agent operator. Not a prompt engineer. Not a data scientist. Someone who understands both the technical architecture of AI agents — MCPs, CLIs, agent configuration files — and the messy reality of enterprise business processes.

The gap this role fills is structural. When you start a company from scratch, you design workflows however you want. When you walk into a Fortune 1000 bank or pharma company, you inherit two decades of fragmented data, legacy systems, regulatory constraints, and employees hardwired into specific processes. Redesigning those workflows for agents — not for people — requires a hybrid skill set that currently doesn't exist in sufficient numbers anywhere.

And the role doesn't end at deployment. The second a new model drops, the workflow probably breaks. Prompting syntax changes, agent behavior shifts, and whoever built the system has to maintain it continuously. That's real technical and business process acumen combined — and there's essentially no one doing it at scale today. People will come from IT, from operations, from engineering. But getting there early, before enterprises figure out what to pay for this, is the move.

Slow enterprise AI adoption isn't a warning sign — it's a decade-long growth window in disguise

Ask a Fortune 500 company to deploy an agent that surfaces the highest-risk contracts in their upcoming renewals. The agent will find 10 different systems that contain contracts. Half of those will be legacy technologies the agent can't properly interface with — network file shares, obsolete document management systems. The other half will be fragmented because employees spent 20 years using whatever tool they preferred, and no one ever had to standardize anything because humans could always muddle through.

Agents can't muddle through. They'll find documents — just as often the wrong ones as the right ones. Cleaning that up, organizing the data, redesigning the workflow, figuring out where human review fits in — that, Levie says, is "10 years of work for Accenture in every enterprise on the planet."

Silicon Valley reads slow adoption as weak demand. Levie reads it as structural lag in a very long diffusion cycle. He watched cloud do the same thing: spiky early enthusiasm, years of grinding implementation, then a market that ended up vastly larger than anyone projected. Diffusion takes longer than the Valley thinks — and the companies bridging the gap, whether that's implementation services, agent observability tools, or workflow redesign specialists, are sitting at the entry point of something durable.

Token budgets moving to OPEX will effectively double global enterprise tech spend — and that's not a blog post thesis, it's a structural shift

Enterprise technology spend sits at roughly 10 to 12% of revenue. Levie thinks it's heading to 20%. The mechanism is something that has almost never happened in the history of enterprise software: a technology that can be sold into line-of-business budgets rather than being capped by the corporate IT allocation.

Every software sale in history ran into the same ceiling — the CIO's budget. Tokens break that. An agent that makes a marketing team 50% more productive can be sold against the marketing OPEX budget. The trade-off isn't "Salesforce licenses vs. compute tokens" — it's "this campaign vs. this automation." That's an entirely different and vastly larger pool of money.

Enterprise budget cycles will slow the realization of this — companies have EPS commitments and annual planning processes, so token maxing across an entire organization overnight isn't how it actually works. But the direction is set. The token budget moves out of IT spend, into regular OPEX, and the addressable market for enterprise AI companies expands accordingly.

Frontier labs are still underpriced — the cloud analogy says the market size question isn't even close to settled

In 2010, AWS generated $500 million in revenue. Azure had just launched. GCP was called Google App Engine and had a turbine logo with wings. Fifteen years later, it's a several-hundred-billion-dollar annual revenue ecosystem — and every major player won enormously.

Levie says he'd still be loading up on frontier rounds at current prices. The OpenAI vs. Anthropic framing — who wins the enterprise race — is the wrong question, just as AWS vs. Azure was the wrong question in 2010. Enterprises don't want single-vendor AI stacks any more than they wanted single-vendor cloud. The market will be multi-model by default, the total size will exceed current projections, and execution will matter less than people think because rising tides really do lift all boats when the tides are this large.

AI security has a permanent structural asymmetry — two new risks for every one defensive tool

Generate more code with AI, and you immediately have more code than anyone can review. Every feature shipped is a potential vulnerability introduced by an agent that may have made a wrong call — opened a port, misconfigured a permission, written something subtly broken. Meanwhile, offensive actors running open models can now scan for exploits across the internet at speeds that weren't possible before.

The math is ugly: two new forms of risk in the development process, one benefit — agents can also review the code. Levie's framing is clean: "Agents are the solution to the problem that agents have caused." That recursive dynamic, where AI security requires AI tools to address AI-generated threats, means the attack surface grows every time a model is released or a feature ships. This isn't a cyclical security concern. It's a permanent expansion of the threat landscape, which makes agentic security one of the most durable investment categories in the current wave.

The SaaS companies agents need most are the ones the market is currently undervaluing

Some software is mostly buttons — 93 features, a UI tuned for human clicking, value that lived in the interaction layer. When agents do the clicking, that value evaporates. But a lot of enterprise software is something different: business logic embedded in APIs that knows who has access to what data, how to automate a supply chain, what compliance looks like for a FINRA-regulated document. None of that logic goes away when agents arrive. If anything, agents need it more than humans ever did.

Box already runs more API call volume than end-user interactions — the "headless" version of the product has existed almost since founding. When agent-driven API calls go up 100x, that's not a threat to the business model. That's the business model scaling. The right frame for evaluating which SaaS companies survive the transition: not seat count, not UI stickiness — depth of business logic in the API layer, and readiness to serve agents as primary users.

The real bet is on the gap between what enterprises need and what Silicon Valley has built

Everything Levie describes — agent operators, 10-year implementation cycles, token budgets migrating to OPEX, compliance-aware APIs — points at the same underlying reality: the value in AI is going to accrue disproportionately to whoever closes the distance between frontier capability and the fragmented, regulated, legacy-encrusted world where most economic activity actually happens. The labs will capture enormous value. So will the people and companies operating in the gap. That gap is wider than Silicon Valley thinks, and it's going to take longer to close than anyone's current model assumes. The 18-month demand window everyone worries about is the wrong unit of time entirely.

---

Topics: enterprise AI, AI agents, future of work, SaaS, token budgets, AI security, frontier models, venture investing, digital transformation, knowledge work