← Wiki

The Expanding Work Frontier

ai-agentsfuture-of-workautomationenterprise-ai

The core thesis

The dominant narrative around AI and work focuses on replacement — which jobs will disappear, which tasks will be automated, how many workers will be displaced. Aaron Levie, CEO of Box, argues this framing misses the larger transformation: “In 5 years from now, probably 95% of the tokens used by AI agents will be used on tasks that humans never did before.”

This reframes AI as a frontier expander — opening categories of work that were previously impossible due to cost, scale, or complexity constraints.

The static-view fallacy

Most predictions about AI’s impact on employment take a static view of what work looks like. They inventory current tasks, estimate which ones AI can perform, and project displacement. Levie identifies this as a fundamental error: “Be skeptical of anyone predicting the future of jobs if they take a static view of work.”

Levie sharpens this in a 2026 post: “What gets missed with AI productivity gains is that by and large, most roles will continue to be as sophisticated as the tools allow. This is why also thinking through ‘today’s jobs will be replaced with AI’ is a fallacy. Everyone thinks the market is static, but it’s not.”

The static view misses three dynamics:

  1. New task creation. When engineers ship features faster with AI, the result is not fewer engineers but more features. When sales reps can build custom demo solutions using AI, they spend more time (not less) because they are performing tasks they never would have attempted before.

  2. Redefined output expectations. AI doesn’t just accelerate existing workflows — it resets what “good enough” looks like. Marketing campaigns become more personalized. Financial analysts model more scenarios. The work expands to fill the newly available capacity.

  3. Geographic and sectoral redistribution. Levie argues the tech industry is “only eight, ten, 15% of GDP in the economy” — and the static-view fallacy is largely a product of Silicon Valley’s self-referential perspective. When AI-enabled coding tools like Claude Code and Codex make software engineering accessible to non-tech industries, “85% of the economy now gets access to engineering tech has always had.” The graduating computer science student of 2030 may go to John Deere, Caterpillar, or Eli Lilly rather than Google — “automating pharmaceutical research” or building “AI for the future of farming and industrial equipment.” The total demand for engineering talent does not shrink; it diffuses across the entire economy.

Automation reveals the next bottleneck

Automating one step in a workflow doesn’t eliminate downstream steps — it exposes them as the new constraint. Levie illustrates this with healthcare referrals (administrative friction removed, but physician supply unchanged) and legal document generation (drafting automated, but court approval and attorney judgment unchanged). Each bottleneck removed reveals a deeper one that was previously invisible. This is one concrete mechanism for why work expands rather than contracts, alongside the Jevons paradox and arms race dynamics below.

For the full treatment of this pattern — including the cybersecurity case and implications for employment predictions — see The Bottleneck Shift.

Why work expands rather than contracts

Historically, for most knowledge work, organizations could not apply more compute to make work go faster. A financial analyst could only read so many reports. A lawyer could only review so many contracts. AI agents change this equation: “One of the big upsides of AI Agents for knowledge work is the ROI changes dramatically because agents can do work that would not have been worth paying a person to do.”

Levie documents concrete examples from Box’s own AI-first transformation:

  • Real estate: AI agents reading and analyzing every lease agreement for business opportunities — work no firm would assign humans to at scale
  • Life sciences: Rapid drug discovery and data quality checks across error-prone datasets
  • Financial services: Retroactive analysis of all past deals for monetization patterns
  • Legal: Executing contracts for previously unprofitable segments

Each example represents work that organizations wanted done but could never justify the human labor cost. AI agents make the marginal cost of this work approach zero, which opens an entirely new category of organizational output.

The 100x thought experiment

Levie proposes a useful frame for strategists: “Consider what things you’d do more of (or differently) if the cost and speed of labor became 100X cheaper and faster.”

This thought experiment reveals the expanding frontier. Most organizations are doing far less than they could because of the cost or limited capacity of talent. The constraint was never ambition or imagination — it was the economics of human attention at scale.

Andrew Wilkinson illustrates this with his claim of hiring “12 new employees” that “work 24/7, never complain, never take a sick day.” The framing is tongue-in-cheek, but the underlying dynamic is real: tasks that were never worth staffing now get done.

Every process was designed around human limitations

The implication is structural. Levie argues: “Every business process today was inherently created around the limitations of scaling labor. We couldn’t have more customer service reps than we did, more financial analysts reviewing deals, more engineers building software features, more supply chain specialists tracking inventory.”

When those constraints lift, the processes themselves must be redesigned. The opportunity is not to run existing processes with fewer humans but to design entirely new processes around the assumption of abundant, cheap cognitive labor.

In Frey’s framework, the enabling path for AI agents is making organizations capable of work they never attempted, not making humans faster at their current tasks. The replacing path — doing the same work with fewer people — captures only a fraction of the potential value.

Academic evidence: new work creation is the norm

Autor’s research provides the academic foundation for the expanding frontier thesis. In forthcoming work (Autor, Chin, Salomons, and Seegmiller), he finds that “the majority of contemporary jobs are not remnants of historical occupations that have so far escaped automation. Instead, they are new job specialties that are inextricably linked to specific technological innovations; they demand novel expertise that was unavailable or unimagined in earlier eras.”

There were no air traffic controllers before radar, no electricians before electrical systems, no gene editors before CRISPR. Innovation does not primarily automate existing tasks — it opens new vistas of work. In 1900, 35% of U.S. employment was in agriculture; by 2022, it was 1%. Those workers didn’t become permanently unemployed — they moved into jobs that didn’t exist when their grandparents farmed.

Autor reinforces this with demographic data: the industrialized world faces a worker shortage, not a job shortage. Plummeting birth rates and aging populations mean that, barring massive immigration changes, rich countries will run out of workers before they run out of work. AI-enabled expansion of what workers can do addresses this demographic constraint directly. See Expertise Democratization for how AI extends expert-level work to more people.

Field confirmation: enterprise leaders say expansion, not replacement

In April 2026, Levie reported on meetings with IT and AI leaders across banking, media, retail, healthcare, consulting, tech, and sports. The consistent finding: “Most companies are not talking about replacing jobs due to agents. The major use-cases for agents are things that the company wasn’t able to do before or couldn’t prioritize.” Examples included software upgrades, automating back-office constraints, and processing large document volumes for business insights. “More emphasis on ways to make money vs. cut costs.”

The field data also confirms an unintuitive outcome: “Unanimous sense that everyone is working more than ever before. AI is not causing anyone to do less work right now.” Far from reducing workloads, AI is generating new categories of demand that absorb the freed capacity and then some.

Levie reinforces Jensen Huang’s observation that scaring people out of engineering, radiology, or healthcare does “a disservice to the next generation.” The reason: “We don’t yet know any way to use AI in a capacity other than augmenting our work where we still eventually have to go and review the work in some form. We haven’t removed humans from the loop. We’ve just changed where they enter the loop.” The loop-entry point shifts upward — from executing tasks to reviewing outputs, orchestrating agents, and handling exceptions — but the loop itself persists.

The Jevons paradox for talent

Cybersecurity offers a concrete case of work expansion through what economists call the Jevons paradox: making a resource cheaper increases total consumption rather than decreasing it. Levie, commenting on Anthropic’s autonomous security research, argues: “AI is going to generate 100X more code, and along with that, there will be an enormous increase in security discoveries.” Autonomous tools surface more real vulnerabilities faster, but each finding requires triage, remediation, and architectural decisions demanding human judgment. “Better AI tooling for security will increase the demand for security talent, not decrease it.”

The pattern generalizes beyond security. Any domain where AI can surface latent demand — hidden contract risks, unreviewed compliance gaps, unmonitored patient data — will see work expansion rather than contraction.

The arms race dynamic

An attorney (@SMB_Attorney) articulates a complementary expansion mechanism: “If you think unprecedented access to legal tools will reduce your dependency on lawyers, you’re looking at it backwards. Guess who else has access? People who hate you and want to extract money from you.” When both sides of a transaction gain AI capabilities, the equilibrium point shifts upward rather than downward. The total volume of legal work increases because the cost of initiating disputes, conducting research, and building cases falls for everyone simultaneously.

The uneven distribution

The expanding frontier does not benefit everyone equally. Greg Isenberg frames the current moment as “a generational moment to start a company and steal market share from billion-dollar incumbents” — because small teams with AI agents can now do what previously required hundreds of employees.

The Bounty and Spread dynamic applies: total output grows (bounty), but the gains accrue disproportionately to those who can deploy AI agents effectively. The frontier expands, but access to it is unevenly distributed.