← Wiki

AI Agents and Job Redefinition

ai-agentsfuture-of-worklabour-marketai-and-jobs

The Prediction Problem

Predicting AI’s impact on employment requires knowing what jobs will look like after AI changes them — which is precisely what cannot be known in advance. Aaron Levie identifies the core difficulty: “The issue that we’re going to have in predicting the future of jobs is that those predicting things don’t have any way of measuring (in advance) the fact that job requirements will just change over time because of AI agents.”

This is not a claim that all jobs are safe. It is a claim that static analysis of job displacement is structurally unreliable because it treats current job descriptions as fixed while only varying the technology.

Jobs as Dynamic Task Bundles

The framing “jobs are not tasks — they are collections of tasks, and AI just automates some of them” captures a crucial distinction. A job is a bundle of tasks, responsibilities, and contextual judgments assembled around organizational needs. AI may automate individual tasks within that bundle, but the bundle itself reconfigures.

Mollick confirms this with empirical precision: research teams analyzing 1,016 professions found that almost all overlap with AI capabilities. Only 36 job categories (dancers, athletes, pile driver operators, roofers) had no overlap — all highly physical jobs requiring embodied spatial movement. The overlap is heaviest at the top: the most highly compensated, most creative, and most educated jobs overlap most with AI. This is the opposite of every previous automation wave, which started with repetitive and dangerous work.

Levie documents this reconfiguration at Box:

RoleBefore AIAfter AINet effect
Software engineerShips features at baseline paceShips features faster with AIMore features, not fewer engineers
Sales repPitches with standard decksBuilds full custom demo solutions with AIMore time per deal, doing work they never did before
Sales engineerSupports demos on requestCreates bespoke technical environmentsNew task category entirely

The sales example is particularly revealing. Sales reps using AI spend more time than before — because they are doing work that was previously impossible. “Ironically this takes them more time than before, because they’re doing a task they never would have done before.”

Skilled Jobs Are Not Going Away

Levie argues directly against the headline-grabbing displacement narrative: “Despite what you hear, skilled jobs are very much not going away because of AI. But the things you do will change a lot.”

The mechanism: AI takes “previously scarce resources and skills and makes them abundant.” When a skill becomes abundant, it loses premium value — but the person who had that skill doesn’t become useless. Instead, their job shifts toward the non-automatable remainder: judgment, context, relationship management, exception handling, and the new tasks that AI makes possible.

The pattern fits Judgement vs Knowledge in the AI Era: knowledge becomes cheap while judgment becomes precious. And Autor’s historical analysis provides the academic foundation: this is the fourth major expertise transition in economic history, from artisanal to mass to elite to AI-democratized expertise (see Expertise Democratization).

The Invisible Unemployment Counter-Argument

Not all views are optimistic. Jason Lemkin warns of “Invisible Unemployment” in tech: “It’s already happening. Lots of very talented folks can’t find roles at the level, scope, and comp they had at their last gig.”

Lemkin’s observation does not contradict the redefinition thesis — it describes the transition cost. When job requirements change, people whose skills matched the old requirements face a gap. The jobs didn’t disappear; the qualifications shifted. The CRO of 2027, Lemkin suggests, may be “the last human standing in GTM” — not because the function disappears but because the role absorbs so much AI-driven capability that far fewer people are needed to fill it.

The Enterprise Adoption Gap

Saanya Ojha contextualizes the MIT finding that 95% of enterprise GenAI pilots fail: “The 95% failure rate isn’t a caution against AI. It’s a mirror held up to how deeply ossified enterprises are.” The redefinition of work is happening, but as Enterprise AI Adoption Lag documents, organizations that can’t restructure around it fail at deployment — not because the technology doesn’t work.

The Dynamic View

The synthesis: AI agents trigger a continuous cycle of job redefinition.

  1. AI automates certain tasks within a job
  2. The human’s remaining tasks shift toward judgment, context, and exception handling
  3. AI simultaneously enables new tasks that were previously impossible (see The Expanding Work Frontier)
  4. These new tasks get absorbed into the job description
  5. The job now has different requirements than before
  6. Repeat as AI capabilities advance

Anyone predicting employment outcomes from a snapshot of current job descriptions will systematically overestimate displacement and underestimate redefinition. Autor reinforces this with demographic data: facing decades of stagnating population growth, the industrialized world faces a shortfall of workers, not a shortfall of jobs. The majority of contemporary jobs are not remnants of historical occupations that escaped automation — they are new specialties inextricably linked to specific technological innovations, demanding expertise that was unavailable or unimagined in earlier eras.

The Jagged Complication

The BCG consultant study (documented in The Jagged Frontier) adds a critical nuance: redefinition is not uniformly positive. Inside AI’s capability boundary, consultants gained 40% in quality. Outside it, they performed worse than those without AI. Job redefinition must account for this jaggedness — the new task bundles include tasks where AI helps and tasks where AI harms, and workers must develop the judgment to distinguish between them. Mollick’s Centaur and Cyborg Work framework provides the practical models for navigating this.