Two Ways AI Changes Your Job: Demand-Limited vs. Collapsable Tasks

The conversation about AI's impact on work tends to collapse all of the ways AI changes jobs into a single question: does AI take your job? The question is too coarse to be useful. AI affects different tasks in different ways, and understanding the mechanisms is what separates strategy from reaction.

Labor economists made this move two decades ago. David Autor, Frank Levy, and Richard Murnane's 2003 paper "The Skill Content of Recent Technological Change" replaced jobs with tasks as the unit of analysis, and most of what economics has usefully said about automation since has come from that substitution. Jobs are bundles of tasks. Technology hits the bundle unevenly, and the fate of the job depends on which strands it hits.

Two categories matter most for AI: demand-limited tasks and collapsable tasks. They require different responses, and confusing them is a reliable path to the wrong investment.

Demand-Limited Tasks

A demand-limited task is one where the constraint on value creation is demand, not execution capacity.

Consider a legal team that drafts contracts. Each contract takes four hours. AI could compress this to forty-five minutes, a significant productivity gain. But if the legal team is already handling the full volume of contracts the company generates, and that volume is fixed by business activity upstream, what does the freed time produce?

The answer depends on whether the freed time can be redirected to higher-value work. If it can, with the lawyers spending the freed hours on more complex legal strategy, client counseling, or the proactive risk identification that previously got squeezed out, then the AI investment produces real value. The role evolves; the demand is constant.

If the freed time cannot be productively redirected, because the work that was crowded out by contract drafting never actually needed doing, the AI investment produces surplus capacity. The organization has more legal capacity than it needs. The economic outcome is either workforce reduction or cost that does not produce proportional value.

Demand-limited tasks are common in roles where the work is well-defined, high-volume, and tied to a business process whose throughput is constrained upstream. The AI makes the execution faster; the demand does not expand to fill the new capacity.

The diagnostic question: If we could do this task in half the time, what would we do with the other half? If you have a clear and valuable answer, the constraint is execution. If the answer is "we'd probably do it less often" or "I'm not sure," the constraint is demand.

Collapsable Tasks

A collapsable task is one that AI can absorb entirely. The human step disappears.

Routine report generation is a collapsable task in many organizations. The report was compiled by a person because that was the only way to get it done. When AI can compile the same report with equal or better quality, the human step collapses. The question shifts from "what does the person do with the freed time?" to "what does this person do now that report compilation has left their job?"

This is categorically different from the demand-limited scenario. Collapsable tasks require a conversation about role redefinition, a level above productivity improvement.

The clearest examples of collapsable tasks:

• Data transcription and normalization: moving data between systems, normalizing formats, cleaning datasets. The human involvement was never about judgment; it was about execution. When AI can execute, the step collapses.
• Routine summarization: daily digests, status summaries, meeting notes. These consumed human hours because humans were the only available engine to generate them, never because they required human judgment.
• First-draft generation at scale: outreach emails, documentation stubs, test case templates. At low volume, a human does this. At scale, a human doing this is expensive relative to the alternatives.

The collapse does not make the person unnecessary. It changes the definition of what they are for. The useful question is what the role is for now that AI has taken this task.

Where the Argument Could Break

Two challenges to the dichotomy deserve attention.

The first is that demand is rarely as fixed as the demand-limited label implies. William Stanley Jevons observed in 1865 that more efficient steam engines increased England's coal consumption, because cheaper power found new uses. The same rebound can hit contracts: when a contract costs forty-five minutes, the business may start papering deals it previously handled with a handshake, and the fixed volume turns out to have been a function of the old price. This is less a refutation than a refinement. The diagnostic question has to be asked at the new cost and re-asked as the cost keeps falling. A task that is demand-limited at four hours per unit may be execution-limited at forty-five minutes.

The second is that collapsable tasks have a habit of leaving a residue. David Autor's work on automation keeps returning to this point: tasks complement each other, and removing the routine ones raises the stakes on the judgment that remains. A report no human compiles is a report no human reads closely, until the quarter it matters. The answer is operational. Collapse the task, keep a quality gate, and write the residual judgment into the redefined role.

The Strategic Response to Each

For demand-limited tasks: plan for role evolution. The freed capacity should go somewhere valuable. Define where before you deploy the AI. If you cannot identify a clear, high-value destination for the freed capacity, either the role is not expanding in the right direction or the demand-limited diagnosis is incorrect.

For collapsable tasks: plan for role redefinition. What was this person's role for, beyond the specific tasks that collapse? What remains that requires human judgment, relationship, or expertise? Is that remainder enough to justify the role? If yes, redefine the role around the remaining work. If no, have the honest conversation about capacity.

The organizations that navigate this transition well do the analysis proactively, before the AI is deployed, while there is still time to redirect investment in people development toward the skills the new role requires.