Will AI Replace Jobs or Redesign Them? A 2035 Perspective

The question "will AI take my job?" is intuitive and almost impossible to answer, because a job is not a unit of work. It is a bundle of tasks, some of which are highly exposed to automation and some of which are not. Radiologists read images, but they also consult with clinicians, weigh ambiguous findings against patient history, supervise trainees and take responsibility for decisions. Automating one task in that bundle changes the job; it does not delete it.

This article takes the task-level view seriously and asks what it implies for 2035. The short answer is that redesign is a more probable dominant outcome than replacement — but that "redesign" can still be painful, unevenly distributed, and slow to show up in aggregate statistics.

What history actually tells us

The historical record supports a specific and limited claim: technological change has repeatedly transformed the composition of employment without producing sustained mass unemployment. Agricultural employment in advanced economies fell dramatically over the twentieth century; office and service employment grew. Bank teller numbers did not collapse immediately after cash machines arrived, because branches shifted toward sales and advice.

Three cautions apply before treating that as reassurance. First, the aggregate can hide severe local damage — specific towns, specific cohorts, specific occupations absorbed permanent losses. Second, the transitions took decades, longer than an individual career can comfortably absorb. Third, past general-purpose technologies mostly automated physical or narrowly routine tasks. AI extends automation into cognitive and, to a degree, non-routine work. That difference is real and is why the historical analogy should inform judgement rather than settle it.

Which tasks are exposed

Exposure is highest where the following overlap: the task is largely informational, the output can be checked, tolerance for occasional error is moderate, and abundant examples exist.

• Drafting standard documents, summaries and correspondence.
• First-pass code, tests and documentation.
• Structured data extraction and reconciliation.
• Translation and transcription.
• Classification, tagging and routing of requests.
• Routine research aggregation and initial literature scans.

Exposure is lowest where tasks involve physical dexterity in unstructured settings, legal or professional accountability, negotiation, care and trust, tacit institutional knowledge, or responsibility for irreversible decisions. Note that the barrier there is often not cognitive difficulty but accountability: someone must be answerable, and current systems cannot be.

Automation and augmentation are not alternatives

Public debate treats these as competing predictions. In practice they are usually simultaneous. When a task inside a job is automated, three things can follow: the freed time is redeployed to higher-value work; the job is redefined around supervising and correcting the automated output; or headcount falls because the remaining bundle no longer justifies the same number of people.

Which of the three happens is determined mostly by demand elasticity. If cheaper output means more output is wanted — more analysis, more software, more customer contact, more design iterations — employment holds or grows while output per worker rises. If demand is fixed, cost savings convert into fewer roles. This is why the same technology reduces employment in one sector and expands it in another.

The measurement problem

Studies of AI assistance in specific occupations have generally found meaningful productivity gains on defined tasks, often with the largest relative gains for less experienced workers. Two cautions matter. Controlled task studies do not capture organisational effects — coordination, rework, quality assurance, error correction downstream. And economy-wide productivity statistics have so far shown nothing resembling a step change. The OECD's employment and AI work documents adoption that is broad in awareness and narrow in deep integration.

Current evidence therefore supports "significant task-level gains, unclear aggregate effect". Anyone claiming to know the aggregate number for 2035 is extrapolating.

Where displacement is genuinely likely

Honest analysis has to name the cases where the optimistic story is weak.

• Narrow-portfolio roles. Where a job consists almost entirely of one exposed task — basic transcription, simple content production, first-line scripted support, routine data entry — the bundle offers little to redesign around.
• Entry-level pathways. If AI absorbs the junior tasks through which people historically learned a profession, the pipeline for developing senior expertise weakens. This may be the most underrated risk in the whole debate, and it is a structural one: it damages capability years later, invisibly.
• Outsourced service work. Locations whose comparative advantage was cost-efficient routine cognitive work face direct competition from systems that are cheaper still.
• Bargaining power. Even where jobs persist, automation of the most easily specified parts can shift the balance of power over pay and conditions.

What new work appears

New occupations tend to be invisible in advance because they emerge from the operational needs of a new technology, not from imagination about it. Categories already forming include evaluation and testing of AI systems, data curation and licensing, AI oversight and assurance roles, integration and workflow engineering, incident investigation, and specialist roles at the boundary between domain expertise and model behaviour.

There is also a quieter pattern worth noting: technology often increases the value of skills it cannot replicate. As routine output becomes abundant and cheap, the premium shifts to judgement, taste, credibility, relationship depth and accountability — the things that determine which output is worth acting on.

Skills and education

If the half-life of a specific technical skill shortens, the returns to durable capabilities rise:

• Problem framing. Deciding what question to ask is now a larger share of the value than producing the answer.
• Verification. The ability to tell a plausible-sounding wrong answer from a correct one is the core competence of working alongside generative systems.
• Domain depth. Enough real expertise to detect errors that sound reasonable.
• Systems literacy. Understanding how a tool fits into a process, where its output goes, and what breaks downstream.
• Communication and trust. Persuading, negotiating and taking responsibility remain human functions in almost every institutional setting.

For education systems the implication is uncomfortable but simple: assessment methods designed to measure the production of text are now measuring something that is cheap. Institutions will need to assess reasoning, verification and defence of conclusions instead. That is a curriculum and workload problem, not a technology problem, which is why it will take years.

Three scenarios for 2035

Scenario A — Redesign dominates (most plausible on current evidence). Adoption is broad but shallow in most sectors. Most occupations change task composition; a minority contract sharply; new roles appear around oversight, integration and data. Aggregate productivity improves modestly and unevenly. The main policy problem is transition support, not unemployment.

Scenario B — Rapid substitution in cognitive services. Reliability improves faster than expected and agentic systems become genuinely dependable across long tasks. Some professional service functions see substantial headcount reduction within a few years, concentrated in junior grades. Aggregate employment holds up through growth elsewhere, but distributional stress and political pressure rise sharply.

Scenario C — Stall. Reliability plateaus, integration proves harder than expected, liability concerns bite, and adoption stays confined to assistive uses. Employment effects are modest; the disappointment is economic rather than social.

Elements of all three will probably coexist across different sectors and countries. That is normal for general-purpose technology and it is why national averages will be misleading.

Indicators to watch

• Job postings by task content, especially the volume of junior roles in exposed professions.
• Wage differentials between roles with and without verification responsibility.
• Sector-level productivity data, not company anecdotes.
• Whether firms report headcount reduction or output expansion after deployment.
• Training investment per employee, which reveals whether firms plan to redesign or replace.

Conclusion

The framing that matters is not human versus machine but which configuration of humans and machines produces reliable, accountable work. On current evidence, the most likely 2035 is one where most jobs have been rearranged rather than removed, where a real minority of roles has been hollowed out, and where the scarce skills are judgement, verification and responsibility.

The risk deserving most attention is not a sudden employment collapse. It is the quiet erosion of the entry-level pathways through which expertise has always been built — a problem that becomes visible only once the people who would have been experts are missing. Thinking clearly about that kind of delayed, second-order effect is precisely what futures thinking is for.