The future of work with AI is a task redesign problem
Jobs are bundles of tasks, relationships, judgment, and accountability. Current evidence points to uneven transformation—not one universal automation story.
AI is more likely to transform task bundles unevenly than replace every exposed occupation at once. Current evidence shows substantial exposure, mixed measured productivity, and differences by occupation, country, experience, and deployment design. Leaders should map tasks, run measured pilots, protect learning and entry-level pathways, and keep accountability with people.
Jobs are not single prompts
An occupation combines information processing, physical work, relationships, tacit knowledge, coordination, legal responsibility, and judgment. A model may draft the document while a person gathers missing facts, negotiates trade-offs, takes responsibility, and acts. Exposure at the task level therefore does not translate directly into a count of jobs removed.
The ILO’s 2025 global index estimates that one in four workers is in an occupation with some generative-AI exposure, while 3.3% of global employment falls in its highest exposure category. Its central conclusion is that transformation is the more likely effect because most occupations contain tasks that still require human input.[1]
Exposure and benefit are uneven
Clerical occupations remain highly exposed, and the ILO reports growing exposure in some professional and technical roles. Exposure also varies by income level and gender because occupational structures differ. A later ILO–World Bank study across 135 countries warned that digital gaps can allow disruption to arrive before productivity benefits in developing economies.[2]
Organizations should therefore avoid one universal adoption playbook. Infrastructure, language coverage, data access, worker voice, and the composition of tasks determine whether a tool removes drudgery, shifts work to reviewers, or makes a service less accessible.
Productivity evidence is real, mixed, and easy to overstate
Controlled studies have found gains in some bounded writing, support, and coding settings, while other realistic work shows friction. METR’s early-2025 randomized study found experienced open-source developers took 19% longer with the tested AI tools on familiar repositories. The authors explicitly said the result did not establish that AI slows most developers; it measured one important setting.[3]
The ILO’s 2026 evidence review concluded that productivity gains are real but often unverified and uneven, and that reported time savings have not consistently translated into measured output, earnings, or employment. This gap can arise when review, coordination, rework, and organizational bottlenecks absorb the local speedup.[4]
Redesign the task before adding the agent
Map the current workflow: inputs, decisions, handoffs, approvals, failure modes, and outcome measures. Decide whether AI should retrieve, draft, critique, classify, simulate, or act. Then redesign the handoff so people receive evidence and exceptions rather than an opaque completed answer.
The strongest candidates have good digital inputs, frequent repetition, and a clear quality check. High-stakes tasks may still benefit from research or drafting, but authority stays with a qualified person. Automating a broken process can increase the volume of bad work.
The durable skills move up and sideways
People need domain judgment to frame the task, source literacy to inspect evidence, statistical literacy to read evaluations, and operational skill to manage tools and exceptions. Clear writing remains important because specifications, rubrics, and review notes are how humans coordinate with systems and each other.
Learning pathways also need protection. Junior tasks often provide the repetitions through which experts develop judgment. If AI absorbs all first drafts, organizations must create deliberate practice: review model errors, rotate ownership, preserve supervised work, and evaluate understanding rather than output volume alone.
Management changes from allocation to verification design
Managers will increasingly decide which work can be delegated, what context an agent may access, when a person must approve, and how evidence is logged. That is not simply ‘managing digital employees.’ It is designing a socio-technical system where responsibility remains legible.
Anthropic’s Economic Index studies distinguish augmentation from automation and show that usage patterns change over time and by product. Its 2026 survey findings are informative but not population-representative; the respondents are Claude users and heavily overrepresent computer and management occupations. Treat such telemetry as one lens, not a global labor forecast.[5]
Use a workforce evidence scorecard
Measure cycle time, accepted quality, worker effort, error severity, customer outcomes, learning, and distribution of gains. Break results down by role and experience. A pilot that saves senior time by transferring invisible cleanup to junior workers is not a clean productivity win.
Publish what changed and invite worker feedback. Provide an escalation path when the system is wrong or a task should not be automated. Revisit access, performance, and job design as tools improve.
| Question | Evidence to collect | Decision |
|---|---|---|
| Does it improve the task? | Time, quality, corrections, failures | Continue, revise, or stop |
| Who gains or loses? | Results by role and experience | Training and workflow changes |
| Is judgment preserved? | Approval and override logs | Change authority boundary |
| Does learning continue? | Skill checks and supervised practice | Protect development pathways |
A reasonable near-term forecast
More knowledge work will begin with machine-generated material and end with human acceptance. Small teams will attempt projects that previously required more coordination. Some roles will shrink, others will grow, and many will change internally faster than job titles reflect. The pace will differ sharply across industries and countries.
The responsible response is neither denial nor a countdown to universal replacement. Build measurement, training, worker participation, and human accountability into adoption now. Those institutions will matter whether capability grows steadily or arrives in bursts.
Frequently asked questions
Will AI replace most jobs?
Current evidence does not support a precise universal forecast. It shows broad but uneven task exposure, with transformation more likely than complete automation for many occupations.
Which jobs are most exposed to generative AI?
The ILO finds clerical work highly exposed and growing exposure in some professional and technical roles, with meaningful variation by country and worker demographics.
What should companies do now?
Map tasks and authority, pilot measurable workflows, involve workers, preserve entry-level learning, monitor distributional effects, and keep people accountable for high-stakes outcomes.
Sources and evidence
Primary and authoritative sources used for factual claims. Company research and executive forecasts are labeled as such in the article.
- 1Generative AI and Jobs: A Refined Global IndexInternational Labour Organization · 2025-05-20
- 2Uneven global impact of generative AI on jobsILO and World Bank · 2026-03-27
- 3Experienced open-source developer productivity RCTMETR · 2025-07-10
- 4GenAI jobs, productivity and work organization evidence reviewInternational Labour Organization · 2026-06-01
- 5Anthropic Economic Index: CadencesAnthropic · 2026-06-26