A task can disappear while a job changes shape.
AI's employment effects begin with tasks, not job titles. An accountant might spend less time classifying documents and more time investigating exceptions. A customer-service team might handle more cases with the same staff, or management might reduce staffing. Technical capability does not select between these outcomes by itself.
The ILO's 2025 occupational-exposure index explicitly examines the relationship between generative AI capabilities and tasks. Exposure should not be read as a count of jobs that will disappear.[1] Adoption costs, demand, regulation, reliability and organizational choices intervene between a capable tool and an employment outcome.
This distinction cuts both ways. It is misleading to announce mass displacement from a capability benchmark, but equally misleading to promise that every saved task will generate an equally good new role. The people who benefit from productivity gains may be different from the people who bear the adjustment costs.
Productivity is not an employment forecast
- Initial labor hours
- 100
- Resulting labor hours
- 84
100 × (1 − time saving) × (1 + demand growth)
Fixed wages, unchanged task quality and no new tasks are assumed. This toy arithmetic is not an employment forecast. Hours are a normalized example, not observed workers.
- Before time saving
- After time saving
Three channels can pull in different directions.
Substitution occurs when a system performs work previously done by people. Complementarity occurs when it makes a worker more capable or makes another human task more valuable. Demand can then change because lower costs, higher quality or new services alter what customers buy. Real organizations can experience all three at once.
There is also work around the system: integration, verification, incident response and dealing with people whose circumstances do not fit the standard pattern. It would be a mistake to assume this work is free. It would also be a mistake to assume it must always offset the labor saved elsewhere.
Risk management helps reveal hidden labor. NIST's framework treats AI as a system that must be evaluated in context, including reliability and impacts on people.[2] A tool that drafts quickly but creates expensive review work may shift effort rather than remove it. The relevant comparison is the complete workflow, including failures.
At the organization
Evaluation covers task changes, retained human judgment, training access, quality, review burdens and treatment of workers whose roles contract.
Across society
Evaluation covers the distribution of gains, access to new employment, income support and training. Aggregate productivity does not resolve those outcomes.
Plan for adjustment without pretending to know its exact size.
Transition planning can identify affected tasks, worker input, bounded trials and stopping criteria without assigning an invented unemployment probability. Evaluation includes total time, quality and error costs; a demonstration does not establish dependable operation.
For workers, transferable judgment and domain knowledge matter, but advice to “reskill” is incomplete when it ignores time, money, location and hiring practices. For employers, a training budget is only useful if people can attend and the resulting skills connect to real roles.
The central political question is who has a voice in the redesign. Employment is a source of income, bargaining power and social connection. Treating those consequences as external to an efficiency calculation makes the calculation narrower than the decision it is being used to justify.
Sources & further reading
Follow the evidence.
- 01ILO: Generative AI and Jobs: A Refined Global Index of Occupational Exposure
2025 task-exposure research; exposure is not a forecast of layoffs.
- 02NIST: AI Risk Management Framework 1.0
2023 voluntary framework; used here for workflow and reliability questions.