Three AI economies worth watching
Jobs research gives us clues, not a finished map. These scenarios separate what has been observed from what we can only imagine.
An illustrative scenario. This story explores a possible arrangement. It is not a report of an existing workplace or a forecast.
The future of work is often compressed into a choice between abundance and unemployment. The evidence is more specific than either slogan. A task can become easier without an occupation disappearing. A company can benefit while a particular group loses. We can examine those possibilities without pretending that one result has already settled the whole economy.
The ILO’s 2025 index estimates some generative-AI exposure in occupations covering roughly one-quarter of global employment. Exposure describes potential task change, not observed job losses.
One published customer-support study found higher issues resolved per hour with AI assistance. Its particular workplace does not settle economy-wide employment or wages.
Other evidence cautions against assuming everyone gains. Acemoglu and Restrepo estimated negative employment and wage effects in American local labor markets more exposed to industrial robots. That historical setting is not current generative AI. Stanford's 2026 AI Index also assembles evidence of uneven labor-market effects, alongside adoption and productivity measures. A collection of indicators is not one experiment demonstrating a single cause.
Scenario one: better services. Imagine organizations using assistance to reduce routine delays while retaining staff for difficult cases and human contact. Customers could receive more attention rather than merely more automated replies. This is a possible use of additional capacity, not a result established by the studies above. Its signature would be improved service quality that survives checking.
Scenario two: concentrated advantage. Imagine the strongest firms capturing most of the gains while entry-level opportunities narrow and bargaining power shifts. This is a risk to investigate, not a prediction about a particular employer. Relevant evidence would include hiring, wages and who receives training, rather than a demonstration of what a model can do.
Scenario three: a busy plateau. Imagine software spending and output volumes rising while correction, supervision and low-value production absorb much of the benefit. Faster generation would then coexist with disappointing useful results. This too is a scenario. It asks whether the unit being counted is an accepted service or simply another draft.
These futures could overlap across sectors and households. The point of comparing them is to make our expectations answerable to evidence. Watch the quality of services, the distribution of opportunity and the work required to check an answer. A better question than who wins the entire AI revolution is which choices are improving whose life, under conditions we can actually observe.
Sources & further reading
Original editorial analysis informed by the sources below. A source’s date, method and limits matter.