AI + Robot Job Impact

Appendix · Data and rules

Every number comes from the file outputs/occupations.json, written by the module aimodel/analysis/occupations.py of the book's software. The book table and this page read the same file.

Data

• AI exposure: share of the occupation's tasks whose time a GPT-type system with software tools can cut by at least half, rated by human experts. Eloundou, Manning, Mishkin and Rock, "GPTs are GPTs", Science 384 (2024); occupation file occ_level.csv, github.com/openai/GPTs-are-GPTs.
• Classic office automation: for office clerks, accountants and customer service the exposure is the higher of that value and the probability of computerisation of Frey and Osborne (2013): 0.96, 0.94 and 0.55. Never above 95%.
• Robots: probability of computerisation of Frey and Osborne for manual and physical jobs. Robots reach the central adoption year five years after AI.
• Demand for the occupation: BLS Employment Projections 2025–35 (published 21 September 2026), percent change in employment 2025–2035, extended at the same yearly pace to 2046 and used for every region, since no comparable projections by occupation exist for the others. BLS attributes to AI and automation only the decline of office and administrative support jobs: for office clerks and customer service demand is held flat, so that automation is not counted twice.
• Observed decline: Stanford Digital Economy Lab / ADP monthly series from 2022, each monthly change weighted by its year: 2026 and 2025 100%, 2024 50%, 2023 5%, 2022 1% (2022 and 2023 reflect post-pandemic over-hiring and higher interest rates). Measured against peers of the same age in the least exposed jobs and averaged over the six age groups: customer service −1.9% a year; software developers +0.4%, because the decline concerns only those under 30 (22–25 −2.4%, 26–30 −1.3%) while older developers grow.

Rules

• Type of work, following the IMF distinction between exposure and complementarity (Pizzinelli et al., WP/23/216; Cazzaniga et al., SDN/2024/001). Routine office and content work: 85% of the exposed work can be substituted and 85% of it is automated by 2046. Design with AI tools (the architect): 70% and 70%. Work with responsibility and human contact (doctor, nurse, teacher, lawyer): 55% and 47.5%, the book's central values. Manual work: 55% and 70% for the part done by AI; for robots the Frey and Osborne probability is already the automatable share, and 70% of it is automated by 2046. Software developers and customer service are counted as routine work because the observed data already show a decline among the young.
• Adoption: the book's S-curve (appendix A1), slope 0.425 a year, starting from zero in 2026. For customer service the central year is set so that the first simulated year, demand included, reproduces the weighted observed decline; office clerks use the customer service series; software developers, with no overall decline, and the other occupations use 2033. Emerging economies adopt 3 years later, low-income countries 6.5 years later.
• New roles: 0.15 new roles inside the occupation for each task lost, within 3 years. New work in other occupations is counted in the book's economy model, not here.
• Result: jobs in the occupation compared with 2026 = demand index × (1 − work taken by AI and robots + new roles). Where demand grows, as for nurses, cooks and electricians, the work taken by machines turns largely into more service with the same people, not into fewer jobs.

Geography

The book uses three groups: advanced, emerging and low-income economies. Each country or continent here takes the timing of its group: USA, Canada, Japan, Europe and Oceania as advanced; China, India, Russia, the rest of Asia and the Americas as emerging; Africa as low-income.