Acemoglu's (2025) accounting with public data for eleven countries: TFP = exposure × profitability × cost saving × labor share, GDP = TFP × a capital multiplier. Move the inputs and the numbers recompute. Part (b) adds Baumol's structural change: the same AI wave evaluated with the sectoral weights that observed productivity trends produce in 10, 20 or 30 years.
Svanberg et al. (2024), Fig. 6b: the share of exposed tasks worth automating within ten years as a function of labor cost relative to the AI system (US = 1). Each country's m is the exposed-labor cost relative to the US divided by its AI-system cost relative to the US.
Each factor relative to the United States (US = 100). GDP = e × q × a; the labor share scales TFP and cancels in GDP when capital keeps a constant capital-output ratio.
Part (a), the accounting. Acemoglu (2025, Economic Policy), equations 14 and 15: TFP gain over ten years = GDP share of the wage bill in exposed tasks (e, from each country's employment by occupation in ILOSTAT and the task exposure of Eloundou et al. 2024) × share of those tasks worth automating (q, read on the Svanberg et al. 2024 curve at each country's labor-cost ratio m) × cost saving (a = 27%, or 7% on hard-to-learn tasks) × labor income share (sL, ILOSTAT). GDP = TFP / sL when capital keeps a constant capital-output ratio; as a robustness check, with a CES technology between capital and labor of elasticity σKL = 0.7 (Oberfield and Raval 2021) the local multiplier is 1 + σKL sK/sL: with complements, capital responds less to a TFP gain. The numbers are level effects cumulated over ten years, not annual growth rates; divide by ten for the growth-rate equivalent.
Part (b), structural change. A Huneeus-Rogerson (2024) model with the nonhomothetic CES of Comin, Lashkari and Mestieri (2021) across eleven sectors (σ = 0.35) and Cobb-Douglas production in capital and labor (CES as robustness), run on each country's observed sectoral productivity trends (CEPALSTAT value added at constant prices, ILOSTAT employment, 2013-2025). The drift factor is the AI gain recomputed with the sectoral GDP weights of 2032, 2042 or 2052, relative to 2022 weights. It is available for seven countries; the others keep the 2022 weights.
Comparisons. Aghion and Bunel (2024) read the same formula with q = 50%, a = 40% and e = 60% (0.68 pp of TFP a year in the US). The IT revolution added about 1.5 pp a year to US output per hour in 1995-2004 (Byrne, Oliner and Sichel 2013) and IT investment about 0.3 pp a year to Latin America's GDP growth (Jorgenson and Vu 2007).
Input-output links. Following Filippucci, Gal and Schief (2024) with OECD TiVA 2025, the bound splits into a direct effect (sectoral gains in gross-output terms, value-added weights) and an input-output multiplier (Domar weights). The two add up to the bound, so the last table column shows how much of it arrives inside the adopting sectors; the rest travels through cheaper inputs. TiVA covers Argentina, Brazil, Chile, Colombia and Peru.
Caveats. Exposure uses US task scores applied to each country's occupational mix at 1 or 2 digits of ISCO; earnings are monthly, not hourly; employer surcharges are statutory; the labor share imputes self-employment income. Everything here is an upper bound on the ten-year effect through existing tasks: new tasks and products, general-equilibrium price responses and effects beyond ten years are outside it.
Federico Huneeus, Duke University. Data: ILOSTAT, CEPALSTAT, Eloundou et al. (2024), Svanberg et al. (2024), Acemoglu (2025). Code and methodological note available on request. The selected country's sliders start at its data; "Reset to data" restores them.