Miklós Koren
Zsófia Bárány
Ulrich Wohak
2026
\[\max_{x} \; F(x) = \left(\sum_{i=1}^N \theta_i^{1/\sigma} x_i^{1-1/\sigma}\right)^{\sigma/(\sigma-1)}\]
subject to
\[ \left( \sum_{i=1}^N (x_i/s_i)^{1+1/\gamma} \right)^{\gamma/(\gamma+1)} \leq B \]
\(x\) = activities, \(\theta\) = job requirements (CES), \(s\) = skills (CET), \(B\) = time in a day.
\(\gamma\) = elasticity of transformation (PPF curvature)
\(\sigma\) = elasticity of substitution (isoquant curvature)
\[ x_i \propto s_i^{(1+\gamma)\sigma/(\gamma+\sigma)} \theta_i^{\gamma/(\gamma+\sigma)} \]
Output per unit of time \[ \Phi = \left[\sum_k (s_k^{\sigma-1} \theta_k)^{(\gamma+1)/(\gamma+\sigma)} \right]^{(\gamma+\sigma)/[(\gamma+1)(\sigma-1)]} \]
Time share on activity \(i\): \[ \omega_i^A = \frac{(s_i^{\sigma-1} \theta_i)^{(\gamma+1)/(\gamma+\sigma)}} { \sum_k (s_k^{\sigma-1} \theta_k)^{(\gamma+1)/(\gamma+\sigma)} } \]
Increases in skill \(s_i\) if \(\sigma > 1\). Relevant range for knowledge work.
\[ \xi_i = \left(\frac{\theta_i}{s_i^{1+\gamma}}\right)^{1/(\gamma+\sigma)}, \]
High for activities that are required (\(\theta_i\) high) but hard to do (\(s_i\) low).
AI can perform any combination of \(\{t_1, \dots, t_N\}\) tasks, but requires supervision.
\[ \max_{x, x_{\text{human}}, \mu} \; F(x) \] \[ x_i = x_{i,\text{human}} + \mu_i t_i \] \[ \sum_i \mu_i + g(x_{\text{human}})\leq B \]
\[ X_{AI} = \{x : \sum_i x_i/t_i \leq \mu B\} \]
\[ X_{\text{human}} = \{x : g(x) \leq (1-\mu)B\} \]
\[ X = X_{\text{human}} + X_{AI} = \{x : x = x_1 + x_2, x_1 \in X_{\text{human}}, x_2 \in X_{AI}\} \]
Whenever the value of AI tasks,
evaluated at the shadow prices of the human PPS,
exceeds your current output.
\[ t_i > \left( \frac{\gamma}{\gamma+1} \right)^{\gamma/(\gamma+1)} s_{i} \omega_{oi}^{-1/(\gamma+1)} \]
\[ \frac{t_i}{s_i} > \left( \frac{\gamma}{\gamma+1} \right)^{\gamma/(\gamma+1)} \omega_{oi}^{-1/(\gamma+1)} > 1 \qquad \text{if } \gamma < \infty \]
Order tasks by highest \(\omega_{oi}^{1/(\gamma+1)}t_i/s_i\).
There exist \(\chi_1 < \chi_2 < \dots < \chi_N\) such that AI is used for the first \(k\) tasks if and only if \[ \chi_{k+1} \geq \omega_{oi}^{1/(\gamma+1)}t_k/s_k \geq \chi_k \]
All results depend on convexity, not on CES–CET:
CES–CET gives closed forms; the geometry works for any \(F\) and \(g\).
Convexity is a reduced-form way to capture task complementarity within a fixed time budget.
Neoclassical theory explains slow adoption of tractors (\(\varepsilon_{\text{horse}, \text{tractor}} < \infty\)).
Human capabilities in strong bundles serve as bottlenecks to AI.
Automation reshapes the task mix of jobs (\(\sigma=1, \gamma=\infty\)).
Complementarity (low \(\sigma\)) matters for progress more than benchmarks (\(t_i\)).
Task success rates, human times and AI times for O*NET tasks, March 2026. Aggregate to 332 Intermediate Work Activities (IWAs).
\[ \tau_i = \frac{t_i}{s_i} = \frac{\text{human time}}{\text{AI time}/\text{success rate}} \]
Use task frequnencies \(\lambda_{oi}\) (daily, weekly, etc) and approximate \[ \omega_{oi} \approx \frac{\lambda_{oi}}{\sum_j \lambda_{oj}} \]


This research was funded by the European Union under the Horizon Europe grant 101061123 and by the National Research, Development and Innovation Office (Forefront Research Excellence Program contract number 144193). Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union, the European Commission, or the National Research, Development and Innovation Office.