Tasks At Work: Comparative Advantage, Technology and Labor Demand
Source: https://www.semanticscholar.org/paper/dc0b124086a988374d6aa71ff1fb07393f1f4bbb ↗
Full text: open-access via OpenAlex ↗
Acemoglu and Restrepo's task-based framework offers the most rigorous economic lens for understanding how AI reshapes work — not just which jobs disappear, but how comparative advantage shifts between humans and machines at the granular level of tasks.
The paper moves beyond automation anxiety to show how new technologies create new tasks even as they automate old ones, providing product directors with a more nuanced view of how AI tools change team dynamics and organizational structure.
This is essential reading for anyone building AI products: it explains why the 'AI will replace everything' narrative is wrong while showing exactly where human-machine collaboration creates value.
The framework helps product people think systematically about which capabilities to automate and which to augment, grounded in rigorous economic theory rather than speculation.
Central argument
Acemoglu, Kong, and Restrepo argue that the standard economic model—which treats all technology as uniformly 'augmenting' labor or capital—is too blunt to explain observed labor market trends. They propose a task-based framework where production requires completing discrete tasks assigned to workers or capital according to comparative advantage, and where four distinct types of technological change (labor-augmenting, capital-augmenting, automation, and new task creation) have fundamentally different effects on wages, labor share, and productivity. The central finding is that automation consistently reduces labor's share of income and can depress real wages, while technologies that create genuinely new tasks for workers raise wages and labor share—making the distinction between these two modes of technological change empirically and strategically decisive.
Critique
The framework's analytical power depends on cleanly categorizing technologies as either automating existing tasks or creating new ones, but in practice—especially with AI—a single technology often does both simultaneously and in ways that shift over time, making the taxonomy difficult to apply ex ante. The paper focuses heavily on aggregate labor market outcomes (wages, employment shares, factor shares) derived from historical data, which means it offers limited guidance on how firms or product leaders should anticipate which category an emerging technology will fall into before its effects are visible in the data. This retrospective orientation is a genuine limitation for decision-makers who need forward-looking frameworks.
Why it matters for product
For a CPO, the paper's core distinction—automation displaces and depresses value for workers, while new task creation raises it—directly maps onto how AI capabilities should be deployed within product teams: using AI to replace existing analyst, researcher, or designer tasks shrinks the value those roles generate, whereas using AI to enable genuinely new workflows (e.g., continuous experimentation at a scale previously impossible) expands it. The concept of 'ripple effects' through a propagation matrix is also relevant to org design: automating one role's tasks doesn't affect only that role but redistributes demand across adjacent functions in ways that compound, meaning decisions about AI tooling for one team (e.g., engineering) will structurally reshape demand for adjacent roles (e.g., QA, product ops) in ways that require deliberate anticipation rather than reactive headcount decisions.