Chaining Tasks, Redefining Work: A Theory of AI Automation

Chaining Tasks: How AI Automation Redefines Job Structure and Worker Specialization

Economics has long studied how machines replace human labor. Most models focus on whether a tool is a substitute for a worker or a complement to their skills. Current models typically treat automation as a series of independent events. An algorithm might assist a professional with a specific calculation. However, this ignores the reality that production is a sequence of interconnected steps. Instead of just helping humans with single tasks, AI can perform multiple steps in a row without human help. This creates "chains." The authors argue that companies will reorganize entire jobs to group these AI steps together. This fundamentally changes the skills workers need and how much they are paid.

The Logic of the AI Chain

The core insight of the paper is that AI does not just automate isolated tasks. It enables "chaining." In a typical production workflow, a job is composed of a sequence of steps. Traditionally, a firm might assign each step to a different specialist to maximize efficiency. But when AI enters the picture, a new possibility emerges. The AI can execute a contiguous block of steps autonomously. It passes its output directly to the next step without human intervention.

This is what the authors call an "AI chain." Within such a chain, the intermediate steps are "automated" (completed by AI end-to-end without human review). Only the final step in the chain is "augmented" (completed by AI but requiring human oversight). This distinction is vital. If a human must inspect every single step, the speed and cost advantages of AI vanish. True productivity leaps occur when the AI handles a whole sequence. The human then acts as a high-level validator rather than a granular supervisor.

The Structural Scaffold of Production

To understand why this matters, one must look at the classical "task-based model" of economics. In these traditional frameworks, production is viewed as a collection of independent tasks. What matters is "comparative advantage" (the idea that a resource should be assigned to a task if it can perform it at a lower relative cost).

The authors, Demirer et al., argue that this logic fails when AI chaining is possible. They model production as a Leontief technology (a system where output depends on the completion of a fixed set of required inputs) over an ordered sequence of steps. They introduce a hierarchy of organization: 1. Steps: The most primitive units of work. 2. Tasks: Contiguous blocks of steps designated for joint execution. 3. Jobs: Groups of tasks assigned to a single worker.

As shown in, a firm must decide how to bundle these steps into tasks and how to partition those tasks into jobs.

This creates a fundamental tension. On one hand, the firm wants to specialize workers to reduce the time spent on individual tasks. On the other hand, specialization increases "hand-off costs" (the time and coordination lost whenever work passes from one worker to another).

The authors demonstrate that as AI improves, firms don't just use it to do things faster. They redesign the architecture of the job itself. They might consolidate several steps into a single AI chain. This allows them to hire a lower-skilled worker to oversee the entire chain. This reduces the total wage bill.

Breaking the Comparative Advantage Rule

The paper’s theoretical strength lies in its demonstration that AI chaining can overturn standard economic intuition. Usually, if a human is slightly better at Step A and an AI is slightly better at Step B, you would split the work. But if Step A and Step B are adjacent, it might be cheaper to let the AI do both as a single chain.

Why? Because appending a step to an existing AI chain adds almost no additional human "verification burden." Verifying the output of a chain is essentially a fixed cost. If you break the chain to let a human perform Step A, you introduce a new "human checkpoint." This incurs a fresh cost of time and oversight. As the authors show in, the decision to automate a step depends heavily on its neighbors.

If a step is surrounded by other AI-executed steps, the pressure to pull it "under the hood" into an automated chain becomes overwhelming.

This leads to a phenomenon the authors call "fragmentation." If the AI-suitable steps in a job are scattered widely apart, they cannot form long, efficient chains. They remain isolated, augmented steps that require constant human oversight. Conversely, if the AI-ready steps are clustered together, the firm can build long, seamless chains. This leads to massive productivity gains.

Non-linear Gains and the Productivity J-Curve

One of the most striking implications of this model is that productivity gains from AI will not be a smooth, linear climb. Instead, they will likely follow a "J-curve."

Because firms optimize their entire organizational structure, they won't change their behavior for every tiny improvement in AI. They will wait until the AI reaches a specific quality threshold. This threshold makes a new, radical reorganization viable. For example, an AI might be good enough to assist with a report. However, it might not be good enough to handle the data collection, the analysis, and the report as a single chain. Once the AI crosses the threshold to handle all three, the firm will abruptly shift its design.

This transition is visualized in . It shows how the marginal benefit of improving AI quality stays relatively flat initially. However, once a reorganization threshold is hit, the cost of production drops sharply. This explains why we might see periods of stagnant productivity despite rapid advances in AI. The technology is improving, but the organizational "leap" hasn't happened yet.

Measuring the Real World

The authors test their predictions using a massive dataset. They combine O*NET task data, Anthropic’s Economic Index, and GPT-generated workflow orderings.

Their empirical findings strongly support the chaining hypothesis: 1. Clustering: AI-executed steps do not appear randomly. They appear in contiguous blocks. This confirms that AI tends to operate in chains . The average AI chain length was found to be 1.45. 2. Fragmentation: Occupations where AI-suitable tasks are highly dispersed show lower levels of actual AI execution. This confirms that "exposure" to AI is a poor predictor of its impact if you ignore the sequence of the workflow. 3. Adjacency: A task is more likely to be executed by AI if its immediate neighbors in the workflow are also handled by AI.

The Limits of the Framework

While the model provides a powerful new lens for labor economics, it has clear boundaries. The authors admit that their model treats "hand-off costs" as fixed. In reality, AI might actually make communication easier. This could potentially reduce the coordination costs between workers.

Furthermore, the empirical analysis relies on Claude usage as a proxy for general AI execution. While the authors argue this is a conservative estimate, it remains a single-tool window. Finally, the model focuses on the "long-run" ability of firms to reorganize. It does not account for the cultural or psychological resistance workers might have to radical changes in their job descriptions.

Figures from the paper