An AI coding assistant can help a software engineer produce more in an hour. But what happens next?
The company might deliver the same amount of software with fewer engineering hours—the same with less. It might use its existing engineering capacity to build more products—more with the same. Or it might redirect that capacity from feature implementation towards product management, architecture and testing.
The same productivity improvement can therefore lead to very different gains. Which one the company realises depends on how it responds.
In economics, the relationship between inputs and output is represented by a production function:
where Y is output, K is capital, L is labour and T represents the available production technology. AI changes that technology, which can change how productive labour and capital are and how the company chooses to combine them.
What the technology changes: technological bias
Technological bias describes which inputs a technology makes relatively more productive. Economists measure this by comparing the inputs’ marginal products: how much extra output comes from adding a little more of one input while keeping the others unchanged.
If a technology increases the productivity of labour more than capital, it is labour-biased. If it increases the productivity of capital more than labour, it is capital-biased. Here, capital means productive assets such as compute, GPUs, software systems and AI infrastructure, not simply money.
The key point is that technological bias describes what the technology makes possible, not what the company ultimately chooses to use. A labour-biased technology, for example, does not necessarily mean the company will employ more labour.
What the company chooses: factor use
Factor use describes how much labour, capital or another input the company actually uses after adopting a technology.
A technology is labour-saving if it leads the company to use less labour, and labour-using if it leads the company to use more.
But this depends on whether we mean labour used per unit of output or in total. A technology can reduce the labour required to produce each unit while total labour use moves in the opposite direction.
The substitution and scale effects
Two forces help explain this: the substitution effect and the scale effect.
The substitution effect describes how a company changes the mix of inputs used to produce a given amount of output. For example, AI may allow capital to perform work that previously required human time, reducing the labour required for each unit of output.
The scale effect describes what happens when lower costs make it worthwhile to produce more. The company may build more products, serve more customers or undertake activities that were previously uneconomic.
Whether total labour use rises or falls depends on which effect is stronger. If the scale effect outweighs the labour saved through substitution, a technology can be labour-saving per unit of output while still being labour-using in total.
How technological bias and factor use fit together
Technological bias and factor use answer different questions:
Technological bias: Which input does the technology make relatively more productive?
Factor use: Does the company end up using more or less of that input?
Applying the framework to generative AI
Consider a generative AI coding assistant that helps software engineers write, test and review code more efficiently.
Suppose it reduces the engineering time required per feature by 30%, but the lower cost of software development leads the company to double the number of features it builds. The company also uses more compute and AI infrastructure to support the tool.
Under these assumptions, the technology could be:
Labour-biased, if it raises the marginal product of engineering labour relative to capital.
Engineering-labour-saving per feature, because each feature requires fewer engineering hours.
Engineering-labour-using in total, because the company expands software production enough to use more engineering hours overall.
Capital-using in total, because it uses more compute and AI infrastructure.
These are not contradictions.
Applying the framework to agentic AI
Generative AI often helps people perform individual tasks. As AI systems become capable of completing longer sequences of work, however, their economic impact can extend from improving individual tasks to reorganising entire workflows.
An agentic system might choose a task, produce an output, test it and escalate uncertain decisions to a person. This makes it useful to distinguish between two types of labour:
Execution work: carrying out tasks such as writing code, testing, documentation and routine coordination.
Complementary work: setting objectives, exercising judgement, reviewing results, handling exceptions and taking responsibility for important decisions.
These are types of work, not fixed categories of workers. The same software engineer may do both.
Suppose an agentic system automates much of the execution work in a software workflow while allowing experienced engineers to oversee a larger amount of AI-generated activity. The company also invests in more compute and agent infrastructure.
Under these assumptions, agentic AI could be:
Biased away from execution labour and towards AI capital and complementary labour, if AI increases the relative productivity of compute and human judgement.
Execution-labour-saving, because less human execution time is needed for each workflow.
Complementary-labour-using, if the expansion in AI-enabled activity creates more demand for direction, evaluation and exception handling.
Capital-using, because the company uses more compute and agent infrastructure.
The important point is that agentic AI can affect different kinds of labour differently.
The important shift may be in the composition of work
The impact of AI is often framed as whether companies will use more or less labour. But the more important change may be what people spend their time doing.
AI may affect tasks within the same job differently. As some activities are delegated to AI, people can spend more time setting direction, applying judgement, resolving ambiguity and evaluating results.
Its economic impact may therefore come not only from doing existing work more cheaply, but from changing how work itself is organised.
What this means for AI and work
A productivity gain alone cannot tell us whether AI will increase or reduce the use of labour. We need to understand which inputs become more productive, how companies reorganise production, and how much additional activity becomes worthwhile as costs fall.
As AI progresses from assisting with individual tasks to carrying out longer workflows, one of its most important impacts may be less about reducing labour and more about expanding what people can achieve. Human work may shift towards the activities where people add the most value, with people playing a more active role in shaping, guiding and taking responsibility for what AI does.
AI disclosure: I use AI to help with structure and drafting. The ideas, judgments, and conclusions are my own, and I review and edit the final text before publication.